content
stringlengths 1
15.9M
|
|---|
\section{The fundamental exterior differential system}
\subsection{Introduction}
\label{sec:Introduction}
The notion of ${\mathrm{SU}}(2)$-structure was introduced by Conti and Salamon in \cite{ContiSalamon} and plays an important role in the theory of generalized Killing spinors. It consists of the geometrical data induced on any hypersurface of a real 6-dimensional manifold endowed with an integrable special-Hermitian or ${\mathrm{SU}}(3)$-structure.
${\mathrm{SU}}(2)$-structures become an independent notion on real 5-manifolds $N$. They are given by three 2-forms $\omega_1,\omega_2,\omega_3$ and a contact 1-form $\theta$ satisfying certain relations between them. These forms induce a Riemannian metric on $N$ and a canonical ${\mathrm{SU}}(2)$-structure on $\ker\theta$. The present article discovers a very useful result concerning the deduction of such metric. Theorem \ref{Teo_themetriconLperp} gives the following identity for the metric
$g_{_{{\mathrm{SU}}(2)}}$ on $\ker\theta$, where $v$ is the volume form ($v=\frac{1}{2}\omega_i\wedge\omega_i$, $\forall i$), which indeed seems to be new:
\begin{equation}\label{themetriconLperp_introduction}
x\lrcorner\omega_1\wedge y\lrcorner\omega_2\wedge\omega_3=\gsudois{x}{y}\,v, \ \ \ \forall x,y\in \ker\theta.
\end{equation}
For hypersurfaces $N$, the induced ${\mathrm{SU}}(2)$-structure is hypo, i.e. satisfies the equations
\begin{equation}
{\mathrm{d}}\omega_1=0,\qquad{\mathrm{d}}(\theta\wedge\omega_2)=0,\qquad{\mathrm{d}}(\theta\wedge\omega_3)=0.
\end{equation}
Conti and Salamon prove the `embeding property', which is almost a reciprocal: an \textit{analytic} ${\mathrm{SU}}(2)$-structure satisfying the hypo system admits an embedding into an integrable special Hermitian manifold. This may eventually be compact, hence a Calabi-Yau 3-fold.
The hypo, the nearly-hypo, and other particular differential systems, imply interesting geometry on $N$. For instance, one easily meets with Sasaki-Einstein manifolds.
This article studies the question raised in \cite{Alb2015a} regarding a certain ${\mathrm{SU}}(2)$-structure defined on the total space $\cals$ of the tangent sphere bundle of a given oriented Riemannian 3-manifold $M$. This shall be referred as the \textit{main example}. We generalise the construction with what we call the natural structures, supported by the fundamental exterior differential system of Riemannian geometry introduced in \cite{Alb2011arxiv} and \cite{Alb2015a}. A classification of ${\mathrm{SU}}(2)$-structures according to first derivatives, which we then follow, was developed in \cite{BedulliVezzoni} and \cite{FIMU}, two references well acquainted with the foundational article of Conti and Salamon.
The exterior differential system discovered in \cite{Alb2011arxiv} depends only on the orientation and the metric on $M$. It consists, in general, of a natural contact 1-form $\theta$ and set of natural differential $n$-forms $\alpha_0,\ldots,\alpha_n$ existing always on the total space $\cals$ of the (unit) tangent sphere bundle $SM\longrightarrow M$ of any given oriented Riemannian $n+1$-manifold $M$. Of course, $\cals$ inherits the induced metric from the well-known canonical or Sasaki metric on $TM$ (not to be confused with Sasakian or Sasaki-Einstein metrics below). The metric plays a central role in defining the $\alpha_0,\ldots,\alpha_n$. The compatible contact structure $\theta$ is due to Tashiro. Applications of the natu\-ral differential system are discussed in \cite{Alb2011arxiv}.
Here we shall consider just the case $n=2$, so that $\cals$ is a 5-dimensional manifold. The fundamental differential system brings up four pairwise-orthogonal 2-forms $\alpha_0,\alpha_1,\alpha_2,{\mathrm{d}}\theta$, sa\-tis\-fying:
\begin{equation}\label{introd_derivadasdastres2formas}
\begin{split}
{\mathrm{d}}\alpha_1=2\,\theta\wedge\alpha_2-r\,\theta\wedge\alpha_0 , \hspace{12mm}\: \\
*\theta=\alpha_0\wedge\alpha_2=-\frac{1}{2}\,\alpha_1\wedge\alpha_1=-\frac{1}{2}\,{\mathrm{d}}\theta\wedge{\mathrm{d}}\theta , \\
{\mathrm{d}}\alpha_0=\theta\wedge\alpha_1, \:\qquad\quad
{\mathrm{d}}\alpha_2 = \calri\alpha_2 . \hspace{6mm}
\end{split}
\end{equation}
The function $r=r(u)={\mathrm{Ric}\,}(u,u),\ u\in\cals$, and the 3-form $\calri\alpha_2$ are curvature dependent tensors. For constant sectional curvature $K$ we have $r=2K$ and $\calri\alpha_2=-K\,\theta\wedge\alpha_1$.
In dimension 3 we have the \textit{nice} coincidence that the $\alpha_i$ are 2-forms like ${\mathrm{d}}\theta$, and then an ${\mathrm{SU}}(2)$-structure naturally takes place. The \textit{main example} is
\begin{equation}
\omega_1={\mathrm{d}}\theta,\qquad\omega_2=\alpha_2-\alpha_0,\qquad\omega_3=\alpha_1
\end{equation}
but many other linear combinations give interesting structures as well. Two distinct types appear with different properties. The distinction seems to be chiefly between those for which ${\mathrm{d}}\theta$ is in the linear span of the $\omega_1,\omega_2,\omega_3$, and those for which it is out. Our further results here concentrate more on the first type.
In \cite{FIMU}, we see that Fern\'andez, Ivanov, Mu\~noz and Ugarte also discovered ${\mathrm{SU}}(2)$-structures on $\cals$ for $M=S^3$, which is the Stiefel-manifold $V_{4,2}$. Our coordinate-free tools lead us to generalise mildly one of those results and also to rediscover the Sasaki-Einstein well-known metrics. More important, \cite{FIMU} introduces the nearly-hypo and double-hypo structures, which have very deep relations with nearly-K\"ahler manifolds and half-flat ${\mathrm{SU}}(3)$-structures. As it is well-known, the latter yield true ${\mathrm{G}_2}$-manifolds. Since we have found below new families of double-hypo structures associated to hyperbolic base $M$, they should lead to interesting results inspired by \cite{FIMU}. New developments from our construction and technique shall be continued in the near future.
Finally, we recall the evolution equations, again due to Conti and Salamon, and solve them in one particular family of hypo manifolds. This leads us to a new integrable ${\mathrm{SU}}(3)$-structure, associated to any given flat base $M$, defined on the space $\cals\times\R_+$. Which is not a trivial analytic manifold.
\subsection{The differential system}
\label{sec:Thedifferentialsystem}
We briefly recall the theory from \cite{Alb2011arxiv}. Let $M$ denote any oriented $n+1$-dimensional Riemannian manifold. Then the space $\calt_M$, i.e. the total space of the vector bundle $\pi:TM\longrightarrow M$, is well-known to be a smooth manifold of dimension $2n+2$. In natural coordinates, we may identify $V:=\ker{\mathrm{d}}\pi$ with $\pi^\star TM$, the tangent to the fibres. Taking the Levi-Civita connection ${\nabla}:\Gamma(M;TM)\longrightarrow\Gamma(M;T^*M\otimes TM)$, we get the canonical decomposition of $T{\calt_M}=H\oplus V\simeq\pi^*TM\oplus\pi^\star TM$. The connection dependent horizontal distribution $H$ identifies again with $\pi^*TM$ via ${\mathrm{d}}\pi$. Hence there exists a vector bundle endomorphism $B:T\calt_M\longrightarrow T\calt_M$ which sends horizontals to verticals and verticals to 0; it is called the \textit{mirror} map. Most important is that $B$ is parallel for the pull-back connection ${\nabla}^*$ by construction. We let $B^{\mathrm{t}}$ denote the adjoint of $B$.
There are two canonical vector fields on $\calt_M$. The first is the tautological vertical vector field $U$, defined by $U_u=u,\ \forall u\in \calt_M$; it is the independent mirror of the second, the geodesic spray, defined on the horizontal distribution and hence connection dependent. Any given frame in $H$, followed by its mirror in $V$, clearly determines a unique orientation on the manifold $\calt_M$.
We recall the map $J=B-B^{\mathrm{t}}$ gives the well-known canonical or Sasaki almost complex structure on $\calt_M$.
Next we consider the well-known canonical or Sasaki metric on the $2n+2$-manifold $\calt_M$. The mirror map becomes an isometry. Any frame at point $u$ arising from an orthonormal frame in $H$ with the first vector equal to $B^tU/\|U\|$, together with the mirror frame in $V$, in fixed order `first $H$, then $V$', is said to be an \textit{adapted frame} of $\calt_M$.
We hence find that $\calt_M\backslash$(zero section) has structure group the Lie group ${\mathrm{SO}}(n)$, cf. \cite[Theorem 1.1]{Alb2015a}. A representation of ${\mathrm{SO}}(n)$, acting diagonally, occurs on the common orthogonal distribution to the geodesic spray and to $U$. On these two directions, the action is of course trivial.
Now we consider the constant radius $s$ tangent sphere bundle of $M$
\begin{equation}
{S}_s{M}=\{u\in TM:\ \|u\|=s\}.
\end{equation}
We let $\cals=\cals_{s,M}$ denote the total space of $S_sM$. We have
$T{\cals}=U^\perp\subset T\calt_M$, because $\ker{\nabla}^*_\cdot U=H$ and ${\nabla}^*_vU=v,\ \forall v\in V$. In particular, this manifold is orientable. The Riemannian submanifold $\cals$ inherits the ${\mathrm{SO}}(n)$-structure, which however is never parallel because $U$ is not parallel.
From the above remarks we have that any orthonormal frame $u,e_1,\ldots,e_n$ on $M$ induces by horizontal and vertical lifts an adapted frame $e_0,e_1,\ldots,e_n,e_{n+1},\ldots,e_{2n}\in T_u\cals$ at point $u\in\cals$, where $e_0=\frac{1}{s}B^{\mathrm{t}}U_u\in H_u$.
We denote by $\theta$ the 1-form on $\cals$ defined as
\begin{equation}\label{thecontactform}
\theta=\langle U,B\,\cdot\,\rangle=s\,e^0.
\end{equation}
It is well-known that $\theta$ and $J$ define a metric contact structure on $\cals$. We also recall the result ${\mathrm{d}}\theta=e^{(1+n)1}+\cdots+e^{(2n)n}$ (from our usual notation: $e^{ij}=e^i\wedge e^j$ and this has norm 1). This is reminiscent of the Liouville form on $T^*M$, and one sees the amazing fact that ${\mathrm{d}}\theta$ no longer depends on $s$.
The ${\mathrm{SO}}(n)$-structure induces the following natural fundamental differential system discovered in \cite{Alb2011arxiv} of global $n$-forms $\alpha_0,\alpha_1,\ldots,\alpha_n$ on $\cals$.
We first write $\pi^\star{\mathrm{vol}}_{_M}$ for the vertical lift of the volume form of $M$ (always a $\pi^\star$ denotes a vertical lift). Then
\begin{equation}\label{alphazero}
\alpha_n = \frac{1}{s}\,U\lrcorner({\pi}^\star{\mathrm{vol}}_{_M})
\end{equation}
and for each $0\leq i\leq n$ we define, $\forall v_1,\ldots,v_n\in T\cals$,
\begin{equation}\label{alpha_itravez}
\alpha_i(v_1,\ldots,v_n) = \frac{1}{i!(n-i)!}\sum_{\sigma\in\mathrm{Sym}(n)}\mathrm{sg}(\sigma)\, \alpha_n(Bv_{\sigma_1},\ldots,B v_{\sigma_{n-i}},v_{\sigma_{n-i+1}},\ldots,v_{\sigma_n}).
\end{equation}
For convenience one also writes $\alpha_{-1}=\alpha_{n+1}=0$; we use the notation
\begin{equation}
R_{lkij} = \langle R^{\nabla}(e_i,e_j)e_k,e_l\rangle=
\langle {\nabla}_{e_i}{\nabla}_{e_j}e_k-{\nabla}_{e_j}{\nabla}_{e_i}e_k-{\nabla}_{[e_i,e_j]}e_k,e_l\rangle.
\end{equation}
\begin{teo}[1st-order structure equations, \cite{Alb2011arxiv}] \label{derivadasdasnforms}
We have
\begin{equation}\label{dalphai}
{\mathrm{d}}\alpha_i=\frac{1}{s^2}(i+1)\,\theta\wedge\alpha_{i+1}+\calri\alpha_i
\end{equation}
where
\begin{equation}\label{Ralphai}
\calri\alpha_i = \sum_{0\leq j<q\leq n}\sum_{p=1}^nsR_{p0jq}\,e^{jq}\wedge
e_{p+n}\lrcorner\alpha_i.
\end{equation}
\end{teo}
Defining $r=\frac{1}{s^2}\pi^\star{\mathrm{Ric}\,}(U,U)=\sum_{j=1}^nR_{j0j0}$, a smooth function on $\cals$ determined by the Ricci curvature of $M$, we find that $\calri\alpha_0=0$ and $\calri\alpha_{1}=-r\,\theta\wedge\alpha_0$. This is
\begin{equation}\label{dalphanen-1}
{\mathrm{d}}\alpha_0=\frac{1}{s^2}\,\theta\wedge\alpha_{1},\qquad\quad
{\mathrm{d}}\alpha_{1}=\frac{2}{s^2}\,\theta\wedge\alpha_{2}-sr\,{\mathrm{vol}}.
\end{equation}
Moreover, the differential forms \,$\theta$,\ $\alpha_n$ and $\alpha_{n-1}$ are always coclosed. In every degree we have
\begin{equation}
\alpha_i\wedge{\mathrm{d}}\theta=0.
\end{equation}
No further assumptions are required, besides orientation and a metric, in order to find the fundamental exterior differential system $\{\theta,\alpha_0,\ldots,\alpha_n\}\Omega^*_\cals$ associated to a given oriented Riemannian manifold.
\subsection{The 3d differential system}
\label{T3dc}
We now consider a 3-dimensional $M$ together with the total space $\cals$ of the tangent 2-sphere bundle of radius $s$ equipped with canonical metric and orientation. We have the contact 1-form, $\theta=s\,e^0$, clearly invariant for the action of ${\mathrm{SO}}(2)$ on $\R^{1+2+2}$, i.e. the trivial action on the 1-dimensional summand and the diagonal action on $\R^{2+2}$.
The global invariant 2-forms, independent of the choice of adapted frame, are
\begin{equation}\label{thefourinvariants}
\alpha_0=e^{12},\qquad \alpha_1=e^{14}-e^{23},\qquad \alpha_2=e^{34},\qquad{\mathrm{d}}\theta=e^{31}+e^{42}.
\end{equation}
We also have
\begin{equation}
\alpha_0\wedge\alpha_1=\alpha_2\wedge\alpha_1=\alpha_i\wedge{\mathrm{d}}\theta=0,\:\ \forall i=0,1,2,
\end{equation}
and
\begin{equation}\label{basicstructurequations}
\frac{1}{s}\,*\theta=\alpha_0\wedge\alpha_2=-\frac{1}{2}\,\alpha_1\wedge\alpha_1=-\frac{1}{2}\,{\mathrm{d}}\theta\wedge{\mathrm{d}}\theta.
\end{equation}
\begin{prop}[\cite{Alb2015a}]\label{prop_decompwedgetwo}
The representation under ${\mathrm{SO}}(2)$ above, induced on the vector bundle $\Lambda^2T^*\cals$, corresponds with the decomposition
\begin{equation}\label{decompwedgetwo}
\Lambda^2\R^5=4\R^1\oplus W_1\oplus W_2\oplus W_3
\end{equation}
where we have the four 1-dimensional invariants from \eqref{thefourinvariants} and three irreducible orthogonal subspaces $W_i$ defined by
\begin{equation}
W_1=\llbracket e^{01},e^{02}\rrbracket, \qquad W_2=\llbracket e^{03},e^{04}\rrbracket,\qquad
W_3=\llbracket \psi_1,\psi_2\rrbracket
\end{equation}
where
\begin{equation}\label{The_f_forms}
\psi_1:=e^{14}+e^{23},\qquad \psi_2:=e^{31}-e^{42}.
\end{equation}
\end{prop}
Other $\Lambda^pT^*\cals$ are easily decomposed. Since the canonical map $\Lambda^1\R^5\otimes\Lambda^2\R^5\longrightarrow\Lambda^3\R^5$ has a kernel of dimension 40, there are many equivalent representations in the space of 3-forms.
The scalar function $r=\frac{1}{s^2}\pi^\star{\mathrm{Ric}\,}(U,U)=R_{1010}+R_{2020}$ (recall 0 stands for the point $u\in\cals$) may be written using scalar and sectional curvatures as $r=\frac{1}{2}{\mathrm{scal}\,}-K(\{e_1,e_2\})=\frac{1}{2}{\mathrm{scal}\,}-R_{1212}$. If $M$ is Einstein, this is ${\mathrm{Ric}\,}=\lambda\langle\cdot,\cdot\rangle$ for some constant $\lambda$, then clearly $M$ has constant sectional curvature $\lambda/2$.
Now recall we have the Sasaki almost complex structure $J$ on $H_0\oplus V_0$, where $H_0=H\cap e_0^\perp=\llbracket e_1,e_2\rrbracket$ and $V_0=V\cap U^\perp=\llbracket e_3,e_4\rrbracket$ are sub-vector bundles of $T\cals$. We may further define $I_+$ and $I_-$, according to $\pm$, to be the unique map defined on any adapted frame as
\begin{equation}\label{segundaestruturacomplexa}
e_0\mapsto 0,\qquad e_1\mapsto e_2 \mapsto-e_1,\qquad
e_3\mapsto \pm e_4 \mapsto -e_3.
\end{equation}
$I_+,I_-$ are commuting endomorphisms of $T\cals$. On one hand,
$JI_+J^t=JI_+\inv{J}=I_+$. On the other, we have that $J$ and $I_-$ anti-commute, giving an $\mathrm{Sp}(1)={\mathrm{SU}}(2)$-structure in the sense of Conti-Salamon, as noticed in \cite{Alb2015a}. It is to these and other similar structures that this article is devoted.
The following 1-form is an important irreducible tensor:
\begin{equation}
\rho=\frac{1}{s}\,U\lrcorner\pi^\star{\mathrm{Ric}\,}=R_{1012}e^4-R_{2012}e^3.
\end{equation}
As complex line bundles, $H_0$ and $V_0$ are very particular to dimension 3. $V_0$ is the holomorphic tangent bundle when restricted to each fibre, $S^2$, with $\alpha_2$ restricting to the K\"ahler class. We have global 1-forms defined by
\begin{equation}\label{osrhostodos}
\begin{split}
\rho =R_{1012}e^4-R_{2012}e^3 , & \\
\rho_1=\rho B =R_{1012}e^2-R_{2012}e^1 , & \\
\rho_2=\rho I_+B =R_{1012}e^1+R_{2012}e^2 , & \\
\rho_3=\rho I_+ =R_{1012}e^3+R_{2012}e^4 . &
\end{split}
\end{equation}
Now, regarding exterior derivatives, from the general formulae in \eqref{dalphanen-1} and recalling $r=R_{1010}+R_{2020}$, we have
\begin{equation}
{\mathrm{d}}\alpha_0 = \frac{1}{s^2}\,\theta\wedge\alpha_1, \label{derivadasdastres2formas_alpha2}
\end{equation}
\begin{equation}
{\mathrm{d}}\alpha_1 = \frac{2}{s^2}\,\theta\wedge\alpha_2-r\,\theta\wedge\alpha_0 .
\label{derivadasdastres2formas_alpha1}
\end{equation}
These are already decomposed into irreducibles. From \cite[Theorem 2.2]{Alb2015a} we have
\begin{equation}\label{dalpha0composto}
{\mathrm{d}}\alpha_2 = \theta\wedge\gamma-\frac{r}{2}\,\theta\wedge\alpha_1+s\,\alpha_0\wedge\rho \quad \in\quad *W_3\oplus\llbracket*\alpha_1\rrbracket\oplus*W_2
\end{equation}
where, by \eqref{The_f_forms}, the 2-form $\gamma$ is defined as
\begin{equation}
\gamma:=R_{1002}\psi_2+\frac{1}{2}(R_{1001}-R_{2002})\psi_1\ \ \in\ W_3.
\end{equation}
The following result shall play a relevant role later on.
\begin{prop}[\cite{Alb2015a}]\label{casocsc}
The following assertions are equivalent on a connected 3-manifold: $M$ has constant sectional curvature; $r$ is constant; $\rho=0$; $\gamma=0$; ${\mathrm{d}}\alpha_2=-\frac{r}{2}\,\theta\wedge\alpha_1$.
\end{prop}
\subsection{Conti-Salamon structures and hypo and nearly-hypo 5-manifolds}
\label{Hanh5m}
${\mathrm{SU}}(2)$-structures on 5 dimensions are understood as the induced metric structures on real hypersurfaces of ${\mathrm{SU}}(3)$ manifolds.
Let $S$ denote any 5-dimensional manifold. It is said that $S$ is endowed with an \textit{${\mathrm{SU}}(2)$-structure} if its frame bundle admits a reduction to ${\mathrm{SU}}(2)=\mathrm{Sp}(1)$ via the canonical plus trivial representation in $\C^2\oplus\R$. Then there exists on $S$ an orientation and a metric such that $TS=L\oplus L^\perp$, where $L\subset TS$ is a real line bundle and $L^\perp$ is endowed with a metric compatible quaternionic structure. The concept was first introduced by D.~Conti and S.~Salamon in \cite{ContiSalamon}. Due to a canonical inclusion of ${\mathrm{SU}}(2)$ in ${\mathrm{SO}}(5)$ and lift into ${\mathrm{Spin}}(5)$, the manifold $S$ must be orientable and spin. ${\mathrm{SU}}(2)$-structures on $S$ are in one-to-one correspondence with pairs of spin structures and unit spinors.
Still following \cite{ContiSalamon}, an ${\mathrm{SU}}(2)$-structure is defined by a 1-form ${\theta}$, such that $L^\perp=\ker{\theta}$, and three 2-forms $\omega_1,\omega_2,\omega_3$ on $S$ such that
\begin{equation}\label{hypostruceq1}
\begin{split}
&\qquad\qquad\quad {\theta}\wedge\omega_1\wedge\omega_1\neq0,\\
&\qquad\quad\quad \omega_i\wedge\omega_j=0,\ \forall i\neq j, \\
& 2v\stackrel{\mathrm{def.}}= \omega_1\wedge\omega_1=\omega_2\wedge\omega_2= \omega_3\wedge\omega_3\neq0
\end{split}
\end{equation}
and
\begin{equation}\label{hypostruceq2}
x\lrcorner\omega_1=y\lrcorner\omega_2\ \Longrightarrow\ \omega_3(x,y)\geq0,\ \ \forall x,y\in TS.
\end{equation}
We remark the system is verified under multiplication by any $\exp{(t\sqrt{-1})},\ t\in\R$, either on $\omega_1+\sqrt{-1}\omega_2$ or on $\omega_2+\sqrt{-1}\omega_3$.
One finds three almost complex structures $\Phi_i$ on $L^\perp$ compatible and positively tame by the respective $\omega_i$, yet inducing a unique positive definite metric
\begin{equation}\label{themetriconLperpintheory}
\gsudois{x}{y}=\omega_i(x,\Phi_iy),\ \forall x,y\in L^\perp.
\end{equation}
Now we have a linear algebra result, which gives a formula for the induced metric without finding any $\Phi_i$.
\begin{teo}\label{Teo_themetriconLperp}
A system $({\theta},\omega_1,\omega_2,\omega_3)$ defines an ${\mathrm{SU}}(2)$-structure on $S$ if and only if it satisfies \eqref{hypostruceq1} and the bilinear map $g_{_{{\mathrm{SU}}(2)}}$ on $L^\perp=\ker{\theta}$ given by
\begin{equation}\label{themetriconLperp}
x\lrcorner\omega_1\wedge y\lrcorner\omega_2\wedge\omega_3=\gsudois{x}{y}\,v, \ \ \ \forall x,y\in L^\perp,
\end{equation}
is positive definite.
\end{teo}
\begin{proof}
We first prove the condition is necessary. By \cite[Corollary 1.3]{ContiSalamon} we see the triplet of the $\omega_i$ forms a frame of self-dual 2-forms of the pair $L^\perp,v$, i.e. there exists an orthonormal frame such that $\omega_1=e^{12}+e^{34},\ \omega_2=e^{13}+e^{42},\ \omega_3=e^{14}+e^{23}$. Writing $x=\sum_ix_ie_i$ and $y=\sum_iy_ie_i$, we have $x\lrcorner\omega_1=x_1e^2-x_2e^1+x_3e^4-x_4e^3$ and $y\lrcorner\omega_2=y_1e^3-y_3e^1+y_4e^2-y_2e^4$ and hence
\begin{eqnarray*}
x\lrcorner\omega_1\wedge y\lrcorner\omega_2 &=& x_1y_1e^{23}+x_1y_3e^{12}-x_1y_2e^{24}-x_2y_1e^{13}-x_2y_4e^{12}+x_2y_2e^{14} \\
& & -x_3y_1e^{34}+x_3y_3e^{14}-x_3y_4e^{24}-x_4y_3e^{13}+x_4y_4e^{23}+x_4y_2e^{34} .
\end{eqnarray*}
The identity of the given bilinear map with the metric follows:
\begin{eqnarray*}
x\lrcorner\omega_1\wedge y\lrcorner\omega_2\wedge\omega_3 &=& (x_1y_1 +x_2y_2+x_3y_3+x_4y_4)e^{1234} .
\end{eqnarray*}
Now let us prove the condition is sufficient. Given the 1- and 2-forms satisfying \eqref{hypostruceq1}, we see there exists a four dimensional sub-vector bundle, say $L^\perp\subset TS$, on which $v$ is non-degenerate. By hypothesis, after symmetrizing, we have a positive definite metric on $L^\perp$, so it is a matter of counting dimensions to see this is uniquely determined. Since we have a volume 4-form, we know a priori that the reduction is in 12 dimensions, from the structure group $\mathrm{SL(4)}$ to ${\mathrm{SU}}(2)$. Now the three 2-forms are written $\omega_i=\sum_{1\leq j<k\leq 4}\omega_{ijk}e^{jk}$. With \eqref{hypostruceq1} we find the $18-6=12$ dimensions. On the other hand, in order to find an orthonormal frame $e_1,\ldots,e_4$, with which one proves the 2-forms to be self-dual for the metric in \eqref{themetriconLperp}, we solve $4+3+2+1=10$ equations. So it is possible to solve these equations, leaving 2 dimensions free due to the remark above.
\end{proof}
Thus the open condition \eqref{hypostruceq2} stands for some choice of ordering of $\omega_1,\omega_2,\omega_3$.
Well understood, we assume $\|{\theta}\|_{{\mathrm{SU}}(2)}=1$ and, on the other hand, that any non-vanishing multiple of ${\theta}$ will define another ${\mathrm{SU}}(2)$-structure.
The first property of such a metric is the relation with ${\mathrm{SU}}(3)$ real submanifold geometry described in the founding article. The above notion is again equivalent to a real, oriented 5-manifold $S$, endowed with a 1-form ${\theta}$, a 2-form $\omega_1$ and a complex 2-form $\phi$, corresponding to $\omega_2+\sqrt{-1}\omega_3$, which is type $(2,0)$ for $\omega_1$, cf. \cite{ContiSalamon}, and which satisfies
\begin{equation}\label{hypostruceq3}
\theta\wedge\omega_1\wedge\omega_1\neq0,\qquad\omega_1\wedge\phi=0,\qquad\phi\wedge\phi=0,\qquad2\omega_1\wedge\omega_1=\phi\wedge\overline{\phi}.
\end{equation}
Reciprocally, since $S$ is oriented, the corresponding ${\mathrm{SU}}(3)$-structure on $S\times\R$ follows as the pair of a real symplectic 2-form $\omega_1+\theta\wedge{\mathrm{d}} t$ and a complex volume 3-form $\phi\wedge(\theta+\sqrt{-1}{\mathrm{d}} t)$.
Let us now recall some further developments from \cite{BedulliVezzoni,CFS,ContiSalamon,dAFFU,FIMU} on the theory of ${\mathrm{SU}}(2)$-structures. Conserved tensors may appear, leading to the characterization of some special Riemannian geometries. Such is the case of \textit{hypo} structures, considered first in \cite{ContiSalamon}:
\begin{equation}\label{eqhypo}
{\mathrm{d}}\omega_1=0,\qquad{\mathrm{d}}(\theta\wedge\omega_2)=0,\qquad{\mathrm{d}}(\theta\wedge\omega_3)=0.
\end{equation}
There it is proved that hypo structures are precisely the ${\mathrm{SU}}(2)$-structures which are induced on a real analytic hypersurface from a complex 3-manifold endowed with an integrable ${\mathrm{SU}}(3)$-structure; `precisely' meaning that any real analytic ${\mathrm{SU}}(2)$-structure satisfying \eqref{eqhypo} arises from such a 3-fold.
\nopagebreak
\begin{figure}
\begin{center}
\includegraphics[height=0.19\textheight]{Hypo2.png}
\end{center}
\end{figure}
Another type of ${\mathrm{SU}}(2)$-structure is considered in \cite{FIMU}. The \textit{nearly-hypo} structures on a 5-manifold $S$ are defined by
\begin{equation}\label{eqnearlyhypo}
{\mathrm{d}}\omega_2=3\theta\wedge\omega_3,\qquad {\mathrm{d}}(\theta\wedge\omega_1)=-2\omega_1\wedge\omega_1.
\end{equation}
Nearly-hypo structures give rise to a general construction of nearly-K\"ahler structures on $S\times\R$, cf. \cite{FIMU}. Structures which are both hypo and nearly-hypo are called \textit{double-hypo}:
\begin{equation}\label{eqdoublehypo}
{\mathrm{d}}\omega_1=0,\qquad{\mathrm{d}}\omega_2=3\theta\wedge\omega_3,\qquad {\mathrm{d}}(\theta\wedge\omega_1)=-2\omega_1\wedge\omega_1,\qquad{\mathrm{d}}(\theta\wedge\omega_2)=0.
\end{equation}
They contain a smaller subset given by the Sasaki-Einstein 5-manifolds, i.e.
\begin{equation}\label{eqSasakiEinstein}
{\mathrm{d}}\theta=-2\omega_1,\qquad{\mathrm{d}}\omega_2=3\theta\wedge\omega_3,\qquad {\mathrm{d}}\omega_3=-3\theta\wedge\omega_2.
\end{equation}
In this case the respective ${\mathrm{SU}}(3)$-fold is a K\"ahler-Einstein manifold.
Many non-trivial examples of the above special geometries are given on products of spheres and Lie groups in \cite{FIMU}. Examples on nilmanifolds are already constructed in \cite{ContiSalamon}.
A fifth special geometry is considered and studied in \cite{BedulliVezzoni,dAFFU}: the \textit{contact-hypo} structures are defined by
\begin{equation}\label{eqcontacthypo}
{\mathrm{d}}\theta=-2\omega_1,\qquad{\mathrm{d}}(\theta\wedge\omega_2)=0,\qquad{\mathrm{d}}(\theta\wedge\omega_3)=0.
\end{equation}
Clearly, contact-hypo are hypo and contain the Sasaki-Einstein structures.
We would also consider the \textit{contact-nearly-hypo} structures as those which satisfy merely ${\mathrm{d}}\theta=-2\omega_1$ and ${\mathrm{d}}\omega_2=3\theta\wedge\omega_3$. However, these consist of the intersection of double-hypo and contact-hypo structures. For instance we see
\begin{eqnarray*}
{\mathrm{d}}(\theta\wedge\omega_2)&=&{\mathrm{d}}\theta\wedge\omega_2-\theta\wedge{\mathrm{d}}\omega_2 \\
&=&-\frac{1}{2}\omega_1\wedge\omega_2-\theta\wedge\theta\wedge\omega_3 \ =\ 0.
\end{eqnarray*}
Last but not least, the invariance of equations under multiplication of $\phi$ by $\exp{(t\sqrt{-1})},\ t\in\R$, is verified in the cases of hypo, contact-hypo and Sasaki-Einstein structures.
\section{On the tangent sphere bundles of 3-manifolds}
\subsection{The natural structures}
\label{TnSU2s}
We consider again the setting from Section \ref{T3dc}, where it is given an oriented Riemannian 3-manifold $(M,g)$. We may then recall the fundamental exterior differential system defined on the associated Riemannian manifold $\cals=\cals_{s,M}$. The canonical or Sasaki metric on $\cals$, also denoted by $g$, is required by the differential system, and so it shall keep its main role in the following and be referred as the \textit{canonical Sasaki metric}. We remark the natural transformations below shall lead to truly non-trivial variations of the canonical metric, cf. Section \ref{Tme}.
Since we are interested in the contact and, in particular, the Sasakian geometry of $\cals$, we shall give the name \textit{canonical Tashiro metric} or \textit{structure} to the almost contact metric and contact structure defined on $\cals$ by
\begin{equation}\label{Tashirometric1}
\check{g}=\frac{1}{4s^2}g,\quad\eta=-\frac{1}{2s^2}\theta,\quad\Phi=B^t-B+\frac{1}{s^2}U\otimes\theta,\quad \xi=-{2}B^tU .
\end{equation}
Of course, $\theta$ comes from \eqref{thecontactform}. This structure $(\check{g},\eta,\Phi,\xi)$ is not quite the so-called standard structure, cf. \cite[Section 9]{Blair} or \cite{Calva}, but is also convenient for many reasons. We have the canonical orientation
\begin{equation}\label{orientation1}
e^{01234}\simeq\eta\wedge({\mathrm{d}}\eta)^2 ,
\end{equation}
\begin{equation}\label{Tashirometric2}
\eta(\xi)=1,\qquad \eta=\xi\lrcorner\check{g}, \qquad \Phi^2=-1_{{T\cals}}+\xi\otimes\eta,
\end{equation}
\begin{equation}\label{Tashirometric3}
\check{g}=\frac{1}{2}{\mathrm{d}}\eta(\Phi\ ,\ ),\qquad \check{g}(\Phi\ ,\Phi\ )= \check{g}-\eta\otimes\eta
\end{equation}
and a reciprocal curious result.
\begin{prop}\label{Prop_Tashirometric}
If $\eta=-p\theta$ for some $p\neq0$ and we are to have (\ref{orientation1}--\ref{Tashirometric3}), then $p=1/2s^2$ and the almost contact metric structure is given by \eqref{Tashirometric1}.
\end{prop}
\begin{proof}
Indeed the conditions are the required for an almost contact structure, cf. \cite{Boyer1,Boyer2}. Recalling the non-linear property of the geodesic 1-form $\theta$, which, for every radius $s$, satisfies ${\mathrm{d}}\theta=g((B^t-B)\otimes 1)$, the result follows by simple computations.
\end{proof}
Notice the previous results are valid in any dimension. A classical result of Tashiro proves $\check{g}$ is Sasakian if and only if $M$ has constant sectional curvature $\frac{1}{s^2}$.
We may also deform the almost contact metric structure along $\xi$, i.e. taking for Reeb vector field a multiple $\lambda$ of $\xi$ different of that for $g$ in $\xi^\perp=H_0\oplus V_0$. It is known that for certain values of $\lambda$ this metric is Sasaki-Einstein. This phenomena shall appear below.
Finally we are in the right moment to recall the main purpose of this article. That is, to study the ${\mathrm{SU}}(2)$-structures on $\cals$ induced by the differential system $\theta,\alpha_0,\alpha_1,\alpha_2$.
First of all we are led to define the 1-form $\tilde{\theta}$ from \eqref{hypostruceq1} as a multiple of the canonical contact 1-form $\theta$; secondly we define the three 2-forms $\omega_1,\omega_2,\omega_3$ as the linear combinations
\begin{equation}\label{definitionofinvariantSU2}
\begin{split}
\omega_1 &=a_0\alpha_0+a_1\alpha_1+a_2\alpha_2+a_3{\mathrm{d}}\theta \\
\omega_2 &=b_0\alpha_0+b_1\alpha_1+b_2\alpha_2+b_3{\mathrm{d}}\theta \\
\omega_3 &=c_0\alpha_0+c_1\alpha_1+c_2\alpha_2+c_3{\mathrm{d}}\theta
\end{split}
\end{equation}
where $a_0,\ldots,a_3,b_0,\ldots,b_3,c_0,\ldots,c_3$ are constant coefficients.
Notice we have $\|\theta\|=s$, but always $\|\tilde{\theta}\|_{{\mathrm{SU}}(2)}=1$. Also, due to \eqref{basicstructurequations}, we have
\begin{equation}
\omega_1\wedge\omega_1=(a_1^2+a_3^2-a_0a_2)\,{\mathrm{d}}\theta\wedge{\mathrm{d}}\theta,
\end{equation}
\begin{equation}
\omega_1\wedge\omega_2=(a_1b_1+a_3b_3-\frac{1}{2}a_0b_2-\frac{1}{2}a_2b_0)\,{\mathrm{d}}\theta\wedge{\mathrm{d}}\theta
\end{equation}
and similar identities with the $a,b,c$'s. Therefore, by \eqref{hypostruceq1}, we must have
\begin{equation}\label{hypostruceq4}
a_1^2+a_3^2-a_0a_2=b_1^2+b_3^2-b_0b_2=c_1^2+c_3^2-c_0c_2\neq0
\end{equation}
and
\begin{equation}\label{hypostruceq5}
\begin{split}
& a_0b_2+a_2b_0-2a_1b_1-2a_3b_3 = b_0c_2+b_2c_0-2b_1c_1-2b_3c_3 =\qquad \qquad\\
& \hspace{47mm} = c_0a_2+c_2a_0-2c_1a_1-2c_3a_3 =0.
\end{split}
\end{equation}
\begin{defi}
A set of differential forms $\tilde{\theta},\omega_1,\omega_2,\omega_3$ as the above defined, with constant coefficients and satisfying (\ref{hypostruceq1},\ref{hypostruceq2}), is called a \emph{natural ${\mathrm{SU}}(2)$-structure} on $\cals$.
\end{defi}
The natural ${\mathrm{SU}}(2)$-structures are natural variations of the Sasaki metric on tangent sphere bundles. They induce `g-natural' metrics in the sense of well-known references, such as \cite{Abb1,AbbKowal}, but they also give a new class of natural metrics in these 5-dimensional manifolds. Our metrics are indeed of a more general kind as we shall see in Proposition \ref{SU2metricsareofmoregeneralkind}.
\vspace{3mm}
\noindent
\textsc{Main example.}
\begin{itemize}
\item This \textit{main example} was first devised in \cite{Alb2015a}. The orientation on $\cals_{s,M}$ is induced by the ordering of any adapted coframe $e^0,e^1,\ldots,e^4$; but that on $\ker\theta$, corresponding to $v$, is $-e^{1234}$. The ${\mathrm{SU}}(2)$-structure is given by (${\mathrm{d}}\tilde{\theta}=-2\omega_1$)
\begin{equation}\label{mainexample}
\tilde{\theta}=-2\theta=-2s\,e^0,\qquad\omega_1={\mathrm{d}}\theta,
\qquad\omega_2=\alpha_2-\alpha_0,\qquad\omega_3=\alpha_1 .
\end{equation}
Indeed, recalling the theory of the intrinsic geometry of Riemannian 3-manifolds, we see that $-2\alpha_0\wedge\alpha_2=\alpha_1\wedge\alpha_1= {\mathrm{d}}\theta\wedge{\mathrm{d}}\theta=-2e^{1234}$. Below we shall prove the ${\mathrm{SU}}(2)$-metric is Sasaki-Einstein if and only if $M$ has constant sectional curvature $K=3$ and $s=\sqrt{3}/3=\sqrt{1/K}$. Also we shall see the induced metric coincides with the canonical metric on $\cals$ if and only if $s=\frac{1}{2}$ (just because of $\tilde{\theta}$). Hence it is not the Sasakian, Tashiro metric $\check{g}$ on $\cals_{s,S^3(s)}$ which is an Einstein metric. Finally we remark the choice of $\tilde{\theta}=\theta,\omega_1=-\frac{1}{2}{\mathrm{d}}\theta$, etc., seems equally keen in the search for hypo equations, but then we would miss the canonical Sasaki metric.
\end{itemize}
\vspace{3mm}
Notice the canonical Tashiro structure cannot be transformed homothetically into the structure of the \textit{main example}, as Proposition \ref{Prop_Tashirometric} shows, except for $s=\frac{1}{2}$.
We remark the general ${\mathrm{SU}}(2)$-structure remains invariant under isometries of $M$ lifted to isometries of the radius $s$ tangent sphere bundle total space $\cals$ with the canonical metric. Indeed, an adapted frame is transformed into an adapted frame. In particular, the new structures descend to a quotient space $\cals/\Gamma\longrightarrow M/\Gamma$ for any discrete subgroup $\Gamma\subset\mathrm{Isom}_+(M)$.
Let us now analyse one of the structural equations.
\begin{lemma}\label{lemadomega1iguala0}
Suppose ${\mathrm{d}}\omega_1=0$. Then:\\
(i) $\omega_1=a_3{\mathrm{d}}\theta$ or\\
(ii) $\omega_1=a_0\alpha_0+a_2\alpha_2+a_3{\mathrm{d}}\theta$, with $a_2\neq0$, and $M$ has constant sectional curvature $K=\frac{a_0}{a_2s^2}$.
\end{lemma}
\begin{proof}
Since ${\mathrm{d}}\alpha_0=\frac{1}{s^2}\,\theta\wedge\alpha_1$, ${\mathrm{d}}\alpha_1=\frac{2}{s^2}\,\theta\wedge\alpha_2-r\,\theta\wedge\alpha_0$ and ${\mathrm{d}}\alpha_2 = \theta\wedge\gamma-\frac{r}{2}\,\theta\wedge\alpha_1+s\,\alpha_0\wedge\rho$, it follows immediately from the hypothesis that either $a_2=0$ or $\rho=0$. Also we find $\frac{a_0}{s^2}\alpha_1+\frac{2a_1}{s^2}\alpha_2-ra_1\alpha_0+a_2\gamma-\frac{ra_2}{2}\alpha_1=0$. Knowing the representation subspaces, this implies $a_1=0$, $2a_0-ra_2s^2=0$ and $a_2\gamma=0$. Now if $a_2=0$, then $a_0=a_1=0$ and we are in case i. If $a_2\neq0$, then $\gamma=0$ and by Proposition \ref{casocsc} we have constant sectional curvature $K$ given by $a_0-Ka_2s^2=0$.
\end{proof}
Now let us study a second main equation, common to all five special ${\mathrm{SU}}(2)$-structures recalled in Section \ref{Hanh5m}. Indeed, $\theta\wedge\omega_3$ must always be closed. Let us consider real constants $c_0,\ldots,c_3$ and
\begin{equation}
\omega_3=c_0\alpha_0+c_1\alpha_1+c_2\alpha_2+c_3{\mathrm{d}}\theta.
\end{equation}
\begin{lemma}\label{lemadtetaomega3}
Suppose ${\mathrm{d}}(\theta\wedge\omega_3)=0$. Then:\\
(i) $\omega_3=c_0\alpha_0+c_1\alpha_1+c_2\alpha_2$, if $M$ has constant sectional curvature, \\
(ii) $\omega_3=c_0\alpha_0+c_1\alpha_1$, if $M$ has non-constant sectional curvature.
\end{lemma}
\begin{proof}
We have ${\mathrm{d}}(\theta\wedge\omega_3)={\mathrm{d}}\theta\wedge\omega_3-\theta\wedge(\sum c_j{\mathrm{d}}\alpha_j)$. From the first summand and the fundamental equations \eqref{basicstructurequations} it follows that $c_3=0$; the remaining summand gives the equivalent condition that $\rho=0$ or $c_2=0$. Applying again Proposition \ref{casocsc}, we have the result.
\end{proof}
Constant coefficients restrict the curvature on the base manifold.
\begin{prop}\label{Excludingnonconstantsectionalcurvature}
Suppose $M$ has non-constant sectional curvature. Then there do not exist natural hypo nor natural nearly-hypo structures on $\cals$.
\end{prop}
\begin{proof}
Regarding the case hypo, by definition and part ii of the above Lemma, we would need two 2-forms $\omega_2=b_0\alpha_0+b_1\alpha_1,\ \omega_3=c_0\alpha_0+c_1\alpha_1$ satisfying the orthogonality relations $b_1^2=c_1^2\neq0$ and $b_1c_1=0$. For the case nearly-hypo, it is not possible also to have three natural 2-forms giving a nearly-hypo sphere bundle, because, in searching for $\omega_2=b_0\alpha_0+b_1\alpha_1+b_2\alpha_2+b_3{\mathrm{d}}\theta$ satisfying \eqref{eqnearlyhypo} and in particular ${\mathrm{d}}\omega_2=3\theta\wedge\omega_3$ for the necessarily $\omega_3=c_0\alpha_0+c_1\alpha_1$ found above, we deduce $\omega_2=b_0\alpha_0$, which has vanishing square.
\end{proof}
\subsection{Structures of type I}
\label{oftypeI}
Following the above conclusions, we assume $M$ has constant sectional curvature. A first candidate for $\omega_1$ is that which is found in case i of Lemma \ref{lemadomega1iguala0}. We thus consider ${\mathrm{SU}}(2)$-structures with
\begin{equation}\label{omega1typei}
\omega_1={\mathrm{d}}\theta .
\end{equation}
\begin{Rema}
For a generalization, if we take $\omega_1={\mathrm{d}}\theta$ and find a hypo structure, then the structure can be adjusted accordingly (simply multiplying $\omega_2,\omega_3$ by the same $a_3$). Notwithstanding, for the nearly-hypo equations it is different. Assuming we have found \eqref{eqnearlyhypo} for the pair $\tilde{\theta},\omega_1={\mathrm{d}}\theta$, then the referred variation of $\omega_1$ together with $\tilde{\tilde{\theta}}=\lambda\tilde{\theta}$, $\lambda\in\R$, yields by
\eqref{eqnearlyhypo}
\begin{equation}
a_3=\lambda a_3\qquad\mbox{and}\qquad \lambda a_3=a_3^2
\end{equation}
implying $a_3=1$. Therefore the solutions are 1-1 dependent on $a_3$. The study then continues in the next section.
\end{Rema}
We shall have a hypo structure and, preferably, a contact-hypo structure, if we let $\tilde{\theta}=-2\theta$ and take any two 2-forms, deduced from case i of Lemma \ref{lemadtetaomega3}, satisfying \eqref{hypostruceq1} and \eqref{themetriconLperp}
\begin{equation}\label{omega2and3foromega1typei}
\omega_2=b_0\alpha_0+b_1\alpha_1+b_2\alpha_2,\qquad \omega_3=c_0\alpha_0+c_1\alpha_1+c_2\alpha_2 .
\end{equation}
These shall be called the ${\mathrm{SU}}(2)$-structures of type I. In sum, as in (\ref{hypostruceq4},\ref{hypostruceq5}), we find the system
\begin{equation}\label{beesecees}
\begin{cases}
b_1^2-b_0b_2=1 \\
c_1^2-c_0c_2=1 \\
b_0c_2+b_2c_0-2b_1c_1=0.
\end{cases}
\end{equation}
A last condition is to be fulfilled by the $b_i,c_i\in\R$: that $\phi=\omega_2+\sqrt{-1}\omega_3$ is $(2,0)$ for $\omega_1$, cf. \eqref{hypostruceq2}. As expected, notice the symmetry $\phi\rightsquigarrow \exp({\sqrt{-1}t})\phi$ leaves the system \eqref{beesecees} invariant.
\begin{prop}[\bf ${\mathrm{SU}}(2)$-structures of type I]\label{ProptheinvariantSU2structures}
The natural ${\mathrm{SU}}(2)$-structures on $\cals$ given by the canonical contact 1-form $\tilde{\theta}$ and by the 2-forms $\omega_1,\omega_2,\omega_3$ in \eqref{omega1typei},\eqref{omega2and3foromega1typei} and \eqref{beesecees} are in one-to-one correspondence with points of the real hypersurface
\begin{equation}\label{planopseudoesfera}
\bigl\{(X,Y,A,B)\in\R^4:\ B^2(1+A^2)^2(X^2+Y^2)=1,\,\ B>0\bigr\},
\end{equation}
via the transformation
\begin{equation}\label{coeficientesSU2estrutura}
\begin{cases}
b_0=(1-A^2)B^2X+2AB^2Y \\
b_1=(1+A^2)B(Y-AX) \\
b_2=-(1+A^2)^2X
\end{cases} \qquad
\begin{cases}
c_0=(1-A^2)B^2Y-2AB^2X \\
c_1=-(1+A^2)B(X+AY) \\
c_2=-(1+A^2)^2Y
\end{cases}.
\end{equation}
\end{prop}
\begin{proof}
Let $e_0,e_1,e_2,e_3,e_4$ be an adapted frame orthonormal for the canonical metric. Since $e_0$ is in the annihilator of all $\omega_i$, it follows the new metric on $\cals$ will have $e_0$ orthogonal to the remaining $e_j$. Since the structure is invariant, the compatible almost complex structures $\Phi_i$ on $\ker\theta$ will be invariant (by isometries of $M$ lifted to $\cals$). For $\Phi_1$ compatible with $\omega_1={\mathrm{d}}\theta$ and respecting formula \eqref{themetriconLperpintheory}, we may hence write $\Phi_1x^h=Ax^h-Bx^v$ with some constants $A,B$ and $B>0$, where $x$ is any vector on $T_{\pi(u)}M$ orthogonal to $u\in\cals$ and $x^h,x^v$ are the canonical lifts. The space of $\Phi_1$ is indeed determined completely by $A$ and $B$ (it agrees with the symmetric space $\mathrm{Sp}(2,\R)/{\mathrm{U}}(1)$, the Siegel domain or Poincar\'e half-plane, as studied e.g. in \cite{AlbRaw}). Thus a basis $\{\beta_1,\beta_2\}$ of $(1,0)$-forms is determined up to factors by
\[ \beta_1=e^1+\sqrt{-1}(\lambda e^3+\mu e^1)\mod\R , \]
such that
\[ \beta_1(e_1+\sqrt{-1}\Phi_1e_1)=0 , \]
and similarly for $\beta_2$ recurring to the mirror pair $e_2,e_4$. Solving for $\lambda,\mu$ and removing denominators, we obtain explicit solutions:
\[ \begin{cases} \beta_1=-Be^1+\sqrt{-1}(ABe^1+(1+A^2)e^3) \\
\beta_2=-Be^2+\sqrt{-1}(ABe^2+(1+A^2)e^4) \end{cases} .\]
The $(2,0)$-form $\beta_1\wedge\beta_2$ is independent of the adapted frame, as expected:
\begin{eqnarray*}
\beta_1\wedge\beta_2 &=& (B^2-A^2B^2)e^{12}-AB(1+A^2)(e^{32}+e^{14})-(1+A^2)^2e^{34} + \\
& &\qquad+\sqrt{-1}\bigl(-AB^2e^{12}
-B(1+A^2)e^{14}-AB^2e^{12}-B(1+A^2)e^{32}\bigr) \\
&=& B^2(1-A^2)\alpha_0-AB(1+A^2)\alpha_1-(1+A^2)^2\alpha_2+ \\
& & \qquad\qquad +\sqrt{-1}\bigl(-2AB^2\alpha_0-B(1+A^2)\alpha_1\bigr).
\end{eqnarray*}
The last condition required by an ${\mathrm{SU}}(2)$-structure is that $\omega_2+\sqrt{-1}\omega_3$ is a form of type $(2,0)$-for $\Phi_1$. In other words, we must have $\omega_2+\sqrt{-1}\omega_3=(X+\sqrt{-1}Y)\,\beta_1\wedge\beta_2$ for some $X,Y\in\R$. Equivalently,
\begin{equation*}
\begin{split}
& \qquad\qquad b_0\alpha_0+b_1\alpha_1+b_2\alpha_2+\sqrt{-1}(c_0\alpha_0+c_1\alpha_1+c_2\alpha_2)= \\
& =XB^2(1-A^2)\alpha_0-XAB(1+A^2)\alpha_1-X(1+A^2)^2\alpha_2
+2YAB^2\alpha_0+YB(1+A^2)\alpha_1 +\\
& \sqrt{-1}\bigl(-2XAB^2\alpha_0-XB(1+A^2)\alpha_1+YB^2(1-A^2)\alpha_0 -YAB(1+A^2)\alpha_1-Y(1+A^2)^2\alpha_2\bigr).
\end{split}
\end{equation*}
This yields formulae \eqref{coeficientesSU2estrutura} for the coefficients $b_0,\ldots,b_2,c_0,\ldots,c_2$. Recalling \eqref{beesecees}, then two short computations on the first rows, $b_1^2-b_0b_2=1$ and $c_1^2-c_0c_2=1$, yield the very same condition which is that defining the set \eqref{planopseudoesfera}. Finally, the last equation is automatically satisfied, as we care to show next. Indeed, we have $b_2c_0-c_1b_1=A$ and $b_0c_2-b_1c_1=-A$. Let us see this last identity:
\begin{equation*}
\begin{split}
b_0c_2-b_1c_1&=(1+A^2)^2B^2\bigl(-(1-A^2)XY-2AY^2
+(Y-XA)(X+YA)\bigr) \\
&= (1+A^2)^2B^2\bigl(-XY+A^2XY-2AY^2+XY+AY^2-AX^2-A^2XY\bigr) \\
&= (1+A^2)^2AB^2(-Y^2-X^2) \\
&=-A.
\end{split}
\end{equation*}
Hence $b_0c_2+b_2c_0-2b_1c_1=0$.
\end{proof}
The above Proposition characterizes completely the 3-dimensional family of natural ${\mathrm{SU}}(2)$-structures of type I. Later we shall see that condition \eqref{hypostruceq2} is assured by
\begin{equation}\label{typeImetricpositivedefinite}
b_1c_0-b_0c_1>0.
\end{equation}
The next result shall also be duely proved in Section \ref{Atcomega_1igualadteta}.
\begin{prop}\label{Propositioncompatiblemetric}
The ${\mathrm{SU}}(2)$-structures of type I which are \emph{compatible} with the canonical metric are given by $A=0,\ B=1,\ X^2+Y^2=1$.
\end{prop}
Recall the set of three 2-forms on the radius $s$ tangent manifold $\cals$ determines the Riemannian structure up to the fixed $\|\tilde{\theta}\|_{{\mathrm{SU}}(2)}=1$ (whereas $\|\theta\|=s$). Hence the meaning of the word \textit{compatible} in the last Proposition: the precisely same metric on $\ker\theta$.
We now state the result which follows from various remarks above.
\begin{teo}[\bf Hypo]\label{Teo_hypo}
A natural ${\mathrm{SU}}(2)$-structure on $\cals$ with $\omega_1={\mathrm{d}}\theta$ is hypo if and only if $M$ has constant sectional curvature and it is of type I. Defining $\tilde{\theta}=-2\theta$ we obtain a contact-hypo structure, i.e. satisfying also ${\mathrm{d}}\tilde{\theta}=-2\omega_1$.
Moreover, for any $X,Y\in\R$ such that $X^2+Y^2=1$, the ${\mathrm{SU}}(2)$-structure given by
\begin{equation}
\omega_2=X\alpha_0+Y\alpha_1-X\alpha_2,\qquad\omega_3=Y\alpha_0-X\alpha_1-Y\alpha_2
\end{equation}
is hypo and compatible with the canonical metric.
\end{teo}
\begin{coro}
For any oriented Riemannian 3-manifold $M$, the \emph{main example}, \eqref{mainexample}, defines a contact ${\mathrm{SU}}(2)$-structure compatible with the canonical metric; which is hypo if and only if $M$ has constant sectional curvature.
\end{coro}
Thus, for each pair $K,s$, there exists a 3 dimensional family of contact-hypo structures. However, notice that, as it happens with the \textit{main example}, the induced metric is the same under symmetry $\phi\rightsquigarrow \exp({\sqrt{-1}t})\phi$.
Let us now find the natural nearly-hypo structures, still with the obvious $\omega_1$. Let us stress that we exclude non-constant sectional curvature due to Proposition \ref{Excludingnonconstantsectionalcurvature}.
\begin{teo}[\bf Nearly-hypo]\label{Teo_nearlyhypo}
Suppose $M$ has constant sectional curvature $K$. Then the natural ${\mathrm{SU}}(2)$-structures on the radius $s$ tangent sphere bundle total space $\cals$, with $\tilde{\theta}=-2\theta$ and $\omega_1={\mathrm{d}}\theta$, are nearly-hypo if and only if they are of the kind given in Proposition \ref{ProptheinvariantSU2structures} and, moreover, of the kind given by
\begin{equation}\label{S_equationsnearlyhypo1}
\omega_2=b_0\alpha_0+b_1\alpha_1+b_2\alpha_2,\qquad
\omega_3=\frac{Kb_1}{3}\alpha_0
+\frac{s^2Kb_2-b_0}{6s^2}\alpha_1-\frac{b_1}{3s^2}\alpha_2
\end{equation}
for any $b_0,b_1,b_2\in\R$ such that $b_1^2-b_0b_2=1$ and
\begin{equation}\label{S_equationsnearlyhypo2}
(b_0+s^2Kb_2)^2+4s^2K=36s^4
\end{equation}
and
\begin{equation}
K>-\frac{b_0^2}{s^2(1+b_1^2)}.
\end{equation}
Moreover, such nearly-hypo structures are always contact-hypo.
The structures are compatible with the canonical metric if and only if (i) $b_2=-b_0$, $b_1\neq0$, $b_0^2+b_1^2=1$, $K=3=s^{-2}$, or (ii) $b_2=-b_0=\pm1$, $b_1=0$, $s^2K+1=6s^2$.
\end{teo}
\begin{proof}
By condition $\omega_1\wedge\omega_2=0$, we must have $\omega_2= b_0\alpha_0+b_1\alpha_1+b_2\alpha_2$, and by the same reason or from Lemma \ref{lemadtetaomega3}, we must have $\omega_3=c_0\alpha_0+c_1\alpha_1+c_2\alpha_2$. Hence such nearly-hypo structures exist if and only if they are of the referred kind, this is, type I or \eqref{beesecees}. Next, we see that we just have to study ${\mathrm{d}}\omega_2=3\tilde{\theta}\wedge\omega_3$. Knowing that the Ricci curvature of $M$ satisfies $r=2K$, we obtain the formula for $\omega_3$:
\begin{eqnarray*}
{\mathrm{d}}\omega_2 &=& b_0{\mathrm{d}}\alpha_0+b_1{\mathrm{d}}\alpha_1+b_2{\mathrm{d}}\alpha_2 \\
&=& \theta\wedge(\frac{b_0}{s^2}\alpha_1+\frac{2b_1}{s^2}\alpha_2-rb_1\alpha_0-\frac{r}{2}b_2\alpha_1)\\
&=& \theta\wedge\bigl(-2Kb_1\alpha_0
+\frac{b_0-s^2Kb_2}{s^2}\alpha_1+\frac{2b_1}{s^2}\alpha_2\bigr) \\
&=& 3\tilde{\theta}\wedge\bigl(\frac{Kb_1}{3}\alpha_0
+\frac{s^2Kb_2-b_0}{6s^2}\alpha_1-\frac{b_1}{3s^2}\alpha_2\bigr).
\end{eqnarray*}
A computation on $c_1^2-c_0c_2=1$ yields $(b_0+s^2Kb_2)^2+4s^2K=36s^4$, and these conditions together with \eqref{typeImetricpositivedefinite} are sufficient. Indeed, a very surprising result, the remaining equation is immediately satisfied:
\[ b_0c_2+b_2c_0-2b_1c_1=-\frac{b_0b_1}{3s^2}+\frac{Kb_2b_1}{3}-\frac{2Kb_1b_2}{6}+\frac{2b_0b_1}{6s^2}=0 . \]
It is trivial to prove that ${\mathrm{d}}(\tilde{\theta}\wedge\omega_2)=0$. Indirectly, we note the structure is contact-nearly-hypo, cf. ending of Section \ref{Hanh5m}. Hence it is double-hypo.
Compatibility with the canonical metric is easily seen to be equivalent to cases i or ii. Only $c_1^2-c_0c_2=1$ needs verification: in case i we have
\[ (b_0+s^2Kb_2)^2+4s^2K=(b_0+b_2)^2+4=36/9=36s^4\]
while case ii is
\[ (b_0+s^2Kb_2)^2+4s^2K=(2-6s^2)^2+4(6s^2-1)=4-24s^2+36s^4+24s^2-4=36s^4 \]
as we wished.
\end{proof}
\vspace{2mm}
\noindent
\textsc{Some examples.}
\begin{itemize}
\item It seems there should exist a $(c_0,c_1,c_3)$ \textit{conjugate} to the class of solutions $(b_0,b_1,b_2)=(b_0,b_0+1,b_0+2)$, for any $b_0\in\R$, of $b_1^2-b_0b_2=1$. Looking at $\omega_3$ above, then the best answer might always depend on $K$. Also notice this example and case ii above both contain the \textit{main example}, $b_0=-1$.
\item Let us see the flat case, $K=0$. The product manifold $\R^3\times S^2(s_0)$ for $s_0=\sqrt{1/6}$ admits two, the author believes non-isometric, ${\mathrm{SU}}(2)$-structures both contact-hypo and double-hypo and not Sasaki-Einstein. The first is the \textit{main example}. The second is the above, necessarily with $b_0^2=36s_0^4=1$. We chose $s_0$ on purpose, because we may then have $b_0=-1$, which indeed returns to the \textit{main example}. But also we may have $b_0=1$ and then find a structure given by $\tilde{\theta}=-2\theta,\ \omega_1={\mathrm{d}}\theta$,
\begin{equation}\label{contacdoublehypoR3xS2nonSE}
\omega_2=\alpha_0+2\alpha_1+3\alpha_2,\quad\omega_3=-\alpha_1-4\alpha_2.
\end{equation}
\item For $M$ a hyperbolic space we may also consider the \textit{main example}, case ii, to find another interesting double-hypo structure. For example, letting $K=-3$ and $s=\frac{1}{3}$, the required inequality holds. We remark that in this case ${\mathrm{d}}\omega_3={\mathrm{d}}\alpha_1=-3\tilde{\theta}\wedge(3\alpha_2+\alpha_0)$.
\end{itemize}
Thus, for each pair $K,s$, there exists a 1-dimensional family of nearly-hypo structures. Now let us see the conditions for the Sasaki-Einstein structures.
\begin{coro}\label{Coro_S_SasakiEinsteincurvaturaebees}
The double-hypo structures in Theorem \ref{Teo_nearlyhypo} are Sasaki-Einstein if and only if $M$ has positive constant sectional curvature
\begin{equation}\label{S_SasakiEinsteincurvaturaebees}
K=9s^2 .
\end{equation}
In particular, of the double-hypo structures compatible with the canonical metric, case i is always Sasaki-Einstein, while case ii implies $K=3=s^{-2}$ --- which is i again.
\end{coro}
\begin{proof}
The condition to be verified is just ${\mathrm{d}}\omega_3=-3\tilde{\theta}\wedge\omega_2=6\theta\wedge\omega_2$ where $\omega_2,\omega_3$ are given by the Theorem. On the left hand side we have
\begin{eqnarray*}
{\mathrm{d}}\omega_3
&=& \frac{Kb_1}{3s^2}\theta\wedge\alpha_1+\frac{s^2Kb_2-b_0}{6s^2} \bigl(\frac{2}{s^2}\theta\wedge\alpha_2-2K\theta\wedge\alpha_0\bigr) +\frac{b_1}{3s^2}K\theta\wedge\alpha_1 \\
&=& 6\theta\wedge\bigl(\frac{Kb_0-s^2K^2b_2}{18s^2}\alpha_0
+\frac{Kb_1}{9s^2}\alpha_1+\frac{s^2Kb_2-b_0}{18s^4}\alpha_2\bigr)
\end{eqnarray*}
and so
\[ Kb_0-s^2K^2b_2=18s^2b_0,\qquad Kb_1=9s^2b_1,
\qquad s^2Kb_2-b_0=18s^4b_2. \]
For $b_1\neq0$,
\[ 9s^2b_0-81s^6b_2=18s^2b_0,\qquad K=9s^2,
\qquad 9s^4b_2-b_0=18s^4b_2 . \]
The first and the last equations are, respectively, $-9s^4b_2=b_0,\ -b_0=9s^4b_2$. But these are both equivalent to $b_0+s^2Kb_2=0$, precisely the condition in \eqref{S_equationsnearlyhypo2}. For $b_1=0$, we have $b_0b_2=-1$, and then we see the remaining two equations yield $K+s^2K^2b_2^2=18s^2$ and $s^2Kb_2^2+1=18s^4b_2^2$ (multiplying by $b_0$ gives equivalent conditions). These two imply $K=9s^2$ and so we may proceed as before.
Finally, case i in the Theorem clearly satisfies $K=3=9s^2$. Case ii yields the very same condition, because the solution to $K=\frac{6s^2-1}{s^2}=9s^2$ is precisely $s^2=\frac{1}{3}$ and $K=3$.
\end{proof}
We describe all natural Sasaki-Einstein structures on $\cals$ with
\begin{equation}
\tilde{\theta}=-2\theta \ \qquad \mbox{and} \ \qquad \omega_1={\mathrm{d}}\theta.
\end{equation}
Since $K=9s^2,\ b_1^2=1+b_0b_2$ and $b_0+s^2Kb_2=0$, we define $Q=Q(s,b_2)=\pm\sqrt{1-9s^4b_2^2}$. Then the two remaining 2-forms satisfying \eqref{eqSasakiEinstein} are
\begin{equation}\label{S_equationsSasakiEinstein}
\begin{split}
\omega_2=-9s^4b_2\alpha_0+Q\alpha_1+b_2\alpha_2, & \\
\omega_3=3s^2Q\alpha_0+3s^2b_2\alpha_1-\frac{Q}{3s^2}\alpha_2 .&
\end{split}
\end{equation}
Below we shall find more information on the metric: it is the same for all $b_2$. Actually this symmetry is the natural invariance on $\exp(t\sqrt{-1})(\omega_2+\sqrt{-1}\omega_3)$.
\vspace{2mm}
\noindent
\textsc{Some examples.}
\begin{itemize}
\item Assuming $Q=0$ (one can also follow $b_2=0$ for this case), equivalently, $b_2=\pm\frac{1}{3s^2}$, we have
\begin{equation}
\omega_2= \mp3s^2\alpha_0\pm\frac{1}{3s^2}\alpha_2,\qquad \omega_3=\pm\alpha_1 .
\end{equation}
In particular, for $\cals_{s,M}$ with ray $s=\sqrt{3}/3$ we obtain the \textit{main example}, \eqref{mainexample}.
\item By an exact sequence of homotopy groups, the simply connected Sasaki-Einstein structures compatible with the canonical metric are given over a unique simply connected base of sectional curvature $K=3$ and tangent sphere bundle with radius $s=\sqrt{3}/3$. This is, precisely the sphere $M=S^3(s)$ since $K=1/s^2$. The condition of equal radius on both base and tangent spheres, in the quest for a \textit{Sasakian} manifold, was first found by Tashiro, cf. \cite{Alb2011arxiv,Alb2015a}. The present metric is different.
\end{itemize}
Our invariant theory, as mentioned earlier, is suitable for any quotient manifold $M/\Gamma$ where $\Gamma$ is a discrete group of isometries. New Sasaki-Einstein metrics on the product of $S^2$ with a lens space may hence be described. We recall that such metrics on such products were found in \cite{GMSW}, with a particular interest on 3-dimensional lens spaces; a coincidence with the metrics above is therefore not to be excluded.
\subsection{Other hypo and nearly-hypo structures and case ii of Lemma \ref{lemadomega1iguala0}}
Let us return to the general construction in Section \ref{TnSU2s}. We may search for natural nearly-hypo structures with $\tilde{\theta}=-2p\theta,\ p\neq0$, and generic $\omega_1$ different from the above. Easy enough, equation ${\mathrm{d}}(\tilde{\theta}\wedge\omega_1)=-2\omega_1\wedge\omega_1$ is equivalent to constant sectional curvature of $M$ and
\begin{equation}
a_3p=a_1^2+a_3^2-a_0a_2 .
\end{equation}
The ${\mathrm{SU}}(2)$-structure requires $a_3\neq0$. Now, given a pair of generic 2-forms $\omega_2,\omega_3$ such that ${\mathrm{d}}\omega_2=3\tilde{\theta}\wedge\omega_3=-6p\theta\wedge\omega_3$, then, recalling the computation in the proof of Theorem \ref{Teo_nearlyhypo}, we see immediately how to write $\omega_3$ in terms of the coefficients of $\omega_2=b_0\alpha_0+b_1\alpha_1+b_2\alpha_2+b_3{\mathrm{d}}\theta$:
\begin{equation}\label{omega3ofcertainlynearlyhypo}
\omega_3=\frac{Kb_1}{3p}\alpha_0+
\frac{s^2Kb_2-b_0}{6s^2p}\alpha_1-\frac{b_1}{3s^2p}\alpha_2 .
\end{equation}
In particular, as found much earlier, we must have $c_3=0$. The solutions for a nearly-hypo structure are thus found within the following system, cf. \eqref{hypostruceq1}:
\begin{equation}\label{generalnearlyhypo}
\begin{cases}
a_1^2+a_3^2-a_0a_2=a_3p \\
b_1^2+b_3^2-b_0b_2=a_3p \\ b_0^2-2s^2Kb_0b_2+s^4K^2b_2^2+4s^2Kb_1^2=36s^4a_3p^3 \\
a_0b_2+a_2b_0-2a_1b_1-2a_3b_3=0 \\
a_0b_1-s^2Ka_2b_1+s^2Ka_1b_2-a_1b_0=0 .
\end{cases}
\end{equation}
A sixth equation would come from $\omega_2\wedge\omega_3=0$, but one sees this is automatically satisfied --- `a very surprising' result already seen above.
Clearly, even the case $a_0=a_1=a_2=0$ is difficult to study.
Now let us look again for hypo structures, just satisfying ${\mathrm{d}}\omega_1=0$.
We are led to case ii of Lemma \ref{lemadomega1iguala0}, necessarily on a base $M$ of constant sectional curva\-tu\-re $K=\frac{a_0}{a_2s^2}$, where $a_2\neq0$, and a closed 2-form, necessarily with $a_1=0$,
\begin{equation}\label{omega1fromLemmacaseii}
\omega_1=a_0\alpha_0+a_2\alpha_2+a_3{\mathrm{d}}\theta.
\end{equation}
It follows by Lemma \ref{lemadtetaomega3} that only
\begin{equation}\label{certainhypo1}
\omega_2=b_0\alpha_0+b_1\alpha_1+b_2\alpha_2, \qquad \omega_3=c_0\alpha_0+c_1\alpha_1+c_2\alpha_2
\end{equation}
may participate in a hypo structure $(\theta,\omega_1,\omega_2,\omega_2)$. The coefficients of these hypo structures of \textit{type II} must further solve the structural equations
\begin{equation}\label{certainhypo2}
\begin{cases}
b_1^2-b_0b_2=c_1^2-c_0c_2=a_3^2-a_0a_2\neq0 \\
b_0a_2+b_2a_0=c_0a_2+c_2a_0=0 \\
b_0c_2+b_2c_0-2b_1c_1=0.
\end{cases}
\end{equation}
It follows easily that case $K=0$ does not admit hypo solutions of type II.
Let us also search for nearly-hypo structures with $\omega_1$ closed, of the type of well-known case ii, i.e. of the previous type. Therefore, over the same base manifold. We have system \eqref{generalnearlyhypo} and in particular $\omega_3$ determined by $\omega_2$. We have $a_1=0$ and we know the curvature, $K=\frac{a_0}{a_2s^2}$, which merely solves automatically the last equation in the system.
Double-hypo structures are the next interesting case. They are given by an extra condition, ${\mathrm{d}}(\theta\wedge\omega_2)=0$, which implies $b_3=0$. The two systems above are then reduced to $a_2,a_3,p\neq0$ and
\begin{equation}\label{certaindoublehypo}
\begin{cases}
K=\frac{a_0}{a_2s^2} \\
a_3^2-a_0a_2=a_3p \\
b_1^2-b_0b_2=a_3p \\
a_0b_2+a_2b_0=0 \\
b_0^2-2s^2Kb_0b_2+s^4K^2b_2^2+4s^2Kb_1^2=36s^4a_3p^3 .
\end{cases}
\end{equation}
We call these structures the natural ${\mathrm{SU}}(2)$-structures on $\cals$ of type II.
\begin{teo}[\bf Double-hypo of type II]\label{noncontactdoublehypo}
The natural ${\mathrm{SU}}(2)$-structures with 1-form $\tilde{\theta}=-2p\theta$ and closed 2-form $\omega_1$ from case ii of Lemma \ref{lemadomega1iguala0} are double-hypo if and only if they are given by (\ref{omega1fromLemmacaseii},\ref{omega3ofcertainlynearlyhypo},\ref{certaindoublehypo}) and $a_0a_2,a_3p>0$. Moreover, in this case $M$ has positive sectional curvature
\begin{equation}
K=9s^2p^2 .
\end{equation}
\end{teo}
\begin{proof}
On the lhs of the last equation in the system, we have
\begin{eqnarray*}
(b_0-s^2Kb_2)^2+4s^2Kb_1^2 &=& (b_0-\frac{a_0b_2}{a_2})^2+4\frac{a_0}{a_2}(a_3p +b_0b_2) \\
&=& \frac{1}{a_2^2}\bigl((b_0a_2-a_0b_2)^2+4a_0a_2a_3p+4a_0a_2b_0b_2 \bigr) \\
&=& \frac{4a_0a_3p}{a_2}.
\end{eqnarray*}
The rhs yields the identity $\frac{a_0}{a_2}=9s^4p^2$ and the result follows. The condition $a_0a_2,a_3p>0$ is required by \eqref{themetriconLperp} and can only be proved later (Proposition \ref{themetric_omega1oftypeii}).
\end{proof}
Thus we are bound to positive sectional curvature.
Notice that ${\mathrm{d}}\tilde{\theta}\neq-2\omega_1$, so these double-hypo structures are not contact hypo. Yet we have the following result, which contrasts, for instance, with the double-hypo in \eqref{contacdoublehypoR3xS2nonSE}.
\begin{prop}
All double-hypo structures of type II satisfy
\begin{equation}
{\mathrm{d}}\omega_3=-3\tilde{\theta}\wedge\omega_2.
\end{equation}
\end{prop}
\begin{proof}
First we notice that $b_0+s^2Kb_2=(b_0a_2+b_2a_0)/a_2=0$. Then we wish to check ${\mathrm{d}}\omega_3=6p\theta\wedge\omega_2$, this is,
\begin{eqnarray*}
\frac{Kb_1}{3ps^2}\theta\wedge\alpha_1+\frac{s^2Kb_2-b_0}{6s^2p}(\frac{2}{s^2}\theta\wedge\alpha_2-2K\theta\wedge\alpha_0)+ \qquad\qquad\ & &\\
+ \frac{b_1}{3ps^2}K\theta\wedge\alpha_1 =6p\theta\wedge(b_0\alpha_0+b_1\alpha_1+b_2\alpha_2) . & &
\end{eqnarray*}
This is equivalent to the system
\[ -2s^2K^2b_2+2Kb_0=36s^2p^2b_0,\quad \frac{2Kb_1}{3s^2p}=6pb_1, \quad 2s^2Kb_2-2b_0=36s^4p^2b_2 \]
or
\[\ \ -s^2Kb_2+b_0=2b_0,\qquad Kb_1=9s^2p^2b_1, \qquad
s^2Kb_2-b_0=2s^2Kb_2 . \]
Since these three equations are satisfied, the result follows.
\end{proof}
In this case it seems there is a real ${\mathrm{SU}}(2)$ rather than ${\mathrm{SO}}(2)$ irreducibility, for in this hypothesis the last result is quite easy to prove from the structure equations and letting ${\mathrm{d}}\omega_3$ be a linear combination of the $\theta\wedge\omega_i$.
\vspace{2mm}
\noindent
\textsc{Examples}.
The following give two double-hypo structures, not contact-hypo.
\begin{itemize}
\item
With $b_0=b_2=0,\ a_3p=1$ and radius $s=1$, we have $\tilde{\theta}=-2\theta,\ b_1^2=1,\ K=9p^2$ and still an interval of solutions; one example is with $K=5$
\begin{equation}\label{exampledoublehyp1}
\begin{split}
\omega_1=2\alpha_0+\frac{2}{5}\alpha_2+\frac{3\sqrt{5}}{5}{\mathrm{d}}\theta , \hspace{7mm} \\ \omega_2=\alpha_1,\qquad\quad
\omega_3=3\alpha_0-\frac{\sqrt{5}}{5}\alpha_2 .
\end{split}
\end{equation}
\item
With $a_0=2,\ a_2=1,\ a_3=2,\ p=1$ and $s^2K=2$. This implies $s=\sqrt[4]{2/9}$ and $K=3\sqrt{2}$. For such arbitrary choices, there remains an interval of solutions; one example is
\begin{equation}\label{exampledoublehypo2}
\begin{split}
\omega_1=2\alpha_0+\alpha_2+2{\mathrm{d}}\theta , \hspace{31mm}\\
\omega_2=-\sqrt{2}\alpha_0\pm\alpha_1+\frac{\sqrt{2}}{2}\alpha_2 ,\qquad\quad
\omega_3=\pm\sqrt{2}\alpha_0+\alpha_1\mp\frac{\sqrt{2}}{2}\alpha_2 .
\end{split}
\end{equation}
\end{itemize}
\vspace{2mm}
Theorem \ref{noncontactdoublehypo} generalizes the ${\mathrm{SU}}(2)$-structure results found in \cite[Proposition 6.3]{FIMU}, which are computed directly on $S^3\times S^2$. Our family of double-hypo structures on $S^3\times S^2$ is one dimension higher. We remark that in \cite{FIMU} an auxiliary global parallel frame field on $S^3$ is used in order to deal with the differential geometry of the unit tangent sphere bundle of the 3-sphere.
\subsection{Evolution equations from hypo structures}
\label{Eefnhs}
Let us recall a question raised in \cite{ContiSalamon} regarding an ${\mathrm{SU}}(2)$ structure on a 5-dimensional manifold $N$ and the associated ${\mathrm{SU}}(3)$ metric defined on $N\times\R$, cf. \eqref{hypostruceq3}.
The fundamental article on the \textit{generalized Killing spinors in dimension 5}, which introduces hypo structures, establishes when a smooth 1-parameter family of hypo structures $(\tilde{\theta},\omega_1,\omega_2,\omega_3)_t$ on $N$, time $t$ dependent, induces an integrable ${\mathrm{SU}}(3)$ (Calabi-Yau) metric on the product manifold via ($\phi=\omega_2+\sqrt{-1}\omega_3$)
\begin{equation}\label{SU3structurefromSU2}
F=\omega_1+\tilde{\theta}\wedge{\mathrm{d}} t,\qquad
\Psi =\Psi_++\sqrt{-1}\Psi_- =\phi\wedge(\tilde{\theta}+\sqrt{-1}{\mathrm{d}} t).
\end{equation}
If $\omega_1,\tilde{\theta}\wedge\omega_2,\tilde{\theta}\wedge\omega_3$ are closed, then the evolution equations
\begin{equation}\label{evolutionequationshypotoKahler}
\begin{cases}
\partial_t\omega_1=-{\mathrm{d}}\tilde{\theta}\\ \partial_t(\omega_2\wedge\tilde{\theta})=-{\mathrm{d}}\omega_3\\ \partial_t(\omega_3\wedge\tilde{\theta})={\mathrm{d}}\omega_2
\end{cases}
\end{equation}
are easily deduced as the integrability equations ${\mathrm{d}} F={\mathrm{d}}\Psi=0$, cf. \cite[Proposition 4.1]{ContiSalamon}. Reciprocally, an integrable product structure arising from a family of ${\mathrm{SU}}(2)$-structures implies the hypo equations \eqref{eqhypo} for all $t$.
In the analytic category, by Cartan-K\"ahler theory, \cite[Theorem 4.4]{ContiSalamon} establishes the existence of solution to \eqref{evolutionequationshypotoKahler}. The question remains open within the smooth category, quite puzzling due to the existence of non-analytic hypersurfaces in Calabi-Yau manifolds.
An explicit solution is immediately provided for Sasaki-Einstein manifolds, on $N\times\R_+$; it is known as the \textit{conical} ${\mathrm{SU}}(3)$-structure:
\begin{equation}\label{SU3structureforSasakiEinstein}
F=t^2\omega_1+t\tilde{\theta}\wedge{\mathrm{d}} t,\qquad
\Psi =t^2\phi\wedge(t\tilde{\theta}+\sqrt{-1}{\mathrm{d}} t).
\end{equation}
Finally, one may consider the evolution equations on natural ${\mathrm{SU}}(2)$-structures on the total space $\cals_{s,M}$ of tangent sphere bundles and try to solve them within the same natural category. It is a quite demanding problem, also because there are other developments of the theory, namely in \cite{FIMU}, which involve the nearly-hypo structures and their own evolution equations now lifted to nearly-K\"ahler complex 3-folds. Interesting findings on double-hypo ${\mathrm{SU}}(2)$-structures and half-flat ${\mathrm{SU}}(3)$-structures lead to constructions of manifolds with ${\mathrm{G}_2}$-holonomy. They all lead to further substantial questions applying on our context, so we leave the subject for the moment and point the reader to a future work.
Nevertheless, we shall give a new solution to the evolution equations of Conti-Salamon for one case on $\cals$ with a natural hypo structure of type I. Given a hypo structure of type $I$ by the usual constant values $p$ in $\tilde{\theta}=-2p\theta$ and $a_3,b_0,b_1,b_2,c_0,c_1,c_2$ in $\omega_1,\omega_2,\omega_3$ over a constant sectional curvature $K$ oriented 3-manifold, we wish to solve the evolution equations within the type I natural hypo structures. In other words, we wish to find $P,A_3,B_0,B_1,B_2,C_0,C_1,C_2$ functions of $t$, such that
\begin{equation}\label{functionswewishtofindforhypoIevolutionproblem}
\begin{split}
& \tilde{\theta}=-2P\theta,\quad\omega_1=A_3{\mathrm{d}}\theta,\quad \omega_2=B_0\alpha_0+B_1\alpha_1+B_2\alpha_2,\quad \omega_3=C_0\alpha_0+C_1\alpha_1+C_2\alpha_2, \\
& \hspace{20mm} B_1^2-B_0B_2=C_1^2-C_0C_2=A_3^2,\ \ B_0C_2+B_2C_0-2B_1C_1=0, \\
& \hspace{37mm} A_3>0,\qquad B_1C_0-B_0C_1>0,
\end{split}
\end{equation}
is a 1-parameter family of ${\mathrm{SU}}(2)$-structures solving \eqref{evolutionequationshypotoKahler} and containing the initial structure. Recall from Theorem \ref{Teo_hypo} that all these structures are automatically hypo.
\begin{prop}
The natural type I evolution equations are equivalent to
\begin{equation}\label{evolutionequationsnaturalstructure}
\begin{cases}
\partial_tA_3=2P \\
\partial_t(PC_0)=KB_1 \\ \partial_t(PC_1)=\frac{s^2KB_2-B_0}{2s^2} \\
\partial_t(PC_2)=-\frac{B_1}{s^2}
\end{cases}
\qquad
\begin{cases}
\partial_t(PB_0)=-KC_1 \\ \partial_t(PB_1)=-\frac{s^2KC_2-C_0}{2s^2} \\
\partial_t(PB_2)=\frac{C_1}{s^2} .
\end{cases}
\end{equation}
\end{prop}
The proof is immediate applying the usual formulae. In particular, if $P=p$ is constant, then both $B_1$ and $C_1$ satisfy
\begin{equation}
\partial_{tt}^2X-\frac{K}{p^2s^2}X=0
\end{equation}
and so all $B_i,C_i$ are in general of the \textit{elliptic} kind. However, the assumption proves not to be so fruitful, because then $A_3=2pt+a_4$ and since the solutions must satisfy $B_1^2-B_0B_2=C_1^2-C_0C_2=A_3^2=4p^2t^2+4pa_4t+a_4^2,\ \forall t$, cf. \eqref{beesecees}, we easily run into contradiction. Clearly an exception occurs with the flat case, $K=0$, a we shall see below.
Another assumption to make would be $P=p_1t+p_2$ with $p_1,p_2$ constant. This leads to quadratic solutions, but only for $K>0$ although not necessarily the Sasaki-Einsten conical solution \eqref{SU3structureforSasakiEinstein}.
\subsection{An integrable special-Hermitian structure}
\label{AiSU3s}
Following the above discussion, we now solve the evolution equations for an oriented Riemannian flat 3-manifold $M$ and a natural hypo structure of type I on $\cals_{s,M}$.
We keep considering any radius $s$ tangent sphere bundle. Indeed, the variable $s$ may enter into the solution as a function of $t$, over the fixed smooth manifold $\cals$. The same is true for the curvature $K$, as long as a conformal change on $M$ carries along conveniently with any changes in $s$. These relations are well established in \cite{Alb2014a}, in particular for space forms.
In the present setting, we have $K=0$ and the initial data of $(\tilde{\theta},\omega_1,\omega_2,\omega_3)$ of type I are the usual constants $p,a_3,b_0,\ldots,c_2$.
Then we have the following solution of system \eqref{evolutionequationsnaturalstructure} with $P=p>0$:
\begin{equation}
\begin{split}
& A_3=2pt+a_4,\qquad B_0=b_0, \qquad C_0=c_0, \\
& B_1=\frac{c_0}{2ps^2}t+b_4,\qquad C_1=-\frac{b_0}{2ps^2}t+c_4, \\
& B_2=-\frac{b_0}{4p^2s^4}t^2+\frac{c_4}{ps^2}t+b_5,\qquad
C_2=-\frac{c_0}{4p^2s^4}t^2-\frac{b_4}{ps^2}t+c_5,
\end{split}
\end{equation}
with $a_4,b_4,c_4,b_5,c_5$ real constants.
The conditions required by ${\mathrm{SU}}(2)$-structures follow:
\begin{equation}
\begin{split} \label{moresystems}
& \hspace{30mm} 2pt+a_4>0 , \\
& \qquad\qquad b_0^2+c_0^2=16p^4s^4 , \qquad b_4c_0-b_0c_4=4p^2s^2a_4 , \\
& b_4^2-b_0b_5=c_4^2-c_0c_5=a_4^2 , \qquad b_0c_5+b_5c_0-2b_4c_4=0.
\end{split}
\end{equation}
These equations come from the second line of \eqref{functionswewishtofindforhypoIevolutionproblem}. For instance, we have
$C_1^2-C_0C_2=A_3^2$ if and only if $\frac{b_0^2}{4p^2s^4}t^2-\frac{b_0c_4}{ps^2}t+c_4^2+\frac{c_0^2}{4p^2s^4}t^2+\frac{c_0b_4}{ps^2}t-c_0c_5 = 4p^2t^2+4pta_4+a_4^2$,
and thus three of the five equations follow. Notice $B_1C_0-B_0C_1>0$ holds trivially.
Also, notice the substitution $a_4=a_3,\ b_4=b_1,\ c_4=c_1,\ b_5=b_2,\ c_5=c_2$ solves the third line and yields the initial structure at time $t=0$.
Regarding the general solution of system \eqref{evolutionequationsnaturalstructure}, notice the $B$s and the $C$s determine each other and, in the end, they determine $A_3$ and so finally $A_3$ determines $P$. Hence the solution is not very far from the above.
Finally we consider the \textit{main example} over an oriented flat 3-manifold $M$. Letting $2ps=1$ and $a_4=b_4=c_4=b_5=c_5=0$ and, moreover, changing $t/s$ for $t$, then we may just as well let $p=\frac{1}{2},\ s=1$. We have the following solution of the natural evolution equations:
\begin{equation}
\tilde{\theta}=-\theta,\quad \omega_1=t{\mathrm{d}}\theta,\quad \omega_2=t^2\alpha_2-\alpha_0, \quad \omega_3=t\alpha_1.
\end{equation}
And so we obtain a \textit{new} integrable ${\mathrm{SU}}(3)$-structure on $Z=\cals_{1,M}\times\R_+$:
\begin{equation}
F=t{\mathrm{d}}\theta-\theta\wedge{\mathrm{d}} t,\quad\qquad \phi=\omega_2+\sqrt{-1}\omega_3,
\end{equation}
\begin{equation}
\begin{split}
& \hspace{34mm} \Psi= \phi\wedge(-\theta+\sqrt{-1}{\mathrm{d}} t) =\\
& = \theta\wedge\alpha_0-t^2\theta\wedge\alpha_2-t\alpha_1\wedge{\mathrm{d}} t-
\sqrt{-1}(t\theta\wedge\alpha_1-t^2\alpha_2\wedge{\mathrm{d}} t+\alpha_0\wedge{\mathrm{d}} t).
\end{split}
\end{equation}
Indeed, in no trivial way becomes $Z$ an open subset of $\C^3$. Nor for any flat trivializing-neighborhood of $M$. We also recall
\begin{equation}
{\mathrm{d}}\alpha_0=\theta\wedge\alpha_1,\quad {\mathrm{d}}\alpha_1=2\theta\wedge\alpha_2,\quad {\mathrm{d}}\alpha_2=0,
\end{equation}
in order to prove ${\mathrm{d}} F={\mathrm{d}}\Psi=0$.
\begin{Rema}
To give a direct proof of the fundamental differential system formulae \eqref{introd_derivadasdastres2formas}, deduced twice in general in \cite{Alb2011arxiv,Alb2015a}, now defined over the Euclidean space, one may use co\-or\-di\-na\-tes $(x^1,x^2,x^3,u^1,u^2,u^3)$ on $\R^3\times S^2$ with $\sum(u^i)^2=1$ and the notation ${\mathrm{d}}^{ijk,}={\mathrm{d}} x^i\wedge{\mathrm{d}} x^j\wedge{\mathrm{d}} x^k$, ${\mathrm{d}}^{ij,k}={\mathrm{d}} x^i\wedge{\mathrm{d}} x^j\wedge {\mathrm{d}} u^k$. Then
\begin{equation}
\theta=\sum u^i{\mathrm{d}} x^i, \quad \alpha_0=\mathop{\kern0.9ex{{+_{123}u^1{\mathrm{d}}^{23,}, \quad\alpha_1=\mathop{\kern0.9ex{{+_{123} u^1({\mathrm{d}}^{2,3}-{\mathrm{d}}^{3,2}), \quad \alpha_2=\mathop{\kern0.9ex{{+_{123} u^1{\mathrm{d}}^{,23} .
\end{equation}
\end{Rema}
\subsection{The metric explicit}
\label{Tme}
We provide some further information on the ${\mathrm{SU}}(2)$ metrics on $\cals_{s,M}$ with the most generic $\omega_i$, linear combination of $\alpha_0,\alpha_1,\alpha_2,{\mathrm{d}}\theta$. For a given a set of coefficients $a_0,\ldots,c_3$, recall the metric induced on $\cals$ is denoted by $g_{_{{\mathrm{SU}}(2)}}$.
The tautological horizontal or Reeb vector field $-\frac{1}{2ps}\,e_0$ on $\cals$ is dual to $\tilde{\theta}=-2ps\,e^0$. We then must have $\|e_0\|_{{\mathrm{SU}}(2)}=2s|p|$ and $\ker\theta=\ker e^0=\ker\gsudois{e_0}{\ }$. Now we need to define the following functions:
\begin{equation}
\begin{split}\label{simplythemetric}
g_{11}=g_{22}&=(a_1b_0-a_0b_1)c_3+(a_0b_3-a_3b_0)c_1+(a_3b_1-a_1b_3)c_0 \\
g_{33}=g_{44} &=(a_2b_1-a_1b_2)c_3+(a_1b_3-a_3b_1)c_2+(a_3b_2-a_2b_3)c_1 \\
g_{12}=g_{34} &=0 \\
g_{13}=g_{24}& =\tfrac{1}{2}(a_3(b_2c_0-b_0c_2)+b_3(a_0c_2-a_2c_0)+
c_3(a_2b_0-a_0b_2)) \\
g_{14}=-g_{23}& =\tfrac{1}{2}(a_1(b_0c_2-b_2c_0)+b_1(a_2c_0-a_0c_2)+
c_1(a_0b_2-a_2b_0)) .
\end{split}
\end{equation}
\begin{prop}\label{Proposition_themetricmatrix}
Let $e_0,e_1,\ldots,e_4$ be an adapted frame on $\cals$, hence orthonormal for the canonical metric. Then the symmetric matrix $G:=[\gsudois{e_i}{e_j}]_{1\leq i,j\leq4}$ equals $[g_{ij}]_{1\leq i,j\leq4}$, this is
\begin{equation} \label{themetricmatrix}
G= \left[\begin{array}{cccc}
g_{11} & 0 & g_{13} & -g_{23} \\
0 & g_{11} & g_{23} & g_{13} \\
g_{13} & g_{23} & g_{33} & 0 \\
-g_{23} & g_{13} & 0 & g_{33}
\end{array} \right].
\end{equation}
\end{prop}
\begin{proof}
A direct application of Theorem \ref{Teo_themetriconLperp}.
\end{proof}
We note that $G$ is indeed invariant of the choice of adapted frame, because that is the case of the fundamental exterior differential system. Or, more plainly, because $\C$ is abelian. Further on, of course we must have condition \eqref{hypostruceq2}, which is equivalent to the metric being positive definite due to Theorem \ref{Teo_themetriconLperp}. A computation first gives
\begin{equation}\label{determinantedeG}
\det G=(g_{11}g_{33}-g_{13}^2-g_{23}^2)^2 .
\end{equation}
Computing the minors of $G$ yields the following result.
\begin{prop}\label{Prop_Gmetricpositivedefinite}
A natural ${\mathrm{SU}}(2)$ metric on $\cals$ being positive definite is equivalent to
\begin{equation}
g_{11}>0,\quad g_{11}g_{33}-g_{13}^2-g_{23}^2>0.
\end{equation}
\end{prop}
The metric matrix $G$ announces a new class of natural metric on tangent sphere bundles of 3-manifolds, which to the best of our knowledge was never considered before. The structure yields the `g-natural' metrics known in the literature, as well as that new class. Recall the term `g-natural' metric, e.g. from \cite{Abb1,AbbKowal}, refers to a metric like the above but \textit{only} involving a constant linear combination of $\theta\otimes\theta$ and $g(x^h,y^h),g(x^h,y^v)$, $g(x^v,y^v)$ for the lifts of any $x,y\in TM$. Hence the importance by the negative of the next result.
\begin{prop}\label{SU2metricsareofmoregeneralkind}
A natural ${\mathrm{SU}}(2)$ metric on $\cals$ is a g-natural metric if and only if $g_{23}=0$.
\end{prop}
We remark there do exist structures with $g_{13}=0$ and $g_{23}\neq0$, cf. Proposition \ref{themetric_omega1oftypeii}.
Next we give a formula for the unique endomorphisms $\Phi_i\in\End{T\cals}$, for $i=1,2,3$, ortho\-go\-nal for the ${\mathrm{SU}}(2)$ metric and such that
\begin{equation}
{\Phi_i}^*\omega_i=\omega_i,\qquad \qquad \Phi_i^2=-1_{T\cals}+e_0\otimes e^0.
\end{equation}
Taking any adapted frame and denoting the matrices of $\Phi_i,\omega_i$ restricted to $\ker\theta$ by the same letters, we have
\begin{equation}
\omega_1=\left[\begin{array}{cc}
a_0J_1 & A_{13} \\ -A_{13}^T & a_2J_1
\end{array}\right]
\end{equation}
where ($k\in\N$)
\begin{equation}\label{definitionofJandA}
J_k=\left[\begin{array}{cc}
0 & 1_k \\ -1_k & 0
\end{array}\right] \qquad\mbox{and}\qquad A_{13}=\left[\begin{array}{cc}
-a_3 & a_{1} \\ -a_1 & -a_3
\end{array}\right] .
\end{equation}
Equivalent notations follow for $\omega_2,\omega_3$, with $B_{13},C_{13}$, respectively, in place of $A_{13}$. Recall there exists a unique $\nu\in\R,\ \forall i$, such that $a_1^2+a_3^2-a_0a_2=\nu$, etc. So we have $A_{13}A_{13}^T=(a_0a_2+\nu)1_2$. Since $A_{13}J_1=J_1A_{13}$, we have $\omega_1\hat{\omega}_1=\nu1_4$ with
\begin{equation}
\hat{\omega}_1=\left[\begin{array}{cc}
a_2J_1 & -A_{13} \\ A_{13}^T & a_0J_1
\end{array}\right].
\end{equation}
The ${\mathrm{SU}}(2)$-structure translates into $\omega_i\Phi_i=G$ for all $i=1,2,3$. This proves the formulae
\begin{equation}\label{Phi_imatrices}
\Phi_i=\frac{1}{\nu}\hat{\omega}_iG .
\end{equation}
Next we deduce when an endomorphism, say $\Phi_1$, does preserve the vertical tangent bundle $V_0$, in which case we say simply $\Phi_1$ \textit{preserves the fibres} or \textit{preserves} $V_0$.
\begin{prop}
$\Phi_1$ preserves the fibres if and only if
\begin{equation}\label{conditionforpreservingthefibres}
\begin{cases}
a_2g_{23}+a_3g_{33}=0 \\ a_2g_{13}-a_1g_{33}=0
\end{cases} .
\end{equation}
In particular, if the ${\mathrm{SU}}(2)$ metric is compatible with the canonical metric and $\Phi_1$ preserves the fibres, then $a_1=a_3=0,\ a_0=-a_2,\ b_0=-b_2,\ c_0=-c_2$.
\end{prop}
\begin{proof}
Combining \eqref{Phi_imatrices} with \eqref{themetricmatrix}, condition $\Phi_1(V_0)\subset V_0$ is equivalent to the vanishing of the top right corner of $\Phi_1$. This is
\[ a_2J_1\left[\begin{array}{cc}
g_{13} & -g_{23} \\ g_{23} & g_{13}
\end{array} \right]-A_{13}g_{33}=0 , \]
and hence the system.
If, furthermore, we have $G=1_4$, then clearly $a_1=a_3=0$. And from $g_{11}=g_{33}=1$, we get $-a_0b_1c_3+a_0b_3c_1=1,\ a_2b_1c_3-a_2b_3c_1=1$ which yields $a_0=-a_2\neq0$ and the determinant $b_1c_3-b_3c_1\neq0$. Now from the formulae for $g_{13},g_{14}$, we find $b_3(a_0c_2-a_2c_0)+c_3(a_2b_0-a_0b_2)=0$ and $ b_1(a_2c_0-a_0c_2)+c_1(a_0b_2-a_2b_0)=0$. In other words, $b_3(c_2+c_0)-c_3(b_0+b_2)=0$ and $b_1(c_0+c_2)-c_1(b_2+b_0)=0$.
\end{proof}
Recall $\ker\theta=H_0\oplus V_0$, so it is only fair to consider the same question for horizontals: we say $\Phi_1$ \textit{preserves} $H_0$ if $\Phi_1(H_0)\subset H_0$. Equivalently,
\begin{equation}\label{conditionforpreservingHorizontals}
\begin{cases}
a_0g_{23}+a_3g_{11}=0 \\ a_0g_{13}-a_1g_{11}=0
\end{cases} .
\end{equation}
A last remark applies only to diagonal metrics, i.e. $g_{13}=g_{23}=0$. We recall the studies in \cite{Alb5,Alb2012} and specially \cite{Alb2014a} regarding a conformal change on the base metric on $M$, a radius $s$ of $\cals_{s,M}$, a conformal change on $H$ and $V_0$ and, moreover, how the previous three must relate, in order to build a homothety with the obvious map between tangent sphere bundles with different radius. Certainly noteworthy results in respect to classifying some of the ${\mathrm{SU}}(2)$ metrics above.
\section{The two distinguished types and evolution equations}
\subsection{The type I metrics}
\label{Atcomega_1igualadteta}
We resume with the natural ${\mathrm{SU}}(2)$-structures of type I, determined in Proposition \ref{ProptheinvariantSU2structures}, Theorems \ref{Teo_hypo} and \ref{Teo_nearlyhypo}. We have $a_0=a_1=a_2=b_3=c_3=0,\ a_3=1,\ b_1^2-b_0b_2=c_1^2-c_0c_2=1$, $b_0c_2+b_2c_0-2b_1c_1=0$, and therefore, reading from Proposition \eqref{Proposition_themetricmatrix}, we prove the next result.
\begin{prop}
The natural ${\mathrm{SU}}(2)$ metrics of type I satisfy
\begin{equation}\label{themetricforomega1oftypei}
\begin{split}
g_{11}=g_{22}&=b_1c_0-b_0c_1 \\
g_{33}=g_{44} &=b_2c_1-b_1c_2 \\
g_{12}=g_{34} &=0 \\
g_{13}=g_{24}& =\tfrac{1}{2}(b_2c_0-b_0c_2) \\
g_{14}=-g_{23}& =0.
\end{split}
\end{equation}
\end{prop}
Recall ``$g_{00}$''$=4s^2$ completes the information on this metric. The nearly-hypo structures of type I satisfy $g_{13}\neq0$ in general. For the particular case of the structure in \eqref{contacdoublehypoR3xS2nonSE}, over a flat base and radius $s$ with square $1/6$, we see
\begin{equation}\label{contacdoublehypoR3xS2nonSEmetrictensor}
g_{11}=1,\quad g_{33}=5,\quad g_{13}=2.
\end{equation}
Regarding the Sasaki-Einstein metrics found in \eqref{S_equationsSasakiEinstein}, we have
\begin{equation}\label{S_equationsSasakiEinstein2}
g_{11}=3s^2,\quad g_{33}=\frac{1}{3s^2},\quad g_{13}=0.
\end{equation}
\begin{prop}\label{Proposition_typeImetricpositivedefinite}
For structures of type I, we have $\det G=1$. Moreover, the metric defined by the matrix $G$ is positive definite if and only if $b_1c_0-b_0c_1>0$.
\end{prop}
\begin{proof}
The detailed computation, requiring \eqref{beesecees} together with the above results applied on \eqref{determinantedeG}, can be obviated if we notice that ${\mathrm{d}}\theta=\omega_1$ induces the same volume-form as the canonical metric. However, there is more; the computation yields $g_{11}g_{33}-g_{13}^2-g_{23}^2=1$. The second assertion then follows by Proposition \ref{Prop_Gmetricpositivedefinite}.
\end{proof}
The following is a restatement of Proposition \ref{Propositioncompatiblemetric}, finally with a proof.
\begin{prop}\label{Propositioncompatiblemetric_withproof}
The ${\mathrm{SU}}(2)$ metric of type I coincides on $\ker\theta$ with the canonical metric if and only if $b_0=-b_2=-c_1,\ b_1=c_0=-c_2$ and $b_0^2+b_1^2=1$.
\end{prop}
\begin{proof}
Conditions \eqref{beesecees} and $G=1_4$ lead to the equivalent relations.
\end{proof}
Immediately we see that $\Phi_1$ arising from $\omega_1={\mathrm{d}}\theta$ does not preserve the fibres. As matrices, we have $\omega_1=-J_2$, thus
\begin{equation}
\Phi_1= J_2G=\left[\begin{array}{cc}
g_{13}1_2 & g_{33}1_2 \\ -g_{11}1_2 & -g_{13}1_2
\end{array}\right] .
\end{equation}
In particular we verify that ${\Phi_1}^2=-1_4$.
Now let us see $\Phi_2$ for general ${\mathrm{SU}}(2)$-structures of type I.
\begin{prop}
$\Phi_2$ preserves the fibres if and only if $c_2=0$.
If moreover the metric is compatible with the canonical metric, then the case is that of the \emph{main example}.
\end{prop}
\begin{proof}
By \eqref{conditionforpreservingthefibres} the condition is equivalent to $b_3g_{33}=0$ and $b_2g_{13}-b_1g_{33}=0$. Recalling \eqref{beesecees}, we have $b_3=0$. On the other hand,
\begin{eqnarray*}
b_2g_{13}-b_1g_{33} &=& \frac{1}{2} b_2^2c_0-\frac{1}{2}b_0b_2c_2-b_1b_2c_1+b_1^2c_2 \\ &=& \frac{1}{2} b_2^2c_0-\frac{1}{2}b_0b_2c_2-\frac{1}{2}b_0b_2c_2
-\frac{1}{2}b_2^2c_0+c_2+b_0b_2c_2 \\
&=& c_2.
\end{eqnarray*}
The result now follows easily.
\end{proof}
One may verify as above that
\begin{equation}
\Phi_2\ \mbox{preserves}\ H_0\ \ \Longleftrightarrow\ \ c_0=0.
\end{equation}
\begin{coro}
$\Phi_2$ preserves $H_0$ and $V_0$ if and only if $\pm b_2>0$ and
\begin{equation}
\omega_1={\mathrm{d}}\theta,\quad\omega_2=b_2\alpha_2-\frac{1}{b_2}\alpha_0,\quad\omega_3=\pm\alpha_1 .
\end{equation}
\end{coro}
\subsection{The double-hypo structures of type II}
\label{TdhtIIm}
We return to the natural non contact double-hypo structures $(\tilde{\theta},\omega_1,\omega_2,\omega_3)$ of type II, found in Theorem \ref{noncontactdoublehypo}, in order to study the induced metric. However, our conclusion will be that this class of metrics on $\cals$ deserves a dedicated study.
\begin{prop}\label{themetric_omega1oftypeii}
Double-hypo structures of type II satisfy
\begin{equation}
g_{11}=\frac{a_0a_3^2}{3a_2s^2},\qquad
g_{33}=\frac{a_3^2}{3s^2}, \qquad g_{13}=0,\qquad
g_{23}=-\frac{a_0a_3}{3s^2}.
\end{equation}
The positive definite condition on the metric corresponds to $a_0a_2>0,a_3p>0$.
\end{prop}
\begin{proof}
Besides $a_1=b_3=c_3=0$ and $a_2,p\neq0$, we have system \eqref{certaindoublehypo} and
\[ c_0=\frac{Kb_1}{3p},\qquad c_1=-\frac{b_0}{3s^2p},\qquad
c_2=-\frac{b_1}{3s^2p} . \]
Therefore $b_0+b_2Ks^2=(b_0a_2+b_2a_0)/{a_2}=0$ and hence $b_2c_0-b_0c_2=0$. On the other hand, by \eqref{simplythemetric}, we find immediately $g_{11}=a_3(b_1c_0-b_0c_1)$, $g_{33}=a_3(b_2c_1-b_1c_2)$, $g_{13}=\frac{1}{2}a_3(b_2c_0-b_0c_2)$ and $g_{23}=\frac{1}{2}(b_1(a_0c_2-a_2c_0)+c_1(a_2b_0-a_0b_2))$ and then the desired identities are trivial to deduce. Regarding the positive definite condition required by Proposition \ref{Prop_Gmetricpositivedefinite} we definitely must have $a_0a_2>0$. Since
\[ g_{11}g_{33}-g_{13}^2-g_{23}^2=\frac{a_0a_3^4}{9a_2s^4}-\frac{a_0^2a_3^2}{9s^4}
=\frac{a_0a_3^2}{9a_2s^4}(a_3^2-a_0a_2)=\frac{a_0a_3^2}{9a_2s^4}a_3p , \]
the result follows.
\end{proof}
Using \eqref{conditionforpreservingthefibres} and \eqref{conditionforpreservingHorizontals} the following is trivial to check.
\begin{prop}
For natural double-hypo structures of type II, neither $\Phi_1,\Phi_2$ or $\Phi_3$ preserve the horizontal or the vertical distributions.
\end{prop}
|
\section{Introduction}
In collision-poor plasmas from space large deviations from thermodynamic equilibrium cannot be relaxed by the
particle-particle (Coulomb) collisions, but can presumably be constrained by the resulting kinetic instabilities.
Thus, if the solar wind expands adiabatically the CGL invariants conserve \citep{chew56} leading to an indefinite
increase of temperature in the direction parallel to the inteplanetary magnetic field, i.e., $T_\parallel > T_\perp$.
However, the in-situ measurements do not confirm such an increase of their parallel temperature with heliocentric
distance, but indicate bounds of the temperature anisotropy of plasma particles \citep{Kasper02, Hellinger06,
St2008}. Because collisions are not efficient, the most invoked mechanism that can limit the increase of parallel
temperature is the firehose instability \citep{Eviatar70,Kasper02,Hellinger06, St2008, La2014}.
The firehose instability driven by the anisotropic electrons with $A \equiv T_\perp / T_\parallel < 1$, also known as the
electron firehose instability (EFHI), is particularly important as it can mediate a resonant transfer of (free) energy from
electrons to protons \citep{Paesold99, Messmer02}. This energy transfer from small to large scales
is facilitated by the quasi-parallel EFH modes, which are left-handed (LH) circularly polarized and have characteristic
frequencies and growth rates in the range of the proton cyclotron frequency. Besides the propagating
(non-zero frequency) modes predominant at small angles (quasi-parallel) with respect to the magnetic field direction,
the firehose instability may destabilize an additional aperiodic (non-propagating) branch which exists only for oblique
directions \citep{Ga2003,Ca2008,Hellinger14}. Although it is well known that the suprathermal populations
are ubiquitous in the solar wind \citep{Lin1998,Pier2010,La2012a}, the anisotropic temperature is in general
quantified by a bi-Maxwellian distribution function, which
is relevant only for the thermal core of the solar wind electrons. In this case the aperiodic FHI
is found to grow faster than the propagating modes, and the instability thresholds approach well enough
the limits of the core anisotropy reported by the observations \citep{St2008}. For anisotropies exceeding these thresholds,
the free energy is dissipated by the resulting instabilities, which may also scatter particles back towards
quasi-equilibrium states and prevent the anisotropy to grow \citep{Ga1994,Ga1998}. Instead, for the suprathermal
electrons from the solar wind the limits of their anisotropy are markedly departed from the instability
thresholds derived for bi-Maxwellian populations, see Figure~6 and the analysis in \cite{St2008}.
This disagreement may simply be motivated by the fact that
suprathermal populations cannot be properly described by the Maxwellian distribution functions, but they can be
accurately reproduced by the Kappa power-laws \citep{Va1968,Ma2005,Pier2010}. Ubiquitous in the solar wind
and subsequent planetary environments, e.g., terrestrial magnetosphere, see the review by \cite{Pier2010},
suprathermal electrons are more dilute but hotter than the core populations. The relaxation through the
particle-particle collisions is even less efficient in this case, but kinetic instabilities are expected to explain
the limits of temperature anisotropy reported by the observations.
In the present paper we propose a refined analysis of the suprathermal electrons by using a bi-Kappa distribution function to
describe the anisotropy of these populations. In the limit of a high power-index $\kappa \to \infty$
the (bi-)Kappa distribution function reduces to a (bi-)Maxwellian. \cite{Ma2005} and \cite{St2008} have used the bi-Kappa
model to quantify the velocity distributions and the principal properties of the suprathermal electrons in the solar wind,
e.g., the components of the anisotropic temperature, parallel ($T_\parallel$) and perpedicular ($T_\perp$) to
the magnetic field direction. The suprathermal electrons are found to be highly anisotropic and with a predominant
excess of parallel temperature susceptible to the FHI. The bi-Kappa model was also extensively invoked in theories
of dispersion and stability by adopting two alternative assumptions for the temperature of Kappa populations, to be either dependent or
independent of the power-index $\kappa$. Studies of the FHI \citep{La2009,La2011} assume $\kappa$-independent
temperatures, and find, contrary to the expectations, that the instability is inhibited by the suprathermals and
the instability thresholds do not approach but depart even more from the anisotropy bounds of the solar wind suprathermal electrons.
However, from a recent analysis on the applicability of Kappa distributions \citep{La2015a,La2016} it
becomes evident that a representation with a $\kappa$-dependent temperature may provide a more natural
interpretation of the suprathermal populations for three fundamental reasons: (1)~it corresponds to a Maxwellian
limit which reproduces more accurately the thermal (core) population enabling for a direct and realistic
comparison \citep{La2015a}; (2)~the kinetic instabilities show a systematic stimulation in the presence of suprathermal
electrons \citep{La2015a, Vinas2015,Sh2016} as one may expect from the excess of free energy acumulated by these populations;
and (3)~the observations show strong evidence of $\kappa$-dependent temperatures, which increase in
the presence of suprathermal populations, i.e., temperatures increase with decreasing the power-index $\kappa$ \citep{Pier2016}.
Motivated by these premises, here we re-analyse the instability of the EFH mode by modeling the suprathermal
electrons with a bi-Kappa approach with
$\kappa$-dependent temperatures. In this preliminary analysis we restrict to the same parallel (non-zero frequency)
modes studied before by \cite{La2009,La2011}. The bi-Kappa approach is introduced in section~2, enabling us
to derive the dispersion relation for the FHI modes. The main features of the instability, are derived
and discussed in section~3. In addition, the EFH thresholds are compared with the observations of the suprathermal
electron anisotropy. The results of the present work are summarized in section 4.
\section{Bi-Kappa electrons. Dispersion relations}
We first introduce the analytical model for the velocity
distributions of suprathermal electrons detected in space
plasmas \citep{Ma2005,St2008}. The suprathermal (halo) electrons are assumed to be
a gyrotropic component (isotropic in the plane transverse to the magnetic field)
with a bi-axis temperature anisotropy $T_\perp \ne T_\parallel$, where
$\parallel$ and $\perp$ denote directions relative to the magnetic field.
The distribution of suprathermal electrons in velocity space with polar coordinates
$(v_{\perp} \cos \phi, v_{\perp} \sin \phi, v_{\parallel}) = (v_x, v_y, v_z)$ is described
by a bi-Kappa distribution function
\begin{align}
F^\kappa (v_{\parallel}, v_{\perp}) = & {1 \over \pi^{3/2}
\theta_{\parallel} \theta_{\perp}^2} \, {\Gamma[\kappa] \over
\kappa^{1/2} \Gamma[\kappa - 1/2]}\notag \\
& \times \left(1 + {v_{\parallel}^2\over \kappa
\theta_{\parallel}^2 } + {v_{\perp}^2\over \kappa
\theta_{\perp}^2 }\right)^{-\kappa-1}, \label{e1}
\end{align}
which is normalized to unity, and where $\theta_{\parallel, \perp}$
are thermal velocities defined by, respectively, the parallel and perpendicular temperatures as moments of second order
\begin{align}
T^\kappa_{\parallel} = {m \over k_B} \int d{\bf v} v_\parallel^2
F^\kappa (v_{\parallel}, v_{\perp}) = {\kappa \over \kappa-3/2}
{m \theta_{\parallel}^2\over 2 k_B}, \label{e2}
\end{align}
\begin{align}
T^\kappa_{\perp} ={m \over 2 k_B} \int d{\bf v} v_\perp^2
F^\kappa (v_{\parallel}, v_{\perp}) = {\kappa \over
\kappa-3/2} {m \theta_{\perp}^2 \over 2 k_B}. \label{e3}
\end{align}
The bi-Kappa simply reduces to a bi-Maxellian in the limit of a very large
$\kappa \to \infty$
\begin{align}
F^M (v_{\parallel}, v_{\perp}) = & {1 \over \pi^{3/2}
\theta_{\parallel} \theta_{\perp}^2} \, \exp \left(- {v_{\parallel}^2\over
\theta_{\parallel}^2 } + {v_{\perp}^2\over \theta_{\perp}^2 }\right), \label{e4}
\end{align}
with
\begin{align}
T^M_{\parallel} = {m \over k_B} \int d{\bf v} v_\parallel^2
F^M(v_{\parallel}, v_{\perp}) = {m \theta_{\parallel}^2\over 2k_B} < T^\kappa_{\parallel} , \label{e5}
\end{align}
\begin{align}
T^M_{\perp} = {m \over 2 k_B} \int d{\bf v} v_\perp^2
F^M(v_{\parallel}, v_{\perp}) = {m \theta_{\perp}^2\over 2 k_B} < T^\kappa_{\perp}. \label{e6}
\end{align}
Notice in case that the temperature of suprathermal electrons decreases with increasing the
power-index $\kappa$ and reaches a minimum for the Maxwellian limit.
In the direction parallel to the magnetic field (${\bf k} \parallel
{\bf B}$), the electromagnetic (EM) modes are decoupled from the electrostatic
oscillations, and are described by the following general
dispersion relation \citep{Ga1993}
\begin{align}
{k^2c^2 \over \omega^2}= & 1 + {4 \pi \over \omega^2} \sum_a {e_a
\over m_a} \int_{-\infty}^{\infty} \, {dv_{\parallel} \over \omega -
k v_{\parallel} \pm \Omega_a}
\int_0^{\infty} \, dv_{\perp} \notag \\
& \times v_{\perp}^2 \left[(\omega - k v_{\parallel}) {\partial
F_{a} \over \partial v_{\perp}} + k v_{\perp} {\partial F_{a} \over
\partial v_{\parallel}} \right], \label{e7}
\end{align}
where $\omega$ and $k$ are respectively, the frequency and the
wavenumber of the plasma modes, $c$ is the speed of light in vacuum,
$\Omega_a = q_a B_0 / (m_a c)$ is the gyrofrequency for the particles of sort $a$, e.g., $a = e$ for electrons
and $a=p$ for protons, respectively, and "$\pm$" describes the
circularly polarized EM modes with right-hand (RH) and
left-hand (LH) polarizations, respectively. For the advanced model
introduced in equation (\ref{e1}) the dispersion relation becomes
\begin{align}
{k^2c^2 \over \omega^2} = 1 & + \sum_a {\omega_{a,h}^2 \over \omega^2}
\Big[A_a-1 \notag \\ & + {(A_a-1)(\omega \pm \Omega_a) + \omega \over k
\theta_{a,\parallel}} \, Z_{\kappa} \left(\omega \pm \Omega_a \over k
\theta_{a,\parallel} \right) \Big], \label{e8}
\end{align}
where $A_a = T_{a,\perp} /T_{a,\parallel}$ is the temperature
anisotropy,
\begin{align}
Z_{\kappa}(f) = &{1 \over \pi^{1/2} \kappa^{1/2}} \, {\Gamma
(\kappa) \over \Gamma
\left( \kappa -{1 \over 2}\right)} \notag \\
& \times \int_{-\infty}^{\infty} dx \, {(1+x^2/\kappa)^{-\kappa }
\over x - f}, \;\;\; \Im(f)>0 \label{e9}
\end{align}
is the Kappa plasma dispersion function \citep{La2008} of
argument
\begin{align}
f_{\kappa} = {\omega \pm \Omega_a \over k \theta_{a,\parallel}}.
\label{e10}.
\end{align}
In the Maxwellian limit this function reduces to the standard plasma
dispersion function \citep{Fr1961}
\begin{align}
Z(f) = & {1 \over \pi^{1/2}}
\int_{-\infty}^{+\infty} dx \, {\exp (-x^2) \over x - f}, \; \;\;
\Im (f) > 0 \label{e11}
\end{align}
of argument
\begin{align}
f = {\omega \pm \Omega_a \over k w_{a}}. \label{e12}
\end{align}
Note that for our model introduced in Eqs.~(\ref{e1})--(\ref{e6}), the anisotropy does not depend on $\kappa$,
i.e., $A = T^\kappa_{\perp}/T^\kappa_{\parallel} = T^M_{\perp}/T^M_{\parallel}$.
We investigate the EFHI, which is a LH EM mode driven unstable by an excess of electron temperature in parallel direction
$T_{e,\parallel} > T_{e,\perp}$, i.e., $A_e < 1$. According to (\ref{e8}),
the dispersion relation describing these modes can be rewritten with normalized
quantities as follows
\begin{eqnarray}
\mu \left[ A_{e}-1+ \frac{ A_e \left(\tilde{\omega}+\mu \right) -\mu }{\tilde{k}
\sqrt{\mu \beta_{e,\parallel}^M}} Z_\kappa\left( \frac{\tilde{\omega}+\mu }{\tilde{k}\sqrt{\mu \beta _{e,\parallel}^M}}\right) \right] \notag \\
+\frac{\tilde{\omega}}{\tilde{k} \sqrt{\beta _{e,\parallel}^M /\Theta}} Z\left( \frac{\tilde{\omega}-1}
{\tilde{k}\sqrt{\beta_{e,\parallel}^M / \Theta}}\right) =\tilde{k}^{2}, \label{e13}
\end{eqnarray}
where protons are assumed Maxwellian and isotropic $A_p = 1$, and $\tilde{\omega}=\omega /\Omega _{p}$,
$\tilde{k}=~kc/\omega_{p,p}$, $\mu =~m_{p}/m_{e}$ is the
proton/electron mass ratio, $\Theta=~$~$T_{e,\parallel}^M/T_{p,\parallel }^M\ $ is
the electron/ proton parallel temperature ratio in the Maxwellian
limit for both species, and $\beta_{e, \parallel}^M=~8\pi
n_{e}k_{B}T_{e,\parallel}^M/B_{0}^{2}$ is the parallel electron beta
parameter in the Maxwellian limit $\kappa \to \infty$. The dispersion relation for
bi-Maxwellian distributed electrons can be obtained from Eq. (\ref{e13}) only by changing $Z_\kappa$
with the Maxwellian plasma dispersion function from (\ref{e11}).
\section{EFHI. Thresholds vs. suprathermal electron anisotropy
We have solved the dispersion relation~(\ref{e13}) numerically, and analyzed the unstable firehose
solutions. In this section we present the main features of the EFHI, namely, growth rates, wave-frequencies
and wave-numbers, as well as the anisotropy thresholds, and restrict our analysis only to the unstable
regimes controlled mainly by the suprathermal electrons. The effects of suprathermal populations are triggered
by their temperature anisotropy and their abundance, which is quantified by the finite (especially low) values of the power-index $\kappa$.
Firstly, we examine the growth rates and the wave-frequency of the EFH instability for different plasma regimes conditioned in principal by the (parallel) plasma beta parameter, $\beta_\parallel$,
the electron anisotropy $A$, and the power-index $\kappa$. The regimes identified in Figures~\ref{f1} and
\ref{f2} are specific to the firehose instability, when a magnetized plasma becomes penetrable
by the LH electromagnetic fluctuations propagating parallel to the magnetic field with frequencies
higher than the proton cyclotron frequency. All the unstable modes, i.e., with $\gamma > 0$ in
Figure~\ref{f1}, exhibit this property that becomes evident in Figure~\ref{f2}, where their
wave-number dispersion extends to high frequencies exceeding $\Omega_p$. In the presence of
suprathermals, i.e., at low values of $\kappa$, the range of unstable wave-nunmbers is
restrained, but the wave-frequencies and the instability growth-rates are enhanced. These
effects are in general stimulated by
increasing the plasma beta parameter $\beta$, the temperature anisotropy and the electron-proton
temperature contrast $\Theta$. Plots evidencing the influence of $\Theta$ are not shown here,
but details about this influence are explicitly given in the text. The unstable solutions
displayed in Figures~\ref{f1}-\ref{f4} are obtained for the same value of this parameter, namely,
for $\Theta=4$ in accordance to the observations in the slow solar wind \citep{Ne1998}.
\begin{figure}
\centering
\includegraphics[width=80mm]{f1a}
\includegraphics[width=80mm]{f1b}
\includegraphics[width=80mm]{f1c}
\caption{Effects of the suprathermal electrons quantified by the power-index $\kappa=$2, 2.3, 3, 6,
$\infty$, on the growth rates of the EFH instability for different plasma beta explicitly given in each
panel.} \label{f1}%
\end{figure}
\begin{figure}
\centering
\includegraphics[width=80mm]{f2a}
\includegraphics[width=80mm]{f2b}
\includegraphics[width=80mm]{f2c}
\caption{Effects of the suprathermal electrons ($\kappa=$2, 2.3, 3, 6, $\infty$) on the wave-frequency of the
EFH instability for the same cases considered in Fig.~\ref{f1}.} \label{f2}%
\end{figure}
At higher values of $\kappa$
the instability conditions may be not satisfied and the electromagnetic modes are damped,
e.g., $\gamma < 0$ for $\kappa \geqslant 6$ in Figure~\ref{f1}, middle and bottom panels.
For these modes, the wave-frequency dispersion curves displayed in Figure~\ref{f2} have
a different allure, showing an asymptotic increase similar to the ion (proton)
cyclotron modes with frequencies always smaller than $\Omega_p$. These are LH modes damped
by the protons and therefore limited only to the large (proton) scales.
At smaller scales controlled by the electrons (higher wave-numbers) these
modes change (mode conversion) to RH polarization (i.e., the wave-frequency displayed in
Figure~\ref{f2} becomes negative) which is more specific to the electron whistlers.
In Figures~\ref{f3} and \ref{f4} we show that these LH-polarized modes with a wave-number
dispersion resembling that of the electromagnetic ion cyclotron (EMIC) modes can be destabilized
by the anisotropic bi-Kappa distributed electrons, see middle and bottom panels.
This is a new regime of the EFHI
destabilizing only the low-frequency branch of the LH modes with wave-frequency showing
an asymptotic increase of their wave-frequencies but remaining always below $\Omega_p$.
To establish this regime the kinetic effects of the electrons are also tempered by considering lower
values of plasma beta, and the instability is triggered only by the anisotropic
distributions with sufficiently low $\kappa$, e.g., $\kappa < 3$ in Figures~\ref{f3} and \ref{f4}.
Furthermore, in this case, both the (maximum) growth-rates and the range of the unstable
wave-numbers are considerably enhanced by increasing the presence of suprathermals, i.e.,
lowering the values of $\kappa$. Again, these features seems to be more specific to the instability
of the cyclotron modes \cite{Shaaban2016JGR}.
The transition between the classical EFH solutions (exemplified in Figures~\ref{f1}
and \ref{f2}) and the new regime of a low-frequency EFHI is suggestively shown by the
top panels in Figures~\ref{f3} and \ref{f4}.
In these panels we have unstable solutions specific to both these regimes: the solid-line
solution obtained for $\kappa = 2$ is a classical firehose, while the next long-dashed-line
solution obtained for $\kappa = 2.3$ is already more specific to the new regime of EFHI.
In this case it is only the power-index $\kappa$ that may switch between these two regimes,
but a direct comparison of the other plasma parameters in Figures~\ref{f1}-\ref{f4}, clearly
shows that these regimes are also conditioned by the temperature anisotropy, the plasma beta,
and the temperature contrast between electrons and protons.
\begin{figure}
\centering
\includegraphics[width=85mm]{f3a}
\includegraphics[width=85mm]{f3b}
\includegraphics[width=85mm]{f3c}
\caption{Effects of the suprathermal electrons quantified by the power-index $\kappa=$2, 2.3, 3, 6,
$\infty$, on the growth rates of the EFH instability for a lower $\beta_\parallel = 0.6$.} \label{f3}%
\end{figure}
\begin{figure}
\centering
\includegraphics[width=80mm]{f4a}
\includegraphics[width=80mm]{f4b}
\includegraphics[width=80mm]{f4c}
\caption{Effects of the suprathermal electrons ($\kappa=$2, 2.3, 3, 6, $\infty$) on the wave-frequency of the
EFH instability for the same cases considered in Fig.~\ref{f3}.} \label{f4}%
\end{figure}
\begin{figure}
\centering
\includegraphics[width=80mm]{f5a} \\
\includegraphics[width=80mm]{f5b}
\caption{Comparison of the anisotropy thresholds (\ref{e14}) for maximum growth rates $\gamma_m / \Omega_p
= 10^{-3}$ with the temperature anisotropy measured in the solar wind which is displayed using a scatter
plot data in the top panel and a histogram data in the bottom panel.} \label{f5}%
\end{figure}
In the second part of this section we analyze the anisotropy thresholds of the instability.
These thresholds represent plasma conditions associated with given values of the maximum growth-rate,
usually small values, e.g., $\gamma_m /\Omega_p = 10^{-2}$, $10^{-3}$, approaching the marginal
stability $\gamma_m/\Omega_p \to 0$. In Figure~\ref{f5} we display the instability thresholds
associated with $\gamma_m/\Omega_p=10^{-3}$ and derived for different values of the electron
power-index $\kappa$. These are isocontours of the electron temperature anisotropy $A$ as a
function of the parallel electron plasma beta $\beta_{\parallel}$, fitted to an inverse
correlation law of the form \citep{Ga1994,Ga1998}
\begin{equation}
A=1+\frac{a}{\beta _\parallel^b}. \label{e14}
\end{equation}
The values obtained for the fitting parameters $a$ and $b$ can be found in Table~\ref{t1}.
For the plasma beta parameter we consider an extended range of values $0.1<\beta_{\parallel}<50$
relevant for the electron halo populations in the solar wind \citep{St2008}. Higher
values of $\beta_{\parallel}$, associated with hotter plasmas or less intense magnetic
fields, means lower deviations from isotropy to trigger the instability.
The effects of suprathermal electrons is reconfirmed here by a systematic stimulation of
the (maximum) growth-rates with decreasing $\kappa$. As a consequence,
the anisotropy thresholds are found to be markedly lowered in the presence of suprathermals,
and this effect may be enhanced by increasing the temperature contrast between electrons
and protons. Larger variations of the anisotropy thresholds are obtained at lower values
of $\kappa$.
The instability thresholds are compared in Figure~\ref{f5} with the observational data of
the electron halo populations in the slow solar wind ($v< 500$ km/s), which are displayed
in the top panel as a scatter plot, and in the bottom panel as a histogram counting the
number of events within a color logarithmic scale. This data set comprises more than 120 000
events detected by three space missions (Helios 1, Cluster II, and Ulysses) at different
heliocentric distances (in the interval 0.3--3.95~AU) in the ecliptic. The details about
the electron analyzers used by these missions, and the methods of correction and
reconstruction of the 3D velocity distribution functions can be found in \cite{St2008}.
These authors have used the same set of events to analyze the temperature anisotropy
of the main electron populations, namely, the thermal core and suprathermal halo, and the
most plausible constraints exercised on their temperature anisotropy by different physical
mechanisms, e.g., collisions and kinetic instabilities. However, \cite{St2008} have limited
to investigate in detail only the bi-Maxwellian core anisotropy finding that the particle-particle
collisions still may have an effect to constrain low levels of anisotropy, while the kinetic
instabilities occur for larger deviations from isotropy, which exceed their thresholds.
Indeed, the instability thresholds predicted by a bi-Maxwellian model are found to shape very
well the limits of the core anisotropy, but the same thresholds cannot explain the limits observed
for the temperature anisotropy of suprathermal electrons. Figure~6 from \cite{St2008} presents
such a comparison between the observational data and the EFHI thresholds predicted by a bi-Maxwellian
approach, which is also reproduced here in Figure~\ref{f5} by the dotted line corresponding to $\kappa \to \infty$.
Moreover, in Figure~\ref{f5} we show that this disagreement may be resolved by the instability thresholds
derived for bi-Kappa models which are more appropriate to describe the suprathermal electrons.
The instability thresholds are markedly changed with decreasing the power index $\kappa$ and
for lower values of $\kappa$ these thresholds are approaching the limits of the temperature
anisotropy observed in the solar wind. What we found even more interesting is that the instability
thresholds also shape very well the isocontours of the observational data, counting the number
of events in the bottom panel.
Developing for large temperature anisotropies exceeding these thresholds, the FHI dissipates
the free energy and enhances the electromagnetic fluctuations, which may also prevent the
anisotropy to grow by scattering particles back towards quasi-equilibrium states.
Given that suprathermal electrons in the solar wind are practically collisionless, such a
good agreement between the instability thresholds predicted by the kinetic theory and the limits of
the temperature anisotropy reported by the observations represents an important confirmation
on the role played by the FHI instability in the relaxation process.
\begin{table}
\centering \caption{Fitting parameters for thresholds $\gamma_{\rm m}/\Omega_p = 10^{-3}$} \label{t1}
\begin{tabular}{c c c c c c c c }
\hline
Fit & $\kappa=2$& $\kappa=2.3$ & $\kappa=3$ & $\kappa=6$ &$\kappa\rightarrow\infty$ \\
\hline
$a$ &-0.1196 & -0.3304 & -0.9615 & -1.3733 & -1.7950 \\
$b$ & 0.8708 & 0.8996 & 1.0009 & 1.0030 & 1.0456 \\
\hline
\end{tabular}
\end{table}
\section{Discussions and conclusions
In this paper we have proposed a refined theory of the electron firehose instability
in anisotropic Kappa distributed plasmas, which provide a new and, in our opinion, valuable evidence
of an extended implication of this instability in the relaxation of the temperature anisotropy
in collision-poor plasmas from space.
Our present study is particularly motivated by the solar wind observations which do not
confirm the indefinite increase of temperature predicted by the solar wind expansion in the
direction parallel to the interplanetary magnetic field, but reveal very clear bounds for the
temperature anisotropy of plasma particles. Previous studies have focused to the
thermal (core) populations of electrons and protons, using standard bi-Maxwellian approaches,
and have shown that large deviations from isotropy are constrained by the kinetic instabilities
\citep{Hellinger06, St2008}. However, the same bi-Maxwellian is not appropriate to describe
suprathermal populations and their anisotropy, and cannot prescribe accurately the resulting
instabilities and their back reaction on these populations.
Here we have assumed the anisotropic electrons well reproduced by the bi-Kappa distribution
function, which is the empirical model invoked by \cite{St2008} to describe the velocity distribution
of suprathermal electrons in the solar wind. In addition, the temperature of the suprathermal
population is considered dependent on the power-index $\kappa$, enabling us a realistic
interpretation of the suprathermals and their effects (theoretical and observational arguments
are detailed in the Introduction). The results of our present study contrast with
those provided by \cite{La2009} and \cite{La2011}, who studied the same EFHI but driven by
bi-Kappa electrons with a $\kappa$-independent temperature. These differences are made clear in
the discussion from this section.
Two distinct regimes of the EFHI are identified in Section~3, and these regimes are differentiated
by the wave-number dispersion laws (curves) obtained for the wave-frequency and growth-rate
of the instability. Thus, more specific to the EFHI are the unstable LH-polarized modes
exemplified in Figures~\ref{f1} and \ref{f2} with frequencies that can significantly exceed
the proton cyclotron frequency $\Omega_p$. For this regime to be established, the kinetic
free energy of the anisotropic electrons
must be sufficiently large, and this usually means a high enough plasma beta or a large
anisotropy. If damped, these modes cannot extend above $\Omega_p$ and
their wave-number dispersion keeps the aspect of low-frequency EMIC modes in the absence of
kinetic anisotropies. At higher wave-numbers (lower scales) these damped modes
can change their polarity converting to the branch of RH-polarized modes (whistlers).
These electromagnetic modes with a wave-number dispersion resembling that of the EMIC modes,
i.e., with wave-frequency increasing asymtotically to $\Omega_p$, can be driven unstable
by the EFHI for conditions approaching marginal stability. A few cases relevant for this new regime are presented in Figures~\ref{f3}
and \ref{f4}, with mention that top panels include unstable solutions representative for a
transition between these two distinct branches of the EFHI.
We should observe that considering plasma parameters with values typical
for the solar wind conditions, e.g., in Figures~\ref{f1}-\ref{f4}, the EFHI develops only in
the presence of suprathermal electrons, i.e., for finite values of $\kappa$,
while for (bi-)Maxwellian limit $\kappa \to \infty$ these modes are damped.
Increasing the presence of suprathermal populations (by lowering $\kappa$)
has opposite effects on the wave-frequency of the unstable modes, which become evident if we compare
for instance Figures~\ref{f2} and \ref{f4}. However, the EFHI is clearly stimulated by the
suprathermal electrons, which enhance the (maximum) growth-rates in both these two regimes.
Noticeable is the significant increase shown by the growth-rates for conditions approaching
the marginal stability (Figure~\ref{f3}), which can also explain the significant decrease of
the instability thresholds shown in Figure~\ref{f5}. These thresholds are markedly lowered with
decreasing $\kappa$, and for lower values of $\kappa$ they shape very well the limits
of temperature anisotropy
reported by the observations in the solar wind. In the previous studies involving bi-Kappa
electrons with a $\kappa$-independent temperature, e.g., in \cite{La2009, La2011}, the existence
of these two distinct regimes was not mentioned, and, in contrast to our present results,
the suprathermals were found inhibiting the EFHI, and departing the instability thresholds
from the limits of temperature anisotropy in the solar wind.
Our present results strongly suggest that the EFHI may efficiently constrain the
temperature anisotropy of the suprathermal electrons in the slow wind, complementing the results by
\cite{St2008}, which showed the same effects of this instability on the core electrons.
A good agreement between the instability thresholds and the bounds of the temperature
anisotropy measured in the solar wind is conditioned by a proper modelling of the velocity
distributions in accord to the observations. In conclusion, the EFHI can be considered a
plausible mechanism of electron energy transfer between the directions parallel and
perpendicular to the uniform magnetic field. From an extended perspective, we can further claim
that the resulting low-frequency fluctuations can establish an energy transfer from small
to large scales, namely from the electrons, especially the energetic or suprathermal electrons
which carry the main heat flux in the solar wind, to the resonant protons. Although suprathermal
populations are not easily captured in numerical experiments, it becomes however clear that
our present results provide valuable premises that may stimulate new and advanced simulations
to confirm these mechanisms.
\section*{Acknowledgments}
The authors acknowledge support from the Katholieke Universiteit Leuven,
Ruhr-University Bochum, and Alexander von Humboldt Foundation.
These results were obtained in the framework of the projects
GOA/2015-014 (KU Leuven), G0A2316N (FWO-Vlaanderen), and C 90347
(ESA Prodex 9). The research leading to these results has also
received funding from the European Commission's Seventh Framework
Programme FP7-PEOPLE- 2010-IRSES-269299 project-SOLSPANET
(www.solspanet.eu). S.M. Shaaban would like to thank the Egyptian Ministry
of Higher Education for supporting his research activities.
|
\section{Introduction}
Quantum entanglement not only plays an elementary role in quantum physics, but also has wide applications in quantum information processing (QIP) tasks, for instance quantum key distribution \cite{gisin}, quantum teleportation \cite{bennett1}, quantum metrology \cite{pezze}, quantum cryptography \cite{bennett2} and quantum computation \cite{steane}.
The structure and the characteristics of bipartite entanglement is well understood, while theoretical and experimental exploration of multipartite entanglement is still a great challenge.
Among many interesting properties of multipartite entanglement, a major feature is the presence of inequivalent classes of multipartite entangled states, such as Greenberger-Horne-Zeilinger (GHZ) states \cite{greenberger} and W states \cite{dur1}, and states in one class cannot be converted into a state in another class by local operations and classical communication \cite{dur2}.
The entanglement of W states is more robust against particle losses than that of GHZ states \cite{dur1}.
A W state of appropriate size is a superior candidate for leader election in anonymous quantum networks \cite{hondt} and required for several secure quantum communication protocols \cite{joo1,wang1,cao,liu}.
Quantum coherence of a W state is also unique and introduces new advantages in quantum thermodynamics.
It has recently been found that in the thermalization process of a single mode cavity by three-atom systems, W states achieve the maximum thermalization, in contrast to GHZ states which provide no thermalization \cite{Ceren1}.
This interesting finding suggests that a large-scale atomic W state, as a quantum information molecule, can be good a candidate to be a novel quantum energy resource in quantum heat engines \cite{Ceren2}. The increasing of the size of the quantum system is essential to overcome the decoherence effects and to increase the thermalization capability \cite{Hardal1,Altintas1,Turkpence1}.
Therefore, besides realizing QIP tasks, quantum heat engines require experimentally accessible schemes for preparing large-scale atomic W states.
Several schemes have been presented for preparing and characterizing multipartite entangled W states in various physical systems, for instance cavity quantum electrodynamics (QED) system \cite{deng1}, nuclear magnetic resonance (NMR) system \cite{dogra1,vandersypen,laamme,dogra2}, optical system \cite{eibl1,shi}, superconducting quantum system \cite{deng2,devorer,huang} and ion traps system \cite{duan,roos,haffner}.
However, the achievable sizes of the prepared W states are far from being large-scale.
It was recently found that a W state of polarization encoded photons can be expanded by adding one or two photons at a time via accessing only one photon of the W state \cite{tashima3,tashima4,tashima5,tashima6} and that two EPR pairs \cite{tashima1} and even two arbitrary size W states can be fused to obtain a larger size W state \cite{tashima2} with simple optical setups.
At a cost of integrating three-qubit gates, it is possible to increase the probability of success of the fusion process \cite{Ozaydin3}, or to fuse several W states to arrive at a large-scale W state with a fewer number of fusion processes, to ease the practical implementations \cite{Ozaydin2,Ozaydin1}.
When it comes to atomic W states, we have used light-matter interface to design expansion and fusion schemes for generating large-scale atomic W states in cavity QED \cite{zang1, zang2, zang3}.
Although the fusion and expansion approaches via accessing a single qubit of the state enable preparing large-scale W states, their probabilistic nature still requires at least a sub-exponential resource \emph{cost} with respect to the target size, in terms of the spent entanglement during the preparation process.
What is more, as Ozaydin \emph{et al.} demonstrated via the Monte Carlo simulation in Ref.\cite{Ozaydin1}, the similar scaling of the number of fusion attempts physically limits the experimental accessibility to large-scale W states.
The probabilistic nature of these approaches also requires detection mechanisms for post-selection to find out whether the attempt has been successful or not, not only introducing additional complexity to the system but also shrinking the final size of the prepared W state in the successful attempts.
An additional inherent weakness of most of the above mentioned schemes is that they require prior entanglement, in particular tripartite W states or at least W-type Bell pairs to start and to sustain the process i.e. they actually require another scheme to prepare these primary resources.
Since the \emph{cost} of the generation process in terms of the total entanglement spent throughout the process considerably depends on the \emph{cost} of the primary resources to be spent, Yesilyurt \emph{et al.} proposed a scheme for deterministic generation of W states of four polarization-based entangled photons, to serve as the primary resource \cite{Can2015Acta}.
Very recently, Yesilyurt \emph{et al.} designed an all-optical expansion mechanism for polarization based entangled photons, which can deterministically expand an $n$-photon W state to a $2n$-photon W state using $n$ ancillary photons \cite{ozaydin4}.
The expansion mechanism used there consists of a controlled-NOT (CNOT) gate and a controlled Hadamard (CH) gate, the latter being decomposable into a CNOT gate and single qubit gates.
Although this mechanism can be realized easily with the current technology for polarization based photonic qubits, this is not the case for the matter qubits, due to the difficulty in experimental realization of the controlled gates.
Therefore, a new, simple, and feasible deterministic expansion mechanism must be designed for preparing matter-qubit W states.
In this paper, we present a simple scheme for deterministically generating large-scale atomic W states by expanding small-size atomic W states in a cavity QED system.
Detuned interactions between a cavity field and two atoms constitute the new deterministic expansion mechanism.
The velocity of the atoms is selected for realizing the deterministic expansion mechanism, which is much easier than realizing the controlled gates, and the feasibility analysis at the end of the paper shows that our scheme can be realized with the current technology.
\section{Expansion mechanism}
The key step of our expanding scheme is the detuned interaction between two identical two-level atoms and a cavity mode.
Assume that two atoms are simultaneously sent through the single-mode cavity, and the interaction Hamiltonian can be described in the interaction picture as \cite{zheng}
\begin{equation}\label{1}
H_i=g\sum_{j=1,2}(e^{-i \delta t} a^\dag S^{-}_j+e^{i\delta t}a S^{+}_j),
\end{equation}
where $S^+_j=|e_j\rangle\langle g_j|$ and $S^-_j=|g_j\rangle\langle e_j|$, with $|g_{j}\rangle$ and $|e_{j}\rangle$ representing the ground and excited states of the $j$th atom, respectively. $a$ and $a^\dagger$ denote, respectively, the annihilation and creation operator of the cavity mode, and $g$ is the coupling strength between the cavity mode and each atom.
The atomic transition frequency is described by the parameter $\omega_0$, $\omega$ is the cavity mode frequency, and $\delta=\omega_0-\omega$ is the detuning between them.
In this scheme, the interaction is detuned, therefor there is no energy exchange between the cavity mode and the atomic system, and the condition is described as $\delta\gg g$.
The effective Hamiltonian can be described as
\begin{equation}\label{2}
H=\lambda[\sum_{j=1,2}(|e_j\rangle\langle e_j|aa^\dag -|g_j\rangle\langle g_j|a^\dag a)+(S^+_1S^-_2+S^-_1S^+_2)],
\end{equation}
where $\lambda=g^2/\delta$. Assume the cavity field is initially prepared in the vacuum state, thus the system effective Hamiltonian reduces to
\begin{equation}\label{3}
H_{eff}=\lambda[\sum_{j=1,2}|e_j\rangle\langle e_j|+(S^+_1S^-_2+S^-_1S^+_2)].
\end{equation}
By solving the Schr\"{o}dinger equation, we can obtain the system evolution of each initial state as \cite{zheng}
\begin{eqnarray}\label{4}
|e_1\rangle|e_2\rangle &\longrightarrow & e^{-i2\lambda t}|e_1\rangle|e_2\rangle,\nonumber \\
|e_1\rangle|g_2\rangle &\longrightarrow & e^{-i\lambda t}(\cos\lambda t|e_1\rangle|g_2\rangle-i\sin\lambda t|g_1\rangle|e_2\rangle), \nonumber \\
|g_1\rangle|e_2\rangle &\longrightarrow & e^{-i\lambda t}(\cos\lambda t|g_1\rangle|e_2\rangle-i\sin\lambda t|e_1\rangle|g_2\rangle), \nonumber \\
|g_1\rangle|g_2\rangle &\longrightarrow & |g_1\rangle|g_2\rangle,
\end{eqnarray}
\noindent constituting the basis of our mechanism.
In the following section, we first present how to use our mechanism to generate a four-atom W state from four atoms initially in a separable state, as an example.
\begin{figure}\label{Fig1}
\centering
\includegraphics[width=0.40\textwidth]{Fig1.eps}
\caption{(color online). The setup for deterministic generation of a four-atom W state from four atoms illustrated with blue, red, red and blue spheres, respectively, initially in the separable state $|g\rangle_{a1} |e\rangle_{a2} |g\rangle_{b2} |g\rangle_{b1}$.
In Step 1, an EPR pair is prepared, which is then expanded in Step 2 to a W state of four atoms, illustrated with green spheres labeled as $a1'$, $a2'$, $b2'$ and $b1'$.}
\end{figure}
\section{Deterministic generation of four-atom W state via generating and expanding an EPR Pair}
As depicted in FIG.~\ref{1}, there are two steps in our deterministic generation scheme: creating a $W_2$ state, and expanding $W_2$ state to $W_4$ state.
The four two-level atoms ($a1$, $a2$, $b1$ and $b2$) to be entangled are supposed to be identical.
In the first step, atoms $a2$ and $b2$ which are initially in the excited and ground states, respectively, are sent through the cavity $C$ simultaneously.
The detuned interaction between the atoms and the cavity field induces a $W_2$ state, given that the velocity of the atoms are adjusted appropriately.
After these two atoms fly out the cavity $C$, the second step starts, where two ancillary atoms, $a1$ and $b1$ which are initially in ground states, are introduced with adjusted velocities.
Atoms $a1$ and $a2$ are sent through cavity $A$, and atoms $b1$ and $b2$ are sent through cavity $B$, simultaneously.
The expansion process is realized by these two detuned interactions, generating a four-atom W state deterministically.
In more detail, in the first step, the atoms $a2$ and $b2$ are initially prepared in the excited state $|e\rangle_{a2}$ and ground state $|g\rangle_{b2}$, respectively, and their velocities are adjusted so that the detuned interaction time satisfies $\lambda t=\pi/4$.
Via Eq.(\ref{4}), it is easy to see that a $W_2$ state
\begin{equation}\label{5}
|W_{2}\rangle= \frac{1}{\sqrt{2}}(|e\rangle_{a2}|g\rangle_{b2}-i|g\rangle_{a2}|e\rangle_{b2})
\end{equation}
is generated between the two atoms $a2, b2$ as flying out of the cavity $C$, with unit probability.
Here a global phase factor has been omitted.
In the second step, two ancilla atoms $a1, b1$ are both prepared in the ground state ($|g\rangle_{a1}$ and $|g\rangle_{b1}$), and their velocities are adjusted so that the two detuned interactions in cavities $A, B$ have the same duration satisfying $\lambda t=\pi/4$ too.
After leaving the cavities, the state of the four-atom system is found as:
\begin{widetext}
\begin{equation}\label{6}
|W_4\rangle=\frac{1}{\sqrt{4}}(|gegg\rangle_{a1^{'}a2^{'}b1^{'}b2^{'}}-i|egggg\rangle_{a1^{'}a2^{'}b1^{'}b2^{'}}-i|ggge\rangle_{a1^{'}a2^{'}b1^{'}b2^{'}}-|ggeg\rangle_{a1^{'}a2^{'}b1^{'}b2^{'}}),
\end{equation}
\end{widetext}
where the global phase factor has been omitted.
Since the four-atom W state is generated with unit probability, no quantum measurement is required in the generation scheme, therefore the complexity of the scheme is reduced considerably.
A consecutive round of four expansion schemes in parallel can expand this four-atom W state to an eight-atom W state, and so on.
Therefore, in principle, our scheme enables the deterministic generation of any even-number-atom W state, starting with atoms in a separable state, i.e. requiring no prior entanglement.
In the next section, we demonstrate how a $2n$-atom W state can be deterministically expanded from an $n$-atom W state.
\section{Deterministic generation of $2n$-atom W state via expanding an $n$-atom W state}
An $n$-atom $W$ state is expressed as $|W_n\rangle=|(n-1)_g,e\rangle/\sqrt{n}=[|(n-1)_g\rangle_{\widetilde{a1}}\otimes |e\rangle_{a1}+\sqrt{n-1}|W_{n-1}\rangle_{\widetilde{a1}}\otimes |g\rangle_{a1}]/\sqrt{n} $, where the subscript $a1$ denotes the $a1$th atom and $\widetilde{a1}$ denotes the rest $(n-1)$ atoms of $|W_n\rangle$.
The evolution of a $W_n$ state together with $n$ ancilla atoms through $n$ parallel expansion mechanisms can be obtained by the following equations
\begin{equation}\label{7}
|(n-1)_g\rangle_{\widetilde{a1}}|e\rangle_{a1}\otimes|n_g\rangle\longrightarrow|2(n-1)_g\rangle|W_2\rangle,
\end{equation}
\begin{equation}\label{8}
|W_{n-1}\rangle_{\widetilde{a1}}|g\rangle_{a1}\otimes|n_g\rangle\longrightarrow|W_{2(n-1)}\rangle|2_g\rangle.
\end{equation}
Therefore, the final state of the whole system is
\begin{equation}\label{9}
|W_{2n}\rangle=\frac{1}{\sqrt{n}}[|2(n-1)_g\rangle|W_2\rangle+\sqrt{n-1}|W_{2(n-1)}\rangle|2_g\rangle],
\end{equation}
which is a standard $2n$-atom W state.
\section{Results and Discussion}
We have presented a new expansion mechanism which can generate a Bell state from a product state of two atoms, and can double the size of a given atomic W state of two or more atoms, deterministically.
It is well-known that if an entangled W state with odd number of atoms ($|W_{2n+1}\rangle$) is required, one can first deterministically generate the state $|W_{2n+2}\rangle$ and then via a projective measurement on any atom of the state $|W_{2n+2}\rangle$, will be left with the remaining $(2n+1)$ atoms in the state $|W_{2n+1}\rangle$, if the measurement result is $|g\rangle$, with a probability of success $1-1/2(n+1)$ which approximates to $1$ for large $n$.
Since the large detuning interaction between the cavity mode and the atoms adiabatically eliminates the excitation states of cavity mode during the interaction, the cavity decay time does not affect the scheme.
The radiative time of a Rydberg atom with principal quantum number $49$, $50$ or $51$ is $T_r=3\times10^{-2} s$, the atom-cavity coupling strength is $g=2\pi\times24$ kHz \cite{zheng,yang1,brune}, and the choice $\delta=10g$ satisfies the large detuning condition, i.e. the interaction time in each step between the cavity mode and atoms is in the order of $\pi\delta/g^2\simeq2\times10^{-4}s$, implying that the time required for accomplishing the whole process is approximately $10^{-3}s$, which is much shorter than $T_r$.
Therefore, the interaction time of our scheme is much shorter than the radiative time of atoms, and thus the deterministic expansion scheme is realizable with the current cavity-QED technology. Actually, the basic expansion mechanism described in Eq.(\ref{1}), i.e. a coherent control of an atomic collision in a detuned cavity has been experimentally demonstrated \cite{Osnaghi}.
Since the radiative time of the Rydberg atom is about $30$ times that of the interaction time, expansion mechanism can be repeated $30$ times in principle, thus the number of atoms of the final W state can reach up to $2^{30}$.
\section{Conclusion}
In conclusion, a new expansion scheme based on detuned interaction between a vacuum cavity mode and two atoms is proposed to double the size of a given atomic W state, with a unit probability of success.
The scheme can not only expand a given W state but it can also generate a W-type Bell state from a product state of two atoms, allowing one to prepare an arbitrarily large scale W state, starting with atoms initially in a separable state, requiring no other setup for preparing any prior entanglement.
No quantum measurements and no controlled operations are required.
The feasibility analysis shows that our expansion scheme is experimentally realizable with the current cavity-QED technology.
Enabling efficient generation of large-scale atomic W states, our scheme promotes advances in quantum information processing, storage and computation tasks, quantum communication protocols, and the extraction of heat and work in quantum heat engines.
We believe that our work may also open new insights in self-testing of atomic W states.
\begin{acknowledgments}
This work is supported by National Natural Science Foundation of China (NSFC) under Grants No.11274010 and No.11374085, the Key Program of the Outstanding Young Talent of Anhui Province under Grant No.gxyqZD2016370, and the personnel department of Anhui province. F. Ozaydin and M. Yang are funded by Isik University Scientific Research Funding Agency under Grant Number: BAP-15B103.
\end{acknowledgments}
|
\section{Introduction}
\label{sec:intro}
Owing to the growing practical importance of space weather, attempts to
predict future levels of solar activity have found much interest in
recent years. A considerable variety of methods has been suggested, of
which the so-called precursor methods proved most successful \citep[see
review by][]{Petrovay:2010}. These methods take the strength of the
magnetic field at the solar polar caps (or some proxy thereof, such as
the open heliospheric magnetic flux, the strength of the radial
interplanetary field, or the level of geomagnetic disturbances) during
activity minimum as indicator for the strength of the subsequent cycle.
The high empirical correlation between precursor level and cycle
strength \citep[e.g.][]{Wang:Sheeley:2009} has a theoretical basis in
the Babcock-Leighton scenario for the solar dynamo
\citep{Charbonneau:2010, Cameron:Schuessler:2015}. Thus a prediction of
the amplitude of the next activity maximum made at a time around
minimum, when the polar fields are fully developed, rests on rather
solid empirical and theoretical ground. The level of uncertainty of such
a prediction can be estimated from the data.
The question is whether and how a prediction at an earlier phase during
a cycle can be made. One possibility is to use surface flux transport
(SFT) simulations, which have quite successfully reproduced the observed
evolution of the large-scale magnetic field at the solar surface in the
course of the solar cycle, particulary also the evolution of the polar
fields and axial dipole moment \citep[e.g.][]{Wang:etal:1989,
Sheeley:2005, Mackay:Yeates:2012, Jiang:etal:2014a,
Upton:Hathaway:2014}. Such simulations also showed that randomness in
the properties of the magnetic flux sources in the form of emerging
bipolar magnetic regions (such as emergence latitude and tilt angle) has
a strong effect: single bipolar regions emerging near the equator can
significantly affect the level of the polar field (or axial dipole
moment) around solar minimum and thus influence the strength of the
subsequent cycle \citep{Jiang:etal:2014b, Jiang:etal:2015}. Therefore,
any prediction of the polar field strength also needs to quantify the
uncertainty that arises from the random scatter of the source
properties.
In the work presented in this paper, we used empirically determined
statistical properties of emerging bipolar regions (sunspot groups)
during the declining phases of solar cycles and ran Monte-Carlo
ensembles of SFT simulations starting from observed synoptic maps of the
surface magnetic field. We thus obtained the axial dipole moment during
solar minima as an ensemble average over the realizations together with
an empirically based estimate of the uncertainty. We prefer to consider
the axial dipole moment and not the polar field strength since it is a
uniquely defined quantity and limits the effect of a hemispherically
asymmetric distribution of magnetic flux. Around activity minima, the
axial dipole moment is dominated by the magnetic flux at the polar caps
and represents the open heliospheric flux. We first tested our approach
by deriving postdictions for the axial dipole moment during the minima
of cycles 21-23, which can be compared with observations. We then ran
simulations for cycle 24 and obtained predictions and statistical
uncertainties for the dipole moment during the upcoming minimum around
2020.
\section{Surface flux transport (SFT) model}
\label{sec:SFT}
\subsection{SFT code}
\label{subsec:SFT_code}
The SFT code used is described in \citet{Baumann:etal:2004} and
\citet{Cameron:etal:2010}. It treats the evolution of the radial
component of the large-scale magnetic field at the solar surface
resulting from passive transport by convection, differential rotation,
and meridional flow. The corresponding magnetohydrodynamic induction
equation is given by
\begin{eqnarray}
\nonumber\frac{\partial B}{\partial t}=& &
-\Omega(\lambda)
\frac{\partial B}{\partial \phi}
- \frac{1}{R_\odot \cos\lambda}
\frac{\partial}{\partial \lambda}[\upsilon(\lambda)
B \cos \lambda] \\ \noalign{\vskip 2mm}
& & +\eta \left[\frac{1}{R_\odot^2 \cos{\lambda}}
\frac{\partial}{\partial \lambda}\left(\cos\lambda
\frac{\partial B}{\partial \lambda}\right) +
\frac{1}{R_\odot^2 \cos^2{\lambda}}\frac{\partial^2 B}{\partial
\phi^2}\right] + S(\lambda,\phi,t),
\label{eqn:SFT}
\end{eqnarray}
where $B(\lambda,\phi,t)$ is the radial component of the magnetic field.
$\lambda$ and $\phi$ are heliographic latitude and longitude,
respectively. We used the profile of latitudinal differential rotation,
$\Omega(\lambda)$, determined by \citet{Snodgrass:1983} and took the
profile of the poleward meridional flow, $\upsilon(\lambda)$, from
\citet{Ballegooijen:1998}. The magnetic diffusivity
$\eta=250$~km$^2$s$^{-1}$ describes the random walk of the magnetic flux
elements as transported by supergranulation flows \citep{Leighton:1964}.
The source term, $S(\lambda,\phi,t)$, represents the emergence of
magnetic flux at the solar surface. The properties of the corresponding
bipolar regions were defined as described in \citet{Baumann:etal:2004}.
Further details are given in \citet{Cameron:etal:2010}.
\subsection{Source selection and initial condition}
\label{subsec:SFT_sources}
Our method requires us to provide synthetic source data (emerging
bipolar magnetic regions, BMRs) for the time period of the
prediction. The properties of these sources were chosen according to
their empirical statistics during the descending phases of
previous activity cycles.
The number of BMRs emerging per month of the simulation was taken to be
equal to $R/2.75$, where $R$ is the monthly sunspot number. This
calibration was carried out on the basis of the sunspot data for cycles
21--24 (between 1976 and 2016). We adopted the functional form
$R_G=f(t)+\Delta f(t)$, where $f(t)$ is given by Eq.~(1) of
\citet{Hathaway:etal:1994} and $\Delta f(t)$ represents the random
scatter of the time evolution. We followed the procedure given in
\citet{Hathaway:etal:1994} to fix the independent parameters $a$
(amplitude) and $t_0$ (starting time) involved in $f(t)$ from
observational data for the cycles considered here. The obtained values
for $(a;t_0)$ are $(0.00351;1976.39)$ for cycle 21, $(0.00336;1986.16)$
for cycle 22, $(0.00229;1996.57)$ for cycle 23, and $(0.00103;2008.84)$
for cycle 24. The standard deviation $\sigma_f(t)$ of $\Delta f(t)$ is
determined from the difference between the fit profiles $f(t)$ and the
observed profiles for cycles 21--23. It is well represented by
$\sigma_f(t) = 0.5 f(t)$.
For the latitudinal distribution of the sources as a function of cycle
phase we used the empirically derived functional forms for mean latitude
and width (standard deviation, $\sigma_\lambda$) of the activity belts
given by Eqs.~(4) and (5) of \citet{Jiang:etal:2011}. The area distribution
for the emerging BMRs and its dependence on cycle phase follows the
empirical profiles represented by Eqs.~(12)--(14) of
\citet{Jiang:etal:2011}.
The mean tilt angle, $\alpha_n(\lambda)$, of the emerging BMRs for cycle
number $n$ is assumed to follow Joy's law in the form
$\alpha_n(\lambda)=T_n\sqrt{|\lambda|}$. $T_n$ was taken from the linear
relation with maximum sunspot number given by Eq.~(15) of
\citet{Jiang:etal:2011}. For the scatter of the tilt angle,
$\sigma_\alpha$ (in degrees), we used the empirical relation with sunspot
umbral area, $A_U$, given by $\sigma_\alpha=-11\log A_U+35.$ as
determined by \citet{Jiang:etal:2014b}. To connect total sunspot area,
$A_S$, and BMR area with umbral area, we used the relation $A_S = 5 A_U$
\citep{Brandt:etal:1990}.
The relevant quantity for the contribution of a BMR to the axial
dipole moment is the latitudinal separation of the two polarities,
which depends on the tilt angle and on the distance between the
polarities. For the latter we employ its empirical relation with the
BMR area as derived by \citet{Cameron:etal:2010}. Here the BMR area is
defined as the sum of sunspot area and plage area, using the
relationship derived from observations by \citet{Chapman:etal:1997}.
The distribution of latitudinal polarity distance derived by this
procedure agrees well with that determined directly from historical
sunspot data. This applies particularly to the tails of the
distribution, which are most important for the buildup of the polar
fields.
The procedure for defining the sources is then as follows: for each
month of the simulation, the number of emerging BMRs is determined
according to the value of $f(t)$ and a random deviation $\Delta
f(t)$. These BMRs are distributed randomly over the days of the
month. For each emerging BMR, its size is chosen randomly from the
(time-dependent) size distribution. Its latitude is given by the
(time-dependent) mean latitude plus a random component drawn from a
distribution with standard deviation $\sigma_\lambda$. The tilt angle
is chosen according to Joy's law plus a random component drawn from a
distribution with standard deviation $\sigma_\alpha$. The resulting
tilt angle is multiplied by a factor 0.7 to account for the effect of
latitudinal inflows towards active regions \citep{Cameron:etal:2010}.
Finally, the BMRs are distributed randomly over longitude and the two
hemispheres.
The introduction of the factor 0.7 is neccessary because the tilt
angle distributions empirically determined by \citet{Jiang:etal:2014b}
refer to the instant of maximum sunspot area during the evolution of
an active region, which typically is a few days after the beginning
of flux emergence. \citet[][see Fig.~8 therein]{Martin:Cameron:2016}
have shown that the lateral inflows towards active regions
significantly reduce the tilt angle after this moment.
Quantitatively, the reduction is consistent with the factor 0.7
assumed here, a value that also provided good reproduction of the
observationally inferred polar fields (open heliospheric flux) during
the activity minima between 1913 and 1986 with our flux transport
model \citep{Cameron:etal:2010}.
Fig.~\ref{fig:emergence} shows an example for one realization of random
sources for cycle 24. The upper panel provides a comparison of the
actual monthly group sunspot number (black), one realization of the
random sources (red) and the average of 50 random realizations (blue).
The lower panel shows a butterfly diagram with the actual sunspot groups
(black crosses) and one realization of the random sources (red crosses).
We started the SFT simulations from observed synoptic magnetograms using
data from NSO/KPVT (available from 8/1976 until 8/2003), NSO/SOLIS
(8/2003 until present), SOHO/MDI (6/1996 until 11/2010), and SDO/HMI
(5/2010 until present). All magnetogram data used were reduced to a
resolution of $1\deg$ in latitude and longitude, respectively.
\subsection{Error estimate}
\label{subsec:SFT_error}
For a sensible prediction, we need to quantify the uncertainty involved
in our extrapolation of the evolution of the axial dipole moment based
on SFT simulations with random sources. Two causes of error are
considered here: 1) uncertainty due to scatter in the properties of the
source BMRs (number, emergence location, size, and tilt angle) and 2)
uncertainty due to measurement error in the magnetogram used as initial
state. While (1) represents the intrinsic uncertainty resulting from the
random component of the solar dynamo process, the contribution of (2) is
due to imperfect measurement and depends on the instrument that provided
the data, the data analysis procedure, etc. Since both contributions to
the total error are uncorrelated, they add quadratically.
The effect of the scatter in the source properties (1) was determined by
repeating each SFT simulation 50 times with different random
realizations of the sources. The uncertainty resulting from the error in
the initial magnetograms (2) was estimated by consideing the
averaged measured radial surface field,
\begin{equation}
\overline{B_r}(t)=\frac{1}{4\pi} \int_{\pi/2}^{-\pi/2}\int_{0}^{2\pi}
B_r(\lambda,t) \cos\lambda d\lambda d\phi,
\label{eq:net_flux}
\end{equation}
which should vanish since the magnetic field is divergence-free. Any
deviation from zero must therefore be due to measurement error (e.g.,
instrumental noise, Zeeman saturation, $B$ angle correction). We
estimate the rms error of $\overline{B_r}$ by considering a 19-month
running mean (indicated by angular brackets), $\varepsilon(t) =
\langle\overline{B_r}^2\rangle^{1/2}$. Fig.~\ref{fig:rms} shows
$\varepsilon(t)$ for the various time series of synoptic magnetograms
considered here. As the instruments were improved over time and became
characterized better, $\varepsilon$ decreased since 1976 by about a
factor of 10 from values over $0.8\,$G to about $0.08\,$G. In the
estimate of the rms error we can safely neglect the average
$\langle\overline{B_r}\rangle$, which is small compared to $\varepsilon$.
In order to estimate the contribution of the measurement error in the
initial magnetogram to the uncertainty of the axial dipole moment
resulting from the SFT simulation, we need to consider the two
hemispheres separately. The relevant quantity is the rms of the {\em
error differences} between the hemispheres, $\varepsilon_{\rm
N-S}(t)$. Assuming that the errors on the two hemispheres are uncorrelated,
we have $\varepsilon_{\rm N-S}(t)=\varepsilon(t)$. If we assume further that
the statistical distribution of the error is uniform over the solar
surface, we can estimate the rms error of the axial dipole moment,
$\varepsilon_{\rm DM}(t)$, by running a SFT simulation without sources
and with an initial radial field equal to $\varepsilon(t_0)$ in the
northern and $-\varepsilon(t_0)$ in the southern hemisphere, where
$\varepsilon(t_0)$ is the rms error at the epoch of the initial
magnetogram. This procedure draws upon the linearity of
Eq.~\ref{eqn:SFT}, which results in $\varepsilon_{\rm DM} \propto
\varepsilon_{\rm N-S} = \varepsilon$.
We note that there is possibly an additional contribution to the
intrinsic uncertainty. \citet{Wang:etal:2009} proposed that
cycle-to-cycle variations of the mean meridional flow could be
responsible for the variability of the polar field amplitude. In the
absence of a sufficiently extended data base for the meridional flow, it
is difficult to judge the validity of this suggestion. If these
variations were of a random nature, they would (quadratically) add to
the uncertainty determined here.
\section{Results}
\label{sec:results}
\subsection{Axial dipole moment for cycles 21--23: postdiction}
\label{subsec:postdiction}
In order to test and validate our procedure, we carried out SFT
simulation runs with random sources for the descending phases of cycles
21, 22, and 23, for which we can directly compare the prediction
(actually: postdiction) from the simulation with the observed actual
evolution of the axial dipole moment. The simulations were started from
synoptic magnetograms obtained about 4 years prior to the subsequent
activity minima. The initial magnetograms correspond to Carrington
rotations CR1729 (Nov 25--Dec 22, 1982; NSO/KPVT) for cycle 21, CR1864
(Dec 24,1992--Jan 22, 1993; NSO/KPVT) for cycle 22, and CR2024 (Dec 5,
2004--Jan 3, 2005; SoHO/MDI) for cycle 23. For each of the three
cycles, we carried out 50 SFT simulations with different realizations of
random sources. The results are shown in Fig.~\ref{fig:cyc21-23}. The
predictions for the axial dipole moment up to the corresponding solar
minimum are given by the averaged evolution for the 50 realizations each
(solid blue curves) in comparison with the actual data obtained from the
observed synoptic magnetograms. The uncertainty of the predictions was
determined according to the procedure outlined in
Sec.~\ref{subsec:SFT_error}. The total $\pm 2\sigma$-error is indicated by
blue shading. The dashed blue lines denote the $\pm 2\sigma$-error due to
source scatter alone, i.e., without the error in the initial
magnetograms.
The results show that the predictions agree with the actual observation
within the $\pm2\sigma$ uncertainty range. For cycle 22 though, this is
only marginally true owing to the early decay of the dipole moment
shortly after activity maximum.Such a deviation is expected to occur occasionally owing to the
significant scatter in the tilt angle and other properties of
BMRs. In particular, the axial dipole moment can be significantly
affected by the emergence of single strongly tilted BMRs near the
equator \citep{Cameron:etal:2013}. In addition, fluctuations in the
background meridional flow (not considered in our model) could also
have contributed.
If a stochastic process indeed underlies the variability of the axial
dipole moment around solar minima, the prediction is expected to be
within the $\pm 2\sigma$ range for about 95\% of the cycles.
Considering only three cycles, of course we cannot make a definite
statement beyond noting that the comparison of our postdictions with the
observations is so far statistically consistent with the assumed
stochastic processes (basically source scatter).
\subsection{Dipole moment for cycle 24: predictor for cycle 25}
\label{subsec:prediction}
We now consider the prediction for the axial dipole moment during the
minimum of the current cycle 24, expected to occur around the year 2020
(cf.~Fig.~\ref{fig:emergence}). As initial state for the SFT simulation
we used the synoptic magnetogram for Carrington rotation 2171 (Nov
28--Dec 25, 2015) from SDO/HMI. Using the same procedure as for the
previous cycles, we determined the predicted evolution of the axial
dipole moment, which is displayed in Fig.~\ref{fig:cyc24}. During the
upcoming minimum around 2020, the expected value and $2\sigma$
uncertainty for the axial dipole moment is $2.5\pm1.1\,$G. The expected
value for the dipole moment is not much higher than that observed at the
end of cycle 23 ($\sim2\,$G) and significantly lower than those for
cycle 21 ($\sim3.5\,$G) and 22($\sim4.1\,$G). However, within the rather
high level of the $2\sigma$ uncertainty, the dipole moment could be as
high as that of cycle 21 or significantly lower than that of cycle 23.
Note also that (for reasons unknown to us) the NSO/SOLIS and SDO/HMI
data drifted apart by about $0.5\,$G. As a result, taking the NSO/SOLIS
magnetogram for CR2171 as initial state leads to a higher expectation
value of $3.1\pm1.1$G for the axial dipole moment around 2020.
Owing to the good (although not perfect) correlation between the open
heliospheric magnetic flux (strongly related to the axial dipole moment)
during solar minimum and the strength of the subsequent cycle
\citep{Wang:Sheeley:2009, Cameron:Schuessler:2012} we therefore expect
that cycle 25 will be of moderate strength, but possibly somewhat more
active than the current cycle. However, the uncertainty of this
prediction is considerable: within the $2\sigma$ range, cycle 25 could
also be as strong as cycle 22 or even considerable weaker than cycle 24.
This reflects the intrinsic limitation of such predictions resulting
from the random nature of flux emergence.
\begin{acknowledgements}
All authors contributed equally to the work presented in this paper.
SOHO is a project of international cooperation between ESA and NASA.
The SDO/HMI data are courtesy of NASA and the SDO/HMI team. The sunspot
records are courtesy of WDC-SILSO, Royal Observatory of Belgium,
Brussels. The National Solar Observatory (NSO)/Kitt Peak data used here
are obtained cooperatively by NSF-NOAO, NASA/Goddard Space Flight
Center, and the NOAA Space Environment Laboratory. NSO/SOLIS data were
courtesy of NISP/NSO/AURA/NSF. J. Jiang acknowledges the support by the
National Science Foundation of China (grants 11522325, 11173033) and by the
Youth Innovation Promotion Association CAS.
\end{acknowledgements}
|
\section{Introduction}
M-theory, that first conjectured
by Edward Witten in the spring of 1995, unifies
all versions of string theory: type I, type IIA,
type IIB, and two flavors of heterotic string theory
(SO(32) and $E_8\times E_8$). Each of these five string theories
is limiting case of M-theory, and should be approximated
by eleven-dimensional supergravity at low energies.
M-theory is connected with the AdS/CFT correspondence
and should describe branes \cite{1.}.
For connection of M-theory with experimental data
is used the mechanism of compactification of its
extra dimensions to four-dimensional world,
that can be verified at the LHC.
One of the fundamental questions of theoretical
high energy physics is the question of phase transitions
of solitonic states, like D-branes.
Due to the higher dimensional models, ADD or RS \cite{2.}
in some scenarios involving extra dimensions of space,
the Planck mass can be as low as the TeV range.
As the Large Hadron Collider (LHC) has energy 14 TeV
for proton-proton collisions it was argued \cite{3.}
that black hole production could be an important
and observable effect at the LHC or future higher-energy colliders.
Such quantum black holes or other high energetic
silitonic objects should decay emitting sprays
of particles that could be seen by detectors.
The purpose of the article is to study
such solitonic objects with the help of effective model - D-branes.
Our attempts for connection string theory with experiment
are focused on the compactification of extra dimensions to $K3\times T^2$
for construction models of our four-dimensional world.
The studied cases involve higher-dimensional spaces - D6-branes
in for-dimensional space, for example. In section 2 we'll consider
the equivalence of vector bundles on $K3\times T^2$ and Calabi-Yau manifolds.
We are dealing with a special type of Calabi-Yau manifolds,
when it is both an elliptic bundle over Hirzebruch surface $F_{2k}$
and K3-fibration over the one dimensional projective space.
In accordance with \cite{4.} there is the correspondence
between the heterotic string compactifications on $K3\times T^2$
and type II string compactifications on Calabi-Yau
threefolds of such type. This is connected with fact
that for compactification of the heterotic string on $K3\times T^2$
is used the embedding, equating the spin
connection of the manifold with the gauge connection.
Through examples, it is possible that the moduli spaces
of many different N=2 heterotic vacua are connected in
a similar way to type II string compactifications
on Calabi-Yau. To each type of Calabi-Yau corresponds
its enhanced symmetry according to Batyrev's toric approximation.
In section 3 will be presented K-theory description
of vector bundles over $K3\times T^2$.
In section 4 we consider D-brane RR charge calculation
corresponding to orbifold. The conclusion is connected with
received fact that breaking of enhanced symmetries of
Calabi-Yau to Standard model is connected with K-theory
description of vector bundles over $K3\times T^2$, which differ
from each other by the rank of the bundle.
\section{Calabi-Yau transitions }
From \cite{5.} we know
that there exist duality between (0, 4)
compactifications of the $E_8\times E_8$ heterotic
string on the manifold $K3\times T^2$ and the type IIA
string compactified on a Calabi-Yau manifold \cite{6.}.
This duality corresponds to the equivalence
of vector bundles on $K3\times T^2$ and Calabi-Yau manifolds.
This Calabi-Yau are of special type - K3-fibration over P$^1$
projective space.
In the context of M- and F-theory \cite{7.}
the dynamics on the moduli space
of string theory would allow to determine the
physical ground state of the string. The criterion
which distinguishes between different vacua of
string theory is the compactifications of these
theories to three and four dimensions, in particular
to CalabiYau fourfolds. Toric description of elliptic
Calabi-Yau manifold according to Batyrev's approximation
\cite{8.} can be realized by dual
polyhedron which is devided by triangle on the top and bottom ,
For fourfolds of type
$X_{18k+18}(1, 1, 1, 3k, 6k + 6, 9k + 9) $
the gauge groups are written in the following way \cite{9.}:
\[H \times SU(1)\ \ \ \mbox{for} k = 1 ,\]
\[H \times SO(8)\ \ \ \mbox{for} k = 2 ,\]
\[H \times E_6 \ \ \ \mbox{for} k = 3 ,\]
\[H \times E_7 \ \ \ \mbox{for} k = 4 ,\]
\[H \times E_8 \ \ \ \mbox{for} k = 5 ,\]
\[H \times E_8 \ \ \ \mbox{for} k = 6 .\]
Thus, to each type of Calabi-Yau corresponds its enhanced symmetry.
The moduli space of string theory vacuum
can be deformed by vevs through breaking
the gauge group. For example, $E_8$ can be completely broken through the chain
\[E_8 \rightarrow E_7 \rightarrow E_6 \rightarrow
SO(10) \rightarrow SU (5) \rightarrow
SU (4) \rightarrow SU (3) \rightarrow SU (2) \rightarrow SU (1)\ .\]
So, the breaking of the gauge group is connected with
phase transitions between different Calabi-Yau manifolds.
\section{K-theory description of vector bundles over $K3\times T2$}
As transitions between Calabi-Yau,
described in such technique, are known,
it would be interesting to understand an
adequate mathematical method for description of transitions between
different vector bundles
over compact base space $K3\times T^2$ according to duality \cite{5.}.
The studying of such vector bundles was performed
by \cite{9.}, where the bundle V on $K3\times T^2$ is fixed as follows
\[V=\oplus_{i}p_{1}^{*}L_{x_{i}}\otimes p_{2}^{*}M_{i} \ ,\]
where $L_{x_{i}}$ are the line bundles on $T^2$
corresponding to $x_{i}$ - a point in the dual
torus $\check{T}^2$ , $p_1$ and $p_2$ are projections from
the product $K3\times T^2$ to the factors $T^2$ and $K3$, respectively.
As is known from \cite{11.},
vector bundles over compact base spaces form the set of
isomorphism classes of vector bundles over $X$ (in our case $X$ is $K3\times T^2$).
This is a semi-ring under the operations of
Whitney sum and tensor product. It contains the disjoint union
\[\mbox{Vect} X=\bigcup\limits_{d=0}^{\infty} \mbox{Vect}_d X\]
where Vect$_d X$ comprises the classes
of vector bundles of rank d. Such construction
of semi-ring Vect $X$ of isomorphism classes of
complex vector bundles over $X$ leads to the ring
$KX:=K(\mbox{Vect} X)$, which is called
a Grothendieck group a contravariant functor
from compact topological spaces to rings.
From \cite{12.}
is known that, there exists the isomorphism of
free sheaves of rank $n$ and classes of
vector bundles of rank $n$ .
The category Vec$_r(X)$ of vector bundles
of rank r on $X$ and the category Loc$_r(X)$ of
locally free sheaves of rank r on $X$ are
equivalent, as defined by Hartshorne, \cite{11.}.
In \cite{13.} is written about D-brane as
locally free sheaf. From \cite{14.} it has
been observed (for example, as in \cite{15.})
that branes supported on complex submanifolds
of complex varieties are naturally described
in terms of coherent sheaves. Therefore, Grothendieck
groups of coherent sheaves, the holomorphic
version of K-theory, can be used to describe
D-branes in the case that all D-branes are
wrapped on complex submanifolds.
\section{String theory and RR charge}
Since we are dealing with M-theory theory
of five string theories, it is necessary to
say that string looks like an ordinary particle,
with its mass and charge. A physical object
that generalizes the notion of a point particle
to higher dimensions is a brane. A point particle
can be viewed as a brane of dimension zero, while
a string can be viewed as a brane of dimension one.
It is also possible to consider higher-dimensional
branes. In dimension p, they are called p-branes.
Branes are dynamical objects which can propagate
through spacetime according to the rules of quantum
mechanics. They can have mass and other attributes
such as RR-charge. Much of the current research in
M-theory attempts to better understand the properties
of these branes. So, the studying of D-brane
classification with the help of K-theory description
of RR-charge is of great importance.
Generally believed of K-theory as a ``poor man's derived category''
that knows only about D-brane charge \cite{13.}.
A D-brane charge corresponding to a vector bundle E is given by formula
\[Q(E)=ch(E)\sqrt{td(X)}\ ,\]
where $ch(E)$ is the Chern character of $E$
and $td(X)$ is the Todd class of the tangent
bundle of $X$. The charge of a D-brane given
by a coherent sheaf can be computed using the
Grothendieck-Riemann-Roch theorem \cite{13.}.
As is stressed in \cite{16.}
blowing-up of $T^4/Z_2$ is K3 space. Therefore, instead
of $K3\times T^2$ space we can consider $T^6/Z_2$ space - orbifold.
D-brane that passes through an orbifold fixed
point carries RR charge. In a supergravity
approximation we can take the large volume limit to
describe the backgrounds and these are RR backgrounds
\cite{17.}.
Large volume charges are connected with RR charges by formula:
\[Q_4=n_1-2n_2+n_3,\ Q_2=-n_1+n_2,\ Q_0=\frac{n_1+n_2}{2}\]
which define the Chern character
\[ch(n_1n_2n_3)=Q_4+Q_2\omega+Q_0\omega^2\ ,\]
$\omega$ is Wu parameter.
The rank of the vector bundle $E$ is $Q_4$, $Q_2=c_1,\ Q_0=\frac{c_1^2}{2}$, where
$c_1$ is the first Chern class.
From \cite{10.} the first Chern class of $SU(N)$ bundle
$E$ over $K3\times T^2$, for example, is zero, therefore $Q_2=0$ and $Q_0 =0$
and vector bundles over $K3\times T^2$ are differ through the rank of the bundle.
\section{Conclusion}
The conclusion is connected with the fact
that morphisms of distinct Calabi-Yau permit
an interpretation in terms of topological K theory
or Grothendieck groups. In spite of the fact
that $K3~T^4/Z_2$ and the full space $K3\times T^2~T^6/Z_2$,
we can't use the notion of Aspinwall \cite{13.}
that D-branes on the orbifold $C_d/G$ and open strings
between them are described
by the derived category of McKay quiver
representations (with relations) because
we cannot have the derived category of a
``compact'' CY manifold represented by a quiver.
Anyway, in spite of the fact that K-theory contains
much less information than the derived category this
is the beginning for understanding of brane
classification when the symmetry is broken from E$_8$
to the Standard molel. According to \cite{18.}
when we focus our attention on supergravity in d=11,
solutions also exist for N=1 in d=10, 9 and 8 dimensions
for which spacetime is Minkowski space and for which
the extra dimensions are $K3\times T2$,
$K3\times S^1$ and $K3$, respectively.
Starting from 64+64 components of N=1 in d=10,
we obtain 96+96 components in d=4 of N=2
supergravity and so on to the Standard model.
|
\section{Introduction}
The OH molecule was discovered using the $\Lambda$ doublet transition
observed between the levels of the ground rotational state
$^{2}\Pi_{3/2}$ J = 3/2 at 18 cm \citep{weinreb1963}; later its
electronic transitions were identified in ultraviolet spectra of
bright OB-stars \citep{crutcher1976,chaffee1977, felenbok1996}. Two
lines of the A$^2\Sigma^+$--X$^2\Pi_i$ band near 3078 and 3082 $\AA$
are available to ground-based observatories and in a series of papers
\citep{weselak2009,weselak2010} estimates of OH column densities along
17 translucent lines of sight from these transitions have been
derived. These observations performed using the high-resolution
VLT/UVES-spectrograph verified the oscillator strengths for the 3078
and 3082\,$\AA$ and also suggested a close correlation between the
column densities of OH and CH molecules. Early chemical models have
suggested production of hydrides like OH and NH due to grain catalysis
reactions \citep{wagenblast1993}. Hence observation of OH to derive
abundances towards translucent sightlines with larger than normal
far-UV extinction is interesting. Interstellar CH is often observed
using its strongest A--X transition at 4300$\AA$. However this line
being saturated, reliable estimates of CH column densities are derived
from the spectral lines due to the weaker B--X system at 3886$\AA$
\citep{krelowski1999,weselak2004}. There exist many studies of CH to
explore its correlation to the diffuse interstellar bands (DIBs) at
5780 and 5797$\AA$ \citep{krelowski1999}. However availability of
data for OH is significantly worse. Observations have shown the
abundances of CH molecule to be tightly correlated with those of the
H$_2$ molecule \citep{mattila1986,weselak2004,sheffer2008}. The
correlation is sufficiently strong so that in many recent studies
column densities of CH have been used to derive estimates for H$_2$
column density \citep{sheffer2008,wiesemeyer2016}. The correlation is
understood in terms of higher formation rates of CH in reactions
involving molecular hydrogen.
The aim of this work is to incorporate the newly available OH and CH
data for 13 lines of sight (5 of which are new) in order to
investigate the correlation between OH and CH in diffuse/translucent
molecular clouds. This correlation is particularly interesting
considering the tight correlation between CH and H$_2$. We shall show
that the oscillator strengths for the C--X system of lines of CH at
3137\,$\AA$ are correct so that these can be equivalently used to
determine the column densities of CH. We shall use simple chemical
models to check whether the observed correlation between OH and CH or
the absence of it is predicted by the chemical models.
\section{UV Spectroscopic datasets}
We extend the analysed sample of \citet{weselak2010} by combining it
with sample of the highly reddened early-type stars which were
analyzed by \citet{bhatt2015}. In order to complete the information on
all the lines of sight as much as possible, we have also analyzed fully
processed UVES/VLT data of both pairs of CH lines, obtained from the
ESO Science Archive
Facility\footnote{\url{http://www.eso.org/sci/observing/phase3.html}}.
\citet{bhatt2015} had identified 30 known features (11 atomic and 19
molecular) and tentatively detected up to 7 new interstellar absorption
lines of unknown origin towards 346 targets. Out of this entire sample
only for 13 sources observations of both the 3078 and 3082\,$\AA$
transitions of OH were available. Out of these 13 sources 8 were
already present in the sample used by \citet{weselak2010}. The
observed data for the two CH transitions B$^2\Sigma^-$--X$^2\Pi_i$ at
3886 and 3890\,$\AA$ for 19 of these sightlines were also presented by
\citet{weselak2010}. \citet{bhatt2015} presented the
C$^2\Sigma^+$--X$^2\Pi_i$ transitions of CH at 3137 and 3143\,$\AA$ for
the 12 out of the 13 sources in which OH was observed. Thus we finally
have 24 sources, 19 from \citet{weselak2010} and 13 from
\citet{bhatt2015}, of which 5 sightlines are new.
\section{Results and discussion}
Table\,\ref{tab_spectro} presents the spectroscopic details,
frequencies, and $f$-values of the transitions of OH and CH considered
here.
\begin{table*}
\caption{Spectroscopic details and $f$-values of OH and CH transitions studied.
CH transitions
\label{tab_spectro}}
\begin{tabular}{crclccc}
\hline
\hline
\multicolumn{1}{c}{Species} &
\multicolumn{1}{c}{Vibrionic Band} &
\multicolumn{1}{c}{Rotational lines} &
\multicolumn{1}{c}{Position} &
\multicolumn{1}{c}{Ref. }&
\multicolumn{1}{c}{$f$-value} &
\multicolumn{1}{c}{ Ref.}\\
\hline
\hline
OH & A$^2\Sigma^+$--X$^2\Pi_i$ (0,0) & Q$_1$(3/2)+$^QP_{21}$(3/2) & 3078.443 & 1 & 0.00105 & 2\\
& (0,0) & P$_1$(3/2) & 3081.6643& 1 & 0.000648 & 2\\
\hline
CH & C$^2\Sigma^+$--X$^2\Pi_i$ (0,0) & R$_2$ & 3137.576 & 3 & 0.00210 & 3 \\
& (0,0) & Q$_2$(1)+$^QR_{12}$(1) & 3143.150 & 3 & 0.00640 & 3 \\
CH & B$^2\Sigma^-$--X$^2\Pi_i$ (0,0) & Q$_2$(1)+$^QR_{12}$(1) & 3886.409 & 3 & 0.00320 & 3 \\
& (0,0) & $^PQ_{12}$(1) & 3890.217 & 3 & 0.00210 & 3 \\
\hline
\end{tabular}
1. Weselak et al. (2009), 2. Felenbok \& Roeuff (1996) and 3. Lien (1984)
\end{table*}
Table\,\ref{tab_meas} presents the observed equivalent widths (EW) of
the OH lines at 3078 and 3082\,$\AA$ and Fig.\,1 shows a plot of the
same. We find that excluding HD\,114213 (with the largest
EW(3082$\AA$)), the equivalent widths of the two transitions show a
correlation of 0.92, which is somewhat worse than the estimates of
\citet{weselak2009}. From the slope of the line of regression we
estimate the ratio of the equivalent widths of the two transitions to
be 1.62$\pm$0.08. The wavelengths of the two transitions being
similar, the equivalent widths are expected to have a ratio similar to
the ratio of the $f$-values of the two transitions. The derived ratio
matches well with the estimate (1.62) based on the $f$-values of the
two transitions.
\begin{figure}
\centering
\includegraphics[width=\columnwidth,angle=0]{fig1.eps}
\caption{Comparison of the equivalent widths of the two transitions of OH at 3078 and
3082\,$\AA$. The straight line corresponds to a linear regression with
a correlation coefficient of 0.92 and a slope of 1.62$\pm$0.08.}
\label{fig_ohcomp}
\end{figure}
For all the unsaturated lines of OH and CH we estimate the column
densities using the relationship proposed by \citet{herbig1968}:
\begin{equation}
N = 1.13\times 10^{20} W_\lambda/(\lambda^2 f)
\end{equation}
where $W_\lambda$ and $\lambda$ are in $\AA$, column density is in
\mbox{cm$^{-2}$}\ and the $f$-values used are taken from
Table\,\ref{tab_spectro}. The total column density for OH is estimated
by adding the column densities of the 3078 and 3082\,$\AA$ transitions
(Table\,\ref{tab_meas}).
Out of the 24 sources, for 19 sources we have measurements of the
3137 and 3143\,$\AA$ transitions and for 22 sources we have
measurements of the 3886 and 3890\,$\AA$ transitions of CH
(Table\,\ref{tab_meas}). We have re-measured all the CH equivalent
widths from the UVES/VLT archival data and find that in most cases our
measurements are consistent with the published work of
\citet{weselak2010,bhatt2015}. We have revised the values of the
equivalent width of the 3886\,$\AA$ CH line towards HD\,27778 and both
the 3886 and the 3890\,$\AA$ CH lines towards HD\,154445.
Figure\,\ref{fig_chcomp} shows a comparison of CH column densities
derived from a combination of the transitions at 3137 \& 3143\,$\AA$
with $N$(CH) derived from a combination of transitions at 3886 \&
3890\,$\AA$ for the 17 sources in which both sets of CH lines have been
observed. The values of $N$(CH) show a correlation of 0.99 with
$N$(CH(3137+3143)/$N$(3886+3890) = 1.03$\pm$0.02. For the rest of the
analysis, we thus use the column densities for the pair of CH
transitions which shows a lower relative uncertainty.
\begin{figure}
\centering
\includegraphics[width=\columnwidth,angle=0]{fig2.eps}
\caption{Comparison of $N$(CH) derived for the 17 sources in which both
pairs of transitions {\em viz.,} (3137 \& 3143\,$\AA$) and (3886 \&
3890\,$\AA$) are available. The straight line shown is fitted to
obtain the ratio between the two column densities.
\label{fig_chcomp}}
\end{figure}
Figure\,\ref{fig_ohchcolden} shows a plot of estimated $N$(OH) as a
function of $N$(CH). For clarity, we have used different symbols
to indicate which of the two pairs of transitions was used to derive
the $N$(CH). We find that for most of the sightlines the column
densities of OH and CH occupy a band in the plot, except towards
BD-14\,5037, HD\,114213, HD\,161056 and HD34078. The overall
correlation coefficient is 0.61 between $N$(OH) and $N$(CH), which is
significantly worse than the value presented by \citet{weselak2009}.
However, the correlation coefficient improves to 0.94 (black dashed
line) when the four outliers are removed, although it is still
less than the 0.99 obtained by \citet{weselak2010}. For the 4 out of 24
sightlines which do not show the correlation between $N$(OH) and
$N$(CH) as seen in the other source, the deviation can not be explained
in terms of the observational uncertainties.
\begin{figure}
\centering
\includegraphics[width=\columnwidth,angle=0]{fig3.eps}
\caption{Variation of $N$(OH) as a function of $N$(CH). Filled circles
and triangles denote $N$(CH) derived from the (3137+3143)$\AA$ and
(3886+3890)$\AA$ transitions, respectively. Black dashed line shows the
linear fit to all datapoints excluding the outliers HD\,34078,
HD\,114213, HD\,161056 and BD-14\,5037, which corresponds to a Pearson
correlation coefficient of 0.94. }
\label{fig_ohchcolden}
\end{figure}
Combined analysis of all available datasets on OH and CH observations
of diffuse clouds enable us to further revise the relation between the
column densities of the two molecules for 20 sightlines to [N(OH)
= 2.41$\pm$0.21 N(CH) +(8.55$\pm$7.17) (in 10$^{12}$\,\mbox{cm$^{-2}$})].
The relationship derived by \citet{weselak2010} had made use of 16
sightlines. The ratio of column densities of OH and CH is found to vary
between 1.5 to 4.1 in diffuse clouds.
We look into the 3 sources HD\,114213, HD\,161056 and BD-14\,5037 for
possible observational reasons behind the very different $N$(CH)-$N$(OH)
relationship, based on the central heliocentric velocities of the
absorption features. The source BD-14\,5037 shows absorptions at two
velocities around -6\,\kms\ and 6\,\kms. Since only -6\,\kms\ component
is seen in the OH(3081\,$\AA$) spectral line, the numbers we have used
correspond to this component. Similarly, the line of sight toward
HD\,114213 shows absorptions at two velocities: CH lines show up at -16
and 6\,\kms, while the OH 3078\,$\AA$ line is at -17\,\kms\ and 3\,\kms\
while the OH(3082\,$\AA$) line is at -20\,\kms\ and 0.3\,\kms. We have
used here the -20\,\kms, which is identified in all the spectral lines
considered here, since the relative shift in the velocity is smaller
than the observed FWHM of the lines \citep{bhatt2015}. The sightline
towards HD\,161056 however shows only one velocity component, although
the heliocentric velocities at 3078, 3082, 3137 and 3146\,$\AA$ are
-9.4$\pm$0.1, -11.0$\pm$0.1, -12.7$\pm$0.3 and -14.0$\pm$0.3\,\kms\
respectively \citep{bhatt2015}. As with the other two sources, the
shifts in central velocities are smaller than the FWHM of the lines.
Thus, for the three sources HD\,114213, HD\,161056 and BD-14\,5037 as
with HD\,34078 the lack of correlation between $N$(CH) and $N$(OH) can
not be explained in terms of any observational uncertainties. It is
clearly obvious that along the sightlines towards HD\,114213, HD\,161056
and BD-14\,5037, the OH/CH ratio are 2--3 times the value seen in the
other diffuse clouds.
\section{Molecular hydrogen fraction in diffuse clouds}
Based on theoretical predictions and the observed correlations between
CH and H$_2$, of late $N$(H$_2$) is often derived from the observed
$N$(CH) assuming CH/H$_2$ = 3.5\,10$^{-8}$ \citep[][and references
therein] {sheffer2008,wiesemeyer2016}. It is also possible to estimate
the column density of hydrogen nuclei using the measured $E(B-V)$ by
using the relationship $N$(H) = 5.8\,10$^{21}$\,$E(B-V)$
\citep{bohlin1978}. For all the lines of sight considered here, since
both $E(B-V)$ and $N$(CH) are available, we have estimated $f_{\rm
H_2}=\frac{N{\rm (H_2)}}{N{\rm (H)}}$ and do not find any
correlation between $N$(CH)/$N$(OH)) and $f_{\rm H_2}$. The sources,
HD\,161056 and HD\,114213, both of which show high $N$(OH)/$N$(OH)
ratios have $f_{\rm H_2}$ of 51\% and 19\% respectively. For 17 lines
of sight $f_{\rm H_2}$ is $\leq 0.4$, with BD-14\,5037 and HD\,34078
showing 11\% and 80\% molecular hydrogen fraction, respectively. We
discuss the anomalously high $N$(CH) shown by HD\,34078 later in the
paper.
\section{Chemical modeling of the diffuse clouds}
We have constructed chemical models for the diffuse/translucent clouds
using the Astrochem code \citep{maret2015} together with the reaction
coefficients from the OSU 2009 and examined the predicted OH and CH
abundances. The models consider a variety of gas phase processes, as
well as simple gas-grain interactions, such as the freeze-out and
desorption via several mechanisms (thermal desorption, cosmic-ray
desorption and photo-desorption). We have run a grid of models
covering a range of values for the relevant physical parameters in
order to ascertain the parameters to which the CH and OH abundances
are the most sensitive. We used the abundances of both CH and OH at
10$^6$\,yr since both attain a constant value beyond 10$^5$\,yr.
Typically in these models, OH abundances are quite sensitive to the
cosmic ray ionization rate ($\zeta_{\rm CR}$), with the abundance
decreasing with increase in $\zeta_{\rm CR}$.
Figure\,\ref{fig_ohchrat} shows a plot of $N$(OH) as a function of
$N$(CH) as predicted by the models which are detailed in the caption.
The purpose of constructing these models is to understand whether
these also predict the observed correlation between $N$(CH) and
$N$(OH) and not to model the diffuse/translucent clouds each line of
sight accurately. The modeling is also aimed to understand whether the
four lines of sight which lie outside of the correlation band
(Fig.\,\ref{fig_ohchcolden}) are generally consistent with these
simple chemical models that consider gas-grain interactions and
gas-phase reactions but do not include the effects of shock or
turbulence.
We have considered multiple values of $N$(CH) and used the CH
abundances predicted by a model corresponding to a set of input
parameters to estimate $N$(H$_2$) and then used this $N$(H$_2$) to
estimate $N$(OH) from the OH abundances predicted by the model. In
Fig.\,\ref{fig_ohchrat} we plot the observed values of $N$(OH) and
$N$(CH) for the 24 sightlines presented here. We find that broadly
the model corresponding to $n_{\rm H}$=500\,\cmcub, $\zeta_{\rm CR}$ =
10$^{-14}$\,s$^{-1}$, \av = 2.0, $f_{\rm H_2}$=0.3, UV radiation
($\chi) = 5$ and $T_{\rm gas}$=30\,K (solid black line in
Fig.\,\ref{fig_ohchrat}) is consistent with many of the observed CH
and OH column densities and we designate this as Model $A$ in the
remaining discussion. For all models discussed here the following
initial abundances relative to H were assumed [He]=0.14, [N]=
2.14\,10$^{-5}$ [O]= 1.76\,10$^{-4}$, [C$^+$]= 7.3\,10$^{-5}$ and
[e$^+$] = 7.3\,10$^{-5}$.
In order to estimate the effect of variation in the input parameters
we also show predictions from models in which the parameters like
$n_{\rm H}$, $\zeta_{\rm CR}$ and \av\ are varied relative to the
parameters of Model $A$, one parameter at a time
(Fig.\,\ref{fig_ohchrat}). Relative to Model $A$, the presence of
correlation between OH and CH does not change drastically when a)
$f_{\rm H_2}$ is varied up to 0.8, b) $T_{\rm gas}$ is assumed to be
10 and 20\,K and c) $\chi$ is reduced to 1. In order to show that it
is indeed possible to find a chemical model which is reasonably
consistent with observed parameters and still reproduce the observed
OH/CH ratio for one of the outliers we have constructed a model for
HD\,114213, one of the outliers. Shown in left bottom panel of
Fig.\,\ref{fig_ohchrat} in green continuous line is the prediction of
a model with \av = 3.0 (consistent with observed $E(B-V)$), $n_{\rm
H}$=500\,\cmcub, $\zeta_{\rm CR}$ = 2.5\,10$^{-16}$\,s$^{-1}$, $T_{\rm
gas}$=30\,K and $f_{\rm H_2}$=0.3, which reproduces the observed
$N$(OH) and $N$(CH) quite well. This is by no means a unique model,
since there are several free parameters and fewer observational
constraints, however this shows that the observed column density
ratios are not inconsistent with the chemical models. Thus overall, we
find that with the exception of HD\,34078, the observed ratio of
column densities of OH and CH can be reproduced in all clouds by the
chemical models which do not involve any shock chemistry. The
extremely low $N$(OH) relative to $N$(CH) as seen in HD\,34078 can not
be explained even by varying the input parameters over a much larger
range than what has been shown in Fig.\,\ref{fig_ohchrat}.
The line of sight towards HD\,34078 has been studied extensively using
FUSE spectra and later millimeter CO observations by
\citet{boisse2005,boisse2009}. This line of sight shows anomalously
high CH/H$_2$ as well as CH$^+$/H$_2$ ratios which have been explained
by the presence of a nascent bow shock around the star, at the
interface between the stellar wind and the ambient interstellar
medium, where the material is strongly compressed \citep{boisse2005}.
There is no corresponding enhancement of OH abundance, which indicates
that the OH absorption is due to the more quiescent H$_2$ gas located
beyond the photodissociation front and shocked region
\citep{boisse2009}.
The correlation of CH with H$_2$ is explained in terms of the higher
formation rates of CH due to reactions involving molecular hydrogen
\citep{federman1982}. The formation of OH on the other hand is
strongly influenced by grain-surface catalysis, which implies that
presence of dust grains (and hence possibly H$_2$) would enhance OH
formation rates \citep{wagenblast1993}. This can be a possible reason
behind the significant correlation between OH and CH in translucent
clouds. However as seen from the chemical models several combinations
of the input parameters can lead to similar abundance ratios for the
two species. Thus in order to obtain a more quantitative view of the
correlation, accurate modeling of individual sources with further
constraining observations is necessary.
\setlength{\tabcolsep}{0.08cm}
\begin{table*}
\caption{Equivalent widths (from literature and archival data) and
calculated column densities of OH and CH transitions
\label{tab_meas}}
\small
\begin{tabular}{llrrrrrrrrrr}
\hline
\hline
\multicolumn{1}{c}{Source} &
\multicolumn{1}{c}{E$_{B-V}$} &
\multicolumn{1}{c}{EW(3078)} &
\multicolumn{1}{c}{EW(3082)} &
\multicolumn{1}{c}{EW(3137)} &
\multicolumn{1}{c}{EW(3143)} &
\multicolumn{1}{c}{EW(3886) }&
\multicolumn{1}{c}{EW(3890)} &
\multicolumn{1}{c}{$N$(OH)} &
\multicolumn{1}{c}{$N$(CH)} &
\multicolumn{1}{c}{$N$(CH)} \\
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{(3137+3143)} &
\multicolumn{1}{c}{(3886+3890)}\\
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{}&
\multicolumn{1}{c}{m$\AA$}&
\multicolumn{1}{c}{m$\AA$}&
\multicolumn{1}{c}{m$\AA$}&
\multicolumn{1}{c}{m$\AA$}&
\multicolumn{1}{c}{m$\AA$}&
\multicolumn{1}{c}{m$\AA$}&
\multicolumn{1}{c}{\mbox{cm$^{-2}$}} &
\multicolumn{1}{c}{\mbox{cm$^{-2}$}} &
\multicolumn{1}{c}{\mbox{cm$^{-2}$}}\\
\hline
HD24398$^1$ & 0.29 & 1.67$\pm$0.08 & 1.11$\pm$0.05 & 3.25$\pm$0.41 & 5.83$\pm$0.36 & 4.96$\pm$0.50 & 3.01$\pm$0.38 & 3.93$\pm$0.18(13) & 2.84$\pm$0.29(13) & 2.23$\pm$0.25(13)\\
HD27778$^{1,3}$ & 0.37 & 5.30$\pm$0.15 & 2.20$\pm$0.10 & 3.15$\pm$1.08 & 9.2$\pm$1.0 & 7.46$\pm$0.60 & 4.84$\pm$0.80 & 1.01$\pm$0.04(14) & 3.366$\pm$0.77(13) & 3.99$\pm$0.41(13)\\
HD34078$^1$ & 0.49 & 1.72$\pm$0.27 & 0.86$\pm$0.21 & 6.92$\pm$1.36 & 19.9$\pm$1.4 & 16.75$\pm$1.30 & 11.28$\pm$0.50 & 3.53$\pm$0.69(13) & 7.34$\pm$0.99(13) & 7.93$\pm$0.48(13)\\
BD-14\,5037$^{a,2}$ & 1.59 & 5.74$\pm$0.43 & 5.71$\pm$0.60 & 3.2$\pm$0.21 & 10.64$\pm$0.33 & 9.47$\pm$0.5 & 6.15$\pm$0.5 & 1.70$\pm$0.16(14) &3.65$\pm$0.17(13) & 4.40$\pm$0.22(13)\\
HD110432$^1$ & 0.48 & 1.81$\pm$0.41 & 1.28$\pm$0.34 &1.83$\pm$0.28 &
5.14$\pm$0.38 & 4.13$\pm$0.20 & 2.40$\pm$0.20 & 4.41$\pm$1.09(13) &
1.92$\pm$0.22(13) & 1.82$\pm$0.12(13)\\
HD114213$^b$ & 1.13 & 9.86$\pm$0.96 & 8.30$\pm$0.87 & 4.48$\pm$0.27 & 10.64$\pm$0.42 & \ldots & \ldots & 2.64$\pm$0.27(14) & 4.35$\pm$0.22(13) & \ldots \\
HD147889 & 1.08 & 12.19$\pm$0.45 & 5.38$\pm$0.20 & 10.8$\pm$0.24 & 22.48$\pm$0.20 & 20.07$\pm$1.20 & 15.02$\pm$0.98 & 2.37$\pm$0.09(14) & 9.92$\pm$0.17(13) & 1.00$\pm$0.06(14)\\
HD147933$^2$ & 0.48 & 3.63$\pm$0.27 & 2.02$\pm$0.31 & 3.14$\pm$0.28 & 5.69$\pm$0.12 & 5.68$\pm$0.35 & 3.46$\pm$0.26 & 7.83$\pm$0.88(13) & 2.73$\pm$0.17(13) & 2.56$\pm$0.17(13)\\
HD148688$^1$ & 0.55 & 0.92$\pm$0.45 & 0.81$\pm$0.54 & 1.54$\pm$0.22 & 4.28$\pm$0.28 & 3.83$\pm$0.20 & 2.75$\pm$0.20 & 2.53$\pm$1.50(13) & 1.61$\pm$0.17(13) & 1.87$\pm$0.12(13)\\
HD149757 & 0.28 & 2.01$\pm$0.67 & 1.25$\pm$0.32 & 3.54$\pm$0.30 & 6.52$\pm$0.17 & 5.45$\pm$0.20 & 3.57$\pm$0.10 & 4.58$\pm$1.35(13) & 3.10$\pm$0.19(13) & 2.54$\pm$0.08(13)\\
HD151932$^1$ & 0.50 & 4.46$\pm$0.71 & 2.13$\pm$0.54 & 2.68$\pm$0.12 & 7.04$\pm$0.47 & 5.97$\pm$0.78 & 3.60$\pm$0.10 & 8.97$\pm$1.80(13) & 2.72$\pm$0.15(13) & 2.68$\pm$0.22(13)\\
HD152236 & 0.66 & 3.50$\pm$0.44 & 2.41$\pm$0.34 & \ldots & \ldots & 5.34$\pm$0.45 & 4.15$\pm$0.56 & 8.40$\pm$1.12(13) & \ldots & 2.72$\pm$0.30(13)\\
HD152249 & 0.48 & 2.81$\pm$1.00 & 1.82$\pm$0.80 &\ldots & \ldots & 3.40$\pm$0.50 & 2.61$\pm$0.40 & 6.53$\pm$2.60(13) &\ldots & 1.72$\pm$0.26(13)\\
HD152270 & 0.50 & 1.95$\pm$0.64 & 2.13$\pm$0.70 & \ldots & \ldots & 3.52$\pm$0.43 & 2.69$\pm$0.45 & 6.12$\pm$2.01(13) & \ldots & 1.78$\pm$0.26(13)\\
HD154368 & 0.47 & 9.25$\pm$0.58 & 4.01$\pm$0.31 & 5.73$\pm$0.26 & 14.79$\pm$0.24 & 12.24$\pm$1.10 & 9.32$\pm$0.40 & 1.79$\pm$0.12(14) & 5.78$\pm$0.19(13) & 6.18$\pm$0.40(13)\\
HD154445$^{1,3}$ & 0.35 & 2.01$\pm$0.31 & 1.52$\pm$0.33 & 2.43$\pm$0.51 & 6.50$\pm$0.29 & 4.35$\pm$0.50 & 3.76$\pm$0.65 & 5.07$\pm$0.96(13) & 2.49$\pm$0.33(13) & 1.96$\pm$0.12(13)\\
HD154811 & 0.66 & 2.43$\pm$0.70 & 1.64$\pm$0.50 & \ldots & \ldots & 5.03$\pm$0.34 & 3.12$\pm$0.56 & 5.77$\pm$1.71(13) & \ldots & 2.29$\pm$0.28(13)\\
HD161056$^2$ & 0.59 & 10.71$\pm$0.61 & 6.78$\pm$0.50 & 5.63$\pm$0.30 &
14.23$\pm$0.45 & 10.3$\pm$0.8 & 9.5$\pm$0.55 & 2.46$\pm$0.16(14) & 5.62$\pm$0.24(13) & 5.79$\pm$0.38(13)\\
HD163800 & 0.61 & 2.00$\pm$0.17 & 1.85$\pm$0.12 & 3.42$\pm$0.15 & 7.97$\pm$0.09 & 6.85$\pm$0.30 & 5.16$\pm$0.40 & 5.67$\pm$0.41(13) & 3.29$\pm$0.10(13) & 3.44$\pm$0.21(13)\\
HD164794 & 0.36 & 1.55$\pm$0.40 & 1.00$\pm$0.65 & \ldots & \ldots & 2.29$\pm$0.54 & 1.83$\pm$0.34 & 3.59$\pm$1.65(13) & \ldots & 1.19$\pm$0.25(13)\\
HD169454 & 1.10 & 6.24$\pm$0.25 & 2.90$\pm$0.22 & 4.52$\pm$0.27 & 11.01$\pm$0.30 & 8.51$\pm$0.54 & 6.47$\pm$0.43 & 1.24$\pm$0.07(14) & 4.44$\pm$0.20(13) & 4.29$\pm$0.28(13)\\
HD170740 & 0.45 & 1.97$\pm$0.43 & 1.73$\pm$0.31 & 2.16$\pm$0.16 & 6.56$\pm$0.22 & 5.10$\pm$0.29 & 2.84$\pm$0.20 & 5.41$\pm$1.06(13) & 2.35$\pm$0.13(13) & 2.20$\pm$0.14(13)\\
HD172028 & 0.79 & 5.05$\pm$0.50 & 4.09$\pm$0.35 & 5.81$\pm$0.35 & 15.59$\pm$0.30 & \ldots & \ldots & 1.32$\pm$0.12(14) & 5.96$\pm$0.24(13) & \ldots\\
HD210121$^2$ & 0.40 & 4.45$\pm$0.28 & 3.03$\pm$0.23 & 3.31$\pm$0.24 &
8.31$\pm$0.19 & 6.51$\pm$0.31 & 5.83$\pm$0.50 & 1.06$\pm$0.07(14) &
3.29$\pm$0.17(13) & 3.59$\pm$0.25(13)\\
\hline
\end{tabular}
$^a$ BD-14\,5037 shows absorptions at two velocities around -5\,\kms\
and 6\,\kms. Since only -5\,\kms\ component is seen in the
OH(3081\,$\AA$) spectral line, the numbers correspond to this
component.\\
$^b$ HD\,114213 shows absorptions at two velocities around -20\,\kms\
and 6\,\kms\ for CH and -20\,\kms\ and 3\,\kms\ for OH. We present
here the -20\,\kms, which is uniquely identified in all the spectral
lines considered here.\\
$^1$ Sources for which we measured EW(3137) and EW(3143).\\
$^2$ Sources for which we measured EW(3886) and EW(3890).\\
$^3$ Sources for which we derived EW(3886) and EW(3890) slightly
different from the published values.
\end{table*}
\begin{figure}
\centering
\includegraphics[width=\columnwidth,angle=0]{fig4.eps}
\caption{Comparison of observed column densities of OH and CH with
computed values from the chemical model Astrochem \citep{maret2015}
used with the chemical network OSU2009. The solid black line in all
panels represents Model $A$ with $n_{\rm H}$=500\,\cmcub, $\zeta_{\rm
CR}$ = 10$^{-14}$\,s$^{-1}$, \av = 2.0, $f_{\rm H_2}$=0.3, $\chi = 5$
and $T_{\rm gas}$=30\,K. {\em Top left}: Model $A$ with \av=1 (dashed) and
\av=5 (dotted); {\em Top Right}: Model $A$ with $\zeta_{\rm CR}$ =
10$^{-16}$\,s$^{-1}$ (dashed) and $\zeta_{\rm CR}$ = 10$^{-13}$\,s$^{-1}$
(dotted); {\em Bottom Left}: Model $A$ with $n_{\rm H}$ = 100\,\cmcub\
(dashed)
and $n_{\rm H}$ = 1000\,\cmcub (dotted). The thick grey line shows the trace of the
model ($n_{\rm H}$=500\,\cmcub, $\zeta_{\rm
CR}$ = 2.5\,10$^{-16}$\,s$^{-1}$, \av = 3.0, $f_{\rm H_2}$=0.3, $\chi = 1$
and $T_{\rm gas}$=30\,K), which explains the OH/CH ratio of HD\,114213.
\label{fig_ohchrat}}
\end{figure}
\section{Conclusions}
The equivalence of the CH column densities derived from a combination
of transitions at 3137 and 3143\,$\AA$ and a combination of
transitions at 3886 and 3890\,$\AA$ has enabled us to combine types of
CH datasets. This has resulted in an increase in the number of
sightlines for which both $N$(OH) and $N$(CH) have been observed.
With the exception of four sightlines, the column density of OH
appears to correlate well with the CH column density in all sources.
The derived relationship between $N$(OH) and $N$(CH) is thus based on
a dataset with larger number of sightlines observed with lower
uncertainties than the previous determinations. These results also
verify the oscillator strengths for all the transitions which have
been observed. We find that the $N$(OH)/$N$(CH) ratio is not
correlated with the fraction of molecular hydrogen in these clouds. We
show that except for the line of sight toward HD\,34078 which is
thought to show signatures of shock-induced increase in CH abundance,
the observed OH/CH ratios in all other clouds are well reproduced by
the chemical models considered here.
\section*{Acknowledgements}
BM sincerely thanks the referee Jacek Krelowski for suggestions which
enhanced the scope of the paper by including additional data. BM thanks
J. P. Ninan for his help with many python related issues. This
research has made use of the VizieR catalogue access tool, CDS,
Strasbourg, France. The original description of the VizieR service was
published in A\&AS 143, 23.
|
\section{Introduction}
\label{intro}
\begin{figure*}[t]
\epsscale{.7}
\plotone{hidens_fourpanel.eps}
\caption{Integrated intensity maps from our new survey of HCN$(1-0)$, HCO$^{+}$$(1-0)$ and HNC$(1-0)$ across M51. For reference, the upper left panel shows the PAWS $^{12}$CO$(1-0)$ single dish map, scaled down by a factor of $30$ to match the intensity scale of the other lines. Gray lines indicate contour levels of $I = 0.2, 0.4, 0.8, 1.6, 3.2$ and $6.4$\,K\,km\,s$^{-1}$. A black line shows the same CO contour ($I = 10$\,K\,km\,s$^{-1}$ in the unscaled map) in each panel. The beam size of 30$\arcsec$ is shown in the lower left corner. The dense gas tracers are particularly bright in the center and along the spiral arms. Their emission correlates well with the bulk molecular gas traced by CO emission.}
\label{fig1}
\end{figure*}
\begin{figure*}[t]
\epsscale{1.1}
\plottwo{lum_hcntir_plot.eps}{lum_hcntir_m51only.eps}
\caption{{\em (left)} IR luminosity, tracing recent star formation, as a function of HCN luminosity, tracing dense gas, for structures from Milky Way cores and Local Group clouds to whole starburst galaxies; compilation from \citet{leroy15}. Along with other recent results, our new M51 data extend the observed correlation into the ``luminosity gap'' corresponding to large parts of galaxies. All data combined scatter about the ratio $L_{\rm TIR} / L_{\rm HCN} \approx 900$~L$_\odot$~(K~km~s$^{-1}$~pc$^2$)$^{-1}$ (gray line; dashed lines show a factor of 3 scatter). ({\em right}) Zoom-in on data for the full-galaxy map of M51 and the pointings from \citet{usero15}. The typical uncertainty on the HCN luminosity for each line of sight is $\sim10^5$K~km~s$^{-1}$~pc$^2$. IR correlates well with HCN luminosity for individual regions, but not all regions show the same IR-to-HCN ratio; both data sets show evidence for a mildly sub-linear slope \citep[see also][]{chen15}, indicating an environment dependent efficiency of star formation in dense gas.}
\label{fig2}
\end{figure*}
\begin{figure*}[t]
\epsscale{1.1}
\plottwo{irhcn_sstar.eps}{irhcn_fmol.eps}
\caption{IR-to-HCN ratio, tracing \mbox{$\rm SFE_{dense}$}\ as a function of ({\em left}) local stellar surface density, $\Sigma_\star$, and ({\em right}) molecular-to-atomic gas ratio, H$_2$/{\sc Hi}$\,\propto$ CO/{\sc Hi}. Gray points with black outlines show points in M51 where HCN is detected at S/N$>3$. Light gray points with no outline show points with lower S/N. Black open points show lower limits - that is, the measured IR/HCN value is above this value or negative (due to $I_{\rm HCN}<0$). A small amount of noise has been added to the $y$-value of the upper limits to distinguish points. Red points show the binned trend for the M51 data; including upper limits. Blue points show results for the \citet{usero15} sample. The disk pointings overlap results for our whole-galaxy map, though M51 has a somewhat lower IR-to-HCN ratio than other galaxies in the survey. Both data sets show that \mbox{$\rm SFE_{dense}$}\ appears to be lower in the high surface density, high molecular gas fraction, central, high pressure parts of galaxies.}
\vspace{0.2in}
\label{fig3}
\end{figure*}
\begin{figure*}[t]
\epsscale{1.1}
\plottwo{hcnco_sstar.eps}{hcnco_fmol.eps}
\caption{HCN-to-CO ratio, tracing \mbox{$\rm f_{dense}$}\ as a function of ({\em left}) local stellar surface density, $\Sigma_\star$, and ({\em right}) molecular-to-atomic gas ratio, H$_2$/{\sc Hi}\,$\propto$ CO/{\sc Hi}. For a description of the individual data points see caption of Figure \ref{fig3}. The disk pointings agree with our results for M51. Both data sets show \mbox{$\rm f_{dense}$}\ to depend strongly on local conditions in a galaxy, with the sense of high dense gas fractions in the high surface density, heavily molecular, inner, high pressure regions of M51.}
\vspace{0.2in}
\label{fig4}
\end{figure*}
\begin{figure*}[t]
\epsscale{1.1}
\plotone{stacks_sstar.eps}
\caption{Stacked HCN spectra for our lowest six stellar surface density bins in Figures \ref{fig3} and \ref{fig4}. In each panel, we sum all spectra in an 0.1~dex-wide bin centered at the indicated stellar surface density. To do this, we normalize all spectra by the local mean CO velocity and then average all spectra, including upper limits. Because we average over a large area in the low $\Sigma_\star$ bins, each bin yields a high significance detection of HCN intensity, despite the modest signal to noise in each individual beam.}
\vspace{0.2in}
\label{fig5}
\end{figure*}
Stars form out of dense molecular gas. This is evident from Milky Way studies that resolve individual clouds and isolate star-forming clumps \citep[e.g., see][]{lada03,heiderman10,lada10,andre14}. It can also be seen from the good correspondence between spectroscopic tracers of dense gas, like the low-$J$ transitions of HCN and HCO$^+$ \citep{gao04a,gao04b,gracia06,juneau09,garcia12,kepley14,usero15,chen15} or higher-J molecular lines \citep[e.g.,][]{zhang14,liu15} to tracers of recent star formation in star-forming galaxies and Milky Way cores \citep{wu05,wu10}. Thus, understanding what sets the equilibrium fraction of gas that is dense and how dense gas relates to star formation across the galaxy population is a crucial next step to understand how galaxies convert their gas reservoirs into stars.
Dense structures within clouds are very compact and thus hard to resolve at extragalactic distances. Spectroscopy of molecules with different density sensitivities offers the best way to systematically pursue this topic in other galaxies. By contrasting emission from lines excited at low density (e.g., low-J CO emission, ${\rm n_{eff}\approx10^2~cm^{-3}}$) with emission from lines excited only at high density (e.g. ${\rm HCN(1-0)~emission, n_{eff}\approx10^5~cm^{-3}}$), one can gauge the fraction of gas in a beam that is dense, $\mbox{$\rm f_{dense}$} = {\rm M_{dense} / M_{gas}}$. Similarly, by comparing tracers of the recent star formation rate (SFR) to lines that trace the dense gas mass, one can explore variations in the efficiency with which dense gas forms stars, $\mbox{$\rm SFE_{dense}$} = {\rm SFR / M_{dense}}$.
With these goals in mind, we are using the IRAM 30-m telescope\footnote{Based on observations carried out with the IRAM 30m Telescope. IRAM is supported by INSU/CNRS (France), MPG (Germany) and IGN (Spain).} to carry out the first wide-area multi-line mapping survey targeting tracers of dense gas across the disks of star-forming galaxies. The ``EMIR Multiline Probe of the ISM Regulating Galaxy Evolution'' (EMPIRE, Bigiel et al., in prep.) is mapping a suite of density-sensitive transitions in the 3-mm atmospheric window over the full area of active star formation in nine nearby disk galaxies. This paper reports the first results of this survey for the prototype target, M51a (NGC 5194).
EMPIRE's main goals include understanding how \mbox{$\rm f_{dense}$}\ and \mbox{$\rm SFE_{dense}$}\ depend on environment within a galaxy. The $\sim 30\arcsec$ beam of the IRAM 30-m, $\approx 1$--$2$~kpc at the distance of our targets, does not resolve individual clouds. However, it does localize quantities related to galactic structure like the stellar surface density, gas surface density, and --- to some degree -- dynamical environment, allowing one to measure how \mbox{$\rm f_{dense}$}\ and \mbox{$\rm SFE_{dense}$}\ depend on these factors.
EMPIRE builds on the seminal work by \citet{gao04a}, who surveyed HCN from the bright regions of active galaxies to show that $\mbox{$\rm SFE_{dense}$}$ varies weakly among galaxies, while $\mbox{$\rm f_{dense}$}$ varies strongly. This provided a main piece of observational evidence for a universal gas density threshold for star formation. Surveying similar systems and using the HCO$^+$ line, \citet{garcia12} showed that this situation appears more complex, with $\mbox{$\rm SFE_{dense}$}$ varying between starburst and normal star-forming galaxies. These surveys targeted whole bright galaxies. For individual pointings across 30 disk galaxies spanning a range of physical conditions, \citet{usero15} showed that both \mbox{$\rm f_{dense}$}\ and \mbox{$\rm SFE_{dense}$}\ within their $\sim 1$--$2$~kpc beam depended strongly on these conditions.
Given the apparent dependence of \mbox{$\rm f_{dense}$}\ and \mbox{$\rm SFE_{dense}$}\ on local conditions, the natural next step must be complete, unbiased maps of a set of galaxies with a wide range of global and local conditions. As the first systematic multi-line mapping survey targeting the whole disks of galaxies, EMPIRE represents this natural next step. This letter demonstrates from complete mapping that both \mbox{$\rm f_{dense}$}\ and \mbox{$\rm SFE_{dense}$}\ vary systematically across the disk of M51.
\section{Observations}
\label{data}
We observed M51 at the IRAM 30-m telescope for 75 hours in seven consecutive runs in July and August 2012 under average summer conditions. We used the 3\,mm band of the sideband-separating dual-polarization EMIR receiver \citep{carter12} and the Fourier transform spectrometers. This yields a bandwidth of 15.6\,GHz per polarization with a channel spacing of 195\,kHz. We tuned to simultaneously measure the ${\rm HCO^+(1-0) [89.19\,GHz]}$, ${\rm HCN(1-0) [88.63 GHz]}$, ${\rm HNC(1-0) [90.66 GHz]}$, H$^{13}$CO$^+(1-0)$ [86.75 GHz], H$^{13}$CN$(1-0)$ [86.34 GHz], and HN$^{13}$C$(1-0)$ [87.09 GHz] transitions. This letter reports first results for HCN emission. Jim\'enez-Donaire et al. (in prep.) present the optically thin isotopologues and report their ratios to other lines.
M51 was mapped in on-the-fly mode, with a dump time of 0.5\,s and a scanning speed of $9\arcsec/$s, which yielded 6 dumps per beam along the scanning direction at the $\approx 28\arcsec$ angular resolution of the ${\rm HCO^+(1-0)}$ line. We covered a field-of-view of $\sim4.2\arcmin\times5.7\arcmin$ at a position angle of $-7.5\degr$ centered on ($\alpha={\rm 13^h29^m52.532^s}$, $\delta={\rm 47^d11^m41.98^s}$). The top left panel of Figure \ref{fig1} shows the footprint of the map on top of the PAWS $^{12}$CO$(1-0)$ map \citep{schinnerer13,pety13}, demonstrating that we cover most of the area of bright CO emission in the galaxy.
We reduced the data using the \textsc{gildas/class} software\footnote{\texttt{http://www.iram.fr/IRAMFR/GILDAS}; for more information see \citet{pety05}.}. First, we extracted a 300\,MHz-wide window around the rest frequency of each line. We flagged spectra with noise larger than expected from the radiometer equation. The rest of the spectra were gridded into a data cube using a Gaussian kernel with a FWHM one third of the telescope beam and a channel width of 7~km~s$^{-1}$. We fit and removed a third order baseline using the signal-free part of each spectrum (determined by comparison to $^{12}$CO emission) in the cube.
We convolve all data to $30\arcsec \approx 1.1$~kpc resolution at the $7.6$~Mpc distance to M51 \citep{ciardullo02} and adopt an inclination of $i\approx22\deg$. We sample all maps using a common set of hexagonally-packed points spaced by $15\arcsec$, or one half beam. For each point, we calculate intensities integrated over the range of velocities with bright $^{12}$CO$(1-0)$ emission. Uncertainties in the integrated intensities are estimated using the noise estimated from the signal-free region and the width ($\sim40$~km~s$^{-1}$) of the integration window. A typical uncertainty for an individual $30\arcsec$ line of sight in the EMPIRE maps is 0.06~K~km~s$^{-1}$.
Note that we do not sigma clip the data or otherwise restrict our measurements to regions of bright HCN emission. We calculate an HCN intensity with an associated uncertainty for every line of sight in the map. In our presentation of results below, we also average these individual lines of sight over large areas (by binning HCN intensity sorted by, e.g., stellar surface density), which yields a signal at high significance. This has the large advantage that our results remain sensitive to faint emission that may not be detected at high significance in individual lines of sight. We rigorously propagate and plot uncertainties in the mean trends, accounting for the oversampling of our data.
In addition to the EMPIRE data, we use the IRAM 30-m map of $^{12}$CO$(1-0)$ emission from PAWS and multi-band {\em Herschel} imaging \citep{mentuch12}, which we combine to estimate the total infrared surface brightness along the line of sight following \citet{galametz13}. We also use the THINGS {\sc Hi} map from \citet{walter08} and a map of stellar surface density estimated from {\em Spitzer} near-IR photometry by \citet{meidt12,querejeta15} using a 3.6$\mu$m mass-to-light ratio of ${\rm \sim 0.5 M_\odot / L_\odot}$, so that ${\rm\Sigma_\star \left[ M_\odot~{pc}^{-2}\right] = 350\,I_{3.6,corr} \,\left[ {MJy~sr}^{-1}\right]}$. The {\em Herschel} and {\em Spitzer} data are extremely high signal to noise, so that the dominant uncertainty is the calibration. This is uncertain at the $\sim 5\%$ level, but systematic. Similarly, the CO and {\sc Hi} data have much higher signal-to-noise than EMPIRE data. The dominant statistical uncertainty throughout the paper is thus that of the HCN intensity; we propagate this and show it in the figures and mean trends.
\section{Results}
\label{results}
Figure \ref{fig1} shows the integrated intensity maps of HCN~(1-0), HCO$^+$~(1-0) and HNC~(1-0), the three main dense gas tracers in EMPIRE. For comparison, the top left panel also shows the $^{12}$CO~(1-0) map from PAWS \citep{pety13}, tracing the overall distribution of molecular gas. We detect emission from all three dense gas tracers across the disk of M51, with emission brightest in the center and along the spiral arms. All three dense gas tracers align well with the structure of the CO emission, with HCN moderately brighter than HCO$^+$, and both brighter than HNC. CO is far brighter than all three molecules; the map that we show has been scaled by the typical HCN-to-CO ratio of 1-to-30 to bring it onto a common intensity scale with the EMPIRE dense gas maps.
Figure \ref{fig2} compares the amount of dense gas, traced by HCN, to the amount of recent star formation, traced by IR luminosity. In the left panel, we convert each individual sampling point to a luminosity and plot these on the global scaling relation between IR luminosity and HCN luminosity following \citet{gao04b} and \citet{wu05}. The figure shows that our M51 data extend the IR-HCN correlation seen for whole galaxies to much lower luminosities, overlapping work for individual pointings by \citet{usero15} and \citet{bigiel15} and maps by \citet[][M82; see also Chen et al. (2015), M51]{kepley14}. These data help to fill in the ``luminosity gap'' between individual Galactic cores \citep{wu10} or clouds \citep{rosolowsky11,brouillet05} and whole active star-forming galaxies \citep{gao04b,garcia12}. Across $\sim 8$ orders of magnitude in luminosity, the data scatter around roughly the same IR-to-HCN ratio, $L_{\rm TIR} / L_{\rm HCN} \approx 900$~L$_\odot$~(K~km~s$^{-1}$)$^{-1}$ \citep[][see also Gao et al. 2007]{gao04b}, shown as a thick gray line.
The right panel zooms in on the luminosity-luminosity relationship for only our new M51 data and the disk pointings of \citet{usero15}. Here each point shows an individual $\sim$ kpc sized beam (sampling point) with the black points and error bars showing a running median and the $1\sigma$ scatter in bins of fixed HCN luminosity. In agreement with the recent work of \citet{chen15}, there is a good correspondence between star formation (IR) and dense gas (HCN) across the disk of M51, with a Spearman rank correlation coefficient of 0.8 relating the two. At a given HCN luminosity, the $1\sigma$ scatter is between a factor of 1.3 at high luminosities and 2 near the low luminosity end. M51 shows similar behavior to the pointings from \citet{usero15}, though it is somewhat brighter in HCN than the other galaxies in the sample at intermediate luminosities. This might be a result of the ongoing merger driving gas in M51 to higher densities but also adding turbulence that prevents it from collapsing. In any case, we expect that comparison with the other full-galaxy maps from EMPIRE will illuminate the origin of this offset.
The right panel in Figure \ref{fig2} also shows that across the disk of M51, the relationship of IR to HCN is not perfectly 1-to-1, showing a high degree of scatter and a mildly sublinear slope. That is, \mbox{$\rm SFE_{dense}$}\ as traced by the IR-to-HCN ratio (fixed for diagonal lines) is lower at high luminosities in both our M51 map (by about 60$\%$) and the \citet{usero15} sample. \citet{usero15} found \mbox{$\rm SFE_{dense}$}\ to depend on environment, being lower at high stellar surface densities and high molecular fractions \citet{chen15} observed qualitatively similar behavior in their analysis of M51, noting that the IR-to-HCN ratio drops with decreasing radius in the galaxy.
In Figure \ref{fig3} we show that the same is true for our whole-galaxy map of M51. The IR-to-HCN ratio varies systematically across the galaxy, becoming lower in regions with high stellar surface densities and high molecular fractions. There is good quantitative agreement with the individual \citet{usero15} pointings for $30$ galaxies. The trend is particularly evident in the binned relation (red); here the error bars reflect propagated statistical uncertainty in the mean, which is dominated by statistical noise in the HCN data.
The amount of dense gas also varies systematically across M51, showing trends opposite to the sense of what we see for \mbox{$\rm SFE_{dense}$}. Figure \ref{fig4} shows that the fraction of gas mass in a dense phase increases with increasing stellar surface density and molecular gas fraction, so that at lower galactocentric radii and higher interstellar pressures, more of the gas is dense. Again, the binned (red) trend for M51 shows continuous variation at high significance. This agrees with the scenario proposed by \citet{helfer97} based on early observations of HCN in the inner regions of a few galaxies, it is quantitatively consistent with the observed results from \citet{usero15} and qualitatively consistent with what is seen in our own Galaxy \citep[e.g.,][]{longmore13}. It presents a clear challenge to the idea that star formation proceeds in a universal way above a gas density threshold similar to the effective density of ${\rm HCN(1-0)}$.
The very significant trends in the red (binned) points in Figures \ref{fig3} and \ref{fig4} emerge from a large set of individual, lower signal-to-noise pointings. Using spectral stacking techniques similar to \citet{schruba11}, we have verified that these averaged HCN measurements reflect an astronomical signal; that is, after averaging it comes from a spectral line feature coincident in velocity with CO. Figure \ref{fig5} shows the results of this stacking for HCN emission from the six lowest stellar surface density bins in Figures \ref{fig3} and \ref{fig4}. In each case, the stacking yields a clear, significant astrophysical line signature. We emphasize that this averaging approach is fundamentally similar to what is done to obtain a deep integration with any single dish telescope. In fact, such deep integrations on individual sparse pointings were carried out by \citet{usero15}. We show here that the two approaches yield consistent results.
Both stellar surface density and molecular gas fraction correlate with interstellar pressure and anti-correlate with galactocentric radius, so that in M51 \mbox{$\rm SFE_{dense}$}\ appears to be lowest and \mbox{$\rm f_{dense}$}\ appears to be highest in the high pressure, high surface density regions near the center of the galaxy. Qualitatively, the environment dependence of \mbox{$\rm f_{dense}$}\ that we observe may be explained if most of the molecular gas is in approximate equilibrium with the hydrostatic pressure in the galaxy \citep[e.g.,][]{helfer97,hughes13}. In this case, the higher stellar surface densities and molecular fractions indicate higher interstellar gas pressures, which in turn shifts the overall gas density distribution to higher values \citep[e.g.,][]{elmegreen93,blitz06,ostriker10,shi11}. This leads a larger fraction of the gas to be dense. The changing \mbox{$\rm SFE_{dense}$}\ may then be explained if star formation only occurs in regions that show a high density contrast relative to the local mean gas density, a scenario consistent with current turbulent models of star formation \citep[e.g., see][for more discussion]{federrath12,usero15}. Then in the dense parts of galaxies, the average gas cloud may be denser and better at emitting HCN, but the HCN-emitting gas may also no longer correspond to the high density, self-gravitating, immediately star-forming tail of the density distribution.
In forthcoming papers, we test this and competing scenarios using the entire sample of nine EMPIRE galaxies and our full suite of density-sensitive molecular lines. Doing so, we expect to make the most thorough test to date of how physical conditions in the molecular gas depend on galactic environment and, in turn, influence star formation.
\section{Summary}
We present new, sensitive maps of M51 in lines sensitive to high density gas. These represent the first result of the IRAM large program ``EMPIRE,'' which aims to use multi-line spectroscopy to relate physical conditions in the cold ISM to star formation in $9$ nearby galaxies. Compared to previous work targeting whole galaxies, we confirm a tight scaling between the amount of dense gas, traced by HCN, and the recent star formation rate, traced by IR emission. But we show that the underlying physics are more complex: we show that both the dense gas fraction and the star formation efficiency of dense gas appear to depend on local conditions in the disk of M51. The sense of this trend is a high \mbox{$\rm f_{dense}$}\ and a low \mbox{$\rm SFE_{dense}$}\ in the high surface density, high molecular fraction parts of the disk. This agrees with work on individual scattered pointings in $30$ galaxy disks by \citet{usero15} but here the results have no bias, they cover the entire galaxy, and by binning the data we show these relations at very high statistical significance. The result is consistent with the ISM pressure playing an important role in setting the density of the ISM and the ability of dense gas to form stars.\\
We thank Ga\"elle Dumas for assisting with data reduction and Fabian Walter for helpful discussions and feedback on the draft. FB, MJ and DC acknowledge support from DFG grant BI 1546/1-1. AH acknowledges support from the Centre National d'Etudes Spatiales (CNES). AU acknowledges support from Spanish MINECO grants AYA2012-32295 and FIS2012-32096. JP thanks the CNRS/INSU programme PCMI for support. NT acknowledges funding from DFG grant SCHI 536/8-2 as part of the priority program SPP 1573 ``Physics of the Interstellar Medium''.
|
\section{Introduction}
\label{intro}
The low-energy physics of a many-body system is governed by the first excitations above its ground state. Often,
these excitations are not directly given by the (dressed) elementary constituents of the system, but are instead complicated objects.
Bound states represent an important class of such excitations.
Superconductivity is probably the best known
example \cite{degennes99}, although not the only one of experimental interest: low dimensional quantum systems, such as 1d cuprate
ladder materials, or the 2d magnetic system SrCu$_2$(BO$_3$)$_2$, are known to also present bound states, whose modes appear in their
measured spectra \cite{uhrig96,trebst00,knetter01,zheng01,windt01}. Beyond the realm of condensed matter, bound states are also very important in the theory of nuclear forces \cite{ZINN,cohen11}, as well as in quantum chemistry \cite{levine09}.
Of particular interest are bound states emerging in strongly correlated systems,
where they are less understood and difficult to characterize using standard perturbative techniques.
Examples of such systems arise in the theory of the strong nuclear interactions, QCD \cite{jaffe77,shifman79}, as well as in the spectrum of quantum excitations in strongly correlated electron systems \cite{vidal02,vidal00,capponi07,sachdev}. There is no need to stress that the study of this class of subjects is ripe for the development of new
approximation methods, capable of dealing with the complexities that emanate from strong correlations.
In this regard, Renormalization Group methods have already shown to be extremely well adapted to this endeavor, given their focus on scale dependent properties, and their proven abilities to deal with strong correlations.
The scalar $\phi^4$ field theory probably represents the simplest example of a strongly correlated system, exhibiting long-range order and a diverging correlation length at the critical point. Near this point, it belongs to the Ising universality class, and a lot of effort has
been put into the understanding of its properties by a myriad of methods,
including Monte-Carlo simulations, high order perturbative expansion \cite{Guida98,hasen10,campos06} and the conformal bootstrap \cite{rychkov12,gliozzi14,rychkov14,kos14,rychkov14b}.
The complex yet manageable behavior of $\phi^4$ theory is thus ideally suited for purposes of benchmarking new approximation methods \cite{BMWpre,BMWlong}, as well as for revisiting a subject of intrinsic interest.
In two spatial dimensions, or, equivalently, in the 1d quantum case at zero temperature \cite{sachdev},
the integrability of the Ising model, and thus of its universality class close to criticality,
allows for the complete determination of the bound state spectrum \cite{zamolo1,zamolo2}. These exact
results stem from the conformal invariance of the theory at criticality. In the language of the classical two-dimensional model, this bound state exists in the presence of a small magnetic field, and when the temperature is exactly the critical temperature. The observed ratio between the mass of the first bound state and of the elementary excitation is $m_1/m_0 = (1+\sqrt{5})/2$ (golden ratio), which has been experimentally confirmed
\cite{exp} in the quasi-1d quantum Ising ferromagnet CoNb$_2$O$_6$, by means of inelastic neutron scattering. Six other bound states are known to exist in this case with only two that are below the mass of the threshold $2m_0$ of the multi-particle continuum.
In the case of the three dimensional $\phi^4$ model, the presence of a bound state in the symmetry broken phase is by now a well-established fact.
A classical argument at $T=0$ shows the existence of a bound state for the Ising model \cite{hasen97}, although this is of course a non-universal result.
For the $\phi^4$ theory, a bound state with a mass ratio $M/m$ around $1.8$ was first detected by Monte Carlo simulations \cite{hasen97,hasen99} of the Ising model at temperatures lower than $T_c$, but still within the scaling regime, and thus expected to be universal. This prompted the use of resummed perturbative calculations by means of a Bethe-Salpether equation, where the leading order yields results compatible with Monte Carlo values \cite{hasen00,hasen02}. However, the next-to-leading order result leads to an unphysical conclusion $M/m < 0$, indicating a strongly non-perturbative behavior for this quantity. Alternatively, this bound state can also be detected studying the quantum 2d Ising model at zero temperature, as was performed using high-order perturbative continuous unitary transformations \cite{vidal10}. More recently, Nishiyama \cite{ponja} found $M/m = 1.84(1)$ using numerical diagonalization methods.
For the three dimensional $\phi^4$ model then, the observed bound state appears to be outside of the standard perturbative Renormalization Group regime. It is interesting however to see whether other, non perturbative, methods based on the RG are able to detect its presence. One of the main difficulties is that a non-trivial
(and in particular, non-analytic) description of the momentum dependence of the correlation
functions of the system is needed to detect bound states, as is discussed below.
Some years ago, a new approximation scheme of the Non Perturbative Renormalization Group
(NPRG, also known sometimes as the Functional or Exact RG), the Blaizot--M\'endez-Galain--Wschebor (BMW) approximation, has been shown to give accurate momentum
dependent results for scalar field theories \cite{BMW,BMWlong}. In this work, we use this approximation to compute the $\phi^4$ bound state mass within the NPRG, for spatial dimensions between $d=2$ and $d=4$. This not only shows the strength of this multi-purpose method, but also allows us to study in a novel way the temperature-dependence of this bound state, even in the non-universal region of the model.
This paper is organized as follows: in section \ref{2point}, we discuss how to check for bound states
using the momentum dependence of the two-point correlation function of the system, and section \ref{BMWpourlesnuls} briefly presents the approximation scheme used for obtaining the full momentum dependence of this function.
Section \ref{num} discusses the numerical implementation, as well as the numerical analytic continuation
procedure, before presenting our main results in section \ref{results}. Finally, we present our conclusions in section \ref{conclusion}.
\section{Signature of a bound state in the spectral function}
\label{2point}
For concreteness, we use the language of classical equilibrium statistical mechanics, but the case of $d-1$
quantum statistical systems at zero temperature corresponds to a trivial renaming of the fields.
The microscopic Euclidean action of the model is written in the well-known Ginzburg-Landau form \cite{ZINN}
\begin{equation}
S[\varphi] = \int d^dx \left\{ \frac 1 2 \big(\nabla \varphi(x) \big)^2 + \frac{r_0}{2} \varphi^2(x) + \frac {u_0}{4!} \varphi^4(x) \right\}.
\ee
When performing Monte Carlo simulations of this system on a lattice, bound states can be most easily
detected by studying the spatial behavior of the two-point connected correlation function.
In the symmetry broken phase, one expects correlations decaying exponentially with distance
\begin{equation}
\label{2pointreal}
\langle\varphi(x) \varphi(0) \rangle_c \underset{x \to \infty}{\sim} Ae^{-mx},
\end{equation}
with $m=\xi^{-1}$ the inverse correlation length, usually termed the ``mass'' in analogy with Quantum Field Theory.
For a theory with a non-trivial spectrum, sub-leading exponentials are expected as well:
\begin{equation}
\label{2pointreal2}
\langle\varphi(x) \varphi(0) \rangle_c \underset{x \to \infty}{\sim} A_0e^{-m x} + A_1e^{-Mx} + \ldots
\end{equation}
which are associated with bound states of the theory, in that they give the sub-leading correlation lengths.
While this is the standard technique for finding bound states when using simulations \cite{hasen97},
one can alternatively study the momentum-dependent spectral function, defined by the Fourier transform
\begin{equation}
G(p) = \int d^dx \,\langle\varphi(x) \varphi(0) \rangle_c \,e^{-ipx}.
\end{equation}
The presence of sub-leading exponential decay terms can also be seen in the analytic
continuation of $G(p)$ to complex values of $p$. Indeed, $G(p)$
behaves in the infrared limit $p\to0$ as
\begin{equation}
G(p)\underset{p \to 0}{\sim} \frac{A'_0}{p^2+m^2}+\frac{A'_1}{p^2+M^2}+\cdots
\end{equation}
This implies that the function $G(\omega=ip)$ has poles
at the values of the masses of the system, with the first mass associated with the correlation length and the
following with bound states or possible many-particle states.
It can be shown \cite{hasen00,hasen02} that at any (finite) order in a perturbative expansion around a free theory,
the ratio of the correlation length to any other length scale must be an integer, forbidding thus the description of bound states.
This issue can be partially solved by performing infinite-order resummations
in some particular channel, but then this expansion seems to be badly behaved \cite{hasen00}.
Not being able to see a non-integer $M/m$ ratio is a problem shared by the simplest approximation schemes within the NPRG, such
as the well-known Local Potential Approximation (LPA), or its higher order generalization dubbed
the Derivative Expansion (DE) \cite{Berges00,Morris94c}.
This approximation amounts to a small momentum expansion of the (vertex) correlation functions.
While it has proven to be very accurate in the low momentum regime, e.g. for the determination of critical
exponents \cite{Berges00,delamotte03,canet03a,canet04a,canet04b,canet05a}, it is not reliable for finite momentum properties and is therefore unable to describe bound states.
As already mentioned in the Introduction, the BMW scheme \cite{BMW} takes into account the full, non-trivial momentum dependency of the correlation functions. This method has been successfully
applied to O$(N)$ scalar field models \cite{BMWlong,BMWONold}, showing excellent results for universal properties such
as critical exponents and momentum-dependent scaling functions. The BMW method has also found applications beyond
the confines of equilibrium statistical mechanics \cite{kpz,adam14,felix}, showing its flexibility
to deal with highly non-trivial momentum dependent quantities. It thus seems very natural to apply this scheme
to the problem of bound states. We present the BMW scheme and its application to $\phi^4$ theory in the following section.
\section{Non-trivial momentum dependence within the NPRG: the BMW approximation}
\label{BMWpourlesnuls}
We start with a brief outline of the NPRG formalism for the case of a scalar field theory \cite{Wetterich92,Ellwanger93,Tetradis94,Morris94b,Berges00}.
The NPRG strategy is to build a family of theories indexed by a
momentum scale $k$,
such that fluctuations are smoothly taken into account as $k$ is lowered
from the microscopic scale $\Lambda$ (e.g. the inverse lattice spacing) down to 0.
In practice, this is achieved by adding to the original
Euclidean action $S$ a ``mass-like'' term of the form
$\Delta S_k[\varphi]= \frac{1}{2} \int_q\: R_k(q^2)\varphi(q)\varphi(-q)$ (here $\int_q\equiv \int \frac {d^dq}{(2\pi)^d}$).
The cut-off function $R_k(q^2)$ is chosen such that $R_k(q^2)\sim k^2$ for $q\lesssim k$, which
effectively suppresses the modes $\varphi(q\lesssim k)$, and such that
it (almost) vanishes for $q\gtrsim k$, leaving
the modes $\varphi(q\gtrsim k)$ unaffected.
One then defines a scale-dependent partition function \begin{equation}
{\cal Z}_k[J] = \int\! {\cal D}\varphi\;
e^{-S[\varphi]- \Delta S_k[\varphi] +\int\! J\varphi} \;,
\label{zk}
\ee
and a scale-dependent
effective action $\Gamma_k[\phi]$ through a (modified) Legendre transform \cite{Berges00},
\begin{equation}
\Gamma_k[\phi] +\log {\cal Z}_k[J] =
\int_q J(q) \phi(-q) -\frac{1}{2} \int_q R_k(q^2) \phi(q)\phi(-q)\,,
\label{legendre}
\ee
with $\phi=\delta \ln{\cal Z}_k/\delta J$ the mean value of the field.
The variation of the effective action $\Gamma_k[\phi]$ as $k$ changes is given by the Wetterich
equation\cite{Wetterich92}:
\begin{equation}
\partial_k\Gamma_k[\phi] = \frac{1}{2} \int_q \partial_k{R}_k(q^2)\, G_k[q,\phi] \;,
\label{rgexact}
\ee
where $G_k[q,\phi]=(\Gamma^{(2)}_k[q,\phi]+R_k(q^2))^{-1}$,
and $\Gamma^{(2)}_k[q,\phi]$ is the second functional derivative of
$\Gamma_k[\phi]$ w.r.t. $\phi$.
With the definitions above, it is easy to show that for $k=\Lambda$, all fluctuations are frozen by the
$\Delta S_k$ term and thus $\Gamma_{k=\Lambda}[\phi]=S[\phi]$. This is the initial condition of the flow equation
(\ref{rgexact}). On the other hand, when $k=0$, $\Delta S_{k=0}\equiv 0$ because $R_{k=0}(q^2)$ vanishes identically
and $\Gamma_{k=0}[\phi]$ is the Gibbs free energy of the original model that we aim to compute.
Differentiating $s$ times Eq.~(\ref{rgexact}) with respect to the field $\phi(q)$
yields the flow equation for the vertex function
$\Gamma^{(s)}_k[q_1,\dots,q_s;\phi]$. Thus, for instance, the flow equation for $\Gamma^{(2)}$
evaluated in a constant field configuration $\phi$ reads:
\begin{equation}
\begin{array}{l}
{\displaystyle\partial_k\Gamma^{(2)}_k(p,\phi)=
\int_q \partial_k{R}_k(q^2) G_k^2(q)\left[
\Gamma^{(3)}_k(p,\!-p\!-\!q,q)\times\right. }\\
\;\;
\left. G_k(p\!+\!q)\Gamma^{(3)}_k(\!-p,p\!+\!q,\!-q)
\!-\! \frac{1}{2}\Gamma^{(4)}_k(p,\!-p,q,\!-q)\right].
\end{array}
\label{rggamma2}
\end{equation}
(Here we have omitted the $\phi$ dependence of the functions $G_k$ and $\Gamma_k^{(n)}$ in the right hand side to alleviate the notation).
Note that the flow equation for $\Gamma^{(s)}_k(q_1,\dots,q_s,\phi)$ involves
$\Gamma^{(s+1)}_k$ and $\Gamma^{(s+2)}_k$, leading to an infinite hierarchy of coupled equations.
The flow equations (\ref{rgexact}) and (\ref{rggamma2}) are exact,
but solving them requires in general approximations. It is precisely one of the virtues of the NPRG to
suggest approximation schemes that are not easily derived in other, more conventional approaches.
In particular, one can develop approximation schemes for the effective action itself, that is, which apply to
the entire set of correlation functions. The BMW approximation\cite{BMW} is such a scheme. It relies on two observations.
First, the presence of the cut-off function $R_k(q^2)$ guarantees
the smoothness of the $\Gamma^{(s)}_k$'s for $k>0$ and limits the internal momentum $q$ in equations such as (\ref{rggamma2})
to $q\lesssim k$. In line with this observation, one neglects the $q$-dependence of the vertex functions in the r.h.s.
of the flow equations (e.g. in $\Gamma^{(3)}$ and $\Gamma^{(4)}$ in Eq.~(\ref{rggamma2})),
while keeping the full dependence on the external momenta $p_i$.
The second observation is that, for uniform fields,
$\Gamma^{(s+1)}_k(p_1,\dots,p_s,0,\phi)=
\partial_\phi \Gamma^{(s)}_k(p_1,\dots,p_s,\phi)$,
which enables one to close the hierarchy of NPRG equations.
At the leading order of the BMW scheme, one keeps the non trivial momentum dependence of the
two-point function and implements the approximations above on Eq.~(\ref{rggamma2}), which
becomes:
\begin{equation}
k\partial_k\Gamma^{(2)}_k(p,\phi) = J_3(p,\phi) \left({\partial_\phi\Gamma^{(2)}_k}\right)^2
\!\! -\! \frac{1}{2} J_2(0,\phi)\, \partial_\phi^2\Gamma^{(2)}_k
\label{BMW}
\end{equation}
with
\begin{equation}
J_n(p,\phi) \! \equiv \! \int_q k\partial_k{R}_k(q^2) \, G_k^{n-1}(q,\phi) G_k(p\!+\!q,\phi)\;.
\label{integraleJ}
\end{equation}
The approximation can be systematically improved: The order $s$ consists in keeping the full momentum
dependence of $\Gamma^{(2)}_k,\dots, \Gamma^{(s)}_k$ and truncating that of
$\Gamma^{(s+1)}_k$ and $ \Gamma^{(s+2)}_k$ along the same lines as those leading to
Eq.~(\ref{BMW}) for the case $s=2$, with a corresponding increase in the numerical complexity.
In order to treat efficiently the low (including zero) momentum sector,
we work with dimensionless and renormalized quantities.
Thus, we measure all momenta in units of $k$:
$\tilde{p}=p/k$. We also rescale $\rho\equiv \frac{1}{2}\phi^2$ according to
$\tilde{\rho}=k^{2-d} Z_k K_d^{-1}\,\rho$ (with the constant $K_d=(2\pi)^{-d} S_d/d$, and $S_d$
being the volume of the unit sphere), and set $\tilde\Gamma^{(2)}_k(\tilde{p},\tilde{\rho})=
k^{-2}Z_k^{-1}\Gamma^{(2)}_k(p,\rho)$.
The running anomalous dimension $\eta_k$ is defined by
$k\,\partial_k Z_k= -\eta_k Z_k$, so that at criticality $\eta_{k\to0}\to\eta$, with $\eta$ the anomalous dimension. Thus,
at a fixed point $Z_k\sim k^{-\eta}$. The absolute normalization of $Z_k$ is fixed by
choosing a point $(\tilde{p}^*,\tilde{\rho}^*)$ where
$\partial_{\tilde{p}^2}\tilde{\Gamma}^{(2)}_k\vert_{\tilde{p}^*,\tilde{\rho}^*}=1$.
Here, we have chosen $\tilde{p}^*=0$ and $\rhot^*=\rhot_{0,k}$,
where $\rhot_{0,k}$ is the $k$-dependent running minimum of the potential.
Then, the flow equation of
$\tilde\Gamma^{(2)}_k(\tilde{p},\tilde{\rho})$ follows trivially from
Eq.(\ref{BMW}).
It is actually more accurate to disentangle the potential part of $\tilde\Gamma^{(2)}_k(\tilde{p},\tilde{\rho})$
from the momentum part and to solve independently the flows of these two quantities \cite{BMWlong}. We thus solve two equations: one for
$\tilde{Y}_k(\tilde{p},\tilde{\rho}) \equiv \tilde{p}^{-2}
[\tilde{\Gamma}^{(2)}_k(\tilde{p},\tilde{\rho})-
\tilde{\Gamma}^{(2)}_k(0,\tilde{\rho})]-1$ and one for
the derivative of the dimensionless effective potential
$\tilde{W}_k(\tilde\rho)=\partial_{\tilde \rho} \tilde{V}_k(\tilde \rho)$, with $\tilde{V}_k(\tilde \rho)=K_d^{-1}k^{-d}V_k(\rho)$.
Note that
$\tilde{\Gamma}^{(2)}_k(0,\tilde{\rho})=
\tilde{W}_k(\tilde\rho) + 2 \tilde\rho\,\tilde{W}_k'(\tilde\rho) $.
Here and below, primes denote derivative w.r.t. $\tilde\rho$, or, in the case of dimensionful variables, w.r.t. $\rho$.
These two equations read (dropping the $k$ index and the $\tilde{\rho}$ and $\tilde{p}$ dependences to alleviate the notation):
\begin{align}
\partial_t \tilde{Y} ={}& \eta_k (1+\tilde{Y})+\tilde{p}\, \partial_{\tilde{p}} \tilde{Y} -(2-d-\eta_k)\tilde{\rho}\,\tilde{Y}'\nonumber\\
& + 2{\tilde{\rho}}\,\tilde{p}^{-2}\left[(\tilde{p}^2\,\tilde{Y}'\!+\!\tilde\lambda_k)^2 \tilde{J}_3(\tilde{p},\tilde{\rho})
-\tilde\lambda_k^2 \tilde{J}_3(0,\tilde{\rho})\right] \nonumber\\
& - \tilde{J}_2(0,\tilde{\rho})(\tilde{Y}'/2+\tilde{\rho}\,\tilde{Y}''),\label{eqY} \\
\partial_t \tilde{W} ={}& (\eta_k \!-\! 2) \tilde{W} +(d \!-\! 2 \!+\! \eta_k)\tilde{\rho}\,\tilde{W}'+ \frac{1}{2} \tilde{J}_1'(0,\tilde{\rho}). \label{eqW}
\end{align}
Here, the renormalization ``time'' $t$ is defined by $t=\log k/\Lambda$, $\partial_t=k\partial_k$,
$\tilde{J}_n(\tilde{p},\tilde{\rho})= K_d^{-1}Z_k^{n-1}k^{2n-d-2}{J}_n({p},{\rho})$
and $\tilde{\lambda}_k(\tilde\rho)=
3\tilde{W}_k'(\tilde\rho)+2\tilde\rho\,\tilde{W}_k''(\tilde\rho)$. The running anomalous dimension
$\eta_k$ is obtained by setting
$\tilde{Y}_k(\tilde{p}^*,\tilde{\rho}^*)=0$ in Eq.(\ref{eqY}) and taking a time derivative,
noting that $ \partial_t \rhot_{0,k}=- \partial_t \tilde{W}_k(\tilde\rho)|_{\rhot_{0,k}}/\tilde{W}'_k(\rhot_{0,k})$.
At the end of the numerical flow, the two-point vertex function $\Gamma^{(2)}(p,\rho)$ can be reconstructed from $\tilde{Y}_k$ and $\tilde{W}_k$:
\[
\Gamma^{(2)}(p,\rho)=\lim_{k\to0}Z_k k^2 \bigg[ \tilde{p}^2\Big(1+\tilde{Y}_k(\tilde p,\tilde \rho)\Big)+\tilde{W}_k(\tilde \rho)+2\tilde\rho \tilde{W}_k'(\tilde \rho) \bigg].
\]
The flow equations (\ref{eqY}) and (\ref{eqW}) can be solved using standard numerical techniques,
see section \ref{num}. Crucially for our purposes, the obtained $k \to 0$ values for
the $\Gamma^{(2)}(p)=G(p)^{-1}$ function must then be extended to the whole complex plane.
This is not a completely well-defined mathematical operation, given that we are extending a
discrete set of numerical values to the whole plane. Nonetheless, this type of extension is often used e.g.
in systems studied using Quantum Monte Carlo methods \cite{Pade,PadeClean}, and are known to yield useful results.
Here, we perform this analytic continuation from Pad\'e approximants of our results, see the next section.
Note that if we neglect non-trivial momentum dependence and set $\tilde{Y}=0$, in the ordered phase $\Gamma^{(2)}(p,\rho=\rho_{0})$ cancels only for $p=\pm i \Delta$,
\begin{equation}
\Delta=\sqrt{\dfrac{2\rho_{0} W_{k=0}'(\rho_0)}{Z_{k=0}}}.
\end{equation}
In this case, $\Delta$ is the mass (the inverse correlation length) and there are no bound states.
Within the BMW approximation, this is no longer true and bound states can exist. The actual mass $m$
of the elementary excitation, that is, the leading correlation length, is close to $\Delta$.
It is therefore useful to retain $\Delta$ as a relevant energy scale, which needs not to be extracted
from an analytic continuation.
It should be mentioned that, recently, an alternative formulation of the NPRG capable of dealing
with the analytic continuation of spectral functions has been proposed in a slightly
different context \cite{vonsmekal14,vonsmekal14b,vonsmekal14c}. In this approach, the analytic
continuation is performed at the level of the flow equations of the NPRG, so that one ends up
with a flow of complex quantities. This is a very promising approach, which has so far only been
applied together with Derivative Expansion-like approximations, resulting in a less accurate
momentum description of the theory than our approach. Ideally, these ideas could be implemented together with a
BMW type of approximation in the near future.
\section{Numerical Procedure}
\label{num}
In this section, we give the key points of the numerical integration of the flow
equations (\ref{eqY},\ref{eqW}), and then we detail the analytic continuation used to extract
the pole of the correlation function, where, as shown below, subtle issues may arise.
\subsection{Flow integration}
The integration of the flow equations is based on well-established numerical analysis methods.
The (renormalization) time evolution is done through explicit Euler integration scheme with a time step $dt=-10^{-4}$.
The momentum dependence of $\tilde{Y}_k(\tilde p,\tilde\rho)$ is studied on the interval $\tilde{p} \in [-\tilde{p}_{\rm max},\tilde{p}_{\rm max}]$
with $\tilde{p}_{\rm max}=10$ and we use a
Chebyshev pseudo-spectral approximation of this function with a variable number of polynomials ranging from 20 to 50.
The field-dependence of $\tilde{Y}_k(\tilde p,\tilde\rho)$ is obtained by discretizing the $\tilde\rho$-space on a finite and regular grid
$\tilde{\rho} \in [0,\tilde{\rho}_{\rm max}]$ with $\tilde{\rho}_{\rm max}$ comprised between 10 and 14.
The lattice spacing of the $\tilde{\rho}$ grid is $d\tilde{\rho} = 0.1$.
We choose an exponential regulator,
\begin{equation}\label{regulator}
R_k(q^2)= \alpha \dfrac{Z_k q^2}{\exp(q^2/k^2)-1},
\end{equation}
where $\alpha$ is an arbitrary parameter that is varied to study the sensitivity of our results with the choice of regulator.
Since the integrals over the momenta involved in the RG flows, Eq.~(\ref{integraleJ}), are all
exponentially cut-off by the (derivative of the) regulator, we restrict their range
to $\vert\tilde p\vert \leq 4$.
These integrals are then computed using a Gauss-Legendre approximation with $40$ points. A good numerical accuracy of the integrals is mandatory to obtain converged results, specially for dimensions $d<3$.
Further difficulties arise when studying the ordered phase. First of all, in this phase,
$\rho_{0,k}$, which is the minimum of the running potential, goes to a constant value $\rho_{0}>0$ as $k$ goes to zero,
since it is half the square of the spontaneous magnetization. Since $\rhot_{0,k} \sim \rho_{0,k}/Z_kk^{d-2}$,
this means that $\rhot_{0,k}$ diverges as $k$ goes to zero. In practice, starting close to criticality,
we observe that in the first stage of the flow $\rhot_{0,k}$ evolves towards its fixed point value. Then,
when $k$ is of the order of the inverse of the correlation length $\xi$, it starts diverging which means
that $\rho_{0,k}$ has almost converged to its final value. Hence, when the flow leaves the critical regime,
we switch to the flow of $\tilde W_k(\rho)$ and $\tilde Y_k(\tilde p,\rho)$ (instead of $\rhot$) while keeping the same number of points on a dimensionful grid in $\rho$.
This allows $\rho_{0,k}$ to remain inside the grid as $k$ goes to zero.
The second difficulty is numerical.
When $k\to0$, the inner part of the potential, $\rho< \rho_{0}$, becomes
flat because of the convexity of the effective potential $V=V_{k=0}$. The convexity of $V$
is reproduced within the BMW approximation and corresponds to the approach of the pole of
the propagator at vanishing momentum
when $k\to0$. Thus, $G_k(p=0,\rho<\rho_{0,k})=\left [R_k(p=0)+W_k(\rho)\right ]^{-1}$ becomes very large at small $k$ which causes numerical instabilities.
The instabilities arise at small $\rho$ since $W_k(\rho)$ is an increasing function of $\rho$.
Since the physics we are interested in corresponds to $\rho=\rho_0$ (there is neither external magnetic field nor
phase coexistence), we eliminate the source of numerical instabilities by eliminating the small values
of $\rho$ from the grid, that is, the values for which $W_k(\rho)$ is `too negative'.
We therefore replace the grid $\rho\in [0,\rho_{\rm max}]$ by a $k$-dependent grid $\rho\in [\rho_{\rm min}(k),\rho_{\rm max}]$
which allows us to continue the flow to smaller values of $k$.
We expect that this supplementary approximation has a small impact on the final results.
However, a difficulty remains for small $k$: $\rho_{\rm min}(k)$ becomes close to $\rho_{0,k}$ and there are no longer
enough points in the $\rho$-grid on the left of $\rho_{0,k}$ to compute the derivatives of the potential at this point. The flow
must then be stopped and the smallest value of $k$ we have been able to reach is typically $k_{\rm min}\simeq 0.1\Delta$.
The function $\Gamma^{(2)}(p,\rho_0)$ is finally obtained using the approximation:
\begin{equation}
\Gamma^{(2)}_{k=0}(p)\simeq \Gamma^{(2)}_{k=p/\tilde{p}_{\rm max}}(p),
\label{approx}
\end{equation}
$k=p/\tilde{p}_{\rm max}$
being the smallest value of $k$ for which $p/k$ is still in the
dimensionless grid $[0,\tilde p_{\rm max}]$. This approximation is justified by the fact that $p$ acts as an effective
infrared cutoff in the flow of $\Gamma^{(2)}_k(p)$ that therefore effectively stops for $k\ll p$. Thus, stopping the flow of $\Gamma^{(2)}_k(p,\rho_{0,k})$ at
$k\ll p$ or at $k=0$ should yield almost the same result. We have checked the validity of the approximation (\ref{approx}) by varying $\tilde{p}_{\rm max}$
and observing that it is indeed almost insensitive to $\tilde{p}_{\rm max}$ when it is of order 10.
\subsection{Analytic continuation}
Let us now detail the Pad\'e approximation procedure used to obtain the spectral function
$G(\omega=ip-0^+)=G_{k=0}(i p-0^+,\rho_0)$. First, we compute the propagator $G(p)$ for
$N$ momentum values $p_i$ evenly
spaced in a window $[\omega_{\rm min}, \omega_{\rm max}]$.
Typically, $N\sim 30-50$, $\omega_{\rm min}\sim \Delta$ and $\omega_{\rm max} \sim 10 \Delta$.
We then construct a $[N-2/N]$ Pad\'e approximant \cite{Pade, PadeClean} $F$, even in $p$, that satisfies
$F(p_i)=G(p_i)$ for all $i$. Once $F$ is known, we evaluate ${\rm Im} \left [F(\omega=ip-0^+)\right ]$
as an approximation of ${\rm Im}[ G(\omega)]$ that shows peaks where $G(ip)$ has poles.
\begin{figure}[ht]
\includegraphics{propeucl.pdf}
\includegraphics{propim.pdf}
\caption{(Color online) Comparison of three different Pad\'e approximants of (top) $\Gamma^{(2)}(p)$, (bottom)
${\rm Im}[G(\omega)]$. Top: we show (in arbitrary units) the vertex $\Gamma^{(2)}(p)$
(full line), obtained from the numerical integration of the flow, as well as its fit by three
different approximants (symbols). Bottom: the spectral functions ${\rm Im} [G(\omega)]$ (in arbitrary units) obtained from
the analytic continuation of these approximants. The three approximants show two poles,
one at a mass $m\simeq 0.75\Delta$ whose position is very stable among the approximants, the other
at a mass $M\simeq 1.35-1.4\Delta$ whose position depends slightly on the approximants.}
\label{fig:padecompared}
\end{figure}
To check the validity of this method, we vary the parameters $N$, $\omega_{\rm min}$ and
$\omega_{\rm max}$ and compare about 20 different approximants, see Fig. \ref{fig:padecompared}.
While they all (almost) coincide for real values of $p$, they vary a lot more when analytically continued,
a signature of the fragility of the Pad\'e procedure with respect to numerical errors.
All approximants show
a remarkable agreement for the pole at $\omega/\Delta$ close to 1,
corresponding to the mass $m$, that is, the inverse correlation length of the system,
see Fig. \ref{fig:padecompared} where all curves are superimposed at this pole. Among the approximants, we eliminate those that exhibit
unphysical spurious behavior, such
as an additional pole at an energy $\omega \ll \Delta$, or a splitting of the mass pole
into two peaks of energy around $m$. Furthermore, we eliminate approximants which
present a mass more than $1\%$ different from the others. Depending on the dimension, between one fourth ($d=3$) and
one half ($d=2$) of the Pad\'e approximants are rejected this way.
We observe that all the remaining approximants present a single second pole
at an energy $M>m$, the value of $M$ varying slightly from approximant to approximant (from $2\%$ in $d=3$ to less than $10\%$ for $d\lesssim 2.6$). Depending on the dimension, we find two possibilities.
In the first case, $M \gtrsim 2m$, and the pole corresponds to two independent single-particle excitations
implying that there are no bound states. In the second case $m<M<2m$ and a bound state exists with mass $M$.
As an additional check of the accuracy of the analytic continuation, we have verified that
the position of the poles varies smoothly with the dimensionality of the system.
\section{Results}
\label{results}
\begin{figure}[t]
\includegraphics{massratio.pdf}
\caption{(Color online) Values of the mass ratio $M/m$ for several values of the reduced temperature,
corresponding to different values of $r_0-r_{0_c}$ measuring the distance to criticality.
For each temperature, the error bar indicates the extremal possible values obtained from the Pad\'e approximants.}
\label{fig:massratio}
\end{figure}
Let us start by discussing the results obtained in $d=3$, where a bound state is clearly present in the broken symmetry phase, and absent in the symmetric phase. The corresponding values of $M/m$ are displayed in Fig. \ref{fig:massratio} as a function of $r_0-r_{0_c}$, where $r_{0_c}$ is the value of the parameter $r_0$ which makes the model critical. For a given value of the reduced temperature (which we identify with $r_0-r_{0_c}$), the value of the ratio $M/m$ varies slightly between different approximants, which is origin of the error bars shown in the figure. To test the accuracy of the method, we have also studied the variation of the results with the parameter $\alpha$ in front of the the regulator function (\ref{regulator}). In all cases this variation turns out to be much smaller than the error bars stemming from the Pad\'e procedure.
Both in the universal regime $r_0 \simeq r_{0_c}$, as well as for larger values of the reduced temperature within the non-universal regime, the ratio does not appear to vary significantly with the reduced temperature. Using a conservative error bar, we find $M/m=1.82(2)$, in agreement with previous results: $1.83(3)$ for Monte Carlo \cite{hasen99}, $1.828(3)$ for the first order approximation of the Bethe-Salpeter equation \cite{hasen02}, $1.84(3)$ for the results of perturbative continuous unitary transformations, and $1.84(1)$ for the most recent and accurate results from numerical diagonalization methods \cite{ponja}.
\begin{figure}[t]
\includegraphics{dimrapp.pdf}
\caption{(Color online) Values of the mass ratio $M/m$ in the critical regime for various
dimensions. For each dimension, the error bar indicates the extremal possible values obtained
from the Pad\'e approximants. The shaded areas denotes the range of dimensions for which it
is certain there are no bound states.}
\label{fig:ratiodim}
\end{figure}
Next, we study the evolution of value of the $M/m$ ratio at criticality as a function of the
dimension, for $2\leq d\leq 4$. The ratio is a smooth function of the dimension, as
shown in Fig. \ref{fig:ratiodim}. It is found that there exists an upper and lower
dimension, $d_r^-\sim2.2-2.3$ and $d_r^+\sim 3.2-3.3$, such that for $d_r^-<d<d_r^+$
there is a bound state, whereas for dimensions outside this interval, there is none.
This is consistent with the fact that there are no bound states in $d=2$ in the
absence of a magnetic field in the critical regime \cite{mccoy}, although they might still be present deeper in the broken symmetry regime \cite{mussardo1,mussardo2,rychkov16}. Furthermore, our results show that no bound state is to be expected in dimension $d=4$.
We have also studied the O$(2)$-symmetric model in $d=3$ along the same lines. Our results show the absence
of a bound state in this case.
\section{Conclusions}
\label{conclusion}
In this work we studied the existence of a bound state in the $\phi^4$ scalar theory in all dimensions between $d=2$ and $d=4$, and for a range of temperatures below the critical point. For $d=3$, our results are within $1\%$ of the previous Monte Carlo and numerical diagonalization values. We use the BMW approximation of the Non-Perturbative Renormalization Group, which allows for the determination of the full-momentum dependence of the spectral function both in the universal and nonuniversal regimes. These results show once again the power of the BMW approximation for dealing with non-trivial physics at arbitrary momentum scales, even in cases where the quantities of interest require to perform analytic continuations of numerical data.
Many generalizations of the present work can be envisaged. First, as the NPRG allows for the computation of nonuniversal quantities, studying
the presence of bound states for lattice models is a priori possible, since this only requires to take into account the lattice dispersion relation
as was already done for the derivative expansion \cite{tristan,adam11}. Second, the dependence of the bound state spectrum on an external magnetic
field can be naturally studied within our formalism, since the $\rho$-dependence of $\Gamma^{(2)}(p,\rho) $ encodes the influence of the external field on the spectral function. Finally, a BMW-type of approximation can also be used to detect bound states in more
complex systems, such as out-of-equilibrium \cite{kpz}, disordered \cite{tarjus, tissier} and quantum systems \cite{felix,adam14}, for which much
less is known via simulations.
\begin{acknowledgments}
We would like to thank N. Dupuis, A. Ran\c{c}on, J. Vidal and N. Wschebor for fruitful discussions.
F.B. wishes to thank the LPTMC for its hospitality during part of this work.
\end{acknowledgments}
|
\section{0pt}{8pt plus 1pt minus 1pt}{4pt plus 2pt minus 2pt}
\begin{document}
\title{Cascading DoS Attacks on IEEE 802.11 Networks}
\author{Liangxiao Xin, David Starobinski, and Guevara Noubir
\thanks{L. Xin, and D. Starobinski are with the Division of Systems Engineering, Boston University, Boston, MA 02215 USA (e-mail: <EMAIL>;
<EMAIL>).}
\thanks{G. Noubir is with the College of Computer and Information Science, Northeastern University, Boston, MA 02115 USA (e-mail:
<EMAIL>).}
}
\maketitle
\begin{abstract}
We unveil the existence of a vulnerability in Wi-Fi (802.11) networks, which allows an adversary to remotely launch a Denial-of-Service (DoS)
attack that propagates both in time and space. This vulnerability stems from a coupling effect induced by hidden nodes. Cascading DoS attacks can
congest an entire network and do not require the adversary to violate any protocol. We demonstrate the feasibility of such attacks through
experiments with real Wi-Fi cards, extensive ns-3 simulations, and theoretical analysis. The simulations show that the attack is effective both in
networks operating under fixed and varying bit rates, as well as ad hoc and infrastructure modes. To gain insight into the root-causes of the
attack, we model the network as a dynamical system and analyze its limiting behavior and stability. The model predicts that a phase transition
(and hence a cascading attack) is possible when the retry limit parameter of Wi-Fi is greater or equal to 7, and characterizes the phase
transition region in terms of the system parameters.
\end{abstract}
\IEEEpeerreviewmaketitle
\section{Introduction}
\label{Introduction}
Wi-Fi (IEEE 802.11) is a
technology widely used to access the Internet. Wi-Fi connectivity is provided by
a variety of organizations operating over a shared RF spectrum. These
include schools, libraries, companies, towns and governments, as well
as ISP hotspots and residential wireless routers.
Wi-Fi traffic is also rapidly rising due to increased
offloading by cellular operators~\cite{lee2010mobile}.
The importance of Wi-Fi networks and the need to strengthen their resilience to intentional and
non-intentional interference have been recognized by companies, such as
Cisco~\cite{Cisco}.
Wi-Fi networks rely on simple, distributed mechanisms to arbitrate access to the shared spectrum and optimize performance. Such mechanisms include
carrier sensing multiple access (CSMA), exponential back-offs, and bit rate adaptation. The behavior of these mechanisms in isolated single-hop
networks has been extensively studied and is generally well-understood (see, e.g.,~\cite{bianchi}). However, due to interference coupling, these
mechanisms result in complex interactions in multi-hop settings. As a consequence, different networks do not always evolve independently, even if
they are located far away.
Figure~\ref{general_case} serves to illustrate this phenomenon at a high level. Suppose that an attacker increases the rate at which it generates
packets, and transmits these packets in accordance with the IEEE 802.11 protocol.
These transmissions may cause packet collisions at nodes concurrently receiving packets from other sources. Due to the infamous hidden node problem,
which is hard to avoid in wireless networks, transmitters may be unable to hear transmission by other nodes, even when using CSMA, and hence keep
retransmitting packets until they reach the so-called retry limit of the back-off procedure. These retransmissions affect other neighbours and may
propagate.
While an optional mechanism, called RTS/CTS, has been designed to combat the hidden node problem,
it increases overhead and latency especially
at high bit rates. Since the cost of the RTS/CTS exchange usually does not
justify its benefits, it is commonly
disabled~\cite{forouzan2004data, gast2005802}. Indeed, most manufacturers of Wi-Fi cards disable RTS/CTS by default and discourage changing this
setting as explicitly stated in~\cite{netgear,tp-link,linksys,d-link}. Therefore, most Wi-Fi systems today operate without RTS/CTS.
The coupling phenomenon induced by interferences
creates multi-hop dependencies, which an adversary can take advantage of to launch a widespread network attack from a single location. We refer to
such an attack as a \emph{cascading Denial-of-Service (DoS) attack}. Cascading DoS attacks are especially dangerous because they affect the entire
network and do not require the adversary to violate any protocol (i.e., the attacks are protocol-compliant).
The contributions of this paper are as follows. First, we unveil the existence of a vulnerability in the IEEE 802.11 standard,
which allows an attacker to launch protocol-compliant cascading DoS attacks. In contrast to
existing jamming attacks, the attacker does not need to be in the vicinity of the victims.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{general_case_v2}
\caption{Illustration of a cascading denial of service attack. Transmissions by an attacker impact nodes located far away, due to interference
coupling caused by hidden nodes.}
\label{general_case}
\end{figure}
Second, we provide a concrete attack that exploits this vulnerability in certain network scenarios.
We demonstrate the attack through experiments on a testbed composed of nodes equipped with real Wi-Fi cards, and through extensive ns-3
simulations.
Third, we show the existence of a \emph{phase transition}. When the packet generation rate of the attacker is
lower than the phase transition point, it has vanishing effect on the rest of the
network. However, once the packet generation rate exceeds the phase transition point, the network becomes entirely congested.
Thus, under a phase transition, the utilization of a remote node experiences no change until it is suddenly forced to
congestion~\cite{saligrama2006macroscopic}.
Finally, we introduce a new analytical model that sheds light into the phase transition observed in the simulations and experiments. We apply fixed point theorems to this model. The analysis predicts for which values of the retry limit a phase transition (and hence a cascading attack)
can occur, and explicitly characterizes the phase transition region in terms of the system parameters. In particular, we show that a phase
transition can occur for the default value of the retry limit in Wi-Fi, which is~7. We carry out a stability analysis and demonstrate that in the
phase transition region the system must have multiple fixed points, one of which being unstable.
The rest of the paper is organized as follows. In Section~\ref{Related_Work}, we discuss related work.
In Section~\ref{Background}, we provide brief background on Wi-Fi, hidden nodes, and Minstrel, and introduce our network model.
We present and discuss experimental and simulation
results in Section~\ref{Simulations and Experiment}. In
Section~\ref{Analysis}, we present an analytical model that explains the behaviour of the network and the impact of various parameters, and compare
the analytical and simulation results. In Section~\ref{Conclusion}, we conclude the paper and discuss possible mitigation methods.
An earlier and shorter version of this paper appeared in the proceedings of the IEEE Conference on Communications and Network Security (CNS 2016)~\cite{cns2016}. This journal version significantly expands the theoretical analysis, including detailed proofs of all the lemmas and theorems, and
new results on stability analysis and heterogeneous traffic load, all of which can be found in Section~\ref{Analysis}. Moreover, new simulation results for infrastructure networks, networks supporting RTS/CTS, ring networks,
networks based on a realistic indoor building model,
and networks with heterogeneous traffic load are presented in Sections~\ref{Simulations} and~\ref{Mitigation}.
\section{Related Work}
\label{Related_Work}
In general, the main goal of a DoS attack is to make communication impossible
for legitimate users. Within the context of wireless networks, a simple and popular means to
launch a DoS attack is to jam the network with high power transmissions of random bits, hence creating
interferences and congestion. Jamming at the physical layer, together with \emph{anti-jamming} countermeasures, have been extensively studied
(cf.~\cite{poisel2011modern} for a monograph on this subject).
More recently, several works have developed and demonstrated \textit{smart jamming} attacks. These attacks exploit protocol
vulnerabilities across various layers in the stack to achieve high jamming gain and energy
efficiency, and a low probability of detection~\cite{pelechrinis2011denial}.
For instance, \cite{lin2005link} shows that the energy consumption of a smart
jamming attack can be four orders of magnitude lower than continuous
jamming. The works in~\cite{noubir2011robustness,
orakcal2014jamming} show that several Wi-Fi bit rate adaptation algorithms, such as
SampleRate, ONOE, AMRR, and RARF, are vulnerable
to smart jamming. However, both conventional and smart jamming attacks are usually non-protocol
compliant. Moreover, they require physical proximity. These limitations can be
used to identify and locate the jammer.
In contrast, in this work we show how a protocol-compliant DoS attack
can be remotely launched by exploiting coupling due to hidden nodes in
Wi-Fi. Rate adaptation algorithms further amplify this attack due to
their inability to distinguish between collisions, interferences, and
poor channels. One potential mitigation is to design a rate
adaptation algorithm whose behaviour is based on the observed interference patterns~\cite{chen2007rate, rayanchu2008diagnosing}. However, to the
best of
our knowledge, none of these rate adaptation algorithms are used in
practice. Our work is based on Minstrel~\cite{Minstrel}, which is the
most recent, popular, and robust rate adaptation algorithm for Linux systems.
The attacks that we are investigating bear similarity to cascading
failures in power transmission systems~\cite{kinney2005modeling, soltan2014cascading}. When one of the nodes
in the system fails, it shifts its load to adjacent nodes. These nodes in
turn can be overloaded and shift their load further. This phenomenon
has also been studied in wireless networks. For
instance,~\cite{haenggi2009stochastic, kong2009wireless} model
wireless networks as a random geometric graph topology generated by a
Poisson point process. They use percolation theory to show that the
redistribution of load induces a phase transition in the
network connectivity. However, the cascading phenomenon that we investigate
in this paper is different from cascading failure studied in those works. In our work, the exogenous generation of
traffic at each node is independent. That is, a node will
not shift its load to other nodes. The amount of traffic measured on the channel increases
due to packet retransmissions caused by packet collisions, rather
than due to traffic redistribution.
The work in~\cite{aziz2009ez,aziz2011understanding} show that interference coupling can affect the stability of multi-hop networks. In the case of a greedy
source, a three-hop network is stable while a four-hop network becomes unstable.
In contrast, in our work, the path of each packet consists of a single-hop. Thus, network instability is not due to multi-hop communication in our case.
The work in~\cite{ray2005performance,saligrama2006macroscopic} show that local coupling due to interferences
can have global effects on wireless networks. Thus, \cite{ray2005performance} proposes
a queuing-theoretic analysis and approximation to predict the probability of a packet
collision in a multi-hop network with hidden nodes. It shows that the sequence of the
packet collision probabilities in a linear network converges to a
fixed point. The work in~\cite{saligrama2006macroscopic} evaluates the impact of rate adaption and finds out that traffic increase
at a single node can congest an entire network, and points out the existence of a phase transition.
Our paper differs in several aspects. First, it considers an adversarial context, and shows how interference-induced coupling can be exploited to
cause denial of service. Second, to our knowledge, it is the first work to demonstrate the existence of such coupling on real commodity hardware.
Third, our simulations are based on a high-fidelity wireless simulator (ns-3), capable of capturing the effects of rate adaptation algorithms and
accurately modeling infrastructure networks. Finally, our analytical model is original and captures the impact of the retry limit and traffic
parameters. A key result is that a cascading attack can be launched for the default value of the retry limit in Wi-Fi, a result validated by the
experiments and simulations.
\section{Background and Model}
\label{Background}
We first review key aspects of IEEE 802.11 and then describe the network model under consideration.
\subsection{Wi-Fi Summary}
\label{Wi-Fi Summary}
Wi-Fi is a wireless local area network (WLAN) technology, which mainly
runs on 2.4~GHz ISM bands and 5~GHz bands~\cite{gast2005802}. The IEEE 802.11 standard is a
series of specifications, such as the media access control (MAC) and
physical layer (PHY) interfaces. The first 802.11 standard
that gained widespread success is 802.11b.
It runs on 2.4~GHz bands and
has up to 11 Mb/s bit rate.
The subsequent standards (e.g., 802.11a, g, n, and
ac) increased the bit rates using higher order modulation along with
coding, OFDM, MIMO, and wider bands.
It is noteworthy
that 802.11b is the only mode that supports communication at 1~Mb/s. Hence, when the
bit rate reduces to 1 Mb/s, Wi-Fi network reverts to the
802.11b mode. Generally, this lower bit rate has higher resistance to
interference during transmission and is able to operate over lower SNR channels.
The IEEE 802.11 standard uses a CSMA/CA mechanism to control access to
the transmission medium and avoid collisions.
After a packet is sent, a
node waits for a short interframe slots (SIFS) period to receive an
ACK. Whenever the channel becomes idle, the node waits for a distributed
interframe space (DIFS $>$ SIFS) period and a random backoff before contending for the channel.
The random backoff consists of a random number of backoff slots, which depends on the so-called contention window.
Specifically, at the $r \geq 1$ retransmission attempt (retry count), the contention window $CW_r$ is given by
\begin{eqnarray}
CW_r = \left\lbrace
\begin{array}{ll}
2^{r-1} (CW_1 + 1) - 1 & CW_r < CW_{max}, \\
CW_{max} & \text{otherwise}.
\end{array}
\right.\label{eq:back-off}
\end{eqnarray}
The number of backoff slots is chosen uniformly at random in the interval $[0, CW_r]$.
For IEEE 802.11b, the initial contention window size is $CW_1 = 31$ , the maximum contention window size is $CW_{max} = 1023$, and the duration of a
backoff slot is $20~\mu s$.
Note that the case $r=1$ corresponds to the initial packet transmission attempt.
\subsection{Hidden Node Problem}
\label{Hidden Node Problem}
A typical instance of the hidden node problem is illustrated in
Figure~\ref{classic_hidden_node}. The figure shows three nodes: a transmitter,
a receiver and a hidden node. The dashed circle represents the
transmission range of the node. Since the transmitter and the hidden
node cannot sense each other, a collision happens when both of them
transmit packets at the same time.
A packet collision triggers a
retransmission. In IEEE 802.11, there is an upper limit on the number
of retransmissions that a packet can incur, called \textit{retry limit} and
denoted by $R$ (the default value is $R=7$). If the retry count $r$ of a packet exceeds the retry
limit, the packet is dropped, the retry count is reset to $r=1$, and a new packet transmission can start. The channel utilization of a node increases with the
probability of a packet collision. In the worst case,
the utilization can be~$R$ times larger than in the absence of packet collisions. Therefore, the access channel of
a node can easily be saturated if it is forced to retransmit packets.
The hidden node problem can in principle be avoided by enabling the RTS/CTS exchange,
which is implemented in Wi-Fi networks. However, the RTS/CTS exchange has not only high overhead, but also does not always fully prevent packet
collisions~\cite{ray2005evaluation}
and may lead to deadlocks in multi-hop configurations~\cite{ray2007false}.
Generally, it is either turned off
\cite{bellardo2003802} or only used for packets whose length exceeds the so-called RTS
threshold. Most manufacturers of Wi-Fi cards, including Netgear~\cite{netgear}, TP-LINK~\cite{tp-link}, Linksys~\cite{linksys} and
D-Link~\cite{d-link}, disable RTS/CTS altogether by setting the RTS
threshold to a sufficiently high default value (e.g., 2346~bytes, which corresponds to the maximum length of an IEEE~802.11 frame).
They furthermore recommend to not change the default setting.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{classic_hidden_node}
\caption{Classical hidden node problem. The transmitter and the hidden
node cannot sense each other. The collision happens when they
transmit simultaneously.}
\label{classic_hidden_node}
\end{figure}
\subsection{Minstrel Rate Adaptation}
\label{Minstrel Rate Adaptation}
Minstrel is a practical, state-of-the-art rate adaptation algorithm that has been
implemented within the MadWiFi project and Linux mac80211
driver framework~\cite{Minstrel}. It chooses the bit rate of a
transmission based on the throughput measured over past transmissions at different rates. Technically, it
selects a bit rate following a retry chain, as shown in
Table~\ref{Minstrel retry chain}.
In Minstrel, 90\% of the packets are
transmitted at a ``normal rate'' (fourth column in Table~\ref{Minstrel retry chain}). The remaining 10\% are
transmitted at a ``lookaround rate'' (second and third columns in Table~\ref{Minstrel retry chain}). Each packet is transmitted at a
rate following a retry chain (rows in Table~\ref{Minstrel retry
chain}). For example, consider a packet being transmitted at
``lookaround rate''. If a random rate is lower than the rate with ``best
throughput'', the packet is first transmitted at the
``best throughput'' rate, then at the ``random rate'', then at the ``best probability'' rate,
and finally at the ``lowest baserate''. The
packet is dropped if the transmission fails at the ``lowest
baserate''. The retry chain table is updated 10 times every second
based on performance statistics.
Therefore, a large amount of packet loss does not
necessarily cause Minstrel to switch to a low bit rate. Another advantage of
Minstrel is that it probes the throughput of different bit rates
randomly. This makes the rate adaptation more robust in complicated
environment and against some adversaries.
\begin{table}
\small
\centering
\caption{Minstrel Retry Chain \cite{Minstrel}}
\begin{tabular}{|c|c|c|c|} \hline \label{Minstrel retry chain}
Try & \multicolumn{2}{|c|}{Lookaround rate} & Normal rate \\
\cline{2-3}
& random $<$ best\footnote{The random rate is lower than the best throughput rate.} & random $>$ best & \\
\hline
1 & Best throughput & Random rate & Best throughput \\
\hline
2 & Random rate & Best throughput & 2nd best throughput \\
\hline
3 & Best probability\footnote{This rate has the highest probability of resulting in a successful transmission.} & Best probability & Best
probability \\
\hline
4 & Lowest baserate & Lowest baserate & Lowest baserate \\
\hline
\end{tabular}
\end{table}
\subsection{Network Model}
\label{Network Model}
The network model considered in this paper is shown in Figure~\ref{linear_topology}. This configuration could arise over different time and space
in
more complex network topologies. We consider
$N+1$ pairs of nodes. Each node $A_i$ ($i=0, 1, 2, \ldots$, $N$)
transmits packets to node $B_i$. The dashed circle represents the
range of transmission. Node $B_{i+1}$ can receive packets from both node
$A_i$ and node $A_{i+1}$. However, node $A_i$ and node $A_{i+1}$
cannot hear each other. That is, node $A_i$ is a hidden node with
respect to node $A_{i+1}$ (and vice-versa). A packet collision happens at node
$B_{i+1}$ when packet transmissions by node $A_i$ and $A_{i+1}$ overlap.
We assume that all the nodes communicate over the same channel. Note that there are only three non-overlapping channels in the 2.4GHz band.
Hence, it is common that several nodes use the same channel over time and space in crowded areas.
\subsection{Cascading DoS attack}
Our goal is to investigate how node $A_0$ can trigger a cascading DoS
attack, resulting in a congestion collapse over the entire network. We start
by increasing the packet generation rate at node
$A_0$. Node $A_0$ transmits packets over its channel, in compliance with the IEEE 802.11 standard. The transmissions by
node $A_0$ cause packet collisions at
node $B_1$. These collisions require node $A_1$ to retransmit packets. The increased amount of packet transmissions and retransmissions by node
$A_1$ impact node $A_2$ and so forth. If this effect keeps propagating and amplifying, then the result is a network-wide denial of service, which
we refer to as a cascading Denial of Service (DoS) attack. Because this attack is protocol-compliant, it is difficult to detect or trace back to the
initiator.
We note here that as a hidden node retransmits its packets, it must back off after each retransmission, which leaves the channel idle for a certain
period of time. However, the duration of the backoff period is generally too short to allow for a successful transmission.
Indeed, a packet transmission is successful only if
\begin{enumerate}
\item The size of the contention window of the hidden node is longer than the packet transmission time.
\item The transmitter starts and ends its transmission entirely during the backoff period of the hidden node.
\end{enumerate}
At 1~Mb/s, the transmission time of an 1500~bytes packet lasts 12~ms. This is longer than
the contention window as long as $CW_{r} < CW_{max} = 1023$. Hence, by Eq.~(\ref{eq:back-off}), a transmission cannot be successful during the
backoff period preceding the $r < 6$ retransmission attempt by a hidden node.
At the $r \geq 6$ retransmission attempt by a hidden node $A_i$, $CW_{r} = CW_{max} = 1023$. Node $A_i$ back-offs for $n$ slots, where $n$ is an integer between 0 and 1023 that is picked uniformly at random (i.e., with probability $1/1024$).
Since the length of a backoff slot is 20~$\mu$s, the backoff delay is $0.02n$~ms. Without loss of generality, assume that node~$A_i$ starts backing off at time $t=0$ and ends its backoff at time $t=0.02n$ (all the time units are in milliseconds). Node $A_i$ then starts a packet transmission, which ends at time $t=0.02n+0.12$.
Node $A_{i+1}$ can transmit a packet successfully only if it starts its transmission during the time interval $[0,0.02n-12]$. This requires $n>600$. Assuming that the starting time of the packet transmission by node $A_{i+1}$ is uniformly distributed in the time interval $[0,0.02n+12]$, the probability that a packet is successfully transmitted by node $A_{i+1}$ is \[\sum_{n=600}^{1023} \frac{1}{1024} \cdot \frac{0.02n-12}{0.02n+12} = 0.059.\]
Thus, the likelihood of a successful packet transmission is low, a result validated by the experimental and simulation results of the next section.
\begin{figure}[!t]
\centering
\includegraphics[width=3.2in]{linear_topology_new2}
\caption{Topology of the network. Node $A_i$ transmits packets to node $B_i$. Node $A_{i}$ is a hidden node with respect to $A_{i+1}$.}
\label{linear_topology}
\end{figure}
\section{Experimental and Simulation Results}
\label{Simulations and Experiment}
In this section, we demonstrate the feasibility of launching cascading DoS attacks both
through experiments and simulations. We first show results on an experimental testbed using real Wi-Fi cards.
We then use ns-3.22 simulations to investigate how this attack can be performed in significantly larger
scale networks, and under different settings (ad~hoc, infrastructure, fixed bit rate, and adaptive bit rate).
\subsection{Experiments}
\label{Experiment}
We set up an experimental testbed composed of six nodes. The testbed
configuration is shown in Figure~\ref{exp_setup}. We establish an
IEEE 802.11n ad hoc network consisting of three pairs of nodes. Each node
consists of a PC and a TP-LINK TL-WN722N Wireless USB Adapter. We use
RF cables and splitters to link the nodes, isolate them from
external traffic, and obtain reproducible results.
We place 70~dB attenuators on links between node $A_i$ and $B_i$ ($i \in 0,1,2$), and 60~dB attenuators
on links between nodes $A_i$ and $B_{i+1}$. The difference in the
signal attenuation of different links ensures that a packet loss occurs if a hidden node transmits. In practice, such a
situation may occur if nodes $A_i$ and $B_{i+1}$ communicate without obstacles, while node $A_i$ and $B_i$ are separated by an office
wall~\cite{stein1998indoor}.
The transmission power of each node is set to 0~dBm.
We use iPerf
\cite{IPERF} to generate UDP data streams and to measure the throughput achieved on
each node. The length of a packet is the default IP packet size of
1500 bytes.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{exp_setup_new2}
\caption{Experimental testbed.}
\label{exp_setup}
\end{figure}
Figure~\ref{implementation_result} demonstrates the cascading DoS
attack on the experimental testbed. At first, the packet generation
rates of nodes $A_0, A_1$ and $A_2$ are set to 400 Kb/s. We observe that the throughput of all the nodes
remains in the vicinity of 400 Kb/s during the first 300 seconds. After 300~seconds,
$A_0$ starts transmitting packets at 1 Mb/s. As a result, the throughput of nodes $A_1$ and $A_2$ suddenly
vanishes. Once node $A_0$ resumes transmitting at 400 Kb/s, the
throughput of node $A_1$ and node $A_2$ recovers.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{implementation_result_new2}
\caption{Throughput performance measurements in testbed. When node
$A_0$ starts increasing its packet generation rate, the throughput of nodes $A_1$
and $A_2$ vanishes.}
\label{implementation_result}
\end{figure}
\subsection{Simulations}
\label{Simulations}
In the previous section, we demonstrated the
feasibility of launching a cascading DoS attack on an experimental
testbed. This testbed relies on commercial cards that are
black boxes for all purposes. For instance, the driver of the Wi-Fi card
and the rate adaptation algorithm are closed-source. There are also
substantial usage restrictions, such as parameter settings.
In order to gain a better insight into the
attack in large-scale networks, we resort to ns-3 simulations, a state-of-the-art simulator which includes high-fidelity wireless libraries. We show
the occurrence of cascading DoS attacks
\begin{enumerate}
\item In ad hoc networks with fixed bit rate;
\item In ad hoc networks under Minstrel rate adaptation;
\item In infrastructure networks;
\item In ring topology networks;
\item In an indoor scenario;
\end{enumerate}
and the countering of cascading DoS attacks
\begin{enumerate}
\setcounter{enumi}{5}
\item In networks with RTS/CTS enabled.
\end{enumerate}
\begin{figure}[!t]
\centering
\subfloat[As the traffic load at node $A_0$ increases, the
utilization of remote nodes (e.g., $A_{20}$ and $A_{40}$) exhibits a
phase transition.]
{\includegraphics[width=3in]{one_node_phase_transaction_new3}
\label{one_node_phase_transaction}}
\vfil \subfloat[Utilization of nodes $A_i$ ($i \geq 1$) for
different traffic loads at node $A_0$. The utilization converges as $i$ gets large. When the load at node $A_0$ changes from $0.4$ to
$0.6$, the sequence of utilization converge to different limits, illustrating the phase transition.]
{\includegraphics[width=3in]{one_node_load_phase_transaction_new3}
\label{one_node_load_phase_transaction}}
\caption{Occurrence of cascading DoS attacks in ad hoc networks with fixed bit rate.}
\label{The occurrence of cascading DoS attack in adhoc networks with fixed bitrate}
\end{figure}
\subsubsection{Fixed bit rate}
\label{EXP:Fixed bitrate}
We first describe the occurrence of a cascading DoS attack in an
ad~hoc network with fixed bit rate. We consider a linear topology
consisting of 41 pairs of nodes (i.e. a sequence of 41 hidden nodes), as shown in Figure~\ref{linear_topology}.
Each packet is transmitted over a single-hop path (similar to Wi-Fi Direct).
We fix the bit rate to 1 Mb/s and the retry limit to $R=7$.
We set up a Wi-Fi network using the standard IEEE 802.11 library in ns-3. At each node $A_i$, $i \geq 1$,
the generation rate of UDP packets is $\lambda_i=8.125$ pkts/s. The
generation rate of UDP packets at node $A_0$, $\lambda_0$, varies from $1.25$ to $61.25$
pkts/s. Packets at each node are generated according to a Poisson process, hence different nodes start transmitting at different times.
The size
of each packet is 2000 bytes. Each node has the same transmission power
(40~mW).
We set the propagation loss between node $A_i$ and $B_i$ to
80~dB and the propagation loss between node $A_i$ and $B_{i+1}$
to 70~dB.
We run each simulation five times for 1,000 seconds, and average out the results.
The \emph{(exogenous) load} at each node $A_i$ is denoted $\rho_i=\lambda_i T$, where $T$ represents the duration of each packet transmission
attempt (0.016 second in our case).
The \emph{utilization} of a node $A_i$, denoted $u_i$, is defined as the
fraction of time the node is busy transmitting bits on the channel.
Figure~\ref{The occurrence of cascading DoS attack in adhoc networks with fixed bitrate}(a) depicts the utilization
$u_1$, $u_{20}$, and $u_{40}$ as a function of
$\rho_0$, the load at node $A_0$. The utilization of node~$A_1$, $u_1$, increases smoothly until it reaches its
upper limit. However, the
utilizations of nodes $A_{20}$ and $A_{40}$ remain low until $u_0$ reaches a certain
threshold around $\rho_0 = 0.5$, at which point $u_{20}$ and $u_{40}$ suddenly jump
to a high value. This sudden jump corresponds to a phase transition, and the critical threshold represents the phase transition point.
Figure~\ref{The occurrence of cascading DoS attack in adhoc networks with fixed bitrate}(b) illustrates the phase transition in a different way.
The figure depicts the
utilization of each node $A_i$ for $i \geq 1$, as $i$ increases. Again, we observe that different values of
$\rho_0$ lead to two completely distinct behaviour for the sequence of
utilizations $(u_i)_{i=0}^{40}$ (i.e., $u_{40} \simeq 0.3$ when $\rho_0 = 0.2$ and $\rho_0 = 0.4$, while
$u_{40} \simeq 0.75$ when $\rho_0 = 0.6$ and $\rho_0 = 0.8$).
Note that the upper limit of the utilization does not reach~1,
due to inter-frame spacing requirements and (random) backoff delays mandated by IEEE 802.11.
\begin{figure}[!t]
\centering
\subfloat[Throughput]{\includegraphics[width=2.5in]{Minstrel-throughput-adhoc-modified}
\label{Minstrel_PHYRateMode_default_onoff_throughput}}
\vfil
\subfloat[Bit rate]{\includegraphics[width=2.5in]{Minstrel-bitrate-adhoc-modified}
\label{Minstrel_PHYRateMode_default}}
\caption{Simulation results with Minstrel rate adaptation. When node
$A_0$ generates packets at 5~Mb/s and transmits, the throughput of nodes $A_{20}$ and $A_{40}$
vanishes. The average bit rates of nodes $A_{20}$ and $A_{40}$ also
reduce to 1 Mb/s. This result indicates that nodes $A_{20}$ and
$A_{40}$ are transmitting packets at the lowest bit rate,
however with no throughput (all their packets collide).}
\label{Simulation Result with Minstrel Rate Adaptation}
\end{figure}
\subsubsection{Rate Adaptation}
\label{Rate Adaption}
We next consider the same network setting as in the previous section,
but this time we assume that nodes can transmit at different bit rates.
We specifically assume that nodes implement the Minstrel rate adaptation algorithm.
In this case, the attack works by coercing the rate adaptation algorithm to reduce the bit rate
to 1 Mb/s at each node, thus leading to similar results to those shown in Section
\ref{EXP:Fixed bitrate}. In our simulations, the parameter $\textit{EWMA}$ of Minstrel is set to 0.25 \cite{xia2013evaluation}.
We set $\lambda_0 =312.5$ pkts/s and
$\lambda_i = 31.25$ pkts/s ($i \geq 1$) for the packet generation rates. As shown in Figure \ref{Simulation Result with Minstrel Rate Adaptation},
packet transmissions at
node $A_0$ start after $t=100$~s.
During the first 100 seconds, the throughput of nodes $A_{20}$ and
$A_{40}$ remain around 0.5 Mb/s, which implies that all the packets are
received.
Once node $A_0$ starts transmitting packets,
the throughput of nodes $A_{20}$ and $A_{40}$ is brought down to close
to zero. We also observe that the bit rates at node $A_{20}$ and $A_{40}$ go down to 1 Mb/s,
due to the repeated packet collisions. Once node
$A_0$ stops transmitting at $t=700$~s, nodes $A_{20}$ and $A_{40}$ recover.
\begin{figure}[!t]
\centering
\subfloat[When
node $A_0$ generates packets at 5~Mb/s and transmits, the throughput of nodes $A_{20}$ and $A_{40}$
vanishes.]{\includegraphics[width=2.5in]{Minstrel-throughput-AP-modified}
\label{Simulation Result under AP mode}}
\vfil
\subfloat[When
node $A_{20}$ generates packets at 5~Mb/s and transmits, the throughput of node $A_{40}$
vanishes while the throughput of node $A_0$ does not.]{\includegraphics[width=2.5in]{Minstrel-throughput-AP-middle-modified}
\label{Simulation Result under AP mode start at middle}}
\caption{Simulation results under AP mode without reassociation. Nodes $A_i$ are stations and
nodes $B_i$ are access points, for $i \in \{0,1,2,\dots\}$.}
\label{Simulation Result under AP}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{Minstrel-throughput-AP-reassociation-modified.eps}
\caption{Simulation results under AP mode with reassociation. When
node $A_{0}$ generates packets at 5~Mb/s and transmits, the throughput of node $A_{20}$ and $A_{40}$ significantly decreases.}
\label{Simulation Result under AP mode with reassociation}
\end{figure}
\subsubsection{Infrastructure networks}
We next show that cascading DoS attacks are also feasible in
infrastructure networks. Since the infrastructure mode is more
widely used than ad hoc in practice, the feasibility of the cascading
DoS attack in infrastructure networks increases its severity and
potential impact. We repeat the simulations of Section~\ref{Rate
Adaption} except that we set nodes $B_i$ as access points, and
nodes $A_i$ as stations. The initial beacon starting time at each AP is a random variable that is uniformly distributed between 0 and $102.4$~ms.
We first investigate the cases where stations do not restart association when beacons are missing.
Toward this end, we set the number of consecutive beacons that must be missed before restarting association, i.e. the attribute {\tt MaxMissBeacons} in ns-3, to a large value.
Otherwise, we use the default settings of ns-3 for the APs \cite{ns3ap} and the stations \cite{ns3sta}. Figure~\ref{Simulation Result under
AP} shows similar results as
in Section~\ref{Rate Adaption}, namely when a cascading DoS attack is launched by node $A_0$, as shown in Figure~\ref{Simulation Result under AP}(a),
the remote nodes $A_{20}$ and $A_{40}$ in the sequence exhibit a phase transition.
If the attacker is node $A_{20}$, the simulation result in Figure~\ref{Simulation Result under AP}(b) shows that the throughput of node
$A_{40}$ vanishes but the throughput of node $A_0$ does not. This result shows that an attack can be launched from any node $A_i$ in the topology
and the following nodes in the sequence (i.e., $A_{i+1}, A_{i+2}, \ldots$) will experience congestion.
We next consider the case where stations restart association when beacons are missing. We set
${\tt MaxMissBeacons} = 10$, which is the default value in ns-3 \cite{ns3sta}. The simulation results are shown in Figure~\ref{Simulation Result under AP mode with reassociation}. When Node $A_0$ starts to transmit packets, we observe a significant throughput degradation at nodes $A_{20}$ and $A_{40}$, but
the throughput does not vanish completely. The reason is that if $A_i$ disassociates from its AP $B_i$ over a certain period
then node $A_{i+1}$ is not affected by interference coupling during that period. This result indicates that reassociations help mitigate cascading DoS attacks, though throughput performance is still significantly impaired.
\subsubsection{Ring topology}
\label{Ring topology}
We investigate cascading DoS attacks in a ring topology with 41 pairs of nodes, as shown in Figure~\ref{ring_topology}. In our previous results for
linear topologies, the effect of an attack disappears once the attacker reduces its packet generation rate. However, the effect of an attack in a
ring topology can last for a long period of time after the attack stops.
Node $A_i$ $(i = 0, 1, \dots)$ generate packets at rate 0.5~Mb/s, following a Poisson process. At time $t=300$ s, node $A_0$ increases its packet generation rate to 11~Mb/s and the
throughput of all the nodes vanishes. Yet, unlike results in linear topologies, the throughput of the nodes does not recover after node $A_0$
reduces its packet generation rate back to 0.5~Mb/s. The cyclic nature of the topology reinforces the attack even after the trigger stops.
This result is illustrated in Figure~\ref{Simulation Result with Minstrel Rate Adaptation circle}.
During the first 100 seconds, all the nodes $A_i$ $(i = 0, 1, \dots)$ generate packets at 0.5~Mb/s. At time $t=300$~s, node $A_0$ increases its packet generation
rate to 11~Mb/s. As a result, the throughput of all nodes vanishes. Yet, unlike results in linear topologies, the throughput of the nodes does not
recover
after node $A_0$ reduces its packet generation rate back to 0.5~Mb/s. The cyclic nature of the topology reinforces the attack even after the trigger
stops.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{ring_topology.eps}
\caption{Ring topology under cascading DoS attack. The dash circle represents the transmission range of the transmitter.}
\label{ring_topology}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{Minstrel-throughput-adhoc-circle-modified.eps}
\caption{Simulation results under a ring topology. When the packet generation rate of node
$A_0$ increases, the throughput of nodes $A_{20}$ and $A_{40}$
vanishes. This effect continues even when the packet generation rate of node $A_0$ decreases.}
\label{Simulation Result with Minstrel Rate Adaptation circle}
\end{figure}
\subsubsection{Building model} \label{Building model}
In this section, we use the ns-3 {\tt HybridBuildingsPropagationLossModel} library~\cite{ns3hybridbuildingspropagationlossmodel} to demonstrate the feasibility of cascading DoS attacks in an indoor scenario. Models in this library realistically characterize the propagation loss across different spectrum bands (i.e., ranging from 200~MHz to 2.6~GHz), different environments (i.e., urban, suburban, open areas), and different node positions with respect to buildings (i.e., indoor, outdoor and hybrid). The building models take into account the penetration losses of the walls and floors, based on the type of buildings (i.e., residential, office, and commercial).
In our simulations, we consider a 20-floor office building with six rooms in each floor, as shown in Figure~\ref{building_model}. We assume that five pairs of Wi-Fi nodes $(A_i, B_i)$ are active in the building, where node $A_i$ transmits packets to nodes $B_i$ ($i=0,1,2,3,4$). The bit rate is set to 1~Mb/s, the retry limit to $R=7$, and the frequency to 2.4 GHz. The generation rate of UDP packets at nodes $A_i$, $i \geq 1$, is $\lambda_i = 8.125$ pkts/s. Packets are 2000~bytes long.
We turn on and off transmissions at node $A_0$ to observe how it impacts the throughput of other nodes.
Simulation results are shown in Figure~\ref{CDoS-1Mbps-adhoc-UDP-building-utilization}. When node $A_0$ does not transmit, the throughput of node $A_4$ is 0.13~Mb/s and it incurs no packet loss. However, when node $A_0$ starts transmitting, the throughput of node $A_4$ collapses. The throughput of node $A_4$ recovers only after node $A_0$ stops transmitting.
\begin{figure}[!t]
\centering
\subfloat[Top view.]{\includegraphics[width=2.5in]{building_model_xy}
}
\vfil
\subfloat[Side view.]{\includegraphics[width=2.5in]{building_model_yz}
}
\caption{Office building model. The building has 20 floors ($z$-axis) and 6 rooms in each floor ($x$ and $y$ axes).}
\label{building_model}
\end{figure}
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{CDoS-1Mbps-adhoc-UDP-building-throughput.eps}
\caption{Simulation results using ns-3 building model. When node $A_0$ transmits, the throughput of remote node $A_4$ collapses.}
\label{CDoS-1Mbps-adhoc-UDP-building-utilization}
\end{figure}
\subsubsection{RTS/CTS}
We next evaluate the impact of enabling RTS/CTS in the topology under consideration. Specifically, we repeat the simulations of Section~\ref{Rate Adaption}, but with RTS/CTS enabled. Figure~\ref{Simulation Result with RTSCTS} shows that transmissions by node $A_0$, which start after $100$~s, have no effect on the throughput of remote nodes $A_{20}$ and $A_{40}$. This shows that RTS/CTS is an effective solution against cascading DoS attacks in this scenario.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{Minstrel-throughput-adhoc-RTSCTS-modified.eps}
\caption{Simulation results when enable RTS/CTS. The increase of the packet generation rate of node $A_0$ does not affect the throughput of nodes
$A_{20}$ and $A_{40}$.}
\label{Simulation Result with RTSCTS}
\end{figure}
\section{Analysis}
\label{Analysis}
In this section, we develop a stylized, analytical model that provides qualitative insight into the network behavior observed in the simulations and experiments for the
linear topology. Specifically, our goal is to explain why and under what conditions the phase transition occurs, and shed light into the roles
played by the retry limit~$R$ and the traffic load at the different nodes.
\subsection{Model}
\label{queueing_model}
We consider the linear topology shown in Figure~\ref{linear_topology}. Packet generations at each node $A_i$ form a
Poisson process with rate~$\lambda_i$. The packet size is fixed and the duration of each packet transmission attempt is $T$ (we assume a fixed bit
rate). A transmission by node $A_{i+1}$ is successful only if does not overlap with any transmission by (hidden) node $A_i$.
If a packet collides, it is retransmitted until either it is successfully received or the retry count reaches the limit $R$. Let $1 \leq \overline{r}_i \leq R$
represent the mean retry count at node $A_i$. Note that the initial packet transmission is included in that count. Then, the mean service time of a packet at node $A_i$ is $\overline{r}_i T$. To keep the analysis
tractable, timing details of Wi-Fi, such as DIFS, SIFS, and back-off inter-frame spacing are ignored. Therefore the upper limit of the utilization equals 1 in our analysis.
We denote the utilization of node $A_i$ by $0 \leq u_i \leq 1$, where $u_i$ represents the fraction of time node $A_i$ transmits. If $u_i=1$, node
$A_i$ is congested and transmits continuously. Otherwise, node $A_i$ is uncongested and transmits packets at rate $\overline{r}_i \lambda$.
Therefore, the utilization of node $A_i$ for all $i \geq 0$ is
\begin{equation} \label{definition of utilization}
u_i=\min\{\overline{r}_i \lambda_i T, 1\}.
\end{equation}
Note that there is no retransmission at node $A_0$ and $\overline{r}_0=1$.
Our model represents a special case of interacting queues, which are notoriously difficult to analyze~\cite{Ephremides09}. To make the analysis
tractable, we \emph{assume} that:
\begin{enumerate}
\item Packet transmissions and retransmissions at each uncongested node $A_i$ form a Poisson process with rate~$\overline{r}_i
\lambda$.\label{poisson arrival}
\item The probability that a packet transmitted by node $A_i$ collides is independent of previous attempts. This probability is denoted
$p_i$.\label{independence}
\end{enumerate}
Though the assumption of Poisson retransmissions is not fully consistent with the Wi-Fi protocol, it is similar to the ``random-look'' model used by Kleinrock and Tobagi in their analysis of (single hop) random access
networks~\cite{kleinrock1975packet} (see also~\cite{bertsekas1992data}[Ch.\ 4]).
The simulations do not incorporate the simplifications used to make the analysis tractable, yet lead to the same effects.
We stress that beside these assumptions, the rest of our analysis is exact.
\subsection{Iterative analysis of the utilization}
\label{model}
Our goal is to find the utilization at each node $i \geq 0$ and in the limit as $i \to \infty$. We consider the same scenario as in our simulations,
whereby node $A_0$ (the attacker) varies its traffic load
\begin{equation}
\rho_0 \triangleq \lambda_0 T,
\end{equation}
while all other nodes $A_i$ ($i \geq 1$) have the same traffic load
\begin{equation} \label{definition of rho}
\rho \triangleq \lambda_i T,
\end{equation}
where $0 < \rho < 1$. We aim to understand if and how changes in the value of $\rho_0$ affect the utilization of nodes that are located far away as
function of the parameters $\rho$ and $R$.
First, we get the utilization at node $A_0$:
\begin{equation}
u_0 = \min\{\rho_0, 1\}.
\end{equation}
We next develop an iterative procedure to derive $u_{i+1}$ from $u_i$.
From (\ref{definition of utilization}) and (\ref{definition of rho}),
\begin{equation} \label{definition of utilization with rho}
u_{i+1}=\min\{ \overline{r}_{i+1} \rho, 1\}.
\end{equation}
We first relate $\overline{r}_{i+1}$ to $p_{i+1}$, the probability that a packet transmitted by node $A_{i+1}$ collides.
Based on Assumption~\ref{independence}, the probability that a packet is successfully received after $1 \leq r \leq R$ attempts is
$(1 - p_{i+1}) (p_{i+1})^{r-1}$ while the probability that a packet fails to be received after $R$ attempts is $(p_{i+1})^R $.
Hence, the
mean retry count at node $A_{i+1}$ is
\begin{eqnarray} \label{LP:average retry count without retry limit}
\overline{r}_{i+1} & = & \sum_{r=1}^{R} r \cdot (1 - p_{i+1}) \cdot (p_{i+1})^{r-1} +R \cdot (p_{i+1})^R \nonumber \\
& = & \sum_{r=1}^{R} (p_{i+1})^{r-1}.
\end{eqnarray}
We next relate $p_{i+1}$ to $u_i$. First, suppose $u_i < 1$ (i.e., node $A_i$ is uncongested).
Assume that node $A_{i+1}$ starts a packet transmission (or retransmission) at some arbitrary time $t=t'$. We compute $p_{i+1}$ by
conditioning on whether or not node $A_i$ is transmitting at time $t'$. Note that due the Poisson Arrivals See Time Averages (PASTA) property, the
transmission state of node $A_i$ at time $t=t'$ is the same as at any random point of time.
If node $A_i$ transmits at time $t'$, which occurs with probability $u_i$, then the packet transmitted by node $A_{i+1}$ collides with probability 1. If node $A_i$
does not transmit at time $t'$, which occurs with probability $1-u_i$, then a collision occurs only if node $A_i$ starts a transmission during the interval
$[t',t'+T]$. Since the packet inter-arrival time on the channel is exponentially distributed with mean $\overline{r}_i T$, such an event
occurs with probability
\begin{equation}
(1 - \mathrm{e}^{-\overline{r}_i \lambda_i T}) = (1 - \mathrm{e}^{-u_i}),
\end{equation}
based on Assumption~\ref{poisson arrival}. Therefore, the unconditional probability that a packet transmitted by node
$A_{i+1}$ collides is
\begin{eqnarray} \label{LP: f(u)}
p_{i+1} & = & 1 \cdot u_i + (1 - \mathrm{e}^{-u_i})\cdot(1-u_i) \nonumber \\
& = & 1 - \mathrm{e}^{-u_i}(1-u_i).
\end{eqnarray}
Next, suppose $u_i = 1$ (i.e., node $A_i$ is congested). In that case, all the transmissions by node $A_{i+1}$ collide and $p_{i+1}=1$. We note that
(\ref{LP: f(u)}) still provides the correct result.
Putting (\ref{definition of utilization with rho}), (\ref{LP:average retry count without retry limit}), and
(\ref{LP: f(u)}) together, we obtain
\begin{equation} \label{LP: convergence of the phase}
u_{i+1} = \min \left\lbrace \rho \sum_{r=1}^{R} \left( 1 - \mathrm{e}^{-u_{i}}(1-u_{i}) \right) ^{r-1}, 1 \right\rbrace.
\end{equation}
\subsection{Limiting behaviour of the utilization}
We next analyze the limiting behaviour of the iteration given by (\ref{LP: convergence of the phase}). The sequence $(u_i)_{i=0}^{\infty}$
corresponds to a discrete non-linear dynamical system~\cite{lynch2004dynamical}. Such systems are generally complex as they may converge to a point,
to a cycle (i.e., they exhibit periodic behaviour), or not converge at all (i.e., they exhibit chaotic behaviour).
The main result of this section is to show that the sequence $(u_i)_{i=0}^{\infty}$ always converges to a point. However, the limit depends on the
initial utilization $u_0$.
To simplify notation, we define the function
\begin{equation}
f(u_{i}) \triangleq \rho \sum_{r=1}^{R} \left( 1 - \mathrm{e}^{-u_{i}}(1-u_{i}) \right) ^{r-1}.
\end{equation}
We then rewrite (\ref{LP: convergence of the phase}) as follows:
\begin{equation} \label{LP: simplification of convergence of the phase}
u_{i+1} = \min \left\lbrace f(u_{i}), 1 \right\rbrace.
\end{equation}
We say that $\omega \in [0,1]$ is a \emph{fixed point} of (\ref{LP: simplification of convergence of the phase})
if
\begin{equation} \label{number of fixed points plus}
\omega = \min \left\lbrace f(\omega), 1 \right\rbrace.
\end{equation}
Suppose (\ref{number of fixed points plus}) has $K$ different fixed points (Theorem~\ref{Thm: three stages} in the sequel will show that $K \geq
1$). We denote by $\Omega$ the ordered set of all the fixed points of (\ref{number of fixed points plus}). That is,
\begin{equation}
\Omega \triangleq\{\omega_1,\ldots,\omega_k,\ldots, \omega_K\},
\end{equation}
where $\omega_1<\ldots<\omega_k<\ldots< \omega_K$.
We are next going to show that for any $u_0\in [0, 1]$, the limit of the sequence $(u_i)_{i=0}^{\infty}$ is one of the elements in $\Omega$. To
prove this result, we will use the following lemma.
\begin{lemma} \label{lemma: f(w)>w}
Let $u, u' \in (\omega_k, \omega_{k+1})$, where $k \in \{1,\ldots,K-1\}$.
If $f(u)>u$, then $f(u') > u'$. If $f(u)<u$, then $f(u') < u'$.
\end{lemma}
\begin{IEEEproof}
The proof goes by contradiction. Let $u, u' \in (\omega_k, \omega_{k+1})$. Suppose $f(u)>u$ and $f(u')<u'$. Since $f$ is continuous in $(\omega_k,
\omega_{k+1})$, then by the intermediate-value theorem there exists a point $u''$ between $u$ and $u'$ such that $f(u'') = u''$ . Thus, $u''$ is a
fixed point of (\ref{number of fixed points plus}). This contradicts the fact that no fixed point exists between $\omega_k$ and $\omega_{k+1}$.
\end{IEEEproof}
We now present the main result of this section.
\begin{theorem} \label{thm: convergence of the fixed point}
\leavevmode
\begin{enumerate}
\item Let $u_0 \in (\omega_k, \omega_{k+1})$, where $k \in \{1,\ldots,K-1\}$.
If $f(u_{0}) > u_{0}$, the sequence $(u_i)_{i=0}^{\infty}$ converges to $\omega_{k+1}$.
If $f(u_{0}) < u_{0}$, the sequence $(u_i)_{i=0}^{\infty}$ converges to $\omega_{k}$.
\item If $u_0 \in [0, \omega_1)$, the sequence $(u_i)_{i=0}^{\infty}$ converges to $\omega_1$.
\item If $\omega_K<1$ and $u_0 \in (\omega_K, 1]$, the sequence $(u_i)_{i=0}^{\infty}$ converges to $\omega_K$.
\end{enumerate}
\end{theorem}
\begin{IEEEproof}
\leavevmode
\begin{enumerate}
\item Let $\omega_k < u_0 < \omega_{k+1}$, where $k \in \{1,\ldots,K-1\}$. Since
$p_i \in (0, 1)$. Therefore,
the function $f$ is continuous and monotonically increasing,
$f(\omega_k) < f(u_0) < f(\omega_{k+1})$. Hence, according to (\ref{LP: simplification of convergence of the phase}) and (\ref{number of fixed
points plus}), we get
\begin{equation} \label{w_k < u_1 < w_k+1}
\omega_k \leq u_1 \leq \omega_{k+1}.
\end{equation}
Now, suppose \mbox{$u_1= f(u_{0}) > u_{0}$}.
If $u_1 = \omega_{k+1}$, then the result is proven.
If $u_1 < \omega_{k+1}$, then by Lemma~\ref{lemma: f(w)>w} and Equation~(\ref{w_k < u_1 < w_k+1}), we have $u_2 = f(u_1) > u_1$. Applying the
same argument inductively, either there exists some value $M \geq 2$ such that $u_i=\omega_{k+1}$ for all $ i \geq M$, or the sequence
$(u_i)_{i=0}^{\infty}$ is monotonically increasing and upper bounded by $\omega_{k+1}$. According to the monotone convergence theorem, the
sequence converges. Since there is no other fixed point between $u_0$ and $\omega_{k+1}$ and $f$ is continuous, the sequence
$(u_i)_{i=0}^{\infty}$ must converge to $\omega_{k+1}$. The case $u_1 = f(u_{0}) < u_{0}$ is handled similarly.
\item Similar to Lemma~\ref{lemma: f(w)>w}, one can show that if there exists $u \in [0, \omega_1)$ such that $f(u) >u$, then $f(u')>u'$ for all
$u' \in [0, \omega_1)$. Since $f(0)=\rho>0$, the sequence $(u_i)_{i=0}^{\infty}$ converges to $\omega_1$.
\item This is handled similarly to case 2.
\end{enumerate}
\end{IEEEproof}
\subsection{Phase transition analysis}
\label{The phase transition phenomenon}
In the previous section, we showed that the limit of the sequence of node utilizations $(u_i)_{i=0}^{\infty}$ must be one of the fixed points in the
set $\Omega$. A phase transition represents a situation where a small change of $u_0$ leads to an abrupt change of the limit. Specifically, we
focus on the case when the limit jumps to 1. Formally:
\begin{definition}[Network congestion] A network is said to be \textit{congested} if $(u_i)_{i=0}^{\infty}$ converges to $1$. Else, the network is
said to be \textit{uncongested}.
\end{definition}
\begin{definition}
[Phase transition] A network experiences a phase transition if there exists a fixed point $\omega \in \Omega$, such that if
$u_0 < \omega$ the network is uncongested, and if $u_0 > \omega$ the network is congested. We refer to $\omega$ as the phase transition point.
\end{definition}
We note that a phase transition can possibly occur only if $\omega_K=1$, since otherwise the network is never congested, irrespective of $u_0$.
A network must fall in one of the following three regimes:
\begin{enumerate}
\item The network is uncongested for all $u_0 \in [0,1]$.
\item The network is congested for all $u_0 \in [0,1]$.
\item A phase transition occurs.
\end{enumerate}
Our goal in the following is to determine what regime prevails under different network parameters.
For this purpose, we investigate the existence and properties of solutions of (\ref{number of fixed points plus}).
First, we investigate the case $\omega = 1$.
\begin{lemma} \label{lemma:overload fixed point is stable}
If $\rho > 1/R$, then
\begin{enumerate}
\item $\omega_K=1$.
\item If $K =1 $, then for all $u_0 \in [0,\omega_K]$ the sequence $(u_i)_{i=0}^{\infty}$ converges to $\omega_K$.
\item If $K \geq 2$, then for all $u_0 \in (\omega_{K-1},\omega_K]$ the sequence $(u_i)_{i=0}^{\infty}$ converges to $\omega_K$.
\end{enumerate}
\end{lemma}
\begin{IEEEproof}
\leavevmode
\begin{enumerate}
\item Let $\rho \geq 1/R$. We compute the RHS of (\ref{number of fixed points
plus}) at $\omega=1$ and obtain $\min\{ f(1), 1\} = \min\{ R\rho, 1\} =1$, which proves that a fixed point indeed exists at $\omega=1$.
\item If $\rho > 1/R$, then $f(1)=R \rho>1$. Since $f(1)>1$, then for all $u_0 \in (0, \omega_K)$ , we have $f(u_0) > u_0$, based on an argument
similar to Lemma~\ref{lemma: f(w)>w}, and the sequence $(u_i)_{i=0}^{\infty}$ converges to $1$, following an argument similar to
Theorem~\ref{thm: convergence of the fixed point}.
\item This is handled similarly to Part 2.
\end{enumerate}
\end{IEEEproof}
Lemma~\ref{lemma:overload fixed point is stable} indicates that the sequence $(u_i)_{i=0}^{\infty}$ can converge to 1 (depending on $u_0$), if $
\rho > 1/R$. Besides this special case, (\ref{number of fixed points plus}) can be rewritten
\begin{equation} \label{LP: the fixed point}
f(\omega) = \omega.
\end{equation}
We look for solutions of (\ref{LP: the fixed point}) that belong to the interval $[0,1]$. Each such solution is an element of $\Omega$.
Equation (\ref{LP: the fixed point}) is difficult to work with because it contains two unknown variables, $\rho$ and $R$. To circumvent this
difficulty, we introduce the function
\begin{equation} \label{lambdaT}
h_R(\omega) \triangleq \frac{\rho \omega}{f(\omega)} = \frac{\omega}{\sum_{r=1}^{R}\left(1 - \mathrm{e}^{-\omega}(1-\omega)\right)^{r-1}}.
\end{equation}
For each value of $\rho$, the solutions of (\ref{LP: the fixed point}) must satisfy
\begin{equation} \label{lambdaT=rho}
h_R(\omega) = \rho.
\end{equation}
We denote the maximum of $h_R(\omega)$
by
\[h^{max}_R \triangleq \max_{0 \leq \omega \leq 1}h_R(\omega).\]
The following theorem establishes the prevailing network regimes for different parameters.
\begin{theorem} \label{Thm: three stages}
\leavevmode
\begin{enumerate}
\item If $\rho < 1/R$, then the network is uncongested for all $u_0 \in [0,1]$.
\item If $h^{max}_R > 1/R$ and $1/R < \rho < h^{max}_R$, then a phase transition occurs and the phase transition point is $\omega_{K-1}$.
\item If $\rho > h^{max}_R$, then the network is congested for all $u_0 \in [0,1]$.
\end{enumerate}
\end{theorem}
\begin{IEEEproof}
\leavevmode
\begin{enumerate}
\item If $\rho < 1/R$, then $R \rho < 1$ and the utilization of each node is always less than 1. Hence, for any $u_0 \in [0,1]$, the network is
always uncongested. Note that since $h_R(0)=0$, $h_R(1)=1/R$, and $h_R$ is continuous, (\ref{lambdaT=rho}) must have at least one solution
(i.e., at least one fixed point exists).
\item Let $\rho \in (1/R, h^{max}_R)$. We know that $h_R(0) = 0$ and $h_R(1)=1/R$. Since the function $h_R$ is continuous, (\ref{lambdaT=rho})
must have at least one solution (i.e, at least one fixed point strictly smaller than 1 exists). Also, because $\rho > 1/R$, a fixed point
point at $\omega=1$ exists (i.e., $\omega_K=1$), by Part 1 of Lemma~\ref{lemma:overload fixed point is stable}. Thus, there are $K \geq 2$
fixed points.
By Part 3 of Lemma~\ref{lemma:overload fixed point is stable}, the sequence $(u_i)_{i=0}^{\infty}$ converges to $\omega_K$ for all $u_0 \in
(\omega_{K-1},\omega_K]$. Moreover, by Theorem~\ref{thm: convergence of the fixed point}, the limit of the sequence $(u_i)_{i=0}^{\infty}$ is no
larger than $\omega_{K-1}$ for all $u_0 \leq \omega_{K-1}$. Hence, a phase transition exists at $\omega_{K-1}$.
\item If $\rho > h^{max}_R$, then (\ref{LP: the fixed point}) has no solution. Moreover, since $\rho > h^{max}_R \geq h_R(1) = 1/R$, we get $\rho
> 1/R$. By Parts 1 and 2 of Lemma~\ref{lemma:overload fixed point is stable}, the sequence $(u_i)_{i=0}^{\infty}$ converges to 1 for any $u_0
\in [0,1]$, and the network is always congested.
\end{enumerate}
\end{IEEEproof}
We next illustrate Theorem~\ref{Thm: three stages} for different values of $R$, using Figure~\ref{number_of_fixed_points}.
First, consider $R = 4$ as shown in Figure~\ref{number_of_fixed_points}(a). Since $h^{max}_R = 1/R=0.25$, there exists no traffic load $\rho$ for
which a phase transition exists. Either the network is always uncongested (for $\rho < 1/R$), or it is always congested (for $\rho > 1/R$).
Next, consider $R = 7$ as shown in Figure~\ref{number_of_fixed_points}(b). There, $h^{max}_R = 0.166 > 1/R = 0.143$. Hence, a phase transition
occurs if $\rho \in (0.143, 0.166)$. For instance, consider the case $\rho = 0.15$. Then, the equation $h_R(\omega)=\rho$ has two solutions.
Including the fixed point $\omega=1$ (since $\rho > 1/R)$, the set $\Omega$ has $K=3$ fixed points: $\{ \omega_1=0.265, \omega_2=0.777,
\omega_3=1\}$. Hence, by Theorem~\ref{Thm: three stages}, the network is uncongested if $u_0 < 0.777$, and congested if $u_0 > 0.777$.
The case $R=10$ also has a phase transition region, as shown in Figure~\ref{number_of_fixed_points}(c). Furthermore, the size of this region is
larger since $(1/R, h^{max}_R) = (0.1, 0.162)$.
\begin{figure}[!t]
\centering
\subfloat[$R=4$]
{\includegraphics[width=1.15in]{Ris4_new5}
\label{R is 4}}
\subfloat[$R=7$]
{\includegraphics[width=1.15in]{Ris7_new5}
\label{R is 7}}
\subfloat[$R=10$]
{\includegraphics[width=1.15in]{Ris10_new5}
\label{R is 10}}
\caption{Illustration of the different network regimes for different values of $R$. For each value of~$\rho$, the fixed points are the solutions of
$h_R(\omega)= \rho$. In addition, the fixed point $\omega =1$ always exists when $\rho > 1/R$. A phase transition region exists if the maximum of
$h_R(\omega)$, $h_R^{max}$, is strictly greater than $h_R(1) = 1/R$.}
\label{number_of_fixed_points}
\end{figure}
\subsection{Sufficient condition for phase transition}
\label{The sufficient condition for phase transition}
In the previous section, we showed that a phase transition exists in the region $1/R < \rho < h^{max}_R$, if $h^{max}_R > 1/R$. In this section, we
derive an explicit lower bound on $h^{max}_R$, which provides a simple condition for the existence of a phase transition. First, we establish a
relationship between the derivatives of $h_{R}(\omega)$ for different values of $R$, but a given value of $\omega$.
\begin{lemma} \label{unstable then unstable}
For $\omega \in [0,1]$, if there exists $R^* \geq 1$ such that $h_{R^*}'(\omega) \leq 0$, then $h_R'(\omega) \leq 0$ for all $R>R^*$.
\end{lemma}
\begin{IEEEproof}
Let $\omega \in [0,1]$.
Since
\begin{equation}\label{h^-1}
\left(h_R^{-1}(\omega)\right)' = - \frac{h_R'(\omega)}{h_R(\omega)^2},
\end{equation}
the sign of $h_R'(\omega)$ is opposite to $\left(h_R^{-1}(\omega)\right)'$. Hence, we investigate the sign of
\begin{equation} \label{1/h'(w)}
\left(h_R^{-1}(\omega)\right)' = \sum_{r=1}^{R}\Psi_r'(\omega),
\end{equation}
where
\begin{equation}\label{Psi_r}
\Psi_r(\omega) \triangleq \frac{\left(1 - \mathrm{e}^{-\omega}(1-\omega)\right)^{r-1}}{\omega}.
\end{equation}
We check the sign of each term $\Psi_r'(\omega)$ in~(\ref{1/h'(w)}), for $r \in \{1, 2, \ldots, R\}$.
For $r=1$, we have
\[\Psi_1'(\omega)=\frac{d}{d \omega}\left(\frac{1}{\omega}\right)=-\frac{1}{\omega^2}<0.\]
For $r\geq2$, we have
\begin{equation}\label{Psi_r'} \Psi_r'(\omega) =-\frac{e^{-\omega} \left(1 - \mathrm{e}^{-\omega}(1-\omega)\right)^{r-2}\Phi_r(\omega)}{\omega^2},
\end{equation}
where \[\Phi_r(\omega) \triangleq -1+e^\omega+(3-2r)\omega+(r-1)\omega^2.\] Clearly, the terms $e^{-\omega}$, $\left(1 -
\mathrm{e}^{-\omega}(1-\omega)\right)^{r-2}$ and $\omega^2$ in~(\ref{Psi_r'}) are all positive. Thus, the signs of $\Phi_r(\omega)$ and
$\Psi_r'(\omega)$ are opposite.
We next investigate the signs of the first and second derivatives of the function $\Phi(\omega)$. We have
\begin{eqnarray}
\Phi_r'(\omega) & = & e^\omega+3-2r+2(r-1)\omega, \label{Phi_r'}\\
\Phi_r''(\omega)& = & e^\omega+2(r-1)>0, \label{Phi_r''}
\end{eqnarray}
for all $\omega \in [0,1]$ and $r \geq 2$.
From (\ref{Phi_r''}), we find that $\Phi_r'(\omega)$ is monotonically increasing with $\omega$.
For any $r\geq2$, we obtain from (\ref{Phi_r'}) that
\begin{eqnarray}
\Phi_r'(0) & = & 4-2r, \\
\Phi_r'(1) & = & e+1.
\end{eqnarray}
We distinguish between three possible cases regarding the sign of $\Phi_r(\omega)$:
\begin{enumerate}
\item For $r=2$, $\Phi_2'(0)=0$. Hence, $\Phi_2'(\omega)>0$. The function $\Phi_2(\omega)$ is monotonically increasing with $\omega$. Since
$\Phi_2(0)=e-1>0$, $\Phi_2(\omega)$ is always positive.
\item For $r=3$, $\Phi_3'(0)<0$. The function $\Phi_3(\omega)$ first decreases then increases as $\omega$ increases from 0 to 1. Since
$\Phi_3(0)=0$ and $\Phi_3(1)>0$, the sign of the function $\Phi_3(\omega)$ turns from negative to positive as $\omega$ increases from $0$ to $1$.
\item For $r>3$, $\Phi_r'(0)<0$. The function $\Phi_r(\omega)$ first decreases then increases as $\omega$ increases from 0 to 1. Since
$\Phi_r(0)=0$ and $\Phi_r(1)<0$, the sign of the function $\Phi_r(\omega)$ is always negative.
\end{enumerate}
Therefore, by~(\ref{1/h'(w)}), for any given $\omega \in [0,1]$, the sign of the function $\Phi_r(\omega)$ turns from being positive to being
negative as $r$ increases. Equivalently, the sign of the function $\Psi_r'(\omega)$ turns from being negative to being positive as $r$ increases.
Thus, by~(\ref{1/h'(w)}), if $\left(h_{R}^{-1}(\omega)\right)'$ is nonnegative for $R=R^*$, then it is also nonnegative for all $R \geq R^*$.
Equivalently, by~(\ref{h^-1}), if $\left(h_R^{-1}(\omega)\right)'$ is nonpositive for $R=R^*$, then it is also nonpositive for all $R \geq R^*$,
which completes the proof.
\end{IEEEproof}
Consider the function $h_R(\omega)$ as $R\rightarrow\infty$:
\begin{eqnarray} \label{lim R infty h_R}
h_\infty(\omega) & = & (1-\left(1 - \mathrm{e}^{-\omega}(1-\omega)\right))\omega \nonumber\\
& = & e^{-\omega}(1-\omega)\omega,
\end{eqnarray}
and its derivative
\begin{equation} \label{lim R infty h_R'}
h_\infty'(\omega) = e^{-\omega}(1-3\omega+\omega^2).
\end{equation}
The next corollary is the logical transposition of Lemma \ref{unstable then unstable}.
\begin{corollary} \label{stability when R infty}
If $h_{\infty}'(\omega) \geq 0$, then $h_R'(\omega) \geq 0$ for all $R \geq 1$.
\end{corollary}
The following lemma establishes that the function $h_R(\omega)$ is always strictly increasing in the interval $[0, \overline{\omega})$, where
\begin{equation}\label{overline_omega}
\overline{\omega} \triangleq \frac{3-\sqrt{5}}{2}.
\end{equation}
\begin{lemma}
\label{stable fixed point}
Let $0 \leq \omega<\overline{\omega}$. Then, $h_R'(\omega)>0$, for all $R \geq 1$.
\end{lemma}
\begin{IEEEproof}
Let the function $h_\infty(\omega)$ and its derivative $h_\infty'(\omega)$ be defined as in (\ref{lim R infty h_R}) and (\ref{lim R infty h_R'}),
respectively.
Since $e^{-\omega}$ is always positive, $h_\infty'(\omega)$ has the same sign as $(1-3\omega+\omega^2)$.
The unique root of $(1-3\omega+\omega^2)=0$ for $\omega \in [0,1]$ is $\bar{w}$ as defined in (\ref{overline_omega}).
Thus, $(1-3\omega+\omega^2)$ is positive when $0 \leq \omega<\overline{\omega}$,
and so is $h_\infty'(\omega)$. By Corollary \ref{stability when R infty}, $h_R'(\omega)>0$ for $0 \leq \omega<\overline{\omega}$ and for all $R \geq
1$.
\end{IEEEproof}
The consequence of Lemma~\ref{stable fixed point} is that for all $R \geq 1$,
\begin{equation} \label{h_max_h_omega}
h^{max}_R \geq h_R(\overline{\omega}).
\end{equation}
This equation provide a lower bound on $h^{max}_R$ that can easily be computed. We then obtain the following sufficient condition for the existence
of phase transition.
\begin{theorem}
\label{Thm:range of lambda T}
Let $\overline{\omega}$ be defined as in (\ref{overline_omega}) and suppose $h_R(\overline{\omega})> 1/R$. Then, a phase transition is guaranteed to
exist for any $\rho \in (1/R, h_R(\overline{\omega}))$.
\end{theorem}
\begin{IEEEproof}
From Theorem \ref{Thm: three stages}, we know that a phase transition exists if $1/R < \rho< h^{max}_R$. By~(\ref{h_max_h_omega}) and the assumption
that $h_R(\overline{\omega})> 1/R$, the proof follows.
\end{IEEEproof}
The next theorem establishes an even more explicit lower bound on $h^{max}_R$.
\begin{theorem}
\label{Thm:specific range of lambda T}
Let $h_\infty(\omega)$ and $\overline{\omega}$ be defined as in (\ref{lim R infty h_R}) and (\ref{overline_omega}), respectively. Then,
$h^{max}_R \geq h_\infty(\overline{\omega}) \simeq 0.161$.
\end{theorem}
\begin{IEEEproof}
By~(\ref{lambdaT}),
\begin{eqnarray}
h_R(\overline{\omega}) & = & \frac{\omega}{\sum_{r=1}^{R}(1-e^{-\omega}(1-\omega))^{r-1}} \nonumber\\
& > & \frac{\omega}{\sum_{r=1}^{\infty}(1-e^{-\omega}(1-\omega))^{r-1}} = h_\infty(\overline{\omega}). \label{hinf_ineq}
\end{eqnarray}
Thus, by~(\ref{h_max_h_omega}) and ~(\ref{hinf_ineq}), $h^{max}_R > h_\infty(\overline{\omega}) \simeq 0.161$. Note that this bound is
asymptotically tight as $R \to \infty$ since $h^{max}_{\infty} = h_\infty(\overline{\omega})$.
\end{IEEEproof}
From Theorems~\ref{Thm: three stages} and~\ref{Thm:specific range of lambda T}, it follows that a phase transition exists if $1/R < 0.161$. Hence:
\begin{corollary}\label{R>=7}
A phase transition is guaranteed to exist for $R\geq7$ and $\rho \in [1/R,0.161]$.
\end{corollary}
We note that the lower bound on $h^{max}_R$ is quite tight. For instance, $h^{max}_7=0.166$. Moreover, $h^{max}_R$ decreases with $R$ (this follows
from~(\ref{lambdaT}), since for any $\omega \in [0,1]$ the denominator increases as $R$ gets larger).
\subsection{Stability of fixed points}
\label{Stability of Fixed point}
In this subsection, we use stability theory to shed further light into the limiting behaviour of the sequence $(u_i)_{i=0}^{\infty}$.
Specifically, the sequence $(u_i)_{i=0}^{\infty}$ converges to \emph{stable} fixed points of $\Omega$ and diverges from \emph{unstable} fixed points
of $\Omega$. We will show that the stability of the fixed points of~(\ref{LP: the fixed point}) are determined by the sign of $h_R'(\omega)$ at
those points.
Informally, a fixed point $\omega$ is stable (or an \emph{attractor}), if there exists a domain containing $\omega$, such that if $u_0$ belongs to
that domain, then $(u_i)_{i=0}^{\infty}$ converges to $\omega$.
\begin{definition}[Stability of a fixed point] \label{def: stable fixed point}
Let $u_0 \in [0,1]$. A fixed point $\omega \in \Omega$ is \textit{stable} if there exists $\epsilon>0$ such that if $|u_0-\omega|< \epsilon$, the
sequence $(u_i)_{i=0}^{\infty}$ converges to $\omega$.
It is \textit{unstable} if for all $u_0 \neq \omega$ the sequence $(u_i)_{i=0}^{\infty}$ does not converge to $\omega$.
\end{definition}
Recall that according to Lemma \ref{lemma:overload fixed point is stable}, a special fixed point of (\ref{number of fixed points plus}) exists at
$\omega = 1$, if $\rho > 1/R$. According to Definition~\ref{def: stable fixed point}, this fixed point is stable.
Besides this special case, the rest of the fixed points satisfy Equation~(\ref{LP: the fixed point}).
To establish the stability of those fixed points, we will employ the following proposition.
\begin{proposition}[{\cite{lynch2004dynamical}}]\label{prop: stable fixed point}
Suppose that a continuously differentiable function $f$ has a fixed point $\omega$. Then, $\omega$ is stable if \mbox{$|f'(\omega)| < 1$} and
unstable if $|f'(\omega)| > 1$.
\end{proposition}
The next theorem provides a criterion to establish the stability of a fixed point $\omega \in \Omega$ with respect to the function $h_R(\omega)$.
\begin{theorem} \label{stability}
Consider a fixed point $\omega \in \Omega$, where $\omega < 1$. Then $\omega$ is stable if $h_R'(\omega) > 0$ and unstable if $h_R'(\omega) < 0$.
\end{theorem}
\begin{IEEEproof}
Let $\omega \in \Omega$. The derivative of $h_R(\omega)$ with respect to $\omega$ is
\begin{equation} \label{Eq: partial of h}
h_R'(\omega) = \frac{1}{\Gamma(\omega)} - \frac{\omega}{(\Gamma(\omega))^2} \cdot \Gamma'(\omega) > 0,
\end{equation}
where
\begin{equation} \label{Eq: Gamma omega}
\Gamma(\omega) \triangleq \sum_{r=1}^{R}\left(1 - \mathrm{e}^{-\omega}(1-\omega)\right)^{r-1} = \frac{f(\omega)}{\rho}.
\end{equation}
If one can show that (\ref{Eq: partial of h}) implies $|f'(\omega)|<1$, then according to Proposition \ref{prop: stable fixed point}, the fixed
point $\omega$ is stable. We multiply both sides of (\ref{Eq: partial of h}) by $(\Gamma(\omega))^2$ and obtain
\begin{equation} \label{Eq: transform of partial of h}
\Gamma(\omega) - \omega \Gamma'(\omega) > 0.
\end{equation}
Using (\ref{Eq: Gamma omega}) and (\ref{LP: the fixed point}), we can rearrange (\ref{Eq: transform of partial of h}) as follows:
\begin{equation} \label{Gamma < 1 rho}
\Gamma'(\omega) < \frac{\Gamma(\omega)}{\omega} = \frac{f(\omega)}{\rho \omega} = \frac{1}{\rho}.
\end{equation}
From (\ref{Eq: Gamma omega}) and (\ref{Gamma < 1 rho}), we get \[f'(\omega) = \rho \Gamma'(\omega) < 1.\]
Since $f(\omega)$ is monotonically increasing with $\omega$, for $\omega \in [0,1]$, we conclude
\[0 < f'(\omega) < 1.\]
Hence, by Proposition \ref{prop: stable fixed point}, $\omega$ is a stable fixed point.
Similarly, $h_R'(\omega) < 0$ implies $f'(\omega) > 1$, which means that $\omega$ is unstable.
\end{IEEEproof}
We next show how the stability analysis of the fixed points helps to determine the limit of the sequence $(u_i)_{i=0}^{\infty}$. Consider, for
instance, the example shown in Figure \ref{stability example} with parameters $R=10$ and $\rho=0.13$. Under these parameters, $\Omega = \{\omega_1,
\omega_2, \omega_3\} = \{0.2, 0.7, 1\}$.
The fixed points $\omega_1$ and $\omega_2$ are the solutions of $h_R(\omega) = \rho$. According to Theorem \ref{stability}, $\omega_1$ is stable and
$\omega_2$ is unstable. The fixed point $\omega_3=1$ exists and is stable, since $\rho > 1/R$.
According to Theorem \ref{Thm: three stages}, $\omega_2$ is a phase transition point. Hence, the sequence $(u_i)_{i=0}^{\infty}$ converges to
$\omega_1$ if $u_0 < \omega_2$ and the network is uncongested. If $u_0 > \omega_2$, the sequence converges to $\omega_3$ and the network is
congested.
\begin{figure}[!t]
\centering
\includegraphics[width=3.2in]{stability_example_4}
\caption{Stability of fixed points with $R=10$. Given a load $\rho = 0.13$ (dash line), $\Omega$ contains three fixed points: $\omega_1 = 0.2$,
$\omega_2 = 0.7$ and $\omega_3=1$. The fixed point $\omega_1$ is stable because $h'_R(\omega_1) > 0$ and $\omega_2$ is unstable because
$h'_R(\omega_2) < 0$. The fixed point $\omega_3=1$ exists and is stable because $\rho > 1/R$. Therefore, the sequence $(u_i)_{i=0}^{\infty}$
converges to $\omega_1$ if $u_0 < \omega_2$, and to $\omega_3$ if $u_0 > \omega_2$.}
\label{stability example}
\end{figure}
\subsection{Heterogeneous traffic load}
\label{sec:heter}
In previous subsections, we assumed that node $A_0$ varies its traffic load
$\rho_0$, but all other nodes $A_i$ ($i \geq 1$) have the same traffic load $\rho$.
We now relax this assumption and assume that nodes $A_i$ ($i \geq 1$) have different traffic loads
$\rho_i = \lambda_i T$.
We next prove that a phase transition still occurs, as long as all the traffic loads fall in the appropriate range.
\begin{theorem}
Suppose $h^{max}_R > 1/R$. If $\rho_i \in (1/R, h^{max}_R)$ for all $i \geq 1$, then a phase transition occurs.
\end{theorem}
\begin{IEEEproof}
Let $\rho_{max} = \max_{i \geq 1} \rho_i$ and $\rho_{min} = \min_{i \geq 1} \rho_i$.
According to Theorem~\ref{Thm: three stages}, the network is uncongested when $\rho_0=0$ and the load at each node $A_i$ is $\rho_{max} <
h_R^{max}$.
Hence, the network must remain uncongested when the load at each node $A_i$ is smaller than $\rho_{max}$.
Similarly, the network is congested when $\rho_0=1$ and the load at each node $A_i$ is $\rho_{min} > 1/R$.
Hence, it must remain congested when the load at each node $A_i$ is larger than $\rho_{min}$.
Thus, a phase transition occurs when $1/R < \rho_i < h_R^{max}$ for all $i \geq 1$.
\end{IEEEproof}
\subsection{Comparison with simulation results}
\label{Mitigation}
We compare the results of our analysis with ns-3 simulations, for different settings of the retry limit $R$ and load $\rho$. For the simulations, we
consider an ad~hoc network composed of 41 pairs of nodes, as described in Section~\ref{EXP:Fixed bitrate}.
\subsubsection{Region of phase transition}
\label{validation of Region of phase transition}
To check whether a phase transition exists for a given $R$, we run simulations both for $\rho_0 =0$ and $\rho_0 = 1$. If the node utilizations in
the limit (i.e., for node $A_{40}$) is the same in both cases, then we assume that there is no phase transition. If the limits are different, then a
phase transition exists.
Figure~\ref{retry_limit} indicates that the existence of a phase transition is related to the retry limit, as predicted by our analysis. For the
case $R=4$, there is no phase transition, while a phase transition occurs in the cases $R=7$ and $R=10$.
in our simulations for any $R \leq 6$.
The analysis also reasonably approximates the phase transition region. For $R=7$, the simulations show that a phase transition exists if $\rho \in
(0.12,0.16)$, while the analysis predicts $\rho \in (0.14, 0.17)$. For $R=10$, the simulation results are $\rho \in (0.08, 0.14)$ while the analysis
predicts $\rho \in (0.10, 0.16)$. We note that the size of the phase transition region increases with $R$, as predicted by the analysis.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{retry_limit_new3}
\caption{Simulation of the limiting behaviour of the node utilization in a network of 41 pairs of nodes. For $R=4$, the limit is the same when
$\rho_0 = 0 $ and $\rho_0 = 1$, hence no phase transition is observed. However, for $R=7$ and $R=10$,
the limits are different, hence showing the existence of a region of load $\rho$ in which a phase transition occurs. }
\label{retry_limit}
\end{figure}
\subsubsection{Heterogeneous traffic load}
\label{validation of Heterogeneous traffic load}
We next show the feasibility of a cascading DoS attack in
a network where the traffic load at different node is heterogeneous, in line with the analysis of Section~\ref{sec:heter}.
Specifically, the traffic load $\rho_i$ at each node $A_i$ ($i \geq 1$) is a continuous random variable that is uniformly distributed between 0.11 and
0.15.
Figure~\ref{CDoS-1Mbps-adhoc-UDP-randomrho-utilization} shows the simulation results for retry limit $R=7$.
When $\rho_0$, the load of node $A_0$, is below 0.5, the network is uncongested and the utilizations of nodes $A_i$ oscillate around 0.35 as $i$ gets large. Note that the sequence does not converge to a fixed value due to the different traffic loads at the different nodes.
However, when $\rho_0$ exceeds 0.6, the sequence of node utilizations converges to its upper limit, implying that the network is congested.
\begin{figure}[!t]
\centering
\includegraphics[width=2.5in]{CDoS-1Mbps-adhoc-UDP-randomrho-utilization}
\caption{Simulation with heterogeneous traffic load in a network with 41 pairs of nodes. The traffic load of nodes $A_i$ ($i \geq 1$) are uniformly
distributed between 0.11 and 0.15. For $R=7$, when the load $\rho_0$ changes from 0.5 to 0.6, the limiting behavior of the sequence of node utilizations differs, thus indicating the occurrence of phase transition.}
\label{CDoS-1Mbps-adhoc-UDP-randomrho-utilization}
\end{figure}
\section{Conclusion}
\label{Conclusion}
We describe a new type of DoS attacks against Wi-Fi networks, called cascading DoS attacks.
The attack exploits a coupling vulnerability due to hidden nodes.
The attack propagates beyond the starting location, lasts for long periods of
time, and forces the network to operate at its lowest bit rate. The attack can be started
remotely and without violating the IEEE 802.11 standard, making it
difficult to trace back.
We demonstrate the feasibility of such attacks, both through experiments on a testbed and extensive ns-3 simulations.
The simulations show that the attack is effective in networks operating under fixed and varying bit rates, as well
as ad hoc and infrastructure modes. We show that a small change
in the traffic load of the attacker can lead to a phase transition of the entire network, from uncongested state to congested state.
We develop an iterative analysis to characterize the sequence of node utilizations, and study its limiting behaviour. We show that the sequence
always converges to stable fixed points while an unstable fixed point represents a phase transition point. Based on the system parameters, we
identify when the system remains always uncongested, congested, or experiences a phase transition caused by a DoS cascading attack.
The analysis predicts that a phase transition occurs for $R \geq 7$ and provides a simple and explicit estimate of traffic load at each node under
which a phase transition occurs (i.e., $\rho_i \in (1/R, 0.161)$ for all $i \geq 1$). The network is always congested when the traffic load is
above the phase transition regime and always uncongested when the traffic load is below the phase transition regime. Although the analysis is based
on some simplifying assumptions, the estimate is not far from the values observed in the simulations.
Exploiting the coupling vulnerability in different network configurations represents an interesting area for future work.
Experience in the security field indeed teaches that once a vulnerability is identified, more potent attacks are subsequently discovered
(consider,
for instance, the history of attacks on WEP~\cite{tews2007breaking} and MD5~\cite{black2006study}). In our case, our simulations for ring topologies
indicate that the presence of a cycle in the topology could reinforce cascading DoS attacks, a result that warrants further investigations.
Several approaches are possible to mitigate cascading DoS attacks. First, one could enable the RTS/CTS exchange, although this solution has several
drawbacks, including major performance degradation under normal network operations, as mentioned in the Introduction. Devising a scheme that triggers RTS/CTS under certain circumstances (e.g., multiple consecutive packet losses) could be an interesting area for future research.
The second approach is to
lower the retry limit. However, this could also negatively impact performance. Other approaches include using short packets, collision-aware rate
adaptation algorithms, dynamic channel selection, and full-duplex radios. We leave the investigation and comparison of these
mitigation techniques as possible areas for future work.
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The classification problem for Cuntz-Krieger algebras has a long and prominent history. Indeed, R\o rdam's classification
(\cite{MR1340839}) of the simple such $C^*$-algebras by appealing to
fundamental results in symbolic dynamics paved the way for the
sweeping generalization by Kirchberg and Phillips (\cite{MR1796912} and \cite{MR1745197}) to all simple,
nuclear, separable and purely infinite $C^*$-algebras in the UCT class, and Restorff's
generalization (\cite{MR2270572}) to the general case of Cuntz-Krieger
algebras with finitely many ideals (equivalent to Cuntz'
Condition (II)) was a key inspiration for the recent surge in results
concerning nonsimple purely infinite $C^*$-algebras.
Until now, almost nothing has been known about the classification of
Cuntz-Krieger $C^*$-algebras having infinitely many ideals --- failing
Condition (II) --- even though the symbolic dynamical systems that
define them are often extremely simple. In this paper, we will
establish classification up to stable isomorphism between the Cuntz-Krieger algebras
defined from a large class of graphs including the pairs of graphs
given in (a) and (b) of Figure \ref{firstexx}, but must leave open the
question concerning some more complicated graphs such as the ones in (c).
\begin{figure}
\begin{center}
\begin{tabular}{|cc|cc|ccc|}\hline
\qquad&&&&&&\\
$\xymatrix{\bullet\ar[d]\ar@(ur,dr)[]\ar@(u,r)[]\\\bullet\ar[d]\ar@(ur,dr)[]\\\bullet\ar@(ur,dr)[]}$\qquad\qquad&
$\xymatrix{\bullet\ar@/_15pt/[dd]\ar@/_/[d]\ar@/^/[d]\ar@(ur,dr)[]\ar@(u,r)[]\\\bullet\ar[d]\ar@(ur,dr)[]\\\bullet\ar@(ur,dr)[]}$\qquad\qquad&$\xymatrix{\bullet\ar[d]\ar@(ur,dr)[]\\\bullet\ar@(u,r)[]\ar[d]\ar@(ur,dr)[]\\\bullet\ar@(ur,dr)[]}$\qquad\qquad\qquad&
$\xymatrix{\bullet\ar[d]\ar@(ur,dr)[]\ar@/_15pt/[dd]\ar@/_25pt/[dd]\\\bullet\ar[d]\ar@(u,r)[]\ar@(ur,dr)[]\\\bullet\ar@(ur,dr)[]}$\qquad\qquad&
$\xymatrix{\bullet\ar[d]\ar@(ur,dr)[]\\\bullet\ar[d]\ar@(ur,dr)[]\\\bullet\ar[d]\ar@(u,r)[]\ar@(ur,dr)[]\\\bullet\ar@(ur,dr)[]}$\qquad\qquad\qquad&
$\xymatrix{\bullet\ar[d]\ar@(ur,dr)[]\\\bullet\ar[d]\ar@(ur,dr)[]\ar@/_15pt/[dd]\ar@/_25pt/[dd]\\\bullet\ar[d]\ar@(u,r)[]\ar@(ur,dr)[]\\\bullet\ar@(ur,dr)[]}$&\\
&&&&&&
\\\hline
\multicolumn{2}{|c|}{(a)}& \multicolumn{2}{|c|}{(b)}& \multicolumn{3}{|c|}{(c)}\\\hline
\end{tabular}
\end{center}
\caption{Six graphs}\label{firstexx}
\end{figure}
We work in the more general (and more natural) setting of graph $C^*$-algebras over finite graphs, where Condition (II) is replaced by the standard Condition (K).
Following \cite{MR3082546} and \cite{MR3047630} we emphasize the question of \emph{geometric} classification, the aim being to generate the equivalence relation on graphs induced by stable isomorphism of the associated $C^*$-algebras as the coarsest equivalence relation containing the class of basic moves on the graphs, resembling the role of Reidemeister moves on knots. These moves are closely related to those defining flow equivalence for shift spaces, apart from the
the so-called \emph{Cuntz splice} which plays a special role and also fails to preserve the canonical diagonal Abelian subalgebra of the graph algebras (cf.\ \cite{MR3276420}, \cite{arXiv:1410.2308v1}).
We will largely approach the problem following the strategy from \cite{MR1340839} and \cite{MR2270572} to reduce the stable classification problem for graph $C^*$-algebras to questions concerning flow equivalence of shifts of finite type. To do so requires three new tools as listed below.
First and foremost, we need to know that the Cuntz splice leaves the $C^*$-algebras in question invariant up to stable isomorphism also in this generality. This we proved in \cite{arXiv:1602.03709v2}. Second, we need to develop the theory of a \emph{gauge invariant prime ideal space} which in our case will serve as a substitute for the standard primitive ideal space. The fact that this space is finite is key to our largely combinatorial approach throughout the paper, and we will equip it with a \emph{temperature map} to help us align the graphs so that the various types of gauge simple subquotients are matched. Finally, we introduce a procedure of \emph{plugging} and \emph{unplugging} sinks to pass between the cases allowing sinks and cases disallowing them, giving us the option to appeal to stronger general classification results in one case and a direct connection to symbolic dynamics in the other.
In the course of proving the above mentioned results, we extract and generalize from \cite{MR2270572} and \cite{MR1990568} some strong results concerning \GLP-equivalence\xspace and \SLP-equivalence\xspace, allowing us from the existence of certain such equivalences to deduce conclusions about the existence of move equivalences or Cuntz move equivalences between the graphs or about the existence of (stable) isomorphisms between the graph \mbox{$C\sp*$-}al\-ge\-bras\xspace, and \emph{vice versa}. This gives us some very concrete and hands-on tools to decide such questions.
In most cases, such as the one illustrated in Figure~\ref{firstexx}~(a), stable isomorphism
of the $C^*$-algebras associated to a pair of graphs allow for a
geometric realization by a finite number of moves, and we crystallize this out via the notion of \emph{Condition (H)} which we introduce here. In
sporadic cases failing this condition, such as the ones illustrated in Figure~\ref{firstexx}~(b) and~(c), we will
establish that no finite sequence of the moves defining the concept of Cuntz move equivalence can connect the two graphs
in each pair, even though the $K$-theoretical invariants of the
associated $C^*$-algebras are the same. In the case of (b), we may in fact
prove by appealing to \emph{ad hoc} classification results that the
$C^*$-algebras are stably isomorphic, proving that stable isomorphism of $C^*$-algebras is not always attainable via the moves hitherto studied.
Condition (H) generalizes Condition (K) and turns out to be met in a lot of other important special cases. When the graph $C^*$-algebras
defined are of type I/postliminary, our results may be refined further and lead to
the classification of a class of quantum lens spaces introduced and
studied by Hong and Szyma\'nski in \cite{MR2015735}. Moreover, specializing
to the graph $C^*$-algebras associated to simple graphs with four
vertices or less, we give a complete classification. These results have bearing on the Abrams-Tomforde conjecture (\cite{MR2775826}).
In forthcoming work
(\cite{Eilers-Restorff-Ruiz-Sorensen-2}) we will introduce a final new move and prove, among
many other things, that indeed all Cuntz-Krieger algebras are
classified by their $K$-theory, because any isomorphism at the level
of $K$-theory may be realized using an enlarged family of moves, all leaving the stabilized $C^*$-algebra invariant. The
present paper is self-contained and does not draw on the much more
complicated approach in \cite{Eilers-Restorff-Ruiz-Sorensen-2}. We will, however, develop basic results in the paper at hand in generality not needed here to anticipate applications in \cite{Eilers-Restorff-Ruiz-Sorensen-2}.
The paper is organized as follows. In Section \ref{genprel} we outline key concepts for the paper, mainly stemming from the theory of graph \mbox{$C\sp*$-}al\-ge\-bras\xspace, and discuss the moves that constitute our fundamental notion of Cuntz move equivalence. In Section \ref{gipis} we develop the idea of the gauge invariant prime ideal space, which is completely essential for everything that follows, and we connect this to $K$-theory, block matrices and partially ordered sets in Section \ref{sec:notation-for-proof}, introducing also the key notion of tempered ideal spaces.
In Section \ref{CC} we then prove a complete characterization of Cuntz move equivalence for finite graphs, drawing heavily on ideas from \cite{MR2270572} augmented with a trick of \emph{plugging} sinks which we also develop there. Section \ref{Ccas} contains our geometric classification theorem for finite graphs with Condition (H), as well as examples showing the necessity of this condition, and in Section \ref{applications} we detail the applications listed above.
\section{General preliminaries}\label{genprel}
In this section, we introduce notation and fundamental concepts concerning graphs and their $C^*$-algebras.
\subsection{\texorpdfstring{$C^*$}{C*}-algebras over topological spaces}
Let $X$ be a topological space satisfying the $T_0$ separation condition and let $\mathbb{O}( X)$ be the set of open subsets of $X$, partially ordered by set inclusion $\subseteq$.
A subset $Y$ of $X$ is called \emph{locally closed} if $Y = U \setminus V$ where $U, V \in \mathbb{O} ( X )$ and $V \subseteq U$.
The set of all locally closed subsets of $X$ will be denoted by $\mathbb{LC}(X)$.
The partially ordered set $( \mathbb{O} ( X ) , \subseteq )$ is a \emph{complete lattice}, that is, any subset $S$ of $\mathbb{O} (X)$ has both an infimum $\bigwedge S$ and a supremum $\bigvee S$, which are for any subset $S$ of $\mathbb{O} ( X )$ defined as
\begin{equation*}
\bigwedge_{ U \in S } U = \left( \bigcap_{ U \in S } U \right)^{\circ} \quad \mathrm{and} \quad \bigvee_{ U \in S } U = \bigcup_{ U \in S } U.
\end{equation*}
Note that if $S$ is empty, these are $X$ and $\emptyset$, respectively.
For a \mbox{$C\sp*$-}al\-ge\-bra\xspace \ensuremath{\mathfrak{A}}\xspace, let $\mathbb{I} ( \ensuremath{\mathfrak{A}}\xspace )$ be the set of closed ideals of \ensuremath{\mathfrak{A}}\xspace, partially ordered by $\subseteq$. The partially ordered set $( \mathbb{I} ( \ensuremath{\mathfrak{A}}\xspace ), \subseteq )$ is a complete lattice. More precisely, for any subset $S$ of $\mathbb{I} ( \ensuremath{\mathfrak{A}}\xspace )$,
\begin{equation*}
\bigwedge_{ \mathfrak{I} \in S } \mathfrak{I} = \bigcap_{ \mathfrak{I} \in S } \mathfrak{I} \quad \mathrm{and} \quad \bigvee_{ \mathfrak{I} \in S } \mathfrak{I} = \overline{ \sum_{ \mathfrak{I} \in S } \mathfrak{I} }.
\end{equation*}
\begin{definition}
Let \ensuremath{\mathfrak{A}}\xspace be a \mbox{$C\sp*$-}al\-ge\-bra\xspace. Let $\operatorname{Prim} ( \ensuremath{\mathfrak{A}}\xspace )$ denote the \emph{primitive ideal space} of \ensuremath{\mathfrak{A}}\xspace, equipped with the usual hull-kernel topology, also called the Jacobson topology.
Let $X$ be a topological space. A \emph{\mbox{$C\sp*$-}al\-ge\-bra\xspace over $X$} is a pair $( \ensuremath{\mathfrak{A}}\xspace
, \psi )$ consisting of a \mbox{$C\sp*$-}al\-ge\-bra\xspace \ensuremath{\mathfrak{A}}\xspace and a continuous map $\ftn{ \psi
}{ \operatorname{Prim} ( \ensuremath{\mathfrak{A}}\xspace ) }{ X }$. \end{definition}
We identify $\mathbb{O} ( \operatorname{Prim} ( \ensuremath{\mathfrak{A}}\xspace ) )$ and $\mathbb{I} ( \ensuremath{\mathfrak{A}}\xspace )$ using the lattice isomorphism
\begin{equation*}
U \mapsto \bigcap_{ \mathfrak{p} \in \operatorname{Prim} ( \ensuremath{\mathfrak{A}}\xspace ) \setminus U } \mathfrak{p}.
\end{equation*}
Let $( \ensuremath{\mathfrak{A}}\xspace , \psi )$ be a \mbox{$C\sp*$-}al\-ge\-bra\xspace over $X$. Then we get a map $\ftn{ \psi^{*} }{ \mathbb{O} ( X ) }{ \mathbb{O} ( \operatorname{Prim} ( \ensuremath{\mathfrak{A}}\xspace ) ) \cong \mathbb{I} ( \ensuremath{\mathfrak{A}}\xspace ) }$ defined by
\begin{equation*}
U \mapsto \setof{ \mathfrak{p} \in \operatorname{Prim} ( \ensuremath{\mathfrak{A}}\xspace ) }{ \psi ( \mathfrak{p} ) \in U }.
\end{equation*}
Using the isomorphism $\mathbb{O} ( \operatorname{Prim} ( \ensuremath{\mathfrak{A}}\xspace ) ) \cong \mathbb{I} ( \ensuremath{\mathfrak{A}}\xspace )$, we get a map from $\mathbb{O}(X)$ to $\mathbb{I}(\ensuremath{\mathfrak{A}}\xspace )$ by
\begin{align*}
U \mapsto \bigcap \setof{ \mathfrak{p} \in \operatorname{Prim} ( \ensuremath{\mathfrak{A}}\xspace ) }{ \psi ( \mathfrak{p} ) \notin U }.
\end{align*}
Denote this ideal by $\ensuremath{\mathfrak{A}}\xspace(U)$. For $Y = U \setminus V \in \mathbb{LC} ( X )$, set $\ensuremath{\mathfrak{A}}\xspace(Y) = \ensuremath{\mathfrak{A}}\xspace (U) / \ensuremath{\mathfrak{A}}\xspace(V)$. By \cite[Lemma~2.15]{MR2545613}, $\ensuremath{\mathfrak{A}}\xspace ( Y)$ does not depend (up to a canonical \mbox{$\sp*$-}iso\-morphism\xspace) on $U$ and $V$.
We can equivalently define an $X$-algebra by giving a map from $\mathbb{O}(X)$ to $\mathbb{O}(\operatorname{Prim}(\ensuremath{\mathfrak{A}}\xspace))$ that preserves finite infima and arbitrary suprema (so the empty set is mapped to the empty set, and $X$ is mapped to $\operatorname{Prim}(\ensuremath{\mathfrak{A}}\xspace)$).
\begin{example}
For any \mbox{$C\sp*$-}al\-ge\-bra\xspace \ensuremath{\mathfrak{A}}\xspace, the pair $( \ensuremath{\mathfrak{A}}\xspace , \ensuremath{\operatorname{id}}\xspace_{ \operatorname{Prim} ( \ensuremath{\mathfrak{A}}\xspace ) } )$ is a \mbox{$C\sp*$-}al\-ge\-bra\xspace over $\operatorname{Prim} ( \ensuremath{\mathfrak{A}}\xspace )$. For each $U \in \mathbb{O} ( \operatorname{Prim} ( \ensuremath{\mathfrak{A}}\xspace ) )$, the ideal $\ensuremath{\mathfrak{A}}\xspace ( U )$ equals $\bigcap_{ \mathfrak{p} \in \operatorname{Prim} ( \ensuremath{\mathfrak{A}}\xspace ) \setminus U } \mathfrak{p}$.
\end{example}
\begin{definition}
Let \ensuremath{\mathfrak{A}}\xspace and \ensuremath{\mathfrak{B}}\xspace be \mbox{$C\sp*$-}al\-ge\-bras\xspace over $X$.
A \mbox{$\sp*$-}ho\-mo\-morphism\xspace $\Phi\colon\ensuremath{\mathfrak{A}}\xspace\rightarrow\ensuremath{\mathfrak{B}}\xspace$ is \emph{$X$-equivariant} if $\Phi ( \ensuremath{\mathfrak{A}}\xspace (U) ) \subseteq \ensuremath{\mathfrak{B}}\xspace ( U )$ for all $U \in \mathbb{O}(X)$.
Hence, for every $Y = U \setminus V$, $\Phi$ induces a \mbox{$\sp*$-}ho\-mo\-morphism\xspace $\Phi_{Y}\colon\ensuremath{\mathfrak{A}}\xspace ( Y ) \rightarrow\ensuremath{\mathfrak{B}}\xspace (Y)$.
Let $\mathcal{C}_X$ be the category whose objects are \mbox{$C\sp*$-}al\-ge\-bras\xspace over $X$ and whose morphisms are $X$-equivariant homomorphisms.
\end{definition}
\subsection{Graphs and their matrices}
By a \emph{graph} we mean a directed graph. Formally:
\begin{definition}
A graph $E$ is a quadruple $E = (E^0 , E^1 , r, s)$ where $E^0$ and $E^1$ are sets, and $r$ and $s$ are maps from $E^1$ to $E^0$.
The elements of $E^0$ are called \emph{vertices}, the elements of $E^1$ are called \emph{edges}, the map $r$ is called the \emph{range map}, and the map $s$ is called the \emph{source map}.
\end{definition}
All graphs considered will be \emph{countable}, \emph{i.e.}\xspace, there are countably many vertices and edges.
We call a graph \emph{finite}, if there are only finitely many vertices and edges.
We will freely identify graphs up to graph isomorphism.
\begin{definition}
A \emph{loop} is an edge with the same range and source.
A \emph{path} $\mu$ in a graph is a finite sequence $\mu = e_1 e_2 \cdots e_n$ of edges satisfying
$r(e_i)=s(e_{i+1})$, for all $i=1,2,\ldots, n-1$, and we say that the \emph{length} of $\mu$ is $n$.
We extend the range and source maps to paths by letting $s(\mu) = s(e_1)$ and $r(\mu) = r(e_n)$.
Vertices in $E$ are regarded as \emph{paths of length $0$} (also called empty paths).
A \emph{cycle} is a nonempty path $\mu$ such that $s(\mu) = r(\mu)$.
We call a cycle $e_1e_2\cdots e_n$ a \emph{vertex-simple cycle} if $r(e_i)\neq r(e_j)$ for all $i\neq j$. A cycle $e_1e_2\cdots e_n$ is said to have an \emph{exit} if there exists an edge $f$ such that $s(f)=s(e_k)$ for some $k=1,2,\ldots,n$ with $e_k\neq f$.
A \emph{return path} is a cycle $\mu = e_1 e_2 \cdots e_n$ such that $r(e_i) \neq r(\mu)$ for $i < n$.
For a loop, cycle or return path, we say that it is \emph{based} at the source vertex of its path.
We also say that a vertex \emph{supports} a certain loop, cycle or return path if it is based at that vertex.
Note that in \cite{MR1988256,MR1989499,MR2023453,MR1914564}, the authors use the term \emph{loop} where we use \emph{cycle}.
\end{definition}
\begin{definition}
A vertex $v\in E^0$ in $E$ is called \emph{regular} if $s^{-1}(v) := \setof{ e \in E^1 }{ s(e) = v }$ is finite and nonempty. We denote the set of regular vertices by
$E_{\mathrm{reg}}^0$. We call the remaining vertices \emph{singular}
and write $E_{\mathrm{sing}}^0=E^0\setminus E_{\mathrm{reg}}^0$.
A vertex $v\in E^0$ in $E$ is called a \emph{source} if $r^{-1}(v) := \setof{ e \in E^1 }{ r(e) = v }$ is the empty set.
A vertex $v\in E^0$ in $E$ is called a \emph{sink} if $s^{-1}(v)=\emptyset$.
An \emph{isolated vertex} is both a sink and a source.
\end{definition}
\begin{definition}
Let $E$ be a graph.
We say that $E$ satisfies \emph{Condition~(K)} if for every vertex $v\in E^0$ in $E$, either there is no return path based at $v$ or there are at least two distinct return paths based at $v$.
\end{definition}
\begin{notation}
If there exists a path from vertex $u$ to vertex $v$, then we write $u \geq v$ --- this is a preorder on the vertex set, \emph{i.e.}\xspace, it is reflexive and transitive, but need not be antisymmetric.
\end{notation}
It is essential for our approach to graph \mbox{$C\sp*$-}al\-ge\-bras\xspace to be able to shift between a graph and its adjacency matrix.
In what follows, we let \ensuremath{\mathbb{N}}\xspace denote the set of positive integers, while $\ensuremath{\mathbb{N}}\xspace_0$ denotes the set of nonnegative integers.
\begin{definition}
Let $E = (E^0 , E^1 , r, s)$ be a graph.
We define its \emph{adjacency matrix} $\mathsf{A}_E$ as an $E^0\times E^0$ matrix with the $(u,v)$'th entry being
$$\left\vert\setof{e\in E^1}{s(e)=u, r(e)=v}\right\vert.$$
As we only consider countable graphs, $\mathsf{A}_E$ will be a finite matrix or a countably infinite matrix, and it will have entries from $\ensuremath{\mathbb{N}}\xspace_0\sqcup\{\infty\}$.
Let $X$ be a set.
If $A$ is an $X \times X$ matrix with entries from $\ensuremath{\mathbb{N}}\xspace_0\sqcup\{\infty\}$, we let $\mathsf{E}_{A}$ be the graph with vertex set $X$ such that between two vertices $x,x' \in X$ we have $A(x,x')$ edges.
\end{definition}
It will be convenient for us to alter the adjacency matrix of a graph in two very specific ways, removing singular rows and subtracting the identity, so we introduce notation for this.
\begin{notation}
Let $E$ be a graph and $\mathsf{A}_E$ its adjacency matrix.
Denote by $\mathsf{A}_{E}^\bullet$ the matrix obtained from $\mathsf{A}_{E}$ by removing all rows corresponding to singular vertices of $E$.
Let $\mathsf{B}_E$ denote the matrix $\mathsf{A}_{E} - I$, and let $\mathsf{B}_{E}^\bullet$ be $\mathsf{B}_E$ with the rows corresponding to singular vertices of $E$ removed.
\end{notation}
\subsection{Graph \texorpdfstring{$C^*$}{C*}-algebras}
We follow the notation and definition for graph \mbox{$C\sp*$-}al\-ge\-bras\xspace in \cite{MR1670363}; this is not the convention used in Raeburn's monograph \cite{MR2135030}.
\begin{definition} \label{def:graphca}
Let $E = (E^0,E^1,r,s)$ be a graph.
The \emph{graph \mbox{$C\sp*$-}al\-ge\-bra\xspace} $C^*(E)$ is defined as the universal \mbox{$C\sp*$-}al\-ge\-bra\xspace generated by
a set of mutually orthogonal projections $\setof{ p_v }{ v \in E^0 }$ and a set $\setof{ s_e }{ e \in E^1 }$ of partial isometries satisfying the relations
\begin{itemize}
\item $s_e^* s_f = 0$ if $e,f \in E^1$ and $e \neq f$,
\item $s_e^* s_e = p_{r(e)}$ for all $e \in E^1$,
\item $s_e s_e^* \leq p_{s(e)}$ for all $e \in E^1$, and,
\item $p_v = \sum_{e \in s^{-1}(v)} s_e s_e^*$ for all $v \in E^0$ with $0 < |s^{-1}(v)| < \infty$.
\end{itemize}
Whenever we have a set of mutually orthogonal projections $\setof{ p_v }{ v \in E^0 }$ and a set $\setof{ s_e }{ e \in E^1 }$ of partial isometries in a \mbox{$C\sp*$-}al\-ge\-bra\xspace satisfying the relations, then we call these elements a \emph{Cuntz-Krieger $E$-family}.
\end{definition}
It is clear from the definition that an isomorphism between graphs induces a canonical isomorphism between the corresponding graph \mbox{$C\sp*$-}al\-ge\-bras\xspace.
\begin{definition}
Let $E=(E^0,E^1,r,s)$ be a graph.
By universality there is a canonical gauge action $\gamma\colon\mathbb{T}\rightarrow\operatorname{Aut}(C^*(E))$ such that for any $z\in\mathbb{T}$, we have that $\gamma_z(p_v)=p_v$ for all $v\in E^0$ and $\gamma_z(s_e)=zs_e$ for all $e\in E^1$.
We say that an ideal $\mathfrak{I}$ of $C^*(E)$ is gauge invariant, if $\gamma_z(\mathfrak{I})\subseteq \mathfrak{I}$ for all $z\in\mathbb{T}$, and we let $\mathbb{I}_\gamma(C^*(E))$ denote the subset of $\mathbb{I}(C^*(E))$ consisting of gauge invariant ideals.
\end{definition}
It is clear that the lattice operations preserve the gauge invariance, so $\mathbb{I}_\gamma(C^*(E))$ is a sublattice. We collect some standard facts about graph \mbox{$C\sp*$-}al\-ge\-bras\xspace below.
\begin{remark}
Every graph \mbox{$C\sp*$-}al\-ge\-bra\xspace (of a countable graph) is separable, nuclear in the UCT class (\cite{MR1738948},\cite{MR2117597}).
A graph \mbox{$C\sp*$-}al\-ge\-bra\xspace is unital if and only if the corresponding graph has finitely many vertices.
A graph \mbox{$C\sp*$-}al\-ge\-bra\xspace is isomorphic to a Cuntz-Krieger algebra if and only if the corresponding graph is finite with no sinks, see~\cite[Theorem~3.12]{MR3391894}.
\end{remark}
\subsection{Moves on graphs}\label{sec:moves}
In this section we describe the moves on graphs we will allow.
\begin{definition}[Move \mbox{\texttt{\textup{(S)}}}\xspace: Remove a regular source]
Let $E = (E^0 , E^1 , r, s)$ be a graph, and let $w\in E^0$ be a source that is also a regular vertex.
Let $E_S$ denote the graph $(E_S^0 , E_S^1 , r_S , s_S )$ defined by
$$E_S^0 := E^0 \setminus \{w\}\quad
E_S^1 := E^1 \setminus s^{-1} (w)\quad
r_S := r|_{E_S^1}\quad
s_S := s|_{E_S^1}.$$
We call $E_S$ the \emph{graph obtained by removing the source $w$ from $E$}, and say $E_S$ is formed by performing Move \mbox{\texttt{\textup{(S)}}}\xspace to $E$.
\end{definition}
\begin{definition}[Move \mbox{\texttt{\textup{(R)}}}\xspace: Reduction at a regular vertex]
Suppose that $E = (E^0 , E^1 , r, s)$ is a graph, and let $w \in E^0$ be a regular vertex with the property that $s(r^{-1} (w)) = \{x\}$, $s^{-1} (w) = \{f \}$, and $r(f ) \neq w$.
Let $E_R$ denote the graph $(E_R^0, E_R^1, r_R , s_R )$ defined by
\begin{align*}
E_R^0&:= E^0 \setminus \{w\} \\
E_R^1&:= \left(E^1 \setminus (\{f \} \cup r^{-1}(w))\right) \cup \setof{e_f}{e \in E^1 \text{ and } r(e) = w} \\
r_R (e) &:= r(e)\text{ if }e \in E^1 \setminus (\{f \} \cup r^{-1}(w)) \quad\text{and}\quad r_R (e_f ) := r(f ) \\
s_R (e) &:= s(e)\text{ if }e \in E^1 \setminus (\{f \} \cup r^{-1}(w)) \quad\text{and}\quad s_R (e_f ) := s(e) = x.
\end{align*}
We call $E_R$ the \emph{graph obtained by reducing $E$ at $w$}, and say $E_R$ is a reduction
of $E$ or that $E_R$ is formed by performing Move \mbox{\texttt{\textup{(R)}}}\xspace to $E$.
\end{definition}
\begin{definition}[Move \mbox{\texttt{\textup{(O)}}}\xspace: Outsplit at a non-sink]
Let $E = (E^0 , E^1 , r, s)$ be a graph, and let $w \in E^0$ be a vertex that is not a sink.
Partition $s^{-1} (w)$ as a disjoint union of a finite number of nonempty sets
$$s^{-1}(w) = \mathcal{E}_1\sqcup \mathcal{E}_2\sqcup \cdots \sqcup\mathcal{E}_n$$
with the property that at most one of the $\mathcal{E}_i$ is infinite.
Let $E_O$ denote the graph $(E_O^0, E_O^1, r_O , s_O )$ defined by
\begin{align*}
E_O^0&:= \setof{v^1}{v \in E^0\text{ and }v \neq w}\cup\{w^1, \ldots, w^n\} \\
E_O^1&:= \setof{e^1}{e \in E^1\text{ and }r(e) \neq w}\cup \setof{e^1, \ldots , e^n}{e \in E^1\text{ and }r(e) = w} \\
r_{E_O} (e^i ) &:=
\begin{cases}
r(e)^1 & \text{if }e \in E^1\text{ and }r(e) \neq w\\
w^i & \text{if }e \in E^1\text{ and }r(e) = w
\end{cases} \\
s_{E_O} (e^i ) &:=
\begin{cases}
s(e)^1 & \text{if }e \in E^1\text{ and }s(e) \neq w \\
s(e)^j & \text{if }e \in E^1\text{ and }s(e) = w\text{ with }e \in \mathcal{E}_j.
\end{cases}
\end{align*}
We call $E_O$ the \emph{graph obtained by outsplitting $E$ at $w$}, and say $E_O$ is formed by
performing Move \mbox{\texttt{\textup{(O)}}}\xspace to $E$.
\end{definition}
\begin{definition}[Move \mbox{\texttt{\textup{(I)}}}\xspace: Insplit at a regular non-source]
Suppose that $E = (E^0 , E^1 , r, s)$ is a graph, and let $w \in E^0$ be a regular vertex that is not a source.
Partition $r^{-1} (w)$ as a disjoint union of a finite number of nonempty sets
$$r^{-1} (w) = \mathcal{E}_1\sqcup \mathcal{E}_2\cdots\sqcup\mathcal{E}_n.$$
Let $E_I$ denote the graph $(E_I^0 , E_I^1 , r_I , s_I )$ defined by
\begin{align*}
E_I^0 &:= \setof{v^1}{v \in E^0\text{ and }v \neq w} \cup \{w^1,\ldots, w^n \} \\
E_I^1 &:= \setof{e^1}{e \in E^1\text{ and }s(e) \neq w} \cup \setof{e^1, \ldots, e^n}{e \in E^1\text{ and }s(e) = w} \\
r_{E_I} (e^i ) &:=
\begin{cases}
r(e)^1 &\text{if }e \in E^1\text{ and }r(e) \neq w \\
r(e)^j &\text{if }e \in E^1\text{ and }r(e) = w\text{ with }e \in \mathcal{E}_j
\end{cases} \\
s_{E_I} (e^i ) &:=
\begin{cases}
s(e)^1 &\text{if }e \in E^1\text{ and }s(e) \neq w \\
w^i &\text{if }e \in E^1\text{ and }s(e) = w.
\end{cases}
\end{align*}
We call $E_I$ the \emph{graph obtained by insplitting $E$ at $w$}, and say $E_I$ is formed by performing Move \mbox{\texttt{\textup{(I)}}}\xspace to $E$.
\end{definition}
\begin{definition}[Move \mbox{\texttt{\textup{(C)}}}\xspace: Cuntz splicing] \label{def:cuntzsplice}
Let $E = (E^0 , E^1 , r , s )$ be a graph and let $v \in E^0$ be a regular vertex that supports at least two distinct return paths.
Let $E_C$ denote the graph $(E_C^0 , E_C^1 , r_C , s_C)$ defined by
\begin{align*}
E_C^0 &:= E^0\sqcup\{u_1 , u_2 \} \\
E_C^1 &:= E^1\sqcup\{e_1 , e_2 , f_1 , f_2 , h_1 , h_2 \},
\end{align*}
where $r_{C}$ and $s_{C}$ extend $r$ and $s$, respectively, and satisfy
$$s_{C} (e_1 ) = v,\quad s_{C} (e_2 ) = u_1 ,\quad s_{C} (f_i ) = u_1 ,\quad s_{C} (h_i ) = u_2 ,$$
and
$$r_{C} (e_1 ) = u_1 ,\quad r_{C} (e_2 ) = v,\quad r_{C} (f_i ) = u_i ,\quad r_{C} (h_i ) = u_i . $$
We call $E_C$ the \emph{graph obtained by Cuntz splicing $E$ at $v$}, and say $E_C$ is formed by performing Move \mbox{\texttt{\textup{(C)}}}\xspace to $E$.
\end{definition}
\begin{definition}\label{def:graph-equivalences}
The equivalence relation generated by the moves \mbox{\texttt{\textup{(O)}}}\xspace, \mbox{\texttt{\textup{(I)}}}\xspace, \mbox{\texttt{\textup{(R)}}}\xspace, \mbox{\texttt{\textup{(S)}}}\xspace together with graph isomorphism is called \emph{move equivalence}, and denoted \ensuremath{\sim_{M\negthinspace E}}\xspace.
The equivalence relation generated by the moves \mbox{\texttt{\textup{(O)}}}\xspace, \mbox{\texttt{\textup{(I)}}}\xspace, \mbox{\texttt{\textup{(R)}}}\xspace, \mbox{\texttt{\textup{(S)}}}\xspace, \mbox{\texttt{\textup{(C)}}}\xspace together with graph isomorphism is called \emph{Cuntz move equivalence}, and denoted \ensuremath{\sim_{C\negthinspace E}}\xspace.
\end{definition}
The following two theorems were essentially proved in \cite{MR2054048}, see also \cite[Propositions~3.1, 3.2 and 3.3 and Theorem~3.5]{MR3082546}.
\begin{theorem}[\cite{MR3082546}]\label{thm:moveimpliesstableisomorphism}
Let $E_1$ and $E_2$ be graphs such that $E_1\ensuremath{\sim_{M\negthinspace E}}\xspace E_2$.
Then $C\sp*(E_1)\otimes \ensuremath{\mathbb{K}}\xspace\cong C\sp*(E_2)\otimes \ensuremath{\mathbb{K}}\xspace$.
\end{theorem}
For the move \mbox{\texttt{\textup{(O)}}}\xspace, we actually obtain isomorphism rather than just stable isomorphism.
\begin{proposition}\label{prop:moveOimpliesisomorphism}
Let $E_1$ and $E_2$ be graphs such that one is obtained from the other using Move \mbox{\texttt{\textup{(O)}}}\xspace, then $C\sp*(E_1)\cong C\sp*(E_2)$.
\end{proposition}
For the move \mbox{\texttt{\textup{(C)}}}\xspace, it has recently been proved in \cite{arXiv:1602.03709v2} that it preserves the Morita equivalence class for arbitrary graphs.
\begin{theorem}[{\cite[Theorem~4.8]{arXiv:1602.03709v2}}]
\label{thm:cuntz-splice-implies-stable-isomorphism}
Let $E$ be a graph and let $v$ be a vertex that supports two distinct return paths. Then $C\sp*(E)\otimes\ensuremath{\mathbb{K}}\xspace\cong C\sp*(E_C)\otimes\ensuremath{\mathbb{K}}\xspace$.
\end{theorem}
We also extend the notation of equivalences to adjacency matrices.
\begin{definition}
If $A,A'$ are square matrices with entries in $\ensuremath{\mathbb{N}}\xspace_0\sqcup\{\infty\}$, we define them to be \emph{move equivalent}, and write $A \ensuremath{\sim_{M\negthinspace E}}\xspace A'$ if $\mathsf{E}_A \ensuremath{\sim_{M\negthinspace E}}\xspace \mathsf{E}_{A'}$.
We define \emph{Cuntz move equivalence} similarly.
\end{definition}
\begin{remark}
The Cuntz move equivalence, $\ensuremath{\sim_{C\negthinspace E}}\xspace$, is called \emph{move prime equivalence} in \cite{MR3082546}. Since the similarity of the two terms could create confusion, we have chosen to use the term \emph{Cuntz move equivalence} instead.
\end{remark}
\subsection{Derived moves}
We now discuss --- following and generalizing \cite[Section~5]{MR3082546} --- ways of changing the graphs without changing their move equivalence class. We will introduce a \emph{collapse} move, and present criteria allowing us to conclude that two graphs are move equivalent when one arises from the other by a row or column addition of the $\mathsf{B}$-matrices.
As we shall see, knowing move invariance of these derived moves dramatically simplifies working with $\ensuremath{\sim_{M\negthinspace E}}\xspace$.
\begin{definition}[Collapse a regular vertex that does not support a loop] \label{def:collapse}
Let $E = (E^0 , E^1 , r , s )$ be a graph and let $v$ be a regular vertex in $E$ that does not support a loop.
Define a graph $E_{COL}$ by
\begin{align*}
E_{COL}^0 &= E^0 \setminus \{v\}, \\
E_{COL}^1 &= E^1 \setminus (r^{-1}(v) \cup s^{-1}(v))
\sqcup \setof{[ef ]}{e \in r^{-1}(v)\text{ and }f \in s^{-1}(v)},
\end{align*}
the range and source maps extend those of $E$, and satisfy $r_{E_{COL}}([ef ]) = r (f )$
and $s_{E_{COL}} ([ef ]) = s (e)$.
\end{definition}
According to \cite[Theorem~5.2]{MR3082546} $E\ensuremath{\sim_{M\negthinspace E}}\xspace E_{COL}$ when $|E^0|<\infty$ --- in fact, the collapse move can be obtained using the moves \mbox{\texttt{\textup{(O)}}}\xspace and \mbox{\texttt{\textup{(R)}}}\xspace. We denote the move \mbox{\texttt{\textup{(Col)}}}\xspace.
Below, we will show how we can perform row and column additions on $\mathsf{B}_E$ without changing the move equivalence class of the associated graphs, when $E$ is a graph with finitely many vertices.
The setup we need is slightly different from what was considered in \cite[Section 7]{MR3082546} --- it was considered in \cite{arXiv:1602.03709v2}.
For the convenience of the reader, we collect the needed results from \cite{arXiv:1602.03709v2} in one proposition. Note that the definition of move equivalence in \cite{arXiv:1602.03709v2} is slightly different from the one above in order to be able to deal with graphs with infinitely many vertices --- but in the case of finitely many vertices they do in fact coincide.
\begin{proposition}[\cite{arXiv:1602.03709v2}]
\label{prop:matrix-moves}
Let $E = (E^0, E^1, r,s)$ be a graph with finitely many vertices.
Suppose $u,v \in E^0$ are distinct vertices with a path from $u$ to $v$.
Let $E_{u,v}$ be equal to the identity matrix except for on the $(u,v)$'th entry, where it is $1$.
Then $\mathsf{B}_E E_{u,v}$ is the matrix formed from $\mathsf{B}_E$ by adding the $u$'th column into the $v$'th column, while $E_{u,v}\mathsf{B}_E $ is the matrix formed from $\mathsf{B}_E$ by adding the $v$'th row into the $u$'th row.
Then the following holds.
\begin{enumerate}[(i)]
\item Suppose $u$ supports a loop or suppose that there is an edge from $u$ to $v$ and $u$ emits at least two edges. Then
\[
\mathsf{A}_E \ensuremath{\sim_{M\negthinspace E}}\xspace \mathsf{B}_E E_{u,v} + I.
\]\label{prop:matrix-moves:I}
\item Suppose $v$ is regular and either $v$ supports a loop or there is an edge from $u$ to $v$. Then
\[
\mathsf{A}_E \ensuremath{\sim_{M\negthinspace E}}\xspace E_{u,v}\mathsf{B}_E + I.
\]\label{prop:matrix-moves:II}
\end{enumerate}
\end{proposition}
\begin{remark} \label{rmk:columnAdd}
As in \cite{arXiv:1602.03709v2}, we can use the above proposition
backwards to subtract columns or rows in $\mathsf{B}_E$ as long as the
addition that undoes the subtraction is legal.
\end{remark}
Since legal row and column additions preserve $\ensuremath{\sim_{M\negthinspace E}}\xspace$, the resulting graph $C^*$-algebras will be stably isomorphic. The column addition in \ref{prop:matrix-moves:I} above rarely preserves the actual isomorphism class, but under modest additional assumptions, the row addition in \ref{prop:matrix-moves:II} does.
\begin{proposition}\label{p:isoaddingrows}
If condition \ref{prop:matrix-moves:II} in Proposition \ref{prop:matrix-moves} is met with $v$ regular, supporting a loop {and} an edge from $u$ to $v$, then $C^*(E)\cong C^*(\mathsf{E}_A)$, where
$A=E_{u,v}\mathsf{B}_E + I$.
\end{proposition}
\begin{proof}
Let $F=\mathsf{E}_A$ and denote the given edge from $u$ to $v$ in $E$ by $f$. Then $F$ is formed by removing $f$ but adding for each $e \in s_{E}^{-1} (v)$ an edge $\overline{e}$ with $s_{F} ( \overline{e} ) = u$ and $r_{F} ( \overline{e} ) = r_{E} ( e )$. Moreover, since $v$ supports a loop, we have that $r_{E}^{-1}( v ) \setminus \{ f \} \neq \emptyset$. Set $\mathcal{E}_{1} = r_{E}^{-1} (v) \setminus \{ f \}$ and $\mathcal{E}_{2} = \{ f \}$. Using this partition, we form $E_{{I}}$, which replaces $v$ with $v^{1}$ and $v^{2}$. The vertex $v^{1}$ receives the edges of $v$ except $f$ and also receives one edge from $v^{2}$ for each loop based at $v$. The vertex $v^{2}$ only receives the edge $f$. Both vertices emit copies of the edges $v$ emitted and do so in such a way that there is no loop based at $v^{2}$. By \cite[Proposition~3.6]{MR3045151}, there exists a \mbox{$\sp*$-}iso\-morphism\xspace $\ftn{\Psi_{1}}{ C^{*} (E) }{ p_{V} C^{*} ( E_{{I}} ) p_{V} }$ where $V = E_I^{0}\backslash\{v^2\}$.
Since $v^{2}$ does not support a loop, we may collapse this vertex, yielding $F$.
Set $q_{w} = p_{w}^{E}$ for all $w \in F^{0}$, $t_{e} = s_{e}^{E}$ for all $e\in E_I^1\setminus (r^{-1}_{E_I}(v^2)\cup s^{-1}_{E_I}(v^2))$ and $t_{[ee']}=s_e^Es_{e'}^E$ for $e\in r^{-1}_{E_I}(v^2)$ and $e'\in s^{-1}_{E_I}(v^2)$.
One can easily check that $\Psi_2 ( p_{v}^{F} ) = q_{v}$ and $\Psi_2 ( s_{e}^{F} ) = t_{e}$ provides a \mbox{$\sp*$-}iso\-morphism\xspace $\ftn{ \Psi_{2} }{ C^{*} (F) }{ p_{V} C^{*} ( E_{{I}} ) p_{V} }$. Hence, $\ftn{ \Phi = \Psi_{2}^{-1} \circ \Psi_{1} }{ C^{*} (E) }{ C^{*} ( F ) }$ is a \mbox{$\sp*$-}iso\-morphism\xspace. \end{proof}
\section{The gauge invariant prime ideal space}\label{gipis}
We now provide definitions and fundamental results concerning the gauge invariant prime ideal spaces of graph \mbox{$C\sp*$-}al\-ge\-bras\xspace. Although this is a very natural thing to do when we have the graph given, we are not aware of any place in the literature where this has been done only using the graph \mbox{$C\sp*$-}al\-ge\-bra\xspace and not the underlying graph. For the benefit of further applications elsewhere, we carry out the analysis in full generality.
\subsection{Structure of graph \texorpdfstring{$C^*$}{C*}-algebras}
It is important for us to view the graph \mbox{$C\sp*$-}al\-ge\-bras\xspace as $X$-algebras over a topological space $X$ that --- in general --- is different from the primitive ideal space. This is due to the fact that when there exist ideals that are not gauge invariant, then there are infinitely many ideals.
The space we choose to work with corresponds to the distinguished ideals being exactly the gauge invariant ideals. We show a \mbox{$C\sp*$-}al\-ge\-bra\xspace{ic} characterization of the gauge invariant ideals, and describe the space $X=\operatorname{Prime}_\gamma(C^*(E))$ in this subsection.
\begin{definition}
Let $E=(E^0,E^1,r,s)$ be a graph.
A subset $H\subseteq E^0$ is called \emph{hereditary} if whenever $v,w\in E^0$ with
$v\in H$ and $v\geq w$, then $w\in H$.
A subset $S\subseteq E^0$ is called \emph{saturated} if whenever $v\in E_{\mathrm{reg}}^0$ with $r(s^{-1}(v))\subseteq S$, then $v\in S$.
For any saturated hereditary subset $H$, the \emph{breaking vertices} corresponding to $H$ are the elements of the set
$$B_H :=\setof{v\in E^0}{|s^{-1}(v)|=\infty\text{ and } 0<|s^{-1}(v)\cap r^{-1}(E^0\setminus H)|<\infty }.$$
It is clear that $\emptyset$ and $E^0$ are both saturated and hereditary subsets. The intersection of any family of hereditary subsets is again hereditary.
Thus, for every subset $S\subseteq E^0$, there exists a smallest hereditary subset of $E^0$ containing $S$ --- this set is called the hereditary subset generated by $S$ and is denoted $H(S)$.
The intersection of any family of saturated subsets is again saturated.
Thus, for every subset $S\subseteq E^0$, there is a smallest saturated subset of $E^0$ containing $S$ --- this set is called the saturation of $S$ and is denoted $\overline{S}$.
The saturation of a hereditary set is again hereditary.
It is also clear that the union of any family of hereditary sets is again hereditary.
This makes the set of saturated hereditary subsets of $E^0$ into a complete lattice.
An \emph{admissible pair} $(H,S)$ consists of a saturated hereditary subset $H\subseteq E^0$ and a subset $S\subseteq B_H$.
We order the collection of admissible pairs by defining $(H,S)\leq (H',S')$ if and only if $H\subseteq H'$ and $S\subseteq H'\cup S'$.
This makes the collection of admissible pairs into a lattice.
\end{definition}
\begin{fact}\label{fact:structure-1}
Let $E=(E^0,E^1,r,s)$ be a graph.
For any admissible pair $(H,S)$, we let $\mathfrak{J}_{(H,S)}$ denote the ideal generated by
$$\setof{p_v}{v\in H}\cup\setof{p_{v_0}^{H}}{v_0\in S},$$
where $p_{v_0}^H$ is the \emph{gap projection}
$$p_{v_0}^H=p_{v_0}-\sum_{\substack{s(e)=v_0 \\ r(e)\not\in H}}s_es_e^*.$$
If $B_H=\emptyset$, for a saturated hereditary subset $H\subseteq E^0$, then we write $\mathfrak{J}_H$ for $\mathfrak{J}_{(H,\emptyset)}$.
The map $(H,S)\mapsto \mathfrak{J}_{(H,S)}$ is a lattice isomorphism between the lattice of admissible pairs and the lattice of gauge invariant ideals of $C^*(E)$ (\emph{cf.}\xspace\ \cite[Theorem~3.6]{MR1988256}).
\end{fact}
\begin{lemma}\label{lem:structure-1}
Let $E=(E^0,E^1,r_E,s_E)$ and $F=(F^0,F^1,r_F,s_F)$ be graphs and let $\mathfrak{I}$ be an ideal of $C^* (E)$. Then $\mathfrak{I}$ is gauge-invariant if and only if $\mathfrak{I}$ is generated by projections. Consequently, every \mbox{$\sp*$-}iso\-morphism\xspace from $C^*(E)$ to $C^*(F)$ will send gauge invariant ideals to gauge invariant ideals and every \mbox{$\sp*$-}iso\-morphism\xspace from $C^*(E)\otimes\ensuremath{\mathbb{K}}\xspace$ to $C^*(F)\otimes\ensuremath{\mathbb{K}}\xspace$ will send gauge invariant ideals to gauge invariant ideals under the identification of the ideal lattice of $C^*(E)$ and $C^*(F)$ with the ideal lattice of $C^*(E)\otimes\ensuremath{\mathbb{K}}\xspace$ and $C^*(F)\otimes\ensuremath{\mathbb{K}}\xspace$, respectively.
\end{lemma}
\begin{proof}
Suppose $\mathfrak{I}$ is a gauge-invariant ideal. Then by Fact~\ref{fact:structure-1}, $\mathfrak{I}$ is generated by vertex projections and gap projections. Suppose $\mathfrak{I}$ is generated by projections $S = \{ p_1, p_2, \dots \}$. By \cite[Theorem~3.4 and Corollary~3.5]{MR3310950}, each $p_i$ is Murray-von~Neumann equivalent to sums of vertex projections and gap projections in $C^*(E)$, where the Murray-von~Neumann equivalence and sums are in $C^*( E) \otimes \ensuremath{\mathbb{K}}\xspace$.
But this implies that $\overline{ C^* (E) p_i C^* (E) }$ is generated by vertex projections and gap projections.
Hence, $\mathfrak{I} = \overline{\operatorname{span} C^* (E) S C^*(E) }$ is generated by vertex projections and gap projections. Since vertex projections and gap projections are fixed by the gauge action, we have that $\mathfrak{I}$ is a gauge-invariant ideal.
Suppose $\Phi \colon C^* (E) \rightarrow C^* (F)$ is a \mbox{$\sp*$-}iso\-morphism\xspace. Let $\mathfrak{I}$ be a gauge-invariant ideal of $C^* (E)$. Then from the first part of the lemma, we have that $\mathfrak{I}$ is generated by projections. Since $\Phi$ is a \mbox{$\sp*$-}iso\-morphism\xspace, we have that $\Phi ( \mathfrak{I} )$ is also generated by projections. Thus, $\Phi ( \mathfrak{I} )$ is a gauge-invariant ideal.
For a \mbox{$C\sp*$-}al\-ge\-bra\xspace \ensuremath{\mathfrak{A}}\xspace, we say an ideal in $\ensuremath{\mathfrak{A}}\xspace \otimes \ensuremath{\mathbb{K}}\xspace$ is generated by projections in \ensuremath{\mathfrak{A}}\xspace if it is generated by projections in $\mathfrak{A} \otimes e_{11}$.
Suppose $\Psi \colon C^* (E) \otimes \ensuremath{\mathbb{K}}\xspace \rightarrow C^* (F) \otimes \ensuremath{\mathbb{K}}\xspace$ is a \mbox{$\sp*$-}iso\-morphism\xspace.
Let $\mathfrak{I}$ be a gauge-invariant ideal of $C^* (E) \otimes \ensuremath{\mathbb{K}}\xspace$, \emph{i.e.}\xspace, $\mathfrak{I} = \mathfrak{J}_{ (H,S) } \otimes \ensuremath{\mathbb{K}}\xspace$.
So, in particular, $\mathfrak{I}$ is generated by projections.
Since $\Psi$ is a \mbox{$\sp*$-}iso\-morphism\xspace, $\Psi ( \mathfrak{I} )$ is generated by projections.
By \cite[Theorem~3.4 and Corollary~3.5]{MR3310950} and using a similar argument as in the first paragraph, we get that $\Psi ( \mathfrak{I} )$ is generated by vertex projections and gap projections in $C^* (F)$.
Hence, $\Psi ( \mathfrak{I} )$ is gauge-invariant.
\end{proof}
\begin{definition}
Let $E = (E^0 , E^1 , r , s )$ be a graph.
Let $\operatorname{Prime}_\gamma(C^*(E))$ denote the set of all proper ideals that are prime within the set of proper gauge invariant ideals, \emph{i.e.}\xspace, $\mathfrak{p}\in\operatorname{Prime}_\gamma(C^*(E))$ if and only if $\mathfrak{p}$ is a proper gauge invariant ideal of $C^*(E)$ and
$$\mathfrak{I}_1\mathfrak{I}_2\subseteq\mathfrak{p}
\Rightarrow\mathfrak{I}_1\subseteq\mathfrak{p}\vee\mathfrak{I}_2\subseteq\mathfrak{p},$$
for all (proper) gauge invariant ideals $\mathfrak{I}_1,\mathfrak{I}_2$ of $C^*(E)$.
Recall that for an ideal $\mathfrak{I}$, we let $\operatorname{hull}(\mathfrak{I})$ denote the set of primitive ideals containing $\mathfrak{I}$, \emph{i.e.}\xspace, $\setof{\mathfrak{p} \in\operatorname{Prim}(C^*(E))}{\mathfrak{p}\supseteq\mathfrak{I}}$.
Similarly, for every ideal $\mathfrak{I}$, we let $\operatorname{hull}_\gamma(\mathfrak{I})$ denote the set $\setof{\mathfrak{p} \in\operatorname{Prime}_\gamma(C^*(E))}{\mathfrak{p}\supseteq\mathfrak{I}}$.
We equip $\operatorname{Prime}_\gamma(C^*(E))$ with a topology similar to the hull-kernel topology for primitive ideals, \emph{i.e.}\xspace, the closure of a subset $S\subseteq\operatorname{Prime}_\gamma(C^*(E))$ is
$$\operatorname{hull}_\gamma(\cap S)=\setof{\mathfrak{p}\in\operatorname{Prime}_\gamma(C^*(E))}{\mathfrak{p}\supseteq\cap S}.$$
To check that this closure operation defines a unique topology, we need only to check that it satisfies the four Kuratowski closure axioms --- but the first two paragraphs of \cite[5.4.6~Theorem]{MR1074574} show this.
With an argument similar to \cite[5.4.7~Theorem]{MR1074574}, it also follows that the topology is $T_0$.
\end{definition}
When $\fct{\Phi}{C^*(E)}{C^*(F)}$ is a \mbox{$\sp*$-}iso\-morphism\xspace, we get by Lemma \ref{lem:structure-1} an induced homeomorphism $\fct{\Phi_\sharp}{\operatorname{Prime}_\gamma(C^*(E))}{\operatorname{Prime}_\gamma(C^*(F))}$.
It is an elementary fact, that every primitive ideal of a \mbox{$C\sp*$-}al\-ge\-bra\xspace is a
(closed) prime ideal (\emph{e.g.}\xspace\ \cite[5.4.5~Theorem]{MR1074574}). For a
separable \mbox{$C\sp*$-}al\-ge\-bra\xspace, the converse is true, which can be seen by showing that
the primitive ideal space of a separable \mbox{$C\sp*$-}al\-ge\-bra\xspace is a Baire space
(\emph{e.g.}\xspace\ \cite[II.6.5.15~Corollary]{MR2188261}), but as shown by Weaver in \cite{MR2003352}
the concepts differ for nonseparable $C^*$-algebras. In fact, there are counterexamples even for nonseparable graph $C^*$-algebras
(see \cite{MR3426227}), but since we only consider countable graphs, this will not be an issue here.
\begin{lemma}\label{lem:prime-ideals-1}
Let $E = (E^0 , E^1 , r , s )$ be a graph.
Every primitive gauge invariant ideal of $C^*(E)$ is in $\operatorname{Prime}_\gamma(C^*(E))$.
Every primitive ideal of $C^*(E)$ that is not gauge invariant has a largest gauge invariant ideal as a subset, and this gauge invariant ideal is in $\operatorname{Prime}_\gamma(C^*(E))$.
If $\mathfrak{I}$ is a proper gauge invariant ideal of $C^*(E)$, then
$$\mathfrak{I}=\cap\setof{\mathfrak{p}\in\operatorname{Prime}_\gamma(C^*(E))}{\mathfrak{p}\supseteq \mathfrak{I}}=\cap\operatorname{hull}_\gamma(\mathfrak{I}).$$
\end{lemma}
\begin{proof}
First, note that all primitive ideals of $C^*(E)$ are described in \cite[Corollary~2.11]{MR2023453} --- we will use the terminology from there. As pointed out in \cite{MR3142035} there is a minor mistake in the description of the topology of the primitive ideal space in \cite{MR2023453}, but this has no consequences for this paper, since we are not using the description of the topology.
So we have a bijection from $\mathcal{M}_\gamma(E)\sqcup BV(E)\sqcup(\mathcal{M}_\tau(E)\times\mathbb{T})$ to $\operatorname{Prim}(C^*(E))$ given by
\begin{align*}
\mathcal{M}_\gamma(E)\ni M&\mapsto \mathfrak{J}_{\Omega(M),\Omega(M)_\infty^\mathrm{fin}}, \\
BV(E)\ni v&\mapsto \mathfrak{J}_{\Omega(v),\Omega(v)_\infty^\mathrm{fin}\setminus\{v\}}, \\
\mathcal{M}_\tau(E)\times\mathbb{T}\ni (N,z)&\mapsto \mathfrak{R}_{N,t},
\end{align*}
where the gauge invariant primitive ideals are exactly the ideals coming from $\mathcal{M}_\gamma(E)$ and $BV(E)$.
Note that every gauge invariant primitive ideal of $C^*(E)$ is also prime in the set of ideals of $C^*(E)$.
Thus every gauge invariant primitive ideal of $C^*(E)$ is in $\operatorname{Prime}_\gamma(C^*(E))$.
Note that for $N\in\mathcal{M}_\tau(E)$ and $z\in\mathbb{T}$, the ideal $\mathfrak{J}_{\Omega(N),\Omega(N)_\infty^\mathrm{fin}}$ is the largest gauge invariant ideal contained in $R_{N,z}$ (\emph{cf.}\xspace\ \cite[Lemma~2.6]{MR2023453}).
Let $N\in\mathcal{M}_\tau(E)$ and assume that $\mathfrak{I}_1\mathfrak{I}_2\subseteq\mathfrak{J}_{\Omega(N),\Omega(N)_\infty^\mathrm{fin}}$ for some gauge invariant ideals $\mathfrak{I}_1,\mathfrak{I}_2$.
Then $\mathfrak{I}_1\mathfrak{I}_2\subseteq\mathfrak{R}_{N,-1}$.
Since $\mathfrak{R}_{N,-1}$ is a primitive ideal in $C^*(E)$, it is prime in the collection of all ideals of $C^*(E)$. Therefore either $\mathfrak{I}_1\subseteq\mathfrak{R}_{N,-1}$ or $\mathfrak{I}_2\subseteq\mathfrak{R}_{N,-1}$.
But since $\mathfrak{J}_{\Omega(N),\Omega(N)_\infty^\mathrm{fin}}$ is the largest gauge invariant ideal contained in $\mathfrak{R}_{N,-1}$, we have $\mathfrak{I}_1\subseteq\mathfrak{J}_{\Omega(N),\Omega(N)_\infty^\mathrm{fin}}$ or
$\mathfrak{I}_2\subseteq\mathfrak{J}_{\Omega(N),\Omega(N)_\infty^\mathrm{fin}}$.
This shows that also $\mathfrak{J}_{\Omega(N),\Omega(N)_\infty^\mathrm{fin}}\in\operatorname{Prime}_\gamma(C^*(E))$ when $N\in\mathcal{M}_\tau(E)$.
Let $\mathfrak{I}$ be a proper gauge invariant ideal of $C^*(E)$.
Then $\mathfrak{I}$ is the intersection of all the primitive ideals containing it.
The only primitive ideals that are not in $\operatorname{Prime}_\gamma(C^*(E))$ are the ideals $\mathfrak{R}_{N,z}$ for $N\in\mathcal{M}_\tau(E)$ and $z\in\mathbb{T}$ --- but if $\mathfrak{I}\subseteq\mathfrak{R}_{N,z}$ then we can replace it in the intersection by the ideal $\mathfrak{J}_{\Omega(N),\Omega(N)_\infty^\mathrm{fin}}\in\operatorname{Prime}_\gamma(C^*(E))$.
So we have shown that
\begin{align*}
\mathfrak{I}&=\bigcap\Big(\setof{\mathfrak{J}_{\Omega(M),\Omega(M)_\infty^{\mathrm{fin}}}}{\mathfrak{J}_{\Omega(M),\Omega(M)_\infty^{\mathrm{fin}}}\supseteq \mathfrak{I},M\in\mathcal{M}_\gamma(E)\cup\mathcal{M}_\tau(E)} \\
&\qquad\qquad\cup\setof{\mathfrak{J}_{\Omega(v),\Omega(v)_\infty^{\mathrm{fin}}\setminus\{v\}}}{\mathfrak{J}_{\Omega(v),\Omega(v)_\infty^{\mathrm{fin}}\setminus\{v\}}\supseteq \mathfrak{I},v\in BV(E)}\Big) \\
&= \bigcap_{\substack{\mathfrak{p}\in\operatorname{Prim}_\gamma(C^*(E)) \\ \mathfrak{p}\supseteq\mathfrak{I}}}\mathfrak{p}
=\cap\operatorname{hull}_\gamma(\mathfrak{I}),
\end{align*}
since the second intersection contains all the sets from the first intersection.
\end{proof}
\begin{lemma}\label{lem:orderreversingprime}
The map
$$\mathfrak{I}\mapsto\operatorname{hull}_\gamma(\mathfrak{I})
=\setof{\mathfrak{p}\in\operatorname{Prime}_\gamma(C^*(E))}{\mathfrak{p}\supseteq\mathfrak{I}}$$
is an order-reversing $1-1$ correspondence between the gauge invariant ideals of $C^*(E)$ and the closed subsets of $\operatorname{Prime}_\gamma(C^*(E))$.
Its inverse map is $S\mapsto \cap S$.
\end{lemma}
\begin{proof}
This proof follows the lines of the proof of \cite[5.4.7~Theorem]{MR1074574}.
\end{proof}
The following lemma tells us exactly how we may consider a graph \mbox{$C\sp*$-}al\-ge\-bra\xspace as an algebra over $\operatorname{Prime}_\gamma(C^*(E))$ such that the distinguished ideals are exactly the gauge invariant ideals.
\begin{lemma}\label{lem:Xaction}
Let $E$ be a graph.
Consider the map $\zeta$ from $\operatorname{Prim}(C^{*}(E))$ to $\operatorname{Prime}_\gamma(C^{*}(E))$ sending each primitive ideal to the largest element of $\operatorname{Prime}_\gamma(C^{*}(E))$ that it contains.
This map is continuous and surjective, and it makes $C^{*}(E)$ into a $\operatorname{Prime}_\gamma(C^{*}(E))$-algebra in a canonical way. Moreover,
\begin{equation}\label{eq:lem:Xaction}
\zeta^{-1}(\operatorname{hull}_\gamma(\mathfrak{I}))
=\operatorname{hull}(\mathfrak{I}),
\end{equation}
for every gauge invariant ideal $\mathfrak{I}$ of $C^*(E)$, so the distinguished ideals under the action are exactly the gauge invariant ideals.
\end{lemma}
\begin{proof}
The validity of the definition of the map $\zeta$ follows from Lemma~\ref{lem:prime-ideals-1}.
First we show \eqref{eq:lem:Xaction}. Then continuity follows since every closed set of $\operatorname{Prime}_\gamma(C^*(E))$ is of the form $\operatorname{hull}_\gamma(\mathfrak{I})$. So let $\mathfrak{I}$ be a gauge invariant ideal.
Let $\mathfrak{p}\in \zeta^{-1}(\operatorname{hull}_\gamma(\mathfrak{I}))$.
Then $\zeta(\mathfrak{p})\supseteq\mathfrak{I}$.
Since, by definition, $\mathfrak{p}\supseteq\zeta(\mathfrak{p})$, it is clear that $\mathfrak{p}\supseteq\mathfrak{I}$. Therefore $\mathfrak{p}\in \operatorname{hull}(\mathfrak{I})$.
Now let $\mathfrak{p}\in \operatorname{hull}(\mathfrak{I})$.
Then $\mathfrak{p}\supseteq\mathfrak{I}$.
If $\mathfrak{p}$ is gauge invariant, then $\zeta(\mathfrak{p})=\mathfrak{p}\supseteq\mathfrak{I}$, so $\mathfrak{p}\in \zeta^{-1}(\operatorname{hull}_\gamma(\mathfrak{I}))$.
If, on the other hand, $\mathfrak{p}$ is not gauge invariant, then $\zeta(\mathfrak{p})$ is the largest gauge invariant ideal contained in $\mathfrak{p}$, \emph{cf.}\xspace~Lemma~\ref{lem:prime-ideals-1}.
Thus $\mathfrak{p}\supseteq\zeta(\mathfrak{p})\supseteq\mathfrak{I}$, so also in this case $\mathfrak{p}\in \zeta^{-1}(\operatorname{hull}_\gamma(\mathfrak{I}))$.
Now we want to show surjectivity of the map. For this we use the notation of \cite{MR1988256} and \cite{MR2023453} and the content of the proof of Lemma~\ref{lem:prime-ideals-1}.
Recall that every gauge invariant ideal $\mathfrak{I}$ of $C^*(E)$ is of the form $\mathfrak{I}=\mathfrak{J}_{H,B}$ for some saturated hereditary subset $H \subseteq E^0$ and some subset $B\subseteq B_H=H_\infty^\textrm{fin}$ --- in fact, if $H_\mathfrak{I}=\setof{v\in E^0}{p_v\in \mathfrak{I}}$ and $B_\mathfrak{I} = \setof{v\in B_{H_\mathfrak{I}}}{p_v^{H_\mathfrak{I}}\in\mathfrak{I} }$, then $\mathfrak{I}=\mathfrak{J}_{H_\mathfrak{I},B_\mathfrak{I}}$.
Note that if $(H,S_1)$ and $(H,S_2)$ are admissible pairs, then $(H,S_1)\wedge(H,S_2)$ is $(H,S_1\cap S_2)$.
Now assume that $\mathfrak{I}\in\operatorname{Prime}_\gamma(C^*(E))$, so $\mathfrak{I}=\mathfrak{J}_{H_\mathfrak{I},B_\mathfrak{I}}$.
Since $\mathfrak{I}$ is a proper ideal, $H_\mathfrak{I}\neq E^0$, so $M=E^0\setminus H_\mathfrak{I}$ is nonempty.
The proof of \cite[Lemma~4.1]{MR1988256} shows that $M$ is a maximal tail.
Note that $\Omega(M)=E^0\setminus M=H_\mathfrak{I}$.
We want to show that $|B_{H_\mathfrak{I}}\setminus B_\mathfrak{I}|\leq 1$.
So assume that $v_1,v_2\in B_{H_\mathfrak{I}}\setminus B_\mathfrak{I}$ with $v_1\neq v_2$.
It follows from \cite[Proposition~3.9]{MR1988256} that
$$\mathfrak{J}_{H_\mathfrak{I},B_\mathfrak{I}\cup\{v_1\}}\cap \mathfrak{J}_{H_\mathfrak{I},B_\mathfrak{I}\cup\{v_2\}}
=\mathfrak{J}_{H_\mathfrak{I},B_\mathfrak{I}}=\mathfrak{I}.$$
But $\mathfrak{J}_{H_\mathfrak{I},B_\mathfrak{I}\cup\{v_i\}} \not\subseteq\mathfrak{J}_{H_\mathfrak{I},B_\mathfrak{I}}=\mathfrak{I}$, for $i=1,2$, which contradicts that $\mathfrak{I}$ is prime within the proper gauge invariant ideals of $C^*(E)$. Hence $|B_{H_\mathfrak{I}}\setminus B_\mathfrak{I}|\leq 1$.
Now assume that $B_{H_\mathfrak{I}}\setminus B_\mathfrak{I}=\{v\}$.
We want to show that $v\in BV(E)$, \emph{i.e.}\xspace, we need to show that $v$ supports a cycle.
So assume that $v$ does not support a cycle. Since $v$ is an infinite emitter, $H_2=\overline{H(v)\setminus\{v\}}$ is a saturated hereditary subset not containing $v$.
Note that $v\not\in B_{H_2}$.
From \cite[Proposition~3.9]{MR1988256}, it follows that
$$\mathfrak{J}_{H_\mathfrak{I},B_{H_\mathfrak{I}}}
\cap\mathfrak{J}_{H_2,B_{H_2}}
\subseteq\mathfrak{J}_{H_\mathfrak{I},B_\mathfrak{I}}=\mathfrak{I}.$$
But $\mathfrak{J}_{H_\mathfrak{I},B_{H_\mathfrak{I}}}\not\subseteq\mathfrak{I}$ and $\mathfrak{J}_{H_2,B_{H_2}}\not\subseteq\mathfrak{I}$, which contradicts that $\mathfrak{I}$ is prime within the proper gauge invariant ideals of $C^*(E)$. Hence $v\in BV(E)$.
Now we also want to show that $\Omega(v)=H_\mathfrak{I}$.
From the definition, it is clear that $\Omega(v)\supseteq H_\mathfrak{I}$.
From \cite[Proposition~3.9]{MR1988256}, it follows that
$$\mathfrak{J}_{\Omega(v),B_{\Omega(v)}\setminus \{v\}}
\cap\mathfrak{J}_{H_\mathfrak{I},B_{H_\mathfrak{I}}}
\subseteq\mathfrak{J}_{H_\mathfrak{I},B_\mathfrak{I}}=\mathfrak{I}.$$
Since $\mathfrak{J}_{H_\mathfrak{I},B_{H_\mathfrak{I}}}\not \subseteq \mathfrak{I}$ and $\mathfrak{I}$ is prime within the proper gauge invariant ideals of $C^*(E)$, it follows that $\mathfrak{J}_{\Omega(v),B_{\Omega(v)}\setminus \{v\}}\subseteq\mathfrak{I}$.
Therefore $\Omega(v)\subseteq H_\mathfrak{I}$.
Now it follows from the proof of Lemma~\ref{lem:prime-ideals-1}, that $\zeta$ is surjective.
\end{proof}
\begin{remark}
Assume that $E$ is a graph with finitely many vertices. Then $E$ satisfies Condition~(K) if and only if $C^*(E)$ has finitely many ideals, and in this case
$\operatorname{Prim}(C^*(E))=\operatorname{Prime}_\gamma(C^*(E))$.
\end{remark}
\subsection{The component poset}
For our purposes, it will be essential to work with block matrices in a way that resembles the ideal structure and the filtered $K$-theory of the graph \mbox{$C\sp*$-}al\-ge\-bras\xspace.
To do this, we need to put the graph in a certain form and to order the vertices in a certain way such that the adjacency matrix has a certain nice block form.
It is also essential to our work, that the topological space $\operatorname{Prime}_\gamma(C^*(E))$ is built into this construction.
For the benefit of possible applications to other settings, we will allow infinite emitters, but it is essential for the exposition that we allow only \emph{finitely many vertices}.
As we shall see, it will be necessary to modify the given graph up to move equivalence to deal with certain complication introduced by transitional and breaking vertices. This will not change the $C^*$-algebras in question up to stable isomorphism, and is hence unproblematic for the work in this paper. But to pave the way for classification of the graph $C^*$-algebras themselves, we keep track of the isomorphism class as far as possible.
\begin{definition}\label{def:structure-a}
Let $E=(E^0,E^1,r,s)$ be a graph with finitely many vertices.
We say that a nonempty subset $S$ of $E^{0}$ is \emph{strongly connected} if for any two vertices $v,w\in S$ there exists a nonempty path from $v$ to $w$.
In particular every vertex in a strongly connected set has to be the base of a cycle.
The maximal strongly connected subsets of $E^{0}$ are all disjoint, and these are called the strongly connected components of $E$.
We let $\Gamma_E$ denote the set of all strongly connected components together with all singletons consisting of singular vertices that are not the base point of a cycle.
The sets in $\Gamma_E$ are all disjoint.
We call the sets in $\Gamma_E$ the \emph{components of the graph} $E$ and the vertices in $E^{0}\setminus\cup\Gamma_E$ the \emph{transition states} of $E$ --- the transition states are by definition all the regular vertices that are not the base point of a cycle. Note that with this terminology, all regular sources are also transition states. A strongly connected component is called a \emph{cyclic component} if one of its vertices (and thus all of its vertices) has exactly one return path.
We define a relation $\geq$ on $\Gamma_E$ by saying that $\gamma_1\geq\gamma_2$ if there exist vertices $v_1\in\gamma_1$ and $v_2\in\gamma_2$ such that $v_1\geq v_2$.
By definition this is the same as for all vertices $v_1\in\gamma_1$ and all vertices $v_2\in\gamma_2$ we have that $v_1\geq v_2$.
Thus it is clear that $\geq$ is a partial order.
We say that a subset $\sigma\subseteq\Gamma_E$ is hereditary if whenever $\gamma_1,\gamma_2\in \Gamma_E$ with $\gamma_1\in \sigma$ and $\gamma_1\geq\gamma_2$, then $\gamma_2\in \sigma$.
We equip $\Gamma_E$ with the topology that has the hereditary subsets as open sets --- this makes $\Gamma_E$ into a $T_0$-space.
For every subset $\sigma\subseteq\Gamma_E$, we let $\eta(\sigma)$ denote the smallest hereditary subset of $\Gamma_E$ containing $\sigma$, \emph{i.e.}\xspace, the set $\setof{\gamma\in\Gamma_E}{\exists \gamma'\in\sigma\colon\gamma'\geq\gamma}$.
\end{definition}
We recall the definition of an Alexandrov space and some of their properties.
\begin{definition}\label{def:structure-b}
A topological space is called an \emph{Alexandrov space} if arbitrary intersections of open subsets are again open. If we have a topological space $X$, then we can define a preorder on $X$ by $x\geq y$ if and only if $x$ is in the closure of $\{y\}$ --- this preorder is called the \emph{specialization preorder}. In the opposite direction, for a preordered set $(X,\geq)$ we can let the sets $F\subseteq X$ satisfying $x\geq y\wedge y\in F\Rightarrow x\in F$ be the closed sets. This topology is the finest topology satisfying that $x\geq y$ if and only if $x$ is in the closure of $\{y\}$. It is also clear that this is an Alexandrov topology.
If an Alexandrov space is given, and we take its specialization preorder, then the Alexandrov topology is uniquely determined from the specialization preorder by the above construction. Thus there is a one-to-one correspondence between Alexandrov topologies and preorders on a space. A map between two Alexandrov spaces is continuous if and only if it is an order preserving map with respect to the specialization preorders.
\end{definition}
Note that often the specialization preorder is written as the opposite order compared to above. Both conventions are used in the literature, while the convention used here is chosen since it fits better with our setup, as we will see now.
\begin{remark}\label{rem:structure-a}
We will mainly consider the topological spaces $\operatorname{Prime}_\gamma(C^*(E))$ and $\Gamma_E$ for graphs with finitely many vertices. Assume that $E$ is a graph with finitely many vertices. Although $\operatorname{Prim}(C^*(E))$ often will be infinite (in the case of a cyclic component), the sets $\operatorname{Prime}_\gamma(C^*(E))$ and $\Gamma_E$ are finite. Thus it is clear that arbitrary intersections of open subsets are again open.
Thus $\operatorname{Prime}_\gamma(C^{*}(E))$ is an Alexandrov space. We see immediately from the definition that $\mathfrak{p}_1$ is in the closure of $\{\mathfrak{p}_2\}$ if and only if $\mathfrak{p}_1\supseteq\mathfrak{p}_2$.
So the specialization preorder $\geq$ is set containment.
Similarly, $\Gamma_E$ is an Alexandrov space and its specialization preorder is exactly the order $\geq$.
\end{remark}
\begin{lemma}\label{lem:structure-a}
Let $E=(E^0,E^1,r,s)$ be a graph with finitely many vertices.
Let $\eta\subseteq\Gamma_E$ be a hereditary subset. Assume that $v\in E^0_\mathrm{reg}$ and that there is no path from $v$ to any of the components in $\Gamma_E\setminus\eta$.
Then $v\in\overline{H(\cup\eta)}$.
\end{lemma}
\begin{proof}
There has to be a path from $v$ to some component --- thus a component in $\eta$.
If $v$ supports a cycle, clearly $v\in\cup\eta\subseteq\overline{H(\cup\eta)}$. Let $H_0=H(\cup\eta)$.
Using the description in \cite[Remark~3.1]{MR1988256}, we get a non-decreasing sequence of hereditary sets $\Sigma_0(H_0)=H_0$, $\Sigma_1(H_0)$, $\Sigma_2(H_0)$, \ldots.
If $v\not\in \Sigma_k(H_0)$, then the length of the longest path from $v$ to $\Sigma_k(H_0)$ is one less than the length of the longest path from $v$ to $\Sigma_{k-1}(H_0)$. Thus eventually $v\in\Sigma_k(H_0)$ for some $k$, \emph{i.e.}\xspace, $v\in\overline{H_0}$.
\end{proof}
\begin{lemma}\label{lem:structure-b}
Let $E=(E^0,E^1,r,s)$ be a graph with finitely many vertices.
Then the map $\eta\mapsto\overline{H(\cup\eta)}$ from the set of hereditary subsets of $\Gamma_E$ to the set of saturated hereditary subsets of $E^0$ is a bijective order isomorphism (with respect to the order coming from set containment).
In fact, $\cup\eta=(\cup\Gamma_E)\cap\overline{H(\cup\eta)}$.
Moreover, for any saturated hereditary subset $H\subseteq E^0$,
the set $(\cup\Gamma_E)\cap H$ is a (disjoint) union of all components that intersect $H$ nontrivially; and if we let $\eta\subseteq\Gamma_E$ be the set of these components, then $\eta$ is hereditary and
$\overline{H(\cup\eta)}=H$.
\end{lemma}
\begin{proof}
Assume that $\eta\subseteq\Gamma_E$ is a hereditary subset.
Clearly, $\cup\eta\subseteq(\cup\Gamma_E)\cap\overline{H(\cup\eta)}$.
Let $v\in(\cup\Gamma_E)\cap\overline{H(\cup\eta)}$.
Suppose that $v\in\gamma_1\in\Gamma_E$ but $\gamma_1\not\in\eta$.
Then $v\not\in H(\cup\eta)$.
Let $H_0=H(\cup\eta)$.
Using the description in \cite[Remark~3.1]{MR1988256}, we get a non-decreasing sequence of hereditary sets $\Sigma_0(H_0)=H_0$, $\Sigma_1(H_0)$, $\Sigma_2(H_0)$, \ldots, such that $v\in\Sigma_k(H_0)\setminus\Sigma_{k-1}(H_0)$, for some $k=1,2,3,\ldots$.
This means that $v\in E_\mathrm{reg}^0$ and $r(s^{-1}(v))\subseteq\Sigma_{k-1}(H_0)$.
Thus, clearly $v$ cannot support a loop. But $v$ cannot either support a cycle, since $\Sigma_{k-1}(H_0)$ is hereditary and all edges that $v$ emit go into $\Sigma_{k-1}(H_0)$.
So we get a contradiction, and therefore $v\in\cup\eta$.
So now it is clear that we have an injective map $\eta\mapsto\overline{H(\cup\eta)}$ from the set of hereditary subsets of $\Gamma_E$ to the set of saturated hereditary subsets of $E^0$. It is also clear that it is order preserving.
Now let there be given a saturated hereditary subset $H\subseteq E^0$.
For each $v\in(\cup\Gamma_E)\cap H$, all $v'$ that belong to the same component as $v$ are elements of $(\cup\Gamma_E)\cap H$.
So let $\eta\subseteq\Gamma_E$ be the (uniquely determined) set such that
$\cup\eta=(\cup\Gamma_E)\cap H$.
Since $\cup\eta\subseteq H$, it is clear that $H(\cup\eta)\subseteq H$.
Let $H_0=\overline{H(\cup\eta)}\subseteq H$.
Suppose $v\in H\setminus H_0$.
Then $v$ needs to be a transition state, so $v\in E_\mathrm{reg}^0$ and $v$ does not support a cycle. Consequently, it has to have a path to at least one component, but it cannot have any path to a component not in $\eta$. Lemma~\ref{lem:structure-a} now implies that $v\in H_0$, which is a contradiction.
Consequently, $H_0=\overline{H(\cup\eta)}= H$, and therefore the map is surjective.
\end{proof}
As an immediate consequence we get the following corollary.
\begin{corollary}\label{cor:structure-a}
Let $E=(E^0,E^1,r,s)$ be a graph with finitely many vertices, and assume that $E$ does not have any transition state.
Then every hereditary subset of $E^0$ is saturated and $\eta\mapsto \cup \eta$ is a lattice isomorphism between the hereditary subsets of $\Gamma_E$ and the saturated hereditary subsets of $E^0$.
\end{corollary}
The following is also clear.
\begin{lemma}\label{lem:structure-c}
Let $E=(E^0,E^1,r,s)$ be a graph with finitely many vertices.
If every infinite emitter emits infinitely many edges to any vertex it emits any edge to, then $B_H=\emptyset$ for every saturated hereditary subset $H \subseteq E^0$.
\end{lemma}
\begin{lemma}\label{lem:structure-d}
Let $E=(E^0,E^1,r,s)$ be a graph with finitely many vertices, and assume that every infinite emitter emits infinitely many edges to any vertex it emits any edge to.
Define a map \fct{{\upsilon}_E}{\Gamma_E}{\operatorname{Prime}_\gamma(C^*(E))} as follows. For each $\gamma_0\in\Gamma_E$, let ${\upsilon}_E(\gamma_0)$ denote the ideal
$$\mathfrak{J}_{\overline{H(\cup\eta_{\gamma_0}})},$$
where
$$\eta_{\gamma_0}=\Gamma_E\setminus\setof{\gamma\in\Gamma_E}{\gamma\geq\gamma_0}.$$
This is in fact an element of $\operatorname{Prime}_\gamma(C^*(E))$ and this makes ${\upsilon}_E$ into a bijection.
Moreover, $\gamma_1\geq\gamma_2$ if and only if ${\upsilon}_E(\gamma_1)\supseteq {\upsilon}_E(\gamma_2)$. Consequently, ${\upsilon}_E$ is a homeomorphism.
\end{lemma}
\begin{proof}
From \cite{MR2023453} and the proof of Lemma~\ref{lem:prime-ideals-1}, it is clear that the ideals in $\operatorname{Prime}_\gamma(C^{*}(E))$ are exactly the ideals $\mathfrak{J}_{E^{0}\setminus M}$, where $M\neq\emptyset$ is a maximal tail.
Assume that $M\neq\emptyset$ and let $H=E^0\setminus M$.
That $M$ is a maximal tail means that $M$ satisfies the conditions (MT1), (MT2) and (MT3) in \cite{MR2023453}.
Condition (MT1) is equivalent to $H$ being hereditary, while (MT2) is equivalent to $H$ being saturated.
Since $E^0$ is assumed to be finite, (MT3) is equivalent to the existence of $w\in M$ such that $v\geq w$ for all $v\in M$, \emph{i.e.}\xspace, $M$ has a least element (we will use this terminology although this is only a preorder and not a partial order in general).
Let $\gamma_0\in\Gamma_E$, and let
$$\eta_{\gamma_0}=\Gamma_E\setminus\setof{\gamma\in\Gamma_E}{\gamma\geq\gamma_0}.$$
It is clear that $\eta_{\gamma_0}$ is hereditary. Clearly, by the definition above ${\upsilon}_E(\gamma_0)$ defines an ideal. Let
$$H_0=\overline{H(\cup\eta_{\gamma_0})}.$$
We want to show that $E^0\setminus H_0$ is a maximal tail. The only thing we need to show is that it has a least element.
Choose $v_0\in\gamma_0$, and let $v\in E^0\setminus H_0$ be given.
Assume that $v\not\geq v_0$.
If $v\in\cup\Gamma_E$, then $v\in \cup\eta_{\gamma_0}$ and thus $v\in H_0$ (which is a contradiction).
Therefore, we would need to have that $v$ is a transition state --- so $v\in E_\mathrm{reg}^0$ and $v$ does not support a cycle.
There exists a path to some component in $\Gamma_E$, and, clearly, no such component can be in $\Gamma_E\setminus\eta_{\gamma_0}=\setof{\gamma\in\Gamma_E}{\gamma\geq\gamma_0}$.
From Lemma~\ref{lem:structure-a} it follows that $v\in\overline{H(\cup\eta_{\gamma_0})}=H_0$, which is a contradiction as well. Therefore $E^0\setminus H_0$ is a maximal tail, and ${\upsilon}_E(\gamma_0)$ is an element of $\operatorname{Prime}_\gamma(C^*(E))$.
From Fact~\ref{fact:structure-1} and Lemma~\ref{lem:structure-a} it follows that ${\upsilon}_E$ is injective.
Given an element of $\operatorname{Prime}_\gamma(C^*(E))$, then it has to be of the form $\mathfrak{J}_{H_0}$ for some saturated hereditary subset $H_0\subsetneq E^0$ with $E^0\setminus H_0$ having a least element $v_0$.
First note that $v_0$ cannot be a transition state, so $v_0\in\gamma_0$ for some $\gamma_0\in\Gamma_E$.
Let $\eta\subseteq\Gamma_E$ be such that $\cup\eta=(\cup\Gamma_E)\cap H_0$.
Clearly $\gamma_0\not\in\eta$.
Let $v\in\gamma\in\Gamma_E\setminus\eta$.
Then $v\geq v_0$, since $v\in\gamma\subseteq E^{0}\setminus H_0$.
Consequently, $\gamma\geq\gamma_0$.
On the other hand, assume that $\gamma\geq\gamma_0$ and let $v\in\gamma$.
Then $v\geq v_0$, so $v\in E^{0}\setminus H_0$.
Consequently, $\Gamma_E\setminus\eta= \setof{\gamma\in\Gamma_E}{\gamma\geq\gamma_0}$.
Thus the map ${\upsilon}_E$ is surjective.
That $\gamma_1\geq\gamma_2$ implies ${\upsilon}_E(\gamma_1)\supseteq {\upsilon}_E(\gamma_2)$ is clear from the definition.
That ${\upsilon}_E(\gamma_1)\supseteq {\upsilon}_E(\gamma_2)$ implies $\gamma_1\geq\gamma_2$ is clear from the definition and Lemma~\ref{lem:structure-b}.
\end{proof}
\begin{lemma}\label{lem:structure-2}
Let $E=(E^0,E^1,r,s)$ be a graph with finitely many vertices.
\begin{enumerate}[(i)]
\item\label{lem:structure-2-2}
If every transition state has exactly one edge going out, then
$H_1\cup H_2=\overline{H_1\cup H_2}$ for all saturated hereditary subsets $H_1,H_2\subseteq E^0$.
\item\label{lem:structure-2-3}
If $\gamma\in\Gamma_E$, then $\overline{H(\gamma)\setminus\gamma}$ is the largest proper saturated hereditary subset of $\overline{H(\gamma)}$.
\item\label{lem:structure-2-4}
If every transition state has exactly one edge going out, then the collection
$\overline{H(\gamma)}\setminus\overline{H(\gamma)\setminus\gamma}$, $\gamma\in\Gamma_E$ is a partition of $E^0$.
\item\label{lem:structure-2-6}
There exists a graph $F$ with finitely many vertices such that every infinite emitter emits infinitely many edges to any vertex it emits any edge to, every transition state has exactly one edge going out, $E\ensuremath{\sim_{M\negthinspace E}}\xspace F$, and $C^*(E)\cong C^*(F)$,
\item\label{lem:structure-2-7}
If every infinite emitter in $E$ emits infinitely many edges to any vertex it emits any edge to and every transition state has exactly one edge going out, then there exists a graph $F$ with finitely many vertices, such that every infinite emitter emits infinitely many edges to any vertex it emits any edge to, $F$ has no transition states $F^0=\cup\Gamma_E\subseteq E^0$, $\Gamma_E=\Gamma_{F}$ and they carry the same order $\geq$,
\begin{equation}\label{eq:structure-2-7-eq1}
s_E^{-1}(\cup\Gamma_E)\cap r_E^{-1}(\cup\Gamma_E)\subseteq s_F^{-1}(\cup\Gamma_{F})\cap r_F^{-1}(\cup\Gamma_{F})
\end{equation}
and there exists an injective \mbox{$\sp*$-}ho\-mo\-morphism\xspace from $C^*(E)$ to $C^*(F)\otimes\ensuremath{\mathbb{K}}\xspace$ such that the image of each ideal $\mathfrak{J}_{\overline{H(S)}}$ is a full corner in $\mathfrak{J}_{H(S)}\otimes\ensuremath{\mathbb{K}}\xspace$ for every hereditary subset $S\subseteq \Gamma_E$.
\item\label{lem:structure-2-8}
In the setting of part \ref{lem:structure-2-7}, we can get all cyclic components of $F$ to be singletons at the cost of \eqref{eq:structure-2-7-eq1} not necessarily holding anymore and only having a canonical identification of $\Gamma_E$ with $\Gamma_F$.
\end{enumerate}
\end{lemma}
\begin{proof}
\ref{lem:structure-2-2}:
Let $H_1$ and $H_2$ be saturated hereditary subsets of $E^0$. Since $H_1 \cup H_2$ is hereditary, it is enough to show that $H_1 \cup H_2$ is saturated. Let $x \in E^0$ be a regular vertex such that $r( s^{-1} (x) ) \subseteq H_1 \cup H_2$. Suppose $x$ is a transitional vertex. Then by assumption $s^{-1}(x) = \{e\}$. Therefore, $r( s^{-1}(x) ) = \{ r( e) \} \subseteq H_1$ or $r( s^{-1}(x) ) = \{ r(e) \} \subseteq H_2$. Since $H_1$ and $H_2$ are saturated, we have that $x \in H_1$ or $x \in H_2$ which implies that $x \in H_1 \cup H_2$. Suppose $x \in \gamma$ for some $\gamma \in \Gamma_E$. Then there exists a path $\mu = \mu_1 \cdots \mu_n$ such that $s( \mu_1 ) = r( \mu_n ) = x$ (we are using the fact that $x$ is a regular vertex). Since $r( s^{-1} (x) ) \subseteq H_1 \cup H_2$ and $\mu_1 \in s^{-1}(x )$, we have that $r( \mu_1 ) \in H_1 \cup H_2$. Since $H_1 \cup H_2$ is hereditary, $x = r( \mu_n ) \in H_1 \cup H_2$.
In both cases, we have shown that $x \in H_1 \cup H_2$. Therefore, $H_1 \cup H_2$ is saturated.
\ref{lem:structure-2-3}:
Let $H=\overline{H(\gamma)\setminus\gamma}$.
Clearly $H$ is saturated and hereditary, and $H\subseteq\overline{H(\gamma)}$.
We want to show that $H$ is a proper subset of $\overline{H(\gamma)}$ and $\gamma\cap H=\emptyset$.
So assume first that $\gamma\cap H\neq\emptyset$.
Then $H\setminus\gamma$ is not saturated.
Thus there exists a $v\in E_{\mathrm{reg}}^{0}$ such that $r(s^{-1}(v))\subseteq H\setminus \gamma$ and $v\not\in H\setminus\gamma$.
Since $H\setminus\gamma\subseteq H$ and $H$ is saturated, $v\in H$.
Thus $v\in\gamma$.
Since $v\in\gamma\in\Gamma_E$, we have that $v$ supports a cycle within $\gamma$ or $v$ is singular --- both being contradictions.
Consequently, we have that $\gamma\cap H=\emptyset$.
Now it is clear that $H$ is a proper subset of $\overline{H(\gamma)}$.
Now we want to show that $H$ is the largest proper saturated hereditary subset of $\overline{H(\gamma)}$.
It is enough to show that for all $v\in \overline{H(\gamma)}\setminus H$, we have that $\overline{H(v)}=\overline{H(\gamma)}$.
So let $v\in \overline{H(\gamma)}\setminus H$ be given. Clearly $\overline{H(v)}\subseteq\overline{H(\gamma)}$.
If $v\in\gamma$, then $H(v)=H(\gamma)$, so $\overline{H(v)}=\overline{H(\gamma)}$. Suppose $v \notin \gamma$.
Since $H(\gamma)\setminus \gamma\subseteq H$, we have that $v\not\in H(\gamma)\setminus \gamma$.
So now assume that $v\in \overline{H(\gamma)}\setminus H$.
Note that $\overline{H(\gamma)}\setminus\gamma$ is saturated, and thus $H\subseteq \overline{H(\gamma)}\setminus\gamma$.
Then $\overline{H(\gamma)}\setminus\{v\}$ cannot be saturated, so $v\in E_\mathrm{reg}^{0}$ and $r(s^{-1}(v))\subseteq \overline{H(\gamma)}\setminus\{v\}$.
By assumption, we must have that $r(s^{-1}(v))\not\subseteq H$.
Using the description of the saturation from \cite[Remark~3.1]{MR1988256}, it follows that there exists a $v_0\in\gamma$ such that $v\geq v_0$. Thus $H(v)\supseteq H(\gamma)$ and $\overline{H(v)}\supseteq\overline{H(\gamma)}$ follows.
\ref{lem:structure-2-4}:
It follows from \ref{lem:structure-2-3} that $\gamma\subseteq\overline{H(\gamma)}\setminus \overline{H(\gamma)\setminus \gamma}$ for each $\gamma\in\Gamma_E$. The transition states are the regular vertices not supporting a cycle. Since we only have finitely many vertices, every transition state will have a path to a component (the sinks are also components). Moreover, since every transition state has exactly one outgoing edge, each transition state has a unique shortest path to a component through transition states. If we have a transition state $v$ and the first component every path from $v$ reaches is $\gamma$, then it follows from \cite[Remark~3.1]{MR1988256} that $v\in\overline{H(\gamma)}$. From the proof of \ref{lem:structure-2-3}, we have that $\gamma\cap \overline{H(\gamma)\setminus\gamma}=\emptyset$, so $v\not\in \overline{H(\gamma)\setminus\gamma}$. Thus we have shown that every vertex belongs to at least one of the sets $\overline{H(\gamma)}\setminus \overline{H(\gamma)\setminus \gamma}$, $\gamma\in\Gamma_E$. Let $\gamma,\gamma'\in\Gamma_E$. If $\gamma\geq\gamma'$ and $\gamma \neq \gamma'$, then $\gamma' \subseteq H(\gamma)$, and therefore $\gamma'\cap(\overline{H(\gamma)}\setminus \overline{H(\gamma)\setminus \gamma})=\emptyset$. If $\gamma\not\geq\gamma'$, then $\overline{H(\gamma)}\setminus \gamma'$ is a saturated set that contains $H(\gamma)$, and, consequently, $\gamma'\cap \overline{H(\gamma)}=\emptyset$. Therefore, the vertices of the components belong to a unique set in the collection $\overline{H(\gamma)}\setminus \overline{H(\gamma)\setminus \gamma}$, $\gamma\in\Gamma_E$. Now let $v$ be a transition state and let $\gamma$ be the first component every path from $v$ reaches. Assume that $v\in\overline{H(\gamma')}\setminus \overline{H(\gamma')\setminus \gamma'}$ for a $\gamma'\in\Gamma_E$ with $\gamma'\neq\gamma$. If $\gamma'\geq\gamma$, then $\gamma\subseteq H(\gamma')\setminus \gamma'$ and therefore $v\in\overline{H(\gamma')\setminus \gamma'}$. So this is a contradiction since $v\in\overline{H(\gamma')}\setminus \overline{H(\gamma')\setminus \gamma'}$. If $\gamma'\not\geq\gamma$, then we have seen that $\gamma\cap \overline{H(\gamma')}=\emptyset$ while $v\in \overline{H(\gamma')}$ implies that $\gamma \subseteq \overline{H(\gamma')}$. So this is a contradiction. Thus we have shown that each transition state belongs to a unique set in the collection $\overline{H(\gamma)}\setminus \overline{H(\gamma)\setminus \gamma}$, $\gamma\in\Gamma_E$ --- namely, the first component every path from it reaches.
\ref{lem:structure-2-6}:
First we show how to modify $E$ to get a graph with the property that if $v$ is an infinite emitter, then $v$ emits infinitely many edges to any vertex it emits any edge to.
Let $v \in E^0$ be an infinite emitter.
If there exists a vertex $u \in E^0$ such that $v$ emits only finitely many edges to $u$, we partition $s^{-1}(v)$ into two sets, $\mathcal{E}_1 = \setof{ e \in s^{-1}(v)}{ |s^{-1}(v) \cap r^{-1}(r(e))| < \infty }$ and $\mathcal{E}_2 = \setof{ e \in s^{-1}(v)}{ |s^{-1}(v) \cap r^{-1}(r(e))| = \infty }$, \emph{i.e.}\xspace, $\mathcal{E}_1$ consists of the edges out of $v$ that only have finitely many parallel edges.
Note that since $E^0$ is finite, $\mathcal{E}_1$ is a finite set.
Hence we can perform Move \mbox{\texttt{\textup{(O)}}}\xspace according to this partition, resulting in a graph $E'$ that is move equivalent to $E$.
Call the vertices $v$ got split into $v_1$ and $v_2$.
In $E'$, $v_2$ is an infinite emitter with the property that it emits infinitely many edges to any vertex it emits any edge to, and any infinite emitter in $E$ that already had that property keeps it.
On the other hand $v_1$ is a finite emitter.
Since $E^0$ is finite, we can do the above process a finite number of times, ending with a graph $G$ that is move equivalent to $E$, and with the property that if $v$ is an infinite emitter, then $v$ emits infinitely many edges to any vertex it emits any edge to.
Let $n\in\ensuremath{\mathbb{N}}\xspace$ and let $v \in G^0$ be a transition state of $G$, \emph{i.e.}\xspace, a regular vertex that is not the base point of a cycle.
Assume that $|s^{-1}(v)|\geq 2$, and that the shortest path from $v$ to a component of $G$ is $n$.
Since $v$ is regular, we can partition $s^{-1}(v)$ into finitely many disjoint singletons $\mathcal{E}_1'$,$\mathcal{E}_2',\ldots,\mathcal{E}_{|s^{-1}(v)|}'$.
Now we can perform Move \mbox{\texttt{\textup{(O)}}}\xspace according to this partition, resulting in a graph $G'$ that is move equivalent to $G$ such that vertices that $v$ got split into are still transition states but each having exactly one outgoing edge, and the shortest path from each of them to a component is at least $n$.
A vertex in $G'$ is a transition state if and only if it is one of the vertices that $v$ got split into or it is a transition state of $G$.
All transition states in $G$ that had exactly one outgoing edge and a path to a component of length $n$ or shorter will still have exactly one outgoing edge and a path of length at most $n$.
Also, every infinite emitter in $G'$ emits infinitely many edges to any vertex it emits any edge to.
We repeat this for all transition states emitting at least two edges and with the shortest path to a component having length $n$.
By induction on $n$, we can get a graph $F$ with finitely many vertices such that every infinite emitter emits infinitely many edges to any vertex it emits any edge to, every transition state has exactly one edge going out.
We got $F$ from $E$ by using Move~\mbox{\texttt{\textup{(O)}}}\xspace a number of times.
Therefore we clearly have that $E\ensuremath{\sim_{M\negthinspace E}}\xspace F$, and it follows from Proposition~\ref{prop:moveOimpliesisomorphism} that $C^*(E)\cong C^*(F)$.
\ref{lem:structure-2-7}:
Let $F$ be the graph obtained by continuing to collapse all transitional vertices of $E$. It is clear from the construction of $F$ that $F^0 = \cup \Gamma_E \subseteq E^0$, $\Gamma_E = \Gamma_F$, they carry the same order, and $s_E^{-1} ( \cup \Gamma_E) \cap r_E^{-1}( \cup\Gamma_E ) \subseteq s_F^{-1} ( \cup \Gamma_F) \cap r_F^{-1}( \cup\Gamma_F )$. Now, there exists an injective \mbox{$\sp*$-}ho\-mo\-morphism\xspace $\Phi_1 \colon C^* (F) \rightarrow C^* (E)$ such that $\Phi_1 ( C^* (F) ) = P C^* (E) P$, where $P$ is the sum of vertex projections of the vertices from $F$. Since $\overline{H(F^0)} = E^0$, we have that $\Phi_1 ( C^* (F) )$ is a full corner of $C^*(E)$. Therefore, $\Phi_1 ( \mathfrak{J}_{ H(S) } ) = P \mathfrak{J}_{ \overline{H(S)} } P$ for every hereditary subset $S \subseteq \Gamma_F$. By \cite{MR0454645} there exists a partial isometry $v$ in $\mathcal{M}( C^* (E) \otimes \ensuremath{\mathbb{K}}\xspace )$ such that $v^*v = \Phi_1 ( 1_{C^*(F)} ) \otimes 1_{\mathbb{B} ( \ell^2 ) }$ and $vv^* = 1_{ \mathcal{M}( C^* (E) \otimes \ensuremath{\mathbb{K}}\xspace ) }$. Set $\Phi_2 = \mathrm{Ad} (v) \circ (\Phi_1 \otimes \mathrm{id}_\ensuremath{\mathbb{K}}\xspace)$. Hence, $\Phi_2 \colon C^* (F) \otimes \ensuremath{\mathbb{K}}\xspace \rightarrow C^* (E) \otimes \ensuremath{\mathbb{K}}\xspace$ is a \mbox{$\sp*$-}iso\-morphism\xspace such that $\Phi_2 ( \mathfrak{J}_{ H(S) } \otimes \ensuremath{\mathbb{K}}\xspace)$ is a full corner of $\mathfrak{J}_{ \overline{H(S)} } \otimes \ensuremath{\mathbb{K}}\xspace$ for every hereditary subset $S \subseteq \Gamma_F$.
Set $\Psi = \Phi_2^{-1} \circ \kappa$, where $\kappa$ is the embedding $C^* (E)$ to $C^* (E) \otimes \ensuremath{\mathbb{K}}\xspace$ given by $a \mapsto a \otimes e_{11}$. Therefore, $\Psi \colon C^* (E) \rightarrow C^* (F) \otimes \ensuremath{\mathbb{K}}\xspace$ is an injective \mbox{$\sp*$-}ho\-mo\-morphism\xspace such that $\Psi ( \mathfrak{J}_{\overline{H(S)} } )$ is a full corner of $\mathfrak{J}_{H(S)}$ for every hereditary subset $S \subseteq \Gamma_F$. Since $\Gamma_F = \Gamma_E$, $S$ is hereditary in $\Gamma_F$ if and only if $S$ is hereditary in $\Gamma_E$. So, $\Psi \colon C^* (E) \rightarrow C^* (F) \otimes \ensuremath{\mathbb{K}}\xspace$ is an injective \mbox{$\sp*$-}ho\-mo\-morphism\xspace such that $\Psi ( \mathfrak{J}_{\overline{H(S)} } )$ is a full corner of $\mathfrak{J}_{H(S)}$ for every hereditary subset $S \subseteq \Gamma_E$.
\ref{lem:structure-2-8}:
In addition to the process in \ref{lem:structure-2-7} of collapsing all transitional vertices of $E$, we also collapse all regular vertices of $E$ that are base points of cyclic components (but not of a loop). Using a similar argument as the proof of \ref{lem:structure-2-7}, we get the desired result.
\end{proof}
\begin{proposition}\label{prop:thetamap}
Let $E$ be a graph with finitely many vertices such that every infinite emitter emits infinitely many edges to any vertex it emits any edge to.
In Lemma~\ref{lem:structure-d} we have defined a homeomorphism ${\upsilon}_E$ from $\Gamma_E$ to $\operatorname{Prime}_\gamma(C^*(E))$.
This homeomorphism induces a lattice isomorphism from the open subsets of $\Gamma_E$ to the open subsets of $\operatorname{Prime}_\gamma(C^*(E))$. We also denote this map ${\upsilon}_E$.
Let $\omega_E$ denote the map given by Lemma~\ref{lem:structure-b} and Fact~\ref{fact:structure-1}, \emph{i.e.}\xspace,
$$\omega_E(\eta)=\mathfrak{J}_{\overline{H(\cup\eta)}}$$
for every hereditary subset $\eta$ of $\Gamma_E$, and let $\varepsilon_E$ denote the map from $\mathbb{O}(\operatorname{Prime}_\gamma(C^*(E)))$ to $\mathbb{I}_\gamma(C^*(E))$ given in Lemma~\ref{lem:orderreversingprime}, \emph{i.e.}\xspace,
$$\varepsilon_E(O)=\cap(\operatorname{Prime}_\gamma(C^*(E))\setminus O)$$
for every open subset $O\subseteq\operatorname{Prime}_\gamma(C^*(E))$.
Then we have a commuting diagram
$$\xymatrix{\mathbb{O}(\Gamma_E)\ar[d]^{{\upsilon}_E}_\cong\ar[r]_-\cong^-{\omega_E} &
\mathbb{I}_\gamma(C^*(E))\ar@{=}[d] \\
\mathbb{O}(\operatorname{Prime}_\gamma(C^*(E)))\ar[r]_-\cong^-{\varepsilon_E}
& \mathbb{I}_\gamma(C^*(E))}$$
of lattice isomorphisms.
\end{proposition}
\begin{proof}
The only new statement in the proposition is the commutativity of the diagram. Note that the inverse of the map $\varepsilon_E$ is also given in Lemma~\ref{lem:orderreversingprime}, and it is $\mathfrak{I}\mapsto\Gamma_E\setminus\operatorname{hull}_\gamma(\mathfrak{I})$.
Let $\eta\subseteq\Gamma_E$ be a hereditary subset. Then $$\varepsilon^{-1}_E\circ\omega_E(\eta)=\operatorname{Prime}_\gamma(C^*(E))\setminus\operatorname{hull}_\gamma(\mathfrak{J}_{\overline{H(\cup\eta)}}).$$
From the description of the elements in $\operatorname{Prime}_\gamma(C^*(E))$ in the proof of Lemma~\ref{lem:structure-d}, we see that this set is exactly the set
$$\setof{\mathfrak{J}_{\overline{H(\cup(\Gamma_E\setminus\setof{\gamma\in\Gamma_E}{\gamma\geq\gamma_0}))}}}{\gamma_0\in\Gamma_E, \overline{H(\cup(\Gamma_E\setminus\setof{\gamma\in\Gamma_E}{\gamma\geq\gamma_0}))}\not\supseteq \overline{H(\cup\eta)}}.$$
But this is exactly the image of
$$\setof{\gamma_0\in\Gamma_E}{\Gamma_E\setminus\setof{\gamma\in\Gamma_E}{\gamma\geq\gamma_0}\not\supseteq \eta}$$
under the homeomorphism ${\upsilon}_E$. Since $\eta$ is hereditary, this set is exactly $\eta$.
\end{proof}
\begin{example} We will now take a look at an example of how we
get the space $\operatorname{Prime}_\gamma(C^{*}(E))$ from the space $\Gamma_E$
for a graph $E$ with finitely many vertices (where all the
infinite emitters emit infinitely many edges to any vertex they
emit any edge to). Let us say that the ordered set $\Gamma_E$
consists of four points $\gamma_1$, $\gamma_2$, $\gamma_3$,
$\gamma_4$ with the relations $\gamma_1\geq\gamma_3$,
$\gamma_2\geq\gamma_3$, $\gamma_3\geq\gamma_4$ (and thus also
$\gamma_1,\gamma_2\geq\gamma_4$), while $\gamma_1\not\geq\gamma_2$
and $\gamma_2\not\geq\gamma_1$. This can be illustrated by the
component graph as in Figure~\ref{figure:component-graph}.
\begin{figure}[h]\begin{center}
$$\xymatrix@=0.3cm{\gamma_1\ar[dr]&&\gamma_2\ar[dl] \\
&\gamma_3\ar[d] \\ & \gamma_4}$$
\end{center}
\caption{The component graph $\Gamma_E$}
\label{figure:component-graph}
\end{figure}
For each $\gamma_i$, $i=1,2,3,4$, we consider the hereditary subset $\eta_i=\Gamma_E\setminus\setof{\gamma\in\Gamma_E}{\gamma\geq\gamma_i}$.
These subsets are illustrated in Figure~\ref{figure:primeexample} by marking the elements of the subset in red.
\begin{figure}[H]\quad\begin{subfigure}[b]{0.4\textwidth}
$$\xymatrix@=0.3cm{\gamma_1\ar[dr]&&{\color{red}\gamma_2\ar[dl]} \\
&{\color{red}\gamma_3}\ar[d] \\ & {\color{red}\gamma_4}}$$
\caption{The set $\eta_1$}
\label{fig:prime-1}
\end{subfigure}\qquad
~ \begin{subfigure}[b]{0.4\textwidth}
$$\xymatrix@=0.3cm{{\color{red}\gamma_1}\ar[dr]&&\gamma_2\ar[dl] \\
&{\color{red}\gamma_3\ar[d]} \\ & {\color{red}\gamma_4}}$$
\caption{The set $\eta_2$}
\label{fig:prime-2}
\end{subfigure}\qquad
~ \begin{subfigure}[b]{0.4\textwidth}
$$\xymatrix@=0.3cm{\gamma_1\ar[dr]&&\gamma_2\ar[dl] \\
&\gamma_3\ar[d] \\ & {\color{red}\gamma_4}}$$
\caption{The set $\eta_3$}
\label{fig:prime-3}
\end{subfigure}\qquad
~ \begin{subfigure}[b]{0.4\textwidth}
$$\xymatrix@=0.3cm{\gamma_1\ar[dr]&&\gamma_2\ar[dl] \\
&\gamma_3\ar[d] \\ & \gamma_4}$$
\caption{The set $\eta_4$ --- it is the empty set}
\label{fig:prime-4}
\end{subfigure}\quad
\caption{The components marked with red show the elements of $\eta_i$, for $i=1,2,3,4$}
\label{figure:primeexample}
\end{figure}
So the corresponding gauge invariant ideals $\upsilon_E(\gamma_i)=\omega_E(\eta_i)$ of $C^*(E)$ are $\mathfrak{J}_{\overline{H(\cup\eta_i)}}$, \emph{i.e.}\xspace, $\mathfrak{J}_{\overline{H(\gamma_2\cup\gamma_3\cup\gamma_4)}}$,
$\mathfrak{J}_{\overline{H(\gamma_1\cup\gamma_3\cup\gamma_4)}}$,
$\mathfrak{J}_{\overline{H(\gamma_4)}}$,
$\mathfrak{J}_{\overline{H(\emptyset)}}=\{0\}$, respectively.
The topology on $\operatorname{Prime}_\gamma(C^*(E))$ is given by the specialization preorder, so we can illustrate it as in Figure~\ref{figure:prime-space}, where an arrow (or path) from $x$ to $y$ indicates that $x$ is in the closure of $\{y\}$.
\begin{figure}[H]\begin{center}
$$\xymatrix@=0.4cm{\mathfrak{J}_{\overline{H(\gamma_2\cup\gamma_3\cup\gamma_4)}}
\ar[dr]&&\mathfrak{J}_{\overline{H(\gamma_1\cup\gamma_3\cup\gamma_4)}}\ar[dl] \\
&\mathfrak{J}_{\overline{H(\gamma_4)}}\ar[d] \\ & \{0\}}$$
\end{center}
\caption{An illustration of $\operatorname{Prime}_\gamma(C^*(E))$}
\label{figure:prime-space}
\end{figure}
\end{example}
\begin{example}\label{linearcase}
In the case that the ordered set $\Gamma_E$ is linearly ordered
\[
\gamma_1\geq \gamma_2\geq \cdots\geq \gamma_n
\]
(where all the infinite emitters emit infinitely many edges to any vertex they emit any edge to) we get hereditary subsets $\eta_i=\setof{\gamma}{\gamma_i >\gamma}$ and prime gauge invariant ideals
\[
\mathfrak p_i=\upsilon_E(\gamma_i)=\omega_E(\eta_i)=\mathfrak{J}_{\overline{H(\cup{\eta_i})}}=
\begin{cases}\mathfrak{J}_{\overline{H(\gamma_{i+1})}}&i<n\\
\{0\}&i=n.
\end{cases}
\]
Note that the $\mathfrak p_i$ decrease as $i$ increases. We denote the corresponding topological space by $X_n$ and note that it is the Alexandrov space of a linear order on a set of $n$ elements.
\end{example}
\subsection{Reduced filtered \texorpdfstring{$K$}{K}-theory}\label{sec:reducedKtheory}
Let $X$ be a topological space satisfying the $T_0$ separation
property and let \ensuremath{\mathfrak{A}}\xspace be a \mbox{$C\sp*$-}al\-ge\-bra\xspace over $X$. For open subsets $U_{1} ,
U_{2} , U_{3}$ of $X$ with $U_{1} \subseteq U_{2} \subseteq U_{3}$, let $Y_{1} = U_{2} \setminus U_{1}, Y_{2} = U_{3} \setminus U_{1},
Y_{3} = U_{3} \setminus U_{2}\in \mathbb{LC}(X)$. Then the
diagram
\begin{equation*}
\xymatrix{
K_{0} ( \ensuremath{\mathfrak{A}}\xspace ( Y_{1} ) ) \ar[r]^{ \iota_{*} } & K_{0} ( \ensuremath{\mathfrak{A}}\xspace ( Y_{2} ) ) \ar[r]^{ \pi_{*} } & K_{0} ( \ensuremath{\mathfrak{A}}\xspace ( Y_{3} ) ) \ar[d]^{\partial_{*}} \\
K_{1} ( \ensuremath{\mathfrak{A}}\xspace ( Y_{3} ) ) \ar[u]^{ \partial_{*}} & K_{1} ( \ensuremath{\mathfrak{A}}\xspace ( Y_{2} ) ) \ar[l]^{ \pi_{*} } & K_{1} ( \ensuremath{\mathfrak{A}}\xspace ( Y_{1} ) ) \ar[l]^{\iota_{*}}
}
\end{equation*}
is an exact sequence. The collection of all such exact sequences is an invariant of the \mbox{$C\sp*$-}al\-ge\-bras\xspace over $X$ often referred to as the \emph{filtered $K$-theory}. We use here a refined notion:
\begin{definition}
Let $X$ be a finite topological space satisfying the $T_0$ separation property and let \ensuremath{\mathfrak{A}}\xspace be a \mbox{$C\sp*$-}al\-ge\-bra\xspace over $X$.
Note that all singletons of $X$ are locally closed.
For each $x\in X$, we let $S_x$ denote the smallest open subset that contains $x$, and we let
$R_x=S_x\setminus\{x\}$, which is an open subset.
As mentioned above, we get a cyclic six term exact sequence in $K$-theory
\begin{equation}\label{eq:sixtermktheory}\vcenter{
\xymatrix{
K_0(\ensuremath{\mathfrak{A}}\xspace(O))\ar[r] & K_0(\ensuremath{\mathfrak{A}}\xspace(U))\ar[r] & K_0(\ensuremath{\mathfrak{A}}\xspace(U\setminus O))\ar[d] \\
K_1(\ensuremath{\mathfrak{A}}\xspace(U\setminus O))\ar[u] & K_1(\ensuremath{\mathfrak{A}}\xspace(U))\ar[l] & K_1(\ensuremath{\mathfrak{A}}\xspace(O))\ar[l] \\
}}
\end{equation}
whenever we have two open subsets $O\subseteq U\subseteq X$.
It follows from \cite[Theorem~4.1]{MR2922394} that the map from $K_0$ to $K_1$ is the zero map whenever $\ensuremath{\mathfrak{A}}\xspace(O)$ and $\ensuremath{\mathfrak{A}}\xspace(U)$ are gauge invariant ideals of a graph \mbox{$C\sp*$-}al\-ge\-bra\xspace.
Let
\begin{align*}
I_0(\ensuremath{\mathfrak{A}}\xspace)&=\setof{R_x}{x\in X,R_x\neq\emptyset}\cup\setof{S_x}{x\in X}\cup\setof{\{x\}}{x\in X},\\
I_1(\ensuremath{\mathfrak{A}}\xspace)&=\setof{\{x\}}{x\in X},
\end{align*}
and let $\operatorname{Imm}(x)$ denote the set
$$\setof{y\in X}{S_y\subsetneq S_x\wedge \not\exists z\in X\colon S_y\subsetneq S_z\subsetneq S_x}.$$
The \emph{reduced filtered $K$-theory} of \ensuremath{\mathfrak{A}}\xspace, $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(X;\ensuremath{\mathfrak{A}}\xspace)$, consists of the families of groups
$(K_0(\ensuremath{\mathfrak{A}}\xspace(V)))_{V\in I_0(\ensuremath{\mathfrak{A}}\xspace)}$ and
$(K_1(\ensuremath{\mathfrak{A}}\xspace(O)))_{O\in I_1(\ensuremath{\mathfrak{A}}\xspace)}$ together with the maps in the sequences
\begin{equation}\label{eq:longtype}
K_1(\ensuremath{\mathfrak{A}}\xspace(\{x\}))\to K_0(\ensuremath{\mathfrak{A}}\xspace(R_x))\to K_0(\ensuremath{\mathfrak{A}}\xspace(S_x))\to K_0(\ensuremath{\mathfrak{A}}\xspace(\{x\}))
\end{equation}
originating from the sequence \eqref{eq:sixtermktheory}, for all $x\in X$ with $R_x\neq\emptyset$, and
the maps in the sequences
\begin{equation}\label{eq:shorttype}
K_0(\ensuremath{\mathfrak{A}}\xspace(S_y))\to K_0(\ensuremath{\mathfrak{A}}\xspace(R_x))
\end{equation}
originating from the sequence \eqref{eq:sixtermktheory}, for all pairs $(x,y)\in X$ with $y\in\operatorname{Imm}(x)$ and $\operatorname{Imm}(x)\setminus\{y\}\neq\emptyset$.
Let \ensuremath{\mathfrak{B}}\xspace be a \mbox{$C\sp*$-}al\-ge\-bra\xspace over $X$.
A \emph{homomorphism} from $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(X;\ensuremath{\mathfrak{A}}\xspace)$ to $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(X;\ensuremath{\mathfrak{B}}\xspace)$ consists of
families of group homomorphisms
$$(\phi_{V,0}\colon K_0(\ensuremath{\mathfrak{A}}\xspace(V))\rightarrow K_0(\ensuremath{\mathfrak{B}}\xspace(V)))_{V\in I_0(\ensuremath{\mathfrak{A}}\xspace)}$$
$$(\phi_{O,1}\colon K_1(\ensuremath{\mathfrak{A}}\xspace(O))\rightarrow K_1(\ensuremath{\mathfrak{B}}\xspace(O)))_{O\in I_1(\ensuremath{\mathfrak{A}}\xspace)}$$
such that all the ladders coming from the above sequences commute. A homomorphism is an \emph{isomorphism} exactly if the group homomorphisms in the family are group isomorphisms.
Analogously, we define the \emph{ordered reduced filtered $K$-theory} of \ensuremath{\mathfrak{A}}\xspace, $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(X;\ensuremath{\mathfrak{A}}\xspace)$, just as $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(X;\ensuremath{\mathfrak{A}}\xspace)$ where we also consider the order on all the $K_0$-groups --- and for a homomorphism respectively an isomorphism, we demand that the group homomorphisms respectively the group isomorphisms between the $K_0$-groups are positive homomorphisms respectively order isomorphisms.
Hereby we get --- in the obvious way --- two functors $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(X;- )$ and $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(X;- )$ that are defined on the category of \mbox{$C\sp*$-}al\-ge\-bras\xspace over $X$.
\end{definition}
\begin{remark}\label{howtocompute}
Let $E$ be a graph.
Then $C^*(E)$ has a canonical structure as a $\operatorname{Prime}_\gamma(C^*(E))$-algebra.
So if $E$ has finitely many vertices --- or, more generally, if $\operatorname{Prime}_\gamma(C^*(E))$ is finite --- then we can consider the reduced filtered $K$-theory, $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(\operatorname{Prime}_\gamma(C^*(E)),C^*(E))$.
We use the results of \cite{MR2922394} to identify the $K$-groups and the homomorphisms in the cyclic six term sequences using the adjacency matrix of the graph.
\end{remark}
\begin{remark}
Let \ensuremath{\mathfrak{A}}\xspace be an $X$-algebra.
Since $\mathfrak{I}\mapsto\mathfrak{I}\otimes\ensuremath{\mathbb{K}}\xspace$ is a lattice isomorphism between $\mathbb{I}(\ensuremath{\mathfrak{A}}\xspace)$ and $\mathbb{I}(\ensuremath{\mathfrak{A}}\xspace\otimes\ensuremath{\mathbb{K}}\xspace)$, the \mbox{$C\sp*$-}al\-ge\-bra\xspace $\ensuremath{\mathfrak{A}}\xspace\otimes\ensuremath{\mathbb{K}}\xspace$ is an $X$-algebra in a canonical way, and the embedding $\kappa_\ensuremath{\mathfrak{A}}\xspace$ given by $a\mapsto a\otimes e_{11}$ is an $X$-equivariant homomorphism from $\ensuremath{\mathfrak{A}}\xspace$ to $\ensuremath{\mathfrak{A}}\xspace\otimes\ensuremath{\mathbb{K}}\xspace$.
Also, it is clear that $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(X;\kappa_\ensuremath{\mathfrak{A}}\xspace)$ is an (order) isomorphism.
Note also that the invariant $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(X;-)$ has been considered in \cite{MR3177344,MR3349327}.
\end{remark}
\begin{remark}\label{lpacomments}
Appealing to \cite{MR2514392} instead of \cite{MR2023453} one
may define $\operatorname{Prime}_\gamma$ also for Leavitt path algebras over
$\ensuremath{\mathbb{C}}\xspace$, and establish most of the results of this section also in a
purely algebraic setting. Since \cite{MR2514392} discusses only
row-finite graphs and we here insist that there are only
finitely many vertices, this applies only to finite graphs.
\end{remark}
\section{Specific preliminaries}
\label{sec:notation-for-proof}
In this section we introduce concepts and notation that are required for the remainder of the paper.
\subsection{Block matrices and equivalences}
\begin{notation}
For $m,n\in\ensuremath{\mathbb{N}}\xspace_0$, we let $\MZ[m\times n]$ denote the set of group homomorphisms from $\ensuremath{\mathbb{Z}}\xspace^n$ to $\ensuremath{\mathbb{Z}}\xspace^m$. When $m,n\geq 1$, we can equivalently view this as the $m\times n$ matrices over $\ensuremath{\mathbb{Z}}\xspace$, where composition of group homomorphisms corresponds to matrix multiplication --- the (zero) group homomorphisms for $m=0$ or $n=0$, we will also call empty matrices with zero rows or columns, respectively.
For $m,n\in\ensuremath{\mathbb{N}}\xspace$, we let \Mplus denote the subset of $\MZ[m\times n]$, where all entries in the corresponding matrix are positive. For an $m\times n$ matrix, we will also write $B>0$ whenever $B\in\Mplus$.
For an $m\times n$ matrix $B$, where $m,n\in\ensuremath{\mathbb{N}}\xspace$, we let $B(i,j)$ denote the $(i,j)$'th entry of the corresponding matrix, \emph{i.e.}\xspace, the entry in the $i$'th row and $j$'th column.\end{notation}
\begin{definition}
Let $m,n\in\ensuremath{\mathbb{N}}\xspace$.
For an $m\times n$ matrix $B$ over \ensuremath{\mathbb{Z}}\xspace, we let $\gcd B$ be the greatest common divisor of the entries $B(i,j)$, for $i=1,\ldots,m$, $j=1,\ldots,n$, if $B$ is nonzero, and zero otherwise.
\end{definition}
\begin{assumption} \label{ass:preorder}
Let $N\in\ensuremath{\mathbb{N}}\xspace$.
For the rest of the paper, we let $\ensuremath{\mathcal{P}}\xspace=\{1,2,\ldots,N\}$ denote a partially ordered set with order $\preceq$ satisfying
$$i\preceq j\Rightarrow i\leq j,$$
for all $i,j\in\ensuremath{\mathcal{P}}\xspace$, where $\leq$ denotes the usual order on \ensuremath{\mathbb{N}}\xspace.
We denote the corresponding irreflexive order by $\prec$.
\end{assumption}
\begin{definition}\label{def:blockmatrices}
Let $\mathbf{m}=(m_i)_{i=1}^{N},\mathbf{n}=(n_i)_{i=1}^{N}\in\ensuremath{\mathbb{N}}\xspace_{0}^N$ be \emph{multiindices}.
We write $\mathbf{m}\leq\mathbf{n}$ if $m_i\leq n_i$ for all $i=1,2,\ldots,N$,
and in that case, we let $\mathbf{n}-\mathbf{m}$ be $(n_i-m_i)_{i=1}^N$.
We also let $\mathbf{m}+\mathbf{n}$ denote $(m_i+n_i)_{i=1}^N$ for any multiindices, and we let $|\mathbf{m}|=m_1+m_2+\cdots+m_N$. We denote the multiindex with $1$ on every entry by $\mathbf 1$.
We let $\MZ$ denote the set of group homomorphisms from $\ensuremath{\mathbb{Z}}\xspace^{n_1}\oplus\ensuremath{\mathbb{Z}}\xspace^{n_2}\oplus\cdots\oplus\ensuremath{\mathbb{Z}}\xspace^{n_N}$ to $\ensuremath{\mathbb{Z}}\xspace^{m_1}\oplus\ensuremath{\mathbb{Z}}\xspace^{m_2}\oplus\cdots\oplus\ensuremath{\mathbb{Z}}\xspace^{m_N}$, and for such a homomorphism $B$, we let $B\{ i,j\}$ denote the component of $B$ from the $j$'th direct summand to the $i$'th direct summand.
We also use the notation $B\{i\}$ for $B\{i,i\}$.
Using composition of homomorphisms, we get in a natural way a category $\mathfrak{M}_N$ with objects $\ensuremath{\mathbb{N}}\xspace_0^N$ and with the morphisms from $\mathbf{n}$ to $\mathbf{m}$ being $\MZ$.
Moreover,
$$(BC)\{ i,j\}=\sum_{k=1}^N B \{ i,k\} C\{ k,j\},$$
whenever $B\in\MZ$ and $C\in\MZ[\mathbf{n}\times\mathbf{r}]$ for a multiindex $\mathbf{r}$.
A morphism $B\in\MZ$ is said to be in $\MPZ$, if
$$B\{i,j\}\neq 0\Longrightarrow i\preceq j,$$
for all $i,j\in\ensuremath{\mathcal{P}}\xspace$.
It is easy to verify, that this gives a subcategory $\mathfrak{M}_\ensuremath{\mathcal{P}}\xspace$ with the same objects but $\MPZ$ as morphisms.
Moreover, for a subset $s$ of \ensuremath{\mathcal{P}}\xspace, we let --- with a slight misuse of notation --- $B\{s\}\in\mathfrak{M}_s((m_i)_{i\in s}\times (n_i)_{i\in s},\ensuremath{\mathbb{Z}}\xspace)$ denote the component of $B$ from $\bigoplus_{i\in s}\ensuremath{\mathbb{Z}}\xspace^{n_i}$ to $\bigoplus_{i\in s}\ensuremath{\mathbb{Z}}\xspace^{m_i}$.
We let $\MZ[\mathbf{n}]$ denote $\MZ[\mathbf{n}\times\mathbf{n}]$, and $\MPZ[\mathbf{n}]$ denote $\MPZ[\mathbf{n}\times\mathbf{n}]$.
For $\mathbf{n}$, we let $\GLPZ$ denote the automorphisms in $\MPZ[\mathbf{n}]$.
Then $U\in\GLPZ$ if and only if $U\in\MPZ[\mathbf{n}]$ and $U\{ i\}$ is a group automorphism (meaning that the determinant as a matrix is $\pm 1$ whenever $n_i\neq 0$, for every $i\in\ensuremath{\mathcal{P}}\xspace$).
An automorphism $U\in\GLPZ$ is in $\SLPZ$ if the determinant of $U\{ i\}$ is $1$ for all $i\in\ensuremath{\mathcal{P}}\xspace$ with $n_i\neq 0$.
\end{definition}
\begin{remark}\label{rem:blockmatrices}
Let $\mathbf{m},\mathbf{n}\in\ensuremath{\mathbb{N}}\xspace_{0}^N$ be \emph{multiindices}.
If $|\mathbf{m}|> 0$ and $|\mathbf{n}|> 0$, we can equivalently view the elements $B\in\MZ$ as block matrices
$$B =
\begin{pmatrix}
B\{1,1\} & \dots & B\{1,N\} \\
\vdots & & \vdots \\
B\{N,1\} & \dots & B\{N,N\}
\end{pmatrix}$$
where $B\{i,j\} \in \MZ[m_i\times n_j]$ with $B\{i,j\}$ the empty matrix if $m_{i} = 0$ or $n_{j} = 0$.
Note that from this point of view, the matrices in \MPZ are upper triangular matrices with a certain zero block structure dictated by the order on \ensuremath{\mathcal{P}}\xspace, and the matrices in \GLPZ (respectively \SLPZ) are matrices in \MPZ with all nonempty diagonal blocks having determinant $\pm 1$ (respectively $1$).
Note that if $B\in\MZ$ and $C\in\MZ[\mathbf{n}\times\mathbf{r}]$ for a multiindex $\mathbf{r}$,
then the matrix product makes sense, and --- as matrices --- we have that
\begin{equation}\label{eq:howtomul}
(BC)\{ i,j\}=\sum_{k\in\ensuremath{\mathcal{P}}\xspace, n_k\neq 0} B \{ i,k\} C\{ k,j\},
\end{equation}
for all $i,j\in\ensuremath{\mathcal{P}}\xspace$ with $m_i\neq 0$ and $r_j\neq 0$.
We will therefore also allow ourselves to talk about matrices with no rows or no columns (by considering it as an element of $\MZ[m\times n]$ with $m = 0$ or $n = 0$); and then $B\{s\}$ for a subset $s$ of \ensuremath{\mathcal{P}}\xspace as defined above is just the principal submatrix corresponding to indices in $s$ (remembering the block structure).
\end{remark}
\begin{definition}\label{def:glpandslpeq}
Let $\mathbf{m}$ and $\mathbf{n}$ be multiindices.
Two matrices $B$ and $B'$ in \MPZ are said to be \emph{\GLP-equivalent\xspace} (respectively \emph{\SLP-equivalent\xspace}) if there exist
$U \in\GLPZ[\mathbf{m}]$ and $V \in\GLPZ$ (respectively $U \in\SLPZ[\mathbf{m}]$ and $V \in\SLPZ$) such that
$$U B V = B'.$$
\end{definition}
\begin{definition}\label{def:iotar}
Let $\mathbf{r}=(r_i)_{i=1}^{N}\in \ensuremath{\mathbb{N}}\xspace_{0}^N$ be a multiindex. We now want to define a functor $\iota_{\mathbf{r}}$ from $\mathfrak{M}_N$
to $\mathfrak{M}_N$.
For objects, we let $\iota_{\mathbf{r}}(\mathbf{n})=\mathbf{n}+\mathbf{r}$, for all multiindices $\mathbf{n}\in\ensuremath{\mathbb{N}}\xspace_0^N$.
We define an embedding $\iota_{\mathbf{r}}$ from \MZ to $\MZ[(\mathbf{m}+\mathbf{r})\times(\mathbf{n}+\mathbf{r})]$, for all multiindices $\mathbf{m}=(m_i)_{i=1}^{N}$, $\mathbf{n}=(n_i)_{i=1}^{N}\in\ensuremath{\mathbb{N}}\xspace_{0}^N$,
as follows.
The block $\iota_{\mathbf{r}}(B)\{i,j\}$ has $B\{i,j\}$ as upper left corner.
Outside this corner this block is equal to the zero matrix if $i\neq j$.
If $i=j$, then the lower right $r_i\times r_i$ corner of this (diagonal) block is the identity matrix and zero elsewhere (outside the upper left and lower right corner).
It is easy to check that $\iota_{\mathbf{r}}$ gives a faithful functor from $\mathfrak{M}_N$ to $\mathfrak{M}_N$ that also induces a faithful functor from $\mathfrak{M}_\ensuremath{\mathcal{P}}\xspace$ to $\mathfrak{M}_\ensuremath{\mathcal{P}}\xspace$.
\end{definition}
Note that this is a generalization of the definitions in \cite{MR1907894,MR1990568} (in the finite matrix case) to the cases with rectangular diagonal blocks or vacuous blocks.
\begin{remark}\label{rem:iotar}
We see that \GLPZ and \SLPZ are groups for all multiindices $\mathbf{n}=(n_i)_{i=1}^{N}\in\ensuremath{\mathbb{N}}\xspace_{0}^N$.
We also see that $\iota_{\mathbf{r}}$ is an injective homomorphisms from \GLPZ to $\GLPZ[\mathbf{n}+\mathbf{r}]$ and from \SLPZ to $\SLPZ[\mathbf{n}+\mathbf{r}]$ preserving the identity, for all multiindices $\mathbf{n}$, $\mathbf{r}\in\ensuremath{\mathbb{N}}\xspace_{0}^N$ (since it is a faithful functor). Moreover, $\iota_{\mathbf{r}'}\circ\iota_{\mathbf{r}}=\iota_{\mathbf{r}+\mathbf{r}'}$, for all multiindices $\mathbf{r}$, $\mathbf{r}'\in\ensuremath{\mathbb{N}}\xspace_{0}^N$, and $\iota_{\mathbf{r}}$ is the identity functor whenever $\mathbf{r}=(0,0,\ldots,0)$.
\end{remark}
\subsection{\texorpdfstring{$K$}{K}-web and induced isomorphisms}\label{UVinduce}
We define the $K$-web, $K(B)$, of a matrix $B \in \MPZ$ and describe how a \GLP-equivalence\xspace $\ftn{(U,V)}{B}{B'}$ induces an isomorphism $\ftn{ \kappa_{(U,V)} }{ K(B) }{K(B')}$.
For an element $B\in\MZ[m\times n]$ (\emph{i.e.}\xspace, a group homomorphism $\ftn{B}{\ensuremath{\mathbb{Z}}\xspace^n}{\ensuremath{\mathbb{Z}}\xspace^m}$), we define as usual $\operatorname{cok} B$ to be the abelian group $\ensuremath{\mathbb{Z}}\xspace^m/B\ensuremath{\mathbb{Z}}\xspace^n$ and $\ker B$ to be the abelian group $\setof{x\in\ensuremath{\mathbb{Z}}\xspace^n}{Bx=0}$. Note that if $m=0$, then $\operatorname{cok} B=\{0\}$ and $\ker B=\ensuremath{\mathbb{Z}}\xspace^n$, and if $n=0$, then $\operatorname{cok} B=\ensuremath{\mathbb{Z}}\xspace^m$ and $\ker B=\{0\}$.
For $m,n\in\ensuremath{\mathbb{N}}\xspace_0$, $B,B'\in\MZ[m\times n]$, $U\in\GLZ[m]$ and $V\in\GLZ$ with $UBV=B'$, it is now clear that this equivalence induces isomorphisms
$$
\xymatrix{\operatorname{cok} B \ar[rr]_{\xi_{(U,V)}}^{ [x] \mapsto [Ux] } & &\operatorname{cok} B'} \quad \text{and} \quad \xymatrix{\ker B \ar[rr]_{\delta_{(U,V)}}^{ [x] \mapsto [V^{-1}x] } & & \ker B'.}
$$
\begin{lemma}\label{lem: Kweb 2 components}
Consider $\ensuremath{\mathcal{P}}\xspace=\ensuremath{\mathcal{P}}\xspace_2 = \{1,2\}$ as a partially ordered set and let $B\in\MPZ$.
Then the following sequence
\[
\xymatrix@C=40pt{
\operatorname{cok} B\{1\} \ar[r]^-{[v] \mapsto \left[ \begin{smallpmatrix} v \\ 0 \end{smallpmatrix} \right] } &
\operatorname{cok} B \ar[r]^-{\left[ \begin{smallpmatrix} v \\ w \end{smallpmatrix} \right] \mapsto [w] } &
\operatorname{cok} B\{2\} \ar[d]^0 \\
\ker B\{ 2 \} \ar[u]^-{ v \mapsto [ B\{1,2\}v ] } &
\ker B \ar[l]^-{ w \mapsfrom \begin{smallpmatrix}v \\ w\end{smallpmatrix} } &
\ker B\{ 1 \} \ar[l]^-{ \begin{smallpmatrix}v \\ 0\end{smallpmatrix}\mapsfrom v } }
\]
is exact.
Moreover, if $B$ and $B'$ are elements of \MPZ and $\ftn{ (U,V) }{ B }{ B' }$ is a \GLP-equivalence\xspace, then $(U,V)$ induces an isomorphism
$$(\xi_{(U\{1\},V\{1\})},\xi_{(U,V)},\xi_{(U\{2\},V\{2\})},\delta_{(U\{1\},V\{1\})},\delta_{(U,V)},\delta_{(U\{2\},V\{2\})})$$
of (cyclic six-term) exact sequences.
\end{lemma}
\begin{proof}
The first part of the lemma follows directly from the snake lemma applied to the diagram
$$\xymatrix{
0\ar[r] & \ensuremath{\mathbb{Z}}\xspace^{n_1}\ar[r]\ar[d]^{B\{1\}} & \ensuremath{\mathbb{Z}}\xspace^{n_1}\oplus\ensuremath{\mathbb{Z}}\xspace^{n_2}\ar[r]\ar[d]^{B} & \ensuremath{\mathbb{Z}}\xspace^{n_2}\ar[d]^{B\{2\} }\ar[r] & 0 \\
0\ar[r] & \ensuremath{\mathbb{Z}}\xspace^{m_1}\ar[r] & \ensuremath{\mathbb{Z}}\xspace^{m_1}\oplus\ensuremath{\mathbb{Z}}\xspace^{m_2}\ar[r] & \ensuremath{\mathbb{Z}}\xspace^{m_2}\ar[r] & 0
}$$
The second part of the proof is a straightforward verification.
\end{proof}
Completely analogous to \cite{MR1990568}, we make the following definitions.
\begin{definition}
A subset $c$ of $\ensuremath{\mathcal{P}}\xspace$ is called \emph{convex} if $c$ is nonempty and for all $k \in \ensuremath{\mathcal{P}}\xspace$,
$$\text{$\{i,j\} \subseteq c$ and $i \preceq k \preceq j \ \implies \ k \in c$.}$$
A subset $d$ of $\ensuremath{\mathcal{P}}\xspace$ is called a \emph{difference set} if $d$ is convex and there are convex sets $r$ and $s$ in $\ensuremath{\mathcal{P}}\xspace$ with $r \subseteq s$ such that $d =s \setminus r$ and
\[
\text{$i \in r$ and $j \in d \ \implies \ j \npreceq i$.}
\]
Whenever we have such sets $r$, $s$ and $d=s\setminus r$, we get a canonical functor from $\mathfrak{M}_\ensuremath{\mathcal{P}}\xspace$ to $\mathfrak{M}_{\ensuremath{\mathcal{P}}\xspace_2}$, where $\ensuremath{\mathcal{P}}\xspace_2=\{1,2\}$ with the usual order if there exist $i\in r$ and $j\in d$ such that $i\preceq j$, and the trivial order otherwise.
Thus such sets will also give a canonical (cyclic six-term) exact sequence as above.
\end{definition}
\begin{definition}
Let $B \in \MPZ$.
The \emph{(reduced) $K$-web} of $B$, $K(B)$, consists of a family of abelian groups together with families of group homomorphisms between these, as described below.
For each $i \in \ensuremath{\mathcal{P}}\xspace$, let $r_i = \setof{j\in\ensuremath{\mathcal{P}}\xspace}{j \prec i}$ and $s_{i} = \setof{ j\in\ensuremath{\mathcal{P}}\xspace }{ j \preceq i }$.
Note that if $r_{i}$ in the above definition is nonempty, then $\{ i \} = s_{i} \setminus r_{i}$ is a difference set.
We let $\mathrm{Imm}(i)$ denote the set of immediate predecessors of $i$ (we say that $j$ is an \emph{immediate predecessor of $i$} if $j \prec i$ and there is no $k$ such that $j \prec k \prec i$).
For each $i \in \ensuremath{\mathcal{P}}\xspace$ with $r_i\neq \emptyset$, we get an exact sequence from Lemma~\ref{lem: Kweb 2 components},
\begin{equation}\label{eq:exact-seq-Kweb}
\ker B\{i\}\rightarrow \operatorname{cok} B\{r_i\}\rightarrow \operatorname{cok} B\{s_i\}\rightarrow \operatorname{cok} B\{i\}.
\end{equation}
Moreover, for every pair $(i, j ) \in \ensuremath{\mathcal{P}}\xspace \times \ensuremath{\mathcal{P}}\xspace$ satisfying $j \in \mathrm{Imm}(i)$ and $\mathrm{Imm}(i) \setminus \{ j \} \neq \emptyset$ is $s_{j} \subsetneq r_{i}$; consequently we have a homomorphism
\begin{equation}\label{eq: Kweb hom}
\operatorname{cok} B\{s_j\} \to \operatorname{cok} B\{r_i\}
\end{equation}
originating from the exact sequence above (\emph{cf.}\xspace\ Lemma~\ref{lem: Kweb 2 components} used on the division into the sets $r_i$, $s_j$ and $r_i\setminus s_j$).
Set
\begin{align*}
I_0^{\ensuremath{\mathcal{P}}\xspace} &=
\setof{ r_i }{ i \in \ensuremath{\mathcal{P}}\xspace \text{ and }r_{i} \neq \emptyset } \cup \setof{ s_i }{ i \in \ensuremath{\mathcal{P}}\xspace } \cup \setof{ \{ i \} }{ i \in \ensuremath{\mathcal{P}}\xspace }, \\
I_1^{\ensuremath{\mathcal{P}}\xspace} &= \setof{i\in\ensuremath{\mathcal{P}}\xspace}{r_i\neq \emptyset}.
\end{align*}
The \emph{$K$-web of $B$}, denoted by $K(B)$, consists of the families $\left( \operatorname{cok} B\{c\} \right)_{ c \in I_0^{\ensuremath{\mathcal{P}}\xspace}}$ and $\left( \ker B\{i\} \right)_{ i \in I_1^\ensuremath{\mathcal{P}}\xspace}$ together with all the homomorphisms from the sequences \eqref{eq:exact-seq-Kweb} and \eqref{eq: Kweb hom}. Let $B'$ be an element of \MPZ[\mathbf{m}'\times\mathbf{n}']. By a \emph{$K$-web isomorphism}, $\ftn{\kappa}{ K(B) }{ K(B') }$, we mean families
$$\left( \ftn{ \kappa_{c,0} }{ \operatorname{cok} B\{c\} }{ \operatorname{cok} B'\{c\} } \right)_{ c \in I_0^\ensuremath{\mathcal{P}}\xspace}$$
and $$\left( \ftn{ \kappa_{i,1} }{ \ker B\{i\} }{ \ker B'\{i\} } \right)_{ i \in I_1^\ensuremath{\mathcal{P}}\xspace }$$ of isomorphisms satisfying that the ladders coming from the sequences in $K(B)$ and $K(B')$ commute.
By Lemma~\ref{lem: Kweb 2 components}, any \GLP-equivalence\xspace $\ftn{ (U,V) }{ B }{ B' }$ induces a $K$-web isomorphism from $B$ to $B'$. We denote this induced isomorphism by $\kappa_{(U,V)}$.
\end{definition}
\begin{remark}\label{diffKwebFK} We note the obvious likeness
between the $K$-web and the reduced filtered
$K$-theory. There are two fundamental differences: In $K(B)$ we
never consider orders, and the groups $\ker B\{i\}$ are only
appearing in $K(B)$ when $\{i\}\not=s_i$, whereas the
corresponding $K_1$-group always appears in
$\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(C^*(\mathsf{E}_{B+I}))$.\end{remark}
\begin{remark}\label{rem:iotarandidentifications}
It is clear that the $K$-webs $K(B)$ and $K(\iota_{\mathbf{r}}(B))$ are canonically isomorphic for all multiindices $\mathbf{m},\mathbf{n},\mathbf{r}\in\ensuremath{\mathbb{N}}\xspace_0^N$ and all $B\in\MPZ$.
Note also that the $K$-webs $K(B)$ and $K(-B)$ are canonically isomorphic, and that $(U,V)$ is a \GLP-equivalence\xspace (respectively \SLP-equivalence\xspace) from $B$ to $B'$ if and only if $(U,V)$ is a \GLP-equivalence\xspace (respectively \SLP-equivalence\xspace) from $-B$ to $-B'$, and they will induce exactly the same $K$-web isomorphisms under the above identification. Note that this identification will change the generators of the cokernels and the kernels.
In this way, we also get a canonical identification of the $K$-webs $K(B)$ and $K(-\iota_{\mathbf{r}}(-B))$ by embedding a vector by setting it to be zero on the new coordinates. This identification preserves the canonical generators of the cokernels and kernels, which will be of importance when we consider positivity.
\end{remark}
\begin{remark} \label{rem:twominusversusnominusiniotar}
The definitions above are completely analogous to the definitions in \cite{MR1990568}, and are the same in the case $m_{i} = n_{i} \neq 0$ for all $i\in\ensuremath{\mathcal{P}}\xspace$.
Note that the last homomorphism in \eqref{eq:exact-seq-Kweb} is really not needed, because commutativity with this map is automatic.
The reason we need to use $K(-\iota_{\mathbf{r}}(-B))$ rather than
$K(\iota_{\mathbf{r}}(B))$ (as in \cite{MR1907894,MR1990568}), is that
we let $B=\mathsf{B}_E^\bullet$, where $\mathsf{B}_E=\mathsf{A}_E-I$ rather than
$I-\mathsf{A}_E$ (as done in \cite{MR1907894,MR1990568}). One of the
benefits with this approach is that it is somewhat more convenient to
work with positive matrices instead of negative matrices --- and Boyle
actually does this partly himself in his proof of the factorization
theorem, \emph{cf.}\xspace\ \cite[Section~4]{MR1907894}). The reason that we do not
define $\iota_\mathbf{r}$ as extending by $-1$'s instead of $1$'s is
crucial. This would force us to have one definition of embeddings for
the \ensuremath{\operatorname{GL}_\calP}\xspace and \ensuremath{\operatorname{SL}_\calP}\xspace-matrices used for \GLP-equivalence\xspace{s} and \SLP-equivalence\xspace{s} and
another for the matrices arriving from the adjacency
matrices. Moreover, such a definition would not give a functor. Both
these problems would be very inconvenient for our work. Thus this is a
matter of choosing either to have the convenience of working with
positive matrices or to not need the two minuses in
$K(-\iota_{\mathbf{r}}(-B))$. We have chosen to use the former
convention.
\end{remark}
\subsection{Block structure for graphs}
\begin{definition}\label{def:circ}
Let $E=(E^0,E^1,r,s)$ be a graph.
We write $\mathsf{B}_E\in\MPZc$ if
\begin{itemize}
\item
\ensuremath{\mathcal{P}}\xspace satisfies Assumption~\ref{ass:preorder}, and there is an isomorphism $\mytheta$ from \ensuremath{\mathcal{P}}\xspace to $\Gamma_E$ such that $\mytheta$ and $\mytheta^{-1}$ are order reversing,\item
$E$ has finitely many vertices,
\item
every infinite emitter emits infinitely many edges to any vertex it emits any edge to,
\item
every transition state has exactly one edge going out,
\item
$\mathsf{B}_E$ is an $\mathbf{n}\times\mathbf{n}$ block matrix where the vertices of the $i$'th block correspond exactly to the set $\overline{H(\mytheta(i))}\setminus\overline{H(\mytheta(i))\setminus \mytheta(i)}$, and
\item
$\mathsf{B}_E^\bullet\in\MPZ$.
\end{itemize}
We write $\mathsf{B}_E\in\MPZcc$ if $\mathsf{B}_E\in\MPZc$ and $E$ does not have any transition states, and we write $\mathsf{B}_E\in\MPZccc$ if $\mathsf{B}_E\in\MPZcc$ and $|\gamma|=1$, for every cyclic component $\gamma\in\Gamma_E$.
According to Lemma~\ref{lem:structure-2}\ref{lem:structure-2-6}, \ref{lem:structure-2-7} and \ref{lem:structure-2-8}, for every graph $E$ with finitely many vertices, there exist graphs $E'$, $E''$ and $E'''$ such that $\operatorname{Prime}_\gamma(C^*(E))\cong\operatorname{Prime}_\gamma(C^*(E'))\cong\operatorname{Prime}_\gamma(C^*(E''))\cong \operatorname{Prime}_\gamma(C^*(E'''))$ in a canonical way and $\mathsf{B}_{E'}\in\MPZc[\mathbf{m}'\times\mathbf{n}']$, $\mathsf{B}_{E''}\in\MPZcc[\mathbf{m}''\times\mathbf{n}'']$, $\mathsf{B}_{E'''}\in\MPZccc[\mathbf{m}'''\times\mathbf{n}''']$, $C^*(E)\cong C^*(E')$, $C^*(E)\otimes\ensuremath{\mathbb{K}}\xspace \cong C^*(E'')\otimes\ensuremath{\mathbb{K}}\xspace$ and $C^*(E)\otimes\ensuremath{\mathbb{K}}\xspace \cong C^*(E''')\otimes\ensuremath{\mathbb{K}}\xspace$ via equivariant isomorphisms.
If we have $\mathsf{B}_E\in\MPZc$ and $\mathsf{B}_{E'}\in\MPZc[\mathbf{m}'\times\mathbf{n}']$, then we say that a \mbox{$\sp*$-}ho\-mo\-morphism\xspace $\Phi$ from $C^*(E)$ to $C^*(E')$ (or from $C^*(E)\otimes \ensuremath{\mathbb{K}}\xspace$ to $C^*(E')\otimes \ensuremath{\mathbb{K}}\xspace$) is \ensuremath{\mathcal{P}}\xspace-equivariant if $\Phi$ is $\operatorname{Prime}_\gamma(C^*(E))$-equivariant under the canonical identification $\operatorname{Prime}_\gamma(C^*(E))\cong\Gamma_E\cong\ensuremath{\mathcal{P}}\xspace\cong\Gamma_{E'}\cong \operatorname{Prime}_\gamma(C^*(E'))$ coming from the block structure.
\end{definition}
Note that the conditions above are not only conditions on the graph --- they are also conditions on the adjacency matrix and how we write it (indexed over $\{1,\ldots,|E^0|\}$). In addition to some assumptions about the graph, we choose a specific order of the vertices and index them over $\{1,\ldots,|E^0|\}$ and we have then implicitly chosen an isomorphism $\Gamma_E\cong\ensuremath{\mathcal{P}}\xspace$, for some appropriate order on $\ensuremath{\mathcal{P}}\xspace=\{1,2,\ldots,|\Gamma_E|\}$.
In general, there might be many different such isomorphisms for the same partially ordered set $\ensuremath{\mathcal{P}}\xspace$ that work depending on the order chosen of the vertices (if $\ensuremath{\mathcal{P}}\xspace$ admits a nontrivial automorphism), and it might also be possible to choose an order reversing isomorphism $\Gamma_E\cong\ensuremath{\mathcal{P}}\xspace'$, where $\ensuremath{\mathcal{P}}\xspace'$ has a different order than $\ensuremath{\mathcal{P}}\xspace$.
\subsection{Reduced filtered \texorpdfstring{$K$}{K}-theory, \texorpdfstring{$K$}{K}-web and \texorpdfstring{\GLP-equivalence\xspace}{GLP-equivalence}}
\label{sec:red-filtered-K-theory-K-web-GLP-and-SLP-equivalences}
Let $(\ensuremath{\mathcal{P}}\xspace, \preceq )$ be a partially ordered set that satisfies Assumption~\ref{ass:preorder}.
We let $\ensuremath{\mathcal{P}}\xspace^\mathsf{T}$ denote the set $\ensuremath{\mathcal{P}}\xspace$ with order defined by $i\preceq^\mathsf{T} j$ in $\ensuremath{\mathcal{P}}\xspace^\mathsf{T}$ if and only if $N+1-j\preceq N+1-i$ in $\ensuremath{\mathcal{P}}\xspace$, for $i=1,2,\ldots,N$.
The partially ordered set $(\ensuremath{\mathcal{P}}\xspace^\mathsf{T},\preceq^\mathsf{T})$ is really the set $\ensuremath{\mathcal{P}}\xspace$ equipped with the opposite order, followed by a permutation to ensure that it satisfies
Assumption~\ref{ass:preorder}.
For every multiindex $\mathbf{m}=(m_1,m_2,\ldots,m_N)$ we let $\mathbf{m}^\mathsf{T}=(m_N,\ldots,m_2,m_1)$ and we let $J_\mathbf{m}$ denote the $|\mathbf{m}|\times|\mathbf{m}|$ permutation matrix that reverses the order.
Now assume that we have a graph $E$ with finitely many vertices such that $\mathsf{B}_E\in\MPZcc[\mathbf{m}_E \times \mathbf{n}_E]$.
It is easy to see, that $\mathsf{B}_{E}^{\bullet} \in \MPZ[\mathbf{m}_E\times \mathbf{n}_E]$ is equivalent to $J_{\mathbf{n}_E}\left(\mathsf{B}_{E}^\bullet\right)^\mathsf{T}J_{\mathbf{m}_E}\in\mathfrak{M}_{\ensuremath{\mathcal{P}}\xspace^\mathsf{T}}(\mathbf{n}_E^\mathsf{T}\times\mathbf{m}_E^\mathsf{T},\ensuremath{\mathbb{Z}}\xspace)$.
Now assume that we also have a graph $F$ with finitely many vertices such that $\mathsf{B}_F\in\MPZcc[\mathbf{m}_F \times \mathbf{n}_F]$.
For notational convenience, we let
\begin{align*}
\mathsf{C}_E&=J_{\mathbf{n}_E}\left(\mathsf{B}_{E}^\bullet\right)^\mathsf{T}J_{\mathbf{m}_E} \\
\mathsf{C}_F&=J_{\mathbf{n}_F}\left(\mathsf{B}_{F}^\bullet\right)^\mathsf{T}J_{\mathbf{m}_F}.
\end{align*}
With the usual description of the $K$-theory and six term exact sequences for graph \mbox{$C\sp*$-}al\-ge\-bras\xspace (\emph{cf.}\xspace\ \cite{MR2922394}), we see that a reduced filtered $K$-theory isomorphism from $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(\ensuremath{\mathcal{P}}\xspace;C^*(E))$ to $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(\ensuremath{\mathcal{P}}\xspace;C^*(F))$ corresponds exactly to a (reduced) $K$-web isomorphism from $K(\mathsf{C}_E)$ to
$K(\mathsf{C}_F)$ together with an isomorphism from
$\ker(\mathsf{C}_E\{i\})$
to $\ker (\mathsf{C}_F\{i\})$ for every $i\in(\ensuremath{\mathcal{P}}\xspace^{\mathsf{T}})_{\min{}}$, where $\ensuremath{\mathcal{P}}\xspace_{\min}=\setof{i\in\ensuremath{\mathcal{P}}\xspace}{ j\preceq i\Rightarrow i=j}$ and $(\ensuremath{\mathcal{P}}\xspace^\mathsf{T})_{\min}=\setof{i\in\ensuremath{\mathcal{P}}\xspace}{ j\preceq^\mathsf{T} i\Rightarrow i=j}$.
Positivity is easy to describe on the gauge simple subquotients. For
components with a vertex supporting at least two distinct return
paths, the positive cone is all of $K_0$ (since the corresponding
subquotient is a Kirchberg algebra in the UCT class). For components
where each vertex supports exactly one return path, the positive cone
is generated by the class of the projections $p_v$, where $v$ are in
this component (the corresponding subquotient is stably isomorphic to
$C(S^1)$). If such a cyclic component is a singleton, the ordered
$K_0$-group is $(\ensuremath{\mathbb{Z}}\xspace,\ensuremath{\mathbb{N}}\xspace_0)$ under the canonical identifications. For
components consisting of a single singular vertex not supporting a
cycle, the positive cone is generated by the class of the projection
$p_v$, where $v$ is the vertex in the component (in this case the
subquotient is stably isomorphic to $\ensuremath{\mathbb{K}}\xspace$). Under the canonical
identifications, the ordered $K_0$-group is also $(\ensuremath{\mathbb{Z}}\xspace,\ensuremath{\mathbb{N}}\xspace_0)$. The
description of the $K_0$-groups for gauge nonsimple subquotients (or
just gauge nonsimple ideals), turns out to be more complicated in
general (when the \mbox{$C\sp*$-}al\-ge\-bra\xspace is not purely infinite) --- see
\cite[Theorem~2.2]{MR1962131} for a general description of the
order. As it turns out, only the order of the gauge simple
subquotients will play a role --- as a result of our classification
result, we see that the information stored in the order of the other
groups is redundant (see
Remark~\ref{onlyorderongs}).
We see that a necessary condition for having an isomorphism between the reduced filtered $K$-theories is that $\mathbf{n}_E-\mathbf{m}_E=\mathbf{n}_F-\mathbf{m}_F$. So assume this holds, and choose $\mathbf{m},\mathbf{n}\in\ensuremath{\mathbb{N}}\xspace_0^N$ such that $\mathbf{m}_E,\mathbf{m}_F\leq\mathbf{m}$ and $\mathbf{n}_E,\mathbf{n}_F\leq\mathbf{n}$, and $\mathbf{n}-\mathbf{m}=\mathbf{n}_E-\mathbf{m}_E=\mathbf{n}_F-\mathbf{m}_F$.
Let $\mathbf{r}_E=\mathbf{m}-\mathbf{m}_E=\mathbf{n}-\mathbf{n}_E$ and let $\mathbf{r}_F=\mathbf{m}-\mathbf{m}_F=\mathbf{n}-\mathbf{n}_F$.
Then the $K$-webs of $K(-\iota_{\mathbf{r}_E^{\mathsf T}}(-\mathsf{C}_E))$ and $K(-\iota_{\mathbf{r}_F^{\mathsf T}}(-\mathsf{C}_F))$ are canonically isomorphic to $K(\mathsf{C}_E)$ and $K(\mathsf{C}_F)$, respectively, and
$\ker(\mathsf{C}_E\{i\})$ and $\ker(\mathsf{C}_F\{i\})$ are canonically isomorphic to $\ker (-\iota_{\mathbf{r}_E^{\mathsf T}}(-\mathsf{C}_E)\{i\})$ and $\ker (-\iota_{\mathbf{r}_F^{\mathsf T}}(-\mathsf{C}_F)\{i\})$ for every $i\in(\ensuremath{\mathcal{P}}\xspace^{\mathsf{T}})_{\min{}}$.
We see that a necessary condition for having a positive isomorphism between the reduced filtered $K$-theories is that under the isomorphisms $\Gamma_E\cong\ensuremath{\mathcal{P}}\xspace\cong\Gamma_{F}$ we have exactly the same strongly connected components, the same cyclic strongly connected components, the same sinks, and the same infinite emitters not supporting a cycle.
It is clear that $(U,V)\mapsto((J_{\mathbf{n}}
VJ_{\mathbf{n}})^\mathsf{T},(J_{\mathbf{m}}
UJ_{\mathbf{m}})^\mathsf{T})$ gives a one-to-one correspondence
between \GLP-equivalence\xspace{s} (respectively \SLP-equivalence\xspace{s}) from
$-\iota_{\mathbf{r}_E}(-\mathsf{B}_E^\bullet)$ to
$-\iota_{\mathbf{r}_F}(-\mathsf{B}_F^\bullet)$ and
$\operatorname{GL}_{\ensuremath{\mathcal{P}}\xspace^\mathsf{T}}$-equivalences (respectively
$\operatorname{SL}_{\ensuremath{\mathcal{P}}\xspace^\mathsf{T}}$-equivalences) from
$-\iota_{\mathbf{r}_E^{\mathsf T}}(-\mathsf{C}_E)$ to $-\iota_{\mathbf{r}_F^{\mathsf T}}(-\mathsf{C}_F)$.
So every \GLP-equivalence\xspace $(U,V)$ from $-\iota_{\mathbf{r}_E}(-\mathsf{B}_E^\bullet)$ to $-\iota_{\mathbf{r}_F}(-\mathsf{B}_F^\bullet)$ will determine a reduced filtered $K$-theory isomorphism from $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(\ensuremath{\mathcal{P}}\xspace;C^*(E))$ to $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(\ensuremath{\mathcal{P}}\xspace;C^*(F))$.
We call this isomorphism $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(U,V)$.
In particular, if $\mathbf{m}=\mathbf{m}_E=\mathbf{m}_F$ and $\mathbf{n}=\mathbf{n}_E=\mathbf{n}_F$, then every \GLP-equivalence\xspace $(U,V)$ from $\mathsf{B}_E^\bullet$ to $\mathsf{B}_F^\bullet$ will determine a reduced filtered $K$-theory isomorphism $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(U,V)$ from $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(\ensuremath{\mathcal{P}}\xspace;C^*(E))$ to $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(\ensuremath{\mathcal{P}}\xspace;C^*(F))$.
Note that $V^\mathsf{T}$ induces the isomorphisms between the $K_0$-groups while $(U^\mathsf{T})^{-1}$ induces the isomorphisms between the $K_1$-groups with the standard identification of the $K$-groups.
Note that the hereditary subsets of vertices --- as usually defined for graphs, when we consider graph \mbox{$C\sp*$-}al\-ge\-bras\xspace{} --- correspond to subsets $S$ of $\ensuremath{\mathcal{P}}\xspace$ satisfying that $i\preceq j$ implies that $j\in S$ whenever $i\in S$ (\emph{cf.}\xspace\ the order reversing bijection between \ensuremath{\mathcal{P}}\xspace and $\Gamma_E$ in Definition~\ref{def:circ}).
This is due to that fact that we generally do not work with the transposed matrix in this paper, since we find it more convenient to work with the non-transposed matrix.
Since we are identifying \ensuremath{\mathcal{P}}\xspace with $\Gamma_E$ using an order reversing isomorphism, we will avoid using terms as minimal, maximal, less than or greater than. We have already introduced the term (immediate) predecessor for elements of \ensuremath{\mathcal{P}}\xspace. We will define (immediate) successors in the analogous way.
But we will use the term that $\gamma_1$ is a predecessor of $\gamma_2$ if and only if $\gamma_2$ is a successor of $\gamma_1$ if and only if $\gamma_1\geq\gamma_2$ and $\gamma_1\neq\gamma_2$. Immediate predecessor and immediate successor in $\Gamma_E$ is defined accordingly. This use of the language also fits better with our usual picture of the component set as a graph: if $\gamma_2$ is a successor of $\gamma_1$ this means that there is a path from component $\gamma_1$ to component $\gamma_2$.
\subsection{Temperatures and standard form}
Let $E$ be a graph with finitely many vertices.
Then $\operatorname{Prime}_\gamma(C^*(E))$ is finite and the gauge simple subquotients are $C^*(E)(\{\mathfrak{p}\})$ for $\mathfrak{p}\in\operatorname{Prime}_\gamma(C^*(E))$.
These are either simple AF algebras, simple purely infinite \mbox{$C\sp*$-}al\-ge\-bras\xspace or nonsimple.
They are stably isomorphic to $C(S^1)$, when they are nonsimple.
\begin{definition}
Let $E$ be a graph with finitely many vertices.
Then we define the \emph{temperature} as the map $\fct{\tau_E}{\operatorname{Prime}_\gamma(C^*(E))}{\{-1,0,1\}}$ defined by
$$\tau_E(\mathfrak{p})
=\begin{cases}
-1,&\text{if }C^*(E)(\{\mathfrak{p}\})\text{ is a simple AF algebra},\\
0,&\text{if }C^*(E)(\{\mathfrak{p}\})\text{ is nonsimple},\\
1,&\text{if }C^*(E)(\{\mathfrak{p}\})\text{ is simple and purely infinite},
\end{cases}$$
where $\mathfrak{p}\in\operatorname{Prime}_\gamma(C^*(E))$.
We call $(\operatorname{Prime}_\gamma(C^*(E)),\tau_E)$ the \emph{tempered (gauge invariant) prime ideal space}.
Let $E$ and $F$ be graphs with finitely many vertices.
Then an isomorphism $\fct{\Theta}{(\operatorname{Prime}_\gamma(C^*(E)),\tau_E)}{(\operatorname{Prime}_\gamma(C^*(F)),\tau_F)}$ is a homeomorphism \fctw{\Theta}{\operatorname{Prime}_\gamma(C^*(E))}{\operatorname{Prime}_\gamma(C^*(F))} satisfying that $\tau_F\circ\Theta=\tau_E$.
We write $(\operatorname{Prime}_\gamma(C^*(E)),\tau_E)\cong(\operatorname{Prime}_\gamma(C^*(F)),\tau_F)$ when such an isomorphism exists.
\end{definition}
We note from the outset that $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(\operatorname{Prime}_\gamma(C^*(E));C^*(E))$ contains the temperature.
\begin{lemma}\label{taufromK}
Let $E$ and $F$ be graphs with finitely many vertices, let $X=\operatorname{Prime}_\gamma(C^*(E))$ and assume that there is a homeomorphism \fctw{\Theta}{X}{\operatorname{Prime}_\gamma(C^*(F))}. View $C^*(E)$ and $C^*(F)$ as $X$-algebras in the canonical way and assume that there is an isomorphism from $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(X;C^*(E))$ to $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(X;C^*(F))$. Then $\tau_F\circ\Theta=\tau_E$, so $(\operatorname{Prime}_\gamma(C^*(E)),\tau_E)\cong (\operatorname{Prime}_\gamma(C^*(F)),\tau_F)$.
\end{lemma}
\begin{proof}
We read off the temperatures from the ordered, reduced filtered $K$-theory as
\begin{eqnarray*}
\tau_E(\mathfrak{p})=-1&\Longleftrightarrow&K_0\not=(K_0)_+\wedge K_1=0,\\
\tau_E(\mathfrak{p})=0&\Longleftrightarrow&K_0\not=(K_0)_+\wedge K_1\not =0, \text{ and}\\
\tau_E(\mathfrak{p})=1&\Longleftrightarrow&K_0=(K_0)_+,
\end{eqnarray*}
where $K_*=K_*(C^*(E)(\{\mathfrak{p}\}))$.
\end{proof}
\begin{remark}\label{countremark}
In the case of graphs with finitely many vertices such that every infinite emitter emits infinitely many edges to any vertex it emits any edge to --- in particular for finite graphs --- we have a canonical homeomorphism \fctw{{\upsilon}_E}{\Gamma_E}{\operatorname{Prime}_\gamma(C^*(E))} (\emph{cf.}\xspace\ Lemma~\ref{lem:structure-d}). Thus we can in this case equally well consider the space $(\Gamma_E,\tau_E\circ{\upsilon}_E)$ as the tempered gauge invariant prime ideal space.
Note that if we for $\gamma\in\Gamma_E$ let
$$\gamma^1=\setof{e\in E^1}{r(e),s(e)\in\gamma}\subseteq E^1,$$
then $(\tau_E\circ{\upsilon}_E)(\gamma)=\operatorname{sgn}(|\gamma^1|-|\gamma|)$ if we use the conventions $\operatorname{sgn}(0)=0$ and $\operatorname{sgn}(\infty)=1$.
\end{remark}
It follows that $(\operatorname{Prime}_\gamma(C^*(E)),\tau_E)\cong(\operatorname{Prime}_\gamma(C^*(F)),\tau_F)$ whenever $E\ensuremath{\sim_{C\negthinspace E}}\xspace F$, because in this case $C^*(E)\otimes\ensuremath{\mathbb{K}}\xspace\cong C^*(F)\otimes\ensuremath{\mathbb{K}}\xspace$. It is not hard, but somewhat tedious, to check directly that the allowed moves will not change the signs of the numbers $|\gamma^1|-|\gamma|$ occurring.
\begin{definition}
Let $E$ be a graph.
We say that $E$ satisfies \emph{Condition (H)} if for any regular vertex $v$ supporting a unique return path, either this path has no exit, or there is a vertex $w\not=v$ which is singular or supports a unique return path so that there is a path from $v$ to $w$, and so that any path from $v$ to $w$ passes through vertices not supporting two distinct return paths (in particular, $w$ cannot support two distinct return paths).
\end{definition}
Under the assumption that the (finite) graphs satisfy Condition (H), we will prove that every stable isomorphism at the level of graph \mbox{$C\sp*$-}al\-ge\-bras\xspace may be realized by the moves defining $\ensuremath{\sim_{C\negthinspace E}}\xspace$. Note that among the graphs in Figure \ref{firstexx}, the two in (a) have Condition (H) whereas the remaining four do not. Also note that Condition (H) in a sense interpolates between Condition (K) and the case when no vertex has more than one return path, and is met in both cases.
\begin{lemma}\label{charKH}
Let $E$ be a graph with finitely many vertices.
Then the following holds.
\begin{enumerate}[(i)]
\item \label{charKH-1}
$E$ satisfies Condition~(K) if and only if $\tau_E(\mathfrak{p})\neq 0$ for every $\mathfrak{p}\in\operatorname{Prime}_\gamma(C^{*}(E))$.
\item \label{charKH-2}
$E$ has no vertices supporting two distinct return paths if and only if $\tau_E(\mathfrak{p})\leq 0$ for every $\mathfrak{p}\in\operatorname{Prime}_\gamma(C^{*}(E))$.
\item \label{charKH-3}
$E$ satisfies Condition (H) if and only if whenever $\tau_E(\mathfrak{p})=0$ then either $\operatorname{Imm}(\mathfrak{p})=\emptyset$ or there is a $\mathfrak{p}'\in\operatorname{Imm}(\mathfrak{p})$ with $\tau_E(\mathfrak{p}')\leq 0$.
\end{enumerate}
If every infinite emitter emits infinitely many edges to any vertex it emits any edge to, then \ref{charKH-3} can be replaced by
\begin{enumerate}[(i')]\addtocounter{enumi}{2}
\item\label{charKH-4}
$E$ satisfies Condition~(H) if and only if whenever $\tau_E({\upsilon}_E(\gamma))= 0$ then either $\gamma$ has no successor in $\Gamma_E$, or
$\gamma$ has an immediate successor $\gamma'\in\Gamma_E$ with $\tau_E({\upsilon}_E(\gamma'))\leq 0$.
\end{enumerate}
\end{lemma}
\begin{proof}
We start by assuming that every infinite emitter in $E$ emits infinitely many edges to any vertex it emits any edge to, so that Remark \ref{countremark} applies. In this case,
\ref{charKH-1} and \ref{charKH-2} are obvious, and \ref{charKH-3} and \ref{charKH-4} are equivalent. For \ref{charKH-4}, assume first that the condition on $\tau_E$ holds. To show (H), let $v$ support a unique return path with an exit and note that $v$ then lies in some $\gamma$ with $\tau_E({\upsilon}_E(\gamma))=0$ where $\gamma$ has a successor in $\Gamma_E$. One such successor $\gamma'$ must be immediate with $\tau_E({\upsilon}_E(\gamma'))\leq 0$, and we take $w\in\gamma'$. Then any path from $v$ to $w$ passes through only transitional vertices and vertices in $\gamma\cup\gamma'$, neither of which supports multiple return paths. Finally, if $w$ is regular, then since it is not a transitional vertex, it must support a unique return path.
In the other direction, assume that the condition on $\tau_E$ fails
and choose $\gamma\in\Gamma_E$ with the property that $\tau_E({\upsilon}_E(\gamma))=0$, $\gamma$ has successors and that all such immediate successors $\gamma'$ have
$\tau_E({\upsilon}_E(\gamma'))=1$. We conclude that any path from $v\in\gamma$ to any $w$ not
a transitional vertex must pass through a vertex supporting at least
two different return paths. It remains to check that $v$ cannot be a singular vertex.
But since $v\in\gamma$ and $\tau_E({\upsilon}_E(\gamma))=0$, $v$ supports a unique return path. Thus $v$ emits finitely many edges to a vertex in $\gamma$; hence $v$ is regular.
For general graphs with finitely many vertices, we note that by Lemma \ref{lem:structure-2}\ref{lem:structure-2-6} (and its proof),
we may replace $E$ by $E'$ with the property that every infinite emitter in $E'$ emits infinitely many edges to any vertex it emits any edge to, in the sense that $C^*(E)\cong C^*(E')$ and $E'$ is obtained from $E$ by a number of moves of type \mbox{\texttt{\textup{(O)}}}\xspace. Since these operations preserve all the conditions on the graphs, the result follows.
\end{proof}
According to \cite{MR2069031}, the
conditions in \ref{charKH-1} above translate exactly to $C^*(E)$ being of real
rank zero. According to \cite{MR2001940}, the conditions in \ref{charKH-2} above
translate exactly to $C^*(E)$ being a type I/postliminal \mbox{$C\sp*$-}al\-ge\-bra\xspace.
\begin{notation}
Let $E$ be a graph with finitely many vertices and assume that $\mathsf{B}_E\in\MPZc$.
This induces a temperature $\mytau=\tau_E\circ{\upsilon}_E\circ\mytheta$ on $\ensuremath{\mathcal{P}}\xspace$.
\end{notation}
It will be extremely convenient for us to know that the adjacency matrices for two graphs are aligned with all components having the same number of vertices. For this, we define:
\begin{definition}\label{standardform}
Let $E$ and $F$ be finite graphs. We say that $(\mathsf{B}_E,\mathsf{B}_F)$ is in \emph{standard form} if $\mathsf{B}_E,\mathsf{B}_F\in\MPZccc$ for some multiindices $\mathbf{m}$ and $\mathbf{n}$ and $\mathcal{T}_{\mathsf{B}_E}=\mathcal{T}_{\mathsf{B}_F}$. This means that the adjacency matrices have exactly the same sizes and block structures, and that the temperatures of the components match up.
\end{definition}
\begin{definition}\label{def: positive matrices}
Define \MPplusZ to be the set of all $B \in \MPZ$ satisfying the following:
\begin{enumerate}[(i)]
\item \label{def: positive matrices-1}
If $i \prec j$ and $B \{ i, j \}$ is not the empty matrix, then $B \{ i , j \} > 0$.
\item \label{def: positive matrices-2}
If $m_{i} = 0$, then $n_{i} =1$.
\item \label{def: positive matrices-3}
If $m_{i} = 1$, then $n_{i} = 1$ and $B \{ i \} = 0$.
\item \label{def: positive matrices-4}
If $m_{i} > 1$, then $B \{ i \} > 0$, $n_{i},m_{i} \geq 3$, and the Smith normal form of $B\{i\}$ has at least two 1's (and thus the rank of $B \{ i \}$ is at least 2). \end{enumerate}
\end{definition}
\begin{lemma}\label{taugivesstd}
Let $E$ and $F$ be two finite graphs. The following are equivalent
\begin{enumerate}[(1)]
\item $(\operatorname{Prime}_\gamma(C^*(E)),\tau_E)\cong(\operatorname{Prime}_\gamma(C^*(F)),\tau_F)$.\label{taugivesstdI}
\item We can choose finite graphs $E'$ and $F'$ so that $(\mathsf{B}_{E'},\mathsf{B}_{F'})$ is in standard form and so that $E\ensuremath{\sim_{M\negthinspace E}}\xspace E'$ and $F\ensuremath{\sim_{M\negthinspace E}}\xspace F'$.\label{taugivesstdII}
\end{enumerate}
In \ref{taugivesstdII}, we may further assume that $\mathsf{B}_{E'}^\bullet,\mathsf{B}_{F'}^\bullet\in\MPplusZ$.
\end{lemma}
\begin{proof}
When $\mathsf{B}_E\in\MPZccc$, we may read off the temperatures of $i\in\ensuremath{\mathcal{P}}\xspace$ by the rules
\begin{eqnarray*}
\mytau(i)=-1&\Longleftrightarrow&m_i=0,\\
\mytau(i)=0&\Longleftrightarrow&m_i=1\text{ and }\mathsf{B}_E\{i\}=0, \text{ and}\\
\mytau(i)=1&\Longleftrightarrow&\text{either }m_i=1\text{ and }\mathsf{B}_E\{i\}>0\text{ or }m_i>1.
\end{eqnarray*}
Thus, \ref{taugivesstdII}$\Longrightarrow$\ref{taugivesstdI} follows from the move invariance of the temperature.
For the other direction, assume \ref{taugivesstdI}. It follows from Lemma~\ref{lem:structure-2} that we can assume that $\mathsf{B}_E\in\MPZccc[\mathbf n'\times\mathbf m']$ and
$\mathsf{B}_F\in\MPZccc[\mathbf n''\times\mathbf m'']$ for appropriate
$\ensuremath{\mathcal{P}}\xspace$, $\mathbf m'$, $\mathbf m''$, $\mathbf n'$ and $\mathbf
n''$ with ${\upsilon}_{F}\circ\mytheta[F]\circ\mytheta^{-1}\circ{\upsilon}_{E}^{-1}$ being the isomorphism given in \ref{taugivesstdI} that intertwines the temperatures. By assumption,
$n_i'=n_i''=m_i'=m_i''=1$ when $\mytau(i)=\mytau[F](i)=0$, and $n_i'=n_i''=1$,
$m_i'=m_i''=0$ when $\mytau(i)=\mytau[F](i)=-1$.
When $\mytau(i)=\mytau[F](i)=1$, we may perform Move \mbox{\texttt{\textup{(Col)}}}\xspace inside each of these components until we get vertices $u_i^E$ and $u_i^F$ which support loops.
Since the components are not cyclic, $u_i^E$ and $u_i^F$ emit at least one other edge than the loop to a vertex in the component, and we may perform \mbox{\texttt{\textup{(R)}}}\xspace in reverse on them successively to increase the sizes of the block to arrive at $n_i'=m_i'=n_i''=m_i''\geq 3$. By doing this (at most) twice more we can ensure that the Smith normal form has at least two ones. After this process $u_i^E$ and $u_i^F$ still support a loop.
We will now show that we may get $\mathsf{B}_E^\bullet\in\MPplusZ$. First we will arrange that all entries are positive in such diagonal blocks. We already have that $u_i^E$ supports a loop, and hence
Proposition~\ref{prop:matrix-moves}\ref{prop:matrix-moves:I} applies to ensure that any vertex in the component which has an edge to $u_i^E$ also supports a loop. Continuing this way, we get that every vertex supports a loop, and we can use Proposition~\ref{prop:matrix-moves}\ref{prop:matrix-moves:II} to ensure that $u_i$ supports two loops. With this, it is easy to arrange that $\mathsf{B}_E\{i\}>0$. Arguing similarly, we may also arrange that $\mathsf{B}_E\{i,j\}>0$ and $\mathsf{B}_E\{k,i\}>0$
for any $i\prec j$ or $k\prec i$.
We continue this process for all $i\in\ensuremath{\mathcal{P}}\xspace$ satisfying $\mytau(i)=\mytau[F](i)=1$.
Any block $\mathsf{B}_E\{j,k\}$ with $j\prec k$ which is not positive after this process must have $\mytau(j)=0$ and $\mytau(k)\leq 0$ and hence will be a $1\times 1$-matrix. Further, $k$ is not an immediate successor of $j$, so we have $j\prec i\prec k$ for some $i$ an immediate successor of $j$. Then $\mathsf{B}_E\{j,i\}>0$, and we may use Proposition~\ref{prop:matrix-moves}\ref{prop:matrix-moves:I} again to arrange that $\mathsf{B}_E\{j,k\}>0$.
We argue similarly for $F$.
\end{proof}
Note that in general there may be several (but finitely many) ways of choosing $\ensuremath{\mathcal{P}}\xspace$ and the isomorphisms from $\ensuremath{\mathcal{P}}\xspace$ to $\Gamma_E$ and $\Gamma_F$ which give the standard forms.
\begin{remark}
Let $E$ be a finite graph. As is well known, we may efficiently describe a partially ordered set such as $\Gamma_E$ by the Hasse
diagram with vertices $\{1,\dots,N\}$ connecting $\gamma$ to $\gamma'$ when
$\gamma'$ is an immediate successor of $\gamma$. Thinking of $\tau_E\circ{\upsilon}_E$ as providing a coloring of the
vertices of the Hasse diagram thus gives an easy way of visualising
the situation. Noting that the color $-1$ can only occur at the vertices with no successors, we see that the smallest
cases of (isomorphism classes of) colored Hasse diagrams \emph{not} meeting Condition (H) are the cases
\begin{equation}\label{twop}
\twop{1}
\end{equation}
when $|\Gamma_E|=2$ and
\begin{center}
\threepin{2}\quad \threepin{3}\quad \threepout{5}\quad\threeplin{1}\\
\threeplin{2}\quad\threeplinmo{2}\quad \threeplin{5}\quad \threeplin{3}
\end{center}
when $|\Gamma_E|=3$ (along with the three cases obtained by adding an unconnected vertex to the one in \eqref{twop}).
\end{remark}
\section{Classifying move equivalence}\label{CC}
In this section we inspect one of the key results from \cite{MR2270572} to conclude that it holds even for those graphs which are finite with no sinks or
sources, essentially corresponding to the case of Cuntz-Krieger
algebras for matrices not necessarily satisfying the Condition (II)
introduced by Cuntz.
As in \cite{MR2270572}, we will appeal to the theory of flow equivalence of shifts of finite type; since we work with graph \mbox{$C\sp*$-}al\-ge\-bras\xspace instead of Cuntz-Krieger algebras, we use the edge shifts, defined from a finite graph $E$ as
\[
{\mathsf X}_E=\{(e_n)\in (E^1)^\ensuremath{\mathbb{Z}}\xspace\mid \forall n:r(e_n)=s(e_{n+1})\}.
\]
This will suffice for our purposes since we may
remove sources via the notion of canonical form, and may replace sinks
by loops as discussed below.
The formal starting point is the following lemma. We must allow for ${\mathsf X}_E=\emptyset$ in the case that no vertex of $E$ supports a return path, and will say that two such empty shift spaces are mutually flow equivalent, and not flow equivalent to any nonempty shift space.
\begin{lemma}\label{flowvsME}
Let $E$ and $F$ be finite graphs. When $E\ensuremath{\sim_{M\negthinspace E}}\xspace F$, then ${\mathsf X}_E$ is flow equivalent to
${\mathsf X}_F$. If neither $E$ nor $F$ have any sinks, the two conditions are equivalent.
\end{lemma}
\begin{proof}
Move \mbox{\texttt{\textup{(S)}}}\xspace does not affect the shift spaces, and the remaining moves
are precisely the ones allowed in \cite{MR0405385}. Any sink of $E$ or
$F$ will not affect the shift space, so it is not possible to infer in
the opposite direction in general, but if there are no sinks, we may
use Move \mbox{\texttt{\textup{(S)}}}\xspace to remove all sources and to remove the vertices that become sources (this process will terminate since $E$ and $F$ are finite graphs) to replace $E$ and $F$
with $E'\ensuremath{\sim_{M\negthinspace E}}\xspace E$ and $F'\ensuremath{\sim_{M\negthinspace E}}\xspace F$ so that neither $E'$ nor $F'$ have
sources. We have ${\mathsf X}_E={\mathsf X}_{E'}$ and ${\mathsf
X}_F={\mathsf X}_{F'}$, so also ${\mathsf X}_{E'}$ and ${\mathsf
X}_{F'}$ are flow equivalent. By \cite{MR0405385}, this flow
equivalence is induced by a finite number of the moves \mbox{\texttt{\textup{(O)}}}\xspace, \mbox{\texttt{\textup{(I)}}}\xspace and \mbox{\texttt{\textup{(R)}}}\xspace.
\end{proof}
\subsection{Plugging sinks}\label{plugging}
We now introduce a way to pass between the case where the
finite graph $E$ has no sinks and the case where the finite graph has
so many sinks that every cycle in $E$ has an exit. The first case is
preferable in the context of symbolic dynamics, whereas the second
case, as we shall see below in Section \ref{unplugging} is preferable
in the operator algebraic context, since it can be used to establish a
certain uniqueness theorem.
We start with the notion of \emph{plugging} sinks. Whenever a graph $E$ is given, $E_\curlywedge$ denotes the graph where a loop has been added to all sinks.
\begin{lemma}\label{CEpassestoplugged}
Let $E$ and $F$ be graphs with finitely many vertices. If $E\ensuremath{\sim_{C\negthinspace E}}\xspace F$, then also $E_\curlywedge\ensuremath{\sim_{C\negthinspace E}}\xspace F_\curlywedge$.
If $E\ensuremath{\sim_{M\negthinspace E}}\xspace F$, then also $E_\curlywedge\ensuremath{\sim_{M\negthinspace E}}\xspace F_\curlywedge$.
\end{lemma}
\begin{proof}
Considering plugging of sinks as a move, one checks that it commutes
with all of the moves defining $\ensuremath{\sim_{C\negthinspace E}}\xspace$ and $\ensuremath{\sim_{M\negthinspace E}}\xspace$. This is obvious in the case of
\mbox{\texttt{\textup{(I)}}}\xspace\ and \mbox{\texttt{\textup{(C)}}}\xspace\ which can never involve a sink. In the case of \mbox{\texttt{\textup{(O)}}}\xspace, \mbox{\texttt{\textup{(S)}}}\xspace,
and \mbox{\texttt{\textup{(R)}}}\xspace, one sees the claim by noting that sinks are involved only as
receivers of edges in such moves.
\end{proof}
\begin{lemma}\label{Ktheoryplug}
Let $E$ and $F$ be graphs with finitely many vertices and assume that
\fctw{\Theta}{\operatorname{Prime}_\gamma(C^*(E))}{\operatorname{Prime}_\gamma(C^*(F))}
is a homeomorphism. Let $X=\operatorname{Prime}_\gamma(C^*(E))$. Since $\operatorname{Prime}_\gamma(C^*(E))$ and $\operatorname{Prime}_\gamma(C^*(F))$ are canonically homeomorphic to $\operatorname{Prime}_\gamma(C^*(E_\curlywedge))$ and $\operatorname{Prime}_\gamma(C^*(F_\curlywedge))$, respectively, we may view $C^*(E)$, $C^*(F)$, $C^*(E_\curlywedge)$, and $C^*(F_\curlywedge)$ as $X$-algebras in the canonical way. Then the following are equivalent.
\begin{enumerate}[(1)]
\item $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(X;C^*(E))$ and $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(X;C^*(F))$ are isomorphic
\item $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(X;C^*(E_\curlywedge))$ and $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(X;C^*(F_\curlywedge))$ are isomorphic, and $\tau_E=\tau_F\circ\Theta$.
\end{enumerate}
\end{lemma}
\begin{proof}
We note that the changes of $E$ and $F$ only affect the
$K_1$-groups. Since we are only recording the $K_1$-groups at sets
$\{x\}$ and the plugging takes place only at components which have no
successors, in fact no sequences \eqref{eq:longtype} or \eqref{eq:shorttype}
are affected. Thus all that happens is that some of the independent $K_1$-groups that
were originally $0$ are changed to $\ensuremath{\mathbb{Z}}\xspace$, and thus the given isomorphism of the $K$-theories of the original graph \mbox{$C\sp*$-}al\-ge\-bras\xspace readily extend to the plugged versions. In the other direction, the temperature assumption is to ensure that the number of sinks and the number of cyclic components of $E$ and $F$ are equal. Thus, an isomorphism of the $K$-theories of the plugged versions restricts to an isomorphism of the $K$-theories of the original graphs.
\end{proof}
Note finally that when $E$ and $F$ are finite graphs with $(\mathsf{B}_E,\mathsf{B}_F)$ in standard form, so is $(\mathsf{B}_{E_\curlywedge},\mathsf{B}_{F_\curlywedge})$.
In this situation, $\mathsf{B}^\bullet_E$ is $\ensuremath{\operatorname{SL}_\calP}\xspace$- or $\ensuremath{\operatorname{GL}_\calP}\xspace$-equivalent to $\mathsf{B}^\bullet_F$ precisely when the relation holds between $\mathsf{B}_{E_\curlywedge}$ and $\mathsf{B}_{F_\curlywedge}$. Indeed, if $U\mathsf{B}^\bullet_E V=\mathsf{B}^\bullet_F$ with $U\in\GLPZ[\mathbf m]$ and $V\in\GLPZ$, we get $\widetilde{U}\in\GLPZ$ so that $\widetilde{U}\mathsf{B}_{E_\curlywedge} V=\mathsf{B}_{F_\curlywedge}$ by padding $U$ with rows and columns from the appropriately sized identity matrix where a plugging has taken place. Conversely, if $\widetilde U$ is given, $U$ is obtained by deleting the relevant rows and columns.
Since $U\in \SLPZ[\mathbf m]$ precisely when $\widetilde{U}\in \SLPZ$, our claim concerning $\ensuremath{\operatorname{SL}_\calP}\xspace$-equivalence is justified.
Further, when $\mathsf{B}^\bullet_E,\mathsf{B}^\bullet_F\in\MPplusZ$, we conclude that $\mathsf{B}^\bullet_{E_\curlywedge},\mathsf{B}^\bullet_{F_\curlywedge}\in\MPplusZ[\mathbf n]$. We will use these observations repeatedly without mention below.
\subsection{Move equivalence versus $\ensuremath{\operatorname{SL}_\calP}\xspace$- and $\ensuremath{\operatorname{GL}_\calP}\xspace$-equivalence}
The results in this section are the key to everything that follows and all depend on the following proposition, which was proved in \cite[Lemma~6.7 and Theorem~6.8]{MR2270572} under the added assumption that the graphs had Condition (K) and no sinks (\emph{i.e.}\xspace, were Cuntz-Krieger algebras with Condition (II)). But since we are working only at components which are neither single cycles nor sinks, the same proof applies.
\begin{proposition}\label{GunnarRULES}
Let $E$ and $F$ be finite graphs and assume that $(\mathsf{B}_E,\mathsf{B}_F)$ is in standard form with $\mathsf{B}^\bullet_E,\mathsf{B}^\bullet_F\in\MPplusZ$. If
$U\in \GLPZ[\mathbf m]$ and $V\in \GLPZ$ are given with
\[
U\mathsf{B}^\bullet_{E}V=\mathsf{B}^\bullet_{F}
\]
and $i\in\ensuremath{\mathcal{P}}\xspace$ is given with $\mytau(i)=1$, then there exist $\mathbf r$, graphs $E'$ and $F'$ with $(\mathsf{B}_{E'},\mathsf{B}_{F'})$ in standard form with $\mathsf{B}_{E'}^\bullet,\mathsf{B}_{F'}^\bullet\in\MPplusZ[(\mathbf{m}+\mathbf{r})\times(\mathbf{n}+\mathbf{r})]$, $U'\in \GLPZ[\mathbf m+\mathbf{r}]$ and $V'\in \GLPZ[\mathbf{n}+\mathbf{r}]$
so that
\begin{enumerate}[(i)]
\item \label{GunnarRULES-1}
$\mathbf r=(r_j)$ with $r_i\leq 3$ and $r_j=0$ for $j\not=i$,
\item \label{GunnarRULES-2}
$E\ensuremath{\sim_{C\negthinspace E}}\xspace E'$, $F\ensuremath{\sim_{C\negthinspace E}}\xspace F'$,
\item \label{GunnarRULES-3}
$U\{j\}=U'\{j\}$, $V\{j\}=V'\{j\}$ for $j\not=i$,
\item \label{GunnarRULES-4}
$\det U'\{i\}=\det V'\{i\}=1$,
\end{enumerate}
and
\[
U'\mathsf{B}^\bullet_{E'}V'=\mathsf{B}^\bullet_{F'}.
\]
\end{proposition}
\begin{proofsk}
The key idea is to note that whenever $U_0BV_0=B'$, then with
\[
\widetilde{U_0}=
\begin{pmatrix}
U_0&\\
&\left(\begin{smallmatrix}-1\end{smallmatrix}\right)
\end{pmatrix}\qquad \widetilde{V_0}=
\begin{pmatrix}
V_0&\\
&\left(\begin{smallmatrix}-1\end{smallmatrix}\right)
\end{pmatrix}
\]
we have
\[
\widetilde{U_0}\begin{pmatrix}
B&\\
&\left(\begin{smallmatrix}-1\end{smallmatrix}\right)
\end{pmatrix}
\widetilde{V_0}
=
\begin{pmatrix}
B'&\\
&\left(\begin{smallmatrix}-1\end{smallmatrix}\right)
\end{pmatrix}
\]
and $\det\widetilde{U_0}=-\det U_0$ and $\det\widetilde{V_0}=-\det V_0$, and with
\[
\overline{U_0}=
\begin{pmatrix}
U_0&\\
&\left(\begin{smallmatrix}0&1\\1&0
\end{smallmatrix}\right)\end{pmatrix}\qquad \overline{V_0}=
\begin{pmatrix}
V_0&\\
&\left(\begin{smallmatrix}-1&0\\0&-1
\end{smallmatrix}\right)
\end{pmatrix}
\]
we have
\[
\overline{U_0}\begin{pmatrix}
B&\\
&\left(\begin{smallmatrix}0&1\\1&0
\end{smallmatrix}\right)\end{pmatrix}\overline{V_0}
=
\begin{pmatrix}
B'&\\
&\left(\begin{smallmatrix}-1&0\\0&-1
\end{smallmatrix}\right)\end{pmatrix}
\]
and $\det\overline{U_0}=-\det U_0$ and $\det\overline{V_0}=\det V_0$. Thus we can adjust the signs of $U'\{i\}$ and $V'\{i\}$ as required in \ref{GunnarRULES-4} at the cost of adding one of the matrices
\[
\begin{pmatrix}-1
\end{pmatrix},
\begin{pmatrix}-1&0\\0&-1
\end{pmatrix},
\begin{pmatrix}-1&0&0\\0&-1&0\\0&0&-1
\end{pmatrix},
\begin{pmatrix}0&1\\1&0
\end{pmatrix},
\begin{pmatrix}0&1&0\\1&0&0\\0&0&-1
\end{pmatrix}
\]
in the appropriate diagonal of the $\mathsf{B}$-matrix, and the proposition is proved as soon as we have established that whenever, say, $E$ as in the statement is given, we can find $E'\ensuremath{\sim_{C\negthinspace E}}\xspace E$ with 1,2, or 3 vertices more than $E$ in the component $\mytheta(i)$ so that $\mathsf{B}^\bullet_{E'}$ is $\ensuremath{\operatorname{SL}_\calP}\xspace$-equivalent to the relevant augmentation of $\mathsf{B}^\bullet_E$. Note that we require that $\mathsf{B}^\bullet_{E'}$ has positive entries wherever it can be nonzero, and that we must take care not to alter the diagonal blocks of $U$ and $V$ away from component $\mytheta(i)$.
To add a single $-1$, we perform Move \mbox{\texttt{\textup{(R)}}}\xspace in reverse on one of the loops at the last vertex of $\mytheta(i)$ to get $\widetilde{E}$ with $\mathsf{B}_{\widetilde{E}}\{i\}$ in the form
\[
\begin{pmatrix}
\mathsf{B}_{{E}}\{i\}&\left(\begin{smallmatrix}0\\\vdots\\0\\1\end{smallmatrix}\right)\\
\left(\begin{smallmatrix}0&\cdots&0&1\end{smallmatrix}\right)&\left(\begin{smallmatrix}-1\end{smallmatrix}\right)
\end{pmatrix}
\]
This matrix is clearly $\operatorname{SL}$-equivalent to the desired one, and it is straightforward to obtain $E'$ which has only positive entries in $\mathsf{B}_{{E'}}\{i\}$ by a number of row or column additions. We can also arrange to have positive entries in added rows and columns in each offdiagonal block $\{i,j\}$ or $\{k,i\}$ where $i\prec j$ or $k\prec i$. By Proposition \ref{prop:matrix-moves}, $E\ensuremath{\sim_{M\negthinspace E}}\xspace E'$. The $\ensuremath{\operatorname{SL}_\calP}\xspace$-matrices implementing the necessary row or column additions will equal the identity at every diagonal block $\{j\}$ with $i\not=j$, so we will not change these blocks as required in \ref{GunnarRULES-3}.
Repeating this process, we can arrange move equivalences taking us from $E$ to $E'$ with $\mathsf{B}^\bullet_{E'}\in\MPplusZ[(\mathbf{m}+k\mathbf{e}_i)\times(\mathbf{n}+k\mathbf{e}_i)]$ being \ensuremath{\operatorname{SL}_\calP}\xspace-equivalent to $-\iota_{k\mathbf e_i}(-\mathsf{B}_E)$ for any $k\in\ensuremath{\mathbb{N}}\xspace$, where $\mathbf e_i$ is the vector that is $1$ at index $i$ and 0 otherwise. Thus all that remains is to note that if we perform Move \mbox{\texttt{\textup{(C)}}}\xspace on the last vertex of $\mytheta(i)$ to get $\overline{E}$ with $\mathsf{B}_{\overline{E}}\{i\}$ in the form
\[
\begin{pmatrix}
\mathsf{B}_{{E}}\{i\}&\left(\begin{smallmatrix}0&0\\\vdots&\vdots\\0&0\\1&0\end{smallmatrix}\right)\\
\left(\begin{smallmatrix}0&\cdots&0&1\\0&\cdots&0&0\end{smallmatrix}\right)&\left(\begin{smallmatrix}0&1\\1&0\end{smallmatrix}\right)
\end{pmatrix}
\]
we again obtain the desired matrix augmentation up to $\ensuremath{\operatorname{SL}_\calP}\xspace$-equivalence, and may arrange for positive entries just as above.
\end{proofsk}
\begin{proposition}\label{fromBwithplug}
Let $E$ and $F$ be finite graphs and assume that $(\mathsf{B}_E,\mathsf{B}_F)$ is in standard form with $\mathsf{B}^\bullet_E,\mathsf{B}^\bullet_F\in\MPplusZ$.
Assume further that $U\in \GLPZ[\mathbf m]$ and $V\in\GLPZ$ are given with
\[
U\mathsf{B}_{E}^\bullet V=\mathsf{B}_{F}^\bullet.
\]
Then
\begin{enumerate}[(i)]
\item \label{fromBwithplug-1}
If $U\in \SLPZ[\mathbf m]$ and $V\in\SLPZ$, then $E\ensuremath{\sim_{M\negthinspace E}}\xspace F$.
\item \label{fromBwithplug-2}
If $V\{i\}=1$ whenever $\mytau(i)\leq 0$ and $U\{i\}=1$ whenever $\mytau(i)=0$, then $E\ensuremath{\sim_{C\negthinspace E}}\xspace F$.
\end{enumerate}
\end{proposition}
\begin{proof}
To prove \ref{fromBwithplug-1}, we pass to the plugged graphs and recall that $\mathsf{B}_{F_\curlywedge}$ and $\mathsf{B}_{E_\curlywedge}$ are also $\ensuremath{\operatorname{SL}_\calP}\xspace$-equivalent. Since $E_\curlywedge$ and $F_\curlywedge$ have neither sinks nor
sources, we may appeal to \cite[Theorem 4.4]{MR1907894} which shows that
$\mathsf{B}_{F_\curlywedge}$ can be obtained from $\mathsf{B}_{E_\curlywedge}$
by a number of elementary row or column additions or subtractions,
never leaving matrices in $\MPplusZ$. In fact, Boyle produces a list of elementary equivalences $E_{u,v}$ as described in Proposition \ref{prop:matrix-moves} where there is a path from $u$ to $v$ throughout, and since we have arranged that any vertex in any graph along the way supports at least one loop, the proposition applies to yield \ref{fromBwithplug-1} when $E$ and $F$ have no
sinks.
When $E$ and $F$ do have sinks, we apply the same sequence of row and column operations to $\mathsf{B}_E$. We
note that in any matrix addition or subtraction implemented by $E_{u,v}$, $u$ will not be one of the plugged sinks, as indeed these provide paths only to themselves. Hence $u$ will not be a sink in the original setup. In the case the matrix implements a column operation, the requirements in
Proposition \ref{prop:matrix-moves} are still met, and thus such an operation remains implemented by moves in the original setup. In the case the matrix implements a row operation, we observe that it has no effect, adding or subtracting a zero row from another row. It may hence be omitted, proving \ref{fromBwithplug-1}.
We prove \ref{fromBwithplug-2} by reducing to \ref{fromBwithplug-1} by Proposition~\ref{GunnarRULES}, changing any negative determinants of the given $U$ and $V$ at blocks $\{i\}$ starting from $\{1\}$ and working downwards. We then get finite graphs $E'$ and $F'$ such that $(\mathsf{B}_{E'},\mathsf{B}_{F'})$ is in standard form with
$\mathsf{B}_{E'}^\bullet,\mathsf{B}_{F'}^\bullet\in \MPplusZ[(\mathbf m+\mathbf r)\times (\mathbf n+\mathbf r)]$ where the multiindex $\mathbf{r}$ has the property that $r_j=0$ for $j$ with
$\mytau(j)\leq 0$ and $r_j\leq 3$ otherwise. We have that
$E\ensuremath{\sim_{C\negthinspace E}}\xspace E'$, $F\ensuremath{\sim_{C\negthinspace E}}\xspace F'$ and that for some $U'\in\SLPZ[\mathbf m+\mathbf r]$ and $V'\in\SLPZ[\mathbf n+\mathbf r]$, we may arrange that
$U'\mathsf{B}_{E'}^\bullet V'=\mathsf{B}_{F'}$. By \ref{fromBwithplug-1}, $E'\ensuremath{\sim_{M\negthinspace E}}\xspace F'$.
\end{proof}
\begin{definition}
Let $E$ and $E'$ be graphs with finitely many vertices and assume that $\mathsf{B}_E\in\MPZc$ and $\mathsf{B}_{E'}\in\MPZc[\mathbf{m}'\times\mathbf{n}']$.
We say that a \mbox{$\sp*$-}iso\-morphism\xspace from $C^*(E)$ to $C^*(E')$ (or from $C^*(E)\otimes\ensuremath{\mathbb{K}}\xspace$ to $C^*(E')\otimes\ensuremath{\mathbb{K}}\xspace$) respects the block structure, if the induced homeomorphism from $\operatorname{Prime}_\gamma(C^*(E))$ to $\operatorname{Prime}_\gamma(C^*(E'))$ commutes with the identification of $\ensuremath{\mathcal{P}}\xspace$ with $\operatorname{Prime}_\gamma(C^*(E))$ and $\operatorname{Prime}_\gamma(C^*(E'))$, respectively.
All of the elementary moves introduced in Section~\ref{sec:moves} induces a canonical stable isomorphism. We say that such an elementary move preserves the block structure if this induced stable isomorphism respects the block structure. We say that a move equivalence or a Cuntz move equivalence respects the block structure, if it is the composition of a series of elementary moves such that the composition of the induced stable isomorphisms respects the block structure.
If the only automorphism of \ensuremath{\mathcal{P}}\xspace is the trivial automorphism, then \mbox{$\sp*$-}iso\-morphisms\xspace, move equivalences and Cuntz move equivalences, respectively, automatically respect the block structure --- we will in particular use that this is the case when \ensuremath{\mathcal{P}}\xspace is linearly ordered.
\end{definition}
\begin{proposition}\label{toBwithplug}
Let $E$ and $F$ be finite graphs and assume that $(\mathsf{B}_E,\mathsf{B}_F)$ is in standard form and has the additional property that $\gcd(\mathsf{B}_E\{i\})=1$ and $\gcd(\mathsf{B}_F\{i\})=1$ at every $i$ with $\mytau(i)=\mytau[F](i)=1$. When $E_\curlywedge\ensuremath{\sim_{C\negthinspace E}}\xspace F_\curlywedge$ respecting the block structure,
there exist $U,V\in \GLPZ[\mathbf n]$ with $U\{i\}=V\{i\}=1$ whenever $\mytau(i)\leq 0$ so that $
U\mathsf{B}_{E_\curlywedge}V=\mathsf{B}_{F_\curlywedge}
$. When $E_\curlywedge\ensuremath{\sim_{M\negthinspace E}}\xspace F_\curlywedge$,
we may choose $U,V\in \SLPZ[\mathbf n]$.
\end{proposition}
\begin{proof}
We may assume without loss of generality that $E \ensuremath{\sim_{C\negthinspace E}}\xspace F$ respecting the block structure and $E$ and $F$ have no sinks. Since $E \ensuremath{\sim_{C\negthinspace E}}\xspace F$, we have a string of moves as follows:
\[
\xymatrix{
E \ar[r]^-{\ensuremath{\sim_{M\negthinspace E}}\xspace} & E_1 \ar[r]^-{\mbox{\texttt{\textup{(C)}}}\xspace} & E_2 \ar[l]\ar[r]^-{\ensuremath{\sim_{M\negthinspace E}}\xspace} & E_3 \ar[r]^-{\mbox{\texttt{\textup{(C)}}}\xspace} &E_4\ar[l]\ar[r]& \cdots \ar[r]^-{\mbox{\texttt{\textup{(C)}}}\xspace} & E_{2n}\ar[l] \ar[r]^-{\ensuremath{\sim_{M\negthinspace E}}\xspace} & F
}
\]
where each move between $E_{2j-1}$ and $E_{2j}$ is either a Cuntz splice or its inverse.
Note that at each stage of the move equivalence, we may have
introduced transitional vertices and we may have increased the number
of vertices in the cyclic components. So, we collapse these
transitional vertices and the cyclic components, to obtain a graph
$F_i$ with no transitional vertices such that $E_i \ensuremath{\sim_{M\negthinspace E}}\xspace F_i$ and $F_i
\in \MPZccc[ \mathbf{n}_i ]$. Note that performing these moves
commutes with any Cuntz splice, since such a move cannot take place at
a cyclic component or at a transitional vertex. Hence, we have a
commuting diagram
\[
\xymatrix{
E \ar[r]^-{\ensuremath{\sim_{M\negthinspace E}}\xspace} & E_1 \ar[r]^-{\mbox{\texttt{\textup{(C)}}}\xspace} \ar[d]^\ensuremath{\sim_{M\negthinspace E}}\xspace & E_2\ar[l] \ar[r]^-{\ensuremath{\sim_{M\negthinspace E}}\xspace} \ar[d]^{\ensuremath{\sim_{M\negthinspace E}}\xspace} & E_3 \ar[r]^-{\mbox{\texttt{\textup{(C)}}}\xspace} \ar[d]^{\ensuremath{\sim_{M\negthinspace E}}\xspace} &E_4\ar[l]\ar[r]\ar[d]^{\ensuremath{\sim_{M\negthinspace E}}\xspace}& \cdots \ar[r]^-{\mbox{\texttt{\textup{(C)}}}\xspace} & E_{2n}\ar[l] \ar[d]^{\ensuremath{\sim_{M\negthinspace E}}\xspace} \ar[r]^-{\ensuremath{\sim_{M\negthinspace E}}\xspace} & F \\
& F_1 \ar[r]^-{\mbox{\texttt{\textup{(C)}}}\xspace} & F_2\ar[l] & F_3 \ar[r]^-{\mbox{\texttt{\textup{(C)}}}\xspace} &F_4\ar[l]& \ar[r]^-{\mbox{\texttt{\textup{(C)}}}\xspace} & F_{2n}\ar[l] & },
\]
where the compositions of the move equivalences all respect the block structure.
Let $F_0=E$ and $F_{2n+1}=F$.
Let $k\in\{0,1,\ldots,n\}$ be given.
Since $F_{2k} \ensuremath{\sim_{M\negthinspace E}}\xspace F_{2k+1}$, we have that the shift spaces $\mathsf{X}_{F_{2k}}$ and $\mathsf{X}_{F_{2k+1}}$ are flow equivalent. By \cite[Theorem~3.1 and Theorem~3.4]{MR1907894}, there exists an \SLP-equivalence\xspace \fctw{(U_{2k}, V_{2k})}{-\iota_{\mathbf{r}_{2k}}(-\mathsf{B}_{F_{2k}})}{-\iota_{\mathbf{r}_{2k}'}(-\mathsf{B}_{F_{2k+1}})}, where $\mathbf{r}_{2k}=(r_{2k,l})_{l\in\ensuremath{\mathcal{P}}\xspace}$ and $\mathbf{r}_{2k}'=(r_{2k,l}')_{l\in\ensuremath{\mathcal{P}}\xspace}$ with $r_{2k,l}=r_{2k,l}'=0$ whenever $\mytau(l) \leq 0$.
Let again $k\in\{0,1,\ldots,n\}$ be given.
A computation based on Restorff's proof of Proposition~\ref{GunnarRULES} shows that there exists a \GLP-equivalence\xspace \fctw{(U_{2k+1} , V_{2k+1} )}{-\iota_{\mathbf{r}_{2k+1}}( -\mathsf{B}_{F_{2k+1} })}{-\iota_{\mathbf{r}_{2k+1}'}(-\mathsf{B}_{F_{2k}} )} such that $U_{2k+1}\{i\} = V_{2k+1}\{i \} = 1$ for all $\mytau[F_{2j+1}](i) \leq 0$, where $\mathbf{r}_{2k+1}=(r_{2k+1,l})_{l\in\ensuremath{\mathcal{P}}\xspace}$ and $\mathbf{r}_{2k+1}'=(r_{2k+1,l}')_{l\in\ensuremath{\mathcal{P}}\xspace}$ with $r_{2k+1,l}=r_{2k+1,l}'=0$ whenever $\mytau(l) \leq 0$.
By \cite[Theorem~3.10]{MR1990568},
the composition of these \ensuremath{\operatorname{SL}_\calP}\xspace- and \ensuremath{\operatorname{GL}_\calP}\xspace-equivalences induces a
$K$-web isomorphism $\kappa$. Hence, by
\cite[Theorem~4.5]{MR1990568}, there exists a \GLP-equivalence\xspace $(U,V) \colon
\mathsf{B}_{E} \rightarrow \mathsf{B}_{F}$ inducing $\kappa$ as in Lemma \ref{lem: Kweb 2 components}. Since the cyclic
components of $E$ and $F$ are $1\times 1$ blocks and $(U,V)$ induces
$\kappa$, we have that $U\{i\}=V\{i\}=1$ whenever $\mathcal{T}_{\mathsf{B}_E}(i)\leq 0$.
We now prove the statement about move equivalence. As above, we may assume that $E$ and $F$ have no sinks and we get an \SLP-equivalence\xspace \fctw{(U,V)}{-\iota_{\mathbf{r}}(-\mathsf{B}_{E})}{-\iota_{\mathbf{r}}(-\mathsf{B}_{F})}, where $\mathbf{r}=(r_l)_{l\in\ensuremath{\mathcal{P}}\xspace}$ with $r_l=0$ whenever $\mathcal{T}_{\mathsf{B}_E}(l)\leq 0$. Now it follows from \cite[Proposition~4.1 and Corollary~4.9]{MR1990568} that there exists an \SLP-equivalence\xspace from $\mathsf{B}_E$ to $\mathsf{B}_F$.
\end{proof}
\begin{theorem}\label{iffcharofCeq}
Let $E$ and $F$ be finite graphs with $(\mathsf{B}_E,\mathsf{B}_F)$ in standard form with $\mathsf{B}^\bullet_E,\mathsf{B}^\bullet_F\in\MPplusZ$. Then the following are equivalent
\begin{enumerate}[(1)]
\item\label{iffcharofCeqI} $E\ensuremath{\sim_{C\negthinspace E}}\xspace F$ respecting the block structure,
\item\label{iffcharofCeqNEW} $E_\curlywedge\ensuremath{\sim_{C\negthinspace E}}\xspace F_\curlywedge$ respecting the block structure,
\item\label{iffcharofCeqII} There exist $U,V\in \GLPZ$ with $U\{i\}=V\{i\}=1$ whenever $\mytau(i)\leq 0$ so that
$U\mathsf{B}_{E_\curlywedge}V=\mathsf{B}_{F_\curlywedge}$,
\item\label{iffcharofCeqIII} There exist $U\in \GLPZ[\mathbf m]$ and $V\in \GLPZ$ with $V\{i\}=1$ whenever $\mytau(i)\leq 0$ and with $U\{i\}=1$ whenever $\mytau(i)=0$ so that
$
U\mathsf{B}_{E}^\bullet V=\mathsf{B}_{F}^\bullet
$.
\end{enumerate}
\end{theorem}
\begin{proof}
Lemma \ref{CEpassestoplugged} proves that \ref{iffcharofCeqI}$\Longrightarrow$\ref{iffcharofCeqNEW}.
Since the $\gcd$ is $1$ at any block with a $1$ in the Smith form, we may apply Proposition \ref{toBwithplug} to prove \ref{iffcharofCeqNEW}$\Longrightarrow$\ref{iffcharofCeqII}. We have noted that \ref{iffcharofCeqII}$\Longleftrightarrow$\ref{iffcharofCeqIII} holds in general, and \ref{iffcharofCeqIII}$\Longrightarrow$\ref{iffcharofCeqI} is the content of Proposition \ref{fromBwithplug}\ref{fromBwithplug-2}.
\end{proof}
\begin{theorem}\label{iffcharofMeq}
Let $E$ and $F$ be finite graphs with $(\mathsf{B}_E,\mathsf{B}_F)$ in standard form with $\mathsf{B}^\bullet_E,\mathsf{B}^\bullet_F\in\MPplusZ$. Then the following are equivalent
\begin{enumerate}[(1)]
\item\label{iffcharofMeqI} $E\ensuremath{\sim_{M\negthinspace E}}\xspace F$ respecting the block structure,
\item\label{iffcharofMeqNEW} $E_\curlywedge\ensuremath{\sim_{M\negthinspace E}}\xspace F_\curlywedge$ respecting the block structure,
\item\label{iffcharofMeqII} There exist $U,V\in \SLPZ$ so that $U\mathsf{B}_{E_\curlywedge}V=\mathsf{B}_{F_\curlywedge}$,
\item\label{iffcharofMeq|||} There exist $U\in \SLPZ[\mathbf m]$ and $V\in \SLPZ$ so that
$U\mathsf{B}_{E}^\bullet V=\mathsf{B}_{F}$.
\end{enumerate}
\end{theorem}
\begin{proof}
The proof is completely analogous to the proof of Theorem~\ref{iffcharofCeq}, where we use Proposition \ref{fromBwithplug}\ref{fromBwithplug-1} in the place of Proposition \ref{fromBwithplug}\ref{fromBwithplug-2}.
\end{proof}
We warn the reader that the implication \ref{iffcharofCeqNEW}$\Longrightarrow$ \ref{iffcharofCeqI} in both results above are only true when the temperatures of $E$ and $F$ match up, as implicitly arranged by the condition of standard form.
\begin{example}\label{notalwayssamepre}
The pair of graphs $E$ and $F$ given in Figure \ref{firstexx}(b) are not Cuntz move equivalent.
\end{example}
\begin{proof}
We see that the vertices $E$ and $F$ may be ordered with $(\mathsf{B}_E,\mathsf{B}_F)$ in standard form with
$\mathsf{B}_E,\mathsf{B}_F\in\MPZccc[\mathbf{1}]$ for $\ensuremath{\mathcal{P}}\xspace=\{1,2,3\}$ ordered
linearly and with $\gcd$ of the blocks at $\{2\}$ equal to $1$. Appealing to Proposition \ref{toBwithplug}, we see that it
suffices to check, which is obviously true, that there is no solution to
\begin{equation}\label{nosol}
\begin{pmatrix}1&x&y\\0&s&z\\0&0&1
\end{pmatrix}
\begin{pmatrix}0&1&2\\0&1&1\\0&0&0
\end{pmatrix}
\begin{pmatrix}1&x'&y'\\0&s'&z'\\0&0&1
\end{pmatrix}
=
\begin{pmatrix}0&1&0\\0&1&1\\0&0&0
\end{pmatrix}
\end{equation}
with $s,s'\in\{-1,1\}$ and $x,x',y,y',z,z' \in \ensuremath{\mathbb{Z}}\xspace$.
\end{proof}
\section{Classifying \mbox{$C\sp*$-}al\-ge\-bras\xspace}\label{Ccas}
\subsection{A classification result}
\begin{theorem}\label{mainthm}
Let $E$ and $F$ be finite graphs and consider the statements
\begin{enumerate}[(1)]
\item $E\ensuremath{\sim_{C\negthinspace E}}\xspace F$,\label{MOV}
\item $C^*(E)\otimes\ensuremath{\mathbb{K}}\xspace\cong C^*(F)\otimes\ensuremath{\mathbb{K}}\xspace$,\label{MOR}
\item There exists a homeomorphism \fctw{\Theta}{X=\operatorname{Prime}_\gamma(C^*(E))}{\operatorname{Prime}_\gamma(C^*(F))} so that when $C^*(E)$ and $C^*(F)$ are considered as $X$-algebras in the canonical way, then $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(X;C^*(E))\cong \ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(X;C^*(F))$.\label{FK}
\end{enumerate}
Then
\[
\text{\ref{MOV}}\Longrightarrow \text{\ref{MOR}}\Longrightarrow \text{\ref{FK}}
\]
and when $E$ and $F$ satisfy
Condition~(H), all
statements \ref{MOV}--\ref{FK} are equivalent.
\end{theorem}
\begin{proof}
The invariance of moves required to prove
\ref{MOV}$\Longrightarrow$\ref{MOR} was established in \cite{MR3082546} and
\cite{arXiv:1602.03709v2}, \emph{cf.}\xspace\ Theorems~\ref{thm:moveimpliesstableisomorphism} and~\ref{thm:cuntz-splice-implies-stable-isomorphism}.
For
\ref{MOR}$\Longrightarrow$\ref{FK} one needs only note, as we did in Lemma~\ref{lem:structure-1}, that any
isomorphism between $C^*(E)\otimes\ensuremath{\mathbb{K}}\xspace$ and $C^*(F)\otimes\ensuremath{\mathbb{K}}\xspace$ must
preserve the gauge invariant ideals even if the isomorphism is not
gauge invariant.
To prove that \ref{FK}$\Longrightarrow$\ref{MOV} under the additional
assumption of Condition (H), we first note that by Lemma
\ref{taufromK}, the tempered gauge prime ideals agree, and hence by
Lemma~\ref{taugivesstd} we may assume that $(\mathsf{B}_E,\mathsf{B}_F)$ is in standard
form, where we may even assume that
$\mathsf{B}_E,\mathsf{B}_F\in\MPplusZ$. Plugging sinks we get
$(\mathsf{B}_{E_\curlywedge},\mathsf{B}_{F_\curlywedge})$ which is also in standard form, having
isomorphic ordered reduced filtered $K$-theories by Lemma \ref{Ktheoryplug}. The
$K$-webs then also agree, and \cite{MR1990568} applies to provide $U,V\in\GLPZ$ with
$U\mathsf{B}_{E_\curlywedge}V=\mathsf{B}_{F_\curlywedge}$. Thus we only need to
arrange that $U$ and $V$ satisfy the conditions in Theorem~\ref{iffcharofCeq}\ref{iffcharofCeqII} to reach the desired conclusion.
In fact, since $V\{i\}$ implements an order isomorphism from
$(\ensuremath{\mathbb{Z}}\xspace,\ensuremath{\mathbb{N}}\xspace_0)$ to $(\ensuremath{\mathbb{Z}}\xspace,\ensuremath{\mathbb{N}}\xspace_0)$ at every $i$ with $\mytau(i)\leq 0$, it must
already be in the desired form. It is straightforward to check
that whenever $U\{i\}=-1$ at some $i$ with no successors, then since
both $\mathsf{B}_{E_\curlywedge}$ and $\mathsf{B}_{F_\curlywedge}$ have zero rows at $i$, the corresponding row of $U$ can be multiplied by $-1$ without affecting the relation that
$U\mathsf{B}_{E_\curlywedge}V=\mathsf{B}_{F_\curlywedge}$.
We claim that in the presence of Condition (H), the remaining blocks
$U\{i\}$ at $i$ with $\mytau(i)\leq 0$ must be of the desired
form. Indeed, choosing an immediate successor $j$ of $i$ with
$\mytau(j)\leq 0$ we assume for contradiction that $U\{i\}=-1$. Note
that $\mathsf{B}_{E_\curlywedge}\{i,j\}=x$ and
$\mathsf{B}_{F_\curlywedge}\{i,j\}=y$ with $x,y>0$ since there must be a
path between the two components, and such a path cannot pass through
any other component. Similarly, we get from the immediate successor
condition that for any $B,B'\in\MPZccc$ and any $k\not\in\{i,j\}$,
either $B\{i,k\}=0$ or $B'\{k,j\}=0$, so that
$(BB')\{i,j\}=B\{i\}B'\{i,j\}+B\{i,j\}B'\{j\}$ (\emph{cf.}\xspace\ \eqref{eq:howtomul}). From this we infer that
$(U\mathsf{B}_{E_\curlywedge})\{i,j\}=-x$ and
$(\mathsf{B}_{F_\curlywedge}V^{-1})\{i,j\}=y$, a contradiction.
\end{proof}
\begin{corollary}\label{maincor}
Let $E$ and $F$ be finite graphs so that $C^*(E)$ and $C^*(F)$ are either of real rank zero or type I/postliminal. Then the
statements \ref{MOV}--\ref{FK} of Theorem~\ref{mainthm} are equivalent.
\end{corollary}
\begin{remark}\label{onlyorderongs}
Inspection of our proof shows that only the order on $K_0(C^*(E)(\{i\}))$ and $K_0(C^*(F)(\{i\}))$ is necessary to conclude that the \mbox{$C\sp*$-}al\-ge\-bras\xspace are stably isomorphic.
It is possible to define a full (ordered) filtered K-theory (see \cite{MR3177344,MR3349327}). Isomorphism of this invariant clearly implies isomorphism of the reduced invariant (both in the case with and without order). As a consequence of the results in \cite{MR1990568}, the opposite holds without order for the cases considered in this paper. From the results in Theorem \ref{mainthm}, it follows that it holds also in the case with order. Thus the full invariant contains the same information about equivalence classes and (stable) isomorphism classes as the reduced one.
\end{remark}
\subsection{Unplugging sinks}\label{unplugging}
\renewcommand{\efvs}[1]{{#1}^0_{\mathrm{iso}}\backslash {#1}^0_{\mathrm{sing}}}
\newcommand{\efiso}[1]{{#1}^0_{\mathrm{iso}}}
For a graph $E$, let $E^0_\mathrm{iso}$ be the set of vertices of $E$ that are either sinks or on a
vertex-simple cycle with no exits (the notation, \emph{cf.}\xspace\ \cite{arXiv:1410.2308v1}, refers to the fact that such vertices give rise to isolated points in the associated path spaces).
Assume that $E$ is a graph with finitely many vertices with $\mathsf{B}_E\in\MPZccc$. Then
every vertex $v \in \efvs{E}$
supports a unique loop $e_v$. Let $E_\curlyvee$ be the graph obtained from $E$ by removing the edges
$e_v$ for all $v \in \efvs{E}$. We note that in general
\[
(E_\curlywedge)_\curlyvee\not=
(E_\curlyvee)_\curlywedge\not=E.
\]
\begin{proposition}\label{prop: unplugging}
Let $E$ and $F$ be graphs with finitely many vertices so that $\mathsf{B}_E\in\MPZccc[\mathbf{m}_E\times\mathbf{m}_E]$ and $\mathsf{B}_F\in\MPZccc[\mathbf{m}_F\times\mathbf{m}_F]$.
If there exists a \mbox{$\sp*$-}iso\-morphism\xspace $\Phi \colon C^* (E_\curlyvee
) \otimes \ensuremath{\mathbb{K}}\xspace \rightarrow C^* (F_\curlyvee ) \otimes \ensuremath{\mathbb{K}}\xspace$ such that
$\mytau[F]\circ\Phi_\sharp=\mytau$, then $C^* ( E ) \otimes \ensuremath{\mathbb{K}}\xspace \cong C^*
( F ) \otimes \ensuremath{\mathbb{K}}\xspace$.
\end{proposition}
\begin{proof}
Note first that whenever $v\in (E_\curlyvee)^0_\mathrm{iso}$ is given, $v$ is a sink, so $\{v\}$ is a saturated and hereditary set. Thus it defines an ideal $\J_{v}$ which is minimal in $C^*(E_\curlyvee)$ and Morita equivalent to $\ensuremath{\mathbb{C}}\xspace$. In fact, any such ideal has this form, and since the same is true for $F_\curlyvee$, we conclude that $\Phi(\J_{v})=\J_w$ for some $w\in (F_\curlyvee)^0_\mathrm{iso}$. Since $\mytau[F]\circ\Phi_\sharp=\mytau$, $w$ will be a sink of $F$ precisely when $v$ is a sink of $E$, and thus a bijection \fct{w}{E^0_\mathrm{iso}}{F^0_\mathrm{iso}} is defined with $w(\efvs{E})=\efvs{F}$.
For any graph $G$, let $SG$ be the stabilized graph, \emph{i.e.}\xspace, for each
vertex $v \in G^0$, we put an infinite head at $v$, \emph{cf.}\xspace\ \cite[Definition~9.4]{MR2775826}. Note that
$\efiso{(SG)} = \efiso{G}$ with $\efvs{(SG)} = \efvs{G}$, so that $w$ may also be considered as a map from $(SE)^0_\mathrm{iso}$ to $(SF)^0_\mathrm{iso}$. Moreover, $v
\in \efiso{G}$ supports a loop if and only if $v \in \efiso{(SG)}$
supports a loop. By the proof of \cite[Proposition~9.3 and
Theorem~9.8]{MR2775826}, there exists a \mbox{$\sp*$-}iso\-morphism\xspace $\chi_G
\colon C^* (SG) \rightarrow C^* (G) \otimes \ensuremath{\mathbb{K}}\xspace$ such that $\chi_G (
p_v ) = p_v \otimes e_{11}$ for all $v \in G^0$. Define \fct{\Psi}{C^* ( SE_\curlyvee )}{ C^* (
SF_\curlyvee )} by $\Psi = \chi_{F_\curlyvee}^{-1} \circ \Phi \circ
\chi_{E_\curlyvee}$.
Note that $ \J_{ v } \cong \ensuremath{\mathbb{K}}\xspace$ in $C^*(SE_\curlyvee)$ and $ \J_{w } \cong \ensuremath{\mathbb{K}}\xspace$ in $C^*(SF_\curlyvee)$ for all $v \in E_{\mathrm{iso}}^{0}$ and for all
$w \in F_{ \mathrm{iso} }^{0}$. Therefore, any generator of $K_{0} (
\J_{v} )_{+}$ is Murray-von Neumann equivalent to $p_{v}$ in $C^*(SE_\curlyvee)$ for all
$v \in E_{\mathrm{iso}}^{0}$ and any generator of $K_{0} ( \J_{w}
)_{+}$ is Murray-von Neumann equivalent to $p_{w}$ in $C^*(SF_\curlyvee)$ for all $w \in F_{
\mathrm{iso} }^{0}$. Consequently,
$\Psi ( p_v ) \sim p_{w(v)}$ in $C^*( SF_\curlyvee)$, so there exists $W_v \in C^*( SF_\curlyvee)$ such that $W_v^* W_v =\Psi( p_v)$ and $W_v W_v^* = p_{w(v)}$. Set $p = \sum_{ v \in (SE)^0_{\mathrm{iso}} } \Psi ( p_v )$ and $q = \sum_{ v \in (SE)^0_{\mathrm{iso }}} p_{w(v)}$. Since $C^*(SF_\curlyvee)$ is a stable $C^*$-algebra, by \cite[Corollary 1.10]{MR869419},
\[
1_{ M( C^*(SF_\curlyvee ) ) } - p \sim 1_{ M( C^*(SF_\curlyvee ) ) } \sim 1_{ M( C^*(SF_\curlyvee ) ) } - q.
\]
Thus, there exists $W \in M( C^*(SF_\curlyvee ))$ such that $W^* W = 1_{ M( C^*(SF_\curlyvee ) } - p$ and $WW^* = 1_{ M( C^*(SF_\curlyvee ) ) } - q$. Set $u = W + \sum_{ v \in (SE)^0_{\mathrm{iso}} } W_v$. A computation shows that $u$ is a unitary in $M( C^*( SF_\curlyvee))$ such that $u \Psi( p_v ) u^* = p_{w(v)}$ for all $v \in SE^0_{\mathrm{iso}}$.
So, without loss of generality, we may
assume that $\Psi ( p_v ) = p_{w(v) }$.
Note that $SE_\curlyvee$ and $SF_\curlyvee$ satisfy Condition~(L) since we have removed all cycles with no exits. Using the universal property and the Cuntz-Krieger Uniqueness Theorem, there are injective \mbox{$\sp*$-}ho\-mo\-morphisms\xspace $\lambda_E \colon C^* ( SE_\curlyvee ) \rightarrow C^* (SE)$ and $\lambda_F \colon C^* ( SF_\curlyvee ) \rightarrow C^* (SF)$ such that $\lambda_E ( s_e ) = s_e$, $\lambda_E ( p_v ) = p_v$ for all $e \in (SE_\curlyvee)^1 \subseteq (SE)^1$ and for all $v \in (SE_\curlyvee)^0 = (SE)^0$ and $\lambda_F ( s_f ) = s_f$, and $\lambda_F ( p_w ) = p_w$ for all $f \in (SF_\curlyvee)^1 \subseteq (SF)^1$ and for all $w \in (SF_\curlyvee)^0 = (SF)^0$. So, using these embeddings, we may assume that $C^* (SE_\curlyvee)$ is a sub-algebra of $C^* (SE)$ and $C^* (SF_\curlyvee)$ is a sub-algebra of $C^* (SF)$.
We now define a Cuntz-Krieger $SE$-family in $C^* ( SF)$. Set $P_v = \Psi ( p_v )$ for all $v \in (SE)^0 = (SE_\curlyvee)^0$ and
\[
S_e =
\begin{cases}
\Psi ( s_e ) &\text{if $e \in (SE_\curlyvee)^1$} \\
s_{ e_{w(v)} } &\text{if $e = e_v$ for some $v \in \efvs{(SE)}$}.
\end{cases}
\]
The only nonobvious Cuntz-Krieger relation is at $v \in \efvs{(SE)}$. But this is also clear since $P_v = \Psi ( p_v ) = p_{w(v)} = s_{ e_{w(v) } }s_{ e_{w(v) } }^*= S_{e_v} S_{e_v}^*$. Therefore, there exists a \mbox{$\sp*$-}ho\-mo\-morphism\xspace $\fct{\Xi}{C^* (SE)}{ C^* ( SF )}$. Since the only vertex-simple cycles in $SE$ with no exits are $e_v$ for all $v \in \efvs{SE}$ and $\Xi ( s_ { e_v } ) = s_ { w(v) }$ has full spectrum, by the General Cuntz-Krieger Uniqueness Theorem in \cite{MR1914564}, we have that $\Xi$ is injective. Note that $\Xi ( C^* ( SE_\curlyvee) ) = \Psi ( C^* (SE_\curlyvee) ) = C^* ( SF_\curlyvee )$. Let $e \in (SF)^1$ such that $e$ is not an element of $(SF_\curlyvee)^1$. Then $e= e_{w}$ for some $w \in \efvs{(SF_\curlyvee)}$. Therefore, there exists $v \in \efvs{(SE_\curlyvee)}$ such that $w(v) = w$. Hence, $\Xi ( s_{ e_v } ) = s_{ e_{w(v) } } = s_e$, so $\Xi$ is surjective, and thus a \mbox{$\sp*$-}iso\-morphism\xspace.
\end{proof}
\subsection{Examples}
In this section we let $E$ and $F$ denote the two graphs given in Figure \ref{firstexx}(b). We note that
Example \ref{linearcase} applies (with $n=3$) to this case. In particular, $\operatorname{Prime}_\gamma(C^*(E))\cong X_3\cong\operatorname{Prime}_\gamma(C^*(F))$.
\begin{example}\label{notalwayssame}
The pair of graphs $E$ and $F$ satisfy condition \ref{FK} of Theorem \ref{mainthm}, but not condition \ref{MOV}. The same is true for the pair of graphs $E_\curlyvee$ and $F_\curlyvee$.
\end{example}
\begin{proof}
We have seen in Example \ref{notalwayssamepre} that $E\not\ensuremath{\sim_{C\negthinspace E}}\xspace F$, and since $E=(E_\curlyvee)_\curlywedge$ and $F=(F_\curlyvee)_\curlywedge$ we conclude that $E_\curlyvee\not\ensuremath{\sim_{C\negthinspace E}}\xspace F_\curlyvee$ by transposition of Lemma \ref{CEpassestoplugged}.
To see that the $K$-theories are isomorphic, we note that
\[
U
\begin{pmatrix}0&1&2\\0&1&1\\0&0&0
\end{pmatrix}
V
=
\begin{pmatrix}0&1&0\\0&1&1\\0&0&0
\end{pmatrix}
\]
with $V=I$ and
\[
U=\begin{pmatrix}-1&2&0\\0&1&0\\0&0&1\end{pmatrix}
\]
This \ensuremath{\operatorname{GL}_\calP}\xspace-equivalence induces an isomorphism $\ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(X_3;C^*(E))\cong \ensuremath{\operatorname{FK}_\mathcal{R}}\xspace(X_3;C^*(F))$ as noted in Section \ref{sec:red-filtered-K-theory-K-web-GLP-and-SLP-equivalences}, and since $V\{i\}=1$ at all blocks, the maps induced by $V^{\mathsf T}$ on the $K_0$-groups are order isomorphisms.
The isomorphism of $K$-theory for $E=(E_\curlyvee)_\curlywedge$ and $F=(F_\curlyvee)_\curlywedge$ follows from Lemma \ref{Ktheoryplug}.
\end{proof}
In fact, in this particular case, reversal of the chain of implications in Theorem~\ref{mainthm} breaks down at \ref{MOR}$\implies$\ref{MOV}. To prove this, we provide an \emph{ad hoc} classification of a small class of \mbox{$C\sp*$-}al\-ge\-bras\xspace of relevance.
Let $\ensuremath{\mathfrak{A}}\xspace$ be a $C^*$-algebra and let $\mathfrak{I}$ be an ideal of $\ensuremath{\mathfrak{A}}\xspace$. Set
\[
\mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace ; \mathfrak{I} ) := \setof{ x \in \mathcal{M}( \ensuremath{\mathfrak{A}}\xspace ) }{ \text{$ax , x a \in \mathfrak{I}$ for all $a \in \ensuremath{\mathfrak{A}}\xspace$}}
\]
and set
\[
\mathcal{Q} ( \ensuremath{\mathfrak{A}}\xspace ; \mathfrak{I} ) := ( \mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace; \mathfrak{I} ) + \ensuremath{\mathfrak{A}}\xspace) / \ensuremath{\mathfrak{A}}\xspace.
\]
\begin{lemma}\label{lem:relative-multiplier-alg}
Let $\ensuremath{\mathfrak{A}}\xspace$ be a stable separable $C^*$-algebra such that $\ensuremath{\mathfrak{A}}\xspace$ has a unique non-trivial ideal $\mathfrak{I}$ with $\mathfrak{I}$ either isomorphic to $\ensuremath{\mathbb{K}}\xspace$ or $\mathfrak{I}$ is a stable, separable, purely infinite simple $C^*$-algebra and $\ensuremath{\mathfrak{A}}\xspace / \mathfrak{I}$ is either isomorphic to $\ensuremath{\mathbb{K}}\xspace$ or is a stable, separable, purely infinite simple $C^*$-algebra. Then $\mathcal{Q} ( \ensuremath{\mathfrak{A}}\xspace , \mathfrak{I} )$ is the unique non-trivial ideal of $\mathcal{Q} ( \ensuremath{\mathfrak{A}}\xspace )$.
\end{lemma}
\begin{proof}
Note that $\mathfrak{I}$ is an essential ideal of $\mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace , \mathfrak{I} )$. Hence, this embedding extends to an embedding $\ftn{\iota}{\mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace , \mathfrak{I} )}{\mathcal{M} (\mathfrak{I} ) }$. We claim that $\iota ( \mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace , \mathfrak{I} ) )$ is a full hereditary subalgebra of $\mathcal{M} (\mathfrak{I} )$.
Let $x = ( L_0, R_0 ) \in \mathcal{M} ( \mathfrak{I} )$ and let $s, t \in \mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace , \mathfrak{I} )$ (where we are using the double centralizer picture of the multiplier algebra). Define $\ftn{L,R}{\ensuremath{\mathfrak{A}}\xspace}{\ensuremath{\mathfrak{A}}\xspace}$ by $L(a) =s ( L_0 (ta) )$ and $R(a) = R_0(as)t$. Note that $L$ and $R$ are well-defined since $ta$ and $as$ are elements of $\mathfrak{I}$ for all $a \in \ensuremath{\mathfrak{A}}\xspace$. A computation shows that $L$ and $R$ are linear and $\| L \|$ and $\| R \|$ are bounded above by $\|s\|\cdot\|x\|\cdot\|t\|$.
Let $\{ e_n \}_{n = 1 }^\infty$ be an approximate identity for $\mathfrak{I}$. For all $a, b \in \ensuremath{\mathfrak{A}}\xspace$, we have that
\begin{align*}
R( ab) &= R_0( ab s) t = \lim_{ n \to \infty } R_0 ( ae_n b s ) t \\
&= \lim_{ n \to \infty} (ae_n) \left( R_0( bs) t \right) = \lim_{n \to \infty } a ( e_n R_0(bs) t) \\
&= a ( R_0(bs) t ) = a R(b), \\
L( ab) &= s L_0 ( tab) = \lim_{n \to \infty } s L_0 ( t a e_n b ) \\
&= \lim_{ n \to \infty } s L_0 ( t a ) (e_n b ) = \lim_{n \to \infty } ( s L_0 ( ta ) e_n ) b \\
&= ( s L_0 ( ta ) ) b = L(a) b,\\
R(a)b &=( R_0 ( as ) t ) b = R_0(as) (tb) \\
&= as L_0 ( t b ) = a ( s L_0 (tb) ) \\
&= a L(b).
\end{align*}
Hence, $y = (L,R)$ defines an element of $\mathcal{M}(\ensuremath{\mathfrak{A}}\xspace)$.
Let $a , b\in \ensuremath{\mathfrak{A}}\xspace$. Then
\[
L_a L (b) = L_a (s L_0 (tb) ) = (a s) L_0 (tb) = R_0( as ) tb = L_{ R_0(as) t }(b)
\]
and
\begin{align*}
R R_a (b) &= R ( ba ) = R_0( bas) t = \lim_{ n \to \infty } R_0 ( b e_n as ) t \\
&= \lim_{ n \to \infty }b (e_n R_0 ( as )t) = b ( R_0 ( as ) t ) = R_{ R_0(as) t }(b).
\end{align*}
Therefore, $( L_a, R_a )( L, R ) = ( L_a L , R R_a ) = ( L_{ R_0(as) t } , R_{ R_0(as) t } ) \in \mathfrak{I}$. Similarly computation shows that $( L, R )( L_a, R_a ) = (L_{s L_0 (ta) } , R_{ s L_0 (ta) } ) \in \mathfrak{I}$. Hence, $(L, R ) \in \mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace , \mathfrak{I} )$. Note that $\iota ( s ) = ( L_s, R_s )$, where we restrict $L_s$ and $R_s$ to $\mathfrak{I}$. Similarly, for $\iota (t)$. Thus,
\[
\iota( s ) x \iota(t) = ( L_s, R_s ) x ( L_t, R_t ) = (L_s, R_s ) ( L_0, R_0 ) ( L_t , R_t ) = (L_s L_0 L_t , R_t R_0 R_s )
\]
and
\begin{align*}
L_s L_0 L_t (z) &= L_s ( L_0 ( t z) ) = s L_0 ( tz) = L(z) \\
R_t R_0 R_s (z) &= R_t ( R_0 ( z s ) ) = R_0 ( zs ) t = R(z)
\end{align*}
for all $z \in \mathfrak{I}$. Hence, $\iota (y) = \iota (s ) x \iota (t)$. Therefore, $\iota ( \mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace , \mathfrak{I} ) )$ is a hereditary subalgebra of $\mathcal{M}( \mathfrak{I} )$.
We claim that $\iota ( \mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace , \mathfrak{I} ) ) \neq \mathfrak{I}$. Let $\{ s_n \}_{ n = 1 }^\infty$ be a collection of isometries in $\mathcal{M}(\ensuremath{\mathfrak{A}}\xspace)$ such that $\sum_{ n = 1}^\infty s_n s_n^*$ converges to $1_{\mathcal{M}(\ensuremath{\mathfrak{A}}\xspace)}$ in the strict topology (note such a collection of isometries exists since $\ensuremath{\mathfrak{A}}\xspace$ is a stable $C^*$-algebra). Let $a \in \mathfrak{I} \setminus \{0\}$. Then $\sum_{ n = 1 }^\infty s_n a s_n^*$ converges in the strict topology of $\mathcal{M}(\ensuremath{\mathfrak{A}}\xspace)$. Therefore, $x = \sum_{ n = 1 }^\infty s_n a s_n^*$ is an element of $\mathcal{M}(\ensuremath{\mathfrak{A}}\xspace)$. In fact, $x \in \mathcal{M}(\ensuremath{\mathfrak{A}}\xspace, \mathfrak{I} )$ since $a \in \mathfrak{I}$. Since $\| s_n a s_n^* \| = \|a\| \neq 0$, we have that $x \notin \ensuremath{\mathfrak{A}}\xspace$. Therefore, $\mathfrak{I} \neq \mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace ; \mathfrak{I} )$. So, $\iota ( \mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace , \mathfrak{I} ) ) \neq \mathfrak{I}$, which proves our claim.
By \cite[Theorem~3.2]{MR1203034}, $\mathcal{M} ( \mathfrak{I} )$ has exactly one non-trivial ideal $\mathfrak{I}$. Therefore, $\iota ( \mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace ; \mathfrak{I} ) )$ is a full hereditary subalgebra of $\mathcal{M} ( \mathfrak{I} )$. Thus, $\mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace ; \mathfrak{I} )$ has exactly one non-trivial, $\mathfrak{I}$. Consequently, $\mathcal{Q} ( \ensuremath{\mathfrak{A}}\xspace ; \mathfrak{I} )$ is a simple $C^*$-algebra.
Let $\ftn{\pi}{\ensuremath{\mathfrak{A}}\xspace}{\ensuremath{\mathfrak{A}}\xspace / \mathfrak{I}}$ be the canonical projection. Then it induces surjective \mbox{$\sp*$-}ho\-mo\-morphisms\xspace $\ftn{\widetilde{\pi}}{\mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace ) }{ \mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace / \mathfrak{I} ) }$ and $\ftn{\overline{\pi}}{ \mathcal{Q} ( \ensuremath{\mathfrak{A}}\xspace )}{ \mathcal{Q} ( \ensuremath{\mathfrak{A}}\xspace / \mathfrak{I} )}$. Note that $\ker( \widetilde{\pi} ) = \mathcal{M} ( \ensuremath{\mathfrak{A}}\xspace ; \mathfrak{I} )$ and $\ker( \widetilde{\pi} ) = \mathcal{Q} ( \ensuremath{\mathfrak{A}}\xspace ; \mathfrak{I} )$. Now, we have an exact sequence
\[
0 \to \mathcal{Q} ( \ensuremath{\mathfrak{A}}\xspace ; \mathfrak{I} ) \to \mathcal{Q} (\ensuremath{\mathfrak{A}}\xspace ) \to \mathcal{Q} ( \ensuremath{\mathfrak{A}}\xspace / \mathfrak{I} ) \to 0.
\]
By \cite[Theorem~3.2]{MR1203034}, $\mathcal{Q} ( \ensuremath{\mathfrak{A}}\xspace / \mathfrak{I} )$ is a simple $C^*$-algebra. Thus, $\mathcal{Q} ( \ensuremath{\mathfrak{A}}\xspace ; \mathfrak{I} )$ must be the unique non-trivial ideal of $\mathcal{Q} (\ensuremath{\mathfrak{A}}\xspace)$.
\end{proof}
\begin{theorem}\label{thm: classification special case}
Let $\ensuremath{\mathfrak{A}}\xspace_{1}$ and $\ensuremath{\mathfrak{A}}\xspace_{2}$ be unital $C^{*}$-algebras equipped with gauge actions. Suppose for each $i$, there exist gauge invariant ideals $\I_{i,1}$ and $\I_{i,2}$ of $\ensuremath{\mathfrak{A}}\xspace_{i}$ such that
\begin{enumerate}[(i)]
\item \label{thm: classification special case-1}
$\I_{i,1} \subseteq \I_{i,2}$,
\item \label{thm: classification special case-2}
$\I_{i,1} \cong \ensuremath{\mathbb{K}}\xspace$,
\item \label{thm: classification special case-3}
$\I_{i,2} / \I_{i,1}$ is isomorphic to the stabilization of a unital, simple purely infinite graph \mbox{$C\sp*$-}al\-ge\-bra\xspace,
\item \label{thm: classification special case-4}
$\ensuremath{\mathfrak{A}}\xspace_{i} / \I_{i,2} \cong C(S^{1})$, and
\item \label{thm: classification special case-5}
$\I_{i,2} / \I_{i,1}$ is an essential ideal of $\ensuremath{\mathfrak{A}}\xspace_{i} / \I_{i,1}$.
\end{enumerate}
If $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace (X_3; \ensuremath{\mathfrak{A}}\xspace_{1} \otimes \ensuremath{\mathbb{K}}\xspace ) \cong \ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace (X_3;\ensuremath{\mathfrak{A}}\xspace_{2} \otimes \ensuremath{\mathbb{K}}\xspace )$, then $\ensuremath{\mathfrak{A}}\xspace_{1} \otimes \ensuremath{\mathbb{K}}\xspace \cong \ensuremath{\mathfrak{A}}\xspace_{2} \otimes \ensuremath{\mathbb{K}}\xspace$.
\end{theorem}
\begin{proof}
Let $\alpha$ be the isomorphism from $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace (X_3; \ensuremath{\mathfrak{A}}\xspace_{1} \otimes \ensuremath{\mathbb{K}}\xspace
)$ to $\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace (X_3; \ensuremath{\mathfrak{A}}\xspace_{2} \otimes \ensuremath{\mathbb{K}}\xspace )$. Let $\mathfrak{e}_{i}$ be
the extension $0 \to \I_{i,2} \otimes \ensuremath{\mathbb{K}}\xspace \to \ensuremath{\mathfrak{A}}\xspace_{i} \otimes \ensuremath{\mathbb{K}}\xspace \to
\ensuremath{\mathfrak{A}}\xspace_{i} / \I_{i,2} \otimes \ensuremath{\mathbb{K}}\xspace \to 0$. We first show that
$\mathfrak{e}_i$ is a full extension. By
Lemma~\ref{lem:relative-multiplier-alg}, the corona algebra $\corona{
\I_{i,2} \otimes \ensuremath{\mathbb{K}}\xspace }$ has exactly one nontrivial ideal. This
ideal is precisely the kernel of the surjective map $\overline{\pi}
\colon \corona{ \I_{i,2} \otimes \ensuremath{\mathbb{K}}\xspace } \rightarrow \corona{ \I_{i,2} /
\I_{i,1} \otimes \ensuremath{\mathbb{K}}\xspace }$ that is induced by the surjective map $\pi
\colon \I_{i,2} \otimes \ensuremath{\mathbb{K}}\xspace \rightarrow \I_{i,2} / \I_{i,1} \otimes
\ensuremath{\mathbb{K}}\xspace$. Therefore, $x \in \corona{ \I_{i,2} \otimes \ensuremath{\mathbb{K}}\xspace}$ is full if and
only if its image in $\corona{ \I_{i,2} / \I_{i,1} \otimes \ensuremath{\mathbb{K}}\xspace }$ is
nonzero.
Note that the diagram
\[
\xymatrix{
0 \ar[r] & \I_{i,2} \otimes \ensuremath{\mathbb{K}}\xspace \ar[r] \ar[d]^{\pi} & \ensuremath{\mathfrak{A}}\xspace_{i} \otimes \ensuremath{\mathbb{K}}\xspace \ar[r] \ar[d] & ( \ensuremath{\mathfrak{A}}\xspace_{i} / \I_{i,2} ) \otimes \ensuremath{\mathbb{K}}\xspace \ar[r] \ar@{=}[d]& 0 \\
0 \ar[r] & ( \I_{i,2} / \I_{i,1} ) \otimes \ensuremath{\mathbb{K}}\xspace \ar[r] & ( \ensuremath{\mathfrak{A}}\xspace_{i} / \I_{i,1} ) \otimes \ensuremath{\mathbb{K}}\xspace \ar[r] & ( \ensuremath{\mathfrak{A}}\xspace_{i} / \I_{i,2} ) \otimes \ensuremath{\mathbb{K}}\xspace \ar[r] & 0
}
\]
is commutative. Hence, with $\boldsymbol{\tau}$ denoting Busby maps, $\overline{\pi} \circ \boldsymbol{\tau}_{\mathfrak{e}_i} = \boldsymbol{\tau}_{\mathfrak{g}_i}$, where $\mathfrak{g}_i$ is the extension
\[
\xymatrix{
0 \ar[r] & ( \I_{i,2} / \I_{i,1} ) \otimes \ensuremath{\mathbb{K}}\xspace \ar[r] & ( \ensuremath{\mathfrak{A}}\xspace_{i} / \I_{i,1} ) \otimes \ensuremath{\mathbb{K}}\xspace \ar[r] & ( \ensuremath{\mathfrak{A}}\xspace_{i} / \I_{i,2} ) \otimes \ensuremath{\mathbb{K}}\xspace \ar[r] & 0.
}
\]
By assumption \ref{thm: classification special case-5}, $\boldsymbol{\tau}_{ \mathfrak{g}_i } ( x )$ is nonzero in $\corona{ (\I_{i,2} / \I_{i,1}) \otimes \ensuremath{\mathbb{K}}\xspace }$ for all nonzero $x \in ( \ensuremath{\mathfrak{A}}\xspace_{i} / \I_{i,1} ) \otimes \ensuremath{\mathbb{K}}\xspace$. Hence, by the above observations, $\boldsymbol{\tau}_{\mathfrak{e}_i}(x)$ is full in $\corona{ \I_{i,2} \otimes \ensuremath{\mathbb{K}}\xspace }$. Since $\I_{i,2} \otimes \ensuremath{\mathbb{K}}\xspace$ has the corona factorization property (see, \emph{e.g.}\xspace, \cite[Proposition 6.1]{MR3056712}) $\mathfrak e_i$ is an absorbing extension.
Since $\ensuremath{\mathfrak{A}}\xspace_i / \I_{i,2} \otimes \ensuremath{\mathbb{K}}\xspace$ is $C( S^{1} ) \otimes \ensuremath{\mathbb{K}}\xspace$, there exists a \mbox{$\sp*$-}iso\-morphism\xspace $\beta_{2}$ from $\ensuremath{\mathfrak{A}}\xspace_{1} / \I_{1,2} \otimes \ensuremath{\mathbb{K}}\xspace$ to $\ensuremath{\mathfrak{A}}\xspace_{2} / \I_{2,2} \otimes \ensuremath{\mathbb{K}}\xspace$ which induces
$\alpha$ restricted to $K_*(\ensuremath{\mathfrak{A}}\xspace_1 / \mathfrak{I}_{1,2})$
(we are using the fact that a positive automorphism on $K_{*} ( C(S^{1}) )$ is induced by $\operatorname{id}_{ C( S^{1} ) \otimes \ensuremath{\mathbb{K}}\xspace}$ or $\psi \otimes \operatorname{id}_\ensuremath{\mathbb{K}}\xspace$ where $\psi$ sends the canonical generator of $C( S^{1})$, denoted by $z$, to $z^{-1}$). Note that $\I_{i,2}$ has a full projection, $\I_{1,2}$ is stably isomorphic to a unital \mbox{$C\sp*$-}al\-ge\-bra\xspace with exactly one nontrivial ideal that is isomorphic to $\ensuremath{\mathbb{K}}\xspace$ and the quotient by this ideal is isomorphic to a unital and simple purely infinite graph \mbox{$C\sp*$-}al\-ge\-bra\xspace. Using this observation together with \cite[Corollary~4.17 and Proposition~4.19]{arXiv:1301.7695v1}, there exists a \mbox{$\sp*$-}iso\-morphism\xspace $\ftn{ \beta_{0} }{ \I_{1,2} \otimes \ensuremath{\mathbb{K}}\xspace }{ \I_{2,2} \otimes \ensuremath{\mathbb{K}}\xspace }$ which induces $\alpha$ restricted to $K_*( \mathfrak{I}_{1,2})$.
Let $\mathfrak{f}_{1}$ be the extension obtained by pushing forward the extension $\mathfrak{e}_{1}$ via the \mbox{$\sp*$-}iso\-morphism\xspace $\beta_{0}$ and let $\mathfrak{f}_{2}$ be the extension obtained by pulling back the extension $\mathfrak{e}_{2}$ by the \mbox{$\sp*$-}iso\-morphism\xspace $\beta_{2}$. Since $\mathfrak{e}_{i}$ is an absorbing extension, we have that $\mathfrak{f}_{i}$ is an absorbing extension. By construction, $K_{*} ( \boldsymbol{\tau}_{ \mathfrak{f}_{1}} ) = K_{*} ( \boldsymbol{\tau}_{ \mathfrak{f}_{2} } )$ as homomorphisms from $K_{*} ( ( \ensuremath{\mathfrak{A}}\xspace_{1} / \I_{1,2} ) \otimes \ensuremath{\mathbb{K}}\xspace )$ to $K_{1-*} (\I_{2,2} \otimes \ensuremath{\mathbb{K}}\xspace )$. Hence, by the UCT of Rosenberg and Schochet \cite{MR894590}, $[ \boldsymbol{\tau}_{ \mathfrak{f}_{1}} ] = [ \boldsymbol{\tau}_{ \mathfrak{f}_{2} } ]$ in $\mathrm{KK}^{1} ( (\ensuremath{\mathfrak{A}}\xspace_{1} / \I_{1,2} ) \otimes \ensuremath{\mathbb{K}}\xspace , \I_{2,2} \otimes \ensuremath{\mathbb{K}}\xspace )$ since $K_{i} ( ( \ensuremath{\mathfrak{A}}\xspace_{1} \otimes \ensuremath{\mathbb{K}}\xspace ) / ( \I_{1,2} \otimes \ensuremath{\mathbb{K}}\xspace ) ) \cong \ensuremath{\mathbb{Z}}\xspace$ for each $i$.
Since $\mathfrak{f}_{i}$ are absorbing extensions, there exists a unitary $U$ in $\mathcal{M} ( \I_{2,2} \otimes \ensuremath{\mathbb{K}}\xspace )$ such that $\mathrm{Ad} ( \pi (U) ) \circ \boldsymbol{\tau}_{ \mathfrak{f}_{1} } = \boldsymbol{\tau}_{ \mathfrak{f}_{2}}$. One checks that $\mathrm{Ad} (U)$ induces a \mbox{$\sp*$-}iso\-morphism\xspace of extensions from $\mathfrak{f}_{1}$ to $\mathfrak{f}_{2}$. Since $\mathfrak{e}_{i}$ is isomorphic to $\mathfrak{f}_{i}$, we have that $\ensuremath{\mathfrak{A}}\xspace_{1} \otimes \ensuremath{\mathbb{K}}\xspace \cong \ensuremath{\mathfrak{A}}\xspace_{2} \otimes \ensuremath{\mathbb{K}}\xspace$.
\end{proof}
It is easy to see that this result applies to conclude that
$C^*(E_\curlyvee)\otimes \ensuremath{\mathbb{K}}\xspace\cong C^*(F_\curlyvee)\otimes \ensuremath{\mathbb{K}}\xspace$ for the
pair of examples in Example \ref{notalwayssame}. To deal with $C^*(E)\otimes \ensuremath{\mathbb{K}}\xspace$
and $C^*(F)\otimes \ensuremath{\mathbb{K}}\xspace$, we apply an unplugging trick to get:
\begin{corollary}\label{cor: classification special case}
Let $E_1$ and $E_2$ be finite graphs with $\operatorname{Prime}_\gamma(C^*(E_i))\cong X_3$ and $\tau_{E_i}(\{0\})\leq0$. If $\frX{X_3}{E_1}\cong \frX{X_3}{E_2}$, then $C^*(E_1) \otimes \ensuremath{\mathbb{K}}\xspace \cong C^*(E_2) \otimes \ensuremath{\mathbb{K}}\xspace$.
\end{corollary}
\begin{proof} Assume that $\frX{X_3}{E_1}\cong \frX{X_3}{E_2}$. We write ${\mathfrak p}^i_j$ for the ideals in $C^*(E_i)$ as in Example \ref{linearcase}. If $\tau_{E_i}({\mathfrak p}^i_1)=1$ we have Condition (H), and the full force of Theorem \ref{mainthm} applies. We may hence assume that $\tau_{E_i}({\mathfrak p}^i_1)=0$.
Again if $\tau_{E_i}({\mathfrak p}^i_2)=0$, we have Condition (H), so we may assume
$\tau_{E_i}({\mathfrak p}^i_2)=1$. When $\tau_{E_i}({\mathfrak p}^i_3)=-1$ we note that all the
conditions of Theorem~\ref{thm: classification special case} are met,
so that this result applies to give the desired conclusion. We thus
need only concern ourselves with the case
$\tau_{E_i}({\mathfrak p}^i_3)=0$.
In this case, we pass to $(E_i)_\curlyvee$ and note that
Theorem~\ref{thm: classification special case} applies. Since the
isomorphism provided by that result must satisfy the conditions of
Proposition \ref{prop: unplugging} because the ideal lattice is linear,
we
get the desired conclusion.
\end{proof}
We conclude:
\begin{example}\label{notalwayssameii}
With $E$ and $F$ the pair of graphs given in Figure \ref{firstexx}(b), we have
\[
C^*(E)\otimes \ensuremath{\mathbb{K}}\xspace\cong C^*(F)\otimes\ensuremath{\mathbb{K}}\xspace,
\]
and
\[
C^*(E_\curlyvee)\otimes \ensuremath{\mathbb{K}}\xspace\cong C^*(F_\curlyvee)\otimes\ensuremath{\mathbb{K}}\xspace.
\]
although (as seen in Example \ref{notalwayssame}) $E\not\ensuremath{\sim_{C\negthinspace E}}\xspace F$ and $E_\curlyvee\not\ensuremath{\sim_{C\negthinspace E}}\xspace F_\curlyvee$.
\end{example}
\section{Applications}\label{applications}
In this section, we give applications of our results.
\subsection{Type I/postliminal $C^{*}$-algebras}
In this section we study further the case where no vertex supports two distinct return paths, \emph{i.e.}\xspace\ the case of type I/postliminal \mbox{$C\sp*$-}al\-ge\-bras\xspace in our class, \emph{cf.}\xspace\ the remarks just after Lemma \ref{charKH}.
It was conjectured by Gene Abrams and Mark Tomforde in
\cite{MR2775826} that if the Leavitt path algebras $L_{\ensuremath{\mathbb{C}}\xspace} (E)$ and
$L_{\ensuremath{\mathbb{C}}\xspace} (F)$ are Morita equivalent, then $C^{*} (E)$ and $C^{*}(F)$
are strongly Morita equivalent (see \cite{MR2417402} for the
definition of $L_{\ensuremath{\mathbb{C}}\xspace} (E)$). Using Theorem~\ref{mainthm}, we can show
that their conjecture holds for finite graphs whose temperatures are
never positive. Moreover, we show that the converse holds as well in that case.
\begin{theorem}\label{coldiso}
Let $E$ and $F$ be finite graphs where $\max\tau_E,\max\tau_F\leq 0$. Then the following are equivalent:
\begin{enumerate}[(1)]
\item \label{coldiso-1}
$E \ensuremath{\sim_{M\negthinspace E}}\xspace F$.
\item \label{coldiso-2}
$L_{{\mathsf k}} ( E )$ and $L_{{\mathsf k}} (F)$ are Morita equivalent for any field ${\mathsf k}$.
\item \label{coldiso-3}
$C^{*} (E)\otimes\ensuremath{\mathbb{K}}\xspace\cong C^{*} (F)\otimes \ensuremath{\mathbb{K}}\xspace$.
\end{enumerate}
If $\tau_{E} =\tau_{F}={{0}}$, then \ref{coldiso-1}--\ref{coldiso-3} are equivalent to
\begin{enumerate}[(1)]\addtocounter{enumi}{3}
\item \label{coldiso-4}
the two-sided shift spaces ${\mathsf X}_{E}$ and ${\mathsf X}_{F}$ are flow equivalent.
\end{enumerate}
If $(\mathsf{B}_E,\mathsf{B}_F)$ is in standard form, then \ref{coldiso-1}--\ref{coldiso-3} are equivalent to
\begin{enumerate}[(1)]\addtocounter{enumi}{4}
\item \label{coldiso-5}
there exist matrices $U,V\in \ensuremath{\operatorname{SL}}\xspace_{\ensuremath{\mathcal{P}}\xspace}(\mathbf 1,\ensuremath{\mathbb{Z}}\xspace)$
so that $U\mathsf{B}_{E_\curlywedge}V=\mathsf{B}_{F_\curlywedge}$.
\end{enumerate}
\end{theorem}
\begin{proof}
By Section~3 of \cite{MR3045151} (see also \cite{MR3082546}), \ref{coldiso-1}
implies \ref{coldiso-2}. We can make sense of $\operatorname{Prime}_\gamma$ also for Leavitt path algebras over finite graphs (see Remark \ref{lpacomments}), and
we have that when $L_{\ensuremath{\mathbb{C}}\xspace} (E)$ and $L_{\ensuremath{\mathbb{C}}\xspace} (F)$ are
Morita equivalent, then $\operatorname{Prime}_\gamma(L_{\ensuremath{\mathbb{C}}\xspace} (E))\cong X\cong \operatorname{Prime}_\gamma(L_{\ensuremath{\mathbb{C}}\xspace} (F))$ for appropriately chosen $X$. Arguing as in the proof of Theorem~4.9
of \cite{MR3188556}, we get that
$\ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(X,C^*(E))\cong \ensuremath{\operatorname{FK}^+_\mathcal{R}}\xspace(X,C^*(F))$. By
this observation together with Theorem~\ref{mainthm}, since obviously
we have Condition~(H), we conclude that \ref{coldiso-2} implies \ref{coldiso-3}. Since $\max\tau\leq 0$, no
vertex supports two different return paths, so Move \mbox{\texttt{\textup{(C)}}}\xspace\ is never
allowed, and we have that $E \ensuremath{\sim_{C\negthinspace E}}\xspace F$ if and only if $E \ensuremath{\sim_{M\negthinspace E}}\xspace F$.
Therefore, Theorem~\ref{mainthm} gives that \ref{coldiso-3} implies \ref{coldiso-1}.
Assuming now that all components of $E$ and $F$ are cyclic, we get that \ref{coldiso-1} and \ref{coldiso-4} are equivalent by Lemma~\ref{flowvsME}.
Finally we get \ref{coldiso-1}$\Longleftrightarrow$ \ref{coldiso-5} by appealing to Theorem \ref{iffcharofMeq}.
\end{proof}
In general (as we shall discuss in \cite{Eilers-Restorff-Ruiz-Sorensen-2}), it may be computationally difficult to determine when two matrices are $\ensuremath{\operatorname{SL}_\calP}\xspace$-equivalent. This is because the problem is equivalent to solving
\begin{gather}
U\mathsf{B}_E = \mathsf{B}_F W\label{linpart}\\
\forall i\in\ensuremath{\mathcal{P}}\xspace: \det U\{i\}=\det W\{i\}=1\label{nonlinpart}
\end{gather}
where \eqref{nonlinpart} is not linear.
But when all blocks are $1\times 1$, the determinant conditions are equivalent to all diagonal blocks being identity matrices, and thus deciding if $U\mathsf{B}_EV=\mathsf{B}_F$ as in \ref{coldiso-5} of Theorem \ref{coldiso} reduces to the linear problem \eqref{linpart} which may readily be decided.
\subsection{Quantum lens spaces}
A class of quantum lens spaces $C(L_q(r;(m_1,\dots, m_n)))$ was
studied in \cite{MR2015735}, \cite{arXiv:1603.04678v1} and proved there to be graph
\mbox{$C\sp*$-}al\-ge\-bras\xspace over finite graphs. We immediately see that the \mbox{$C\sp*$-}al\-ge\-bras\xspace are postliminal/type I with every vertex supporting a loop. To decide any isomorphism question among two such \mbox{$C\sp*$-}al\-ge\-bras\xspace one hence need only to compare their $\Prime_\gamma$-spaces, and if these are homeomorphic, arrange that the corresponding matrices are in standard form and decide $\ensuremath{\operatorname{SL}_\calP}\xspace$-equivalence as in Theorem \ref{coldiso}\ref{coldiso-5} (for each possible homeomorphism).
As an immediate application, we shall see that in fact in some cases there are
several different quantum lens spaces associated to different choices
of secondary parameters $m_i$ even when the dimension $n$ and the
primary parameter $r$ are fixed. Although our classification result applies in the general setting of \cite{arXiv:1603.04678v1}, we will here consider only the original setup from \cite{MR2015735} where $\Prime_\gamma$ becomes the Alexandrov space of a linear order.
We emphasize the fact that even though the $K$-groups of the quantum lens spaces carry important information (\emph{cf.}\xspace\ \cite{MR2015735}, \cite{MR3448329}, \cite{arXiv:1603.04678v1}), they are not complete invariants. It follows from Theorem \ref{mainthm} that the reduced ordered filtered $K$-theory is complete, but as we shall see it is much more convenient to work with \ensuremath{\operatorname{SL}_\calP}\xspace-equivalence in this setting.
\begin{definition}
For each $n \in \ensuremath{\mathbb{N}}\xspace$, define the directed graph $L_{2n-1}$ as the graph with $n$ vertices, $L_{2n - 1}^{0} = \{ v_{1} , \dots, v_{n} \}$, and $\frac{ n( n+1 ) }{2}$ edges $\bigcup_{ i = 1}^{n} \{ e_{i,j} \mid j = i, i+1, \dots, n \}$ with $s ( e_{i,j} ) = v_{i}$ and $r( e_{i,j} ) = v_{j}$. For example, $L_{5}$ is the graph
\begin{align*}
\xymatrix{
v_{1} \ar@(ul,ur)[]^{e_{1,1} } \ar[rr]^{ e_{1,2} } \ar@/_1pc/[rrrr]_{e_{1,3}} & & v_{2} \ar@(ul,ur)[]^{ e_{2,2} } \ar[rr]^{e_{2,3}} & & v_{3} \ar@(ul,ur)[]^{e_{3,3}}
}
\end{align*}
\end{definition}
\begin{definition}
For each $r , n \in \ensuremath{\mathbb{N}}\xspace$ and ${\underline{m}}=( m_{1} , \dots, m_{n} ) \in \ensuremath{\mathbb{N}}\xspace^{n}$, we define the directed graph $L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r}$ as follows:
\begin{enumerate}[(i)]
\item The set of vertices is
\begin{align*}
( L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r} )^{0} = L_{2n-1}^{0} \times \ensuremath{\mathbb{Z}}\xspace_{r}.
\end{align*}
\item The set of edges is
\begin{align*}
( L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r} )^{1} = L_{2n-1}^{1} \times \ensuremath{\mathbb{Z}}\xspace_{r}.
\end{align*}
\item $s( e_{i,j} , k ) = ( v_{i} , k - m_{i} )$ and $r( e_{i,j} , k ) = ( v_{j} , k )$
\end{enumerate}
\end{definition}
For each $i$, let $( L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r} ) \langle i \rangle$ be the subgraph with vertex set $\{ v_{i} \} \times \ensuremath{\mathbb{Z}}\xspace_{r}$ and edge set $\{ e_{i,i} \} \times \ensuremath{\mathbb{Z}}\xspace_{r}$. For each $i_{1} \leq i_{2}\leq \cdots \leq i_{t}$, let $( L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r} ) \langle i_{1} , i_{2} , \dots, i_{t} \rangle$ be the subgraph with vertex set $\bigcup_{ l = 1}^{t} \{ v_{l} \} \times \ensuremath{\mathbb{Z}}\xspace_{r}$ and edge set the set of all edges $e$ in $L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r}$ such that $s(e), r(e) \in \bigcup_{ l = 1}^{t} \{ v_{l} \} \times \ensuremath{\mathbb{Z}}\xspace_{r}$.
\begin{definition}
Let $r \in \ensuremath{\mathbb{N}}\xspace$ and $( m_{1} , m_{2} , \dots, m_{n} ) \in \ensuremath{\mathbb{N}}\xspace^{n}$ with $r \geq 2$ and $\gcd( m_{i} , r ) =1$ for all $i$. A path $\alpha = ( e_{i_{1} , j_{1} } , k_{1} ) \cdots ( e_{i_{r},j_{r}} , k_{\ell} )$ in $L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r}$ is called \emph{$0$-simple} if $k_{1} = m_{i_{1}}$, $k_{a} \neq 0$ for $a \neq \ell$, and $k_{\ell} = 0$.
Note that for each $0$-simple path $\alpha= ( e_{i_{1} , j_{1} } , k_{1} ) \cdots ( e_{i_{\ell},j_{\ell}} , k_{\ell} ) $, we have that $s( \alpha ) = ( v_{i_{1}} , 0 )$ and $r( \alpha ) = ( v_{j_{\ell}} , 0 )$. Thus the $0$-simple paths may be thought of as paths starting and ending at vertices of the form $(v,0)$, but avoiding all such vertices along the way.
A $0$-simple path $\alpha = ( e_{i_{1} , j_{1} } , k_{1} ) \cdots ( e_{i_{\ell},j_{\ell}} , k_{\ell} )$ is called \emph{$k$-step} if there exist positive integers $ t_{1} < t_{2} < \dots < t_{k+1}$ such that $t_{1} = i_{1}$, $t_{k+1} = j_{\ell}$, and for each $2 \leq q \leq k$, we have that
\begin{align*}
\setof{ r( ( e_{i_{s} , j_{s} } , k_{s} ) ) }{ 1 \leq s \leq \ell } \cap ( ( L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r} )\langle t_{q} \rangle )^{0}\neq \emptyset
\end{align*}
and
\begin{align*}
\setof{ r( ( e_{i_{s} , j_{s} } , k_{s} ) )}{ 1 \leq s \leq \ell } \subseteq \bigcup_{ i = 1}^{k+1} ( ( L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r} )\langle t_{i} \rangle )^{0}.
\end{align*}
\end{definition}
\begin{definition}
Let $r \in \ensuremath{\mathbb{N}}\xspace$ and ${\underline{m}}=( m_{1} , m_{2} , \dots, m_{n} ) \in \ensuremath{\mathbb{N}}\xspace^{n}$ with $\gcd( m_{i} , r ) =1$ and $r \geq 2$. Define $L_{2n-1}^{(r; {\underline{m}} )}$ to be the graph with vertices $\{ (v_{1}, 0) , \dots, (v_{n}, 0 ) \}$, the edges of $L_{2n-1}^{(r; {\underline{m}} )}$ consisting of all $0$-simple paths in $L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r}$, and the range and source maps extending the range and source maps of $L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r}$.
\end{definition}
Note that by our assumption on the $m_i$, they are always units in $(\ensuremath{\mathbb{Z}}\xspace_r\backslash\{0\},\cdot)$. We denote by $m_i^{-1}$ any representative in $\ensuremath{\mathbb{Z}}\xspace$ of a multiplicative inverse to $m_i$ modulo $r$.
\begin{lemma}\label{l:PathsQuantum}
Let $r \in \ensuremath{\mathbb{N}}\xspace$ and ${\underline{m}}\in \ensuremath{\mathbb{N}}\xspace^{n}$ with $\gcd( m_{i} , r ) =1$ and $r \geq 2$.
\begin{enumerate}[(i)]
\item\label{l:PathsQuantum:1} For each $i, j$ with $i +1 \leq j$, the number of $1$-step $0$-simple paths from $( v_{i} , 0 )$ to $( v_{j} , 0 )$ is $r$.
\item\label{l:PathsQuantum:2} For each $i,j$ with $i+2 \leq j$, the number of $2$-step $0$-simple paths from $( v_{i} , 0 )$ to $( v_{j} , 0 )$ is $\frac{ r ( r - 1 ) }{ 2 }( j - i - 1 )$.
\item\label{l:PathsQuantum:3} For each $i$, the number of $3$-step $0$-simple paths from $( v_{i} , 0 )$ to $( v_{i+3} , 0 )$ is congruent to $- m_{i+2}^{-1} m_{i+1} \left( \frac{ r ( r - 1 )( r - 2 ) }{ 3 } \right)$ modulo $r$
\end{enumerate}
Consequently, the number of $0$-simple paths from $( v_{i} , 0 )$ to $(v_{i+2}, 0 )$ is $\frac{ r ( r + 1 ) }{ 2 }$ and the number of $0$-simple paths from $( v_{i} , 0 )$ to $(v_{i+3}, 0 )$ is congruent to $$- m_{i+2}^{-1} m_{i+1} \left( \frac{ r ( r - 1 )( r - 2 ) }{ 3 } \right)$$ modulo $r$.
\end{lemma}
\begin{proof}
We first prove \ref{l:PathsQuantum:1}. Note that for each $0 \leq k < r$, there is exactly one edge from $( v_{i} , k )$ to $(L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r})\langle j \rangle$. Since there is exactly 1 path from $( v_{j}, l )$ to $( v_{j} , 0 )$ which passes through $( v_{j} , 0 )$ once, we have that the number of $1$-step $0$-simple paths from $( v_{i} , 0 )$ to $( v_{j} , 0 )$ is equal to the number of edges in the subgraph $(L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r})\langle i \rangle$. This is equal to $r$, so \ref{l:PathsQuantum:1} holds.
We now prove \ref{l:PathsQuantum:2}. Let $V$ be the set of all $2$-step $0$-simple paths from $( v_{i}, 0 )$ to $( v_{j} , 0 )$. For each $l$ with $1 \leq l \leq j-i-1$, let $V_{l}$ be the set of all $2$-step $0$-simple paths from $( v_{i} , 0 )$ to $( v_{j} , 0 )$ that goes through the subgraph $( L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r} ) \langle i+l\rangle$. Then $V = \bigsqcup_{ l = 1}^{j-i-1} V_{l}$. By symmetry $| V_{l} | = | V_{1} |$, so $| V | = | V_{1} | (j - i - 1)$.
Let $\alpha = \alpha_{1} \cdots \alpha_{t} \in V_{1}$ and recall that for all $k$, $r(\alpha_k)\neq (v_i,0)$ and $r(\alpha_k)\neq (v_{i+1},0)$. Since for each $1 \leq l \leq r-1$,
there is exactly one path from $(v_i,0)$ to $(v_i,l-m_i)$ that does not come back to $(v_i,0)$, and there is exactly one edge from $(v_i,l-m_i)$ to $(v_{i+1},l)$,
we have that
\begin{align*}
| V_{1} | = \sum_{ l = 1}^{r-1} P_{l}
\end{align*}
where $P_{l}$ is the number of paths from $( v_{i+1} , m_{i+1}l )$
to $( v_{j} , 0 )$ in the subgraph $( L_{2n-1} \times_{{\underline{m}}} \ensuremath{\mathbb{Z}}\xspace_{r} )\langle i+1, j \rangle$ that do not go through $( v_{i+1} , 0 )$. Clearly, $P_l=r-l$, so
\begin{align*}
| V_{1} | = \sum_{ l = 1}^{r-1} P_{l} = \sum_{ l = 1}^{r-1} ( r - l ) = r (r-1) - \frac{ r(r-1)}{2} = \frac{ r ( r - 1 ) }{ 2 }.
\end{align*}
Therefore, \ref{l:PathsQuantum:2} holds.
We now prove \ref{l:PathsQuantum:3}. For each $1\leq l\leq r-2$, we have an edge $(e_{i+1},m_{i+1}(l+1))$ from $(v_{i+1},m_{i+1}l)$ to $(v_{i+2},m_{i+1}(l+1))$.
We let $Q_l$ be the number of paths from $(v_{i+2},m_{i+1}(l+1))$ to $(v_ {i+3},0)$ that do not go through $(v_{i+2},0)$ and only go once through $(v_{i+3},0)$. Since there are exactly $l$ paths from $(v_i,0)$ to $(v_{i+1},m_{i+1}l)$ that do not come back to $(v_i,0)$ and do not go through $(v_{i+1},0)$, we have that the number of $3$-step $0$-simple paths from $(v_i,0)$ to $(v_{i+3},0)$ is $\sum_{l=1}^{r-2} l Q_l$ .
Recall that $m_{i+2}^{-1}$ is a representative of the multiplicative inverse of $m_{i+2}$ modulo $r$, and let $s_l$ be the integer such that $0<m_{i+2}^{-1}m_{i+1}(l+1)+rs_l<r$. Since $m_{i+1}(l+1)$ is congruent to $m_{i+2}(m_{i+2}^{-1}m_{i+1}(l+1)+rs_l)$ modulo $r$, it follows from the proof of part \ref{l:PathsQuantum:2} that
$$Q_l=r-(m_{i+2}^{-1}m_{i+1}(l+1)+rs_l).$$ Hence, the number of $3$-step $0$-simple paths from $( v_{i} , 0 )$ to $( v_{i+3} , 0 )$ is
\begin{align*}
&\sum_{ l = 1}^{r-2} l (r - m_{i+2}^{-1} m_{i+1} (l+1) - r s_{l} ) \\
&\qquad \equiv \sum_{ l = 1}^{r-2}( - m_{i+2}^{-1} m_{i+1} l(l+1) ) \mod r \\
&\qquad \equiv -m_{i+2}^{-1} m_{i+1} \frac{ r( r-1)( r-2)}{3} \mod r.
\end{align*}
Hence, \ref{l:PathsQuantum:3} holds.
For the last part of the lemma, by \ref{l:PathsQuantum:1} and \ref{l:PathsQuantum:2}, we have that the number of $0$-simple paths from $( v_{i} , 0 )$ to $( v_{i+2} , 0 )$ is equal to $ r + \frac{ r ( r - 1 ) }{2} = \frac{ r ( r + 1 ) }{ 2 }$ and by \ref{l:PathsQuantum:1}, \ref{l:PathsQuantum:2}, \ref{l:PathsQuantum:3}, we have that the number of $0$-simple paths from $( v_{i}, 0 )$ to $( v_{i+3} , 0 )$ is congruent to $r + \frac{ r ( r - 1 ) }{2} + \frac{ r ( r - 1 ) }{2} - m_{i+2}^{-1} m_{i+1} \left( \frac{ r ( r - 1 )( r - 2 ) }{ 3 } \right)$ modulo $r$. It is now clear that the conclusion holds.
\end{proof}
\begin{corollary}\label{Kthy}
$K_0(C^{*} ( L_{2n-1}^{(r ; {\underline{m}} )} ))\cong \ensuremath{\mathbb{Z}}\xspace\oplus G$ for $G$ some group of order $|G|=r^{n-1}$
\end{corollary}
\begin{proof}
The first row and the last column of $(\mathsf{B}_{L_{2n-1}^{(r;{\underline{m}}')}})^{\mathsf T}$ are zero. The remaining $(n-1)\times (n-1)$ submatrix is upper triangular and has $r$ in the diagonal as seen in Lemma~\ref{l:PathsQuantum}\ref{l:PathsQuantum:1}, and thus the determinant is $r^{n-1}$. Now the lemma follows (\emph{e.g.}\xspace\ by using the Smith normal form).
\end{proof}
Determining $G$ exactly is a difficult problem, \emph{cf.}\xspace\ \cite{MR3448329}.
Since we obviously have
\[
|\Prime_\gamma(C^*(L_{2n-1}^{(r;{\underline{m}})}))|=|\Gamma_{L_{2n-1}^{(r;{\underline{m}})}}|=n,
\]
the isomorphism class of
$C^*(L_{2n-1}^{(r;{\underline{m}})})\otimes\ensuremath{\mathbb{K}}\xspace$ determines $n$ and hence, by Corollary \ref{Kthy}, also $r$. Further, Lemma \ref{l:PathsQuantum}\ref{l:PathsQuantum:1} and \ref{l:PathsQuantum:2} show that the graphs and their adjacency matrices are the same irrespective of ${\underline{m}}$ when $r$ is fixed and $n\leq 3$. When $n=4$, something new happens precisely when $r$ is a multiple of 3.
\begin{theorem}\label{t:QuantumIso}
Let $r\geq 2$ be given and let ${\underline{m}}=( m_{1} , m_{2} , m_{3} , m_{4} )$ and ${\underline{n}}=( n_{1} , n_{2} , n_{3} , n_{4} )$ be given in $\ensuremath{\mathbb{N}}\xspace^4$ such that $\gcd( m_{i} , r)=\gcd( n_{i} , r ) =1$ for all $i$. Then the following are equivalent:
\begin{enumerate}[(1)]
\item \label{t:QuantumIso-1}
$C^{*} ( L_{7}^{(r; {\underline{m}})} ) \cong C^{*} ( L_{7}^{(r; {\underline{n}})} )$,
\item \label{t:QuantumIso-2}
$C^{*} ( L_{7}^{(r ; {\underline{m}} )} )\otimes\ensuremath{\mathbb{K}}\xspace\cong C^{*} ( L_{7}^{(r; {\underline{n}} )} )\otimes \ensuremath{\mathbb{K}}\xspace$,
\item \label{t:QuantumIso-3}
$\left( m_{3}^{-1} m_{2} - n_{3}^{-1} n_{2} \right) \left( \frac{ r ( r - 1)( r - 2 ) }{ 3 } \right) \equiv 0 \mod r$.
\end{enumerate}
\end{theorem}
\begin{proof}
Let $\mathsf{A}_{{\underline{m}}}$ be the adjacency matrix for $L_{7}^{(r ;{\underline{m}} )}$ and let $\mathsf{A}_{{\underline{n}}}$ be the adjacency matrix for $L_{7}^{(r ;{\underline{n}} )}$, with $\mathsf{B}_{{\underline{m}}}$ and $\mathsf{B}_{{\underline{n}}}$ obtained by subtraction of the identity matrix as usual. By Lemma~\ref{l:PathsQuantum},
\begin{align*}
\mathsf{A}_{{\underline{m}}} =
\begin{pmatrix}
1 & r &\frac{ r ( r + 1 ) }{2} & x \\
0 & 1 & r&\frac{ r ( r + 1 ) }{2}\\
0&0&1&r\\
0&0&0& 1
\end{pmatrix}
\quad \text{and} \quad \mathsf{A}_{{\underline{n}}} =
\begin{pmatrix}
1 & r &\frac{ r ( r + 1 ) }{2} & y \\
0 & 1 & r&\frac{ r ( r + 1 ) }{2}\\
0&0&1&r\\
0&0&0& 1
\end{pmatrix}
\end{align*}
where $x \equiv - m_{3}^{-1} m_{2} \frac{ r(r-1)( r-2) }{3 } \mod r$ and $y \equiv - n_{3}^{-1} n_{2} \frac{ r(r-1)( r-2) }{3 } \mod r$.
We first show that \ref{t:QuantumIso-2} implies \ref{t:QuantumIso-3}. By Theorem~\ref{coldiso}\ref{coldiso-5}, there
exist $U, V \in \operatorname{SL}_{\ensuremath{\mathcal{P}}\xspace_4} (\mathbf{1}, \ensuremath{\mathbb{Z}}\xspace )$ such that $U \mathsf{B}_{{\underline{m}}}
V = \mathsf{B}_{{\underline{n}}}$, with $\ensuremath{\mathcal{P}}\xspace_4=\{1,2,3,4\}$ ordered linearly. Note that $U, V$ are upper triangular matrices and
$U\{ i \} = V \{ i \} = 1$ for $i = 1, 2, 3, 4$.
A computation implies that
\begin{align*}
x + r s_{1} + \frac{ r ( r+1) }{2} s_{2} = y
\end{align*}
for some $s_{1} , s_{2} \in \ensuremath{\mathbb{Z}}\xspace$. Since $y \equiv - n_{3}^{-1} n_{2} \frac{ r(r-1)( r-2) }{3 } \mod r $ and $x \equiv - m_{3}^{-1} m_{2} \frac{ r(r-1)( r-2) }{3 } \mod r $, we have that
\begin{align*}
( m_{3}^{-1}m_{2} - n_{3}^{-1} n_{2} ) \frac{ r(r-1)( r-2) }{3 } \equiv \frac{r(r+1)}{2}s_{2} \mod r.
\end{align*}
Thus,
\begin{align}\label{eq:QuantumIso1}
( m_{3}^{-1}m_{2} - n_{3}^{-1} n_{2} ) \frac{ r(r-1)( r-2) }{3 } + r m = \frac{r(r+1)}{2}s_{2}
\end{align}
for some $m \in \ensuremath{\mathbb{Z}}\xspace$.
Suppose $r$ is odd. Then $\frac{r(r+1)}{2}s_{2} \equiv 0 \mod r$ and hence \ref{t:QuantumIso-3} holds. Suppose $r$ is even, say $r = 2^{t} k$ where $\gcd( k , 2 ) = 1$. Dividing Equation~(\ref{eq:QuantumIso1}) by $2^{t-1}$, we get
\begin{align}\label{eq:QuantumIso2}
( m_{3}^{-1}m_{2} - n_{3}^{-1} n_{2} ) \frac{ 2k(r-1)( r-2) }{3 } + 2m k = k(r+1)s_{2}.
\end{align}
Since $3$ divides $r( r - 1)( r - 2)$, we have that $3$ divides $k(r-1)(r-2)$. Therefore, $\frac{ k ( r - 1)( r - 2) }{3} \in \ensuremath{\mathbb{Z}}\xspace$. Hence, the left hand side of Equation~(\ref{eq:QuantumIso2}) is divisible by $2$ which implies that $2$ divides $(r+1) s_{2}$. Since $r$ is even, $2$ divides $s_{2}$. Thus, $\frac{ r ( r+1) }{2} s_{2} \equiv 0 \mod r$. Hence, \ref{t:QuantumIso-3} holds.
We now show that \ref{t:QuantumIso-3} implies \ref{t:QuantumIso-1}. Since $\left( m_{3}^{-1} m_{2} - n_{3}^{-1} n_{2} \right) \left( \frac{ r ( r - 1)( r - 2 ) }{ 3 } \right) \equiv 0 \mod r$, $x \equiv - m_{3}^{-1} m_{2} \frac{ r(r-1)( r-2) }{3 } \mod r$, and $y \equiv - n_{3}^{-1} n_{2} \frac{ r(r-1)( r-2) }{3 } \mod r$, we have that $x \equiv y \mod r$. Therefore, $x + r s = y + r t$ for some positive integers $s, t$.
Consider the matrix
\begin{align*}
C =
\begin{pmatrix}
1 & r &\frac{ r ( r + 1 ) }{2} & x+rs \\
0 & 1 & r&\frac{ r ( r + 1 ) }{2}\\
0&0&1&r\\
0&0&0& 1
\end{pmatrix} = \begin{pmatrix}
1 & r &\frac{ r ( r + 1 ) }{2} & y+rt\\
0 & 1 & r&\frac{ r ( r + 1 ) }{2}\\
0&0&1&r\\
0&0&0& 1
\end{pmatrix} \end{align*}
By applying Proposition~\ref{p:isoaddingrows}, $s$ times (note that $x>0$), we get that $C^{*} ( L_{7}^{(r; {\underline{m}})} ) \cong C^{*} ( \mathsf{E}_{C} )$. Similarly, we can apply Proposition~\ref{p:isoaddingrows}, $t$ times, we get that $C^{*} ( L_{7}^{(r; {\underline{n}})} ) \cong C^{*} ( \mathsf{E}_{C} )$.
\end{proof}
It is in fact true in general (also in the general setting of \cite{arXiv:1603.04678v1}) that whenever two quantum lens spaces are stably isomorphic, they are isomorphic. We will pursue this in \cite{Eilers-Restorff-Ruiz-Sorensen-2}.
\begin{corollary}\label{c:QuantumIso}
If $3$ does not divide $r$, then
\begin{align*}
C^{*} ( L_{7}^{(r ; (1,1,1,1) )} ) \cong C^{*} ( L_{7}^{(r ; {\underline{m}} )} )
\end{align*}
for all ${\underline{m}}=( m_{1} , m_{2} , m_{3}, m_{4} ) \in \ensuremath{\mathbb{N}}\xspace^{4}$ with $\gcd( m_{i} , r ) = 1$.
Suppose $r = 3s$ and let ${\underline{m}}\in \ensuremath{\mathbb{N}}\xspace^{4}$ with $\gcd( m_{i} , r ) = 1$ be given. Then
\begin{align*}
C^{*} ( L_{7}^{(r ; {\underline{m}} )} ) \cong C^{*} ( L_{7}^{(r ; (1,1,1,1) )} )
\end{align*}
if and only if $m_{2} \equiv m_{3} \mod 3$ and
\begin{align*}
C^{*} ( L_{7}^{(r ; {\underline{m}} )} ) \cong C^{*} ( L_{7}^{(r ; (1, 1, r-1 ,1 ))} )
\end{align*}
if and only if $m_{2} \not\equiv m_{3} \mod 3$.
\end{corollary}
The isomorphism question for quantum lens spaces was introduced in \cite{MR2015735} and some $K$-groups were explicitly computed there. We note here that the $K$-groups in their own right do not contain sufficient information to classify, even if one takes the order into account.
\begin{remark}
The triple
\begin{align*}
\left( K_{0} ( C^{*} ( L_{7}^{(r ; {\underline{m}} )} ) ) , K_{0} ( C^{*} ( L_{7}^{(r ; {\underline{m}} )} ) )_{+} , K_{1} ( C^{*} ( L_{7}^{(r ; {\underline{m}} )} ) ) \right)
\end{align*}
is not a complete isomorphism invariant.
Set $E = L_{7}^{(3; (1, 1,1,1)) }$ and $F= L_{7}^{(3;(1,1,2,1))}$ with adjacency matrices
\[
\mathsf{A}_E=\begin{pmatrix}
1 & 3 &6& 10 \\
0 & 1 & 3&6\\
0 & 0& 1 & 3 \\
0 & 0 & 0 & 1
\end{pmatrix}
\qquad
\mathsf{A}_F=\begin{pmatrix}
1 & 3 &6& 11 \\
0& 1 & 3&6\\
0 & 0& 1 & 3 \\
0 & 0 & 0 & 1
\end{pmatrix}
\]
By Corollary~\ref{c:QuantumIso}, we have that $C^{*} (E)$ and $C^{*} (F)$ are not stably isomorphic. We will show that
\begin{align*}
\left( K_{0} ( C^{*} ( E ) ) , K_{0} ( C^{*} ( E ) )_{+} , K_{1} ( C^{*} ( E ) ) \right) \cong \left( K_{0} ( C^{*} ( F ) ) , K_{0} ( C^{*} ( F ) )_{+} , K_{1} ( C^{*} ( F ) ) \right)
\end{align*}
Because of the symmetry in the antidiagonal of these two matrices, we have $\mathsf{C}_E=\mathsf{B}_E$ and $\mathsf{C}_F=\mathsf{B}_F$ and may hence consider the $K$-groups as given by the kernels and cokernels of $\mathsf{B}_E$ and $\mathsf{B}_F$ themselves (see Remark \ref{howtocompute} and Section \ref{sec:red-filtered-K-theory-K-web-GLP-and-SLP-equivalences})
Let $e_{i}$ be the vector with $1$ in the $i$-th coordinate and zero
elsewhere, let $[ e_{i} ]_{E}$ be the class in $\operatorname{cok}(
\mathsf{B}_{E})$, and let $[ e_{i} ]_{F}$ be the class in
$\operatorname{cok}( \mathsf{B}_{F})$. Under our identification of the $K_0$-groups of $C^*(E)$ and $C^*(F)$ with cokernels of $\mathsf{B}_E$ and $\mathsf{B}_F$, the positive cones become exactly
\begin{align*}
S_{E} = \{ n_{1} [ e_{1} ]_{E} + n_{2} [ e_{2} ]_{E} + n_{3} [ e_{3} ]_{E} + n_{4} [ e_{4} ]_{E} : n_{i} \in \ensuremath{\mathbb{N}}\xspace_{ 0 } \}
\end{align*}
and
\begin{align*}
S_{F} = \{ n_{1} [ e_{1} ]_{F} + n_{2} [ e_{2} ]_{F} + n_{3} [ e_{3} ]_{F} + n_{4} [ e_{4} ]_{F} : n_{i} \in \ensuremath{\mathbb{N}}\xspace_{ 0 } \},
\end{align*}
respectively.
Hence, it is enough to show that
\begin{align*}
( \operatorname{cok}( \mathsf{B}_{E}) , S_{E} , \ker( \mathsf{B}_{E} )) \cong ( \operatorname{cok}( \mathsf{B}_{F}) , S_{F} , \ker( \mathsf{B}_{F} )).
\end{align*}
Set
\begin{align*}
U=
\begin{pmatrix}
10&-18&9&0\\ 6&-11&6&0\\3&-6&4&0\\0&0&0&1
\end{pmatrix}
\qquad
W=
\begin{pmatrix}
1&0&0&0\\0&-1&0&-1\\0&0&1&0\\0&3&0&2
\end{pmatrix}
\end{align*}
A computation shows that $U$ and $W$ are in $\ensuremath{\operatorname{GL}}\xspace_{4} (\ensuremath{\mathbb{Z}}\xspace)$ and
$U \mathsf{B}_{ E } = \mathsf{B}_{F} W$.
Thus, $U$ induces an isomorphism from $\operatorname{cok}( \mathsf{B}_{E})$ to $\operatorname{cok}( \mathsf{B}_{F})$ and $W$ induces an isomorphism from $\ker( \mathsf{B}_{E})$ to $\ker( \mathsf{B}_{F})$ as described in Section \ref{UVinduce}.
It is clear that $U ( [e_{i}]_{E} ) \in S_{F}$ for all $i \neq 2$.
Note that in $\operatorname{cok}( \mathsf{B}_{F})$, we have that
\begin{align*}
U ( [ e_{2} ]_{E} ) =
\begin{pmatrix}
-18\\-11\\-6\\0
\end{pmatrix} =
\begin{pmatrix}
15 \\
7\\
3 \\
0
\end{pmatrix}
+ \mathsf{B}_{F}\begin{pmatrix}
0 \\
0 \\
0 \\
-3
\end{pmatrix} \in S_{F}
\end{align*}
In the other direction, since
\[
U^{-1} = \begin{pmatrix}
-8&18&-9&0\\-6&13&-6&0\\-3&6&-2&0\\0&0&0&1\end{pmatrix}
\]
we may argue similarly.
\end{remark}
\subsection{Atlas of graph $C^{*}$-algebras of small graphs}
Inspired by a similar undertaking for Leavitt path algebras (\cite{MR3201827}), we end by a complete analysis of the stable isomorphism problem for small graphs, focusing on simple graphs with no more than 4 vertices. Although our invariant may be efficiently computed by methods outlined in \cite{serj:ciugc} we do not know an efficient general procedure for deciding whether or not an isomorphism exists between a pair of invariants, and further we will attempt to study also the few cases where our Condition (H) is not met, so instead of appealing exclusively to our invariant we will proceed by defining two equivalences on the set of graphs under investigation, approximating stable isomorphism of the associated graph algebras on both sides. The number of cases in need of further study is then so small that we may resolve it case by case.
\begin{definition}
The \emph{$K$-temperature} of a finite graph $E$ is the map \fctw{\mathfrak{t}_E^K}{\Gamma_E}{\{0,-1\}\cup\boldsymbol{Ab}} given by
$$\mathfrak{t}_E^K(\gamma)=\begin{cases}\tau_E(\upsilon_E(\gamma)), & \tau_E(\upsilon_E(\gamma))<1,\\ K_0(C^*(E)(\{\upsilon_E(\gamma)\})),&\tau_E(\upsilon_E(\gamma))=1.\end{cases}$$
\end{definition}
Note that when $\mathsf{B}_E\in\MPZccc$, then $K_0(C^*(E)(\{\upsilon_E(\gamma)\}))\cong \operatorname{cok}((\mathsf{B}_E^\bullet\{\mathcal{Y}_{\mathsf{B}_E}^{-1}(\gamma)\})^{\mathsf T})$.
\begin{definition}
We say that two graphs $E$ and $F$ with $(\mathsf{B}_E,\mathsf{B}_F)$ in standard form are \emph{outer equivalent}, and write $E\equiv_O F$, if
\begin{enumerate}[(i)]
\item $\operatorname{cok}((\mathsf{B}_E^\bullet)^{\mathsf T})\cong\operatorname{cok}((\mathsf{B}_F^\bullet)^{\mathsf T})$, and
\item for some order isomorphism \fct{h}{\Gamma_E}{\Gamma_F}, $\mathfrak{t}_F^K(h(\gamma))$ and $\mathfrak{t}_E^K(\gamma)$ are either isomorphic Abelian groups or equal numbers for all $\gamma\in\Gamma_E$
\end{enumerate}
\end{definition}
We will say that a row or column addition in a matrix $\mathsf{B}_E$
representing a simple graph (\emph{i.e.}\xspace, all diagonal entries are in $\{-1, 0\}$ and all other entries are in $\{0, 1\}$) is \emph{legal} if
it meets the requirements of Proposition \ref{prop:matrix-moves} and produces another such matrix.
Similarly, we say that a Move \mbox{\texttt{\textup{(Col)}}}\xspace is a legal collapse if it is applied to a regular vertex not supporting a loop, and if it takes a simple graph to another simple graph.
\begin{definition}
Fix an integer $M$ and let $E$ and $F$ be simple graphs both with finite numbers of vertices $m,n\leq M$ respectively. We say that $E$ and $F$ are \emph{elementary equivalent through simple graphs of size $M$} if either $m=n$ and one of
\begin{enumerate}[(i)]
\item $E$ is isomorphic to $F$,\label{iifirst}
\item $F$ arises from $E$ by performing a legal row addition in $\mathsf{B}^\bullet_E$,
\item $F$ arises from $E$ by performing a legal column addition in $\mathsf{B}^\bullet_E$,
\end{enumerate}
holds, or if $m=n+1$ and
\begin{enumerate}[(i)]\addtocounter{enumi}{3}
\item $F$ arises from $E$ by deleting a regular source,
\item $F$ arises from $E$ by a legal collapse.\label{iilast}
\end{enumerate}
The coarsest equivalence relation containing elementary equivalence through simple graphs of size $M$ is called \emph{$M$-inner equivalence}, and we write $E\ii{M} F$ when $E$ and $F$ are $M$-inner equivalent.
\end{definition}
The following is now clear.
\begin{proposition}\label{iioorel}
When $E$ and $F$ are finite simple graphs both with $M$ vertices or less, we have
\[
\xymatrix@C=4mm{
E\ii{M} F\ar@{=>}[r]&E\ensuremath{\sim_{M\negthinspace E}}\xspace F\ar@{=>}[d]\ar@{=>}[r]&L_{{\mathsf k}} ( E )\sim_{\text{Morita}}L_{{\mathsf k}} (F)
\ar@{=>}[r]& E\equiv_O F\\
& E\ensuremath{\sim_{C\negthinspace E}}\xspace F\ar@{=>}[rr]&&C^*(E)\otimes \ensuremath{\mathbb{K}}\xspace\cong C^*(F)\otimes \ensuremath{\mathbb{K}}\xspace\ar@{=>}[u]}
\]
\end{proposition}
Although counting the number of nonisomorphic graphs of a certain size $M$ is easy by Burnside's lemma (\emph{cf.}\xspace\ \cite{oeis} A595), producing lists of them is rather computationally demanding. The most efficient way to obtain such lists is provided by McKay and Piperno (\cite{MR3131381}).
Developing algorithms to decide $M$-inner equivalence is then straightforward by testing for elementary equivalence and partitioning the set (using, \emph{e.g.}\xspace, Warshall's algorithm) by the smallest equivalence relation containing the relations found. Drawing on methods developed in \cite{serj:ciugc} it is not much harder to design an algorithm to decide outer equivalence. At $M=4$,
it then only takes a few minutes of computing time to partition these sets of graphs into $\ii{4}$- and $\equiv_O$-classes, obtaining the numbers listed in Table~\ref{numberofiioo}. At $M=5$ we have not attempted a complete analysis, as it takes hours even to compute all the $K$-temperatures and divide the graphs into $\equiv_O$-classes.
\begin{table}
\begin{center}
\begin{tabular}{|c|||c|c|c|c|c|}\hline
$M$&1&2&3&4&5\\\hline\hline
nonisomorphic graphs&2&10&104&3044&291968\\
$\ii{M}$-classes&2&8&35&218&?\\
$\equiv_O$-classes&2&8&35&199&1310\\\hline
\end{tabular}
\end{center}
\caption{Number of classes for $M\in\{1,2,3,4,5\}$}\label{numberofiioo}
\end{table}
It follows directly from Proposition \ref{iioorel} that $\equiv_O$-classes are unions of $\ii{M}$-classes, and that when they coincide, they also coincide with $\ensuremath{\sim_{M\negthinspace E}}\xspace$-classes, $\ensuremath{\sim_{C\negthinspace E}}\xspace$-classes or stable isomorphism classes of the \mbox{$C\sp*$-}al\-ge\-bras\xspace. Consequently, the \emph{ad hoc} invariant defining outer equivalence is complete whenever the graph has 1, 2 or 3 vertices. Note that this confirms the Abrams-Tomforde conjecture in these special cases.
In the case with $M=4$ vertices, the notions differ by 17 $\equiv_O$-classes being divided into a total of 36 $\ii{4}$-classes, which we now address. We organize these classes into four groups as indicated in Figures \ref{groupI}--\ref{groupIV}, drawing one representative for each $\ii{4}$-class and indicating the boundaries of each $\equiv_O$-class by triple vertical lines. In the cases, explained below, where the graphs fail to be $\ensuremath{\sim_{C\negthinspace E}}\xspace$-equivalent we draw a vertical line between them.
\begin{figure}
\begin{center}
\begin{tabular}{c c c c c c c c}\hline
\multicolumn{1}{|||c|}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 2
\fed 1 3
\flo 2
\flo 3
\fed 4 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 2 1
\flo 2
\fed 2 3
\fed 2 4
\flo 3
\fed 4 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 3
\flo 2
\fed 2 3
\flo 3
\fed 4 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 2 1
\flo 2
\fed 2 4
\fed 3 1
\flo 3
\fed 4 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 3
\fed 2 1
\flo 2
\fed 2 4
\flo 3
\fed 4 1
\end{tikzpicture}
}
\\\hline
\multicolumn{1}{|c|||}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 2 1
\flo 2
\fed 2 4
\fed 3 2
\flo 3
\fed 4 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 4
\flo 2
\fed 2 4
\fed 3 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\fed 1 4
\fed 2 1
\flo 2
\fed 2 4
\fed 3 1
\flo 3
\end{tikzpicture}
}
&
\multicolumn{1}{c|}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 3
\fed 1 4
\fed 2 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 2
\fed 1 3
\fed 1 4
\fed 2 3
\end{tikzpicture}
}
\\\hline
\multicolumn{1}{|||c|}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 4
\fed 2 1
\flo 2
\fed 2 3
\fed 3 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 2
\fed 1 4
\fed 2 4
\fed 3 1
\flo 3
\end{tikzpicture}
}
&
\multicolumn{1}{c|}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 2
\fed 1 4
\flo 2
\fed 3 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 2 1
\flo 2
\fed 2 3
\fed 2 4
\fed 3 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 2
\fed 1 3
\fed 1 4
\fed 2 4
\flo 3
\end{tikzpicture}
}
\\\hline
\end{tabular}
\end{center}
\caption{Group I}\label{groupI}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{c c c c c c c c}\hline
\multicolumn{1}{|||c}
{
\begin{tikzpicture}
\fournodes
\fed 1 2
\fed 1 3
\fed 2 1
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 2
\fed 3 4
\fed 4 1
\end{tikzpicture}
}
&
\multicolumn{1}{c}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 2
\fed 1 3
\fed 1 4
\fed 2 1
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 2
\fed 3 4
\fed 4 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\fed 1 2
\fed 1 3
\fed 1 4
\fed 2 1
\flo 2
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 4
\fed 4 1
\fed 4 2
\end{tikzpicture}
}
&
\multicolumn{1}{c}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 2
\fed 1 3
\fed 1 4
\fed 2 1
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 4
\fed 4 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\fed 1 2
\fed 1 3
\fed 2 1
\flo 2
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 2
\fed 3 4
\fed 4 1
\end{tikzpicture}
}
\\\hline
\multicolumn{1}{|||c}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 4
\flo 2
\fed 2 4
\flo 3
\fed 3 4
\fed 4 1
\fed 4 2
\fed 4 3
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 3
\fed 1 4
\flo 2
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 2
\flo 3
\fed 4 1
\fed 4 2
\flo 4
\end{tikzpicture}
}
&
\multicolumn{1}{c}
{
\begin{tikzpicture}
\fournodes
\fed 1 2
\fed 1 3
\fed 1 4
\fed 2 1
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 2
\fed 3 4
\fed 4 1
\fed 4 2
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 2
\fed 1 3
\fed 1 4
\fed 2 1
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 2
\fed 3 4
\fed 4 1
\fed 4 2
\fed 4 3
\end{tikzpicture}
}
&
\multicolumn{1}{c}
{
\begin{tikzpicture}
\fournodes
\fed 1 2
\fed 1 3
\fed 1 4
\fed 2 1
\fed 2 3
\fed 3 1
\fed 3 4
\fed 4 1
\fed 4 2
\end{tikzpicture}
}
\\\hline
\multicolumn{1}{c|}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 2
\fed 1 3
\fed 1 4
\fed 2 1
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 4
\fed 4 1
\fed 4 2
\end{tikzpicture}
}
&
\multicolumn{1}{|||c}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 2
\fed 1 3
\fed 2 1
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 2
\fed 3 4
\fed 4 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\fed 1 2
\fed 1 4
\fed 2 1
\flo 2
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 4
\fed 4 1
\fed 4 2
\end{tikzpicture}
}
&\multicolumn{1}{c}
{
\begin{tikzpicture}
\fournodes
\fed 1 2
\fed 1 3
\fed 1 4
\fed 2 1
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 4
\fed 4 1
\fed 4 2
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 2
\fed 1 3
\fed 1 4
\fed 2 1
\fed 2 3
\fed 2 4
\fed 3 1
\fed 3 2
\fed 3 4
\fed 4 1
\fed 4 2
\end{tikzpicture}
}\\\hline
\end{tabular}
\end{center}
\caption{Group II}\label{groupII}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{c c}\\\hline
\multicolumn{1}{|||c}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 2
\fed 2 1
\fed 3 1
\fed 4 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 3
\fed 1 4
\flo 2
\flo 3
\fed 2 3
\fed 3 1
\fed 3 2
\fed 4 1
\end{tikzpicture}
}\\\hline
\end{tabular}
\end{center}
\caption{Group III}\label{groupIII}
\end{figure}
\begin{figure}
\begin{center}
\begin{tabular}{c c c c}\\\hline
\multicolumn{1}{|||c}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 3
\fed 1 4
\fed 2 1
\flo 2
\fed 3 1
\end{tikzpicture}
}
&
\multicolumn{1}{|c|||}
{
\begin{tikzpicture}
\fournodes
\fed 1 3
\fed 1 4
\fed 2 1
\flo 2
\fed 2 4
\fed 3 1
\flo 3
\end{tikzpicture}
}
&
\multicolumn{1}{c|}
{
\begin{tikzpicture}
\fournodes
\flo 1
\fed 1 3
\fed 1 4
\fed 2 1
\flo 2
\flo 3
\fed 4 1
\end{tikzpicture}
}
&
\multicolumn{1}{c|||}
{
\begin{tikzpicture}
\fournodes
\flo 1
\flo 2
\fed 2 4
\fed 3 1
\flo 3
\fed 3 4
\fed 4 1
\fed 4 2
\end{tikzpicture}
}\\\hline
\end{tabular}
\end{center}
\caption{Group IV}\label{groupIV}
\end{figure}
\begin{observation}
None of the graphs in the outer equivalence classes listed in Group I are $\ensuremath{\sim_{C\negthinspace E}}\xspace$-equivalent, and none of them give stably isomorphic \mbox{$C\sp*$-}al\-ge\-bras\xspace.
\end{observation}
\begin{proof}
Since Theorem~\ref{coldiso} applies, this follows directly by checking that no solution
to the small linear systems in \eqref{linpart} exists.
\end{proof}
In this case, the $\ii{4}$-classes coincide with the $\ensuremath{\sim_{M\negthinspace E}}\xspace$-classes as well as with the classes giving stably isomorphic graph \mbox{$C\sp*$-}al\-ge\-bras\xspace, and the invariant used to define outer equivalence fails to be complete. This is simply because the information needed to distinguish the matrices up to \ensuremath{\operatorname{SL}_\calP}\xspace-equivalence may not be reconstructed from the partial data contained in the $K_0$-group of the whole system and of the irreducible components.
\begin{observation}
All graphs in the outer equivalence classes listed in
Group II are mutually $\ensuremath{\sim_{M\negthinspace E}}\xspace$-equivalent.
\end{observation}
\begin{proof}
In every case the given graph defines an irreducible SFT, and hence by \cite{MR758893} (see also \cite{MR3082546}), since we know that the Bowen-Franks groups are the same in each outer equivalence class, we just need to check --- which is easily done --- that the signs of the determinants match up.
\end{proof}
This observation contains the result that indeed the $\equiv_O$-classes coincide with the $\ensuremath{\sim_{C\negthinspace E}}\xspace$-classes and $\ensuremath{\sim_{M\negthinspace E}}\xspace$-classes as well as the classes with stably isomorphic graph \mbox{$C\sp*$-}al\-ge\-bras\xspace.
The explanation of the lack of success of our approach to establish elementary equivalence through simple graphs is that since the graphs have so many edges, there is not room for enough row or column additions to pass from one to another. Indeed, all the graphs in each outer equivalence class turn out to be $\ii{5}$-equivalent.
\begin{observation}
The graphs in Group III are $\ensuremath{\sim_{C\negthinspace E}}\xspace$-equivalent without being $\ensuremath{\sim_{M\negthinspace E}}\xspace$-equivalent. The graphs in the outer equivalence classes listed in
Group IV fail to be $\ensuremath{\sim_{C\negthinspace E}}\xspace$-equivalent, yet produce stably isomorphic \mbox{$C\sp*$-}al\-ge\-bras\xspace.
\end{observation}
\begin{proof}
For the first claim, we see that the two graphs given are clearly move equivalent to the graph given by the matrix $(2)$ and its Cuntz splice.
For the second, we note that we get the four graphs considered in Examples \ref{notalwayssamepre}, \ref{notalwayssame}, \ref{notalwayssameii} after applying Move \mbox{\texttt{\textup{(Col)}}}\xspace\ to the unique regular vertex not supporting a loop.
\end{proof}
Combining these results, we get
\begin{observation}
The 3044 different simple graphs with four vertices are divided into 210 different $\ensuremath{\sim_{M\negthinspace E}}\xspace$-classes and 209 different $\ensuremath{\sim_{C\negthinspace E}}\xspace$-classes. They define a total of 207 different graph \mbox{$C\sp*$-}al\-ge\-bras\xspace, identified up to stable isomorphism.
\end{observation}
The number of different Leavitt path algebras (say with $\mathsf k=\ensuremath{\mathbb{C}}\xspace$) defined, identified up to Morita equivalence, is not known, but must be in the range $\{207,208,209,210\}$ since for all the graphs giving isomorphic stabilized $C^*$-algebras except the ones in Group III and IV we have established $\ensuremath{\sim_{M\negthinspace E}}\xspace$, which implies Morita equivalence of the Levitt path algebras as well.
\section*{Acknowledgements}
This work was partially supported by the Danish National Research Foundation through the Centre for Symmetry and Deformation (DNRF92), by VILLUM FONDEN through the network for Experimental Mathematics in Number Theory, Operator Algebras, and Topology, by a grant from the Simons Foundation (\# 279369 to Efren Ruiz), and by the Danish Council for Independent Research | Natural Sciences.
The third and fourth named authors would like to thank the School
of Mathematics and Applied Statistics at the University of Wollongong
for hospitality during their visit there, and the first and second named authors likewise thank the Department of
Mathematics, University of Hawaii, Hilo. The initial work was carried out at these two long-term visits, and
it was completed while
all four authors were attending the research program
\emph{Classification of operator algebras: complexity, rigidity, and
dynamics} at the Mittag-Leffler Institute, January--April 2016. We
thank the institute and its staff for the excellent work conditions
provided.
The authors would also like to thank Mike Boyle and James Gabe for many fruitful discussions.
\newcommand{\etalchar}[1]{$^{#1}$}
\providecommand{\bysame}{\leavevmode\hbox to3em{\hrulefill}\thinspace}
\providecommand{\MR}{\relax\ifhmode\unskip\space\fi MR }
\providecommand{\MRhref}[2]{%
\href{http://www.ams.org/mathscinet-getitem?mr=#1}{#2}
}
\providecommand{\href}[2]{#2}
|
\section{Introduction}
The decade of 1970's will be remembered as the one in which Solar Physicist were able to really start to infer the
magnetic and thermodynamic properties of the solar plasma from the observations. At that specific time, there was a
sweet coincidence. On the one hand, the theory of radiative transfer for polarized light was already in its maturity. On
the other, computers started to be available for researchers in general and powerful enough to carry out complex calculations.
It was then the time at which the ideas of non-linear inversion codes were set \citep{harvey72,auer_heasly_house77,skumanich_lites87}.
Inversion algorithms are able to extract information about the magnetic and thermodynamical properties of the solar plasma from the analysis of spectropolarimetric observations.
They function by proposing a specific model to explain the observations and then defining a merit function (usually the $\chi^2$ function,
valid under the presence of uncorrelated Gaussian noise). The model parameters are iteratively modified for optimizing the merit
function. The first inversion codes were relatively simple and based on strongly simplifying assumptions, like the
Milne-Eddington (ME) approximation to analytically solve the radiative transfer equation \citep[e.g.,][]{harvey72,auer_heasly_house77,landi_landolfi04}.
Such inversion codes are still used today, like VFISV \citep{borrero07,borrero_vfisv10}, MILOS \citep{orozco_hinode07} or
MERLIN \citep{skumanich_lites87,lites07}.
An enormous step forward was introduced by \cite{sir92}, who developed SIR (Stokes Inversion based on Response functions), an inversion
code that recovers the optical depth stratification of the physical quantities (temperature, magnetic field, velocity, etc.)
from the interpretation of the Stokes profiles. These codes are based on the idea of response function \citep{landi_response77}, that
allows the user to link the perturbation in the emergent Stokes parameters with perturbations in the physical parameters. One of the
key ingredients that facilitated the development of such inversion codes was the possibility to find an analytical
expression for the response functions in local thermodynamic equilibrium (LTE) \citep{sanchez_almeida92,ruizcobo_deltoro94}.
Based on the seed of SIR, several non-linear inversion codes are now available, some of them even dealing with the
much more difficult case of the inversion of spectral lines in non-LTE \citep[NICOLE;][]{socas_trujillo_ruiz00,socas_nicole14}.
All these 1D inversion codes are based on the concept of ``nodes'', that need to be defined a-priori. These nodes mark positions
along the optical depth axis where the value of the physical parameters will be modified to fit the Stokes profiles. The
full stratification of the atmosphere between the nodes, which is needed
to accurately integrate the radiative transfer equation and to derive the gas pressure scale, is interpolated using a piece-wise polynomial.
The complexity of the solution then critically
depends on the number of nodes that are employed to describe each of the physical parameters.
After more than a decade without any fundamental improvement \citep[except perhaps the introduction of Bayesian inference
into the field;][]{asensio_martinez_rubino07,asensio_hinode09,asensioramos_modcomp12}, we are nowadays living another
sweet era, again driven by the improvements in computational power. On one side, \cite{vannoort12} has developed a spatially
coupled two-dimensional inversion code in which the effect of the telescope point spread function (PSF) is taken into account.
The PSF couples nearby pixels so that deconvolution and inversion is done at the same time. Following a similar motivation,
\cite{ruizcobo_asensioramos13} have used a regularized deconvolution of the Stokes profiles based on the principal
component analysis (PCA) and have used SIR to invert the deconvolved Stokes parameters.
\begin{figure*}
\includegraphics[width=\textwidth]{examples.pdf}
\caption{Examples of situations in which systematic effects are important when extracting information from
spectral lines.}
\label{fig:examples}
\end{figure*}
Arguably, the last step in the evolution of inversion codes has been the introduction of regularization ideas based on the
concept of sparsity or compressibility\footnote{Given an $n$-dimensional vector $\mathbf{x}$, sparsity means that the majority of the elements of the vector are
strictly zero, while compressibility means that the elements of such vector, when ordered by absolute value, fulfill $|x_i| \leq C i^{-1/r}$.}
\cite[e.g.,][]{starck10}. \cite{asensio_delacruz15} have presented an inversion code
for the inversion of Stokes profiles that uses $\ell_0$ or $\ell_1$ regularization on a transformed space\footnote{The $\ell_p$-norm
is given by: $\Vert \mathbf{x} \Vert_p = \left( \sum_i |x_i|^p \right) ^{1/p}$, with $p\geq0$.}. The two-dimensional
maps of parameters are linearly transformed (Fourier, wavelet or any other appropriate transformation can be used) and assumed to
be sparse in the transformed domain. This introduces two important constraints. First, the sparsity assumption reduces the number
of unknowns in the problem, avoiding overfitting. Second, the global character of the transformations that are routinely used, spatial
correlation of the result is automatically taken into account. Consequently, the Stokes parameters observed at every pixel potentially
introduces constraints onto every other pixel of the observed map. This is the first time that the inherent spatial correlation of
the physical parameters has been taken into account in inversion codes. This results into much more stable inversion codes that
do not produce spurious pixel-by-pixel variations of the maps of physical quantities. Additionally, the complexity of the solution
is automatically adapted, producing more structure where it is needed.
In the conclusions of \cite{asensio_delacruz15} we pointed out that the sparsity regularization can be applied to invert
Stokes profiles with systematic effects. This is precisely what we present in this paper.
A customary way of dealing with systematic effects in current inversions is to downweight the influence
on the merit function of the parts of the spectrum that are affected by these effects. Although it works in
practice, it depends on a set of parameters (e.g., region and factor of the downweighting) that make it
quite subjective. Instead, we propose several basis sets
(orthogonal and non-orthogonal) to absorb the systematic effects (for instance, telluric lines) in the Stokes profiles and
use a proximal projection algorithm \citep[e.g.,][]{parikh_boyd14} to make the solution automatically adapt to the necessary complexity.
We will show that the modifications needed in existing inversion codes are minimal so that
this approach can be introduced without much effort.
\section{Sparsity regularization}
Our objective is to fit the observed Stokes parameters, that are discretized at $N$ finite wavelength
points $\lambda_j$. For simplicity, we stack the four Stokes parameters in a long vector
of length $4N$ so that
$\mathbf{O}=\left[I(\lambda_1),\ldots,I(\lambda_N),Q(\lambda_1),\ldots,U(\lambda_1),\ldots,V(\lambda_1),\ldots,V(\lambda_N)\right]^\dag$
(with $\dag$ the transpose). The fit is carried out using a model atmosphere, with the aim of extracting useful
thermal, dynamic and magnetic information
from them. Additionally, we make the assumption that these Stokes parameters are perturbed by some uncontrollable
systematic effects. These systematic effects can be, for instance, telluric lines produced by absorption
in the Earth atmosphere, variations along the spectral direction produced by an incorrect illumination of the camera or
a low-quality flatfielding, etc. Examples of these situations are found in Fig. \ref{fig:examples}. The left panel shows the
well-known region around 6302 \AA\, that contains two \ion{Fe}{i} lines, which are blended with two telluric lines. This
is of special importance for the Fe \textsc{i} line at 6302.5 \AA. Extracting physical information from the line
requires then to avoid the far wings, which can lead to problems in the deepest parts of the atmosphere. The
right panel shows the region around 10830 \AA, containing a \ion{He}{i} multiplet used for chromospheric diagnostics, which
is blended with a photospheric \ion{Si}{i} line and a telluric line. The interpretation of the He \textsc{i} multiplet, given
that it forms on the extended wings of the Si \textsc{i} line, usually requires a previous analysis of the photospheric line.
Additionally, the strong chromospheric dynamics induces that the He \textsc{i} multiplet sometimes blends with
the telluric absorption.
In very general terms, we can explain the observations with the following generative model:
\begin{equation}
\mathbf{O} = \mathbf{S}_\mathrm{atm}(\mathbf{q}) + \mathbf{D} \mbox{\boldmath$\alpha$} + \mbox{\boldmath$\epsilon$},
\label{eq:generative}
\end{equation}
where $\mbox{\boldmath$\alpha$}$ is a vector of dimension $d$ that describes the systematic effects via a dictionary
$\mathbf{D} \in \mathbb{R}^{4N \times d}$, also known as synthesis operator. We remind that a dictionary is just a collection
of (possibly non-orthogonal) functions that are used to describe the systematic effects. The contribution
$\mathbf{S}_\mathrm{atm}(\mathbf{q})$ contains the Stokes parameters that emerge from a model
atmosphere depending on a vector of model parameters $\mathbf{q}$, ordered like in $\mathbf{O}$.
The noise components $\mbox{\boldmath$\epsilon$}$ are considered to be Gaussian random variables with zero mean and
diagonal covariance matrix.
Note that when the size of the dictionary equals the number of observed spectral points
($d=4N$) and $\mathbf{D}= \mathbf{1}$ (with $\mathbf{1}$ the identity matrix), the model for the systematic effects described in
Eq. (\ref{eq:generative}) turns out to be very flexible because $\mbox{\boldmath$\alpha$}$ refers to the particular value of the systematic effects at each
sampled wavelength point.
From this generative model, the merit function one has to optimize is the $\chi^2$ with diagonal covariance, in our case given by:
\begin{equation}
\chi^2_\mathbf{D}(\mathbf{q},\mbox{\boldmath$\alpha$}) = \frac{1}{4N} \sum_{j=1}^{4N} w_j
\frac{\left[S_{\mathrm{atm},j}(\mathbf{q})+ (\mathbf{D} \mbox{\boldmath$\alpha$})_{j}-O_j\right]^2}{\sigma_{j}^2},
\label{eq:chi2_q}
\end{equation}
where we make explicit the dependence of the merit function on the election of the dictionary.
The previous merit function considers a potentially different noise variance $\sigma_{j}^2$ for every
Stokes parameter and wavelength position, which is surely the case for very strong lines in which the number of photons
in the core is much more absorbed than the wings. Additionally, it is customary to introduce weights $w_j$ for each Stokes
parameter to increase the sensitivity to some parameters when carrying out the inversion\footnote{Note that the role
of the weights is to modify the value of the noise variance. If the noise variance is artificially increased for one Stokes
parameter, its constraining power decreases.}. These weights are precisely the ones that are currently used to
deal with systematic effects, by decreasing their values for specific wavelength points.
The inferred physical parameters are typically found by solving the following maximum-likelihood problem:
\begin{equation}
\mathop{\mathrm{argmin}}\limits_\mathbf{q,\mbox{\boldmath$\alpha$}} \, \chi^2_\mathbf{D}(\mathbf{q},\mbox{\boldmath$\alpha$}),
\label{eq:problem_l2}
\end{equation}
where the operator ``$\mathop{\mathrm{argmin}}\limits$'' returns the value of $\mathbf{q}$ and $\mbox{\boldmath$\alpha$}$ that minimizes the
$\chi^2_\mathbf{D}(\mathbf{q},\mbox{\boldmath$\alpha$})$ function. Note that Eq. (\ref{eq:problem_l2}) is, in general, non-convex\footnote{When the
function to minimize is convex in the variables, a local minimum is also the global minimum. This cannot be
guaranteed in the optimization of non-convex functions, which is typically the case in the inversion
of Stokes parameters} for the inversion of Stokes parameters.
Therefore, we only aspire to reach one of the local minima and later check for the physical relevance of the
solution. Problem (\ref{eq:problem_l2}) is usually solved by direct application of the Levenberg-Marquardt (LM) algorithm
\citep{levenberg44,marquardt63}, which is
especially suited to the optimization of such $\ell_2$-norms. Although problem (\ref{eq:problem_l2}) is stated without any constraint,
it is clear from the introduction that it is customary to regularize the solution by imposing some regularity of the solution through
the use of nodes. Consequently, the vector of parameters $\mathbf{q}$ contains the value of the physical parameters (temperature, magnetic field, velocity, \ldots)
at a small number of depths in optical depth in the atmosphere. In between these points, the physical properties are
interpolated from the values at these nodes.
Without any additional constraint for $\mbox{\boldmath$\alpha$}$, it is sure that we will encounter cross-talk between
$\mbox{\boldmath$\alpha$}$ and $\mathbf{q}$. The fundamental reason is that our regressor for the systematic effects is so flexible that
it can potentially fit the observations perfectly, including the specific noise realization. As noted above,
if $d=4N$ and $\mathbf{D}=\mathbf{1}$, a solution
to Eq. (\ref{eq:problem_l2}) is $\mbox{\boldmath$\alpha$}=\mathbf{O}$ and $\mathbf{S}_\mathrm{atm}(\mathbf{q})=0$
for every observed wavelength. This trivial solution is of no interest because it does not extract any
relevant physical information from the observations. In order to overcome this possibility, we follow recent
ideas \citep{candes06,starck10,asensio_delacruz15} and regularize the problem by imposing a
sparsity constraint on $\mbox{\boldmath$\alpha$}$ or, in general, in any transformation of $\mbox{\boldmath$\alpha$}$.
Two fundamental approaches to include this sparsity penalty have been proposed \citep{starck10}. Each
one has advantages and disadvantages, as we will show in Sec. \ref{sec:examples}.
\subsection{The analysis penalty approach}
This approach is based on having as many degrees of freedom as observed data points ($d=4N$) and $\mathbf{D}=\mathbf{1}$ and solving the
following problem:
\begin{equation}
\mathop{\mathrm{argmin}}\limits_\mathbf{q,\mbox{\boldmath$\alpha$}} \, \chi^2_\mathbf{1}(\mathbf{q},\mbox{\boldmath$\alpha$}), \,\,
\mathrm{subject\, to\,} \Vert \mathbf{W} \mbox{\boldmath$\alpha$} \Vert_p \leq s,
\label{eq:analysis_problem0}
\end{equation}
where $\mathbf{W}$ is the matrix associated to any linear transformation of interest, either orthogonal or not, while
$s$ is a predefined threshold. We also remind the reader that $\Vert \mathbf{x} \Vert_p = \left( \sum_i |x_i|^p \right) ^{1/p}$
is the $\ell_q$-norm. For instance, the $\ell_0$ norm of a vector is just the number of non-zero elements, while
the $\ell_1$ norm is the sum of the absolute value of its elements.
Put in words, solving Eq. (\ref{eq:analysis_problem0}) requires to seek the pair $(\mathbf{q},\mbox{\boldmath$\alpha$})$
that better fits our observations, imposing that
the projection of the systematic effects on the transformed domain defined by $\mathbf{W}$ is sparse. The
name \emph{analysis penalty} comes from the fact that $\mathbf{W}$ is the analysis operator, that
carries the vector $\mbox{\boldmath$\alpha$}$ to the transformed sparsity-inducing domain \citep{elad07,starck10}. It
is especially suited to deal with cases in which $\mathbf{W}$ is non-orthogonal and/or overcomplete.
The solution to the previous problem when $p=0$
\citep[or equivalently when $p=1$ under some conditions;][]{candes06,donoho06} is known to
coincide with the exact solution when it exists. Sparsity is, therefore, a very convenient
regularization. The only degree of freedom is to find the appropriate transformation $\mathbf{W}$.
The same problem can be equivalently written with the addition of a regularization parameter ($\lambda$):
\begin{equation}
\mathop{\mathrm{argmin}}\limits_\mathbf{q,\mbox{\boldmath$\alpha$}} \, \left[ \frac{1}{4N} \sum_{j=1}^{4N} w_j
\frac{\left[S_{\mathrm{atm},j}(\mathbf{q})+ \alpha_{j}-O_j\right]^2}{\sigma_{j}^2} + \lambda \Vert \mathbf{W} \mbox{\boldmath$\alpha$} \Vert_p \right].
\label{eq:problem_l0}
\end{equation}
\subsection{The synthesis penalty approach}
This approach is based on having $\mathbf{D}=\mathbf{W}^T$, and imposing
a sparsity constraint on the $\mbox{\boldmath$\alpha$}$ vector itself. Consequently, we have to solve
the problem
\begin{equation}
\mathop{\mathrm{argmin}}\limits_\mathbf{q,\mbox{\boldmath$\alpha$}} \, \chi^2_\mathbf{W^T}(\mathbf{q},\mbox{\boldmath$\alpha$}), \,\,
\mathrm{subject\, to\,} \Vert \mbox{\boldmath$\alpha$} \Vert_p \leq s,
\end{equation}
which, in lagrangian form, becomes:
\begin{equation}
\mathop{\mathrm{argmin}}\limits_\mathbf{q,\mbox{\boldmath$\alpha$}} \, \left[ \frac{1}{4N} \sum_{j=1}^{4N} w_j
\frac{\left[S_{\mathrm{atm},j}(\mathbf{q})+ (\mathbf{W}^T \mbox{\boldmath$\alpha$})_{j}-O_j\right]^2}{\sigma_{j}^2} + \lambda \Vert \mbox{\boldmath$\alpha$} \Vert_p \right],
\label{eq:problem_l0_synthesis}
\end{equation}
The name \emph{synthesis penalty} comes from the fact that $\mathbf{W}^T$ is the synthesis operator, that
generates the systematic effects from the vector $\mbox{\boldmath$\alpha$}$ living on the sparsity-inducing transformed domain.
It was noted by \cite{elad07} that both approaches are equivalent when $\mathbf{W}$ is an
orthogonal transform (like Fourier, wavelet, \ldots), because it fulfills $\mathbf{W}^T \mathbf{W}=\mathbf{1}$. However, they solve completely different
problems when $\mathbf{W}$ is non-orthogonal.
\subsection{Transformations}
We will consider in this paper three options for the regularization term.
The first one uses as regularization an orthogonal wavelet transform. The fact that
$\mathbf{W}$ is orthogonal will slightly simplify the algorithms. The second one is a
non-orthogonal overcomplete isotropic undecimated wavelet transform using the B$_3$-spline \citep{starck10}.
Finally, we will use a hand-made non-orthogonal transform made of Voigt functions centered
at specific locations in the spectrum, that will be used to absorb the systematic effects. We defer the detailed
description of each one until Sec. \ref{sec:proximal}.
\section{The proposed optimization}
Given the special structure of both problems defined in Eqs. (\ref{eq:problem_l0}) and (\ref{eq:problem_l0_synthesis}),
in which the regularization only occurs for the $\mbox{\boldmath$\alpha$}$ variables,
we propose to use an alternating optimization method. If $\mbox{\boldmath$\alpha$}$ is fixed, Eq. (\ref{eq:problem_l0})
becomes the traditional least-squares problem for $\mathbf{q}$, that is solved efficiently with Newton-type
methods like the Levenberg-Marquardt algorithm. This method uses second-order information given by $\mathbf{H}_q$, the
Hessian of the merit function with respect to the $\mathbf{q}$ variables:
\begin{equation}
\mathbf{q}_{i+1} = \mathbf{q}_i - \hat{\mathbf{H}}_q^{-1} \nabla_q \chi^2_\mathbf{1}(\mathbf{q}_i,\mbox{\boldmath$\alpha$}_i) \qquad \,\,\,
\qquad \text{with $\mbox{\boldmath$\alpha$}_i$ fixed}.
\label{eq:q_update}
\end{equation}
Conforming to the prescriptions of the Levenberg-Marquardt algorithm, we
use a modified Hessian matrix by enhancing its diagonal by a factor $\beta$, so
that $\hat{\mathbf{H}}_q=\mathbf{H}_q+ \beta \mathrm{diag}(\mathbf{H}_q)$.
This hyperparameter is modified during the
iteration to shift between the gradient descent method (large $\beta$) and Newton-type method (small $\beta$).
We note that inverting the Hessian matrix with a truncated singular value decomposition \citep{sir92}
introduces an extra regularization on the physical parameters at the nodes that is often needed.
\subsection{The analysis penalty}\label{sec:prior}
On the other hand, when $\mathbf{q}$ is fixed, Eq. (\ref{eq:problem_l0}) becomes a
standard sparsity-constrained linear problem for $\mbox{\boldmath$\alpha$}$.
This problem can be solved efficiently using proximal algorithms \citep{parikh_boyd14,asensio_delacruz15}, which
are especially suited to solve problems of the type
\begin{equation}
\mathop{\mathrm{argmin}}\limits_{\mbox{\boldmath$\alpha$}} f(\mbox{\boldmath$\alpha$}) := g(\mbox{\boldmath$\alpha$}) + h(\mbox{\boldmath$\alpha$}),
\label{eq:problem_general}
\end{equation}
where $g(\mbox{\boldmath$\alpha$})=\chi^2_\mathbf{1}(\mathbf{q},\mbox{\boldmath$\alpha$})$, is a smooth function and
$h(\mbox{\boldmath$\alpha$})=\lambda \Vert \mathbf{W}\mbox{\boldmath$\alpha$} \Vert_p$ is a convex
but not necessarily smooth function (note that the derivative of the $\Vert \mathbf{W} \mbox{\boldmath$\alpha$} \Vert_p$ term is not
continuous). We propose the following first-order iterative scheme to solve the problem:
\begin{equation}
\mbox{\boldmath$\alpha$}_{i+1} = \mathrm{prox}_{p,\lambda,\mathbf{W}} \left[\mbox{\boldmath$\alpha$}_i -
\tau \nabla_{\alpha} \chi^2_\mathbf{1}(\mathbf{q}_{i+1},\mbox{\boldmath$\alpha$}_i) \right]
\quad \text{with $\mathbf{q}_{i+1}$ fixed},
\label{eq:iteration_alpha1}
\end{equation}
where the operator $\mathrm{prox}_{p,\lambda,\mathbf{W}}$ is the proximal projection operator \citep{parikh_boyd14}
associated with the constraint $\Vert \mathbf{W} \mbox{\boldmath$\alpha$} \Vert_p$, that we show how to efficiently compute in
Sec. \ref{sec:examples}. The election of the step size $\tau$ is important for the convergence
of the algorithm. It is known to converge provided the step size fulfills $\tau < 2 / \Vert \mathbf{H}_\alpha \Vert^2$, where
$\Vert \mathbf{H}_\alpha \Vert$ is the spectral norm of the Hessian of the merit function with respect to $\mbox{\boldmath$\alpha$}$ (given by the
square root of the maximum eigenvalue of $\mathbf{H}_\alpha^T \mathbf{H}_\alpha$). We note that faster algorithms like
FISTA \citep{beck_teboulle09,asensio_delacruz15} can also be used.
Given the simple dependence of $\chi^2_{\mathbf{q},\mbox{\boldmath$\alpha$}}$ on $\mbox{\boldmath$\alpha$}$,
it can be proven \citep[e.g.,][]{starck10} that Eq. (\ref{eq:iteration_alpha1}) can be simplified to finally obtain:
\begin{equation}
\mbox{\boldmath$\alpha$}_{i+1} = \mathrm{prox}_{p,\lambda,\mathbf{W}} \left[ \mathbf{S}_{\mathrm{atm}}(\mathbf{q}_{i+1})-\mathbf{O} \right]
\quad \text{with $\mathbf{q}_{i+1}$ fixed}.
\label{eq:iteration_alpha}
\end{equation}
In other words, the estimation of the systematic effects for a new iteration is very simple and reduces to
computing the proximal projection of the residual between the observed Stokes profiles and the current modeled ones. We think
that this approach gives a very transparent and intuitive understanding of what the algorithm is doing. We note that the best results have been
found applying Eq. (\ref{eq:iteration_alpha}) every $\sim 3$ iterations of the LM algorithm.
\subsection{The synthesis penalty}
When $\mathbf{q}$ is fixed, Eq. (\ref{eq:problem_l0_synthesis}) becomes again a sparsity-constrained problem
for $\mbox{\boldmath$\alpha$}$. In this case, the sparsity penalty $h(\mbox{\boldmath$\alpha$})=\Vert \mbox{\boldmath$\alpha$} \Vert_p$ is much simpler, but the function
$g(\mbox{\boldmath$\alpha$})=\chi^2_{\mathbf{W}^T}(\mathbf{q},\mbox{\boldmath$\alpha$})$ is
more complex because of the presence of the synthesis operator $\mathbf{W}^T$. We propose to solve
the problem using the following first-order iterative scheme:
\begin{equation}
\mbox{\boldmath$\alpha$}_{i+1} = \mathrm{prox}_{p,\lambda,\tau}
\left[\mbox{\boldmath$\alpha$}_i - \tau \nabla_{\alpha} \chi^2_\mathbf{W^T}(\mathbf{q}_{i+1},\mbox{\boldmath$\alpha$}_i) \right]
\quad \text{with $\mathbf{q}_{i+1}$ fixed},
\label{eq:iteration_alpha2}
\end{equation}
where $\tau < 2 / \Vert \mathbf{H}_\alpha \Vert^2$, which can be computed from the spectral norm of the transformation
matrix $\mathbf{W}$, and the proximal operator $\mathrm{prox}_{p,\lambda,\tau}$ is described in the following. We point
out that a more complex second-order iterative scheme is discussed in the Appendix.
\subsection{Computing the proximal operators}
\label{sec:proximal}
The previous iterative schemes rely on the existence of algorithms for the computation
of the proximal projection operators $\mathrm{prox}_{p,\lambda,\tau}(\mbox{\boldmath$\alpha$})$ and $\mathrm{prox}_{p,\lambda,\mathbf{W}}(\mbox{\boldmath$\alpha$})$.
\subsubsection{Computing $\mathrm{prox}_{p,\lambda,\tau}(\mbox{\boldmath$\alpha$})$}
This operator is very simple to compute for
typical choices of $p$ \citep{parikh_boyd14}. Useful cases of the regularization term are the $\ell_0$-norm ($p=0$) or
the $\ell_1$-norm ($p=1$). In the case of the $\ell_0$-norm, the proximal
operator reduces to the hard-thresholding operator, which is trivially given by
\begin{equation}
\mathrm{prox}_{p=0,\lambda,\tau} (\mbox{\boldmath$\alpha$}) =
\begin{cases}
\mbox{\boldmath$\alpha$} & \quad |\mbox{\boldmath$\alpha$}| > \tau \lambda \\
\mathbf{0} & \quad \mathrm{otherwise}.
\end{cases}
\label{eq:proximal_l0}
\end{equation}
where $\tau$ is the step-size defined in \S\ref{sec:prior} and $\lambda$ is the regularization parameter that we introduced in Eq.~\ref{eq:problem_l0}.
For the $\ell_1$-norm, it reduces to the soft-thresholding operator, which is given by
\begin{equation}
\mathrm{prox}_{p=1\lambda,\tau}(\mbox{\boldmath$\alpha$}) = \mathrm{sign}(\mbox{\boldmath$\alpha$}) (|\mbox{\boldmath$\alpha$}|-\tau \lambda)_+,
\label{eq:proximal_l1}
\end{equation}
where $(\cdot)_+$ denotes the positive part. Other proximal operators with analytical expressions
can be found in \cite{parikh_boyd14}.
\subsubsection{Computing $\mathrm{prox}_{p,\lambda,\mathbf{W}}(\mbox{\boldmath$\alpha$})$}
The solution of Eq. (\ref{eq:iteration_alpha}) for any $\mathbf{W}$, either orthogonal or not, is slightly
more complicated and requires some elements of proximal calculus \citep{parikh_boyd14}.
A general algorithm for the solution of the proximal projection of Eq. (\ref{eq:iteration_alpha})
has been developed by \cite{fadili09}, that we reproduce in Alg. \ref{alg:proximal} for completeness. We note that
the algorithm is just a simple iterative scheme that has been proven to converge to the
solution provided the step size $\tau < 2 / \Vert \mathbf{W} \Vert^2$.
\begin{algorithm}[!t]
\KwData{$\mathbf{W}$ and $\mbox{\boldmath$\alpha$}$}
\KwResult{$\mathrm{prox}_{p,\lambda,\mathbf{W}}(\mbox{\boldmath$\alpha$})$}
Initialization: $\mathbf{y}_1=\mbox{\boldmath$\alpha$}$ and suitable step $\tau < 2 / \Vert \mathbf{W} \Vert^2$\;
\While{not converged}{
1. $r=\mbox{\boldmath$\alpha$} - \mathbf{W} \mathbf{y}_{i}$ \;
2. $\mathbf{s} = \tau \mathbf{y}_i + \mathbf{W}^T \mathbf{r}$ \;
3. $\mathbf{y}_{i+1} = \mbox{\boldmath$\alpha$} - \tau^{-1} \mathbf{W} \left[ \mathbf{s} - \mathrm{prox}_{p,\lambda,\tau}(\mathbf{s}) \right]$ \;
}
\Return $\mathbf{y}_i$
\caption{Algorithm for proximal projection, extracted from \cite{fadili09}.}
\label{alg:proximal}
\end{algorithm}
It is interesting to point out that, when $\mathbf{W}$ is an orthogonal transform, Alg. \ref{alg:proximal} hugely simplifies and
the solution to Eq. (\ref{eq:iteration_alpha}) can be obtained with \citep{starck10}:
\begin{equation}
\mbox{\boldmath$\alpha$} = \mathbf{W}^T \left[ \mathrm{prox}_{p,\lambda,\tau} \left( \mathbf{W} \mbox{\boldmath$\alpha$} \right) \right].
\label{eq:proximal_wavelet}
\end{equation}
In other words, the vector $\mbox{\boldmath$\alpha$}$ is first transformed, it is then thresholded using the appropriate
proximal operator, and finally transformed back to the original domain. If the transformation is
unitary, so that $\mathbf{W}^T \mathbf{W}=\mathbf{1}$, the result
of the application of Eq. (\ref{eq:proximal_wavelet}) leaves the value of $\mbox{\boldmath$\alpha$}$ unchanged for $\lambda=0$.
Another interesting use of Eq. (\ref{eq:proximal_wavelet}) is that it is equivalent to the first iteration
of Alg. \ref{alg:proximal}. We carry out experiments in the following sections to verify if using this simple
approximation in the general case of non-orthogonal $\mathbf{W}$ gives good results.
\subsection{Selection of $\lambda$}
As we show in the next section, $\lambda$ has a strong impact on the sparsity of the final solution. When $\lambda$ is too small,
overfitting clearly appears. On the contrary, if $\lambda$ is large, the fit is typically of bad quality. Therefore,
the selection of $\lambda$
requires some fine-tuning, but it is possible to have an order of magnitude estimation using very simple arguments.
As noted in Eqs. (\ref{eq:proximal_l0}) and (\ref{eq:proximal_l1}), the thresholding happens for values of the
$\mbox{\boldmath$\alpha$}$ parameters larger than $\tau \lambda$. Therefore, if we want to avoid values of $\mbox{\boldmath$\alpha$}$ smaller than $\eta$,
then $\lambda \sim \eta \Vert \mathbf{W} \Vert^2$. For the orthogonal unitary transforms, we find that
the spectral norm is $\Vert \mathbf{W} \Vert=1$.
\begin{algorithm}[!t]
\KwData{Stokes profiles, model atmosphere, transform $\mathbf{W}$ and $k$.}
\KwResult{Regularized solution}
Initialization: $\mathbf{q}_0$ and $\mbox{\boldmath$\alpha$}_0=0$, first estimation of solution\;
\While{not converged}{
1. Compute gradient $\nabla_q \chi^2_\mathbf{1}(\mathbf{q}_i,\mbox{\boldmath$\alpha$}_i)$ and Hessian matrices $\mathbf{H}_q$\;
2. Modify Hessians: $\hat{\mathbf{H}}_q=\mathbf{H}_q+ \beta \mathrm{diag}(\mathbf{H}_q)$\;
3. Update $\mathbf{q}$ : $\mathbf{q}_{i+1} = \mathbf{q}_{i} - \hat{\mathbf{H}}_q^{-1} \nabla_q \chi^2_{\mathbf{1}}(\mathbf{q}_i,\mbox{\boldmath$\alpha$}_i)$ \;
4. Update $\mbox{\boldmath$\alpha$}$: $\mbox{\boldmath$\alpha$}_{i+1} = \mathrm{prox}_{p,\lambda,\mathbf{W}}
\left[ \mathbf{S}_{\mathrm{atm}}(\mathbf{q}_{i+1})-\mathbf{O} \right]$ using Alg. \ref{alg:proximal} every $k$ iterations.
}
\Return $\mathbf{q}$, $\mbox{\boldmath$\alpha$}$
\caption{Proximal Levenberg-Marquardt algorithm for analysis penalty.}
\label{alg:proximal_analysis}
\end{algorithm}
\subsection{Summary}
The full step-by-step algorithms are described in Algs. \ref{alg:proximal_analysis} and \ref{alg:proximal_synthesis},
together with Alg. \ref{alg:proximal} for the application of the proximal operator
described in step 4 of Alg. \ref{alg:proximal_analysis}. We want to clarify to any potential user of this method that
the only real difference between current inversion codes for the Stokes parameters and our approach
is in point 4 of Alg. \ref{alg:proximal_analysis}, together with the necessity to include the systematic effects in the
calculation of the gradient and Hessian with respect to the $\mathbf{q}$ variables. This step is very
easy to carry out and can be implemented in any existing inversion code with only a few lines of programming.
\section{Examples}
\label{sec:examples}
In this section we demonstrate the capabilities of Algs. \ref{alg:proximal_analysis} and \ref{alg:proximal_synthesis} in the inversion of
Stokes profiles using several transformation matrices $\mathbf{W}$ and two datasets.
For simplicity and for the purpose of clarity, we only focus on Stokes $I$, for which the
systematic effects are usually more important. Throughout this section, we use the $\ell_0$-norm
as regularization, so we apply the proximal operator of Eq. (\ref{eq:proximal_l0}).
The codes used in this paper can be obtained from \texttt{https://github.com/aasensio/proxStokesSystematics}.
\begin{figure*}
\includegraphics[width=\textwidth]{hazel_wavelet.pdf}
\caption{Left panels: observed (blue) and fitted (green) Stokes $I$ profile for different values of the
regularization parameters $\lambda$ and using an orthogonal wavelet transform as the sparsity-inducing
transformation. The dashed red curve shows the inferred systematic effects. The spectral range corresponds to that around
the He \textsc{i} multiplet at 10830 \AA. Right panels: first 100 active wavelet coefficients of the set of
512 total coefficients. The percentage of active functions is shown in each panel.}
\label{fig:wavelet}
\end{figure*}
\subsection{Observations}
To show examples of application, we choose two spectral regions of special interest that suffer from
the problems described in this paper. The first one is the region around 10830 \AA, which
contains the He \textsc{i} multiplet at 10830 \AA\ and is displayed in the
right panel of Fig. \ref{fig:examples}. This multiplet is
used for diagnosing the magnetic properties of chromospheric
material. Due to the potentially large Doppler shifts in the chromosphere \citep[e.g.,][]{lagg07}, it is interesting
to apply our approach to this multiplet.
The observations that we analyze have been obtained with the Vacuum Tower Telescope (VTT) on the
Observatorio del Teide with the TIP-II instrument \citep{collados_tipII07}. The profiles have been
extracted from a plage region close to a pore. The He \textsc{i} multiplet is
synthesized with the help of \textsc{Hazel}\ \citep{asensio_trujillo_hazel08}, which gives the $\mathbf{S}_\mathrm{atm}(\mathbf{q})$ part of Eq. (\ref{eq:generative}).
The spectral lines are characterized by the optical depth on the red component of the multiplet $\Delta \tau$,
the Doppler broadening of the line $v_\mathrm{th}$ and the bulk velocity of the plasma $v$.
The second region is the well-known region around 6301-6302 \AA\ that contains
an Fe \textsc{i} doublet. This region also contains two telluric absorptions. The Stokes profiles,
shown in the left panel of Fig. \ref{fig:examples}, have been obtained from the ground with the
POlarimetric LIttrow Spectrograph \citep[POLIS;][]{beck_polis05} and belong to the observations
of the quiet Sun analyzed by \cite{marian08}.
To fit the profiles, we choose an atmosphere in local thermodynamic equilibrium and we infer the depth
stratification of the temperature, velocity along the line-of-sight and microturbulent
velocity. For this experiment we fix the number of nodes of the parameterization: 5 nodes for
the temperature, 3 for the bulk velocity and 1 for the microturbulent velocity. Like virtually
all inversion codes with depth stratification \citep[e.g.,][]{sir92}, we place the nodes equispaced in the $\log \tau_c$ axis,
with $\tau_c$ the continuum optical depth at 5000 \AA.
\subsection{Orthogonal wavelet regularization}
The orthogonal wavelet transform \citep{ripples01} is a very powerful
sparsity-inducing transformation in cases in which the signal is smooth. One of the advantages
of the orthogonal wavelet transform is that a fast algorithm to compute the direct and inverse
transformation exists, without ever computing the transformation matrix $\mathbf{W}$.
\begin{figure*}
\includegraphics[width=\textwidth]{hazel_iuwt.pdf}
\caption{Observed (blue), fitted (green) and systematic effects (dashed red) obtained using the IUWT
in the region around 10830 \AA\ and using the updated version of \textsc{Hazel}. The left panels
correspond to using the full Alg. \ref{alg:proximal}, while the right panel shows the results when using the
simplified Eq. (\ref{eq:proximal_wavelet}).}
\label{fig:iuwt_hazel}
\end{figure*}
\begin{algorithm}[!t]
\KwData{Stokes profiles, model atmosphere, transform $\mathbf{W}$ and $k$.}
\KwResult{Regularized solution}
Initialization: $\mathbf{q}_0$ and $\mbox{\boldmath$\alpha$}_0=0$, first estimation of solution\;
\While{not converged}{
1. Compute gradient $\nabla_q \chi^2_\mathbf{W^T}(\mathbf{q}_i,\mbox{\boldmath$\alpha$}_i)$ and Hessian matrices $\mathbf{H}_q$\;
2. Modify Hessians: $\hat{\mathbf{H}}_q=\mathbf{H}_q+ \beta \mathrm{diag}(\mathbf{H}_q)$\;
3. Update $\mathbf{q}$ : $\mathbf{q}_{i+1} = \mathbf{q}_{i} - \hat{\mathbf{H}}_q^{-1} \nabla_q \chi^2_\mathbf{W^T}(\mathbf{q}_i,\mbox{\boldmath$\alpha$}_i)$ \;
4. Update $\mbox{\boldmath$\alpha$}$: $\mbox{\boldmath$\alpha$}_{i+1} = \mathrm{prox}_{p,\lambda,\tau}
\left[ \mbox{\boldmath$\alpha$}_i - \tau \nabla_\alpha \chi^2_\mathbf{W^T}(\mathbf{q}_{i+1},\mbox{\boldmath$\alpha$}_i) \right]$\;
}
\Return $\mathbf{q}$, $\mbox{\boldmath$\alpha$}$
\caption{Proximal Levenberg-Marquardt algorithm for synthesis penalty.}
\label{alg:proximal_synthesis}
\end{algorithm}
The specific approach for orthogonal transformations is irrelevant because both the
analysis and synthesis penalties are equivalent. For convenience, we choose to
do this study under the analysis penalty case.
We impose the sparsity constraint using the transformation matrix $\mathbf{W}$ associated with the
Daubechies-8 orthogonal wavelet and for different values of the regularization parameter $\lambda$.
The results are shown in Fig. \ref{fig:wavelet}. The left panel displays the observed data in
blue, the final fit in green (including systematic effects) and the inferred systematic effects
in dashed red. The right panel shows the wavelet coefficients of the systematic effects, showing only the first
100 coefficients of the potential 512 coefficients (the wavelength axis contains 512 sampled points).
Each panel contains the percentage of active (non-zero) wavelet coefficients.
Note that the fit quality is strongly affected by the value of the regularization parameters $\lambda$.
The important point is that it is possible to find values of $\lambda$ that lead to a good fit of the
He \textsc{i} multiplet and simultaneously fit the systematic effects. Of special relevance in this case is the
extended wing of the Si \textsc{i}, which sometimes makes it difficult to set a continuum level
for the 10830 \AA\ multiplet. Using this
approach, the continuum level is automatically obtained from the fit.
\begin{figure*}
\includegraphics[width=\textwidth]{lte_iuwt.pdf}
\caption{Observed (blue), fitted (green) and systematic effects (dashed red) obtained using the IUWT
in the 6301-6302 \AA\ spectral region and using the inversion code that assumes local thermodynamic equilibrium. The left panels
correspond to using the full Alg. \ref{alg:proximal}, while the right panel shows the results when using the
simplified Eq. (\ref{eq:proximal_wavelet}).}
\label{fig:iuwt_lte}
\end{figure*}
When $\lambda$ is small, a strong overfitting of the data occurs, in which even the noise is
absorbed by the systematic effects. For very large $\lambda$, the method cannot fit the observations.
For intermediate values of $\lambda$, a very nice fit is obtained. The optimal thresholding
of the wavelet coefficients is similar to the expected noise level, as pointed out by \cite{starck10}.
Using only 10\% of the wavelet coefficients is probably enough to have a fit of the whole spectral
region. It is interesting to note that the model that we impose for the He \textsc{i} multiplet is probably not enough for explaining
the observations at the noise level. This is the reason why some ``extra absorption'' is added by the systematic effects to the
wings of the red component. If this behavior is undesirable, it is potentially possible to introduce an extra
regularization to avoid them (for instance, an additional $\ell_0$ or $\ell_1$ regularization for the
vector $\mbox{\boldmath$\alpha$}$). However, we demonstrate in the following that a better option is to regularize
using different $\mathbf{W}$ transformations.
\begin{figure*}
\includegraphics[width=\textwidth]{synthesis_lte.pdf}
\caption{Observed (blue) and fitted (green) Stokes $I$ profile for different values of the
regularization parameters $\lambda$ and using a dictionary made of Voigt functions centered
at every pixel. The dashed red curve shows the inferred systematic effects. The left panels show the
results for the Fe \textsc{i} lines, while the right panels shows the active functions, with the
label indicating the percentage of active functions.}
\label{fig:lte_hazel_voigt}
\end{figure*}
\subsection{Isotropic undecimated wavelet regularization}
The isotropic undecimated wavelet transform \citep[IUWT;][]{starck94,starck10} algorithm is a non-orthogonal
redundant multiscale transform that is well suited to objects that are more or less isotropic. This transform
has found great success for the denoising of astronomical images \citep[e.g.,][and references therein]{starck10}.
Recently, we have witnessed examples of applications to one-dimensional spectra \citep{machado13}.
Given that this is a redundant non-orthogonal transform, the analysis penalty approach is more
efficient from a computational point of view.
The IUWT can be efficiently applied using the \hbox{\emph{\`a-trous}} algorithm, which proceeds as follows.
The original data $I(\lambda)$ are filtered, in our case using the B$_3$-spline given
by the filter $[1,4,6,4,1]/16$. The filtered data are substracted from the original ones, obtaining
what is commonly known as the \emph{detail}, $w_i(\lambda)$. The filtered data are again iteratively filtered (with scaled
filters) up to a depth $d$, computing the detail in each scale. The final smoothed signal will be
termed $c_d(\lambda)$. At the end, we have a smoothed version
of the original signal and a set of \emph{details} at each depth. The size of the IUWT is $d+1$ times
the original data, so the information is really encoded on the correlation among the transformed coefficients.
The data are reconstructed from the transformed data simply using:
\begin{equation}
I(\lambda) = c_d(\lambda) + \sum_{i=1}^d w_i(\lambda).
\end{equation}
The results shown in Fig. \ref{fig:iuwt_hazel} for the He \textsc{i} multiplet and \hbox{Fig. \ref{fig:iuwt_lte}}
for the Fe \textsc{i} doublet have been obtained using the IUWT up to $d=6$ and
applying the hard thresholding operator to all the details. The left panel shows the results when
the full \hbox{Alg. \ref{alg:proximal}} is used to apply the proximal algorithm, while the right
panels give the result when only the first iteration of this algorithm is used, which corresponds
to using Eq. (\ref{eq:proximal_wavelet}). The application of Eq. (\ref{eq:proximal_wavelet}) cannot
assure convergence to the correct solution to the problem, but we have tested that it
does a very good job. This simplifies the application of our proposed algorithm to any
existing inversion code because Eq. (\ref{eq:proximal_wavelet}) can be implemented in
only one or a few lines of code.
When the value of $\lambda$ is small, we witness again the overfitting, which also includes
the addition of some broadening of the red component of the He \textsc{i}. When $\lambda$ is
large, the fit is of low quality. In the intermediate range close to the expected noise
standard deviation, we find an excellent fit of the whole profile, with a flat contribution
exactly where the He \textsc{i} multiplet is located.
It is interesting to note that we have empirically found that applying Eq. (\ref{eq:proximal_wavelet}) instead of solving
the full problem via Alg. \ref{alg:proximal} gives better and more robust results (see Figs. \ref{fig:iuwt_hazel} and \ref{fig:iuwt_lte}).
The reason has to be found on the fact that we are solving the optimization problem of Eqs. (\ref{eq:problem_l0}) or
(\ref{eq:problem_l0_synthesis}) by separating
it into two simpler problems via an alternating optimization method. It is a general characteristic
of these methods that they work better in practice if none of the two problems is solved with
full precision at each iteration, but only approximately. If any of the two problems is solved with precision in
any iteration, one can produce some amount of overfitting that is very difficult to compensate
for in later iterations.
\subsection{Voigt functions}
The final example uses another non-orthogonal and redundant transformation, in this case not of general
applicability, but tailored to explain the systematic effects. The Hermite functions described
by \cite{hermite_deltoro03} is a good option, but we prefer to utilize a basis set made of Voigt functions centered
at every sampled wavelength point:
\begin{equation}
\phi_i(\lambda) = H \left( \frac{\lambda-\lambda_i}{\Delta}, a\right).
\end{equation}
Even though the basis is non-orthogonal, it is much easier in this case to work in
the synthesis prior approach, because the results are much more transparent. The reason
is that we directly impose the sparsity constraint on the coefficients associated with
all the Voigt functions.
In this case, we focus on the 6301-6302 \AA\ spectral region. The systematic effects are modeled with Voigt functions
placed at every sampled wavelength between 6301.15 and 6302.93 \AA, which includes both Fe \textsc{i} lines and
both telluric contaminations. The damping constant is fixed to
$a=0.8$, which gives line wings similar to those observed in the telluric lines. To boost sparsity,
we consider two different widths, $\Delta=14.8, 29.6$ m\AA, although more fits and damping parameters
can be easily considered with ease.
The results are displayed in Fig. \ref{fig:lte_hazel_voigt} for different values of $\lambda$. The
left panel shows the observations and the fit, together with the inferred systematic effects. The right panel shows the active
Voigt functions. When $\lambda$ is too small, a very good fitting is found, but we can see that the optimization induces
corrections on the spectral lines of interest. This is, in principle, undesired, although it is
clearly stating that the model proposed for the spectral lines cannot fit them with enough
precision. When $\lambda$ is large, the fit of the systematic effects is bad, which negatively impacts
the fit of the spectral lines of interest. For intermediate values of the regularization parameter, we find
an acceptable fit, where only the telluric absorption lines are fitted by the systematic effects with a very
sparse solution.
\section{Conclusions}
In the present study we have developed a new method that allows to deal with systematic effects in data inversions.
These systematic effects can include a variety of calibration defects, spectral lines from the Earth's atmosphere or spectral features that
are present in the observations that our model atmosphere cannot reproduce. Our method builds upon the assumption that these
systematic effects are not random noise and therefore they have an inherent \emph{level of sparsity} in
a suitable basis set.
We have shown that the assumption of sparsity induces a very powerful regularization because the resulting cross-talk between the
inversion model and the systematic effects can be made very low. To do so, we have introduced a regularization parameter $\lambda$ that
must be adjusted according to the needs of each problem based on experience. However, we argue that this step is not very different
from selecting the weights of the observations in current implementations of inversions codes and a rough estimate can be obtained \emph{a-priori}.
Although the mathematical characterization of the algorithm can appear a bit convoluted and depends on
recent advances on the optimization of non-smooth functions, a few extra lines of code should suffice to make
it work with current inversion codes. Our experience with the SIR and Hazel codes has positively confirmed it. When these systematic effects are not
corrected, inversion codes usually try to overcompensate their effect by converging the parameters of the model to a wrong solution. Our method
minimizes the impact of such effects in the convergence of the solution when an adequate value of the regularization parameter is selected.
Our method has a few interesting advantages over the downweighting technique
for dealing with systematic effects. Arguably the most important one concerns the assumptions
imposed on the generative model. Using the $\chi^2$ merit
function requires that the uncertainty of the residual between the observations
and the model is Gaussian with zero mean and a certain variance. This is not true unless one is
able to \emph{model all} expected signals. Only if everything is modeled (to a certain level, of course), one
is sure that the results can be interpreted appropriately (for instance, error bars). A second advantage
concerns the reduction in the subjectivity on the election of parameters. One only needs to choose
the value of $\lambda$ and the method automatically adapts the solution to the systematic effects.
|
\section{Introduction}
Future cellular networks require the support for high data rate video and content delivery.
Many researchers have recently focused on
proposing robust solutions to efficiently address the bandwidth utilization problem.
For example, the authors in \cite{chandrasekhar2008femtocell} proposed to create
home sized femtocells to overcome this issue.
Golrezaei et. al \cite{golrezaei2012femtocaching} proposed an alternate solution by introducing the concept of
femtocaching. In their solution, several {\em helper} nodes with high storage capabilites
are deployed in each cell to create a distributed wireless caching infrastructure.
These nodes will reduce the communication burden on the base station by satisfying many of
the {\em User Terminal (UT)} requests using the stored contents in their caches. Therefore, the storage capability
of helper nodes is used to increase the overall network capacity.
Currently, many researchers recommend to utilize high bandwidth Device-to-Device (D2D)
and Machine-to-Machine (M2M) communication capabilities for UTs. Current IEEE 802.11ad standard \cite{ieee80211ad} and
the millimeter-wave proposal for future 5G networks
\cite{boccardi2014five,rappaport2013millimeter}
are examples of such high bandwidth D2D communications which can enable up to hundreds of GHz of bandwidth.
Authors in \cite{kiskani2015multihop} suggest to use
this abundant bandwidth to deliver the contents from the helper nodes to the UTs through multihop communications.
Therefore, they extend the solution in \cite{golrezaei2012femtocaching} to allow
multihop communication between the helper and the UTs. This approach can significantly reduce network deployment
and maintenance costs without imposing restrictions on content delivery.
On the other hand, multihop communication between the helper node and the UTs
together with the use of UTs' storage capabilities can improve the overall network capacity.
Current improvements on the storage capacity of mobile devices show that future
UTs will have considerable under-utilized storage capabilities which can be effectively used to improve the network
content delivery. Utilizing the storage capability of UTs allows future cellular networks to move toward a
distributed D2D wireless caching network without imposing significant communication burden on the base station.
In this paper, we consider a wireless cellular network in which several helper nodes are deployed throughout
the network to create a wireless distributed caching infrastructure. Each helper is serving a
wireless ad hoc network of UTs through multihop communications as proposed in
\cite{kiskani2015multihop}.
We assume that helpers are connected to the base station through a high bandwidth backhaul infrastructure and have
access to all contents. They will use multihop communications to deliver the contents to the UTs. We assume
that the UTs also use their under-utilized storage capacity to improve network content delivery.
We will compute the capacity of such networks under decentralized random
coded and uncoded cache placement algorithms.
In decentralized cache placement algorithms, each UT's cache is populated independently of other UTs.
In a random decentralized uncoded cache placement algorithms, contents are chosen randomly and stored in UTs cache locations.
However, in a random decentralized coded cache placement algorithm, each UT stores a combination of multiple contents in its cache.
The UTs will follow this process until
their caches are fully populated. Coded cache placement is of interest in systems when the storage capacity of each node is
limited compared to the total number of contents in the network.
This paper computes the capacity of cellular networks with multihop communications using helper and relay nodes for both uncoded and coded random decentralized cache placement algorithm. Our prior work \cite{kiskani2015multihop} focused on multihop communications with helper nodes but without using the contents stored by the relay nodes. In this paper, the requests can be satisfied either by the helper node or a relay on the path between requesting node and the helper.
As far as we know, this is the first paper
to prove that coded caching which is originally motivated by the lack of sufficient storage capacity in UTs \cite{lee2015index}
can also increase the network capacity.
The rest of the paper is organized as follows. In section \ref{relwork}, the related work is discussed and
section \ref{netmod} describes the network model considered in this paper. Section \ref{uncoded_sec} focuses on the capacity computation of
wireless cellular networks operating under a decentralized random
uncoded cache placement algorithm and section \ref{coded_sec} reports the capacity for a
random coded cache placement algorithm. Simulation results are reported in
section \ref{sim_sec} and the paper is concluded in
section \ref{conc_sec}.
\section{Related Work}
\label{relwork}
The femtocaching network model is proposed in
\cite{golrezaei2012femtocaching, shanmugam2013femtocaching} and the capacity improvement for single-hop communication is computed.
In \cite{kiskani2015multihop}, the authors considered a femtocaching D2D network with multihop relaying of information from the helper to the UTs. They proposed a solution based on index coding in which the helper is utilizing the side information in the UTs to create index codes which are to be multicasted to the UTs. This way, they reduce bandwidth utilization by grouping multiple unicast transmissions into multicast transmission. However, that paper does not consider the case of coded side information and also it assumes that the relayed message from the helper cannot be changed based on the information in the relaying UTs.
Caching has been a subject of recent interest to many researchers. The fundamental limits of caching is studied
in \cite{maddah2014fundamental}. The results in \cite{maddah2014fundamental} has been extended to include
decentralized coded caching strategies in \cite{maddah2013decentralized,pedarsani2014online,hachem2014multi,
karamchandani2014hierarchical}.
Other researchers studied the problem of caching in wireless and D2D networks. Among them are the works of
authors in \cite{ji2014fundamental,ji2013wireless,jeon2015wireless}.
The authors in \cite{jeon2015wireless} have studied the capacity of wireless
D2D networks with caching in certain regimes. Our work is essentially different from all of these works in
the sense that the UT always request the content from helper (femtocache) while in these papers, a wireless ad hoc network is
considered where UTs' requests can be satisfied by any of the nodes in the network. Clearly, such network model requires significant overhead to locate the nearest UT with the requested content while in our approach, the request always is sent toward the helper.
Coded caching has been
previously suggested \cite{lee2015index,chen2014fundamental} as an efficient caching technique for devices with small storage capacity. Our results demonstrate that apart from the practical importance of coded caching in small storage
systems, it can be useful in increasing the capacity of cached networks.
\section{Network Model}
\label{netmod}
In this paper we will study the capacity of cellular networks utilizing a distributed femtocaching
infrastructure as proposed in \cite{golrezaei2012femtocaching}. In these networks, it is assumed that
several helpers with high storage capacity are deployed throughout the network to assist
the base station in delivering the contents to the UTs. The UTs can receive contents from helpers using D2D communications
through either single hop \cite{golrezaei2013femtocaching} or multiple hops \cite{kiskani2015multihop}.
\iffalse
\begin{figure}[H]
\centering
\begin{tikzpicture}[->,>=stealth',shorten >=1pt,auto,node distance=1.5cm, semithick]
\node[rectangle,draw] (H) {H};
\node[circle,draw] (A0) [right of=H] {}
\node[circle,draw] (A1) [above right of=H] {}
\node[circle,draw] (A2) [below right of=H] {}
\node[circle,draw] (A3) [right of=A0] {}
\node[circle,draw] (A4) [above right of=A3] {}
\node[circle,draw] (A5) [below right of=A3] {}
\node (T1) [right of = A1, xshift=-1cm] {\footnotesize{$N_1$}};
\node (T2) [right of = A2, xshift=-1cm] {\footnotesize{$N_3$}};
\node (T3) [above of = A0, yshift=-1.1cm] {\footnotesize{$N_2$}};
\node (T4) [above of = A3, yshift=-1.1cm] {\footnotesize{$N_4$}};
\node (T5) [right of = A4, xshift=-1cm] {\footnotesize{$N_5$}};
\node (T6) [right of = A5, xshift=-1cm] {\footnotesize{$N_6$}};
\pat
(H) edge node {} (A0);
\pat
(H) edge node {} (A1);
\pat
(H) edge node {} (A2);
\path (A0) edge node {} (A3);
\path (A3) edge node {} (A4);
\path (A3) edge node {} (A5);
\end{tikzpicture}
\caption{Example of a wireless multihop network being served by the helper $H$. Each arrow represents a link with high bandwidth
D2D communication capability.}
\label{fig_ex_netwoork}
\end{figure}
\fi
Assume that a helper is serving a D2D network of $n$ nodes. To analyze the capacity of this network, we will use the deterministic
routing
approach proposed in \cite{kulkarni2004deterministic}. Without loss of generality, it is assumed that the UTs are distributed on a
square of area
one and the helper is located at the center of the square as shown in Figure \ref{fig_model}.
\begin{figure}
\center
\includegraphics[scale=0.2,angle=0]{squarlet2.png}
\caption{The helper node which is located at the center of the unit square is serving $n$
UTs which are randomly distributed on a unit square. The square is divided into
$\Theta(\frac{n}{\log n})$ square-lets of area $\Theta(\frac{\log n}{n})$.
Gray square-lets can transmit simultaneously. Around each grey square-lets
there is a ``silence'' region of square-lets that are not allowed to transmit
in the given time slot.}
\label{fig_model}
\end{figure}
When a UT requests a content from
the helper, the content is routed from the helper to the UT in a sequence of horizontal and vertical
square-lets that are crossing the straight line which
connects the helper to the UT.
It is proved in \cite{kulkarni2004deterministic} that if
the UTs are uniformly distributed over the unit square area and
the area is divided into $\Theta(\frac{n}{\log n})$
square-lets each with area $\Theta(\frac{\log n}{n})$, then with a probability close to one, each square-let contains
$\Theta(\log n)$ UTs.
A minimum transmission range of $s(n) = \Theta(\sqrt{\frac{\log n}{n}})$ ensures network
connectivity \cite{penrose1997longest} in such a dense network. Therefore, assuming a transmission range of $s(n) = \Theta(\sqrt{\frac{\log n}{n}})$, the
proposed routing algorithm is proved to converge and all the UTs will be able to receive their requested contents with probability one.
To avoid multiple access interference, a
{\em Protocol Model} is considered \cite{xue2006scaling} for the successful communication between UTs.
According to this model, if the UT $i$ is placed at the coordinates $Y_i$, then a transmission from $i$ to another UT $j$ is
successful if $|Y_i-Y_j | < s(n)$ and for any other UT $k$ transmitting on the same frequency band, $|Y_k-Y_j| > (1 + \Delta)s(n)$
for a fixed guard zone factor $\Delta$. A Time Division Multiple Access (TDMA) scheme is assumed for the transmission between the
square-lets. With the assumption of Protocol Model,
it was shown \cite{kulkarni2004deterministic} that if the square-lets have a side length of $C_1 s(n)$ for a fixed
constant $C_1$ and if the square-lets with a distance of $C_2=\frac{2+\Delta}{C_1}$
square-lets apart from each other transmit simultaneously,
then there will be no interference between the concurrent transmissions.
Lets denote the data rate for each UT by $\lambda$, the number of hops between each UT and its helper by $x$, its average
value by $\mathbb{E}[x]$, and the total network throughput by $n \lambda$.
Therefore, on average the network delivers $n \lambda \mathbb{E}[x]$ bits in a unit of time. There are exactly
$\frac{1}{(C_2 C_1 s(n))^2}$ square-lets at any time slot available for transmission and if the total network bandwidth is
$W$ which is a constant value independent of $n$, then the total number of bits that the network is capable of delivering is
upper bounded \cite{azimdoost2013} by $\frac{W}{(C_2 C_1 s(n))^2}$. Hence,
\begin{equation}
\label{capa}
\lambda \le \lambda_{\max}=\dfrac{W}{n \mathbb{E}[x] (C_2 C_1 s(n))^2} = \Theta
\left(\dfrac{1}{\mathbb{E}[x] \log n} \right).
\end{equation}
This result implies that the maximum throughput can be derived by computing $\mathbb{E}[x]$. The capacity problem
is therefore reduced to computing the average number of hops traveled between the UTs and the helper.
We assume the number of contents in the network is $m$ which grows polynomially
with $n$ \cite{jeon2015wireless} as $m=C_3 n^{\alpha}$. We denote the set of indices of all contents by $\xi=\{1,2,\dots,m\}$.
Without loss of generality we assume that the contents with lower indices are more popular compared to the ones with higher indices.
We further assume that the contents can be categorized into two groups of highly popular contents and less popular contents.
Let's denote the requested content by $r$, then the probability that $r$ belongs to the highly popular group of contents should
be close to one. The highly popular and less popular groups can be defined as
\begin{mydef}
{\em For $\epsilon \in (0,1)$, define $h_{\epsilon}$ as the smallest integer such that if $\xi_{1-\epsilon} = \{1,2,\dots,h_{\epsilon}\}$ and
$\xi_{\epsilon} = \{h_{\epsilon}+1,h_{\epsilon}+2,\dots,m\}$, then $\textrm{P}[r \in \xi_{1-\epsilon}] \ge 1-\epsilon$.
}
\end{mydef}
We refer to $\xi_{1-\epsilon}$ as the group of highly popular contents and $\xi_{\epsilon}$ as the group of
less popular contents.
We assume that the helper has access to all the contents in $\xi$ but the UTs are assumed to have a limited cache of size
$M=C_4 n^{\beta}$. For the purpose of this paper, we assume that all UTs have the same cache size and the helper (or base station) is applying a decentralized caching strategy
to populate a UT cache independently of other UTs. Since the UTs have limited cache size, we assume that only popular contents in $\xi_{1-\epsilon}$ are stored in UTs caches.
Any request for the less
popular contents from $\xi_{\epsilon}$ will be satisfied directly by the helper or base station.
When a UT $i$ requests a content, if that specific content or a
group of coded contents which can be used to decode the content are available in the caches of the UTs in the routing
path between the UT $i$ and the helper, then the helper informs the UTs which have the coded contents in their caches to
send the content to UT $i$. If the content or a set of coded contents do not exist
in the caches of the UTs between UT $i$ and helper, then the content is routed to UT $i$ from the helper through on average
$s(n)^{-1} = \Theta( \sqrt{\frac{n}{ \log n}})$ hops. Since majority of the requests are from popular contents, these requests can be satisfied by the UTs instead of helper which reduces the average number of transmissions per request.
Therefore,
provided that the content request probability distribution is known, the average number of traveled hops in the network
can be written as
\begin{eqnarray}
\mathbb{E}[x] &=& \mathbb{E}[x | r \in \xi_{\epsilon}] \textrm{P}[r \in \xi_{\epsilon}] + \mathbb{E}[x | r \in \xi_{1-\epsilon}] \textrm{P}[r \in \xi_{1-\epsilon}], \nonumber \\
&=& \epsilon \sqrt{\frac{ n}{\log n}} + (1-\epsilon) \mathbb{E}[x | r \in \xi_{1-\epsilon}].
\label{ex112}
\end{eqnarray}
\begin{rem}{\em
By choosing $\epsilon=\frac{1}{\sqrt{n}}$,
the average hop count of the contents in $\xi_{\epsilon}$ will become less than one and therefore
the total average hop count can be approximated by the average hop counts of the
files in $\xi_{1-\frac{1}{\sqrt{n}}}$, i.e.,
\begin{eqnarray}
\mathbb{E}[x] &=& \frac{1}{\sqrt{n}} \sqrt{\frac{ n}{\log n}} +
(1-\frac{1}{\sqrt{n}}) \mathbb{E}[x | r \in \xi_{1-\frac{1}{\sqrt{n}}}] \nonumber \\
&=& \Theta( \mathbb{E}[x | r \in \xi_{1-\frac{1}{\sqrt{n}}}] )
\label{ex1532}
\end{eqnarray}
}\label{dmfbfj}
\end{rem}
For many web applications \cite{breslau1998implications,breslau1999web}, the content request popularity follows Zipfian-like distributions.
Although we will express our results in general form without any specific assumption,
we will later compute explicit capacity results assuming a Zipfian content popularity distribution.
Our main results in proving the gain of coded caching over uncoded caching
is independent of the content popularity distribution.
For a Zipfian content popularity distribution with parameter $s$,
the probability of requesting a content with popularity index $i$ will have the
form
$ \textrm{P}[r = i] = \frac{i^{-s}}{\sum_{j=1}^m j^{-s}} = \frac{i^{-s}}{H_{m,s}},
$
where $H_{m,s}$ represents the generalized harmonic number with parameter $s$.
\begin{rem}{\em
In case of Zipfian distribution with $s>1$,
when few popular contents are widely
requested by the UTs, we have
\begin{eqnarray}
\textrm{P}[r \in \xi_{\epsilon}] = \sum_{i=h_{\epsilon}+1}^m \frac{i^{-s}}{H_{m,s}} \le
\frac{(m-h_{\epsilon})(h_{\epsilon})^{-s}}{H_{m,s}}.
\label{enbg3}
\end{eqnarray}
Assuming that $m=C_3 n^{\alpha}$ is a large number, $H_{m,s}$ converges to Reimann Zeta function $\zeta(s)$. Since
the number of popular contents is negligible compared to the total number of contents, $\frac{m-h_{\epsilon}}{H_{m,s}}$ can be upper bounded by
$\frac{2m}{\zeta(s)}$ and therefore in case of a Zipfian distribution with $s>1$, we have
\begin{eqnarray}
\textrm{P}[r \in \xi_{\epsilon}] \le \frac{2C_3 n^{\alpha}(h_{\epsilon})^{-s}}{\zeta(s)}.
\label{enbg33dsf}
\end{eqnarray}
In order to compute $ h_{\frac{1}{\sqrt{n}}} $ such that $\textrm{P}[r \in \xi_{\frac{1}{\sqrt{n}}}] \le \frac{1}{\sqrt{n}}$, it is sufficient to have
\begin{equation}
h_{\frac{1}{\sqrt{n}}} = \Theta \left(n^{\frac{1}{s}(\alpha + \frac{1}{2})} \right).
\label{jdfdh}
\end{equation}
\label{derreol1}
Since we implicitly assume that $h_{\frac{1}{\sqrt{n}}} = O(m) = O (C_3 n^{\alpha})$, equation \eqref{jdfdh} is valid when
$\alpha > \frac{1}{2(s-1)}$.
}
\end{rem}
\begin{rem}{\em
In case when $\beta > \frac{1}{s}(\alpha + \frac{1}{2})$,
the average number of traveled hops can be zero since in that case, all UTs
can store all the popular
contents in their caches. Therefore, in this case, the maximum per node capacity $\Theta(1)$ is trivially achievable.
\label{mdgfb}
}\end{rem}
For the rest of paper, we compute capacity assuming that the number of popular contents $h_{\epsilon}$
is known. The capacity for the special case of Zipfian distribution will be derived as well.
\section{Decentralized Uncoded Caching}
\label{uncoded_sec}
This section focuses on computing the capacity of cellular networks when UTs cache uncoded contents in a distributed fashion. It is assumed that the UTs only cache the most popular contents from $\xi_{1-\epsilon}$.
\begin{lem}
{\em If a content is drawn uniformly at random from the set of most popular contents in
$\xi_{1-\epsilon}$, then the average required number of requests to have at least one copy
of each content from $\xi_{1-\epsilon}$ is equal to
\begin{equation}
\mathbb{E}[l] = h_{\epsilon} H_{h_{\epsilon}} = h_{\epsilon} \sum_{i=1}^{h_{\epsilon}} \frac{1}{i}
= \Theta(h_{\epsilon} \log h_{\epsilon}),
\label{codhf}
\end{equation}
where $H_{h_{\epsilon}}$ is the $h_{\epsilon}^{th}$ harmonic number.
This problem is similar to the well-known coupon collector problem.
}
\label{leme0}
\end{lem}
\begin{proof}
Denote $t_i$ as the number of required requests to collect the $i^{th}$ content after $(i-1)^{th}$ content have been
collected. Notice that the probability of collecting a new content given that $i-1$ contents have been collected is
equal to $p_i = \frac{{h_{\epsilon}}-(i-1)}{{h_{\epsilon}}}$. Therefore, $t_i$ has
geometric distribution with expected value of
$\frac{1}{p_i}=\frac{{h_{\epsilon}}}{{h_{\epsilon}}-(i-1)}$. By the linearity of expectation we have:
\begin{eqnarray}
\mathbb{E}[l]=\sum_{i=1}^{{h_{\epsilon}}} \mathbb{E}[t_i] =
\sum_{i=1}^{{h_{\epsilon}}} \frac{{h_{\epsilon}}}{{h_{\epsilon}}-(i-1)}
= {h_{\epsilon}} \sum_{i=1}^{{h_{\epsilon}}} \frac{1}{i} = h_{\epsilon} H_{h_{\epsilon}} \nonumber
\label{bsdhgtr}
\end{eqnarray}
\end{proof}
\begin{rem}{\em
Notice that
the contents in UT caches should be stored such that each UT does not
cache a content more than once. Therefore, this problem cannot
exactly be modeled by the coupon collector problem but
when $h_{\epsilon} >> M$, the probability of having
multiple instances of the same content in one UT goes
to zero and hence the above argument is valid and
$\mathbb{E}[l]= h_{\epsilon} H_{h_{\epsilon}}$.
}\label{rem_explain}
\end{rem}
Note that during placement phase, we cache contents from the popular set $\xi_{1-\epsilon}$ inside the UTs independently and with uniform distribution. The distribution of placement of contents inside the UTs is different from the popularity distribution of the contents.
\begin{thm}{\em
In a cellular network with femtocaching operating under a
decentralized uncoded caching assumption, the average number of traveled hops is
\begin{equation}
\mathbb{E}[x]=\mathbb{E}[x | r \in \xi_{1-\frac{1}{\sqrt{n}}}] =
\Theta \left(\frac{h_{\frac{1}{\sqrt{n}}} \log h_{\frac{1}{\sqrt{n}}}}{M} \right).
\label{ex_uncoded}
\end{equation}
Therefore, the following capacity is achievable
\begin{equation}
\lambda_{\textrm{uncoded}} = \Theta(\frac{1}{ \mathbb{E}[x] \log n})=
\Theta \left(\frac{M}{h_{\frac{1}{\sqrt{n}}} \log h_{\frac{1}{\sqrt{n}}} \log n} \right).
\label{capa_uncoded}
\end{equation}
}
\label{thm_uncoded}
\end{thm}
\begin{proof}
Lemma \ref{leme0} shows that the average number of cache places needed so that all of the requests can
be satisfied is $\Theta(h_{\frac{1}{\sqrt{n}}} \log h_{\frac{1}{\sqrt{n}}})$. Since each UT has a cache size
$M$, it is obvious that the average number of users needed such that at least one copy of each content is available
in their caches is $\Theta(\frac{h_{\frac{1}{\sqrt{n}}} \log h_{\frac{1}{\sqrt{n}}}}{M})$. Hence, along the routing
path to the helper, the average number of required hops needed so that the UT can reach its desired content is
$\Theta(\frac{h_{\frac{1}{\sqrt{n}}} \log h_{\frac{1}{\sqrt{n}}}}{M})$. This proves the theorem. Equation
\eqref{capa} can be used to compute the capacity by replacing $\mathbb{E}[x]$ with the above result.
\end{proof}
We can use equation \eqref{jdfdh} to simplify the results of theorem
\ref{thm_uncoded}
to the case of Zipfian content request distribution.
\begin{corol}{\em
In a cellular femtocaching network with Zipfian content request distribution with parameter $s>1$ and assuming
$\alpha > \frac{1}{2(s-1)}$, the following capacity result is achievable.
\begin{equation}
\lambda_{\textrm{uncoded}}^{\textrm{Zipf}} =
\Theta \left(n^{\beta- \frac{1}{s} (\alpha + \frac{1}{2}) } \frac{1}{(\log n)^2}\right)
\label{capa_uncoded_zipfian}
\end{equation}
\label{mikloplo}
}\end{corol}
\vspace{-0.4in}
\section{Decentralized Coded Caching}
\label{coded_sec}
In this section we will find capacity results assuming that the UTs are caching coded contents from the
set of popular contents in $\xi_{1-\epsilon}$ independently of
other UTs. We propose a random coding strategy and we will prove that if UTs follow this random coding
strategy, the capacity will be increased by a factor of $\log h_{\epsilon}$. The result proves that not only coded
caching is more efficient in small storage scenarios \cite{lee2015index}, but also it increases the capacity. We first describe the random coding cache placement and the decoding procedure.
{\bf Coded cache placement:} For the purposes of this paper, we assume that random coding is done over
Galois Field GF(2). For each encoded file,
the helper node (or base station) randomly selects each one of contents from the set $\xi_{1-\epsilon}$
with probability $\frac{1}{2}$ and then add all the selected contents to create one encoded file. For a UT with
cache size $M$, the helper node creates $M$ of these encoded files. Therefore, each one of the contents in $\xi_{1-\epsilon}$ has
been used on average $\frac{M}{2}$ times to create the $M$ coded files.
{\bf Coded file reconstruction:} When a UT requests a content, if the content is among the set of popular contents
$\xi_{1-\epsilon}$, it sends the request to the helper. The helper then decides to send the file through a routing
path as proposed in \cite{kulkarni2004deterministic}. However, it is highly possible that the content can be reconstructed
using the coded contents in the caches of UTs between the requesting UT and the helper along the routing path. If that
is the case, then the helper sends appropriate coding information to the relaying UTs along the routing path and each
relay UT that has useful information, add that information to the file that is being relayed to the requesting UT. This
procedure continues hop by hop until the content reaches the requesting UT. After the requesting UT receives this file,
it can reconstruct the desired content by applying its own coding gains to the received coded file. By doing so, there will
not be multiple transmissions by relaying UTs to construct the requested content.
To prove our results
we will first prove the following lemma.
\begin{lem}
{\em
If for a vector $v_i \in \mathbb{F}_2^{{h_{\epsilon}}}$, every element is equal to 1 with probability
$\frac{1}{2}$ and equal to 0 with
probability $\frac{1}{2}$ and $\{ v_1,v_2,\dots,v_q \}$ span the vector space of
$\mathbb{F}_2^{h_{\epsilon}}$, then
the average required number of such vectors to span the set $\mathbb{F}_2^{{h_{\epsilon}}}$ equals
$ \mathbb{E}[q] = {h_{\epsilon}} + \sum_{i=1}^{h_{\epsilon}}
\frac{1}{2^{i}-1} = {h_{\epsilon}} + \gamma $ where
$\gamma \approx 1.6067$ is called the Erdős–Borwein constant.
}
\label{leme1}
\end{lem}
\vspace{-0.2cm}
\begin{proof}
We can form a Markov chain to model the problem.
The states of this Markov chain are equal to the dimension of the space spanned by vectors $v_1,v_2,\dots,v_l$.
Let $k_l$ ($k_l \leq h_{\epsilon}$) denote the dimension of the space spanned by vectors $v_1,v_2,\dots,v_l$.
Therefore, the Markov chain will have $k_l+1$ distinct states. Assuming that we are in state $k_l$,
we want to find the
probability that adding a new vector will change the state to $k_l+1$. When we are in state $k_l$, adding
$2^{k_l}$ vectors out of the total $2^{h_{\epsilon}}$ possible vectors will not change the dimension while adding any one of
$2^{h_{\epsilon}} - 2^{k_l}$ new vectors will change the
dimension to $k_l+1$. Therefore, the Markov chain can be represented as the one in Figure \ref{markovchain}.
\begin{figure}[H]
\begin{center}
\begin{tikzpicture}[->,>=stealth',auto,semithick,node distance=1.8cm]
\tikzstyle{every state}=[fill=white,draw=black,thick,text=black,scale=1]
\node[state] (k0) {};
\node[state] (k1)[right of=k0] {};
\node[state] (k2)[right of=k1] {};
\node[state] (k3)[right of=k2] {};
\node[state] (kn)[right of=k3] {};
\node [label={[label distance=0.5cm]$k_l=0$}] (t0)[below of=k0]{};
\node [label={[label distance=0.5cm]$k_l=1$}] (t1)[below of=k1] {};
\node [label={[label distance=0.5cm]$k_l=2$}] (t2)[below of=k2] {};
\node [label={[label distance=0.5cm]$k_l=3$}] (t3)[below of=k3] {};
\node [label={[label distance=0.5cm]$k_l={h_{\epsilon}}$}] (tn)[below of=kn] {};
\path (k0) edge[loop above] node{$\frac{1}{2^{h_{\epsilon}}}$} (k0);
\path (k0) edge[above] node{$1-\frac{1}{2^{h_{\epsilon}}}$} (k1);
\path (k1) edge[loop above] node{$\frac{2}{2^{h_{\epsilon}}}$} (k1);
\path (k1) edge[above] node{$1-\frac{2}{2^{h_{\epsilon}}}$} (k2);
\path (k2) edge[loop above] node{$\frac{2^2}{2^{h_{\epsilon}}}$} (k2);
\path (k2) edge[above] node{$1-\frac{2^2}{2^{h_{\epsilon}}}$} (k3);
\path (k3) edge[loop above] node{$\frac{2^3}{2^{h_{\epsilon}}}$} (k3);
\path (k3) edge[dashed] node{} (kn);
\path (kn) edge[loop above] node{$1$} (kn);
\end{tikzpicture}
\end{center}
\vspace{-1cm}
\caption{The state space of the Markov chain used in proof of lemma \ref{leme1}.}
\label{markovchain}
\end{figure}
The state transition matrix for this Markov chain
\iffalse
is
\begin{equation}
P=\begin{bmatrix}
\frac{1}{2^{h_{\epsilon}}} & 1- \frac{1}{2^{h_{\epsilon}}} & 0 & \cdots & 0 &0 \\
0 & \frac{2}{2^{h_{\epsilon}}} & 1- \frac{2}{2^{h_{\epsilon}}} & \cdots & 0 &0 \\
0 & 0 & \frac{2^2}{2^{h_{\epsilon}}} & \cdots & 0 &0 \\
\vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\
0 & 0 & 0 & \cdots & 1- \frac{2^{{h_{\epsilon}}-2}}{2^{h_{\epsilon}}} & 0\\
0 & 0 & 0 & \cdots &\frac{2^{{h_{\epsilon}}-1}}{2^{h_{\epsilon}}} & 1-\frac{2^{{h_{\epsilon}}-1}}{2^{h_{\epsilon}}}\\
0&0&0&\cdots&0& 1\\ \end{bmatrix}, \nonumber
\end{equation}
which
\fi
can be written in the form of a discrete phase-type distribution as
\begin{equation}
P=\begin{bmatrix}
T & T_0 \\
0 & 1 \\ \end{bmatrix},
\label{msdbf}
\end{equation}
where
\begin{equation}
T=\begin{bmatrix}
\frac{1}{2^{h_{\epsilon}}} & 1- \frac{1}{2^{h_{\epsilon}}} & 0 & \cdots & 0 &0 \\
0 & \frac{2}{2^{h_{\epsilon}}} & 1- \frac{2}{2^{h_{\epsilon}}} & \cdots & 0 &0 \\
0 & 0 & \frac{4}{2^{h_{\epsilon}}} & \cdots & 0 &0 \\
\vdots & \vdots & \vdots & \ddots & \vdots &\vdots \\
0 & 0 & 0 & \cdots & \frac{2^{{h_{\epsilon}}-2}}{2^{h_{\epsilon}}} & 1- \frac{2^{{h_{\epsilon}}-2}}{2^{h_{\epsilon}}} \\
0 & 0 & 0 & \cdots & 0 &\frac{2^{{h_{\epsilon}}-1}}{2^{h_{\epsilon}}} \\
\end{bmatrix},
\end{equation}
\begin{equation}
T_0^t= [ 0 \hspace{0.1in} 0 \ldots 1-\frac{2^{{h_{\epsilon}}-1}}{2^{h_{\epsilon}}}],
\end{equation}
and $t$ denotes transpose operation.
If $e$ denotes all one vector of size ${h_{\epsilon}}$,
since $P$ is a probability distribution we have
$P\begin{bmatrix}e \\ 1 \end{bmatrix}=\begin{bmatrix}e \\ 1 \end{bmatrix}.
$ This
implies that $Te + T_0 = e$, hence $ T_0 = (I - T)e$. Therefore, it is easy to show by induction that the state transition matrix in $l$ steps can
be written as
\begin{equation}
P^l=\begin{bmatrix}
T^l & (I-T^l)e \\
0 & 1 \\ \end{bmatrix}.
\label{mskit}
\end{equation}
This equation implies that if we define the absorption time as
\begin{equation}
q = \inf \{l \ge 1 ~|~ k_l={h_{\epsilon}}\},
\end{equation}
and if $l$ is strictly less than the absorption time, the probability of transitioning from state $i$ to state $j$ by having $l$ new vectors
can be found from the submatrix $T^l$ of $P^l$. In other words,
\begin{equation}
\mathrm{P}_i^l[k_l = j, l < q] = (T^l)_{ij}.
\label{probres}
\end{equation}
Therefore, starting from state $i$, if $t_j^i$ denotes the time spent in state $j$ before absorption,
$t_j^i$ can be written as
\begin{equation}
t_j^i = \sum_{l=0}^{q-1} \mathrm{1}\{k_l=j\}
\label{msdhlsdjfusdg}
\end{equation}
Therefore, starting from state $i$, the average time spent in state $j$ will be equal to
\begin{eqnarray}
\mathbb{E}[t_j^i] = \mathbb{E} \left[\sum_{l=0}^{q-1} \mathrm{1}\{k_l=j\} \right] = \sum_{l=0}^{q-1} \mathbb{E} \left[ \mathrm{1}\{k_l=j\} \right]. \nonumber
\label{expected1}
\end{eqnarray}
Since $\mathbb{E} \left[ \mathrm{1}\{k_l=j\} \right] = \mathrm{P}_i^l(k_l = j, l \le q-1) $, we have
\begin{equation}
\mathbb{E}[t_j^i] = \sum_{l=0}^{q-1} \mathrm{P}_i^l(k_l = j, l \le q-1)
= \sum_{l=0}^{\infty} \mathrm{P}_i^l(k_l = j, l \le q-1) \nonumber
\end{equation}
\begin{equation}
= \sum_{l=0}^{\infty} \mathrm{P}_i^l(k_l = j, l < q) = \sum_{l=0}^{\infty} (T^l)_{ij}.
\label{expected2}
\end{equation}
Since the probability is nonzero up to $q-1$, then we can extend the summation to infinity adding zero terms in \eqref{expected2}.
Notice that the equality in the last line comes from equation \eqref{probres}.
If we denote matrix $U=(\mathbb{E}[t_j^i])_{ij}$, using equation \eqref{expected1} and using matrix algebra, we have
\begin{equation}
U = \sum_{i=0}^{\infty}T^i =(I-T)^{-1}.
\end{equation}
It is not difficult to verify that
\begin{equation}
\nonumber \\
U=(I-T)^{-1}=
\begin{bmatrix}
\frac{2^{h_{\epsilon}}}{2^{h_{\epsilon}}-1} & \frac{2^{{h_{\epsilon}}-1}}{2^{{h_{\epsilon}}-1}-1} & \frac{2^{{h_{\epsilon}}-2}}{2^{{h_{\epsilon}}-2}-1}& \cdots &2 \\
0 & \frac{2^{{h_{\epsilon}}-1}}{2^{{h_{\epsilon}}-1}-1} & \frac{2^{{h_{\epsilon}}-2}}{2^{{h_{\epsilon}}-2}-1}& \cdots&2\\
0 & 0 & \frac{2^{{h_{\epsilon}}-2}}{2^{{h_{\epsilon}}-2}-1}& \cdots&2\\
\vdots & \vdots & \vdots & \ddots & \vdots \\
0 & 0 & 0 & \cdots & 2
\end{bmatrix}
\end{equation}
Therefore, starting at $k_l=0$, the average time it takes to get to absorption is equal to
\begin{eqnarray}
\mathbb{E}[q] &=&
\begin{bmatrix}1 & 0 & \cdots & 0\end{bmatrix} U e \nonumber \\
&=& \begin{bmatrix}1 & 0 & \cdots & 0\end{bmatrix} (I-T)^{-1} e \nonumber \\
&=& \sum_{i=1}^{h_{\epsilon}} \frac{2^i}{2^i-1} ={h_{\epsilon}} + \sum_{i=1}^{h_{\epsilon}} \frac{1}{2^i-1}
\end{eqnarray}
This proves the lemma.
\label{prof1}
\end{proof}
This lemma shows that each UT's request can be satisfied in a smaller number of hops compared to an uncoded caching
strategy. Therefore, the capacity will be increased. The following theorem formalizes this.
\begin{thm}{\em
In a cellular network with femtocaching,
our proposed decentralized coded caching in which each popular content in $\xi_{1-\epsilon}$ is
present in any cache location with probability $\frac{1}{2}$ reduces the required number of
traveled hops for each request by UTs to at most
\begin{equation}
\mathbb{E}[x]=\mathbb{E}[x | r \in \xi_{1-\frac{1}{\sqrt{n}}}] =
\Theta \left( \frac{h_{\frac{1}{\sqrt{n}}}}{M}\right).
\label{ex_coded}
\end{equation}
Therefore, the following capacity is achievable through decentralized coded caching.
\begin{equation}
\lambda_{\textrm{coded}} = \Theta(\frac{1}{ \mathbb{E}[x] \log n})=
\Theta \left(\frac{M}{h_{\frac{1}{\sqrt{n}}} \log n}\right)
\label{capa_coded}
\end{equation}
}
\label{thm_coded}
\end{thm}
\begin{proof}
Lemma \ref{leme1} shows that to be able to decode a requested content, on average $\Theta( h_{\frac{1}{\sqrt{n}}})$
coded contents are required. Since each UT has a cache of size $M$, we need $\Theta( \frac{h_{\frac{1}{\sqrt{n}}}}{M})$ UTs to be able to
decode the desired content. This means that along the routing path, we only need to travel
a distance of $\Theta (\frac{ h_{\frac{1}{\sqrt{n}}}}{M})$ hops away from each UT to find
all the contents that the UT requires for decoding its desired content.
Notice that individual UTs do not need to separately send their coded content to the requesting node. Each UT can combine the appropriate encoded files to the received file along the
route to the requesting node.
\end{proof}
Similarly, the results of theorem \ref{thm_coded} can be simplified by using
equation \eqref{jdfdh} for the case of Zipfian content request distribution.
\begin{corol}{\em
In a cellular femtocaching network with Zipfian content request distribution with parameter $s>1$ and assuming
$\alpha > \frac{1}{2(s-1)}$, the following capacity result
is achievable through decentralized coded caching.
\begin{equation}
\lambda_{\textrm{coded}}^{\textrm{Zipf}} = \Theta \left(n^{\beta- \frac{1}{s} (\alpha + \frac{1}{2} )}
\frac{1}{\log n}\right)
\label{capa_coded_zipfian}
\end{equation}
\label{miklo}
\vspace{-0.4cm}
}\end{corol}
{ Our proposed coded caching strategy can be done with
insignificant overhead as the coding instructions sent from the helper
is negligible compared to the size of the files. The computational
complexity of in each UT (XOR operation) is also not significant. However, the
helper requires to have high computational complexity capability. Future works should
concentrate on the ways to reduce the complexity and delay for this
approach.}
\vspace{-0.5cm}
\section{Simulations}
\label{sim_sec}
The simulation results compare the performance of our proposed decentralized
random coded caching with decentralized random uncoded caching. We assume a helper which is serving $n=2500$ UTs. The
Zipfian content request probability parameter is $s=2.5$, $\alpha = 1.5$, and $C_3 =8$ which means that a
total of $m=1000000$ contents exist in the network and 523 popular contents are considered for this simulation. The cache size parameter
$\beta$ is ranging from $0.3$ to $0.8$ while $C_4=1$. Figure \ref{fig_sim22} shows the simulation results comparing the average number of hops
required to decode the content in both decentralized coded and uncoded caching. As can be seen from this figure, our proposed
decentralized random coded cache placement algorithm can significantly reduce the average number of traveled hops compared to
decentralized uncoded cache placement.
Further, the theoretical results match the simulation results for both cases.
\begin{figure}
\center
\includegraphics[scale=0.4,angle=0]{fig_sim.eps}
\vspace{-0.1in}
\caption{Simulation results show that the expected number of hops in case of decentralized coded caching
is significantly lower than the expected number of hops for decentralized uncoded caching.}
\label{fig_sim22}
\end{figure}
\vspace{-0.3cm}
\section{Conclusions}
\label{conc_sec}
In this paper, we studied the content delivery capacity in cellular networks with femtocaching
with decentralized uncoded and coded caching for UTs. We computed the capacity of random
decentralized uncoded caching. We then proposed a random
coded caching strategy for network users and proved that this random coded caching technique can improve the capacity.
Note that we did not consider the possibility of congestion near helper node since all contents are moving toward that node. In the future work, we intend to study the effects of congestion on the capacity of the network.
\bibliographystyle{plain}
|
\section{Introduction}
In the 1930's, Seifert \cite{Seifert35} introduced the study of a knot $K$ via a compact connected oriented surface now called a \textbf{Seifert surface} bounded by $K$. While any knot admits many different Seifert surfaces, the study of any Seifert surface for a fixed knot results in interesting invariants. For example, Seifert surfaces are used to compute The Alexander module \cite{Alexander28}, the Levine-Tristram signature function \cite{L5}, and the Conway polynomial \cite{Kauffman81}. See also \cite{NS03, Seifert35, Seifert50}.
In \cite{Cooper82}, Cooper defines a generalization of a Seifert surface called a \textbf{C-complex} (or {clasp-complex}). Informally, a C-complex is a collection of embedded surfaces in $S^3$ which might intersect each other in clasps. An example is depicted in Figure~\ref{fig:Ccomplex}. Similar to a Seifert surface, these objects are not themselves invariants of links, yet out of them many important invariants of links can be understood. See for example, \cite{Cimasoni2004, CimFlo, Cooper82}. A formal definition is given in Section~\ref{sect: C-complexes}.
\begin{figure}
\centering
\includegraphics[width=14em]{Ccomplex.pdf}
\caption{An example of a C-complex consisting of two embedded disks which intersect each other in four clasps.}
\label{fig:Ccomplex}
\end{figure}
Notice that for any knots $K_1$ and $K_2$, if $K_1$ bounds a genus $g_1$ Seifert surface $F_1$, $K_2$ bounds a genus $g_2$ Seifert surface $F_2$, and $g_1<g_2$ then by stabilizing $F_1$ as in Figure~\ref{fig:Genus} enough times we create a new Seifert surface $F_1'$ for $K$ such that $g(F_1') = g(F_2)$. Since the genus and the number of boundary components give a complete set of invariants of compact oriented connected surfaces we conclude that $K_1$ and $K_2$ bound homeomorphic Seifert surfaces.
\begin{figure}
\begin{picture}(190,30)
\centering
\put(0,0){\includegraphics[height=2em]{PreAddGenus.pdf}}
\put(90,5){\huge{$\rightsquigarrow$}}
\put(130,0){\includegraphics[height=2em]{AddGenus.pdf}}
\end{picture}
\caption{Stabilizing to increase the genus of a surface.}
\label{fig:Genus}
\end{figure}
The goal of this paper is to ask when the same is true of C-complexes. Given two $n$-component links $L$ and $J$, when do there exist C-complexes $F$ and $G$ for these links which are related by an orientation preserving homeomorphism? To what extent is the homeomorphism type of a C-complex an invariant of the link? We will call two C-complexes $F=F_1\cup \dots \cup F_n$ and $G=G_1\cup \dots \cup G_n$ \textbf{equivalent} if there is a homeomorphism $\Phi:F\to G$ which restricts to orientation preserving homeomorphisms from the components of $F$ to the components of $G$ and preserves the signs of the clasps. See Definition~\ref{def: equiv} in Section~\ref{sect: C-complexes} for more detail.
Since the pairwise linking numbers of a link can be computed by counting clasps in a C-complex with sign, it is clear that the pairwise linking number is an obstruction to two links admitting equivalent C-complexes. In the case of 2-component links linking number is the only obstruction.
\begin{theorem}\label{thm: 2-components}
Let $L=L_1 \cup L_2$ and $J=J_1 \cup J_2$ be 2-component links. Then $\operatorname{lk}(L_1 , L_2) = \operatorname{lk}(J_1,J_2)$ if and only if $L$ and $J$ admit equivalent C-complexes.
\end{theorem}
In \cite{M2}, Milnor produced a family of \textbf{higher order linking numbers}. For an $n$-component link $L$, we will be most interested in the \textbf{triple linking number} $\overline{\mu}_{ijk}(L)\in \mathbb{Z}$ (with $1\le i<j<k\le n$). In \cite{MellorMelvin2003}, Mellor and Melvin give a means of computing Milnor's triple linking number from a C-complex. We recall this result in Section \ref{sect:triple linking}. Their formulation depends only on the equivalence class of a C-complex for the link. Hence we see that triple linking number gives an obstruction to links admitting equivalent C-complexes. Indeed, in the case of links with vanishing pairwise linking numbers, the triple linking numbers form a complete obstruction.
\begin{theorem}\label{thm:main}
Let $L$ and $J$ be $n$-component links with vanishing pairwise linking numbers. Then the following are equivalent
\begin{enumerate}
\item For all $1\le i<j<k\le n$, $\overline{\mu}_{ijk}(L)=\overline{\mu}_{ijk}(J)$
\item $L$ and $J$ admit equivalent C-complexes
\item There exist unknotted curves $\gamma_1,\dots,\gamma_k$ disjoint from $L$ such that $\operatorname{lk}(L_i,\gamma_j)=0$ for all $i,j$ and $J$ is obtained from $L$ by some surgery on $\gamma_1,\dots,\gamma_k$.
\end{enumerate}
\end{theorem}
The implication (2) $\implies$ (1) follows immediately from the formulation of $\overline \mu_{ijk}(L)$ in \cite{MellorMelvin2003}. We recall this formulation in subsection \ref{sect:triple linking}. The implication (1) $\implies$ (3) relies on a necessary and sufficient condition due to Martin \cite[Theorem 1]{MartinThesis} for two links to be related by a sequence of band pass moves. See Figure \ref{fig:BPSurgery}.
For links with nonvanishing pairwise linking number, $\overline{\mu}_{ijk}$ is only well defined modulo the greatest common divisor of $\operatorname{lk}(L_i, L_j)$, $\operatorname{lk}(L_i, L_k)$ and $\operatorname{lk}(L_j, L_k)$. On our way to proving Theorem~\ref{thm:main} above we gain the following result.
\begin{proposition}\label{Prop:mu123}
Let $L$ and $J$ be $n$-component links. If $L$ and $J$ admit equivalent C-complexes, then for all $1\le i<j<k$, $\overline{\mu}_{ijk}(L) = \overline{\mu}_{ijk}(J)$ (modulo the greatest common divisor of $\operatorname{lk}(L_i, L_j)$, $\operatorname{lk}(L_i, L_k)$ and $\operatorname{lk}(L_j, L_k)$).
\end{proposition}
Note that our results fail to address the question posed in the title of this document in the case of links with more than two components and non-vanishing pairwise linking numbers. The complete solution will require an answer to the following question.
\begin{question}
Let $L$ and $J$ be $n$-component links. Suppose that $\operatorname{lk}(L_i,L_j)=\operatorname{lk}(J_i,J_j)$ for all $i,j$ but that $\operatorname{lk}(L_i,L_j)\neq 0$ for some $i,j$. Suppose also that $\overline{\mu}_{ijk}(L)=\overline{\mu}_{i,j,k}(L)$ modulo the greatest common divisor of $\operatorname{lk}(L_i, L_j)$, $\operatorname{lk}(L_i, L_k)$ and $\operatorname{lk}(L_j, L_k)$. Does it follow that $L$ and $J$ admit equivalent C-complexes?
\end{question}
The genus of a knot $K$ is defined to be the minimum genus of all Seifert surfaces for $K$. Similarly, for a link $L$ one can define $\beta(L)$ to be the minimum first Betti number of all C-complexes for $L$. The following question asks how this measure of complexity behaves for equivalent C-complexes.
\begin{question}
Suppose that $L$ and $J$ are $n$-component links which admit equivalent C-complexes. Let $\beta(L,J)$ the the minimum first betti number of all C-complexes for $L$ which are equivalent to some C-complex for $J$. Do there exist links $L$ and $J$ for which $\beta(L,J)>\max(\beta(L),\beta(J))$?
\end{question}
\subsection{Organization of paper.}
In Section \ref{sect: C-complexes} we state formally the definition of a C-complex and what it means for two C-complexes to be equivalent. We then study the relationship between linking numbers and C-complexes. In subsection \ref{sect:2-components} we prove Theorem~\ref{thm: 2-components}. In Section~\ref{sect:prelim} we recall the meaning of Milnor's triple linking number and briefly recall surgery. Finally, in Section~\ref{sect: many components} we prove Theorem~\ref{thm:main}.
\section{C-complexes and linking numbers}\label{sect: C-complexes}
We begin by recalling the definition of a $C$-complex appearing in \cite{CimFlo}.
\begin{definition}
An $n$-component C-complex $F=F_1\cup\dots\cup F_n$ is a union of compact oriented connected oriented surfaces in $S^3$ such that
\begin{enumerate}
\item For all $i$, $\partial F_i$ is a simple closed curve.
\item For $i\neq j$, $F_i\cap F_j$ is a union of embedded arcs running from a point on $\partial F_i$ to a point on $\partial F_j$. These arcs are called \textbf{clasps}. See Figure \ref{fig:clasps}.
\item For $1\le i<j<k\le n$, $F_i\cap F_j\cap F_k = \emptyset$.
\end{enumerate}
\end{definition}
Given an oriented link $L = L_1\cup\dots \cup L_n$, we say that $F=F_1\cup \dots \cup F_n$ is a C-complex for $L$ if $\partial F_i = L_i$ for $i=1,\dots n$. In \cite[Lemma 1]{Cimasoni2004}, Cimasoni proves that every link admits a C-complex.
\begin{figure}
\begin{picture}(320,100)
\put(70,0){\includegraphics[width=8em]{PosClasp.pdf}}
\put(200,0){\includegraphics[width=8em]{NegClasp.pdf}}
\end{picture}
\caption{Left: A positive clasp. Right: A negative clasp.}
\label{fig:clasps}
\end{figure}
\begin{figure}
\begin{picture}(170,120)
\put(0,0){
\includegraphics[height=12em]{BRingsAnotherCcomplex.pdf}
}
\put(0,47){$L_3$}
\put(50,25){$L_1$}
\put(170,40){$L_2$}
\put(0,85){$c_3^{-1}$}
\put(40,85){$c_4^{+1}$}
\put(70,81){$c_1^{-1}$}
\put(70,35){$c_2^{+1}$}
\end{picture}
\caption{A C-complex for the Boromean Rings with the components and clasps labeled.}
\label{fig:BR}
\end{figure}
\begin{definition}
\label{def: equiv}
Two C-complexes $F = F_1\cup \dots \cup F_n$ and $G = G_1\cup \dots \cup G_n$ in $S^3$ are called \textbf{equivalent} if there exists a homeomorphism $\Phi:F\to G$ such that
\begin{enumerate}
\item For $k=1,\dots, n$, the restriction $\Phi|_{F_k}$ is an orientation preserving homeomorphism from $F_k$ to $G_k$
\item For every clasp $\c\subseteq F_a\cap F_b$, $\Phi(c) \subseteq G_a\cap G_b$ is a clasp with the same sign as $c$.
\end{enumerate}
The homeomorphism $\Phi$ is called an \textbf{equivalence} between $F$ and $G$.
\end{definition}
\begin{remark}
If $N(F_i)$ and $N(G_i)$ are regular neighborhoods of $F_i$ and $G_i$, then the condition that the signs of the clasps agree implies that $\Phi$ extends to an orientation preserving homeomorphism from $N(F_1)\cup\dots\cup N(F_n)$ to $N(G_1) \cup\dots\cup N(G_n)$.
\end{remark}
We proceed by discussing the classification of C-complexes up to this notion of equivalence. For a C-complex $F = F_1\cup \dots \cup F_n$ bounded by $L_1,\dots, L_n$ let $g(F_k)$ be the genus of $F_k$, the $k$'th component of $F$. Since the genus of a surface is an invariant of that surface, $g(F_k)$ is an invariant of the equivalence class of $F$.
Another invariant can be seen by recording the clasps of $F$. For a C-complex $F$, let $\{c_1,\dots c_k\}$ refer to the set of clasps. Any clasp $c_i\subseteq F_a\cap F_b$ is assigned a sign $\epsilon_i = \operatorname{sign}(c_i) = \pm1$ depending on the intersection between $L_a$ and $F_b$ at $c_i$. See Figure~\ref{fig:clasps}. In order to encode the sign of the clasps in the notation, we will say $c_i^{\epsilon_i}$ in place of $c_i$. {For example, consider the C-complex for the Boromean rings of Figure \ref{fig:BR}. It has clasp set $\{c_1^{-1}, c_2^{+1}, c_3^{-1}, c_4^{+1}\}$. }
After picking a basepoint $p_k\in \partial F_k$ away from the clasps one can build a word $\omega_k(F)$ in the letters $\{c_1^{\epsilon_1},\dots,c_\ell^{\epsilon_\ell}\}$ by following the boundary $L_k$ of $F_k$ and recording $c_i^{\epsilon_i}$ whenever $L_k$ passes through the clasp $c_i^{\epsilon_i}$. We call these \textbf{claspwords}. Notice that each clasp $c_i^{\epsilon_i}\in \mathcal{C}^F$ appears in precisely two of these claspwords in which it appears once. A change of basepoint alters $\omega_k$ by a cyclic permutation. Notice that since the assignment of labels to the clasps was arbitrary, we can change $\{\omega_k(F)\}_{k=1}^n$ by any permutation of the names of the clasps.
For example, consider again the C-complex for the Boromean rings in Figure~\ref{fig:BR}. It has
$$
\omega_1(F)=c_1^{-1}c_3^{-1}c_2^{+1}c_4^{+1},~
\omega_2(F)=c_1^{-1}c_2^{+1},~
\omega_2(F)=c_3^{-1}c_4^{+1},~g(F_1)=g(F_2)=g(F_3)=0.
$$
These invariants give a complete description of the equivalence of C-complexes.
\begin{proposition}\label{prop:equiv C-comp}
Let $F=F_1\cup \dots \cup F_n$ and $G=G_1\cup \dots \cup G_n$ be $n$-component C-complexes. Then $F$ is equivalent to $G$ if and only if for all $k$, $g(F_k)=g(G_k)$ and $\omega_k(F)=\omega_k(G)$ (up to a cyclic permutation and relabeling the clasps.)
\end{proposition}
\begin{proof}
If $\Phi:F\to G$ is an equivalence, then $\Phi|_{F_k}$ is an orientation preserving homeomorphism. Since genus is an invariant of surfaces, $g(F_k)=g(G_k)$. Suppose that $\omega_k(F) = c^{\epsilon_{a_1}}_{a_1}\dots c^{\epsilon_{a_k}}_{a_k}$. Since $\Phi$ restricts to a homeomorphism from $\partial F_k$ to $\partial G_k$ and preserves the signs of the clasps, $\omega_k(F) = \Phi(c_{a_1})^{\epsilon_{a_1}}\dots \Phi(c_{a_k})^{\epsilon_{a_k}}$. Thus, up to a relabeling of the clasps, $\omega_k(F) = \omega_k(G)$.
Now suppose that for some labeling of the clasps, some choice of basepoints and all $k$, $\omega_k(F)=\omega_k(G)$ and $g(F_k) = g(G_k)$. Let $A_k(F)$ be a closed annular neighborhood of $\partial F_k$ containing all of the clasps in $F_k$ and $F_k^0\subseteq F_k$ be the closure of the complement of $A_k(F)\subseteq F_k$. Since $\omega_k(F)=\omega_k(G)$, there is a homeomorphism $\Phi^A_k: A_k(F)\to A_k(G)$ preserving the clasps. Since $F_k^0$ and $G_k^0$ are surfaces with the same genus and one boundary component each, the restriction $(\Phi^A_k)|_{\partial F_k^0}$ extends to a homeomorphism $F^0_k\to G^0_k$. Thus, we have a homeomorphism $\Phi_k:F_k\to G_k$.
The map $\Phi$ given by $\Phi(x)= \Phi_k(x)$ for $x\in F_k$ gives an equivalence between $F$ and $G$.
\end{proof}
If $L = L_1\cup\dots \cup L_n$ is an $n$-component link with C-complex $F_1\cup \dots \cup F_n$, then it is clear that $g(F_k)$ and $\omega_k(F)$ are definitely not invariants of $L$. For instance, the move of Figure \ref{fig:clasppass} merely isotopes the underlying link, yet alters both the claspword and the genus. In order to build an obstruction theory to a pair of links admitting equivalent C-complexes, we find quantities depending on the claspwords and genera of a C-complex which are invariants of the link. One such quantity is the linking number $\operatorname{lk}(L_i, L_j)$. Recall that $\operatorname{lk}(L_i, L_j)$ can be computed in terms of bounded surfaces. See for example, \cite[Chapter 5, Section D]{Rolfsen}. In particular, if $F_j$ is a surface bounded by $L_j$ then $\operatorname{lk}(L_i, L_j)$ is given by counting the number of positive intersections between $L_i$ and $F_j$ and then subtracting the number of negative intersections. This is the same as counting the number of positive clasps in $F_i\cap F_j$ and then subtracting the number of negative clasps. This clearly depends only on the claspword $\omega_i(F)$.
\subsection{The proof of Theorem~\ref{thm: 2-components}}\label{sect:2-components}
In this subsection, we prove that if a pair of 2-component links have the same linking number, then these links admit equivalent C-complexes.
\begin{figure}
\begin{picture}(320,100)
\put(70,0){\includegraphics[width=8em]{ClaspPassBefore.pdf}}
\put(170,40){\huge{$\rightsquigarrow$}}
\put(200,0){\includegraphics[width = 10em]{ClaspPassAfter.pdf}}
\end{picture}
\caption{By isotoping $L=L_1 \cup L_2$, we modify $\omega_1$ by the transposition of two consecutive clasps. Notice that this increases the genus of $F_2$ by 1. }\label{fig:clasppass}
\end{figure}
\begin{reptheorem}{thm: 2-components}
Let $L=L_1 \cup L_2$ and $J=J_1 \cup J_2$ be 2-component links. Then $\operatorname{lk}(L_1 , L_2) = \operatorname{lk}(J_1,J_2)$ if and only if $L$ and $J$ admit equivalent C-complexes.
\end{reptheorem}
\begin{proof}
That the linking number provides an obstruction is trivial, as the linking number can be computed in terms of the number of positive and negative clasps. It suffices then to suppose that $\operatorname{lk}(L_1,L_2) = \operatorname{lk}(J_1,J_2)$ and construct equivalent C-complexes for $L$ and $J$. Consider any two C-complexes $F = F_1\cup F_2$ and $G=G_1\cup G_2$ for $L$ and $J$ respectively. We begin by modifying these C-complexes so that their claspwords become very simple.
Suppose that $F$ has $m$ positive and $n$ negative clasps. Then $F$ has a total of $k=m+n$ clasps. Since each of these clasps must then be between $F_1$ and $F_2$, every clasp appears once in $\omega_1(F)$ and once in $\omega_2(F)$. By labeling the clasps according to the order they appear in $L_1$ we arrange that
$$\omega_{1}(F)=c_{1}^{\epsilon_1}\dots c_k^{\epsilon_k}
\text{ and }
\omega_{2}(F) = c_{\sigma(1)}^{\epsilon_{\sigma(1)}}\dots c_{\sigma(k)}^{\epsilon_{\sigma(k)}}$$
for some permutation $\sigma$.
Consider the move of Figure~\ref{fig:clasppass}. It transposes two adjacent clasps in $\omega_1(F)$ and does not change $\omega_2(F)$. Therefore, $\omega_{1}(F)$ can modified to read $c_{\rho(1)}^{\epsilon_{\rho(1)}} \dots c_{\rho(k)}^{\epsilon_{\rho(k)}}$ for any permutation, $\rho$. Pick $\rho$ such that $\epsilon_{\rho(1)} = \dots = \epsilon_{\rho(m)} = +1$ and $\epsilon_{\rho(m+1)} = \dots = \epsilon_{\rho(k)} = -1$. By relabeling the clasps we now have that
$$\omega_{1}(F)=c_{1}^{+1}\dots c_{m}^{+1}c_{m+1}^{-1} \dots c_k^{-1}.$$
We can now similarly permute $\omega_2(F)$ without altering $\omega_1(F)$ until
$$\omega_{2}(F)=c_{1}^{+1}\dots c_{m}^{+1}c_{m+1}^{-1}\dots c_k^{-1}.$$
Similarly we arrange that $$\omega_{1}(G) = \omega_2(G) = c_{1}^{+1}\dots c_{m'}^{+1}c_{m'+1}^{-1} \dots c_{k'}^{-1}$$
for some $m',k'\in \mathbb{N}$
By modifying $F$ or $G$ as in Figure~\ref{fig:SL}, we increase the number of positive and negative clasps until $m=m'$. Since $L$ and $J$ have identical linking numbers, it must then follow that $F$ and $G$ have the same number of negative clasps also. Thus, we have that $k=k'$. Notice then that $\omega_1(F)=\omega_1(G)$ and $\omega_2(F)=\omega_2(G)$. Finally, using the modification in Figure~\ref{fig:Genus} we may assume that $g(F_1)=g(G_1)$ and $g(F_2)=g(G_2)$. Proposition~\ref{prop:equiv C-comp} now allows us to conclude that $F$ and $G$ are equivalent.
\begin{figure}
\begin{picture}(320,100)
\put(70,0){\includegraphics[height=8em]{PreSL.pdf}}
\put(190,40){\huge{$\rightsquigarrow$}}
\put(215,0){\includegraphics[height=8em]{SL.pdf}}
\end{picture}
\caption{Addition of a pair of canceling clasps.}
\label{fig:SL}
\end{figure}
\end{proof}
\section{Preliminaries to the proof of Theorem~\ref{thm:main}}\label{sect:prelim}
In this section we recall the concepts needed for our study of C-complexes for links with more than two components. More precisely, we recall a reformulation of Milnor's triple linking number due to Mellor and Melvin \cite{MellorMelvin2003}. We go on to informally discuss the notion of surgery. Since our proof also makes use of the Arf invariant of knots and the Sato-Levine invariant of 2-component links we recall these also in this section.
\subsection{Milnor's triple linking number}\label{sect:triple linking}
Let $J=J_1 \cup J_2 \cup J_3$ be a 3 component link admitting a C-complex $F=F_1\cup F_2\cup F_3$ with claspwords $\omega_1(F)$, $\omega_2(F)$, $\omega_3(F)$. Out of these claspwords we generate new words $u_1(F)$, $u_2(F)$, and $u_3(F)$ in the variables $x_1^{\pm1}, x_2^{\pm1}, x_3^{\pm1}$ as follows. If $c_\ell^{\epsilon_\ell}$ is a clasp between $F_i$ and $F_j$ then in $\omega_i(F)$, replace $c_\ell^{\epsilon_\ell}$ with $x_j^{\epsilon_\ell}$ and in $\omega_j(F)$ replace $c_\ell^{\epsilon_\ell}$ with $x_i^{\epsilon_\ell}$. The words $u_1(F)$, $u_2(F)$, and $u_3(F)$ are obtained from $\omega_1(F)$, $\omega_2(F)$, and $\omega_3(F)$ by making this replacement for every clasp.
Finally, consider $u_1(F)$, $u_2(F)$, and $u_3(F)$ and take their Magnus expansions. That is, construct the formal power series in non-commuting variables $M_1(F), M_2(F), M_3(F)\in \mathbb{Z}[[h_1,h_2,h_3]]$ by sending $x_i\mapsto (1+h_i)$ and $x_i^{-1}\mapsto (1-h_i+h_i^2-h_i^3\dots)$. For $\{i,j,k\}=\{1,2,3\}$, define $\epsilon_{ijk}(F)\in \mathbb{Z}$ as the coefficient in front of $h_ih_j$ in $M_k(F)$.
\begin{definition}[See Theorem 1 of \cite{MellorMelvin2003}]\label{defn:mu123}
For the 3-component link $J = J_1 \cup J_2 \cup J_3$ admitting C-complex $F=F_1\cup F_2\cup F_3$, Milnor's triple linking number is given by
$$
\overline{\mu}_{123}(J)=\epsilon_{123}(F)+\epsilon_{312}(F)+\epsilon_{231}(F).
$$
It is well defined as an invariant of $L$ modulo the greatest common divisor of $\operatorname{lk}(J_1,J_2)$, $\operatorname{lk}(J_1,J_3)$, and $\operatorname{lk}(J_2,J_3)$. For an $n$-component link $L = L_1 \cup L_2 \dots \cup L_n$ and any three distinct numbers $i,j,k\in \{1,\dots,n\}$ consider the 3-component sub-link $L_i \cup L_j \cup L_k$. Define
$$
\overline{\mu}_{ijk}(L)=\overline{\mu}_{123}(L_i \cup L_j \cup L_k).
$$
\end{definition}
The triple linking number satisfies that
$
\overline{\mu}_{ijk}(L)=
\overline{\mu}_{jki}(L)=
\overline{\mu}_{kij}(L)=
-\overline{\mu}_{ikj}(L)=
-\overline{\mu}_{kji}(L)=
-\overline{\mu}_{jik}(L)
$,
so that for our purposes it will suffice to restrict to $1\le i<j<k\le n$.
Notice that Definition~\ref{defn:mu123} depends only on the claspwords, $\omega_1(F)$, $\omega_2(F)$ and $\omega_3(F)$. In turn, these depend only on the equivalence class of the C-complex, $F$. As a consequence we see the the triple linking number provides an obstruction to 3-component links admitting equivalent C-complexes.
\begin{repproposition}{Prop:mu123}
Let $L$ and $J$ be $n$-component links. If $L$ and $J$ admit equivalent C-complexes, then for all $1\le i<j<k$, $\overline{\mu}_{ijk}(L) = \overline{\mu}_{ijk}(J)$ (modulo the greatest common divisor of $\operatorname{lk}(L_i, L_j)$, $\operatorname{lk}(L_i, L_k)$ and $\operatorname{lk}(L_j, L_k)$).
\end{repproposition}
\begin{example}
For the sake of illustration, we perform this computation for the C-complex $F$ for $B$, the Boromean rings depicted in Figure~\ref{fig:BR}.
It has clasp-words
$$
\omega_1(F)=c_1^{-1}c_3^{-1}c_2^{+1}c_4^{+1}, ~
\omega_2(F)=c_1^{-1}c_2^{+1},~
\omega_3(F)=c_3^{-1}c_4^{+1}.
$$
Performing the replacement described in Definition~\ref{defn:mu123} we get
$$
u_1(F)=x_2^{-1}x_3^{-1}x_2^{+1}x_3^{+1}, ~
u_2(F)=x_1^{-1}x_1^{+1}=1,~
u_3(F)=x_1^{-1}x_1^{+1}=1
$$
and Magnus expansions
$$
\begin{array}{l}
M_1(F)=(1-h_2+\dots)(1-h_3+\dots)(1+h_2)(1+h_3), \\
M_2(F)=1,~
M_3(F)=1.
\end{array}
$$
Expanding,
$$
M_1(F)=1-h_2^2-h_3^2-h_3h_2+h_2h_3+\dots, ~
M_2(F)=1, ~
M_3(F)=1.
$$
The ellipses denote terms of degree at least $3$. Hence,
$$
\epsilon_{123}=0,~\epsilon_{312}=0,~\epsilon_{231}=1
$$
and $\overline\mu_{123}(B)=1$.
\end{example}
Since the three component unlink has $\overline{\mu}_{123}(U)=0$ we have the following corollary
\begin{corollary}
There does not exist a C-complex for the Boromean Rings which is equivalent to a C-complex for the three component unlink.
\end{corollary}
\subsection{Surgery} Given a link $\gamma = \gamma_1,\dots,\gamma_k$ and a sequence of rational numbers in reduced form $\frac{p_1}{q_1},\dots, \frac{p_k}{q_k}$ one can form a new 3-dimensional manifold by cutting out tubular neighborhoods of the curves $\gamma_1,\dots, \gamma_k$ and gluing in their place solid tori $V_1,\dots, V_k$ such that for $i=1,\dots, k$ the meridian of $V_i$ is identified to the curve consisting of $p_i$ meridians of $\gamma_i$ and $q_i$ longitudes. For a more complete description of surgery the reader is directed to \cite[Chapter 9]{Rolfsen}.
Given a link $L$ disjoint from the surgery curves $\gamma$ of the preceding paragraph then one sees a new link by taking the image of $L$ in the space resulting from the surgery. If the resulting link is isotopic to $J$, then we say that $J$ is \textbf{obtained} from $L$ by surgery along $\gamma$.
\begin{figure}
\begin{picture}(320,100)
\put(240,80){-1}
\put(240,5){-1}
\put(70,0){\includegraphics[height=8em]{trefoil.pdf}}
\put(200,0){\includegraphics[height=8em]{trefoilViaSurgery.pdf}}
\end{picture}
\caption{Left: The trefoil knot. Right: Performing -1 surgery on these curves in the complement of an unknot produces the trefoil.}
\label{fig:Trefoil}
\end{figure}
\begin{figure}
\begin{picture}(350,100)
\put(105,25){$\dots$}
\put(94,0){\small{n-full twists}}
\put(0,5){\includegraphics[height=8em]{nTwistedWhitehead.pdf}}
\put(180,5){\includegraphics[height=8em]{WhiteheadSurgery.pdf}}
\put(270,-5){$-1/n$}
\end{picture}
\caption{Left: The $n$-twisted Whitehead link. Right: $\frac{-1}{n}$ surgery along this curve sends the unlink to the $n$-twisted Whitehead link. }
\label{fig:Whitehead}
\end{figure}
Many interesting surgeries on $S^3$ produce $S^3$. There are two of particular interest to us in this paper. First, if $\gamma$ is unknotted then $\frac{-1}{n}$ surgery along $\gamma$ produces $S^3$ and puts $n$ positive full twists in the strands of $L$ passing through a disk bounded by $\gamma$ \cite[Chapter 9]{Rolfsen}. For example, the trefoil of Figure \ref{fig:Trefoil} is obtained from the unknot by $-1=\frac{-1}{1}$ surgery along two curves in the complement. Similarly, the $n$-twisted Whitehead link of Figure~\ref{fig:Whitehead} is obtained from the unlink by a $\frac{-1}{n}$ surgery. The other move of interest to us is the band pass move of Figure~\ref{fig:BPSurgery}, which can be obtained by $0=\frac{0}{1}$ surgery along two Hopf linked curves. While this result is well known the proof in its full detail can be found in the proof of \cite[Lemma 1]{MartinThesis}.
\begin{figure}
\begin{picture}(320,100)
\put(120,76){0}
\put(115,10){0}
\put(70,0){\includegraphics[width=8em]{PreBandPassNoCComplexSurgeryCurves.pdf}}
\put(200,0){\includegraphics[width=8em]{AfterBandPassNoCComplex.pdf}}
\end{picture}
\caption{Left: A Hopf link in the complement of a link. Right: Performing $0$-surgery along the components of this Hopf link accomplishes a band pass. }
\label{fig:BPSurgery}
\end{figure}
\subsection{The Arf invariant and the Sato-Levine invariant}
The Arf invariant is an invariant of knots which takes values in $\mathbb{Z}/2$. For a complete discussion of the Arf invariant, see \cite[Chapter X]{OnKnots}. To see a rigorous discussion on the Sato-Levine invariant for a 2-component link in terms of Milnor's invariant the reader is directed to \cite{C4}. Our work will require the following elementary properties of these invariants.
\begin{proposition}\label{Arf}
For any knot $J$, let the knot $J'$ be obtained by band summing with the Trefoil. Then the Arf-invariant satisfies that $\operatorname{Arf}(J') = \operatorname{Arf}(J)+1$.
\end{proposition}
\begin{proposition}\label{Mu}
Let $\operatorname{WH}_n$ be the $n$-twisted Whitehead link of of Figure \ref{fig:Whitehead} and let $L=L_1\cup L_2$ be a 2-component link with vanishing pairwise linking number. Let $L'$ be the 2-component link obtained by band summing with $\operatorname{WH}_n$. Then the Sato-Levine invariant satisfies that $\overline{\mu}_{1122}(L') = \overline{\mu}_{1122}(L)+n$.
\end{proposition}
For an $n$-component link $L=L_1\cup \dots\cup L_n$, $\overline{\mu}_{iijj}(L) = \overline{\mu}_{1122}(L_i, L_j)$ is the Sato-Levine invariant of the 2-component sublink $L_i, L_j$.
According to \cite[Theorem 1]{MartinThesis}, the Arf-invaraint, the Sato-Levine invariant and the triple linking number are a complete set of obstructions to a pair of links being related by a sequence of band pass moves, as in Figure~\ref{fig:BPSurgery}.
\begin{figure}
\begin{picture}(350,100)
\put(0,5){\includegraphics[height=8em]{UnLinkCcplx.pdf}}
\put(290,25){$\dots$}
\put(270,0){$n$ full twists}
\put(180,5){\includegraphics[height=8em]{nTwistedWhiteheadCcplx.pdf}}
\end{picture}
\caption{Left: A C-complex for the 2-component unlink. Right: An equivalent C-complex for the $n$-twisted Whitehead link. }
\label{fig:WhiteheadCplx}
\end{figure}
\section{The proof of Theorem \ref{thm:main}. }\label{sect: many components}
We close this paper with the proof of Theorem \ref{thm:main}.
\begin{reptheorem}{thm:main}
Let $L$ and $J$ be $n$-component links with vanishing pairwise linking numbers. Then the following are equivalent
\begin{enumerate}
\item For all $1\le i<j<k\le n$, $\overline{\mu}_{ijk}(L)=\overline{\mu}_{ijk}(J)$.
\item $L$ and $J$ admit equivalent C-complexes.
\item There exist curves $\gamma_1,\dots,\gamma_k$ disjoint from $L$ such that $\operatorname{lk}(L_i,\gamma_j)=0$ for all $i,j$ and such that $J$ is obtained from $L$ after performing some surgery on $\gamma_1,\dots,\gamma_k$.
\end{enumerate}
\end{reptheorem}
\begin{proof}
The claim that (2) implies (1) is the content of Proposition~\ref{Prop:mu123}. We begin by showing that (3) implies (2). Let $L=L_1,\dots, L_n$ and $J=J_1,\dots, J_n$ be $n$-component links. Suppose that there exist curves $\gamma=\gamma_1,\dots,\gamma_k$ in the complement of $L$, for which $\operatorname{lk}(L_i,\gamma_j)=0$ and a sequence of rational numbers $p_1/q_1,\dots p_k/q_k$ such that $J$ is obtained by modifying $L$ by $p_j/q_j$ surgery along $\gamma_j$ for all $1\le j \le k$.
Let $F$ be a C-complex for $L$. Consider any component $F_i$ of $F$ and $\gamma_j$ of $\gamma$. Since $\operatorname{lk}(L_i, \gamma_j)=0$ by assumption, either $\gamma_j$ is disjoint from $F_i$ or there exists some pair of intersection points $p,q\in F_i\cap \gamma_j$ with opposite sign such that the arc $\alpha\subseteq \gamma_j$ running from $p$ to $q$ is disjoint from $F_i$. We proceed to modify $F_i$ using a tube following $\alpha$ from $p$ to $q$. See Figure~\ref{fig: stabilize to disjoint}. This modification may introduce some simple closed curves in the intersection $F_i\cap F_\ell$ for some $\ell\neq i$. The resulting collection of surfaces is no longer a C-complex. By pushing $F_i$ and $F_\ell$ along some arcs as in Figure \ref{fig: fix double loop}, we replace each intersection circle with a pair of clasps. In doing so we merely isotope $L_i$ and $L_\ell$.
\begin{figure}[b]
\begin{picture}(190,105)
\put(0,5){\includegraphics[height=8em]{SurgeryCurve1.pdf}}
\put(6,8){$\gamma_j$}
\put(-13,40){$F_i$}
\put(35,98){$F_\ell$}
\put(30,0){(a)}
\put(120,5){\includegraphics[height=8em]{SurgeryCurve2.pdf}}
\put(126,8){$\gamma_j$}
\put(107,40){$F_i$}
\put(155,98){$F_\ell$}
\put(150,0){(b)}
\end{picture}
\caption{(a) $\gamma_j$ intersects of $F_i$ in two points with opposite sign. (b) Stabilizing $F_i$ removes these intersections, but adds a simple closed curve in $F_i\cap F_\ell$.}
\label{fig: stabilize to disjoint}
\end{figure}
\begin{figure}
\begin{picture}(330,105)
\put(0,10){\includegraphics[height=8em]{DoubleLoop.pdf}}
\put(-5,28){$F_i$}
\put(-5,70){$F_j$}
\put(30,0){(a)}
\put(120,10){\includegraphics[height=8em]{DoubleArc.pdf}}
\put(115,28){$F_i$}
\put(115,70){$F_j$}
\put(130,0){(b)}
\put(240,10){\includegraphics[height=8em]{TwoClasps.pdf}}
\put(235,28){$F_i$}
\put(235,70){$F_j$}
\put(230,0){(c)}
\end{picture}
\caption{(a) A double loop. The darker red here indicates the opposite side of $F_j$ (b) A ``finger move'' replaces this loop with an arc with both endpoints on $L_j$. (c) Another ``finger move'' reduces the arc to two clasps.}
\label{fig: fix double loop}
\end{figure}
By assumption, the link $J$ is obtained by starting with $L$, removing neighborhoods of $\gamma_1,\dots, \gamma_k$ from $S^3$, and gluing back in solid tori $V_1, \dots, V_k$ such so that the curve given by $p_i$ longitudes of $\gamma_i$ and $q_i$ meridians bounds the meridional disk for $V_i$. Since $F$ is disjoint from $\gamma$, this cut and paste process preserves the equivalence class of $F$. We have produced a C-complex for $J$ equivalent to a C-complex for $L$, as required by the theorem.
It remains to prove that (1) implies (3). Suppose that for all $1\le i<j<k\le n$, $\overline{\mu}_{ijk}(L)=\overline{\mu}_{ijk}(J)$. For $i=1,\dots, n$ if $\operatorname{Arf}(L_i)\neq \operatorname{Arf}(J_i)$ then modify $L_i$ by band summing with the trefoil knot. Call the resulting link $L^0$. By Proposition~\ref{Arf}, $\operatorname{Arf}(L_i^0) = \operatorname{Arf}(L_i)+1=\operatorname{Arf}(J_i)$, since the Arf invariant takes values in $\mathbb{Z}/2$. Notice that as in Figure~\ref{fig:Trefoil} each band sum with the trefoil may be obtained by first band summing with the unknot (which does not change the link type of $L$) and then performing $-1$ surgery along two curves each of which have zero linking number with every component of $L$.
Next, for all $1\le i\le j\le n$, if $\overline\mu_{iijj}(L^0)\neq \overline\mu_{iijj}(J)$ then we band sum the components $L_i^0$ and $L_j^0$ of $L^0$ with the $m_{ij}$ twisted Whitehead link where $m_{ij}=\overline\mu_{iijj}(J) - \overline\mu_{iijj}(L^0)$. Call the resulting link $L^1$. By Proposition~\ref{Mu},
$
\overline{\mu}_{iijj}(L^1) = \overline{\mu}_{iijj}(L^0) + m_{ij} =
\overline{\mu}_{iijj}(J).
$
This band sum can be obtained by first band summing with the 2-component unlink (which does not change the link type of $L^0$) and then performing $\frac{-1}{m_{ij}}$ surgery along a single curve as in Figure~\ref{fig:Whitehead} which has zero linking numbers with all components of $L^0$.
By design, we now have that for all $i,j,k$, $\overline{\mu}_{ijk}(L^1) =\overline{\mu}_{ijk}(L) = \overline{\mu}_{ijk}(J)$, $\operatorname{Arf}(L_i^1)=\operatorname{Arf}(J_i)$ and $\overline\mu_{iijj}(L^1) = \overline{\mu}_{iijj}(J)$. According to \cite[Theorem 1]{MartinThesis}, $L^1$ and $J$ are be related by a sequence of band pass moves, depicted in Figure~\ref{fig:BPSurgery}. Thus, there exist a collection of curves each of which has zero linking number with every component of $L^1$ such that $J$ is the result of modifying $L^1$ via $0$-surgery along these curves.
Let $\gamma$ be the collection of all of the curves of the preceding three paragraphs. We see that $J$ can be obtained by modifying $L$ by surgery along these curves each of which has zero linking number with every component of $L$. This completes the proof.
\end{proof}
\bibliographystyle{plain}
|
\section{Introduction}
Ongoing experiments on relativistic heavy-ion collisions at high energies, like the Relativistic Heavy Ion Collider
(RHIC) and the Large Hadron Collider (LHC), have been designed to reach a new state of matter known as quark and gluon
plasma (QGP). It is a deconfined strongly-interacting plasma behaving like an almost perfect fluid. The bulk properties of this
high-temperature phase are governed by the light quarks and gluons. However, charm and bottom quarks (collectively denoted as heavy quarks) are
responsible for several observables which are essential to probe the QGP properties. The reason is that these heavy quarks are witnesses of
the entire plasma evolution as they are produced in the initial hard scatterings and remain abiding until hadronization.
In their final state they appear as constituents of heavy hadrons, mainly $D$ and $B$ mesons. Indeed, these states have generated significant interest in the
recent past because they serve as indicators of QGP dynamics~\cite{Prino:2016cni} due to the suppression of their momentum distribution at large $p_T$ in the thermal medium, reflected
in a low nuclear suppression factor $R_{AA}$, and a sizable value of the elliptic flow $v_2$, a measure of the azimuthal
anisotropy in the plasma. Noticeably $R_{AA}$ and $v_2$ have similar values like the light hadrons.
Thanks to the last upgrades in the experimental detectors, RHIC and LHC can reconstruct $D$ mesons from their hadronic decay products (like $D^0 \rightarrow K^- \pi^+$),
instead of collecting nonphotonic electrons coming from semileptonic decays. With recent experimental results from STAR (RHIC)~\cite{Adamczyk:2014uip} as well
as ALICE (LHC)~\cite{ALICE:2012ab,Abelev:2013lca,Abelev:2014ipa,Adam:2015sza}, one can now contrast the predictions of the theory groups, which have computed
the $R_{AA}$ and $v_2$ of heavy mesons, using numerical simulations for the heavy-ion evolution under different models~\cite{Linnyk:2008hp,Das:2010tj,
Mazumder:2011nj,Lang:2012yf,He:2012df,He:2011qa,Das:2013kea,Alberico:2011zy,Uphoff:2011ad,Uphoff:2012gb,Cao:2013ita,Cao:2015hia,
Gossiaux:2008jv,Nahrgang:2013xaa,Nahrgang:2014vza,Ozvenchuk:2014rpa,Song:2015sfa,Song:2015ykw}.
In addition to heavy mesons, future upgrades in the ALICE detector will allow to study $\Lambda_c$ and $\Lambda_b$ baryons within the so-called Run 3 of
LHC~\cite{Abelevetal:2014dna,Andronic:2015wma,Dainese:2016dea} (see~\cite{Aaij:2013mga} for a recent study on the $\Lambda_c$ baryon
reconstruction in $p+p$ collisions by the LHCb
collaboration). As presented in Ref.~\cite{Abelevetal:2014dna}, the ALICE collaboration plans to study several observables related to $\Lambda_c$ baryons, namely
the $R_{AA}$, $v_2$ and $\Lambda_c/D$ ratio. Given some key upgrades in the ALICE detector capabilities, the $\Lambda_b$ physics in heavy-ion collisions has
also been considered for the Run 3~\cite{Abelevetal:2014dna,Andronic:2015wma}.
These experimental advances on heavy baryons ($\Lambda_c$ and $\Lambda_b$) are of general interest because they will allow us to
have novel information on the hadronization mechanism and, more specifically, on the evaluation of the heavy
baryon-to-meson ratio ~\cite{Oh:2009zj,Lee:2007wr}. In the light and strange sectors, this ratio has shown an anomalous enhancement
with respect to $p+p$ collisions.
Also an enhancement would affect the $R_{AA}$ of non-photonic electrons~\cite{Sorensen:2005sm,Ayala:2009pe}. This is because the branching
ratio of heavy-baryon decay~\cite{MartinezGarcia:2007hf} to electrons is smaller than the branching ratio of heavy meson
to electrons. Furthermore the heavy baryon to meson ratio, ($\Lambda_c/D$ and $\Lambda_b/B$), is very
fundamental for the understanding of the in-medium hadronization~\cite{Greco:2003vf,Fries:2008hs}.
In a recent work~\cite{Tolos:2016slr} some of us have extensively studied the microscopical details of the $\Lambda_c$ and
$\Lambda_b$ interactions with light mesons, such as $\pi,K,\bar{K},\eta$ (see also Ref.~\cite{Ghosh:2014oia} for a
first study in this direction). In Ref.~\cite{Tolos:2016slr} the authors used an effective field theory at low energies to
describe the hadronic interactions, which are, in addition, unitarized to account for the required unitarity property of the
scattering amplitudes. They presented the typical cross sections for both $\Lambda_c$ and $\Lambda_b$ baryons, containing
many resonant states. Then, the authors computed the relevant transport coefficients, drag force and diffusion coefficients, as a function of the
temperature and heavy baryon momentum for the conditions expected after the hadronization in high-energy heavy ion collisions.
Either the individual cross sections, or the
transport coefficients themselves, can be readily used in transport simulations to account for the $R_{AA}$ or $v_2$ of heavy
baryons at heavy-ion collisions at RHIC and LHC.
As an application of the findings in~\cite{Tolos:2016slr}, we now present predictions for
$R_{AA}$ as well as $v_2$ of $\Lambda_c$ and $\Lambda_b$ baryons for RHIC and LHC energies for an eventual
comparison with experimental results and to understand
if we can describe both the heavy meson and heavy baryon observables
simultaneously. We also present prediction for the
$\Lambda_c/D$ ratio for RHIC and LHC energies. We will accommodate a Langevin equation
for the momentum evolution of the heavy particles (equivalent to a Fokker-Planck realization) whose parameters are related to the
drag and diffusion coefficients. These are taken from a quasiparticle model~\cite{Das:2015ana} for the heavy quark
propagation, and from Ref.~\cite{Tolos:2016slr} for the hadronic phase.
This manuscript is organized as follows. In Sec.~\ref{sec:langevin} we describe the Langevin equations to be solved for the dynamics of the heavy particle.
In particular, we explain how the coefficients of the equations of motion are related to the transport coefficients computed in our previous work~\cite{Tolos:2016slr}.
In Sec.~\ref{sec:model} we provide some details about the practical implementation of our model: we describe our prescription for the initial state, the quasiparticle
model used for the propagation of heavy quarks in the hot plasma, the hadronization mechanism for the confined phase transition, and the freeze-out condition.
Our results are presented in Sec.~\ref{sec:observables}, where we give our predictions for $R_{AA}$~(\ref{sec:raa}) and $v_2$~(\ref{sec:v2}).
Sec.~\ref{sec:ratio} is devoted for heavy baryon to meson ratio. Finally, we draft our conclusions in Sec.~\ref{sec:conclusions}.
\section{Langevin equation for heavy particles~\label{sec:langevin}}
The standard approach to heavy-quark dynamics in the QGP and the propagation of open-heavy hadrons in the hadronic medium is to
follow their evolution by means of a Fokker-Planck equation solved stochastically by the Langevin equations~\cite{Prino:2016cni}.
The relativistic Langevin
equations of motion for the time evolution of the position and momentum of the
heavy quarks/heavy hadrons can be written in the form
\begin{equation} \left\{ \begin{array}{rcl}
dx_i & = & \frac{p_i}{E}dt \ , \\
dp_i & = & -F(p) p_i dt+C_{ij}(p)\rho_j\sqrt{dt} \ ,
\label{lv1} \end{array}
\right.
\end{equation}
where $dx_i$ and $dp_i$ are the shift of the coordinate and momentum in each discrete time step $dt$.
$F(p)$ and $C_{ij}(p)$ are the drag force and the
covariance matrix respectively. $\rho$ is the noise which obeys the probability distribution of independent Gaussian-normal
distributed random variables, $P(\rho)=(2\pi)^{-3/2}e^{-\rho^2/2}$, along with
the relations $<\rho_i \rho_j>=\delta_{ij}$ and $<\rho_i>=0$.
The covariance matrix is related to the transverse and longitudinal diffusion coefficients,
\begin{eqnarray}
C_{ij}=\sqrt{2\Gamma_0(p)} \Delta_{ij}+\sqrt{2\Gamma_1(p)} \ \frac{p_i p_j}{p^2} \ ,
\label{cmmm}
\end{eqnarray}
where $\Delta_{ij} = \delta_{ij}-p_i p_j/p^2$ is the transverse projector operator.
Under the assumption, $\Gamma_0 (p)=\Gamma_1 (p)=\Gamma (p)$,
Eq~(\ref{cmmm}) reduces to $C_{ij}=\sqrt{2 \Gamma(p)} \delta_{ij}$. Such an assumption, strictly valid
for $p\rightarrow 0$, is usually employed at finite $p$ for heavy quark dynamics
in the QGP~\cite{Moore:2004tg,vanHees:2005wb,Cao:2011et,He:2012df,Das:2010tj,Mazumder:2011nj,Lang:2012yf}.
With the knowledge of $F(p)$ and $\Gamma(p)$ as functions of $T$ and $p$, the Langevin equation is ready to be solved.
We use pre-Ito discretization scheme for the numerical implementation of the Langevin dynamics.
\section{Dynamical model~\label{sec:model}}
To solve the Langevin equation in the QGP/hadronic phase one needs the drag and diffusion coefficients of
heavy quarks/heavy baryons as a function of temperature and momentum
in the QGP/hadronic medium. The drag and diffusion coefficients of the heavy quarks in the QGP are calculated
inspired by the quasi-particle model
(QPM)~\cite{Das:2012ck,Berrehrah:2013mua,Berrehrah:2014kba}. The quasi-particle approach accounts for
the non-perturbative dynamics
by means of temperature-dependent quasi-particle masses for light quarks and gluons, respectively,
\begin{eqnarray} m^2_q & = & \frac{2N_c + N_f}{12} g^2(T) T^2 \ , \\
m^2_g & = & \frac{N_c^2-1}{8N_c} g^2(T) T^2 \ , \end{eqnarray}
as well as a $T$-dependent background field known as bag constant.
The strong coupling constant is obtained by a fit of the lattice energy density and is parametrized as follows:
\begin{equation} g^2(T)= \frac{48 \pi^2}{(11N_c-2N_f) \ln \left[ \lambda \left(T/T_c-T_s/T_c \right) \right]^2}\ , \end{equation}
with $N_c=N_f=3$, $\lambda=2.6$ and $T_s/T_c=0.57$~\cite{Plumari:2011mk}.
The quasi-particle scheme is able to successfully reproduce the thermodynamics of lattice-QCD~\cite{Plumari:2011mk}
by fitting the strong coupling $g(T)$. For the evaluation of the drag and diffusion coefficients in the QGP medium, we
use the QPM approach recently addressed in Ref.~\cite{Das:2015ana} to describe heavy quark $R_{AA}$ and $v_2$
at RHIC and LHC energies. A self-consistent dynamical treatment should include the finite width of the quasiparticle, however
the drag and diffusion are not significantly affected, see Ref~\cite{Berrehrah:2013mua,Berrehrah:2014kba}.
The drag has been calculated in Ref.~\cite{Das:2015ana} and show a very mild $T$ dependence in comparison with
perturbative QCD (pQCD) or AdS/CFT. We notice that a similar dependence is found in the T-matrix approach~\cite{vanHees:2007me,Riek:2010fk}.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=17pc,clip=true]{Dx_p100_c.eps}\hspace{2pc}
\includegraphics[width=17pc,clip=true]{Dx_p100_b.eps}\hspace{2pc}
\caption{$D_x$ as a function of $T$ for $c$ quark and $\Lambda_c$ (left panel) and
for $b$ quark and $\Lambda_b$ (right panel).}
\label{figx}
\end{center}
\end{figure}
Within the Fokker-Planck approach, the spatial diffusion coefficient~\cite{Moore:2004tg,Abreu:2011ic}, $D_x$,
can be calculated in the static limit ($p\rightarrow 0$) in two different ways. It can be obtained from the diffusion
coefficient in momentum space,
\begin{equation} \label{eq:DxG} D_x = \frac{T^2}{\Gamma} \ , \end{equation}
or from the drag coefficient using the Einstein relation ($\Gamma=MFT$),
\begin{equation} \label{eq:DxF} D_x = \frac{T}{MF} \ . \end{equation}
However, the Einstein relation may not be strictly valid at high temperatures in the QGP phase.
Hence, we are using both approaches (\ref{eq:DxG}) and (\ref{eq:DxF}) to evaluate the spatial diffusion coefficient in the QGP phase.
In Fig.~\ref{figx} the spatial diffusion coefficient in the QGP phase~\cite{Das:2015ana} is compared with the one for heavy baryons in hadronic matter~\cite{Tolos:2016slr}.
In the left panel of this figure, we show $2\pi D_x T$ as a function of the temperature for $c$ quarks (high temperature) and $\Lambda_c$ baryons (low temperature).
We find that the $\Lambda_c$ diffusion coefficient also supports a continuous evolution with a minimum around $T_c$ like the heavy-meson case ($D$ meson)~\cite{He:2012df,Ozvenchuk:2014rpa,Berrehrah:2014tva}.
The differences of the $D_x$ in the QGP phase by the two different approaches are due to the violation of the Einstein relation. In the hadronic sector we have observed
that the Einstein relation is satisfied for all temperatures~\cite{Tolos:2016slr}.
In the right panel of Fig.~\ref{figx} we show $2\pi D_x T$ as a function of temperature for $b$ quarks (high temperature) and for $\Lambda_b$ baryons (low temperature).
In this case we also find an almost continuous evolution with a minimum around $T_c$. For the bottom case, the calculations of the spatial diffusion coefficient
using Eq.~(\ref{eq:DxG}) and Eq.~(\ref{eq:DxF}) are in better agreement than for the charm case. The reason is that due to the heavier mass of the $b$ quark, the Einstein relation is
better satisfied than the charm case (violations are more severe at high temperature), and the two ways of computing $D_x$ are practically equivalent.
Once the temperature of the QGP phase goes below $T_c$, the QGP phase give way to the hadronic phase.
In this phase heavy hadrons, produced after hadronization, suffer from collisions with light mesons.
To fully characterize the QGP phase, the impact of hadronic phase should be then taken into account. Several attempts
have been made in this direction to study the hadronic medium interaction and their impact on heavy mesons ($D$ and $B$)
observables at RHIC and LHC energies. However, little efforts have been given to the study of heavy
baryon interaction in the hadronic phase.
Here, we use the recent results in Ref.~\cite{Tolos:2016slr}, where some of us analyzed the heavy baryon
interaction with the hadronic medium consisting of light mesons ($\pi$, $K$, $\bar K$ and $\eta$) within unitarized
interactions from effective field theories that respect chiral and heavy-quark symmetries.
With these interactions, we have obtained the heavy-baryon transport coefficients (drag and diffusion) as a function
of temperature and momentum. In the present work we aim at studying the heavy baryon evolution in the hadronic
phase within the Langevin dynamics using the drag and diffusion coefficients calculated in the previous paper and highlight its
impact on several observables potentially measurable at RHIC and LHC energies, in particular on $R_{AA}$ and $v_2$.
\subsection{Initialization and heavy quark dynamics}
The solution of the Langevin equation needs a background medium describing the space-time evolution of the bulk matter.
To describe the expansion and cooling of the bulk matter and its elliptic flow $v_2(p_T)$ at both RHIC and LHC colliding
energies, we have employed a 3D+1 relativistic transport code
with an initial condition given by a standard Glauber model. Such a model allow us to describe the evolution of
a fluid with a fixed $\eta/s$ in the same way as it is done by viscous hydrodynamical simulation. For more details we refer the
reader to Refs.~\cite{Ruggieri:2013bda,Ruggieri:2013ova,Ferini:2008he,Greco:2008fs}.
In this work we have performed simulations of $Au+Au$ collisions at $\sqrt{s}= 200$ AGeV for the minimum bias. The initial
conditions for the bulk evolution in the coordinate space are given by the Glauber model.
In momentum space we use a Boltzmann-J\"uttner distribution
function up to a transverse momentum $p_T=2$ GeV, while at larger momenta mini-jet distributions as calculated within
pQCD at Next-to-leading order (NLO) order~\cite{Greco:2003xt,Greco:2003mm}.
At RHIC energies for $Au+Au$ at $\sqrt{s}=200$ AGeV, the maximum initial temperature of the
fireball at the center is $T_0=340$ MeV and
the initial time for the fireball simulations is chosen as $\tau_0=0.6$ fm/c (according to
the criterion $\tau_0 \cdot T_0 \sim 1$ and to standard setting in hydrodynamics).
We have also extended our calculation to study the heavy baryons $R_{AA}$ and $v_2$ at LHC energies
performing simulations of $Pb+Pb$ collisions
at $\sqrt{s}= 5.5$ ATeV energy. In this case the maximum initial temperature at the center of
the fireball is $T_0=610$ MeV and the initial time for the
simulations is chosen as $\tau_0\sim 1/T_0 =0.25$ fm/c. We have performed simulations for $0-20\%$ centrality class.
The heavy-quark distribution in momentum space, both for RHIC and LHC, is taken in accordance with the charm distribution
in $p+p$ collisions, calculated within Fixed Order + Next-to-Leading Log (FONLL), taken from Ref.~\cite{Cacciari:2005rk,Cacciari:2012ny},
where in the coordinate space they are distributed according to number of binary nucleon-nucleon collisions ($N_{coll}$)
from the Glauber model for both RHIC and LHC energies. We solve the Langevin dynamics to study the time evolution
of heavy quark momentum in QGP created in $Au+Au$ collision as discussed in Sec.~\ref{sec:langevin}. The interaction
between the heavy quarks and the bulk has been embedded through the drag and diffusion coefficients calculation within
the QPM approach discussed at the beginning of this section.
\subsection{Hadronization and hadronic evolution\label{sec:frag}}
Another important aspect of a heavy-ion collision is the hadronization mechanism, when heavy quarks combine
into color-neutral objects. Hadrons are formed when the temperature reaches $T=T_c=160$ MeV~\cite{Bazavov:2011nk}.
One of the basic mechanisms of hadronization, widely
considered in this context, is the fragmentation of an individual quark where the hadron momentum is a fraction $z$
of the quark momentum. For gluons and light quarks the fragmentation
functions are rather broad distributions around $z=0.5$, but for heavy quarks the fragmentation
functions become rather sharply peaked towards $z=1$.
The charm quark fragmentation for $D$ meson and $\Lambda_c$ can be described using the
Peterson fragmentation function~\cite{Pet},
\begin{equation}
f(z) \propto
\frac{1}{ z \lbrack 1- \frac{1}{z}- \frac{\epsilon_c}{1-z} \rbrack^2 } \ ,
\end{equation}
where $\epsilon_c$ is a free parameter to fix the shape of the fragmentation function in comparison
with the experimental data in $p+p$ collision. Unlike $D$ meson, the heavy baryon fragmentation function is not precisely
known as it is yet to be measured in $p+p$ collisions. The $D$ meson spectra in $p+p$ collision at RHIC energy using FONLL
calculation for
the initial charm production can be reproduced using $\epsilon_c=0.01$. The $D$ meson spectra at LHC
energy can be also reproduced using $\epsilon_c=0.01$. In the absence of the $p+p$ data for the $\Lambda_c$ production
at RHIC and LHC energies, we are using the electron-positron annihilation data to fix the shape of the $\Lambda_c$
fragmentation. In electron-positron annihilation, the $\epsilon_c$ for the $\Lambda_c$ is about a factor two larger
than the $D$ meson one~\cite{hffs}.
This means that the $\Lambda_c$ fragmentation function is softer than the $D$ meson fragmentation.
This is because $\Lambda_c$ contains one heavy quark and two light quarks, whereas
$D$ meson has one heavy quark and one light anti-quark. So in accordance with the electron-positron annihilation data,
we are using $\epsilon_c=0.02$ for the $\Lambda_c$, a factor two larger than the $D$ meson.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=17pc,clip=true]{fg.eps}\hspace{2pc}
\caption{Variation of the fragmentation function with the fraction of momentum.}
\label{fig1}
\end{center}
\end{figure}
In Fig.~\ref{fig1} we show the variation of $\Lambda_c$ fragmentation function with the fraction of momentum together with the
$D$ meson fragmentation function. As expected, the $\Lambda_c$ fragmentation function is softer than the $D$ meson as it
takes more energy to pop-up two quarks from the vacuum in the fragmentation picture.
For $\Lambda_b$ we use $\epsilon_c=0.006$, a factor two larger than the $B$-meson fragmentation function.
After the hadronization from the charm and bottom quarks to $\Lambda_c$ and $\Lambda_b$, respectively,
we solve the Langevin dynamics for the propagation of $\Lambda_c$ and $\Lambda_b$ in an hadronic bath that consists
of $\pi$, $K$, $\bar K$ and $\eta$. The interaction between the heavy baryons with the bath has been treated
within unitarized interactions based on effective field theories that respect chiral and heavy-quark symmetries.
Specifically, the interaction of $\Lambda_c$ and $\Lambda_b$ scattering off $\pi$, $K$, $\bar K$ and $\eta$ mesons is obtained
within a unitarized meson-baryon coupled-channel model that incorporates heavy-quark spin symmetry
~\cite{GarciaRecio:2008dp, Gamermann:2010zz, Romanets:2012hm,GarciaRecio:2012db,Garcia-Recio:2013gaa,Tolos:2013gta}.
This is a predictive model for four flavors including all basic hadrons (pseudoscalar and vector mesons, and $1/2^+$
and $3/2^+$ baryons) which reduces to the Weinberg-Tomozawa interaction in the sector where Goldstone bosons
are involved. This scheme has $\mbox{SU}(6)\times {\rm HQSS}$ symmetry, i.e., spin-flavour symmetry in the
light sector and HQSS in the heavy (charm/bottom) sector, and
it is consistent with chiral symmetry in the light sector. For more details of the hadronic interaction
we refer to the earlier work \cite{Tolos:2016slr}. The time evolution of heavy
baryons within the hadronic phase is continued until the temperature reaches $T_{kin}=120$ MeV~\cite{Das:2013lra}, at
the kinetic freeze out.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=17pc,clip=true]{LambdaC.eps}\hspace{2pc}
\caption{Variation of $\Lambda_c$ spectra in $p+p$ and $Pb+Pb$ collision at LHC colliding energy in arbitary normalization.}
\label{fig1_1}
\end{center}
\end{figure}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=17pc,clip=true]{LambdaB.eps}\hspace{2pc}
\caption{Variation of $\Lambda_b$ spectra in $p+p$ and $Pb+Pb$ collision at LHC colliding energy in arbitary normalization.}
\label{fig1_2}
\end{center}
\end{figure}
In Fig.~\ref{fig1_1} and Fig.~\ref{fig1_2} we show the variation of
$\Lambda_c$ and $\Lambda_b$ spectra in $p+p$ and $Pb+Pb$ at LHC colliding energies in arbitary normalization.
In the $Pb+Pb$ collision, due to interaction between heavy quarks and the bulk in the QGP phase as well as heavy baryons
and the bulk in the hadronic phase, the heavy baryons rearrange their spectra with larger population at low momentum.
\section{Results: experimental observables \label{sec:observables}}
The heavy-baryon observables which are going to be measured at LHC Run 2 and 3~\cite{Abelevetal:2014dna,Tieulent:2015wec}
are the nuclear suppression factor ($R_{AA}$) and the elliptic flow ($v_2$). We evaluate these
observables at both RHIC and LHC energies using Peterson fragmentation function as described above. One of our main motivations
is to highlight the impact of the hadronic medium rescattering on heavy baryon observables.
\subsection{Nuclear modification factor, $R_{AA}$~\label{sec:raa}}
One of the key observables related to heavy quark propagation, which is measured at RHIC and LHC energies,
is the nuclear suppression factor $R_{AA}$. It measures the depletion of
high transverse momentum ($p_T$) hadrons ($D$ and $B$ mesons) produced in nucleus+nucleus collisions with
respect to those produced in
proton+proton collisions scaled with the number of binary collision.
The ALICE physics programme for Runs 3 and 4~\cite{Abelevetal:2014dna,Tieulent:2015wec} is going to measure the
nuclear suppression factor $R_{AA}$ for heavy baryons.
Keeping this in mind, we are keen to study the $R_{AA}$ of heavy baryons highlighting the possible impact of
the hadronic medium. To access the effects of the QGP phase without hadronic interaction,
we take the initial distribution of heavy quarks $f_i$ at $t=\tau_i$, and compare it with the distribution of
heavy baryons right after the heavy quark fragmentation takes place ($f_{QGP\rightarrow HP}$ at $T_c$), that is
\begin{equation} R_{AA}^{QGP \rightarrow HP}(p)=\frac{f_{QGP \rightarrow HP} (p)}{f_i (p)} \ . \end{equation}
Similarly, the suppression factor in the hadronic phase alone can be written as
\begin{equation} R_{AA}^{HP}(p)=\frac{f_{HP}(p)}{f_{QGP \rightarrow HP }(p)} \ , \end{equation}
where $f_{HP}$ is the solution of the Langevin equation describing the evolution in the hadronic phase at
the freeze out $T_{kin}=120$ MeV.
Notice that in the absence of any hadronic rescattering effect $R_{AA}^{HP}=1$.
The net suppression of the heavy mesons during the entire evolution process, from the beginning
of the QGP phase to the end of the hadronic phase is given by:
\begin{equation}
R_{AA}(p)=R_{AA}^{QGP \rightarrow HP}(p) \times R_{AA}^{HP}(p)= \frac{f_{HP} (p)}{f_i (p)} \ ,
\end{equation}
which in the absence of genuine hadronic effects $R_{AA} (p) \simeq R_{AA}^{QGP \rightarrow HP} (p)$.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=17pc,clip=true]{RAA_wH_LambdaC_FM_QP.eps}\hspace{2pc}
\includegraphics[width=17pc,clip=true]{RAA_wH_LambdaB_FM_QP.eps}\hspace{2pc}
\caption{$R_{AA}$ as a function of $p_T$ for $\Lambda_c$ (left panel) and $\Lambda_b$ (right panel) at RHIC energy.}
\label{fig2}
\end{center}
\end{figure}
In Fig.~\ref{fig2} we show the variation of $R_{AA}$ as a function $p_T$ for the $\Lambda_c$ (left panel)
and $\Lambda_b$ (right panel) in the QGP as well as in the QGP+HP at RHIC energy.
For the $\Lambda_c$ the suppression is stronger as we increase $p_T$ than the $\Lambda_b$, mainly due to
the different interaction of c and b quarks in the QGP phase. We find that the
role of the hadronic phase on both the $\Lambda_c$ and the $\Lambda_b$ $R_{AA}$ is almost unnoticeable.
This can be explained because $R_{AA}$ is very sensitive to the early stages of the
expansion (at high temperatures) where the energy density
is the highest~\cite{Das:2015hla}. Therefore, collisions take place at a high rate in the early stages, before hadronization.
This translates into a strong initial suppression ($R_{AA}$) which then gets saturated within 3-4 fm due
to the radial flow that is able to compensate the baryon energy loss.
Hence, further rescattering in the hadronic medium is unable to alter this spectrum.
Note that the spectra of $\Lambda_c$ and $\Lambda_b$ baryons is obtained here from the fragmentation of
high-energy charm and bottom quarks using the Peterson fragmentation function.
Such mechanism of hadronization may not be valid for low-momentum hadrons which are expected to be produced
from the coalescence of a heavy quark with thermal light partons~\cite{Greco:2003vf}.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=17pc,clip=true]{RAA_LambdaC_5.5TeV.eps}\hspace{2pc}
\includegraphics[width=17pc,clip=true]{RAA_LambdaB_5.5TeV.eps}\hspace{2pc}
\caption{$R_{AA}$ as a function of $p_T$ for $\Lambda_c$ (left panel) and $\Lambda_b$ (right panel) at LHC energy.}
\label{fig3}
\end{center}
\end{figure}
We have also extended our calculation to study $R_{AA}$ of $\Lambda_c$ and $\Lambda_b$ at LHC colliding energy
by performing simulations of $Pb+Pb$ at $\sqrt{s}$ = 5.5 ATeV.
These are our predictions for the upcoming heavy-baryon data at ALICE energy. In Fig.~\ref{fig3} we
present the variation of $R_{AA}$ as a function of $p_T$
for the $\Lambda_c$ (left panel) and $\Lambda_b$ (right panel) in the QGP as well as in the QGP+HP at LHC energy.
As seen before for RHIC energies, the suppression is stronger for $\Lambda_c$ than $\Lambda_b$.
This is mainly due to the larger drag coefficient of charm quark than bottom quark, which shifts the high-$p_T$ particles
to lower $p_T$ resulting in a higher population at low $p_T$. In addition, the bottom quark initial distribution is harder
than the initial charm quark distribution. We have not considered the effect of shadowing~\cite{Eskola:2008ca}
in the initial charm distribution which could be significant at low momentum.
In the case of LHC energy we find that the role of the hadronic phase on $R_{AA}$ is almost unnoticeable for both $\Lambda_c$ and $\Lambda_b$.
\subsection{Elliptic flow, $v_2$~\label{sec:v2}}
Another key observable related to heavy quarks measured at the RHIC and LHC energies is the elliptic flow induced by the spatial
anisotropy of the bulk medium. It can be calculated as
\begin{equation}
v_2=\left\langle \frac{p_x^2 -p_y^2}{p_T^2}\right\rangle =\left\langle \frac{p_x^2 -p_y^2}{p_x^2+p_y^2}\right\rangle\ . \qquad \qquad
\end{equation}
We define the $v_2$ generated in QGP phase taking $p_x$, $p_y$ and $p_T$ as the momenta of the
heavy baryons at $T_c$.
The $v_2$ for the heavy baryons during the entire evolution process, from the
beginning of the QGP phase to the end of the hadronic phase, is computed by
taking $p_x$, $p_y$ and $p_T$ the momenta at the freeze-out $T_f$.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=17pc,clip=true]{v2_wH_LambdaC_FM_QP.eps}\hspace{2pc}
\includegraphics[width=17pc,clip=true]{v2_wH_LambdaB_FM_QP.eps}\hspace{2pc}
\caption{$v_2$ as a function of $p_T$ for $\Lambda_c$ (left panel) and $\Lambda_b$ (right panel) at RHIC energy.}
\label{fig4}
\end{center}
\end{figure}
In Fig.~\ref{fig4} we see the variation of $v_2$ as a function of $p_T$ for the $\Lambda_c$ (left panel)
and $\Lambda_b$ (right panel) in the QGP as well as in the QGP+HP at RHIC energy.
We find that the $v_2$ is enhanced due to the presence of the hadronic phase.
As mentioned earlier, the $R_{AA}$ is quite sensitive to the early stages of the expansion
(at high $T$) where the energy density is the highest, and therefore
collisions take place at a higher rate. However such a strong interaction will not be accompanied
by a build-up of $v_2$ because the bulk medium has not yet developed
a sizable part of its elliptic flow. First, the bulk will generate its own $v_2$ and then the
bulk will transfer it to the heavy quarks. This usually happens at the later
stage of the evolution. Hence, the $v_2$ is sensitive to the heavy particle-bulk interaction.
It should be mentioned that the heavy baryons develop a substantial part
of their $v_2$ mainly from the interaction they suffer at the quark level (as $c$ or $b$ quarks)
in the QGP phase as well as due to their interaction in the hadronic phase. But they also can
get some part of their $v_2$ (mainly at low momentum) from the
thermal light quarks during hadronization by coalescence, which cannot be captured using only
fragmentation.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=17pc,clip=true]{V2_5.5TeV_Lc.eps}\hspace{2pc}
\includegraphics[width=17pc,clip=true]{V2_5.5TeV_LB.eps}\hspace{2pc}
\caption{$v_2$ as a function of $p_T$ for $\Lambda_c$ (left panel) and $\Lambda_b$ (right panel) at LHC energy.}
\label{fig5}
\end{center}
\end{figure}
We have also extended the calculation for LHC energies at Run 2. In Fig.~\ref{fig5} we have shown the variation of $v_2$ as a
function of $p_T$ for the $\Lambda_c$ (left panel) and $\Lambda_b$ (right panel) in the QGP as
well as in the QGP+HP phase at $\sqrt{s} = 5.5$ ATeV. We find
the $v_2$ further enhanced upto 15 $\%$ due to the presence of the hadronic phase.
We find the enhancement of the $v_2$ due to the presence of hadronic phase is larger for RHIC
colliding energy than the LHC energy, which is clearly shown on the $v_2$ plots.
The difference in the magnitude of $v_2$ due to the hadronic phase contribution
at the RHIC and LHC energies can be understood from the magnitude of the drag and
diffusion coefficients in the hadronic medium as well as from the initial distribution.
The coefficients and the initial distribution are inputs in Langevin dynamics at the beginning of the hadronic phase.
The temperature of the hadronic medium for both the RHIC and LHC colliding energies varies from $T_c$ to $T_f$
(from 160 to 120 MeV), and therefore the values of the drag and diffusion
coefficients will not change much. However, the input initial distribution to the hadronic matter
is harder at the LHC energy than at the RHIC energy, resulting in less $v_2$ at the LHC energy.
Also the lifetime of the hadronic phase remains the same for both RHIC and LHC energies, whereas the lifetime
of the QGP phase is longer at LHC energy than RHIC, hence, having the hadronic phase
less impact at LHC energy. Indeed, the effect of the hadronic phase on the $v_2$ will be more significant
for low-energy nuclear collision due to the diminishing lifetime of the QGP phase.
Note that, as compared to $v_2$, the impact of the hadronic phase in $R_{AA}$ is much smaller both at RHIC and LHC energy.
Hence, the $R_{AA}$ may play a unique role in characterizing QGP phase.
It is also important to mention that the impact of the hadronic medium is almost mass independent, i.e. the
impact of the hadronic medium in the $v_2$ of $\Lambda_b$ and $\Lambda_c$ is similar.
\section{Heavy baryon to meson ratio~\label{sec:ratio}}
The heavy baryon to heavy meson ratios, ($\Lambda_c/D$ and $\Lambda_b/B$), are fundamental for
the understanding of in the medium hadronisation~\cite{Greco:2003vf} with respect to the light flavored baryon to meson ratio~\cite{Greco:2003xt,Greco:2003mm}.
Enhancement of $\Lambda_c/D$ and $\Lambda_b/B$ in Au+Au/Pb+Pb collisions compared to p+p collisions affects
the non-photonic single-electron spectra resulting from semileptonic
decays of hadrons containing heavy flavors, hence, their nuclear suppression factor ($R_{AA}$)~\cite{Sorensen:2005sm,MartinezGarcia:2007hf,Ayala:2009pe,Fries:2008hs}.
This is because the branching ratio for the decay process $\Lambda_c \rightarrow e + X (4.5\% \pm 1.7\%)$ is
smaller than the decay process $D \rightarrow e + X (17.2\% \pm 1.9\%)$, resulting in less electrons
from decays of $\Lambda_c$ baryon than D meson. Hence, enhancement of $\Lambda_c/D$
ratio in Au+Au/Pb+Pb collisions will affect the observed non-photonic single-electrons, hence, the $R_{AA}$.
In this manuscript we investigate the $\Lambda_c/D$ ratio for both RHIC and LHC energies.
We study the possible impact of hadronic medium, if any, on
heavy baryon to heavy meson ratio. This investigation is very timely, because LHC is preparing for Run 2 and 3
having major interest on heavy baryon to meson ratio. It becomes particularly
appealing to study if heavy baryons observables are carrying signature of the QGP phase or QGP+Hadronic
phase.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=17pc,clip=true]{LcbyD_RHIC.eps}\hspace{2pc}
\includegraphics[width=17pc,clip=true]{LcbyD_LHC.eps}\hspace{2pc}
\caption{$\Lambda_c/D$ as a function of $p_T$ at RHIC (left panel) and at LHC energy (right panel).}
\label{fig6}
\end{center}
\end{figure}
To evaluate the heavy baryon to meson ratio we use the fragmentation as well as fragmentation plus coalescence model for heavy qurak hadronization.
The coalescence mechanism we employ for D meson and $\Lambda_c$ is similar to one used for the
hadronization of light quarks in \cite{Minissale:2015zwa,Greco:2003xt,Greco:2003mm}.
Given the momentum distribution of the heavy quarks obtained solving the Langevin dynamics, the contribution due to
coalescence can be evaluated as follows:
\begin{eqnarray}
\label{eq-coal}
\frac{d^{2}N_{H}}{dP_{T}^{2}}&=& g_{H} \int \prod^{n}_{i=1} \frac{d^{3}p_{i}}{(2\pi)^{3}E_{i}} p_{i} \cdot d\sigma_{i} \; f_{q_i}(x_{i}, p_{i})
f_{H}(x_{1}..x_{n}, p_{1}..p_{n})\, \delta^{(2)} \left(P_{T}-\sum^{n}_{i=1} p_{T,i} \right)
\label{eq_coal}
\end{eqnarray}
where $d\sigma_{i}$ represents an element of a space-like hypersurface, $n$ is the number of quarks, $g_{H}$ is a statistical
factor to form a colorless hadron from the spin 1/2 quark and antiquark. $f_{q_i}$ are the quark/anti-quark distribution functions.
$f_{H}$ is the Wigner function and describes the coordinate and momentum distribution of quarks/anti-quarks in a hadron.
The Wigner function depends in principle on the overlap of the quark and anti-quark distribution functions with the
wave function of the meson/baryons as well as the interactions of emitted virtual partons, which are needed for the
energy and momentum balance, with the QGP. If one neglecte the off-shell effects then the coalescence probability
function is simply the covariant hadron Wigner distribution function. The longitudinal momentum distributions of the quarks and
antiquarks are assumed to be boost-invarian. For details, we refer to Ref.\cite{Greco:2003mm}.
In the Greco-Ko-Levai (GKL) approach~\cite{Greco:2003mm} for a heavy meson the Wigner function is taken as a Gaussian of
radius $\Delta_{x}$ in the coordinate and $\Delta_{p}$ in the
momentum space, these two parameters being by the uncertainty principle $\Delta_{x}\Delta_{p}=1$,
\begin{eqnarray}
f_{M}(x_{1}, x_{2}; p_{1}, p_{2}) &=& 8
\exp(x_r^2/(2\Delta_x^2))\exp((p_r^2-\Delta m_{12}^{2})/(2\Delta_p^2))
\end{eqnarray}
where $x_{r}=x_{1} - x_{2}$ and $p_{r}=p_1-p_2$ are the quadri-vectors for the relative coordinates
and $\Delta m_{12}=m_1-m_2$ is the scalar
We use $\Delta_x$= 1.06 fm for D meson.
To extend the calculations for mesons to the formation of baryons from the parton
distribution functions, we take the baryon coalescence probability function as,
\begin{eqnarray}
&&F_B(x_1,x_2,x_3;p_1,p_2,p_3) \nonumber \\
&&=8^2\exp(x_r^2/(2\Delta_x^2))\exp((p_r^2-\Delta m_{12}^{2})/(2\Delta_p^2)) \nonumber\\
&&\times \exp(\frac{1}{6}(x_1+x_2-2x_3)^2/(2\Delta_x^2)) \nonumber \\
&&\times \exp(\frac{1}{6}((p_1+p_2-2p_3)^2-(m_1+m_2-2m_3)^2)/(2\Delta_p^2))
\end{eqnarray}
We use $\Delta_x$= 0.98 fm for $\Lambda_c$.
Starting from the charm quark distributions, coalescence probability of $D$ and $\Lambda_c$ has been calculated using Eq.~\ref{eq-coal} at $T_c$
with the appropriate choice of Wigner function.
The charm quarks that do not coalescence, are eventually fragmented in accordance with the fragmentation functions of $D$ and $\Lambda_c$
discussed in subsection ~\ref{sec:frag}. In the present calculations, we have included the contributions from resonances decay coming from $\Sigma_c$,
$\Lambda(2526)$, $\bar \Sigma_c$ and $\bar \Lambda_c$. It can be mentioned that the contribution from resonance decays affect the ratio (heavy baryon to meson)~\cite{Oh:2009zj}
as it involve the ratio of two different hadron species having different contribution from the resonance decays. But the impact of resonances
decays on $R_{AA}$ is negligible, even if not vanishing, as it affects similarly the numerator and denominator of the ratio because its impact is similar
in p+p and Au+Au/Pb+Pb. After the hadronization, we perform the time evolution of heavy hadrons ($D$ and $\Lambda_c$) within the
hadronic phase until the temperature reaches $T_{kin}=120$ MeV.
In Fig.~\ref{fig6} we see the variation of $\Lambda_c/D$ as a function of $p_T$ at RHIC (left panel) and at LHC energy (right panel) for QGP
and QGP+Hadronic phase within coalescence plus fragmentation and fragmentation. The impact of coalescence is mainly restricted to the
low $p_T$ region (within $p_T$~1-5 GeV) above which the hadronization mechanism is dominated by fragmentation. The coalescence probability involves the
product of two distribution functions which fall very fast at high momentum making way for the fragmentation as the dominant mechanism of hadronization.
As shown, the impact of coalescence is less significant at LHC energy than RHIC energy. This is because impact of coalescence depends on the slope of the charm
quark $p_T$ distribution. For a harder distribution the gain in momentum reflects in a smaller increase of the slope of the charm quark distribution, instead
if the distribution decreases fast in momentum then the same momentum gain due to coalescence will result in a stronger increase of the spectrum.
For a hard distribution, which is the case at LHC energy (in contrast to RHIC),the impact of coalescence will be less pronounce.
The present study gives the possibility to disentangle/understand different hadronization mechanisms of heavy quarks once the data will be available.
More significantly, the $\Lambda_c/D$ is independent of the hadronic phase both at RHIC and LHC energies. As discussed earlier, the impact of hadronic
phase is almost unnoticeable on $R_{AA}$ of $\Lambda_c$, hence, on the spectra.
Thus, the impact of the hadronic medium on another ratio, such as $\Lambda_c/D$, is negligible, which enables $\Lambda_c/D$ as a noble probe of QGP phase
dynamics including its hadronization. We prefer to ignore
the bottom case to avoid uncertainty in the final ratios due to lack of knowledge of the resonance feed-down from higher states.
\section{Conclusions~\label{sec:conclusions}}
We have studied the $R_{AA}$ and $v_2$ of heavy baryons highlighting the impact of the hadronic medium
on these observables within a Langevin dynamics. The QGP medium interaction of the $c$ and $b$ quarks with
the light quarks and gluons have been treated within a quasi-particle model~\cite{Das:2015ana}.
To fix the shape of heavy baryon fragmentation function, we have used the information available
from electron-positron annihilation data on heavy baryon fragmentation function.
Heavy baryon fragmentation function is softer than the heavy meson fragmentation function as
the baryon involves three-body fragmentation whereas the meson a two-body fragmentation. We find that the impact of
hadronic medium on the $R_{AA}$ for heavy baryons ($\Lambda_c$ and $\Lambda_b$) is almost unnoticeable while
the hadronic medium contribution is sizable on $v_2$, which is about $20\%$. We have also calculated the
$\Lambda_c/D$ ratio at RHIC and LHC energies. The heavy hadron suppression does not
change in the hadronic phase, hence the spectra. Thus, the impact of the hadronic medium on another
ratio, such as $\Lambda_c/D$, is negligible, which enables $\Lambda_c/D$ as a noble probe of QGP phase.
So the enhancement of heavy baryon to meson ratio, if any,
in $Au+Au$/$Pb+Pb$ collisions with respect to $p+p$ collisions would be an indication of QGP phase dynamics
including its hadronization. Furthermore, the $\Lambda_c/D$ can serve as a tool to disentangle different
hadronization mechanisms once the data will be available.
\section{Acknowledgements}
JMTR thanks Jan Wagner for interesting discussions on heavy-quark physics.
SKD, FS and VG acknowledge the support by the ERC StG under the QGPDyn Grant no. 25968.
JMTR acknowledges the financial support from programme TOGETHER from R\'egion Pays de la Loire,
and from a Helmholtz Young Investigator Group VH-NG-822 from the Helmholtz Association and GSI.
LT acknowledges support from the Ram\'on y Cajal research programme.
JMTR and LT also acknowledge support by Grants FPA2010-16963 and FPA2013-43425-P (Spain).
|
\section{Introduction}
\IEEEPARstart{I}{t} has been shown that short low-rate codes with a good performance can be constructed from generalized low-density parity-check (GLDPC) codes with hybrid check nodes (e.g. \cite{Liva_Hamming}, \cite{Liva_QC}).
Liva and Ryan \cite{Liva_Hamming} were first who defined {\it{doping}} to refer to substituting some single parity checks by super checks corresponding to a stronger linear block code and constructed check-hybrid GLDPC (CH-GLDPC) codes using Hamming codes as component codes. In another work by Liva {\it{et al.}} \cite{Liva_QC}, low-rate GLDPC codes are constructed by doping quasi-cyclic (QC)-LDPC codes with Hamming codes. It was shown that the constructed codes have a remarkable performance both in the waterfall and the error-floor regions on the additive white Gaussian noise (AWGN) channel.
Paolini {\it{et al.}} \cite{Paolini_1}, \cite{Paolini_2} studied the GLDPC and doubly-GLDPC codes with Hamming or BCH codes as the component codes and proposed a method for the asymptotic analysis of doubly-GLDPC codes on the binary erasure channel (BEC). They also considered CH-GLDPC codes and showed that the asymptotic threshold of hybrid GLDPC codes outperforms that of the LDPC codes.
In another work \cite{Paolini_3}, Paolini {\it{et al.}} analyzed the asymptotic exponent of both the weight spectrum and the stopping set size spectrum for the CH-GLDPC codes and provided a simple formula for the asymptotic exponent of the weight distribution of the CH-GLDPC codes.
Two common features of the methods given in the previous work are: (i) replacing the super checks based on degree distribution or density evolution of the resulting CH-GLDPC codes, and (ii) significant reduction of the rate of CH-GLDPC codes compared to the original LDPC code. In this paper, we propose a method to construct CH-GLDPC codes; however, our approach is different in that the super checks are chosen specifically to address the error floor issue and is based on the knowledge of failures of the global
LDPC code on the BSC under the parallel bit flipping (PBF) algorithm. The PBF algorithm is a simple algorithm with low complexity and hence suitable for high-speed applications. This algorithm is also appropriate for the analysis of failures of iterative decoding algorithms of LDPC codes, first identified by Richardson and denoted as ``trapping sets" \cite{Richardson}. While trapping sets of the LDPC codes over the binary erasure channel (BEC) are well characterized as ``stopping sets", they are more complicated to define over the BSC and the AWGN channel. In \cite{Vasic}, the most harmful structures of column-weight three LDPC codes on the BSC using Gallager A/B and the PBF algorithms have been identified. It was also shown that the trapping sets are short cycles or can be obtained as the union of short cycles in the Tanner graph.
One important aspect of this work is to provide a guidance in order to jointly design the Tanner graph of the proposed CH-GLDPC codes, and assign the location of the component codes with the objective of lowering the error floor. Our construction of the CH-GLDPC codes is decomposed in two steps: we start with a classical LDPC code design (QC, protograph, etc.), and the knowledge of its small trapping sets, then, instead of randomly choosing super checks, we place the super checks corresponding to a 2-error correcting component codes at those check nodes so that the PBF decoder can correct the errors on a trapping set. For an efficient check-hybrid code design, it is also desirable to find the minimum number of super checks such that the rate loss of the constructed check-hybrid codes be reduced. In this paper, we study the minimum number of such critical super checks, denoted as the {\it{splitting number}} and provide upper bounds on the splitting number for some dominant trapping sets. The LDPC codes that are used in this paper are column-weight three and column-weight four LDPC codes. We first focus on trapping sets of column-weight three LDPC codes and provide an algorithm to find critical checks in a trapping set and also provide upper bounds on the splitting number of trapping sets. Furthermore, we study the error correction capability of two classes of CH-GLDPC codes using a column-weight three LDPC code as the global code and show that a CH-GLDPC code in which each variable node is connected to 2 super checks is able to correct up to 5 errors. The results obtained for the critical checks, splitting number and error correction capability of CH-GLDPC codes with column-weight three LDPC codes as the global code and the PBF decoding algorithm are generalized when the Gallager B decoding algorithm is used.
The rest of the paper is organized as follows. In Section \ref{Pre}, we provide the notations and definitions that are used throughout the paper. In section \ref{SuperChecks}, we characterize the effect of super checks in terms of trapping sets elimination. In section \ref{results}, we present our main results on CH-GLDPC codes free of small trapping sets. In section \ref{GEC}, we give the guaranteed error correction capability of the constructed CH-GLDPC codes. In section \ref{Discussion}, we extend some of our results for column-weight four global LDPC codes and also Gallager B decoder.
Section \ref{conclusion} concludes the paper.
\section{Preliminaries}
\label{Pre}
In this section, we first establish the notations and then give a brief summary on the definitions and concepts of LDPC and GLDPC codes. We also define trapping sets and fixed sets for the iterative decoding algorithms.
\subsection{Graph Theory Notations }
Let $G(U,E)$ be an undirected simple graph with the set of vertices $U$ and the set of edges $E$. An edge $e$ is an unordered pair $(u_1,u_2)$. The edge $e=(u_1,u_2)$ is said to be incident on $u_1$ and $u_2$ and the two vertices $u_1$ and $u_2$ are said to be adjacent (neighbors). The set of neighbors of the vertex $u$ is denoted by ${\cal{N}}(u)$. The degree of each vertex $d(u)$ is defined as the number of vertices in its neighborhood. The length of the shortest cycle is called the girth of the graph and is denoted by $g$. A bipartite graph $G(V \cup C,E)$ is graph with two disjoint sets of vertices; variable nodes $V$ and check nodes $C$. An edge $e$ is incident on a variable node $v\in V$ and a check node $c \in C$. A bipartite graph is called $(\gamma,\rho)$-regular if the degree of each variable node is $\gamma$ and the degree of each check node is $\rho$. The girth of a bipartite graph is even.
The parity check matrix $H$ of a linear code $C$ can be represented with a bipartite graph called the Tanner graph. Each column in the parity check matrix is shown by a variable node and each row is denoted by a check node in the Tanner graph. A variable node $v_j$ and a check node $c_i$ are adjacent if and only if $H_{i,j}=1$. A vector ${\bf{v}}=(v_1,v_2,...,v_n)$ is a codeword if and only if $H{\bf{v}}^{T}={\bf{0}} ~~({\rm{mod ~2}})$. A linear code is called $(\gamma,\rho)$-regular if its parity check matrix is $(\gamma,\rho)$-regular. This code has rate $r \geq 1-\frac{\gamma}{\rho}$ \cite{Gallager}.
\subsection{LDPC codes, GLDPC and CH-GLDPC codes }
LDPC codes were first introduced by Gallager in his landmark work \cite{Gallager} where he proposed different methods for constructing parity check matrices of LDPC codes and provided different hard decision algorithms for decoding of LDPC codes. LDPC codes are usually defined by their Tanner graphs. A $(\gamma,\rho,g)$ LDPC code is a $(\gamma,\rho)$-regular code of girth $g$.
GLDPC codes were introduced by Tanner in \cite{Tanner_GLDPC} where he proposed a method to construct longer error-correcting codes from shorter error-correcting codes. In GLDPC codes, each super check node is satisfied if its neighboring variable nodes form a codeword of a linear code called {\it{component code}}. That is if $c_i$ is a single parity check node in the Tanner graph of the global code and $\{v_{i_1},v_{i_2},...,v_{i_n}\}$ with values $\{x_1,x_2,...,x_n\}$ are the neighbors of $c_i$, then in the GLDPC code, the super check corresponding to $c_i$ is satisfied if $(x_1,x_2,...,x_n)$ be a codeword of the component code.
The parity check matrix of GLDPC codes is constructed using the parity check matrix of the longer code also known as the global code and the parity check matrix of the component code. To construct the parity check matrix of the GLDPC code, it is enough to replace each one in each row of the parity check matrix of the global code by one column of the parity check matrix of the component code. Each zero in each row will be replaced by a zero-column in the parity check matrix.
A CH-GLDPC code has two types of check nodes: single parity checks and super checks corresponding to a component code. As in GLDPC codes, a super check node is satisfied when its neighboring variable nodes be codeword of the component code, while the single parity check is satisfied when the modulo-2 sum of its neighboring variable nodes is zero.
The component codes in GLDPC and CH-GLDPC codes can be chosen arbitrarily and possibly from different block codes. However, in this paper, GLDPC and CH-GLDPC codes are constructed from the same component code and the global codes are chosen from the family of $(\gamma,\rho)$-regular codes.
\subsection{Decoding Algorithms and Trapping Sets }
The decoding algorithms for decoding LDPC codes include a class of iterative algorithms such as bit flipping algorithms (parallel and serial) and messages passing algorithms like Gallager A/B and belief propagation decoding algorithms.
The notion of ``trapping sets" was first introduced by Richardson \cite{Richardson} as the structures in the Tanner graph of LDPC codes responsible for failures of decoders. Before we characterize the trapping sets of bit flipping decoding algorithm, we provide definitions and assumptions.
In this paper, we consider transmission over the BSC. We also consider that the all-zero codeword is sent. Under this assumption, a variable node is said to be correct if its received value is 0; otherwise it is called corrupt. The support of a vector ${\bf{x}}=(x_1,x_2,...,x_n)$ denoted by ${\rm{supp}}({\bf{x}})$ is the set $\{i ~|~x_i \neq 0\}$. The decoder runs until the maximum number of iterations $M$ is reached or a codeword is found. Let ${\bf{y}}=(y_1,y_2,...,y_n)$ be a received vector after transmitting the all-zero codeword and let ${\bf{y}^{(l)}}=(y_1^{(l)},y_2^{(l)},...,y_n^{(l)})$ be the output of the decoder after the $l$-th iteration. A variable node $v$ is said to be eventually correct if there exists an integer $L>0$ such that for all $l\geq L$, $v \notin {\rm{supp}}({\bf{x}}^l)$. The decoder fails on decoding ${\bf{y}}$ if there does not exist $l \leq M$ such that $|{\rm{supp({\bf{x}})}}| = 0$. For the received word $\bf{y}$, the set of variable nodes which are not eventually correct is called a trapping set and is denoted by ${T}(\bf{y})$. If ${T}(\bf{y}) \neq \emptyset$, then ${T}({\bf{y}})$ is called an $(a,b)$ trapping set and is denoted by ${\cal{T}}(a,b)$ if the number of variable nodes in ${T}({\bf{y}})$ equals $a$ and the number of odd degree check nodes in the subgraph induced by ${T}({\bf{y}})$ is $b$. For the trapping set $T({\bf{y}})$, ${\rm{supp({\bf{y}})}}$ is an induced set. ${\cal{T}}(a,b)$ is called an {\it{elementary trapping set}} if the degree of each check node in the subgraph induced by the set of variable nodes is one or two and there $b$ check nodes of degree one.
Chilappagari {\it{et al.}} \cite{Shashi_Error_Floor} introduced the notion of ``critical number" as the minimum number of variable nodes on a trapping set that need to be initially in error such that the decoder fails. It was shown that the harmfulness of a trapping set depends on its critical number; the smaller the critical number, the more harmful a trapping set. In this paper, we say that a trapping set is {\it{harmful}} if the decoder fails to decode at least one error pattern on the trapping set; Otherwise, it is called {\it{harmless}}.
While trapping sets can have different induced sets, a class of trapping sets called {\it{fixed sets}} have the fixed induced set. A fixed set $F$ is the set of variable nodes that are corrupt at the beginning and at the end of decoding iterations, while the variable nodes that are initially correct remain correct after decoding. A vector $\bf{y}$ is called a fixed point if ${\rm{supp(y)}}=F$. From definition of the fixed set and trapping set, it is clear that a fixed set is always a trapping set while a trapping set is not necessarily a fixed set. Fixed sets of an LDPC code with the column-weight $\gamma$ are the set of variable nodes ${\cal{I}}$ such that every variable node in ${\cal{I}}$ is connected to at least $\left\lceil \gamma /2 \right\rceil$ of check nodes of even-degree and no $\left\lfloor \gamma /2 \right\rfloor$ check nodes of odd-degree share a variable node outside ${\cal{I}}$ \cite{Vasic}.
Chilappagari {\it{et al.}} defined fixed sets for the PBF algorithm of GLDPC codes as follows:
{\bf{Fact 1}}:(\cite{shashi_GLDPC} Theorem 6) Let ${\cal{C}}$ be a GLDPC code with $(\gamma,\rho)$-regular global code and a $t$-error correcting component code. Let ${\cal{I}}$ be a subset of variable nodes with the following properties: (a) The degree of each check node in ${\cal{I}}$ is either 1 or $t+1$; (b) Each variable node in ${\cal{I}}$ is connected to $\left\lceil \gamma /2 \right\rceil$ checks of degree $t+1$ and $\left\lfloor \gamma /2 \right\rfloor$ check nodes of degree 1; and (c) No $\left\lfloor \gamma /2 \right\rfloor +1$ checks of degree $t+1$ share a variable node outside ${\cal{I}}$. Then, ${\cal{I}}$ is a fixed set.
\section{Effect of Super checks on trapping sets}
\label{SuperChecks}
Let us start by some observations on the effect of replacing single parity checks by super checks. In fact, we show how trapping sets responsible for the failure of the PBF are not harmful anymore when some selected single checks are replaced by super checks \cite{RDV_14_ITA}, \cite{RDV_14_ISIT}. We first describe the PBF algorithm for the CH-GLDPC codes and use it throughout the paper for our analysis. We mention that the decoding algorithm at each super check is the bounded distance decoding (BDD). The BDD is capable of correcting $t$ errors when the minimum distance of the code is at least $2t+1$.
%
\begin{algorithm}
\caption{The PBF algorithm for decoding CH-GLDPC codes \cite{RDV_14_ITA}, \cite{RDV_14_ISIT}.}
\label{Alg1}
\begin{algorithmic}
\STATE {\bf{In each iteration:}}
\begin{itemize}
\item Variable nodes send their current estimates to the neighboring single parity check and super check nodes.
\end{itemize}
\STATE {\bf{~~Updating rule at check nodes:}}
\begin{itemize}
\item Each super check node performs the BDD on the incoming messages. If a codeword is found, then the check node sends flip messages to all variable nodes which differ from the codeword. If not, then the check node does not send any flip messages.
\end{itemize}
\begin{itemize}
\item At each single parity check, the modulo-2 sum of the incoming messages is calculated. If the sum is not zero, then the check node sends flip messages to the neighboring variable nodes. If the sum is zero, then the check node does not send any flip messages.
\end{itemize}
\STATE {\bf{~~Updating rule at variable nodes:}}
\begin{itemize} \item A variable node flips if it receives more than $\gamma/2$ flip messages.
\end{itemize}
\end{algorithmic}
\end{algorithm}
Let $\cal{C}$ be a $(3,\rho,8)$ LDPC code.
Fig. \ref{trapsets} shows some small trapping sets of a column-weight three LDPC codes of girth $g=8$ namely the $(4,4)$ trapping set, the $(5,3)$ trapping set and a $(6,4)$ trapping set. In this paper, $\circ$ denotes a variable node and $\Box$ denotes a check node.
It can be easily seen that if all single parity checks in the Tanner graph corresponding to the parity check matrix of ${\cal{C}}$ are replaced by super checks of a 2-error correcting component code, then the PBF decoding algorithm for GLDPC codes can correct all errors on the trapping sets. This result can be explained by the fact that in all elementary trapping sets, the degree of each check node is at most two and since they are replaced by a 2-error correcting component code, the BDD at each super check can correct all errors. Fig. \ref{AllSC} shows how the PBF corrects all errors located on the (5,3) trapping set when all single checks are replaced by super checks. In this paper, a $\blacksquare$ denotes a super check and flip messages are shown with $\rightarrow$. However, as we show in the following, it is not necessary to replace all super checks in a trapping set for the decoder to correct the errors. We show that it is possible to make the trapping set
harmless by replacing only some selected single checks by super checks. We say a trapping set is {\it{eliminated}} if by replacing super checks, the trapping set is not harmful anymore.
\begin{figure}[t]
\begin{center}
\subfigure[]{
\centering\includegraphics[width=1.1in]{ts44cropped1.png}
}
\subfigure[]{
\centering \includegraphics[width=1in]{ts53cropped1.png}
}
\subfigure[]{
\centering\includegraphics[width=1.8in]{ts642cropped1.png}
}
\end{center}
\caption{Tanner graph representation of trapping sets for column-weight three and girth $g=8$ LDPC codes; (a) the (4,4) trapping set, (b) the (5,3) trapping set, (c) a (6,4) trapping set.}
\label{trapsets}
\end{figure}
\begin{figure}[t]
\begin{center}
\subfigure[]{
\centering\includegraphics[width=1in]{AllSC53cropped1.png}
}
\subfigure[]{
\centering \includegraphics[width=1in]{AllSC53bcropped1.png}
}
\end{center}
\caption{ The (5,3) trapping set is eliminated when all single parity checks are replaced by super checks corresponding to a 2-error correcting component code. (a) flip messages from super checks to corrupt variable nodes in the first iteration of the PBF algorithm, (b) all variable nodes are corrected after the first iteration. }
\label{AllSC}
\end{figure}
Let consider the (5,3) trapping set. Fig. \ref{fig:ts(5,3)} shows how the PBF algorithm corrects all errors located on the trapping set in which only two single parity checks of degree 2 are replaced by super checks.
\begin{figure}[t]
\begin{center}
\subfigure[]{
\centering\includegraphics[width=1.2in]{ts531cropped1.png}
}
\subfigure[]{
\centering \includegraphics[width=1in]{ts532cropped1.png}
}
\subfigure[]{
\centering\includegraphics[width=1in]{ts533cropped1.png}
}
\subfigure[]{
\centering\includegraphics[width=1in]{ts534cropped1.png}
}
\end{center}
\caption{The (5,3) trapping set in a column-weight three code is eliminated if two super checks corresponding to a 2-error correcting component code are replaced. Arrows show flip messages from check nodes to corrupt variable nodes in each iteration of the PBF algorithm: (a) flip messages from checks in the first iteration, (b) flip messages from checks to 3 variable nodes that are still in error, (c) flip messages from checks to the only one corrupt variable, (d) all variable nodes are corrected after the third iteration.}
\label{fig:ts(5,3)}
\end{figure}
It should be noted that not all pairs of super checks in the (5,3) trapping set can be helpful for the decoder to correct the errors on the (5,3) trapping set. Fig. \ref{exceptions} shows three possible cases that by replacing the super checks the trapping sets remain harmful \cite{RDV_14_ITA}. In Fig. \ref{exc1} and \ref{exc2} only the variable node $v_5$ will be corrected, while in Fig. \ref{exc3} all variable nodes will remain incorrect.
\begin{figure}[t]
\begin{center}
\subfigure[]{
\centering\includegraphics[width=1.1in]{except1cropped1.png}
\label{exc1}
}
\subfigure[]{
\centering \includegraphics[width=1.1in]{except2cropped1.png}
\label{exc2}
}
\subfigure[]{
\centering \includegraphics[width=1.1in]{except3cropped1.png}
\label{exc3}
}
\end{center}
\caption{Possible super-check replacements which are not helpful for the decoder to correct all errors on the (5,3) trapping set.}
\label{exceptions}
\end{figure}
The above examples show that not only the number of super checks, but also the positions of super checks in a trapping set are important for the decoder to successfully correct the errors. Since the rate of the GLDPC codes decreases by replacing single parity checks by super checks, we are interested in replacing the minimum number of super checks such that the resulting Tanner graph will be free of small trapping sets. In the next section, we first provide an algorithm to find a set of such critical checks in a trapping set and then
we present upper bounds on the minimum number of super checks that need to be replaced in the parity check matrix such that the resulting Tanner graph will be free of small trapping sets.
\section{Critical sets and the splitting number}
\label{results}
In this section, we provide our main results on CH-GLDPC codes in which the trapping sets responsible for the failure of the PBF algorithm have been eliminated. In this section, whenever not stated, the global LDPC code of the CH-GLDPC codes is a $(3,\rho,8)$ LDPC code.
\subsection{ Critical sets and minimal size of critical sets }
As shown in Section \ref{SuperChecks}, a trapping set can be eliminated by judiciously replacing check nodes in the original global code. A set of such checks is called a {\it{critical set}} and defined as follows \cite{RDV_14_ITA},\cite{RDV_14_ISIT}.
\begin{definition}
Let ${\cal{T}}(a,b)$ be an elementary trapping set. Let $C=\{c_1,c_2,...,c_k\}$ where $k\leq b$ be a set of check nodes of degree 2 in ${\cal{T}}$. A set $S \subseteq C$ is called critical if by converting the single parity checks in $S$ to the super checks, the trapping set is eliminated.
\end{definition}
We note that a critical set is not unique and there are many possible critical sets with different sizes in a trapping set.
\begin{definition}
\label{sp}
Let ${\cal{T}}(a,b)$ be an elementary trapping set. The minimum size of a critical set in ${\cal{T}}$ is denoted by $s_{(a,b)}({\cal{T}})$.
\end{definition}
As an example,$s_{(5,3)}({\cal{T}})=2$ as can be seen in Fig. \ref{fig:ts(5,3)}.
In Algorithm \ref{Alg2}, we provide a method to find one of many possible critical sets in a trapping set. The motivation behind finding a critical set using Algorithm \ref{Alg2} is based on the role of super checks in elementary trapping sets. When a single parity check of degree-2 is replaced by a super check, then the super check sends a flip message to a neighboring variable node if and only if the variable node is corrupt. Thus, each super check plays the role of 2 equivalent and isolated single parity checks, one for each of connected variable nodes. Breaking the cycles in a trapping set by splitting the super check into two single parity checks is the basis for finding a critical set in Algorithm \ref{Alg2}.
Fig. \ref{cycle breaking} shows an alternative view of the effect of a super check to eliminating a trapping set.
\begin{figure}[t]
\centering \includegraphics[width=2.4in]{CycleBreak53cropped1.png}
\caption{Super checks corresponding to a 2-error correcting component code can be considered as two single parity checks of degree-1. These replacements break the cycles responsible for the failure of decoding.}
\label{cycle breaking}
\end{figure}
\begin{algorithm}
\caption{Finding a critical set in a trapping set ${\cal{T}}(a,b)$ \cite{RDV_14_ITA}, \cite{RDV_14_ISIT}.}
\label{Alg2}
\begin{algorithmic}
\STATE {\bf{initialization}:} Let ${\cal{T}}'={\cal{T}}$ be the $(a,b)$ trapping set.
\WHILE{Number of variable nodes in ${\cal{T'}}$ is greater than 0}
\IF{there exists a variable node $v$ in ${\cal{T'}}$ which is connected to exactly one degree-1 check node and two degree-2 checks }
\STATE Replace one of the check nodes of degree-2 connected to $v$ by a super check
corresponding to a 2-error correcting code. Split the super check into two single checks. Remove the variable node $v$ and all edges connected to it.
\ELSE
\STATE Choose a variable node $v$ in ${\cal{T'}}$. Replace one check node of degree-2 connected to $v$ by a super check and split the super check node to 2 single parity checks.
\ENDIF
\WHILE{Number of variable nodes connected to at least two single parity checks of degree-1 is greater than 0}
\STATE Remove variable nodes connected to at least two single parity checks of degree-1 and all edges connected to them.
\ENDWHILE
\ENDWHILE
\end{algorithmic}
\end{algorithm}
As we explained, the number of cycles in a trapping set plays a key role in finding the number of critical checks of a trapping set. This fact helps us to find the number of critical checks in some trapping sets without using Algorithm \ref{Alg2}. If a trapping set ${\cal{T'}}(a',b')$ has been obtained by adding some variable and check nodes to another trapping set ${\cal{T}}(a,b)$ such that the new variable and check nodes do not create a new cycle, then $s_{(a',b')}({\cal{T'}})$ and $s_{(a,b)}({\cal{T}})$ are equal. To be more precise, we first provide the following definitions.
\begin{definition}
A subdivision of a simple graph $G$ is a graph resulting from the subdivision of edges in G. In other words, a subdivision of a graph is a graph obtained by adding
at least one vertex on an edge of the graph.
\end{definition}
Fig. \ref{subdivision} shows a simple graph (Fig. \ref{simple}) and a particular subdivision in Fig. \ref{subd}.
\begin{figure}[t]
\begin{center}
\subfigure[]{
\centering\includegraphics[width=1.6in]{simplegraphcropped1.png}
\label{simple}
}
\subfigure[]{
\centering \includegraphics[width=1.6in]{subdivisioncropped1.png}
\label{subd}
}
\end{center}
\caption{(a) A simple graph, (b) a subdivision of the graph given in (a).}
\label{subdivision}
\end{figure}
We define a graph induced by the set of the variable nodes of a bipartite graph and then we generalize the definition of subdivision of a graph for bipartite graphs.
\begin{definition}
Let $G(V \cup C,E)$ be a bipartite graph. The simple graph $G'(V,E')$ induced by the set of variable nodes $V$ is a graph with $|V|$ vertices in which two vertices $v_1$ and $v_2$ are connected to each other if and only if there exists a check node $c$ in $C$ such that $v_1$ and $v_2$ are neighbors of $c$.
\end{definition}
As an example, consider the (5,3) trapping set as a bipartite graph. The simple graph induced by the set of variable nodes of the (5,3) trapping set is shown in Fig. \ref{varinduced}.
\begin{figure}[t]
\begin{center}
\subfigure[]{
\centering\includegraphics[width=1.1in]{ts53cropped1.png}
\label{1}
}
\subfigure[]{
\centering \includegraphics[width=0.9in]{ts53inducedcropped1.png}
\label{2}
}
\end{center}
\caption{(a) The (5,3) trapping set as a bipartite graph, (b) the simple graph induced by the 5 variable nodes of the (5,3) trapping set.}
\label{varinduced}
\end{figure}
\begin{definition}
Let ${\cal{T}}(a,b)$ be a trapping set. The trapping set ${\cal{T'}}(a+1,b+1)$ is called a subdivision of ${\cal{T}}$ if the simple subgraph induced by the set of variable nodes of ${\cal{T'}}$ is a subdivision of the simple graph induced by the set of variable nodes of ${\cal{T}}$.
\end{definition}
Fig. \ref{subdivision2} shows two trapping sets, a (6,4) trapping set and a (7,5) trapping set, in which the (7,5) trapping set is a subdivision of the (6,4) trapping set.
\begin{figure}[t]
\begin{center}
\subfigure[]{
\centering\includegraphics[width=1.9in]{ts64cropped1.png}
}
\subfigure[]{
\centering \includegraphics[width=1.9in]{ts75cropped1.png}
}
\end{center}
\caption{(a) A (6,4) trapping set, (b) a (7,5) trapping set which is a subdivision of the (6,4) trapping set given in (a).}
\label{subdivision2}
\end{figure}
\begin{corollary}
Let ${\cal{T'}}(a+1,b+1)$ be a trapping set which is a subdivision of the trapping set ${\cal{T}}(a,b)$. Then $s_{(a+1,b+1)}({\cal{T'}})=s_{(a,b)}({\cal{T}})$.
\end{corollary}
As we want to reduce the rate-loss caused by converting single checks to super checks, we now study the minimum number of super checks that are required to be replaced in a Tanner graph of an LDPC code such that the decoder can correct all error patterns on all $(a,b)$ trapping sets.
\begin{definition}
Let ${\cal{C}}$ be a $(3,\rho,8)$-LDPC code with the parity check matrix $H$ and let ${\cal{T}}(a,b)$ be an elementary trapping set in $H$.
The minimum number of super checks corresponding to a 2-error correcting component code that are required for eliminating all $(a,b)$ trapping sets in $H$ is called the {\it{splitting number}} of the $(a,b)$ trapping sets in $H$ and is denoted by $s_{(a,b)}(H)$.
\end{definition}
\subsection{Upper bounds on the splitting number}
Now, we provide upper bounds on the splitting number of trapping sets in the parity check-matrices based on permutation matrices. Permutation-based LDPC codes are $(\gamma,\rho)$- regular codes constructed from permutation matrices. A permutation matrix is any square matrix in which the weight of each row and each column is one. If the permutation matrix is cyclic, the permutation matrix is called a circulant permutation matrix and the LDPC code becomes quasi-cyclic \cite{Marc}. The parity check matrix of a quasi-cyclic LDPC code can be represented by an array of circulant permutation matrices as follows \cite{Marc}:
\begin{equation}
\label{matrix}
H=\left[ \begin{array}{cccc}
I_0 & I_0 & \cdots & I_0 \\
I_0 & I_{p_{1,1}} & \cdots & I_{p_{1,\rho-1}}\\
\vdots & & \ddots & \vdots\\
I_0 & I_{p_{\gamma-1,1}} & \cdots & I_{p_{\gamma-1,\rho-1}} \end{array} \right]
\end{equation}
where for $1 \leq j \leq \gamma-1$ and $1 \leq l \leq \rho-1$, $I_{p_{j,l}}$ represents the circulant permutation matrix with a one at column-$(r+p_{j,l})$ mod $p$ for the row $r$ ($0 \leq r \leq p-1$). If for $1 \leq j \leq \gamma-1$ and $1 \leq l \leq \rho-1$, $I_{p_{j,l}}$ is not circulant, then $H$ is just a $(\gamma,\rho)$-regular matrix based on permutation matrices.
\begin{lemma}
\label{Lem1}
Let ${\cal{C}}$ be a $(3,\rho,8)$ LDPC code with the parity-check matrix $H$ based on permutation matrices of size $p$. Then, $s_{(a,b)}(H)\leq 2p$, for all $a$ and $b$.
\end{lemma}
Proof: Suppose the first $2p$ rows of $H$ are replaced by super checks. The first $2p$ rows of $H$ correspond to the first two rows of blocks in equation (\ref{matrix}). Thus, each variable node is connected to exactly 2 super checks and 1 single parity check. It results that each variable node receives at least 2 correct messages from its neighbors. In fact, by converting two single parity checks to super checks and then splitting each super check into two single parity check nodes, all cycles in all elementary trapping sets are eliminated. {\it{Q.E.D.}}
According to Lemma \ref{Lem1}, all elementary trapping sets are eliminated when each variable node is connected to exactly two super checks. Thus, the trapping sets for this class of CH-GLDPC codes are non-elementary trapping sets.
We now exhibit a fixed set for the PBF algorithm for the CH-GLDPC code in the case that the super checks have been replaced such that each variable node is connected to exactly two super checks.
\begin{theorem}
\label{fixedset}
Let ${\cal{T}}$ be a subset of variable nodes with the induced subgraph ${\cal{I}}$. Then, ${\cal{T}}$ is a fixed set if (a) The degree of each check node in ${\cal{I}}$ is either 1 or 3 and; (b) Each variable node in ${\cal{I}}$ is connected to 2 check nodes of degree 3 and 1 check node of degree 1 where the check nodes of degree 3 have been replaced by super checks of the 2-error correcting component code and; (c) No 2 check nodes share a variable node outside ${\cal{I}}$.
\end{theorem}
Proof: Since the check nodes of degree 3 have been replaced by super checks of a 2-error correcting component code and since the decoding in the component codes is the BDD, the super checks of degree 3 do not send any flip messages to the variable nodes in ${\cal{I}}$. Also, since any variable node in ${\cal{I}}$ is connected to 2 super checks, it remains corrupt. Furthermore, no variable node outside ${\cal{I}}$ receives more than 1 flip message because no 2 check nodes share a variable node outside ${\cal{I}}$. Thus, the variable nodes outside ${\cal{I}}$ that are originally correct will remain correct. Consequently, ${\cal{I}}$ is a fixed set. {\it{Q.E.D.}}
Fig. \ref{fig:fixedset} shows a fixed set in a $(3,\rho,8)$-LDPC code in which each variable node is connected to exactly 2 super checks. We note that conditions (a) and (c) are similar to the corresponding conditions in Fact 1. The main difference is in condition (b) where in Theorem \ref{fixedset}, the constraint on the position of super checks is a stronger condition on ${\cal{I}}$ to be a fixed set. We also note that if this condition is not satisfied, ${\cal{I}}$ may not be either a trapping set or a fixed set.
Fig. \ref{fig:notfixedset} shows a subgraph satisfying all conditions of Theorem \ref{fixedset} except the condition (b) which is not a trapping set nor a fixed set.
\begin{figure}[t]
\centering
\includegraphics[width=2.1in]{fixedsetcropped1.png}
\caption{A fixed set for a $(3,\rho,8)$-LDPC code in which each variable node is connected to exactly two super checks.}
\label{fig:fixedset}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=2.1in]{notfixedsetcropped1.png}
\caption{An example of a subgraph in a $(3,\rho,8)$-LDPC code which satisfies all conditions of Theorem \ref{fixedset} except the condition (b). This structure is not harmful for the PBF algorithm. }
\label{fig:notfixedset}
\end{figure}
Although all elementary trapping sets are eliminated when each variable node is connected to two super checks, there are trapping sets that are eliminated if each variable node is connected to exactly one super check. Fig. \ref{fig:split} depicts a possible way for replacing super checks in ${\cal{T}}(5,3)$ and ${\cal{T}}(7,3)$, such that each variable node is connected to exactly one super check and the trapping sets are not harmful anymore.
Thus, for a permutation-based LDPC code ${\cal{C}}(3,\rho,8)$ with the parity-check matrix $H$, if the parity checks corresponding to the first $p$ rows of $H$ are replaced by super checks, then all ${\cal{T}}(5,3)$ and ${\cal{T}}(7,3)$ trapping sets are eliminated and hence $s_{(5,3)}(H) \leq p$ and $s_{(7,3)}(H) \leq p$.
\begin{figure}[t]
\begin{center}
\subfigure[]{
\centering\includegraphics[width=1.15in]{ts53slesspcropped1.png}
\label{ts5,3}
}
\subfigure[]{
\centering \includegraphics[width=1.5in]{ts73slesspcropped1.png}
\label{ts7,3}
}
\end{center}
\caption{Some trapping sets in a column-weight 3 LDPC codes that can be eliminated if each variable node is connected to exactly one super check. The graphs in (a) and (b) correspond to the (5,3) and (7,3) trapping sets, respectively.}
\label{fig:split}
\end{figure}
It is easy to see that the smallest trapping set, the (4,4) trapping set, may not be eliminated if each variable node is connected to exactly one super check. In fact, the $(4,4)$ trapping set will remain harmful if the single parity checks of degree-1 are replaced by super checks (Fig. \ref{ts44harmful}). The following Theorem provides a condition on the parity check matrix $H$ in which all $(4,4)$ trapping sets are eliminated if each variable node is connected to exactly one super check.
\begin{theorem}
\label{Th2}
Let ${\cal{C}}$ be a $(3,\rho,8)$ QC-LDPC code with the parity check matrix $H$. Suppose the first $p$ rows of $H$ are replaced by super checks. Then, $s_{(4,4)}(H) \leq p$ if the girth of the Tanner graph corresponding to the last $2p$ rows of $H$ is 12.
\end{theorem}
Proof: If in ${\cal{T}}(4,4)$ the single parity checks of degree-1 are replaced by super checks, then due to the existence of a cycle of length 8, the PBF cannot correct the errors. However, if the girth of the subgraph induced by the single parity checks is greater than 8, then there will not be any 8-cycle and consequently all $(4,4)$ trapping sets will be eliminated. According to Corollary 2.1 in \cite{Marc}, the girth of a $(2,\rho)$-regular QC-LDPC code is $4i$ for some integer $i>0$. Moreover, the girth of $H$ cannot be more than 12 as shown in \cite{Marc}. Thus, if the girth of the subgraph induced by the last $2p$ rows of $H$ is 12, it results that all 8-cycles in $H$ will contain at least one super check of degree 2, and henceforth the 8-cycles are not the harmful $(4,4)$ trapping sets. {\it{Q.E.D.}}
\begin{figure}[h]
\centering
\includegraphics[width=1.2in]{ts44harmfulcropped1.png}
\caption{The (4,4) trapping set is still harmful if each variable is connected to exactly one super check where have been replaced instead of degree-1 single parity checks in the trapping set.}
\label{ts44harmful}
\end{figure}
We finish this section by providing a lower bound on the rate of the CH-GLDPC codes.
\begin{lemma}
Let ${\cal{C}}$ be a $(\gamma,\rho)$-regular LDPC code with the parity-check matrix $H_{M \times N}$. Let $C$ be a $t$-error correcting component code of rate $r$ with a full-rank parity-check matrix $H'_{m \times \rho}$. If $\kappa$ be the number of single parity checks in $H$ that are replaced by super checks corresponding to $C$, then the rate of the CH-GLDPC code
$$R \geq 1-\frac{\gamma}{\rho}-\kappa \lambda (1-r)$$
where $\lambda = \frac{\rho}{N}$.
\end{lemma}
Proof: If $\kappa$ be the number of super checks that are replaced in $H$, then there will be $(\kappa m+ (M-\kappa))$ rows in the parity check matrix of the CH-GLDPC codes. Thus, the rate of the CH-GLDPC code is:
\begin{eqnarray}
R &\geq &1- \frac{(\kappa m+ (M-\kappa))}{N} \nonumber\\
& \geq &1-\frac{\gamma}{\rho}-\kappa (1-r)\frac{\rho}{N}\nonumber
\end{eqnarray}
where the last inequality follows from the fact that $1-\frac{M}{N}=1-\frac{\gamma}{\rho}$ and $m-1 <\rho(1-r)$.
Assuming $\lambda = \frac{\rho}{N}$ proves the result. {\it{Q.E.D.}}
\begin{corollary}
\label{rate}
Let ${\cal{C}}$ be a $(\gamma,\rho)$-regular LDPC code and let $H_{M \times N}$ be the parity-check matrix based on permutation matrices with size $p$. Let $C$ be a $t$-error correcting component code of rate $r$ with a full-rank parity-check matrix $H'_{m \times \rho}$. If $\kappa=\alpha p$ be the number of single parity checks in $H$ that are replaced by super checks corresponding to $C$, where $\alpha$ is an integer and $0 \leq \alpha \leq \gamma$, then the rate of the CH-GLDPC code is:
$$R \geq 1-\frac{\gamma}{\rho}-\alpha (1-r).$$
\end{corollary}
To see how tight the lower bound on the rate of the CH-GLDPC codes given in Corollary \ref{rate} is consider a permutation-based ${\cal{C}}(3,31,8)$ LDPC code of rate 0.9034. If each variable node is connected to 1 super check corresponding to the BCH(31,21), then the actual rate of the CH-GLDPC code is 0.6130 while the lower bound given in Corollary \ref{rate} is 0.5806.
If each variable node is connected to 2 super checks of the BCH(31,21), then the actual rate is 0.3236 and the lower bound is 0.2580.
\section{Guaranteed Error Correction Capability of the CH-GLDPC codes }
\label{GEC}
In this section, we study the error correction capability of the CH-GLDPC codes in which the global code is a $(3,\rho,8)$ regular LDPC code and the component code is a 2-error correcting code. The code families that are studied are i) CH-GLDPC codes in which each variable node is connected to exactly 2 super checks and ii) the CH-GLDPC codes in which each variable node is connected to exactly 1 super check. For simplicity, we denote the first code family with ${\cal{C}}^{I}$ and the second code family with ${\cal{C}}^{II}$.
\begin{theorem}
\label{GuarErr}
Consider a CH-GLDPC code \textsc{C} from the code family ${\cal{C}}^{I}$. Then the PBF can correct up to 5 errors in \textsc{C}.
\end{theorem}
Proof: See Appendix.\\
\begin{corollary}
Consider a CH-GLDPC code \textsc{C} in ${\cal{C}}^{I}$. Then, there exists an error pattern of size 6, in which the PBF fails on correcting the errors.
\end{corollary}
Proof: Figure \ref{6UncorrectableError} shows an example in which the PBF fails to correct 6 errors while every variable node is connected to 2 super checks corresponding to
a 2-error correcting component code. {\it{Q.E.D.}}
\begin{figure}[ht]
\centering
\includegraphics[width=2.2in]{6UncorrectableErrorcropped1.png}
\caption{An uncorrectable error pattern of size 6 in a CH-GLDPC code that each variable node is connected to 2 super checks.}
\label{6UncorrectableError}
\end{figure}
As shown in Theorem \ref{GuarErr}, when each variable node is connected to 2 super checks of a 2-error correcting component code, then the CH-GLDPC code can correct up to 5 errors. The following Corollary proves the guaranteed error correction capability of the CH-GLDPC codes in ${\cal{C}}^{II}$.
\begin{corollary}
Lets suppose a CH-GLDPC code \textsc{C} in ${\cal{C}}^{II}$. Then the PBF can correct up to 1 error in \textsc{C}.
\end{corollary}
Proof: It is easy to see that if there exist 2 errors on a (4,4) trapping set in which each degree-1 check node is replaced by a super check (as shown in Fig. \ref{ts44harmful}), then the PBF fails. Thus, the guaranteed error correction capability of a CH-GLDPC code in ${\cal{C}}^{II}$ is equal to the error correction capability of PBF for LDPC codes. {\it{Q.E.D.}}
\section{Splitting numbers of ${\cal}(4,\rho,6)$ LDPC codes and trapping sets elimination using the Gallager B decoding algorithm }
\label{Discussion}
In this section, we generalize our results on critical sets and splitting number of $(3,\rho,8)$ LDPC codes to $(4,\rho,6)$ LDPC codes and the Gallager B decoding algorithm
\subsection{Elimination of trapping sets by super checks in $(4,\rho,6)$ LDPC codes}
In Section \ref{SuperChecks}, we provided a method to eliminate harmful (elementary) trapping sets in $(3,\rho,8)$ LDPC codes and provided upper bounds on the splitting number of trapping sets in permutation based LDPC codes. In this section, we extend our results for $(4,\rho,6)$ LPDC codes. Fig. \ref{TS4} shows some small trapping sets in a $(4,\rho,6)$ LDPC code.
\begin{figure}[t]
\begin{center}
\subfigure[]{
\centering\includegraphics[width=1in]{ts436cropped1.png}
\label{ts4,3,6}}
\subfigure[]{
\centering \includegraphics[width=0.7in]{ts444cropped1.png}
\label{ts4,4,4}}
\subfigure[]{
\centering\includegraphics[width=0.97in]{ts446cropped1.png}
\label{ts4,4,6}}
\end{center}
\caption{Some small trapping sets in column-weight four LDPC codes with girth 6. (a) the (3,6) trapping set, (b) the (4,4) trapping set and (c) the (4,6) trapping set.}
\label{TS4}
\end{figure}
Fig. \ref{TS4harmless} shows a possible replacement of super checks corresponding to a 2-error correcting component code that eliminates the trapping sets. In Algorithm \ref{Alg3} we provide a method to find critical sets in an elementary trapping set of a $(4,\rho,6)$ LDPC code.
\begin{figure}[t]
\begin{center}
\subfigure[]{
\centering\includegraphics[width=1in]{ts436harmlesscropped1.png}
\label{ts4,3,6harmless}}
\subfigure[]{
\centering \includegraphics[width=0.7in]{ts444harmlesscropped1.png}
\label{ts4,4,4harmless}}
\subfigure[]{
\centering\includegraphics[width=0.9in]{ts446harmlesscropped1.png}
\label{ts4,4,6harmless}}
\end{center}
\caption{A possible replacement of super checks to eliminate the trapping sets in column-weight four LDPC codes.}
\label{TS4harmless}
\end{figure}
\begin{algorithm}
\caption{Finding a critical set in a trapping set ${\cal{T}}(a,b)$ in a $(4,\rho,6)$ LDPC code.}
\label{Alg3}
\begin{algorithmic}
\STATE {\bf{initialization}:} Let ${\cal{T}}'={\cal{T}}$ be the $(a,b)$ trapping set.
\WHILE{Number of variable nodes in ${\cal{T'}}$ is greater than 0}
\IF{there exists a variable node $v$ in ${\cal{T'}}$ which is connected to two degree-1 check nodes and two degree-2 checks }
\STATE Replace one of the check nodes of degree-2 connected to $v$ by a super check
corresponding to a 2-error correcting code. Split the super check into two single checks. Remove the variable node $v$ and all edges connected to it.
\ELSE
\IF{there exists a variable node $v$ in ${\cal{T'}}$ which is connected to one degree-1 check node and three degree-2 checks}
\STATE Replace two check nodes of degree-2 connected to $v$ by super checks
corresponding to a 2-error correcting code. Split the super checks into two single checks. Remove the variable node $v$ and all edges connected to it.
\ENDIF
\ENDIF
\WHILE{Number of variable nodes connected to at least two single parity checks of degree-1 is greater than 0}
\STATE Remove variable nodes that are connected to at least two single parity checks of degree-1 and all edges connected to them.
\ENDWHILE
\ENDWHILE
\end{algorithmic}
\end{algorithm}
Following the same methodology used in Section \ref{results}, Lemma \ref{Lem1} can be generalized for column-weight four LDPC codes as follows.
\begin{lemma}
\label{Lem2}
Let ${\cal{C}}$ be a $(4,\rho,6)$ LDPC code with the parity-check matrix $H$ based on permutation matrices of size $p$. Then, $s_{(a,b)}(H)\leq 3p$, for all $a$ and $b$.
\end{lemma}
Proof: The proof is similar to the proof of Lemma \ref{Lem1}. If the first $3p$ rows of $H$ are replaced by super check corresponding to a 2-error correcting component code, then each variable node receives at least 3 correct messages from its neighbors and hence all cycles in all trapping sets are broken by the super checks. {\it{Q.E.D.}}
We now present a fixed set for the PBF algorithm for the CH-GLDPC code in which each variable node in a $(4,\rho,6)$ LDPC code is connected to exactly 3 super checks of a 2-error correcting component code.
\begin{corollary}
\label{fixedset2}
Let ${\cal{T}}$ be a subset of variable nodes with the induced subgraph ${\cal{I}}$. Then, ${\cal{T}}$ is a fixed set if (a) The degree of each check node in ${\cal{I}}$ is either 1 or 3 and; (b) Each variable node in ${\cal{I}}$ is connected to 3 check nodes of degree 3 and 1 check node of degree 1 where the check nodes of degree 3 have been replaced by super checks of the 2-error correcting component code and; (c) No 2 check nodes share a variable node outside ${\cal{I}}$.
\end{corollary}
The following result provides a condition on the parity check matrix $H$ in which all $(3,6)$ trapping sets are eliminated if each variable node is connected to one super check.
\begin{theorem}
\label{Condition}
Let ${\cal{C}}$ be a $(4,\rho,6)$ QC-LDPC code with the parity check matrix $H$. Suppose the first $p$ rows of $H$ are replaced by super checks. Then, $s_{(3,6)}(H) \leq p$ if the girth of the Tanner graph corresponding to the last $3p$ rows of $H$ is at least 8.
\end{theorem}
Proof: The proof is similar to the proof of Theorem \ref{Th2}. If each variable node is connected to 1 super check and the girth of the subgraph induced by the single parity checks is greater than 6, then there is not any $(3,6)$ trapping set. {\it{Q.E.D.}}
We may note that under the condition in Theorem \ref{Condition}, the 8-cycles may not broken and so the other small trapping sets shown in Fig. \ref{TS4}, i.e. the $(4,4)$ and the $(4,6)$ trapping sets may remain harmful.
\begin{theorem}
\label{GuarErr2}
Let ${\cal{C}}$ be a $(4,\rho,6)$-regular LDPC code. Lets suppose in a CH-GLDPC code constructed using ${\cal{C}}$ as the global code, each variable node is connected to 3 super checks corresponding to a 2-error correcting component code. Then the PBF can correct at least 3 errors in the CH-GLDPC code obtained by replacing super checks.
\end{theorem}
Proof: The proof is similar to the proof of Theorem \ref{GuarErr}. All elementary trapping sets are eliminated when each variable node is connected to 3 super checks. Thus, it is enough to consider the cases that there exists at least one check (single check or super check) connected to more than 2 errors. Recall that since the decoding algorithm of the component codes is the BDD, each super check sends at most two flip messages to the variable nodes in its neighborhood. If a super check is connected to more than 2 corrupt variable nodes, we consider the worst case scenario and assume that the super check sends 2 flip messages to correct variable nodes in its neighborhood. We may note that the errors on a tree subgraph are eventually corrected. It can be easily seen that when a super check or a single check is connected to 3 corrupt variable nodes, all errors are eventually corrected. {\it{Q.E.D.}}
We may note that if in a $(4,\rho,6)$ LDPC code, each variable node is connected to exactly 1 super check of a 2-error correcting component code, or if is connected to exactly 2 super checks, the error correction capability of the CH-GLPDC code is equal the error correction capability of a $(4,\rho,6)$ LDPC code which is 2. This is due to the fact that in these classes of CH-GLDPC codes, the smallest trapping sets in a $(4,\rho,6)$ LDPC code are not necessarily eliminated.
\subsection{Elimination of trapping sets by super checks using Gallager B decoding algorithm}
In this section, we show that the method used for eliminating the trapping sets of a column-weight 3 LDPC code with the PBF algorithm can also be used for eliminating the trapping sets with the Gallager B decoding algorithm. To show how the results of the PBF algorithm can be generalized for the Gallager B decoding algorithm, we first explain the decoding algorithm of the CH-GLDPC codes using the Gallager B decoding algorithm on the global code and the BDD on the component codes in Algorithm \ref{Alg4}.
\begin{algorithm}
\caption{The Gallager B decoding algorithm for CH-GLDPC codes.}
\label{Alg4}
\begin{algorithmic}
\STATE {\bf{Initialization}}
The variable nodes send their received values to the
neighboring single checks and super checks over the incident edges.
\STATE {\bf{In each iteration:}}
\STATE {\bf{~~Updating rule at check nodes:}}
\begin{itemize}
\item Each super check node performs the BDD on the incoming messages. If a codeword is found, then the check node sends the values of the BDD decoder to the variable nodes. If not, then the check node sends the value of each variable node to itself.
\end{itemize}
\begin{itemize}
\item At each single parity check, the message sent from a check to a neighboring variable is the sum of all incoming messages except the one arriving from the variable.
\end{itemize}
\STATE {\bf{~~Updating rule at variable nodes:}}
\begin{itemize} \item The
message sent from a variable to a neighboring check is the
majority (if it exists) among all incoming messages except the
one arriving from the check. If a majority does not exist, then
the received value corresponding to the variable is sent to the
check.
\end{itemize}
\end{algorithmic}
\end{algorithm}
Fig. \ref{GalB} shows how the Gallager B decoding algorithm can correct all errors on the $(5,3)$ trapping set.
\begin{figure}[t]
\begin{center}
\subfigure[]{
\centering\includegraphics[width=1in]{GallagerBts53cropped1.png}
\label{GalBts5,3}}
\subfigure[]{
\centering \includegraphics[width=1in]{GallagerBts532cropped1.png}
\label{GalBts5,32}}
\subfigure[]{
\centering\includegraphics[width=0.97in]{GallagerBts533cropped1.png}
\label{GalBts5,33}}
\subfigure[]{
\centering \includegraphics[width=1in]{GallagerBts534cropped1.png}
\label{GalBts5,34}}
\subfigure[]{
\centering\includegraphics[width=0.97in]{GallagerBts535cropped1.png}
\label{GalBts5,35}}
\subfigure[]{
\centering\includegraphics[width=0.97in]{ts534cropped1.png}
\label{GalBts5,36}}
\end{center}
\caption{The (5,3) trapping set in a column-weight three code that is eliminated if two super checks corresponding to a 2-error correcting component
code are replaced. Arrows show messages from check nodes to
variable nodes in each iteration of the Gallager B decoding algorithm: (a)-(d) messages from
checks in iterations 1 to 4, (e) all variable nodes are corrected after the fourth iteration.}
\label{GalB}
\end{figure}
It is easy to see that the role of a super check of a 2-error correcting component code in trapping set using Gallaber B decoding algorithm is similar to the role of a super check of a 2-error correcting component code in trapping set using the PBF. Hence, by carefully replacing the super checks in the trapping set, the cycles responsible for the failure of the Gallager B decoder are broken. Thus, the results obtained for finding a critical set and the upper bounds on the splitting number of the trapping sets with the PBF are also correct for the Gallager B decoder. For the guaranteed error correction capability of the CH-GLDPC codes using the Gallager B decoding, it can be easily seen that the Theorem \ref{GuarErr} is also true for the Gallager B decoding algorithm. A single check that sends a flip message to variable nodes in the PBF, sends 0 to a variable node that is in error and sends 1 to a correct variable node in the Gallager B decoding algorithm. Thus, the same analysis used in the proof of Theorem \ref{GuarErr} can be used to prove it for the Gallager B decoding algorithm.
The subgraph shown in Fig. \ref{6UncorrectableError} is also an error pattern of size 6 for the failure of the Gallager B decoding algorithm. We note that in the CH-GLDPC codes in which each variable node is connected to exactly 1 super check, the error correction capability of the code with the Gallager B decoding algorithm is the same as the error correction capability of a $(3,\rho,8)$ LDPC code with the Gallager B decoding. In this case, the error correction capability of the CH-GLDPC code is 2.
\section{Conclusion }
\label{conclusion}
In this paper, we introduced a method for constructing CH-GLDPC codes in which the super checks corresponding to a 2-error correcting component code are chosen based on the knowledge of trapping sets of a column-weight 3 global LDPC code. By replacing the super checks, we eliminated harmful trapping sets of the PBF algorithm while minimizing the rate loss caused by adding more constraints on check nodes of the component code. We also studied the guaranteed error correction capability in the CH-GLDPC codes. The results were also extended to the Gallager B decoding algorithm and column-weight 4 LDPC codes.
\begin{appendix}
Proof of Theorem \ref{GuarErr}: To prove this theorem, we first note that according to Lemma \ref{Lem1}, all elementary trapping sets are eliminated when each variable node is connected to two super checks. Thus, it is enough to consider the cases that there exists at least one check (single check or super check) connected to more than 2 errors. We also mention that since the decoding algorithm of the component codes is the BDD, each super check sends at most two flip messages to the variable nodes in its neighborhood. If a super check is connected to more than 2 corrupt variable nodes, we consider the worst case scenario and assume that the super check sends 2 flip messages to correct variable nodes in its neighborhood. We may note that the errors on a tree subgraph are eventually corrected. In other words, every trapping set must contain at least one cycle.
Recall that as we showed in Fig. \ref{cycle breaking}, super checks corresponding to a 2-error correcting component codes break the cycle if they are connected to at most 2 corrupt variable nodes. Using these facts, we show that all error patterns of size 5 are corrected as their corresponding subgraph can be transformed to a tree.
We first consider all possible subgraphs in which a super check node is connected to more than 2 corrupt variable nodes and show that the subgraphs with different error patterns can be transformed to a tree. Then, we repeat it for a single check connected to more than 2 corrupt variable nodes. To construct all the subgraphs in which a super check node is connected to more than 2 corrupt variable nodes, we consider a super check as a root (level 1) and expand it. The root check node is connected to at least 3 corrupt variable nodes to this super check (level 2). As we mentioned before, we consider the worst case scenario and assume that the super check node sends 2 flip messages to 2 correct variable nodes. The other variable nodes connected to the root check node always send 0 to the root check node, therefore, it is sufficient to connect the root check node to the corrupt variable nodes and two correct variable node to which the flip massages are sent. We then expand this graph by connecting 2 check nodes (one single and one super check) to each variable node of level 2 to construct the level 3 check nodes. We note that since girth of the global code is 8, a cycle can only be made in at least 5th level of constructing the subgraph. However, as we show, the cycles are broken due to super check nodes connecting to at most two corrupt variable nodes and the graph forms a tree.
Figure \ref{5Err_5} shows how 5 corrupt variable nodes that are connected to one super check are corrected in one iteration.
\begin{figure}[t]
\centering\includegraphics[width=1.8in]{5Err5cropped1.png}
\caption{A correctable error pattern of size 5 in which all corrupt variable nodes are connected to one super check.}
\label{5Err_5}
\end{figure}
Fig. \ref{5Err_4} is one of the cases that an error pattern of size 5 is considered in which 4 errors are connected to one super check and one error is connected to one single check.
It is easy to see that all the other error patterns in which 4 errors are connected to one super check can be corrected in at most 2 iterations since this graph is transformed a tree.
\begin{figure}[t]
\centering\includegraphics[width=1.8in]{5Err42cropped1.png}
\caption{A correctable error pattern of size 5 in which 4 corrupt variable nodes are connected to one super check and one error is connected to one single check.}
\label{5Err_4}
\end{figure}
In Fig. \ref{5Err_3_1}, an error pattern of size 5 is considered in which 3 errors are connected to one super check and 2 errors are connected to one single check and one super check. The cycle is broken when each super check is replaced by 2 degree-1 single check. Figures \ref{5Err_3_2}-\ref{5Err_3_6} show other possible error patterns of size 5 in which 3 errors are connected to one super check.
\begin{figure}[ht]
\begin{center}
\subfigure[]{
\centering \includegraphics[width=2.6in]{5Err31cropped1.png}
\label{5Err_3_1}}
\subfigure[]{
\centering\includegraphics[width=2.6in]{5Err32cropped1.png}
\label{5Err_3_2}}
\subfigure[]{
\centering \includegraphics[width=2.6in]{5Err33cropped1.png}
\label{5Err_3_3}}
\subfigure[]{
\centering \includegraphics[width=2.6in]{5Err34cropped1.png}
\label{5Err_3_4}}
\subfigure[]{
\centering \includegraphics[width=2.6in]{5Err35cropped1.png}
\label{5Err_3_5}}
\subfigure[]{
\centering \includegraphics[width=2.6in]{5Err36cropped1.png}
\label{5Err_3_6}}
\caption{Some possible correctable error patterns of size 5 in which one super check is connected to 3 corrupt variable nodes. }
\label{3errors}
\end{center}
\end{figure}
Now, we consider the case that one single check is connected to more than 2 corrupt variable nodes and all the other super check nodes are connected to at most two corrupt variables. Therefore, all the super checks can break the cycles and the errors are corrected in one iteration. Note that the case in which one single check and one super check are connected to more than 2 corrupt variable nodes, is already included in the previous cases (super check as the root check node).
\end{appendix}
|
\section{Introduction}
The issue of matching a service system's capacity to stochastic demand induced by its clients arises in many practical settings. Typically, the resources available to satisfy demand are scarce and hence expensive. This forces the manager to consider a trade-off between the system efficiency and the quality of service perceived by its clients. In this paper, we focus on this trade-off in the context of the $M/G/1$ queue, in which the variable amendable for optimization is the server speed $\mu$.
In general, optimizing the server speed $\mu$ in a single-server queue in time-homogeneous environment, while trading off congestion levels against capacity allocation cost, does not pose any technical challenges. Typically, the objective function to be minimized, the total cost function, has the shape
\begin{equation}\label{eq:intro}
\Pi_\infty(\mu) = \mathbb{E}[Q_\mu(\infty)] + \alpha\mu = \frac{\lambda\mathbb{E}[B^2] }{2(\mu-\lambda\mathbb{E}[B])} + \alpha\mu,
\end{equation}
where $\mathbb{E}[Q_\mu(\infty)]$ denotes the expected steady-state amount of work given server speed $\mu$, and $B$ describes the service requirement per arrival. The parameter $\alpha>0$ represents the relative capacity allocation costs incurred by deploying service rate $\mu$. This one-dimensional optimization problem yields the optimizer
\begin{equation}
\mu_\infty^\star = \lambda \mathbb{E}[B] + \sqrt{\frac{\lambda\mathbb{E}[B^2]}{2\alpha}}.
\end{equation}
Despite the simplicity and tractability of the problem described above, the presence of the \emph{steady-state} measure in the cost function in \eqref{eq:intro} should be handled carefully. By employing this particular cost structure, one automatically agrees with the underlying assumption of the system being sufficiently close to its steady state.
However, referring the practical applications of the single-server model, system parameters rarely remain constant over time. Moreover, planning periods for the optimization problem are naturally finite. Hence, the \emph{true} expected costs incurred, which we denote by $\Pi_T(\mu)$, depend on the length of the planning period $T$. Consequently, the usage of steady-state models for decision making needs to be justified by a more elaborate time-dependent or \emph{transient} analysis for these type of settings.
The time-dependent behavior of the single-server queue received much attention in queueing theory. First efforts to analyze the time-dependent properties of the $M/G/1$ queue date back to the 1950s and 1960s, e.g \cite{Kendall1951,Takacs1955,Takacs1962,Gaver1959,Benes1957}. The analyses in these papers mostly yield implicit expressions for performance characteristics through Laplace transforms, integro-differential equations and infinite convolutions.
More specifically, there is vast literature on the transient analysis of the $M/M/1$ queue, with the goal to derive explicit expressions for queue length characteristics, see e.g. \cite{Prabhu1964,Cohen1969,Pegden1982,Abate1987}.
These works provide a variety of explicit expressions for the transient dynamics, although the complexity of the resulting expressions, typically involving Bessel functions, expose the intricate intractability of the matter. Consequently, approximation methods for insightful quantification of the dynamics based on numerical \cite{Neuts1966} or asymptotic methods, have become prevalent in more recent literature.
The asymptotic methods either exploit knowledge on the evolution of the queueing process as time $t$ grows large \cite{Newell1982,Odoni1983,Abate1987}, or the arrival rate $\lambda$ is increased to infinity \cite{Gaver1968,Abate1987a,Abate1987b}.
It is noteworthy that a substantial contribution to the transient literature is made by Abate and Whitt \cite{Abate1987a,Abate1987b,Abate1987,Abate1994} who exploit the existence of a decomposition of the mean transient queue length and obtain expressions for the moments of the queue length and virtual waiting through probabilistic arguments in several queueing models.
More recently, asymptotic methods have been used to justify the application of stationary performance measures in Markovian environments or to refine them, see e.g. \cite{Green1991,Whitt1991}.
Other approximative methods under the name of uniform acceleration expansions \cite{Massey1998} have been developed to reveal the asymptotic behavior of the single-server queue as a function of $t$, which are moreover able to capture time-varying arrival rates.
The majority of the works mentioned above do reflect on the error imposed by usage of steady-state performance metrics instead of the correct time-dependent counterpart. However, no light has been shed on the accumulation of this error over a finite period of time. To the best of our knowledge, the only work that addresses this issue is the paper by Steckley and Henderson \cite{Steckley2007}, who compute an approximation for the error accumulated between the steady-state and transient delay probability. Our analysis on the other hand is centered around the mean workload, which requires a different approach. In addition, the focus in \cite{Steckley2007} is on performance measures only, while the main goal of our paper is to investigate the quality of staffing rules.
Although the $M/G/1$ queue serves as the leading example in our analysis, we choose to use a more general framework for the arrival process of the queue. Namely, we let the server face a L\'evy process.
This gives the advantage that once we have obtained the results, we can apply them to broader queue input classes, such as Brownian motion and the Gamma process.
To shed light on the influence of the transience of the queueing process on traditional staffing questions, we will study the capacity allocation problem in the context of cost minimization in which the objective function is $\Pi_T(\mu)$, i.e. a function of both $\mu$ and $T$. We investigate how the invalidity of the stationary assumption is echoed through the operational cost accounting for congestion-related penalties.
Furthermore, we establish a result on the strict convexity of the function $\Pi_T(\mu)$, for almost all values of $T$ (with a few minor exceptions for certain deterministic initial states), which is an essential property for convergence of both cost function and corresponding minimizer to their stationary counterparts.
As it will appear that an exact analysis of this disparity is intractable, we will present an explicit approximate correction to the conventional stationary objective function given by $\Psi(\mu)/T$ and prove that
\begin{equation}
\Pi_T(\mu) = \Pi_\infty(\mu) + \frac{\Psi(\mu)}{T} + O(1/T^2),
\end{equation}
with the help of recent results from the fluctuation theory of L\'evy processes.
Based on this refinement we ultimately examine how incorporating transient effects reflects in setting the optimal capacity level and propose a refinement to the steady-state capacity allocation rule,
\begin{equation}
\mu_T^\star = \mu_\infty^\star + \frac{\mu_\bullet}{T} + o(1/T).
\end{equation}
We moreover deduce an explicit expression for $\mu_\bullet$ in terms of the initial state and the first three moments of the service requirement per arrival.
It is noteworthy that similar refined square-root staffing rules have been proposed for multi-server queues in the Halfin-Whitt regime, see e.g. \cite{Janssen2008,Janssen2011,Janssen2015,Randhawa2014,Zhang2012}. In those cases, the relevant decision value is the number of servers and refinements are derived for $\lambda\to\infty$, whereas we consider the regime $T\to\infty$.
Building upon the insights gained through the analysis of this optimality gap, we reflect on the parameter settings of the underlying queueing process in which our refined capacity sizing rule yields significant improvement and in which cases it has little effect. Special emphasis is put on the relationship between the accuracy of the standard procedure and the length of the planning period.
The remainder of the paper is structured as follows. Section 2 is devoted to the model description and presents some preliminary results. The main result will be given in Section 3 and results regarding the optimization problem will be discussed in Section 4, followed by the validation of our novel techniques through numerical experiments in Section 5. We will give some concluding remarks and topics for further research in Section 6. We have deferred all proofs to the Appendix.
\section{Model description}
\subsection{A queueing model with L\'evy input \label{sec:levymodel}}
The model that inspired our study is the standard $M/G/1$ queue starting out of equilibrium. Customers arrive to the queue according to a Poisson process with rate $\lambda$ and each arrival has service requirement $B_i$, stemming from a common random variable $B$.
Without loss of generality we will assume $\mathbb{E}[B] = 1$ throughout. The server is able to remove $\mu$ amounts of work from the system per time unit; a variable we will refer to as the \emph{server speed}.
E.g. if $\mu = 3$ and two customers are in the system with remaining service times $4$ and $2$, then the queue will be empty 2 time units later, provided that no new arrivals occur in the meantime.
Let $N_\lambda(t)$ denote the number of arrivals until time $t$.
Accordingly, the total work generated by the customers is given by
\begin{equation}
Z_\lambda(t) = \sum_{i=1}^{N_\lambda(t)} B_i.
\end{equation}
Furthermore, define $X_{\mu}(t) = Z_\lambda(t) - \mu t$. We call $X_{\mu}$ the \emph{net-input process}.
More generally, we assume throughout the paper that $X_{\mu}$ is a L\'evy process.
Specifically, we let $Z_\lambda$ be of the form $Z_\lambda(t) = U(\lambda t)$, where $U$ is a spectrally positive L\'evy process generated by the triplet $(a,\sigma,\nu)$ and $\mathbb{E}[U(1)] = 1$.
This restriction to spectrally positive processes is equivalent to stating $\nu(-\infty,0)=0$ and is a vital assumption to our analysis.
Subsequently, we assume the net-input process $X_{\mu}$ to be
\begin{equation}
\label{eq:Xlmprocess}
X_{\mu}(t) = U(\lambda t) - \mu t, \qquad t \geq 0.
\end{equation}
Note that by setting $a=\sigma=0$ and $\nu = \lambda\, F_B$, where $F_B$ is the cumulative distribution function of $B$, we recover the original $M/G/1$ queue.
The stochastic process central to our analysis is the \emph{workload process} $Q_{\mu}(t)$, $t\geq 0$, which describes the amount of work the server is facing at time $t$.
The net-input process $X_{\mu}$ completely determines the trajectory of $Q_{\mu}$, namely
\begin{equation}\label{eq:Qlm}
Q_{\mu}(t) = \max\{ Q(0) + X_{\mu}(t), \sup_{s\in[0,t]} [X_{\mu}(t)-X_{\mu}(s)]\}, \qquad t\geq 0,
\end{equation}
where $Q(0)$ is the initial workload in the system.
In fact, $Q_{\mu}$ is the reflected version of $X_{\mu}$ with reflection barrier at zero.
Careful inspection of the structure also reveals that $X_{\mu}(t) \equiv X_{\lambda/\mu,1}(\mu t) \equiv X_{1,\mu/\lambda}(\lambda t)$, so that
\begin{equation}
\label{eq:Qidentity}
Q_{\mu}(t) \,{\buildrel d \over =}\, Q_{\lambda/\mu,1}(\mu t) \,{\buildrel d \over =}\, Q_{1,\mu/\lambda}(\lambda t)
\end{equation}
for all $\lambda,\mu,t>0$.
This identity will prove to be convenient for numerical analysis in Section \ref{sec:numerics}.
The process $Q_{\mu}$ is a natural indicator of the level of congestion in the system and therefore a good choice for quantifying the Quality of Service (QoS) received by a client.
We remark that alternative processes characterizing congestion in the system can be directly deduced from $Q_{\mu}(t)$. For example, consider the virtual waiting time process $V_{\mu}(t)$, which is the waiting time a customer would experience if he arrives at time $t$. This satisfies the relation $V_{\mu}(t) \equiv Q_{\mu}(t)/\mu$ for all $t\geq 0$.
Likewise, the expected number of the customers in the system $L_{\mu}(t)$ at time $t\geq 0$ is given by Little's law
\begin{equation}
\mathbb{E}[L_{\mu}(t)] = \lambda\, \mathbb{E}[V_{\mu}(t)] = \frac{\lambda}{\mu}\, \mathbb{E}[Q_{\mu}(t)].
\end{equation}
To facilitate our investigation of the queueing model, we end this subsection by introducing some notation regarding the net-input and workload process and by stating a useful preliminary result concerning the stationary process $Q_{\mu}(\infty)$.
Throughout the paper we assume $\mu>\lambda$ to ensure ergodicity of the queue and existence of the limit
\begin{equation}
Q_{\mu}(\infty) := \lim_{t\to\infty} Q_{\mu}(t),
\end{equation}
for any initial state $Q(0)$. This random variable necessarily coincides with the stationary distribution of $Q_{\mu}(t)$.
By $\kappa_U(\cdot)$ and $\kappa_{\mu}(\cdot)$ we denote the L\'evy exponents of the processes $U$ and $X_{\mu}$, respectively:
\begin{equation}
\kappa_{\mu}(\theta) = \log \mathbb{E}[e^{\theta X_{\mu}(1)}] = \log \mathbb{E}[e^{\theta(U(\lambda) - \mu)}] = \lambda \kappa_U(\theta) - \mu \theta.
\end{equation}
Furthermore, define $u_k = \mathbb{E}[\{U(1) - \mathbb{E} U(1)\}^k]$ for $k=2,3,...$.
Using this representation we obtain the following preliminary result.
\begin{lemma}\label{lemma:workloadmoments}
Let $\mathbb{E}|U(1)|<\infty$, $u_2, u_3 < \infty$ and $\mu > \lambda$. If $Q_{\mu}(\infty)$ represents the steady-state distribution of the workload process, then
\begin{equation}
\mathbb{E}[Q_{\mu}(\infty)] = \frac{\lambda u_2}{2(\mu-\lambda)},\qquad \mathbb{E}[Q_{\mu}^2(\infty)]=\frac{\lambda^2u_2^2}{2(\mu-\lambda)^2} + \frac{\lambda u_3}{3(\mu-\lambda)}.
\end{equation}
\end{lemma}
\subsection{Finite horizon}
For the purpose of this paper, we are interested in the dynamics of the workload process within a fixed time frame of length $T>0$.
For all $0\leq t \leq T$, we assume that the parameters of the queue, $\lambda,\mu,u_2,u_3$, remain unchanged.
If at $t=0$ the queue is not in steady-state corresponding to the specified parameters of the starting period, the process $\{ Q_{\mu}(t)\,:t\in[0,T] \}$ differs from its stationary counterpart $Q_{\mu}(\infty)$.
To illustrate this, Figure \ref{fig:transientmeans} depicts the expected value $Q_{\mu}$ in a $M/M/1$ queue as a function of time for several initial workloads $Q(0)$ for a particular setting of $\lambda$ and $\mu$.
Clearly, transient behavior of $\mathbb{E}[Q_{\mu}(t)]$, for $Q(0) \neq Q_{\mu}(\infty)$, differs significantly from the steady-state mean with the same system parameters.
Note that even if $Q(0) \equiv \mathbb{E}[Q_{\mu}(\infty)]$, the time-dependent mean does not coincide with the steady-state mean. Moreover, $\mathbb{E}[Q_{\mu}(t)]$ is not even a strictly increasing nor decreasing function of time. This phenomenon is a consequence of the decomposition of the transient mean into one strictly increasing, and a strictly decreasing term for $Q(0)>0$, as was studied in \cite{Abate1987}.
Nonetheless, $Q_{\mu}(t)$ converges in distribution to $Q_{\mu}(\infty)$ as $t\to\infty$, if $\mu>\lambda$.
\begin{figure}
\centering
\begin{tikzpicture}[xscale=0.2,yscale=0.3]
\draw (0,0) -- coordinate (x axis mid) (50,0);
\draw (0,0) -- coordinate (y axis mid) (0,21);
\node[right] at (51,0) {$t$};
\node[rotate=90, above=0.7 cm] at (y axis mid) {$\mathbb{E}[Q_{\mu}(t)]$};
\draw[dashed, thick, gray] (0,10) -- coordinate (eq) (51,10);
\definecolor{col1}{rgb}{0.368417, 0.506779, 0.709798}
\definecolor{col2}{rgb}{0.880722, 0.611041, 0.142051}
\definecolor{col3}{rgb}{0.560181, 0.691569, 0.194885}
\definecolor{col4}{rgb}{0.922526, 0.385626, 0.209179}
\draw[->] (24,6.4) -- coordinate (a1) (21.65,8.49574);
\node[right=0.6cm,below=0.3cm] at (a1) {$Q(0)\equiv 0$};
\draw[->] (14,6.2) -- coordinate (a2) (13.,8.49434);
\node[right=0.2cm,below=0.3cm] at (a2) {$Q(0)\equiv10$};
\draw[->] (9,15.5) -- coordinate (a3) (7.5,13.7648);
\node[right=0.6cm,above=0.1cm] at (a3) {$Q(0)\equiv20$};
\draw[->] (40,12.5) -- coordinate (a4) (38,10.9712);
\node[right,above=0.3cm] at (a4) { $Q(0)\sim \exp\left(\tfrac{1}{15}\right)$ };
\foreach \x in {0,10,...,50}
\draw (\x,1pt) -- (\x,-10pt)
node[anchor=north] {\x};
\foreach \y in {5,10,15,20}
\draw (1pt,\y) -- (-20pt,\y)
node[anchor=east] {\y};
\draw[thick,color = col1] plot
file { means0.txt};
\draw[thick,color = col2] plot
file { means10.txt};
\draw[thick,color = col3] plot
file { means20.txt};
\draw[thick,color = col4] plot
file { meansExp.txt};
\end{tikzpicture}
\caption{Time-dependent mean workload in $M/M/1$ queue with $\lambda = 10$ and server speed $\mu=11$ for different initial states $Q(0)$. The dashed line depicts $\mathbb{E}Q_{\mu}(\infty)$.}
\label{fig:transientmeans}
\end{figure}
Since the time horizon of our analysis is limited to $t\leq T$, the process may not approach the steady-state distribution sufficiently close to appropriately use its steady-state properties for capacity allocation.
To overcome this disparity, we propose a way to include the influence of this transient phase in the capacity allocation problem.
\subsection{Cost structure}
As mentioned before, we are interested in balancing the QoS and efficiency of the queue by choosing the optimal server speed $\mu$.
The adjective \emph{optimal} indicates that we intend to choose the speed according to some objective function.
In our case, we conduct our analysis based on a cost function, which consists of a part accounting for the penalty for congestion in the system and a part for staffing cost. The cost value of both parts is governed by the variable $\mu$.
The instantaneous cost incurred at time $t$ equals
\begin{equation}
\mathbb{E}[Q_{\mu}(t)] + \alpha \mu,
\end{equation}
where $\alpha$ is a positive constant defining the \emph{relative staffing cost}.
Hence, the cost structure we apply is a combination of the transient mean of the workload process and a linear staffing cost.
Accumulated and normalized over the period $[0,T]$, the cost function on which the rest of this paper will be based equals
\begin{equation}\label{eq:PiT}
\Pi_{T}(\mu) := \frac{1}{T}\int_0^T\left( \mathbb{E}[Q_{\mu}(t)] + \alpha\mu\, \right) {\rm d} t
= \frac{1}{T} \int_0^T \mathbb{E}[Q_{\mu}(t)]\, {\rm d} t + \alpha\mu.
\end{equation}
We use shorthand notation for the normalized congestion costs:
\begin{equation}\label{eq:CTmu}
C_{T}(\mu) := \frac{1}{T}\int_0^T \mathbb{E}[Q_{\mu}(t)] {\rm d} t,
\end{equation}
and $C_{\infty}(\mu) = \mathbb{E}[Q_{\mu}(\infty)]$.
In order to compare the actual costs incurred over the interval $[0,T]$ to the cost function of the queue in stationary conditions, we define
\begin{equation}\label{eq:PiInf}
\Pi_{\infty}(\mu) := C_{\infty}(\mu) + \alpha \mu = \mathbb{E}[Q_\mu(\infty)] + \alpha\mu,
\end{equation}
which allows an explicit expression by Lemma \ref{lemma:workloadmoments}.
Also, note that by dominated convergence theorem
\begin{equation}
\lim_{T\to\infty} \Pi_{T}(\mu) = \Pi_{\infty}(\mu),
\end{equation}
for all $\mu$.
Rewriting \eqref{eq:PiT} gives the relation
\begin{align}
\Pi_{T}(\mu) &= \frac{1}{T}\int_0^{T} \left( \mathbb{E}[Q_{\mu}(t)] - \mathbb{E}[Q_{\mu}(\infty)] \right)\, {\rm d} t + \mathbb{E}[Q_{\mu}(\infty)] + \alpha\mu = \Omega_{T}(\mu) + \Pi_{\infty}(\mu).
\label{eq:decomp}
\end{align}
Section \ref{sec:analysis} is concerned with the analysis of the correction factor $\Omega_{T}(\mu)$.
Ultimately, we are concerned with the additional costs incurred by choosing the server speed through minimization of $\Pi_{\infty}(\mu)$ instead of $\Pi_{T}(\mu))$.
Therefore, we formulate the exact and approximate optimization problems as follows
\begin{equation}\label{eq:muStar}
\mu_T^\star := \arg\min_{\mu\geq 0} \Pi_{T}(\mu), \qquad \qquad \mu_\infty^\star := \arg\min_{\mu\geq 0} \Pi_{\infty}(\mu),
\end{equation}
\begin{equation}\label{eq:piStar}
\Pi_{T}^\star = \Pi_{T}(\mu_T^\star), \qquad \qquad \Pi_{\infty}^\star = \Pi_{\infty}(\mu_\infty^\star).
\end{equation}
In Section \ref{sec:optimization} we turn to the comparison of $\mu_T^{\star}$ and $\mu_\infty^\star$ as well as the \emph{optimality gap} $\Pi_{\infty}^\star - \Pi_{T}^\star$.
For sake of clarity, we omit the subscript $\lambda$ in our expressions if no ambiguity is possible.
\section{Analysis of the objective function \label{sec:analysis}}
From \eqref{eq:decomp} it is evident that, for finding an explicit characterization of $\Pi_{T}(\mu)$, it suffices to study the term $\Omega_T(\mu)$ in more detail. We start by stating the main result of this section, which describes the leading order behavior of $\Omega_T(\mu)$ as $T$ increases.
\begin{thm}
Let $X_\mu(t)$ be of the form \eqref{eq:Xlmprocess}. If $\mathbb{E}[Q(0)^2], \mathbb{E}[Q(0)^3] < \infty$ and $u_2,u_3 < \infty$, then
\begin{equation}
\Omega_T(\mu) = \frac{1}{2T(\mu-\lambda)}\left( \mathbb{E}[Q(0)^2] - \frac{\lambda^2 u_2^2}{2(\mu-\lambda)^2} - \frac{\lambda u_3}{3(\mu-\lambda)}\right) + O\left(\frac{1}{T^2}\right),
\end{equation}
for $\mu>\lambda$.
\end{thm}
Note that this expression provides an \emph{approximation} of the actual cost function $\Pi_T(\mu)$. We elaborate on the implications of this additional information on the optimization problem in Section 4.
In the remainder of this section we provide a detailed description of the steps taken to obtain this outcome. Proofs of the intermediate results can be found in Appendix A.
\subsection{Constructing a coupling}
Before starting our analysis with the correction term $\Omega_{T}(\mu)$ we introduce some auxiliary notation.
By $Q_\mu^A(t)$ we denote the workload process as described in Subsection \ref{sec:levymodel} with $Q(0)\,{\buildrel d \over =}\, A$ and $\mathbb{E}_A$ the expectation with respect to the random variable $A$.
To be able to compare $\mathbb{E}[Q_{\mu}^Z(t)]$ and $\mathbb{E}[Q_\mu(\infty)]$ as in $\Omega_T(\mu)$, we will use a coupling technique.
For brevity, denote by $Z$ a random variable for which $Z\,{\buildrel d \over =}\, Q_{\mu}(\infty)$. Then $Q_{\mu}(\infty) \,{\buildrel d \over =}\, Q_{\mu}^{Z}(t)$ for all $t \geq 0$ and $\mathbb{E}[Q_{\mu}(\infty)] = \mathbb{E}_Z[Q_{\mu}^{Z}(t)]$.
Hence, quantifying the difference between the transient and stationary mean is equivalent to comparing the workload processes of two queues starting in two different (random) states at $t=0$.
For now, assume $Q(0)\equiv x \geq 0$. Later, we relax this by replacing $x$ by the random variable $Q(0)$. In this subsection, we will omit the subscript $\mu$ for brevity.
Equation \eqref{eq:Qlm} shows that all randomness in $Q$ originates from the process $X(t)$.
With this in mind, we couple the processes $Q^x(t)$ and $Q^{Z}(t)$ on a sample path level by feeding both queues the same net-input process $X(t)$ for $t\geq 0$.
This allows us to compare the processes in the same probability space,
\begin{align}
\mathbb{E}[Q^x(t)] - \mathbb{E}[Q(\infty)] &= \mathbb{E}_{X}[Q^x(t)] - \mathbb{E}_{X}\hspace{-4pt}\left[\mathbb{E}_Z[Q^{Z}(t)]\right]\nonumber\\
&= \mathbb{E}_{Z}\left[\mathbb{E}_{X}\hspace{-4pt}\left[ Q^x(t) - Q^{Z}(t)\right]\right].
\end{align}
For brevity, we also replace $Q_{\mu}(\infty)$ by the variable $y$. At the end of our analysis we will obtain the original form by randomization.
Define
\begin{equation}
Y^{x,y}(t) := Q^x(t) - Q^y(t).
\end{equation}
Then
\begin{equation}
\Omega_{T}^{x,y} := \frac{1}{T}\,\int_0^T \mathbb{E}\left[Y^{x,0}(t)\right] \, {\rm d} t
\end{equation}
and
\begin{equation}
\Omega_{T} = \mathbb{E}_{Q(\infty)}\left[ \Omega_T^{x,Q_{\mu}(\infty)} (\mu) \right].
\end{equation}
A possible sample path triple for $Q^x(t)$, $Q^0(t)$ and $Y^{x,y}(t)$ is depicted in Figure \ref{fig:samplePaths}. As we see from this figure, $Y^{x,y}(t)$ has nice structural properties which we will exploit in the next subsection.
\begin{figure}
\centering
\begin{tikzpicture}[y=0.8cm, x=0.013cm]
\draw (0,0) -- coordinate (x axis mid) (800,0);
\draw (0,0) -- coordinate (y axis mid) (0,6.5);
\node[below=0.2cm] at (x axis mid) {$\to t$};
\node[rotate=90, above=0.2cm] at (y axis mid) {$Q(t)$};
\node[above=1.3cm,left =0.08 cm] at (y axis mid) {$x$};
\draw plot
file { samplePathLevy.txt};
\draw[color = gray] plot
file { samplePathLevy2.txt};
\draw[thick,color=red] plot
file { runningMinimumLevy.txt};
\end{tikzpicture}
\caption{Sample path visualization of the processes $Q^x(t)$ (solid), $Q^0(t)$ (gray) and $Y^{x,0}(t)$ (red).}
\label{fig:samplePaths}
\end{figure}
\subsection{Difference process and leading order behavior of the correction term}
We further examine the \emph{difference process} $Y^{x,y}(t)$. Let us assume that $x>y$. Recall from \eqref{eq:Qlm},
\begin{equation}\label{eq:Wz}
Q^z(t) = \max\{ z + X(t),\, \sup_{0<s\leq t} [X(t)-X(s)]\} = X(t) + \max\{ z, -\inf_{0\leq s\leq t} X(s)\},
\end{equation}
where $X(t)$ is a L\'evy process with no negative jumps.
Let $\tau^x(z)$, $z<x$ denote the first passage time of level $z$ by the process $Q^x$, i.e.
\begin{equation}
\tau^x(z) := \inf \left\{ t \geq 0\, |\, Q^x(t) < z \,\right\}.
\end{equation}
Then it is easily seen that
\begin{equation}
Q^z(t) = \left\{
\begin{array}{ll}
z + X(t), & {\rm if }\ t <\tau^z(0), \\
\sup_{0<s\leq t} [X(t)-X(s)], & {\rm if }\ t \geq \tau^z(0).
\end{array}\right.
\end{equation}
Consequently,
\begin{equation}\label{eq:Yxy}
Y^{x,y}(t) = \left\{
\begin{array}{ll}
x - y, & \text{if }t < \tau^y(0),\\
\inf_{0<s\leq t} \{ x+X(s)\}, & \text{if }\tau^y(0) \leq t < \tau^x(0),\\
0, & \text{if }\tau^x(0) \leq t.
\end{array}\right.
\end{equation}
Using this representation we can identify
\begin{equation}
\Omega^{x,y}_T = \frac{1}{T}\,\mathbb{E}\left[\int_0^{\tau^x(0)\wedge T} Y^{x,y}(t) {\rm d} t\right],
\end{equation}
where $\wedge$ denotes the minimum operator, due to the fact $Y^{x,y}(t) = 0$ for $t\geq \tau^x(0)$.
Subsequently, we decompose $\Omega_T^{x,y}$ into two terms
\begin{equation}
\Psi^{x,y}_T := \frac{1}{T} \int_0^\infty \mathbb{E}[Y^{x,y}(t)]\, {\rm d} t \qquad
\text{and}
\qquad
\Delta_T^{x,y} := \Omega_T^{x,y} - \Psi_T^{x,y}.
\label{eq:Deltaxy}
\end{equation}
Note that $\Psi_T^{x,y}$ is obtained by replacing $T$ by $\infty$ only in the integration bound.
This decomposition is insightful, because $\Psi_T^{x,y}$ prescribes the leading order behavior of $\Omega_T^{x,y}$, while $\Delta_T^{x,y}$ captures the smaller order error term.
In this section, we only consider $\Psi_T^{x,y}$. Subsection \ref{sec:trunc} investigates the magnitude of $\Delta_T^{x,y}$.
The next preliminary result presents a useful property of $\Psi_T^{x,y}$.
\begin{lemma}
Let $x>y$. If $\mathbb{E}[\tau^x(0)]<\infty$, then
\begin{equation}\label{eq:H(x,y)}
\Psi^{x,y}_T = \frac{1}{T}\,\mathbb{E}[\tau^{y}(0)](x-y) + \Psi^{x-y,0}_T.
\end{equation}
\end{lemma}
This leaves us with two unknowns $\mathbb{E}[\tau^y(0)]$ and $\Psi_T^{x-y,0}$.
The next lemma gives an equivalent form for the latter.
\begin{lemma}
For $z\geq 0$ and $\mathbb{E}[\tau^z(0)] < \infty$,
\begin{equation}\label{eq:H(x,0)}
\Psi^{z,0}_T = \int_0^z \mathbb{E}[\tau^y(0)]\, {\rm d} y.
\end{equation}
\end{lemma}
As the term $\mathbb{E}[\tau^y(0)]$ appears in many of the preliminary results, we devote attention to this in the next subsection.
\subsubsection*{First passage time}
When studying the first passage time $\tau^x(y)$ of the workload process starting $x$, we first observe that $\{\tau^x(z-y)\}_{y=0}^x$ is a spectrally positive L\'evy process itself.
More precisely, it is a subordinator, i.e. a L\'evy process whose paths are almost surely non-decreasing \cite{Kyprianou2006}.
In order to calculate $\mathbb{E}[\tau^x(x-y)]$ we use theory presented in \cite[Section 46]{Sato1999}, although results presented there are valid for spectrally \emph{negative} L\'evy processes, as opposed to the absence of negative jumps in our case.
Nonetheless, our setting is easily transformed into this framework by observing that $\hat{X} \equiv -X$, that is $\hat{X}(t) = -X(t)$ for all $t\geq 0$, is spectrally negative.
Furthermore, let
\begin{equation}
\label{eq:transformedTau}
\hat{\tau}^0(y) := \inf\{ t \geq 0\,:\, \hat{X}(t) > y\} = \inf\{ t \geq 0\,:\, x+X(t)< x-y\} = \tau^x(x-y).
\end{equation}
For completeness, we cite \cite[Thm~46.3]{Sato1999}.
\begin{thm}
Let $\hat{X}(t)$ be a spectrally negative L\'evy process with generating triplet $(-a,\sigma,\hat{\nu})$ and $\hat{\tau}^0(y)$ its corresponding hitting time process. Define $\Upsilon(\theta)$ for $\theta\geq 0$ as
\begin{equation}\label{eq:thmCharExp}
\Upsilon(\theta) = -a\theta + \tfrac{1}{2}\sigma^2\theta^2 + \int_{-\infty}^0 (e^{\theta x}-1-\theta x{\bf 1}_{[-1,0)}(x))\, \hat{\nu}({\rm d} x).
\end{equation}
Then $\Upsilon(\theta)$ is strictly increasing and continuous, $\Upsilon(0)=0$, and $\Upsilon(\theta)\to\infty$ as $\theta\to\infty$. For $x\geq 0$ and $0\leq u < \infty$ we have
\begin{equation}\label{eq:invCharExp}
\mathbb{E}[\exp(-u\hat{\tau}^0(y))] = \exp(-y\,\Upsilon^{-1}(u)),
\end{equation}
where $\theta=\Upsilon^{-1}(u)$ is the inverse function of $u=\Upsilon(\theta)$.
\end{thm}
This immediately induces an expression for $\Psi^{z,0}$.
\begin{cor}\label{cor:Psixy}
Let $X(t)$ be a spectrally positive L\'evy process defined as in \eqref{eq:Xlmprocess} with $\mu > \lambda$. Let $\Psi^{z,0}_T$ as in \eqref{eq:H(x,0)}. Then
\begin{equation}
\Psi^{z,0}_T = \frac{z^2}{2T(\mu-\lambda)}.
\end{equation}
Furthermore, if $x,y\geq 0$,
\begin{equation}\label{eq:mainResult}
\Psi^{x,y}_T = \frac{x^2-y^2}{2T(\mu-\lambda)}.
\end{equation}
\end{cor}
\subsubsection*{Randomization}
As we stated before, we easily obtain the original $\Omega_T$ from $\Omega_T^{x,y}$ through substitution of $x$ and $y$ by $Q(0)$ and $Q(\infty)$, respectively, and taking the expectation.
In the previous paragraph, we deduced an explicit expression for $\Psi_T^{x,y}$, the leading order term for $\Omega_T^{x,y}$.
Therefore we equivalently get an approximation for $\Omega_T$, given by
\begin{equation}
\Psi_T := \frac{1}{T} \int_0^\infty \left( \mathbb{E}[Q(t)]-\mathbb{E}[Q(\infty)] \right)\, {\rm d} t,
\end{equation}
through randomization of $x$ and $y$ in $\Psi_T^{x,y}$.
By combining the results in Corollary \ref{cor:Psixy} and Lemma \ref{lemma:workloadmoments} we directly prove the result in Theorem 1.
\subsection{Truncation error}\label{sec:trunc}
In order to get a better comprehension of the properties of $\Psi_T$, we depict the value in terms of the (infinite) region between the curves $\mathbb{E}[Q(t)]$, $\mathbb{E}[Q(\infty)]$ and the vertical axis for the case $Q(0)\equiv 0$ in Figure \ref{fig:PsiVisualization}.
In this figure, $\Omega_T$ is given by the area enclosed by the two curves, the vertical axis and the line $t=T$.
One can see that the main contribution to the correction term $\Omega_T$ is given for small $t$.
As $t$ increases, the difference between transient and stationary mean decreases.
Hence for moderate values of $T$, the contribution to the integral in \eqref{eq:Deltaxy} is only minor compared to the contribution over the interval $[0,T]$.
\begin{figure}
\centering
\begin{tikzpicture}[xscale=0.13,yscale=0.3]
\node[below=0.4cm,right=0.5cm] at (x axis mid) {$\to t$};
\draw[dashed, thick, fill =gray!30] (0,0) rectangle coordinate (eq) (50,10);
\node[] at (-7,10) {$\mathbb{E}[Q(\infty)]$};
\node[] at (-3,0) {$0$};
\draw[->] (18,6.4) coordinate (a1) -- (21.65,8.49574);
\node[below] at (a1) {$x=0$};
\foreach \x in {30}
\draw (\x,1pt) -- (\x,-10pt)
node[anchor=north] {$T$};
\foreach \y in {10}
\draw (1pt,\y) -- (-20pt,\y);
\draw[thick,color = gray,fill=white] plot
file {means0_2.txt};
\draw[thick] (0,0) -- coordinate (x axis mid) (50,0);
\draw[thick] (0,0) -- coordinate (y axis mid) (0,12);
\draw[color=white,very thick] (50,0.05) -- (50,9.56);
\draw[very thick, dotted] (30,0) -- (30,10);
\draw[->] (18,6.8) coordinate (delta) -- (17,8.7);
\node[below] at (delta) {$\Psi_{T}$};
\draw[->] (38,8.1) coordinate (delta) -- (36,9.7);
\node[below] at (delta) {$\Delta_{T}$};
\end{tikzpicture}
\caption{Visualization of $\Omega_T$ and $\Psi_T$ as the area between the curves $E[Q(t)]$, $\mathbb{E}[Q(\infty)]$ for $Q(0) = 0$.}
\label{fig:PsiVisualization}
\end{figure}
Recall the definition of $\Delta^{x,y}_T$ as in \eqref{eq:Deltaxy}. As we eluded to in Subsection 3.2
we claim the contribution of $\Delta^{x,y}_T$ to $\Omega_T^{x,y}$ is negligible compared to $\Psi^{x,y}_T$. Also note that
\begin{equation}
\label{eq:Delta}
\Delta_T := \Omega_T - \Psi_T = {-}\frac{1}{T} \int_T^\infty \mathbb{E}[Q(t)] - \mathbb{E}[Q(\infty)]\,{\rm d} t.
\end{equation}
can be derived through $\Delta^{x,y}_T$ in a similar manner as we did for $\Psi^{x,y}_T$ to obtain $\Psi_T$.
To substantiate our claim, we compute an upper bound for $\Delta^{x,y}_T$ of order $1/T^2$. The existence of such an upper bound poses a limit on the error this tail integral contributed to the cost structure as a whole.
\begin{prop} \label{prop:truncation_error}
Let $x,y\geq 0$ and $\mathbb{E}[\max\{Q(0),Q_\mu(\infty)\}^3] < \infty$. Then
\begin{equation}
|\Delta^{x,y}_T| \leq \frac{1}{T^2}\left(\frac{\max(y,x)^3}{3(\mu-\lambda)^2}+\frac{u_2 \max(y,x)^2}{2(\mu-\lambda)^3}\right)
\end{equation}
and
\begin{equation}
|\Delta_T| \leq \frac{1}{T^2}\left(\frac{\mathbb{E}[\max(Q(0),Q_\mu(\infty))^3]}{3(\mu-\lambda)^2}+\frac{u_2 \mathbb{E}[\max(Q(0),Q_\mu(\infty))^2]}{2(\mu-\lambda)^3}\right).
\end{equation}
\end{prop}
\section{Optimization \label{sec:optimization}}
The result in Theorem 1, characterizing the leading order behavior of $\Omega_T(\mu)$, also reveals the behavior of $\Pi_T(\mu)$ in leading order. Namely,
\begin{equation}
\Pi_T(\mu) = \Pi_\infty(\mu) + \Psi_T(\mu) + O(1/T^2).
\end{equation}
In fact, this representation naturally gives rise to an \emph{approximation} of the actual cost function:
\begin{align}\label{eq:decomposition}
\hat{\Pi}_{T}(\mu) := \Pi_{\infty}(\mu) + \Psi_T(\mu)
\end{align}
(We again include $\mu$ in the descriptions of variables derived in previous sections, because of the central role this decision variable will be playing within this section.)
Denote the corresponding minimizer of $\hat{\Pi}_{T}$ by
\begin{equation}\label{eq:muhat}
\hat{\mu}_T^\star := \arg\min_{\mu\geq 0} \hat{\Pi}_{T}(\mu), \qquad \hat{\Pi}_{T}^\star := \hat{\Pi}_{T}(\hat{\mu}_T^\star)
\end{equation}
in addition to the definitions in \eqref{eq:muStar} and \eqref{eq:piStar}.
This section is devoted to the analysis of the minimizers $\mu_T^\star$, $\hat{\mu}_T^\star$ and $\mu_\infty^\star$, and the optimality gap for the two approximations.
Throughout this section, we assume that $u_2, u_3 <\infty$ and $\mathbb{E}[Q(0)^2] <\infty$.
By its definition in \eqref{eq:PiInf} and Lemma \ref{lemma:workloadmoments}, we have an optimal expression for the steady-state cost function
\begin{equation}
\Pi_{\infty}(\mu) = \frac{\lambda u_2}{2(\mu-\lambda)} + \alpha\mu.
\end{equation}
It is easily verified that $\Pi_{\infty}$ is strictly convex in $\mu$, e.g. by observing that $\Pi_{\infty}''(\mu) > 0$ for all $\mu > \lambda$. Therefore $\Pi_{\infty}$ has a unique global minimizer and
\begin{equation}
\label{eq:muInf}
\mu_\infty^\star = \lambda + \sqrt{\frac{\lambda u_2}{2\alpha}}, \qquad \Pi_{\infty}^\star = \alpha\lambda + \sqrt{2\alpha\lambda u_2}.
\end{equation}
We are interested in the relation between $\mu_\infty^\star$ and $\mu_T^\star$, and $\hat{\mu}_T^\star$ and $\mu_T^\star$.
Since $\Pi_{T}(\mu) = \Pi_{\infty}(\mu) + O(1/T)$ for all $\mu > \lambda$, we have pointwise convergence of the sequence $\Pi_{T}$, as well as $\hat{\Pi}_{T}$, to $\Pi_{\infty}$ for $T\to\infty$, we also expect $\mu_T^\star \to \mu_\infty^\star$ and $\hat{\mu}_T^\star\to\mu_\infty^\star$ for $T\to\infty$.
Before proving that this convergence indeed holds, we result a result on the strict convexity of the function $\Pi_{T}$.
\begin{lemma}
Let $\mu\geq 0$. The function $\Pi_{T}(\mu)$ is
\begin{itemize}
\item convex in $\mu$, if $Q(0)\equiv x$, $T<x/\mu$ and $\sigma=0$,
\item strictly convex in $\mu$, otherwise.
\end{itemize}
\end{lemma}
Building upon strict convexity of both $\Pi_T(\mu)$ and $\Pi_\infty(\mu)$ for $\mu>\lambda$, we derive the following convergence result. The proof can be found in Appendix B.
\begin{prop}
Let $\mu_T^\star$, $\hat{\mu}_T^\star$ and $\mu_\infty^\star$ be as defined in \eqref{eq:muStar} and \eqref{eq:muhat}. Then
\begin{equation}
\mu_T^\star \to \mu_\infty^\star\, \qquad \text{\rm and } \qquad \hat{\mu}_T^\star \to \mu_\infty^\star,
\end{equation}
for $T\to\infty$.
\end{prop}
The next result describes a refinement of $\mu_T^\star$ in terms of $\mu_\infty^\star$.
\begin{prop}\label{prop:muBullet}
For $T$ sufficiently large,
\begin{equation}
\mu_T^\star = \mu_\infty^\star + \frac{\mu_\bullet}{T} + o(1/T),
\end{equation}
where
\begin{equation}\label{eq:muBullet}
\mu_\bullet = \frac{\mathbb{E}[Q(0)^2]}{\sqrt{8\lambda u_2\alpha}} - \frac{u_3}{3 u_2} - 3\sqrt{\frac{\alpha\lambda u_2}{8}}.
\end{equation}
\end{prop}
Based on Proposition \ref{prop:muBullet} we propose a \emph{corrected staffing rule}, accounting for the finite horizon
\begin{equation}
\label{eq:correctedMu}
\tilde{\mu}_T^\star = \left[\mu_\infty^\star + \frac{\mu_\bullet}{T}\right]^+,
\end{equation}
with $\mu_\bullet$ as in \eqref{eq:muBullet}.
Here $[x]^+ := \max\{x,0\}$, which ensures the value of $\tilde{\mu}_T^\star$ is non-negative and thus is a feasible solution of the optimization problem.
This refined capacity allocation rule is expected to reduce the costs incurred in transient settings.
However, the value we are particularly interested in is the cost increase for using either one of the approximations rather than the actual minimum $\mu_T^\star$, that is, the \emph{optimality gap}.
As it happens, we deduce the order of the optimality gap for $\mu_\infty^\star$ with the help of the explicit form of $\mu_\bullet$ given in \eqref{eq:correctedMu}, which is stated in the next proposition. The proof is given in Appendix \ref{sec:proofProp4}
\begin{prop}\label{prop:optimalitygap_mui}
Let $\mu_\infty^\star$ be as in \eqref{eq:muInf}. Then,
\begin{equation}
\Pi_T(\mu_\infty^\star) - \Pi_T^\star = O(1/T^2).
\end{equation}
\end{prop}
\section{Numerical experiments}
\label{sec:numerics}
\subsection{Influence of $\Omega_{T}(\mu)$}
We first assess the contribution of the correction to the cost function provided by Theorem 1. In other words, we investigate whether $\hat{\Pi}_{T}(\mu)$ as in \eqref{eq:PiT} yields a significantly better fit to $\Pi_{T}(\mu)$, than $\Pi_{\infty}(\mu)$ does.
Note that these three functions only differ in the costs describing the congestion.
Therefore, we limit our study in this subsection to the evaluation of $C_T(\mu)$ as in \eqref{eq:CTmu} with stationary equivalent $C_{\infty}(\mu) = \mathbb{E}[Q_{\mu}(\infty)]$.
Our novel approximation hence reads
\begin{equation}
\hat{C}_{T}(\mu) := C_{\infty}(\mu) + \Omega_{T}(\mu),
\end{equation}
with $\Omega_{T}(\mu)$ given in \eqref{eq:mainResult}.
We conduct our numerical experiments based on three models, namely:
\begin{enumerate}
\item \underline{$M/M/1$ queue}: $U(t)$ is a unit rate compound Poisson process with exponentially distributed increments. We have $u_2 = 2$, $u_3=3$, so that
\begin{equation}\label{eq:MM1cor}
\hat{C}_{T}(\mu) = \frac{\lambda}{\mu-\lambda} + \frac{1}{T(\mu-\lambda)} \left(\frac{x^2}{2} - \frac{\lambda^2}{(\mu-\lambda)^2} - \frac{\lambda}{\mu-\lambda}\right).
\end{equation}
\item \underline{$M/{\rm Pareto}/1$ queue}: $U(t)$ is a unit rate compound Poisson process with Pareto increments. The Pareto distribution deserves special attention due to its heavy-tailed nature, having tail probability $\bar{F}(x) = (x/k)^{-\gamma}$, if $x\geq k$ and 1 otherwise.
It is well-known that heavy-tailed service times lead to long relaxation time. For our purposes, we fix shape parameter $\gamma = 16/5$ and scale parameter $k=11/16$, so that $\beta = 1$, $u_2 = 121/96$, $u_3 = 1331/256$ and $u_k=\infty$ for all $k>3$. Hence,
\begin{equation}
\label{eq:MP1cor}
\hat{C}_{T}(\mu) = \frac{121\lambda}{192(\mu-\lambda)} + \frac{1}{2T(\mu-\lambda)}
\left( x^2 - \frac{(121\lambda/96)^2}{2(\mu-\lambda)^2} - \frac{ 1331\lambda/256 }{2(\mu-\lambda)}\right)
\end{equation}
\item \underline{Reflected Brownian motion}: $U(t)$ is Brownian motion with drift 1 and infinitesimal variance $\sigma^2$. We have $u_2 = \sigma^2$, $u_3=0$, so that
\begin{equation}\label{eq:RBMcor}
\hat{C}_{T}(\mu) = \frac{\lambda\sigma^2}{2(\mu-\lambda)} + \frac{1}{2T(\mu-\lambda)} \left( x^2 - \frac{\lambda^2\sigma^4}{2(\mu-\lambda)^2}\right).
\end{equation}
\end{enumerate}
Let $C_T^{\lambda}(\mu)$ denote the cost function given arrival rate $\lambda$.
Although we want to explore a variety of parameter settings for these three settings, one can deduce from the identity in \eqref{eq:Qidentity} that $C_{T}^{\lambda}(\mu) \equiv C_{\lambda T}^{1}(\mu/\lambda)$ and $\Omega_{T}^\lambda(\mu) \equiv \Omega_{\lambda T}^1(\mu/\lambda)$.
This implies that it suffices to evaluate the systems for $\lambda=1$, since we directly obtain the measures for any other value of $\lambda$ by scaling the variable $\mu$ and parameter $T$ appropriately.
For the $M/M/1$ and $M/{\rm Pareto}/1$ queue, we obtained the function $C_{T}(\mu)$ with $\lambda=1$ through simulation and are accurate up until a 95\% confidence interval of width $10^{-3}$. For reflected Brownian motion, we used the explicit distribution function given in \cite{Harrison1985} for double numerical integration. The results for several values of $T$ and two different starting states are depicted in Figures 4-6. These plots also include the approximated functions $\hat{C}_{T}(\mu)$.
\begin{figure}[h]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[scale=0.55]{MM1_L1_x0}
\caption{$x=0$}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[scale=0.55]{MM1_L1_x25}
\caption{$x=2.5$}
\end{subfigure}
\caption{$C_{T}(\mu)$ as a function of $\mu$ for $M/M/1$ for $T=2$ (blue), $T=5$ (yellow) and $ T=10$ (green) with their approximative equivalents $\hat{C}_{T}(\mu)$ (dashed) and $C_{\infty}(\mu)$ (black).}
\end{figure}
\begin{figure}[h!]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[scale=0.55]{MP1_L1_x0}
\caption{$x=0$}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[scale=0.55]{MP1_L1_x25}
\caption{$x=2.5$}
\end{subfigure}
\caption{$C_{T}(\mu)$ as a function of $\mu$ for $M/{\rm Pareto}/1$ for $T=2$ (blue), $T=5$ (yellow) and $T=10$ (green) with their approximative equivalents $\hat{C}_{T}(\mu)$ (dashed) and $C_{\infty}(\mu)$ (black).}
\end{figure}
\begin{figure}[h!]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[scale=0.55]{RBM_L1_x0}
\caption{$x=0$}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[scale=0.55]{RBM_L1_x25}
\caption{$x=2.5$}
\end{subfigure}
\caption{$C_{T}(\mu)$ as a function of $\mu$ for RBM with $\sigma=1$ for $T=2$ (blue), $T=5$ (yellow) and $T=10$ (green) with their approximative equivalents $\hat{C}_{T}(\mu)$ (dashed) and $C_{\infty}(\mu)$ (black).}
\end{figure}
We name a few observations based on these figures.
First, we indeed note the pointwise convergence of $\hat{C}_{T}(\mu)$ to $\hat{C}_{\infty}(\mu)$ as $T$ grows, for all $\mu$ in all three cases. However, the difference between the stationary costs and those for small values of $T$ can be significant. This is most clear in the plots with $x=2.5$ and when $\mu$ is close to $\lambda$, i.e. it is in heavy-traffic. In these scenarios, it is evident that refinements of the stationary costs are needed. $\hat{C}_{T}(\mu)$ does a fairly good job at providing such correction, especially for moderate values of $\mu$.
Furthermore, we note that $C_{T}(\mu)$ approaches $C_{\infty}(\mu)$ from below for $x=0$ for any value of $\mu$, while this is not strictly the case for $x>0$.
$\hat{C}_{T}(\mu)$ correctly captures the sign of this correction.
Finally, observe that $\hat{C}_{T}(\mu)\to -\infty$ as $\mu$ approaches $\lambda$. This divergence is clear from the expressions in \eqref{eq:MM1cor}-\eqref{eq:RBMcor}.
Our correction term relies on the premise that under the coupling scheme, the sample paths of the two queues starting from different states have hit with high probability.
This is equivalent to saying that the `largest' of the two queues is has emptied at least once before time $T$. However, as $\mu$ approaches $\lambda$, the system enters heavy traffic, and hence the hitting time of the zero barrier is set to run off to infinity.
Consequently, this causes our approximation to be inaccurate for small values of $\mu$.
\subsection{Validation of corrected staffing rule}
In this section, we examine whether the corrected staffing rule $\tilde{\mu}_T^\star$ as in \eqref{eq:correctedMu} indeed yields a significant cost reduction over the choice of $\mu_\infty^\star$ by comparing their true costs $\Pi_{T}(\tilde{\mu}_T^\star)$ and $\Pi_{T}(\mu_\infty^\star)$.
We conduct this comparison for different values of the parameters, $\alpha$, $T$ and starting state $x$ through numerical experiments.
The three models on which we do our calculations are the $M/M/1$ queue, the $M/$Pareto$/1$ queue and the reflected Brownian motion, as introduced in the previous subsection.
Extending the reasoning of the previous subsection, saying that the cost function for general $\lambda>0$ can be fit into the case $\lambda=1$ while applying additional scaling, we focus on the latter scenario only.
For each of the three models, we adhere to the following set-up. The quality of both staffing rules is assessed for $\alpha = 0.1, 1$ and 2, resembling three modes of valuation of the QoS in the system.
As a benchmark, observe that the expected workload in steady-state conditions with staffing level $\mu_\infty^\star$ equals
\begin{equation}
C_\infty(\mu_\infty^\star) = \sqrt{\frac{\alpha\lambda u_2}{2}}.
\end{equation}
For each value of $\alpha$, we consider two scenarios: one in which the system starts empty, i.e. $x=0$, and one in which the initial state is double this benchmark value, thus $x=\sqrt{2\alpha\lambda u_2}$.
The numerics will be presented for each model separately. Afterwards, we discuss the conclusions we are able to draw from these results.
\subsubsection*{$M/M/1$ queue}
As we discussed before, if $U$ is a unit rate compound Poisson process with exponentially distributed increments, then $Q_{\mu}$ describes the workload process in an $M/M/1$ queue.
For this setting we get
\begin{equation}
\mu_\infty^\star = \lambda + \sqrt{\frac{\lambda}{\alpha}},\qquad \tilde{\mu}_T^\star = \left[\lambda + \sqrt{\frac{\lambda}{\alpha}} + \frac{1}{T}\left( \frac{x^2}{4\sqrt{\lambda\alpha}} - 1 - \frac{3}{2} \sqrt{\lambda\alpha}\right)\right]^+.
\end{equation}
Table \ref{tab:mm1} presents the actual costs corresponding to these two staffing levels for different value of $x$ and $\alpha$.
\begin{table}[h!]
\centering\resizebox{15cm}{!} {
\begin{tabular}{|c|r|rr|rr|r||rr|rr|r|}
\cline{3-12}
\multicolumn{1}{c}{} & \multicolumn{1}{r|}{} & \multicolumn{5}{c||}{$x = 0$} & \multicolumn{5}{c|}{$x = 2\sqrt{\alpha}$} \\
\hline
$\alpha$ & $T$ & $\mu_\infty^\star$ & $\Pi_T(\mu_\infty^\star)$ & $\tilde{\mu}_T^\star$ & $\Pi_T(\tilde{\mu}_T^\star)$ & \% & $\mu_\infty^\star$ & $\Pi_T(\mu_\infty^\star)$ & $\tilde{\mu}_T^\star$ & $\Pi_T(\tilde{\mu}_T^\star)$ & \% \\
\hline
\multicolumn{1}{|c|}{\multirow{4}[2]{*}{0.1}} & 1 & 4.162 & 0.620 & 2.688 & 0.536 & 0.136 & 4.162 & 0.682 & 2.688 & 0.536 & 0.214 \\
\multicolumn{1}{|c|}{} & 2 & 4.162 & 0.669 & 3.425 & 0.641 & 0.041 & 4.162 & 0.700 & 3.425 & 0.641 & 0.085 \\
\multicolumn{1}{|c|}{} & 5 & 4.162 & 0.706 & 3.867 & 0.703 & 0.005 & 4.162 & 0.719 & 3.867 & 0.703 & 0.022 \\
\multicolumn{1}{|c|}{} & 10 & 4.162 & 0.719 & 4.015 & 0.719 & 0.001 & 4.162 & 0.726 & 4.015 & 0.719 & 0.010 \\
\hline
\hline
\multicolumn{1}{|c|}{\multirow{4}[2]{*}{1}} & 1 & 2.000 & 2.309 & 0.000 & 0.500 & 0.783 & 2.000 & 3.500 & 0.500 & 2.750 & 0.214 \\
\multicolumn{1}{|c|}{} & 2 & 2.000 & 2.461 & 0.750 & 1.480 & 0.398 & 2.000 & 3.218 & 1.250 & 3.125 & 0.029 \\
\multicolumn{1}{|c|}{} & 5 & 2.000 & 2.675 & 1.500 & 2.400 & 0.103 & 2.000 & 3.043 & 1.700 & 2.968 & 0.025 \\
\multicolumn{1}{|c|}{} & 10 & 2.000 & 2.810 & 1.750 & 2.726 & 0.030 & 2.000 & 3.007 & 1.850 & 2.980 & 0.009 \\
\hline
\hline
\multicolumn{1}{|c|}{\multirow{4}[2]{*}{2}} &1 & 1.707 & 3.744 & 0.000 & 0.500 & 0.866 & 1.707 & 5.889 & 0.000 & 3.328 & 0.435 \\
\multicolumn{1}{|c|}{} &2 & 1.707 & 3.924 & 0.146 & 1.232 & 0.686 & 1.707 & 5.547 & 0.854 & 4.682 & 0.156 \\
\multicolumn{1}{|c|}{} &5 & 1.707 & 4.209 & 1.083 & 3.343 & 0.206 & 1.707 & 5.114 & 1.366 & 4.910 & 0.040 \\
\multicolumn{1}{|c|}{} &10 & 1.707 & 4.424 & 1.395 & 4.108 & 0.071 & 1.707 & 4.945 & 1.536 & 4.868 & 0.016 \\
\hline
\end{tabular}}
\caption{Comparison of costs for the $M/M/1$ queue for steady-state and corrected staffing rules.}
\label{tab:mm1}
\end{table}
\subsubsection*{M/{\rm Pareto}/1 queue}
In case the service requirements follow a Pareto distribution with shape parameter $\gamma = 16/5$, the staffing rules become
\begin{equation}
\mu_\infty^\star = \lambda + \frac{11}{8}\sqrt{\frac{ \lambda }{3 \alpha}},\qquad \tilde{\mu}_T^\star = \left[\lambda + \frac{11}{8}\sqrt{\frac{ \lambda }{3 \alpha}} + \frac{1}{T}\left( \frac{2 x^2}{11\sqrt{\lambda\alpha/3}} - \frac{11}{8} - \frac{11}{16} \sqrt{3\lambda\alpha}\right)\right]^+.
\end{equation}
The numerical results are given in Table \ref{tab:mp1}.
\begin{table}[h!]
\centering\resizebox{15cm}{!} {
\begin{tabular}{|c|r|rr|rr|r||rr|rr|r|}
\cline{3-12}
\multicolumn{1}{c}{} & \multicolumn{1}{r|}{} & \multicolumn{5}{c||}{$x = 0$} & \multicolumn{5}{c|}{$x = 11/4\cdot \sqrt{\alpha/3}$} \\
\hline
$\alpha$ & $T$ & $\mu_\infty^\star$ & $\Pi_T(\mu_\infty^\star)$ & $\tilde{\mu}_T^\star$ & $\Pi_T(\tilde{\mu}_T^\star)$ & \% & $\mu_\infty^\star$ & $\Pi_T(\mu_\infty^\star)$ & $\tilde{\mu}_T^\star$ & $\Pi_T(\tilde{\mu}_T^\star)$ & \% \\
\hline
\multicolumn{1}{|c|}{\multirow{4}[2]{*}{0.1}} & 1 & 3.510 & 0.524 & 1.759 & 0.461 & 0.120 & 3.510 & 0.573 & 2.010 & 0.562 & 0.019 \\
\multicolumn{1}{|c|}{} & 2 & 3.510 & 0.555 & 2.635 & 0.539 & 0.029 & 3.510 & 0.580 & 2.760 & 0.574 & 0.010 \\
\multicolumn{1}{|c|}{} & 5 & 3.510 & 0.580 & 3.160 & 0.578 & 0.003 & 3.510 & 0.591 & 3.210 & 0.589 & 0.002 \\
\multicolumn{1}{|c|}{} & 10 & 3.510 & 0.590 & 3.335 & 0.590 & 0.000 & 3.510 & 0.596 & 3.360 & 0.595 & 0.001 \\
\hline
\hline
\multicolumn{1}{|c|}{\multirow{4}[2]{*}{1}} & 1 & 1.794 & 2.076 & 0.000 & 0.500 & 0.759 & 1.794 & 2.989 & 0.000 & 2.088 & 0.302 \\
\multicolumn{1}{|c|}{} & 2 & 1.794 & 2.190 & 0.511 & 1.291 & 0.411 & 1.794 & 2.790 & 0.610 & 2.588 & 0.072 \\
\multicolumn{1}{|c|}{} & 5 & 1.794 & 2.345 & 1.281 & 2.108 & 0.101 & 1.794 & 2.638 & 1.320 & 2.607 & 0.012 \\
\multicolumn{1}{|c|}{} & 10 & 1.794 & 2.441 & 1.537 & 2.371 & 0.029 & 1.794 & 2.597 & 1.557 & 2.585 & 0.005 \\
\hline
\hline
\multicolumn{1}{|c|}{\multirow{4}[2]{*}{2}} & 1 & 1.561 & 3.427 & 0.000 & 0.500 & 0.854 & 1.561 & 5.087 & 0.000 & 2.745 & 0.460 \\
\multicolumn{1}{|c|}{} & 2 & 1.561 & 3.567 & 0.032 & 1.050 & 0.706 & 1.561 & 4.832 & 0.172 & 3.417 & 0.293 \\
\multicolumn{1}{|c|}{} & 5 & 1.561 & 3.779 & 0.950 & 3.012 & 0.203 & 1.561 & 4.499 & 1.006 & 4.313 & 0.041 \\
\multicolumn{1}{|c|}{} & 10 & 1.561 & 3.935 & 1.255 & 3.356 & 0.147 & 1.561 & 4.351 & 1.284 & 4.304 & 0.011 \\
\hline
\end{tabular}
}
\caption{Comparison of costs for the $M/{\rm Pareto}/1$ queue for steady-state and corrected staffing rules.}
\label{tab:mp1}
\end{table}
Just as in the results for the $M/M/1$ queue, we observe a higher reduction for larger value of $\alpha$ and $T$. Also, again $\tilde{\mu}_T < \mu_\infty^\star$. Hence, the conclusions for the $M/{\rm Pareto}/1$ queue are similar to those of the $M/M/1$ queue.
\subsubsection*{Reflected Brownian motion}
In case the input process $U$ is Brownian motion with drift 1 and infinitesimal variance $\sigma^2$, the steady-state staffing rule and its corrected version reduce to
\begin{equation}
\mu_\infty^\star = \lambda + \sqrt{\frac{\lambda\sigma^2}{2\alpha}}, \qquad
\tilde{\mu}_T^\star = \left[\lambda + \sqrt{\frac{\lambda\sigma^2}{2\alpha}} + \frac{1}{2\sqrt{2}\,T}\left(\frac{x^2}{\sqrt{\lambda \alpha}\sigma} - 3\sigma\sqrt{\alpha\lambda} \right)\right]^+.
\end{equation}
In Tables \ref{tab:rbm1} and \ref{tab:rbm2}, the costs obtained through numerical evaluation are presented for several values of $x$, $T$. We also vary $\sigma$ to examine the influence of the volatility of arrival process on the quality of the staffing rules.
The observations on the influence of $\alpha, x$ and $T$ are similar to those of the $M/M/1$ queue and the $M/{\rm Pareto}/1$ queue.
However, here we see little improvement induced by the corrected staffing rule for small values of $\alpha$ for both values of $x$.
The results in Tables \ref{tab:rbm1}-\ref{tab:rbm2} also suggest that the reduction is smaller for larger values of $\sigma$.
\begin{table}
\centering
\resizebox{15cm}{!} {
\begin{tabular}{|c|r|rr|rr|r||rr|rr|r|}
\cline{3-12}
\multicolumn{1}{c}{} & \multicolumn{1}{r|}{} & \multicolumn{5}{c||}{$x = 0$} & \multicolumn{5}{c|}{$x = \sqrt{2\alpha} $} \\
\hline
$\alpha$ & $T$ & $\mu_\infty^\star$ & $\Pi_T(\mu_\infty^\star)$ & $\tilde{\mu}_T^\star$ & $\Pi_T(\tilde{\mu}_T^\star)$ & \% & $\mu_\infty^\star$ & $\Pi_T(\mu_\infty^\star)$ & $\tilde{\mu}_T^\star$ & $\Pi_T(\tilde{\mu}_T^\star)$ & \% \\
\hline
\multicolumn{1}{|c|}{\multirow{4}[2]{*}{0.1}} & 1 & 3.236 & 0.525 & 2.901 & 0.518 & 0.013 & 3.236 & 0.565 & 3.124 & 0.564 & 0.001 \\
\multicolumn{1}{|c|}{} & 2 & 3.236 & 0.536 & 3.068 & 0.534 & 0.003 & 3.236 & 0.556 & 3.180 & 0.556 & 0.000 \\
\multicolumn{1}{|c|}{} & 5 & 3.236 & 0.543 & 3.169 & 0.542 & 0.000 & 3.236 & 0.551 & 3.214 & 0.551 & 0.000 \\
\multicolumn{1}{|c|}{} & 10 & 3.236 & 0.545 & 3.203 & 0.545 & 0.000 & 3.236 & 0.549 & 3.225 & 0.549 & 0.000 \\
\hline
\hline
\multicolumn{1}{|c|}{\multirow{4}[2]{*}{1}} & 1 & 1.500 & 3.420 & 0.000 & 0.833 & 0.756 & 1.500 & 4.741 & 1.000 & 3.984 & 0.160 \\
\multicolumn{1}{|c|}{} & 2 & 1.500 & 3.539 & 0.750 & 2.386 & 0.326 & 1.500 & 4.579 & 1.250 & 4.293 & 0.063 \\
\multicolumn{1}{|c|}{} & 5 & 1.500 & 3.707 & 1.200 & 3.363 & 0.093 & 1.500 & 4.335 & 1.400 & 4.274 & 0.014 \\
\multicolumn{1}{|c|}{} & 10 & 1.500 & 3.820 & 1.350 & 3.705 & 0.030 & 1.500 & 4.190 & 1.450 & 4.175 & 0.004 \\
\hline
\hline
\multicolumn{1}{|c|}{\multirow{4}[2]{*}{2}} & 1 & 1.500 & 3.420 & 0.000 & 0.833 & 0.756 & 1.500 & 4.741 & 1.000 & 3.984 & 0.160 \\
\multicolumn{1}{|c|}{} & 2 & 1.500 & 3.539 & 0.750 & 2.386 & 0.326 & 1.500 & 4.579 & 1.250 & 4.293 & 0.063 \\
\multicolumn{1}{|c|}{} & 5 & 1.500 & 3.707 & 1.200 & 3.363 & 0.093 & 1.500 & 4.335 & 1.400 & 4.274 & 0.014 \\
\multicolumn{1}{|c|}{} & 10 & 1.500 & 3.820 & 1.350 & 3.705 & 0.030 & 1.500 & 4.190 & 1.450 & 4.175 & 0.004 \\
\hline
\end{tabular}}
\caption{Comparison of costs for RBM with $\sigma = 1$ for steady-state and corrected staffing rules}
\label{tab:rbm1}
\end{table}
\begin{table}
\centering
\resizebox{15cm}{!} {
\begin{tabular}{|c|r|rr|rr|r||rr|rr|r|}
\cline{3-12}
\multicolumn{1}{c}{} & \multicolumn{1}{r|}{} & \multicolumn{5}{c||}{$x = 0$} & \multicolumn{5}{c|}{$x = 2\sqrt{2\alpha} $} \\
\hline
$\alpha$ & $T$ & $\mu_\infty^\star$ & $\Pi_T(\mu_\infty^\star)$ & $\tilde{\mu}_T^\star$ & $\Pi_T(\tilde{\mu}_T^\star)$ & \% & $\mu_\infty^\star$ & $\Pi_T(\mu_\infty^\star)$ & $\tilde{\mu}_T^\star$ & $\Pi_T(\tilde{\mu}_T^\star)$ & \% \\
\hline
\multicolumn{1}{|c|}{\multirow{4}[2]{*}{0.1}} & 1 & 5.472 & 0.950 & 4.801 & 0.936 & 0.015 & 5.472 & 1.030 & 5.249 & 1.029 & 0.001 \\
\multicolumn{1}{|c|}{} & 2 & 5.472 & 0.972 & 5.137 & 0.968 & 0.003 & 5.472 & 1.012 & 5.360 & 1.012 & 0.000 \\
\multicolumn{1}{|c|}{} & 5 & 5.472 & 0.985 & 5.338 & 0.985 & 0.000 & 5.472 & 1.002 & 5.427 & 1.002 & 0.000 \\
\multicolumn{1}{|c|}{} & 10 & 5.472 & 0.990 & 5.405 & 0.990 & 0.000 & 5.472 & 0.998 & 5.450 & 0.998 & 0.000 \\
\hline
\hline
\multicolumn{1}{|c|}{\multirow{4}[2]{*}{1}} & 1 & 2.414 & 3.176 & 0.293 & 1.546 & 0.513 & 2.414 & 4.633 & 1.707 & 4.228 & 0.087 \\
\multicolumn{1}{|c|}{} & 2 & 2.414 & 3.356 & 1.354 & 2.690 & 0.199 & 2.414 & 4.375 & 2.061 & 4.247 & 0.029 \\
\multicolumn{1}{|c|}{} & 5 & 2.414 & 3.573 & 1.990 & 3.411 & 0.045 & 2.414 & 4.094 & 2.273 & 4.073 & 0.005 \\
\multicolumn{1}{|c|}{} & 10 & 2.414 & 3.689 & 2.202 & 3.646 & 0.012 & 2.414 & 3.966 & 2.344 & 3.962 & 0.001 \\
\hline
\hline
\multicolumn{1}{|c|}{\multirow{4}[2]{*}{2}} & 1 & 2.000 & 4.839 & 0.000 & 1.339 & 0.723 & 2.000 & 7.481 & 1.000 & 5.967 & 0.202 \\
\multicolumn{1}{|c|}{} & 2 & 2.000 & 5.078 & 0.500 & 2.773 & 0.454 & 2.000 & 7.158 & 1.500 & 6.585 & 0.080 \\
\multicolumn{1}{|c|}{} & 5 & 2.000 & 5.414 & 1.400 & 4.726 & 0.127 & 2.000 & 6.670 & 1.800 & 6.549 & 0.018 \\
\multicolumn{1}{|c|}{} & 10 & 2.000 & 5.639 & 1.700 & 5.409 & 0.041 & 2.000 & 6.380 & 1.900 & 6.349 & 0.005 \\
\hline
\end{tabular}}
\caption{Comparison of costs for RBM with $\sigma = 2$ for steady-state and corrected staffing rules.}
\label{tab:rbm2}
\end{table}
\subsection{Discussion}
Based upon these numerical results in Tables \ref{tab:mm1}-\ref{tab:rbm2}, we make a few remarks. The three models roughly exhibit similar behavior as $T$, $x$ and $\alpha$ are varied.
Non-surprisingly, we note that $\tilde{\mu}_T$ approaches $\mu_\infty^\star$ with increasing $T$, which also implies that the cost reduction achieved by the corrected staffing rule vanishes as $T\to\infty$.
Also, we observe that in all scenarios examined, the cost reduction increases with $\alpha$. This can be explained through investigation of the objective function $\Pi_T$ as function of $\mu$. Namely, for $\alpha$ small, the curve is relatively flat around the true optimum $\mu_T^\star$. Hence, in this case a moderate deviation from $\mu_T^\star$ will likely not lead to a significant cost increase. However, as $\alpha$ becomes larger, i.e. the server efficiency is valued more than minimization of congestion, the curve becomes more sharp around $\mu_T^\star$, and hence more accurate approximations of $\mu_T^\star$ are required to achieve an acceptable cost level. Hence, the corrected staffing rule \eqref{eq:correctedMu} proves particularly useful in these cases.
Another point we want to highlight is that the relative improvement is higher for $x=0$, as opposed to $x=\sqrt{2\alpha\lambda u_2}$. Moreover, even though the initial state of the system is above the optimal equilibrium, $\tilde{\mu}_T$ is smaller than $\mu_\infty^\star$. This is somewhat counter-intuitive. In fact, from \eqref{eq:muBullet} it follows that $\mu_\bullet$ positively contributes to the corrected staffing function if
\begin{equation}
\mathbb{E}[Q^2(0)] > 3\alpha\lambda u_2 + \frac{2 u_2}{3 u_3}\,\sqrt{2\alpha\lambda u_2}.
\end{equation}
Even more surprisingly, obverse that if $Q(0)\equiv Q(\infty)$ with $\mu \equiv \mu_\infty^\star = \lambda + \sqrt{\lambda u_2/(2\alpha)}$ we get with Lemma \ref{lemma:workloadmoments}
\begin{equation}
\mathbb{E}[Q^2(0)] = \alpha\lambda u_2 + \frac{u_3}{3} \sqrt{\frac{2\alpha\lambda}{u_2}},
\end{equation}
so that
\begin{equation}
\mu_\bullet = {-}\frac{u_3}{2 u_2} - \sqrt{2\alpha\lambda u_2} < 0.
\end{equation}
This suggests that even when the process is started in equilibrium with the corresponding optimal steady-state speed $\mu_\infty^\star$, it is more cost efficient to change the server speed. This seems strange, but we provide an explanation for this phenomenon.
In out particular setting, we strictly focus on the period $[0,T]$, and do not care about what happens after time $T$. Hence, it might be beneficial to let the queue build up towards the end of the period, thereby employing a smaller server speed than stipulated by the steady-state optimum.
Naturally, this effect diminishes with $T$.
\section{Conclusion \& further research}
Motivated by the time-varying nature of queues in practical applications, we studied the impact that the transient phase has on traditional capacity allocation questions.
By defining a cost minimization problem, in which the objective function contains a correction accounting for the transient period, we identified the leading and second-order behavior of the cost function as a function of the interval length $T$.
As a by-product, this result provides an approximation for the actual cost function, which is a refinement to its stationary counterpart.
Our numerical experiments in Section 5.1 demonstrate the improved accuracy achieved by this approximation in a number of settings.
By perturbation analysis of the optimization problem, this furthermore gives rise to a correction to the steady-state optimal capacity allocation of order $1/T$.
The necessity of the refined capacity allocation level is substantiated by the numerics in Section 5.2, which show the cost reduction that can be achieved in the number of settings, compared to settings in which stationary metrics are used.
Especially for small values of $T$ and large values of $\alpha$ this reduction is significant.
Additionally, these results also indicate that it is relatively safe to use the stationary cost when $T$ is moderate, or $\alpha$ is small.
The latter reflects the scenario in which QoS to clients is much more valued than service efficiency.
This observation links to the flat nature of the cost function around its optimal value for $\alpha$ small, a statement on the optimality gap that we formally proved in Proposition 4.
Besides the validation of our theoretical results of Sections 3 and 4, the numerical results also reveal some phenomena that require more investigation.
As noted, our corrected capacity allocation level $\tilde{\mu}_T^\star$ is in most studied cases less than the steady-state optimal value $\mu_{\infty}^\star$. This implies that congestion levels tends to be higher under our staffing scheme then under stationary staffing.
A possible explanation for this may be the fact that the planning period under consideration is finite.
Clearly, in the setting we analyzed, anything that happens after time $T$ is neglected.
Therefore, it might be beneficial from the cost perspective to end the period with a higher expected congestion level, as it does not need to be canceled out in the future.
Related to this observation, it would be interesting to look at the setting in which staffing decisions need to be made in consecutive periods of equal length, in which the arrival rate changes at the start of each period.
This case requires careful consideration of the correlation among the staffing decisions within the separate periods.
Another question that arises concerns the translation of our (qualitative) findings to more general queues, in particular the $M/M/s$ queue.
Whereas in our analysis, the central decision variable is the server speed $\mu$, the variable of interest in multi-server queues is typically the number of servers.
It may well be that similar explicit corrections to staffing levels can be deduced to account for transience.
Since our analysis heavily relies on the comparibility of the sample paths of two single-server queues, which is due to the equal negative drift for the two processes, another approach must be taken to tackle this extension.
The analysis and findings for the single-server queue with L\'evy input presented in this paper may serve a stepping stone for investigation of these more elaborate problems.
\bibliographystyle{plain}
|
\section{Introduction}
{\it Unique path partitions} were introduced by Olsson in \cite{OlssAA}.
Their study is motivated from the Murnaghan--Nakayama rule for the
calculation of the value of characters of the symmetric group.
They were completely characterised by Bessenrodt, Olsson and
Sellers in \cite{BeOSAA}. They used this characterisation to derive
a formula for the generating function for the number $u(n)$ of all
unique path partitions of~$n$. This formula reads
(cf.\ \cite[Remark~3.6]{BeOSAA})
\begin{align} \notag
\sum_{n\ge1}u(n)q^n&=2\sum_{i\ge1}q^{2^i-1}(1+q^{2^{i-1}})
\prod _{j=0} ^{i-2}\frac {1} {1-q^{2^j}}\\
&=
2\left(
{q(1+q)}
+\sum_{i\ge2}\frac {q^{-1}+1} {1-q^2}\cdot
\frac {q^{2^i}(1+q^{2^{i-1}})} {\prod_{j=1}^{i-2}(1-q^{2^{j}})}
\right).
\label{eq:UGF}
\end{align}
The final part in \cite{BeOSAA} concerns congruences modulo~8 for $u(n)$.
The corresponding
main result \cite[Theorem~4.6]{BeOSAA} provides a complete description
of the behaviour of $u(n)$ modulo~8 (in terms of the related sequence
of numbers $w(n)$; see the next section for the definition of~$w(n)$).
The arguments to arrive at this result are mainly of a recursive nature.
The purpose of this note is to show that a more convenient and more
powerful method to derive congruences (modulo powers of~$2$ is by an analysis
of the generating function \eqref{eq:UGF}. Not only are we able to
recover the result from \cite{BeOSAA}, but in addition
we succeed in determining the congruence class of $u(n)$ modulo~16,
see \eqref{eq:wu} and Theorem~\ref{thm:w}, thus solving the problem
left open in the last paragraph of \cite{BeOSAA}.
We point out that the approach presented here is very much
inspired by calculations in \cite[Appendix]{KrMuAE}, where expressions
similar to the one on the right-hand side of \eqref{eq:UGF} appear,
with the role of the prime number~2 replaced by~3, though.
\section{An equivalent expression for the generating function}
We start with the observation (already made in \cite{BeOSAA}) that,
first, all numbers $u(n)$ are divisible by~$2$, and, second, we have
$u(2n)=u(2n-1)$ for all $n$. This is easy to see from the right-hand
side of \eqref{eq:UGF} since it has the form $2(1+q)f(q^2)$, where
$f(t)$ is a formal power series in~$t$.
We therefore divide the right-hand side of \eqref{eq:UGF} by $2(1+q^{-1})$,
subsequently replace $q$ by $q^{1/2}$,
and consider the ``reduced" generating function
\begin{align*}
\notag
\sum_{n\ge2}w(n)q^n&=\sum_{i\ge2}q^{2^{i-1}}(1+q^{2^{i-2}})
\frac {1} {(1-q)\prod _{j=0} ^{i-3}(1-q^{2^j})}.
\end{align*}
In other words, we have
\begin{equation} \label{eq:wu}
2w(n)=u(2n)=u(2n-1)
\end{equation}
for all~$n$.
Using the convention
\begin{equation} \label{eq:SUM}
\sum _{k=M} ^{N-1}\text {\rm Expr}(k)=\begin{cases}
\hphantom{-}
\sum _{k=M} ^{N-1} \text {\rm Expr}(k),&N>M,\\
\hphantom{-}0,&N=M,\\
-\sum _{k=N} ^{M-1}\text {\rm Expr}(k),&N<M.\end{cases}
\end{equation}
for sums, we rewrite the above equation in the following way:
\begin{align}
\notag
\sum_{n\ge2}w(n)q^n&=\sum_{i\ge2}q^{2^{i-1}}
\frac {1+q^{2^{i-2}}} {(1-2q+q^2)\prod _{j=1} ^{i-3}(1-q^{2^j})}\\
\notag
&=\sum_{i\ge2}q^{2^{i-1}}
\frac {1+q^{2^{i-2}}}
{(1-\frac {2q} {1+q^2})(1+q^2)\prod _{j=1} ^{i-3}(1-q^{2^j})}\\
\notag
&=\sum_{i\ge2}q^{2^{i-1}}
\frac {1+q^{2^{i-2}}}
{(1-\frac {2q} {1+q^2})(1-q^4)\prod _{j=2} ^{i-3}(1-q^{2^j})}\\
\notag
&=\sum_{i\ge2}q^{2^{i-1}}
\frac {1+q^{2^{i-2}}}
{(1-\frac {2q} {1+q^2})(1-q^4)^2\prod _{j=3} ^{i-3}(1-q^{2^j})}\\
\notag
&\hphantom{{}={}}\hbox to 6cm{\leaders\hbox to .3cm{\hss.\hss}\hfill}\\
\notag
&=\sum_{i\ge1}q^{2^{2i-1}}
\frac {1+q^{2^{2i-2}}}
{(1-\frac {2q} {1+q^2})(1-\frac {2q^4} {1+q^8})\cdots(1-\frac {2q^{2^{2i-4}}}
{1+q^{2^{2i-3}}})
(1-q^{2^{2i-2}})}\\
\notag
&\kern1cm
+\sum_{i\ge1}q^{2^{2i}}
\frac {1}
{(1-\frac {2q} {1+q^2})(1-\frac {2q^4} {1+q^8})\cdots(1-\frac {2q^{2^{2i-2}}}
{1+q^{2^{2i-1}}})
}\\
&=\sum_{i\ge1}q^{2^{2i-1}}
\frac {1+\frac {2q^{2^{2i-2}}} {1-q^{2^{2i-2}}}}
{\prod _{j=0} ^{i-2}(1-\frac {2q^{2^{2j}}}
{1+q^{2^{2j+1}}})
}
+\sum_{i\ge1}q^{2^{2i}}
\frac {1}
{\prod _{j=0} ^{i-1}(1-\frac {2q^{2^{2j}}}
{1+q^{2^{2j+1}}})}
.
\label{eq:WGF2}
\end{align}
From the last expression it is immediately obvious that $w(n)$ is odd
if and only if $n$ is a power of~$2$, thus recovering the first
assertion of \cite[Cor.~4.3]{BeOSAA}. The above mentioned mod-4 result
\cite[Theorem~4.6]{BeOSAA} for $w(n)$ --- which, by \eqref{eq:wu},
translates into a mod-8 result for~$u(n)$ ---
can also be derived within a few lines from the above expression.
In the next section, we show how to obtain congruences modulo~$8$
for~$w(n)$, which, by \eqref{eq:wu}, translate into congruences
modulo~16 for the unique path partition numbers~$u(n)$.
\section{Congruences modulo powers of $2$}
In what follows, we write
$$f(q)=g(q)~\text {modulo}~2^\gamma$$
to mean that the coefficients
of $q^i$ in $f(q)$ and $g(q)$ agree modulo~$2^\gamma$ for all $i$.
We apply geometric series expansion in \eqref{eq:WGF2}, and at the
same time we neglect terms which are divisible by~$8$. For example,
we expand
$$
\frac 1 {1-\frac {2q} {1+q^2}}=1+\frac {2q} {1+q^2}+
\frac {4q^2} {(1+q^2)^2}\quad
\text{modulo }8.
$$
In this manner, we obtain the congruence
\begin{align}
\notag
\sum_{n\ge2}w(n)q^n&=
\sum_{i\ge1}
q^{2^{2i-1}}\Bigg(
1+\frac {2q^{2^{2i-2}}} {1-q^{2^{2i-2}}}
+2\sum_{j=0}^{i-2}\frac {q^{2^{2j}}}
{1+q^{2^{2j+1}}}\\
\notag
&\kern1.5cm
+4\frac {q^{2^{2i-2}}} {1-q^{2^{2i-2}}}
\sum_{j=0}^{i-2}\frac {q^{2^{2j}}}
{1+q^{2^{2j+1}}}
+4\sum_{0\le s\le t\le i-2}^{}\frac {q^{2^{2s}+2^{2t}}}
{(1+q^{2^{2s+1}})(1+q^{2^{2t+1}})}
\Bigg)\\
\notag
&\kern.5cm
+\sum_{i\ge1}q^{2^{2i}}
\Bigg(1
+2\sum_{j=0}^{i-1}\frac {q^{2^{2j}}}
{1+q^{2^{2j+1}}}
+4\sum_{0\le s\le t\le i-1}\frac {q^{2^{2s}+2^{2t}}}
{(1+q^{2^{2s+1}})(1+q^{2^{2t+1}})}
\Bigg)\\
\notag
&\kern11.5cm
\text{modulo }8.
\end{align}
After rearrangement, this becomes
\begin{align}
\notag
\sum_{n\ge2}w(n)q^n&=
\sum_{i\ge1}q^{2^i}
+\frac {2q^3} {1-q}
+2\sum_{j\ge1}\frac {1}
{1-q^{2^{2j}}}\Bigg(
q^{2^{2j}+2^{2j+1}}
+q^{2^{2j-2}}(1-q^{2^{2j-1}})\sum_{\ell\ge 2j}q^{2^\ell}
\Bigg)\\
\notag
&\kern1cm
+4\sum_{1\le s<t}\frac {q^{2^{2s-2}+2^{2t-2}}}
{(1-q^{2^{2s-1}})(1-q^{2^{2t-1}})}
\Bigg(
q^{2^{2t-1}}(1+q^{2^{2t-2}})
+\sum_{\ell\ge2t}q^{2^\ell}
\Bigg)\\
&\kern1cm
+4\sum_{s\ge1}\frac {q^{2^{2s-1}}}
{(1-q^{2^{2s}})}
\sum_{\ell\ge2s}q^{2^\ell}
\quad \quad \quad \quad
\text{modulo }8.
\label{eq:Wcong1}
\end{align}
We must now analyse the individual sums in \eqref{eq:Wcong1}.
\begin{lemma} \label{lem:S1}
Let $n\ge2$, and write
$n=\sum_{i=a}^e n_i\cdot 2^i$, with $0\le n_i\le1$ for all~$i$
and $n_a\ne0\ne n_e$.
Then the coefficient of $q^n$ in
\begin{equation} \label{eq:S1}
\sum_{j\ge1}\frac {q^{2^{2j}+2^{2j+1}}}
{1-q^{2^{2j}}}
\end{equation}
is equal to $\fl{a/2}$ if $n$ is not a power of\/ $2$, and it is
equal to $\max\{\fl{a/2}-1,0\}$ otherwise.
\end{lemma}
\begin{proof}
By geometric series expansion, we see that the coefficient of $q^n$ in
\eqref{eq:S1} is equal to the number of possibilities to write
$n=(k+3)2^{2j}$ for some $j\ge1$ and $k\ge0$. For fixed $j$, we can
find a suitable $k$ if and only if $n\ge 3\cdot 2^{2j}$.
If $n$ is not a power of $2$,
this is equivalent to the condition that $2j\le a$.
The claim follows immediately.
\end{proof}
\begin{lemma} \label{lem:S2}
Let $n\ge2$, and write
$n=\sum_{i=a}^e n_i\cdot 2^i$ as in Lemma~\ref{lem:S1}.
Then the coefficient of $q^n$ in
\begin{equation} \label{eq:S2}
\sum_{j\ge1}
\frac {q^{2^{2j-2}}}
{1-q^{2^{2j}}}
\sum_{\ell\ge 2j}q^{2^\ell}
\end{equation}
is equal to $e-2j+1$ if $a=2j-2$, $n_{a+1}=n_{2j-1}=0$,
and $n$ is not a power of $2$,
and it is equal to $0$ otherwise.
\end{lemma}
\begin{proof}
By geometric series expansion, we see that the coefficient of $q^n$ in
\eqref{eq:S2} is equal to the number of possibilities to write
$n=2^{2j-2}+k\cdot 2^{2j}+2^\ell$ for some $j\ge1$, $\ell\ge2j$, and $k\ge0$.
The claim follows immediately.
\end{proof}
\begin{lemma} \label{lem:S3}
Let $n\ge2$, and write
$n=\sum_{i=a}^e n_i\cdot 2^i$ as in Lemma~\ref{lem:S1}.
Then the coefficient of $q^n$ in
\begin{equation} \label{eq:S3}
\sum_{j\ge1}
\frac {q^{2^{2j-2}+2^{2j-1}}}
{1-q^{2^{2j}}}
\sum_{\ell\ge 2j}q^{2^\ell}
\end{equation}
is equal to $e-2j+1$ if $a=2j-2$, $n_{a+1}=n_{2j-1}=1$,
and it is equal to $0$ otherwise.
\end{lemma}
\begin{proof}
By geometric series expansion, we see that the coefficient of $q^n$ in
\eqref{eq:S3} is equal to the number of possibilities to write
$n=2^{2j-2}+2^{2j-1}+k\cdot 2^{2j}+2^\ell$ for some $j\ge1$,
$\ell\ge2j$, and $k\ge0$.
The claim follows immediately.
\end{proof}
\begin{lemma} \label{lem:S5}
Let $n\ge2$, and write
$n=\sum_{i=a}^e n_i\cdot 2^i$ as in Lemma~\ref{lem:S1}.
Then the coefficient of $q^n$ in
\begin{equation} \label{eq:S5}
\sum_{s\ge1}
\frac {q^{2^{2s-1}}}
{1-q^{2^{2s}}}
\sum_{\ell\ge 2s}q^{2^\ell}
\end{equation}
is equal to $e-2s+1$ if $a=2s-1$
and $n$ is not a power of $2$,
and it is equal to $0$ otherwise.
\end{lemma}
\begin{proof}
By geometric series expansion, we see that the coefficient of $q^n$ in
\eqref{eq:S5} is equal to the number of possibilities to write
$n=2^{2s-1}+k\cdot 2^{2s}+2^\ell$ for some $s\ge1$, $\ell\ge2s$, and $k\ge0$.
The claim follows immediately.
\end{proof}
\begin{lemma} \label{lem:S4}
Let $n\ge2$, and write
$n=\sum_{i=a}^e n_i\cdot 2^i$ as in Lemma~\ref{lem:S1}.
Then the coefficient of $q^n$ in
\begin{equation} \label{eq:S4}
\sum_{1\le s<t}\frac {q^{2^{2s-2}+2^{2t-2}}}
{(1-q^{2^{2s-1}})(1-q^{2^{2t-1}})}
\sum_{\ell\ge2t-1}q^{2^\ell}
\end{equation}
is congruent to
\begin{equation} \label{eq:S4erg}
e\sum_{i=a+2}^{e-\chi(e\text{ \em even})}
n_{i}
-a\cdot n_{a+2}
+\fl{\tfrac {1} {2}(e-a-1)}
\quad \quad
\text{\em (mod~2)},
\end{equation}
where $\chi(\mathcal S)=1$ if $\mathcal S$ is
true and $\chi(\mathcal S)=0$ otherwise.
\end{lemma}
\begin{proof}
By geometric series expansion, we see that the coefficient of $q^n$ in
\eqref{eq:S4} is equal to the number of possibilities to write
\begin{equation} \label{eq:k1k2a}
n=(2k_1+1)2^{2s-2}+(2k_2+1)2^{2t-2}+2^{2t-1+k_3}
\end{equation}
for some $s$ and $t$ with
$1\le s<t$ and $k_1,k_2,k_3\ge0$.
Clearly, we need $a$ to be even in order that
the number of these possibilities be non-zero.
Given that $a=2s-2$,
we just have to count the number of possible triples $(t,k_2,k_3)$ in
\eqref{eq:k1k2a}, since the appropriate $k_1$ can certainly be found.
If we fix $t$ and $k_3$, the number of possible $k_2$'s is
$$
\fl{\frac {1} {2}\cdot \frac {n-2^{2t-1+k_3}} {2^{2t-2}}+\frac {1} {2}}
=
\fl{\frac {n} {2^{2t-1}}+\frac {1} {2}}-2^{k_3}.
$$
This needs to be summed over all $t$ and $k_3$
with $\frac {1} {2}(a+2)=s<t\le \frac {1} {2}(e+1)$ and $0\le k_3\le e-2t+1$.
We obtain
\begin{align*}
\sum_{t=s+1}^{\fl{\frac {1} {2}(e+1)}}
&\sum_{k_3=0}^{e-2t+1}
\left(\fl{\frac {n} {2^{2t-1}}+\frac {1} {2}}-2^{k_3}\right)\\
&
\equiv
\sum_{t=\frac {1} {2}(a+4)}^{\fl{\frac {1} {2}(e+1)}}
\sum_{k_3=0}^{e-2t+1}
\big\lfloor
n_a\cdot 2^{a-2t+1}+\dots+(n_{2t-2}+1)\cdot 2^{-1}\\
&\kern4cm
+n_{2t-1}+n_{2t}\cdot 2+\dots +n_e\cdot 2^{e-2t+1}
\big\rfloor
-\fl{\tfrac {1} {2}(e-a-1)}\\
&
\equiv
\sum_{t=\frac {1} {2}(a+4)}^{\fl{\frac {1} {2}(e+1)}}
(e-2t+2)
(n_{2t-2}+n_{2t-1})
+\fl{\tfrac {1} {2}(e-a-1)}\\
&
\equiv
e\sum_{t=\frac {1} {2}(a+4)}^{\fl{\frac {1} {2}(e+1)}}
n_{2t-1}
+e\sum_{t=\frac {1} {2}(a+4)}^{\fl{\frac {1} {2}(e-1)}}
n_{2t}
+(e-a)n_{a+2}
+\fl{\tfrac {1} {2}(e-a-1)}
\quad \quad (\text{mod }2).
\end{align*}
\end{proof}
\begin{lemma} \label{lem:S5a}
Let $n\ge2$, and write
$n=\sum_{i=a}^e n_i\cdot 2^i$ as in Lemma~\ref{lem:S1}.
Then the coefficient of $q^n$ in
\begin{equation} \label{eq:S5a}
\sum_{1\le s<t}\frac {q^{2^{2s-2}+2^{2t}}}
{(1-q^{2^{2s-1}})(1-q^{2^{2t-1}})}
\end{equation}
is congruent to
\begin{equation} \label{eq:S5aerg}
\sum_{t=\frac {1} {2}(a+4)}^{\fl{\frac {1} {2}(e+1)}}
n_{2t-1}
+\fl{\tfrac {1} {2}(e-a-1)}.
\quad \quad
\text{\em (mod~2)}.
\end{equation}
\end{lemma}
\begin{proof}
By geometric series expansion, we see that the coefficient of $q^n$ in
\eqref{eq:S5a} is equal to the number of possibilities to write
\begin{equation} \label{eq:k1k2a2}
n=(2k_1+1)2^{2s-2}+(k_2+2)2^{2t-1}
\end{equation}
for some $s$ and $t$ with
$1\le s<t$ and $k_1,k_2\ge0$.
Clearly again, we need $a$ to be even in order that
the number of these possibilities be non-zero.
Given that $a=2s-2$,
we just have to count the number of possible pairs $(t,k_2)$ in
\eqref{eq:k1k2a2}, since the appropriate $k_1$ can certainly be found.
If we fix $t$, the number of possible $k_2$'s is
$$
\fl{\frac {n-2^{2t}} {2^{2t-1}}+1}
=
\fl{\frac {n} {2^{2t-1}}}-1.
$$
This needs to be summed over all $t$
with $\frac {1} {2}(a+2)=s<t\le \frac {1} {2}(e+1)$.
We obtain
\begin{align*}
\sum_{t=s+1}^{\fl{\frac {1} {2}(e+1)}}
\left(\fl{\frac {n} {2^{2t-1}}}-1\right)
&\equiv
\sum_{t=\frac {1} {2}(a+4)}^{\fl{\frac {1} {2}(e+1)}}
\big\lfloor
n_a\cdot 2^{a-2t+1}+\dots+n_{2t-2}\cdot 2^{-1}\\
&\kern3cm
+(n_{2t-1}-1)+n_{2t}\cdot 2+\dots +n_e\cdot 2^{e-2t+1}
\big\rfloor\\
&
\equiv
\sum_{t=\frac {1} {2}(a+4)}^{\fl{\frac {1} {2}(e+1)}}
n_{2t-1}
-\fl{\tfrac {1} {2}(e-a-1)}
\quad \quad (\text{mod }2).
\qedhere
\end{align*}
\end{proof}
We are finally in the position to state and prove our main result.
It expresses the congruence class of $w(n)$ modulo~8 --- and thus,
by \eqref{eq:wu}, the congruence class of the unique path partition
number~$u(n)$ modulo~16 --- in terms of the binary digits of~$n$.
We point out that the assertion \eqref{eq:w1} already appeared
in \cite[Prop.~4.5]{BeOSAA}.
\begin{theorem} \label{thm:w}
Let $n\ge2$, and write
$n=\sum_{i=a}^e n_i\cdot 2^i$ as in Lemma~\ref{lem:S1}.
Then, if $a=e$ {\em(}i.e., if $n$ is a power of~$2${\em)},
the number $w(n)$ is congruent to
\begin{equation} \label{eq:w1}
2\fl{a/2}+1\quad \text{\em(mod~8)},
\end{equation}
while it is congruent to
\begin{multline}
2+2\fl{a/2}+2\chi(a\text{ \em even})(1-2n_{a+1})(e-a-1)
+4\chi(a\text{ \em odd})(e-a)\\
+4\chi(a\text{ \em even})\Bigg(e\sum_{i=a+2}^{e-\chi(e\text{ \em even})}
n_{i}
+a\cdot n_{a+2}
+\sum_{t=\frac {1} {2}(a+4)}^{\fl{\frac {1} {2}(e+1)}}
n_{2t-1}
\Bigg)\quad \text{\em(mod 8)}
\label{eq:w2}
\end{multline}
otherwise.
\end{theorem}
\begin{proof}
Let first $n=2^a$. We must then read the coefficient of $q^n$ on the
right-hand side of \eqref{eq:Wcong1} and reduce the result modulo~8.
Non-zero contributions come from the very first sum, from the series
$2q^3/(1-q^2)$, and from the series which is discussed in Lemma~\ref{lem:S1}.
Altogether, we obtain
$$1+2\chi(a\ge2)+2\max\{\fl{a/2}-1,0\},$$
which can be simplified to \eqref{eq:w1}.
Now let $n$ be different from a power of $2$.
The non-zero contributions when reading the coefficient of $q^n$ on
the right-hand side of \eqref{eq:Wcong1} come again from the series
$2q^3/(1-q^2)$, and from the series discussed in
Lemmas~\ref{lem:S1}--\ref{lem:S5a}. These contributions add up to
\begin{multline*}
2\chi(n\ge3)+2\fl{a/2}+2\chi(a\text{ even, $n_{a+1}=0$})(e-a-1)\\
+2\chi(a\text{ even, $n_{a+1}=1$})(e-a-1)
+4\chi(a\text{ odd})(e-a)\\
+4\chi(a\text{ even})\Bigg(e\sum_{i=a+2}^{e-\chi(e\text{ even})}
n_{i}
-a\cdot n_{a+2}+\fl{\tfrac {1} {2}(e-a-1)}
+\sum_{t=\frac {1} {2}(a+4)}^{\fl{\frac {1} {2}(e+1)}}
n_{2t-1}+\fl{\tfrac {1} {2}(e-a-1)}
\Bigg).\kern-4pt
\end{multline*}
This expression can be simplified to result in \eqref{eq:w2}.
\end{proof}
It is clear that, in the same way,
one could also derive a result for $w(n)$ modulo~16, 32, \dots,
albeit at the cost of considerably more work.
|
\section{Introduction}
The design of compliant surfaces for turbulent skin friction reduction has attracted significant attention since the early experiments of Kramer \cite{Kramer1961}. However, despite many experimental \cite[e.g.\,][]{Bushnell1977,GadelHak1984,Lee1993,Choi1997,Zhang2015} and numerical \cite[e.g.\,][]{Endo2002,Xu2003,Fukagata2008,Kim2014} efforts, there are few definitive results. Broadly, the direct numerical simulations (DNS) and experiments both show that \textit{softer} surfaces often give rise to energetic two-dimensional (i.e.\, spanwise constant) wave-like motions, which can cause a substantial increase in skin friction. \textit{Harder} surfaces appear to have little impact on the flow, although some qualitative flow visualization experiments hint at an intermittent relaminarization-like phenomenon \cite{Lee1993}.
One of the major challenges associated with developing performance-enhancing surfaces is the extent of the parameter space to be explored. Even the simplest spring-damper walls considered in DNS depend on three independent parameters: a mass ratio, a spring constant and a scholasdamping coefficient. The viscoelastic layers tested frequently in experiments \cite{GadelHak1984,Lee1993} depend on at least five different parameters: two elastic constants which determine the shear- and longitudinal wave speeds, the mass density, a viscous relaxation time, and the layer thickness. Independent evaluation and optimization of these parameters in experiments would be very time-consuming and expensive, while current computational capabilities limit DNS-based compliant wall design to low Reynolds numbers or small domain sizes. The limitations of DNS-based design of compliant walls are well illustrated by the evolutionary optimization of anisotropic compliant walls pursued by Fukagata et al. \cite{Fukagata2008}. Specifically, these simulations showed that the best walls obtained in channel flow DNS over a small domain of length $3h$ ($h$ is the channel half-height), led to a near-$200\%$ increase in drag when the domain length was doubled to $6h$. Further, these DNS were limited to low Reynolds numbers. The bulk Reynolds number was $Re_B = 2U_Bh/\nu = 3300$, where $U_B$ is the bulk-averaged flow speed and $\nu$ is kinematic viscosity.
The impracticality of experimental or numerical approaches in designing performance-enhancing walls indicates the need for a computationally inexpensive theoretical framework to study turbulence-compliant wall interactions. In an effort to address this need, Luhar et al. \cite{Luhar2015} recently extended the resolvent formulation proposed by McKeon and Sharma \cite{McKeon2010}. Under this formulation, the turbulent flow field is expressed as a superposition of propagating velocity response modes, identified via a gain-based decomposition of the Navier-Stokes equations (NSE). Compliant surfaces are introduced via changes in the kinematic and dynamic boundary conditions. In particular, a complex wall admittance is used to define the relationship between the pressure and wall-normal velocity at the wall. This change in the boundary conditions leads to a change in the gain and structure of the modes, whereby a reduction in gain is interpreted as mode suppression.
Luhar et al. \cite{Luhar2015} show that this approach predicts the amplification of the quasi two-dimensional structures observed recently in DNS \cite{Kim2014} with minimal computation. Further, the formulation also enables an optimization of surface properties (i.e.\, wall admittance) to suppress flow structures known to be energetic in wall turbulence. This material-blind optimization suggests that walls with negative damping are required to suppress the near-wall (NW) cycle, identified by various researchers as essential to controlling wall turbulence \cite[e.g.\,][]{Bushnell1977}. However, walls with positive damping could be effective against the so-called superstructures or very-large-scale motions (VLSMs) that appear at high Reynolds number. Unfortunately, Luhar et al. \cite{Luhar2015} show that the optimal walls identified via this procedure also have negative effects elsewhere in spectral space, with slow-moving spanwise-constant structures particularly susceptible to further amplification.
The purpose of the present paper is to build on the above findings and evaluate the effect of varying wall models in greater detail, looking closely at the sensitivity to two-dimensional structures. While Luhar et al. \cite{Luhar2015} focused primarily on a spring-damper wall, this paper introduces the effects of tension, stiffness and anisotropy, and considers the effects of varying mass ratios to contrast aerodynamic and hydrodynamic applications. In addition, Reynolds number effects are explored briefly, and the framework is used to provide further insight into results from the aforementioned DNS-based optimization of anisotropic compliant walls pursued by Fukagata et al. \cite{Fukagata2008}, hereafter referred to as F2008.
One of the limitations of the resolvent formulation in its present form is the requirement of a mean velocity profile in the construction of the resolvent operator. As such, the smooth and compliant wall mean velocity profiles from F2008 are also used to evaluate the sensitivity of the resolvent-based predictions to the specific form of the mean profile.
\section{Theory}
This section provides a brief review of the resolvent formulation proposed by McKeon and Sharma \cite{McKeon2010}, the extension to account for the effects of compliant walls developed by Luhar et al. \cite{Luhar2015}, and the wall model employed in DNS by F2008 \cite{Fukagata2008}.
\subsection{Resolvent Formulation}\label{sec:theory-resolvent}
The resolvent formulation proposed by McKeon and Sharma \cite{McKeon2010} considers the full turbulent velocity field, $\mathbf{u}$, to be a superposition of highly amplified velocity structures, or modes, identified via a gain-based decomposition of the Fourier-transformed Navier-Stokes equations (NSE). For each wavenumber-frequency combination $\mathbf{k} = (\kappa_x,\kappa_z,c=\omega/\kappa_x)$, where $\kappa_x$ and $\kappa_z$ are the streamwise and spanwise wavenumbers, $\omega$ is the frequency and $c$ is the phase speed, the NSE are interpreted as a forcing-response system\footnote{This paper focuses on turbulent channel flows but the approach can be generalized to pipe and boundary layer flows as well.}:
\vspace{-0.5cm}
\begin{align}\label{eqNSE}
\left[\begin{array}{c} \mathbf{u_k} \\ p_\mathbf{k} \end{array}\right]
=
\left( -i\omega \left[\begin{array}{cc} \mathbf{I} & \\ &0\\ \end{array}\right] -
\left[\begin{array}{cc} \mathbf{\cal L}_\mathbf{k} & -\nabla_\mathbf{k}\\ \nabla_\mathbf{k}^T & 0\\ \end{array}\right] \right)^{-1}
\left[\begin{array}{cc} \mathbf{I} \\ 0\\ \end{array}\right] \mathbf{f_k}
=
\mathbf{\cal H}_\mathbf{k} \mathbf{f_k}
\end{align}
{\noindent}The nonlinear terms are interpreted as the forcing to the system, $(\mathbf{u} \cdot \nabla \mathbf{u})_\mathbf{k} = \mathbf{f_k}(y) \exp {i(\kappa_x x + \kappa_z z - \omega t)}$, and the resolvent operator, $\mathbf{\cal H}_\mathbf{k}$, maps this forcing to the velocity and pressure responses, e.g. $\mathbf{\hat{u}}_\mathbf{k} = \mathbf{u_k}(y) \exp {i(\kappa_x x + \kappa_z z - \omega t)}$. Here, $x$, $y$ and $z$ are the streamwise, wall-normal and spanwise coordinates, respectively, and $t$ is time. A subscript $\mathbf{k}$ denotes an individual Fourier component. In Eq.~\ref{eqNSE}, $\nabla_\mathbf{k}$ and $\nabla_\mathbf{k}^T$ represent the Fourier-transformed gradient and divergence operators, and $\mathbf{\cal L}_\mathbf{k}$ is the linearized Navier-Stokes operator:
\vspace{-0.5cm}
\begin{align}\label{eqLinearOperator}
\mathbf{\cal L}_\mathbf{k}
&=
\left[\begin{array}{ccc}
-i\kappa_x U + \Ret^{-1} \nabla_\mathbf{k}^2 & -\partial U/\partial y & 0 \\
0 & -i\kappa_x U + \Ret^{-1} \nabla_\mathbf{k}^2 & 0 \\
0 & 0 & -i\kappa_x U + \Ret^{-1} \nabla_\mathbf{k}^2 \\
\end{array}\right],
\end{align}
{\noindent}where $U(y)$ is the mean velocity profile and $\Ret = u_\tau h / \nu$ is the friction Reynolds number. The variables $u_\tau$, $h$, and $\nu$ represent the friction velocity, channel half-height, and kinematic viscosity, respectively. $\nabla_\mathbf{k}^2 = [-\kappa_x^2 + \partial^2/\partial y^2 - \kappa_z^2]$ is the Fourier-transformed Laplacian.
A singular value decomposition (SVD) of the discretized resolvent operator $\mathbf{\cal H}_\mathbf{k} = \sum_{m}^{} \psi_m(y)\sigma_m \phi^*_m(y)$ yields a set of orthonormal forcing ($\phi_m$) and response ($\psi_m$) modes, ordered based on the input-output gain ($\sigma_1>\sigma_2>\sigma_m>...$). Forcing in the direction of the $m^{th}$ forcing mode with unit amplitude results in a response in the direction of the $m^{th}$ response mode amplified by factor $\sigma_m$. Thus, forcing $\mathbf{f_k}(y)=\phi_1(y)$ creates a response $[\mathbf{u_k}(y),p_\mathbf{k}(y)]^T = \sigma_1 \psi_1(y)$. Note that the resolvent operator is scaled prior to performing the SVD to enforce an $L^2$ norm for the velocity, $\mathbf{u_k}$, and forcing, $\mathbf{f_k}$ \cite{Luhar2015}.
In general, for $\mathbf{k}$ combinations energetic in natural turbulence, the resolvent operator tends to be low rank \cite{McKeon2010,Moarref2013}. A limited number of input directions are highly amplified, often with $\sigma_1 \gg \sigma_2$, and so the velocity and pressure fields can be reasonably approximated by the first response mode $[\mathbf{u_k}(y),p_\mathbf{k}(y)]^T \sim \psi_1(y)$. Recent studies show that this rank-1 approximation captures many of the key features of wall-bounded turbulent flows, including the emergence of coherent structures and their footprint in the wall pressure field \cite{Sharma2013,Luhar2014b}. Further, the rank-1 modes also form useful building blocks for low-order models of flow control \cite{Luhar2014a}. As a result, the rest of this paper only considers the first singular values and modes, dropping the subscript $1$ for convenience. Extending the analysis to consider further singular values and modes is straightforward.
As discussed below, the effect of the compliant wall is introduced in this framework via the boundary conditions for the velocity and pressure fields. This change in the boundary conditions modifies the mode structure and singular value relative to the rigid wall case. A reduction in $\sigma$ is interpreted as mode suppression, which is deemed beneficial for control purposes. Keep in mind that this approach essentially focuses on how the compliant wall modifies the linear amplification mechanisms in the flow. The effect of the compliant walls on nonlinear interactions between modes, and the forcing generated due to these nonlinear interactions, is neglected.
The discretized resolvent operator in Eq.~\ref{eqNSE} is constructed using a spectral collocation method on Chebyshev points. The differentiation matrices are computed using the MATLAB differentiation matrix suite developed by Weideman and Reddy \cite{Weideman2000}. The SVD of the resolvent operator generally yields pairs of structurally similar response modes with near-identical singular values but differing symmetry along the channel centerline \cite{Moarref2013}. To avoid any confusion arising from this mode pairing and to make the computation more efficient, the grid is restricted to $N$ points in the lower half-channel, with user-specified mode symmetry across the centerline, $y=1$ ($y$ is normalized by the channel half-height $h$). The specific grid resolution required for convergence tends to be Reynolds number and wave speed dependent. For the results presented in this paper, we employ $N = 100$ at $\Ret = 2000$ and $N = 200$ at $\Ret = 2\times 10^4$. In both cases, the singular values had converged to $O(10^{-4})$. For greater details on numerical implementation and convergence, the reader is referred to \cite{Luhar2015}. As a rough estimate of computational expense, construction of the resolvent operator and computing the SVD takes approximately 0.1s on a single core of a laptop for each wavenumber frequency combination at $N=100$ and 0.5s at $N = 200$.
Note that construction of the linear operator $\mathbf{\cal L}_\mathbf{k}$ in Eq.~\ref{eqLinearOperator}, and hence $\mathbf{\cal H}_\mathbf{k}$, requires knowledge of the mean velocity profile $U(y)$. The exact form of this mean profile is important since high amplification in the resolvent framework results from two mechanisms: (i) localization of the modes around the critical layer, $y_c$, where the mode speed matches the mean velocity $U(y_c)=c$, and (ii) energy transfer from the mean flow to the turbulence via the so-called lift-up mechanism, which depends on the interaction between mean shear and wall-normal velocity, $\propto v_\mathbf{k} (\partial U/\partial y)$ \cite{McKeon2010,McKeon2013,Sharma2013}. For the modeling and optimization efforts described in \S\ref{sec:theory-BC}-\ref{sec:theory-optimal}, the mean velocity profile is generated using a well-known turbulent eddy viscosity model for smooth-walled flows \cite{Reynolds1967}. However, when comparing model predictions with simulation results from F2008 in \S\ref{sec:theory-F2008}, the mean velocity profiles from DNS are used. Use of the DNS mean profiles enables \textit{a posteriori} analysis of model sensitivity to the assumed $U(y)$. In other words, we evaluate how the predicted mode amplification changes when using the mean velocity profiles from compliant wall DNS relative to the smooth wall DNS.
\subsection{Boundary Conditions}\label{sec:theory-BC}
The effect of the compliant wall is introduced by changing the boundary conditions on velocity and pressure within the resolvent (Eq.~\ref{eqNSE}) before computing the SVD. For wall displacement $\eta(x,z,t)$ constrained to be in the wall-normal ($\mathbf{e_y}$) direction, the kinematic boundary conditions at the wall, $\mathbf{u}(y=\eta) = (\partial \eta / \partial t) \mathbf{e_y}$, can be expressed as the following Fourier-transformed, linearized Taylor series expansions:
\vspace{-0.5cm}
\begin{align}\label{eqKinematicBCu}
u_\mathbf{k}(\eta) & \approx u_\mathbf{k}(0) + \eta_\mathbf{k} \frac{\partial U}{\partial y}\Big|_{0}
+& \sum\limits_{\mathbf{k} = \mathbf{k}_a - \mathbf{k}_b} \bcancel{\eta_{\mathbf{k}_a} \frac{\partial u^*_{\mathbf{k}_b}}{\partial y}\Big|_{0}} + ... &=& 0, \\ \label{eqKinematicBCv}
v_\mathbf{k}(\eta) & \approx v_\mathbf{k}(0)
+& \sum\limits_{\mathbf{k} = \mathbf{k}_a - \mathbf{k}_b} \bcancel{\eta_{\mathbf{k}_a} \frac{\partial v^*_{\mathbf{k}_b}}{\partial y}\Big|_{0}} + ... &=& -i\omega \eta_\mathbf{k}, \\ \label{eqKinematicBCw}
w_\mathbf{k}(\eta) & \approx w_\mathbf{k}(0)
+& \sum\limits_{\mathbf{k} = \mathbf{k}_a - \mathbf{k}_b} \bcancel{\eta_{\mathbf{k}_a} \frac{\partial w^*_{\mathbf{k}_b}}{\partial y}\Big|_{0}} + ... &=& 0,
\end{align}
{\noindent}where $\eta_\mathbf{k}$ represents the Fourier coefficient for the wall displacement at wavenumber-frequency combination $\mathbf{k} = (\kappa_x,\kappa_x,\omega)$. The neglected quadratic terms are shown for reference.
The use of these linearized kinematic boundary conditions is one of the key limitations of the present approach since the neglected higher-order terms in Eq.~\ref{eqKinematicBCu}-\ref{eqKinematicBCw} can become important for large wall deflection. However, retaining terms of quadratic or higher order in the fluctuations would require a coupled nonlinear model allowing for interactions between resolvent modes across all wavenumber-frequency combinations that can interact and force the Fourier mode of interest \cite{Luhar2015,Duvvuri2015triadic}, which is outside of the scope of the current effort. Similarly, note that the nonlinear terms arising from the Fourier mode being considered would appear in the boundary conditions for higher harmonics. For instance, the quadratic terms $(1/2)\eta_\mathbf{k}^2 (\partial^2 U/\partial y^2)_{y=0} + \eta_\mathbf{k} (\partial u_\mathbf{k}/\partial y)_{y=0}$ would appear in the kinematic boundary condition for the streamwise velocity of the mode with wavenumber-frequency combination $(2\kappa_x,2\kappa_z,2\omega)$. It can be shown that the magnitude of the first term in the above expression only becomes important relative to the retained linear term, $\eta_\mathbf{k} (\partial U/\partial y)_{y=0}$, when the wall deformation is $O(1)$. However, the magnitude of the second term can be significant for energetic modes with near-wall gradients, $(\partial u_\mathbf{k}/\partial y)_{y=0}$, comparable to the mean velocity gradient, $(\partial U/\partial y)_{y=0}$. Thus, while these terms do not directly influence the Fourier mode being considered, they can be important elsewhere in spectral space. Also keep in mind that the linearized boundary conditions require an estimate of the mean shear at the wall (Eq.~\ref{eqKinematicBCu}), which is assumed to correspond to the prescribed, smooth wall mean velocity profile. This assumption breaks down if the compliant wall significantly alters the near-wall mean flow.
The dynamic boundary condition at the wall is expressed as a mechanical admittance, $Y$, linking wall-normal velocity and pressure:
\vspace{-0.5cm}
\begin{equation}\label{eqDynamicBC}
v_\mathbf{k}(0) = Y p_\mathbf{k}(0).
\end{equation}
{\noindent}$Y$ dictates the relative phase and amplitude of the wall-normal velocity and the pressure at the wall. As such, it can be used to represent walls of known material properties. For example, the most commonly used model for compliant walls involves a tensioned plate supported on a bed of springs and dampers. For such walls, the admittance can be expressed as \cite{Xu2003}:
\vspace{-0.5cm}
\begin{equation}\label{eqAdmittance}
Y = \frac{i \omega}{-C_m \omega^2 - i \omega C_d + C_{ke}}
\end{equation}
{\noindent}where $C_m$ and $C_d$ are the dimensionless mass ratio and damping coefficient, and
\vspace{-0.5cm}
\begin{equation}\label{eqEffectiveSpring}
C_{ke} = C_k + C_t k^2 + C_s k^4
\end{equation}
{\noindent}is a wavenumber-dependent effective spring constant, with $k^2 = (\kappa_x^2 + \kappa_z^2)$. The parameters $C_k$, $C_t$ and $C_s$ represent the dimensionless spring constant, tension and flexural rigidity. All of the above parameters are normalized based on the channel half-height $h$, friction velocity $u_\tau$ and fluid density $\rho$.
\subsection{Optimal Walls}\label{sec:theory-optimal}
In addition to evaluating the effects of the wall parameters individually, the resolvent framework can also be used to solve the inverse problem: finding an optimal $Y$ that leads to the most favorable effect on the turbulent flow structures of interest. Luhar et al. \cite{Luhar2015} pursued this optimization for modes resembling the NW-cycle and VLSMs at friction Reynolds number $\Ret = u_\tau h/\nu = 2000$. The NW-cycle was represented by the wavenumber-frequency combination $\mathbf{k} = (\kappa_x,\kappa_z,c^+) = (12,120,10)$ and the VLSMs were represented by $\mathbf{k} = (1,10,16)$. These wavenumbers translate into structures of streamwise and spanwise wavelength $(\lambda^+_x,\lambda^+_z) \approx (1050,105)$ and $(\lambda^+_x,\lambda^+_z) \approx (12500,1250)$, respectively. Optimality was defined in two different ways: walls that lead to the greatest mode suppression (i.e.\, lowest $\sigma_{\mathbf{k}}$) or the largest reduction in the channel-integrated Reynolds stress contribution from the mode \cite[per][]{Fukagata2002}. Throughout this paper, a superscript $+$ denotes normalization with respect to $u_\tau$ and $\nu$.
For brevity, this paper focuses primarily on the optimal gain-reducing wall for modes resembling the VLSMs at $\Ret=2000$. A simple pattern search procedure shows that a wall with admittance $Y = -2.0385 - 0.4387 i$ leads to the greatest reduction in singular value for such modes, with the ratio of compliant to rigid-wall (null-case) singular values being $\sigma_{\mathbf{k}c} / \sigma_{\mathbf{k}0} = 0.52$. Note that this optimization is blind to the physical properties of the compliant walls. Designing a wall with the appropriate admittance would then become an engineering problem (or perhaps one for material scientists). For walls characterized by Eq.~\ref{eqAdmittance}-\ref{eqEffectiveSpring}, this optimal admittance can be realized through any combination of springs, tension and stiffness. To evaluate how these factors affect performance, particularly with respect to the excitation of spanwise-constant modes, we test the different walls listed in Table~\ref{tab:walls}, each of which has admittance $Y = -2.0385 - 0.4387 i$ for $\mathbf{k} = (1,10,16)$.
\begin{table}
\centering
\caption{Different walls optimized to suppress resolvent modes resembling VLSMs at $Re_\tau = 2000$. The damping coefficient is $C_d = 0.4688$ in all cases.}
\begin{tabular}{lcccc}
\hline
Case & $C_m$ & $C_k$ & $C_s$ & $C_t$ \\
\hline
base & 2 & 510.4 & 0 & 0 \\
low $C_m$ & 0.2 & 49.59 & 0 & 0 \\
high $C_m$ & 20 & 5118 & 0 & 0 \\
tension & 2 & 0 & 0 & 5.053 \\
stiffness & 2 & 0 & 0.0500 & 0 \\
anisotropy & 2 & 0 & 0 & $C_{tx} = 288$\\
& & & & $C_{tz} = 2.224$\\
\hline
\end{tabular}%
\label{tab:walls}%
\end{table}%
The base case is the wall evaluated by Luhar et al. \cite{Luhar2015}, which represents a simple spring-damper system such that $C_{ke}=C_k$ and $C_m = 2$. The high and low mass ratio cases ($C_m =20$ and $C_m=0.2$, respectively) are similar but require different spring constants to counteract the changes in $C_m$. The next two cases in Table~\ref{tab:walls} remove the spring support but introduce the effects of tension and stiffness, such that $C_{ke} = C_t (\kappa_x^2 + \kappa_z^2)$ and $C_{ke} = C_s (\kappa_x^4 + 2\kappa_x^2 \kappa_z^2 + \kappa_z^4)$, respectively. The last case introduces the effects of anisotropy through differing streamwise and spanwise tension, such that $C_{ke} = C_{tx}\kappa_x^2 + C_{tz}\kappa_z^2$. Despite the physical differences, all of the walls are resonant just below the mode frequency $\omega = 16$ for $\kappa_x = 1$ and $\kappa_z = 10$. Specifically, the resonant frequency is $\omega_r = \omega_n \sqrt{1-2\zeta^2} = 15.97$, where $\omega_n = \sqrt{C_{ke}/C_m}$ is the undamped natural frequency of the wall and $\zeta = C_d/(2\sqrt{C_{ke} C_m})$ is the damping factor.
One of the key advantages of the resolvent formulation is that it can be extended to higher Reynolds numbers with limited computational penalty. Therefore, to test for Reynolds number effects, we also consider compliant walls optimized for VLSM-type modes at $\Ret = 2\times 10^4$, which we assume are characterized by wavenumber-frequency combination $\mathbf{k} = (\kappa_x,\kappa_z,c^+) = (1,10,19)$ \cite{Marusic2010}. For these modes, the pattern search algorithm suggests that walls with admittance $Y = -0.418 + 0.099 i$ lead to the greatest reduction in singular value, with $\sigma_{\mathbf{k}c}/\sigma_{\mathbf{k}0} = 0.30$. For a spring-damper wall with mass ratio $C_m = 2$, this optimal admittance translates into stiffness $C_k (=C_{ke})= 732$ and damping $C_d = 2.26$. The mean velocity profile for this high Reynolds number case is again obtained using an eddy viscosity formulation \cite{Reynolds1967}, while the grid resolution is increased to $N=200$ points for convergence.
\subsection{DNS-based Wall Optimization by Fukagata et al. 2008}\label{sec:theory-F2008}
The anisotropic wall model employed by F2008 was introduced by Carpenter and Morris \cite{Carpenter1990anisotropic} to mimic earlier experiments performed by Grosskreutz \cite{Grosskreutz1971}. As shown in Fig.~\ref{fig:F2008}, this wall involves an elastic plate that rests on spring-supported links that rotate about an equilibrium angle of $\theta \le \pi/2$ relative to the horizontal plane. Similar to Eq.~\ref{eqKinematicBCu}-\ref{eqEffectiveSpring}, the boundary conditions for this wall can be expressed in terms of a single displacement variable $\eta_\mathbf{k}$, which is defined as the displacement of the tip of the rigid link in this case. Specifically, the kinematic boundary conditions are:
\vspace{-0.2cm}
\begin{equation}\label{eqKinematicBCF2008}
u_\mathbf{k}(0) = -i \omega \eta_\mathbf{k} \sin \theta; \:\: v_\mathbf{k}(0) = -i \omega \eta_\mathbf{k} \cos \theta; \:\: w_\mathbf{k}(0) = 0,
\end{equation}
{\noindent}while the dynamic boundary condition can be expressed as:
\vspace{-0.2cm}
\begin{equation}\label{eqDynamicBCF2008}
\left[ -\rho_m b \omega^2 - i \omega C_d + \left\{ \frac{E b^3}{12(1-\nu_p^2)} k^4 \cos^2 \theta + E b k^2 \sin^2 \theta + C_k \right\} \right] \eta_\mathbf{k}
= g_\mathbf{k},
\end{equation}
{\noindent}where $\rho_m$, $b$, $E$, $\nu_p$ are the dimensionless plate density, thickness, elastic modulus, and Poisson's ratio, respectively. The forcing function on the right-hand side, $g_\mathbf{k}$, is defined in Eq.~\ref{eqForcingF2008} below. As before, $k^2 = (\kappa_x^2 + \kappa_z^2)$, $C_d$ is the damping coefficient, and $C_k$ is the spring stiffness. The quantity inside the curly brackets in Eq.~\ref{eqDynamicBCF2008} can once again be considered a wavenumber-dependent effective stiffness:
\vspace{-0.2cm}
\begin{equation}\label{eqCkeF2008}
C_{ke} = \frac{E b^3}{12(1-\nu_p^2)} k^4 \cos^2 \theta + E b k^2 \sin^2 \theta + C_k,
\end{equation}
{\noindent}leading to an undamped natural frequency of $\omega_n = \sqrt{C_{ke}/(\rho_m b)}$.
\begin{figure}
\begin{center}
\includegraphics[width=6cm]{figures/Fig1-F2008.jpg}%
\caption{Anisotropic compliant wall model employed in F2008 (image reproduced from \cite{Fukagata2008}).}
\label{fig:F2008}
\end{center}
\end{figure}
Note that, unlike the compliant wall model discussed in \S\ref{sec:theory-BC}, here all variables are normalized using twice the bulk-averaged velocity, $2U_B$. As a result, the bulk-averaged Reynolds number $Re_B = 2U_Bh/\nu = 3300$, which was kept constant in the DNS, replaces the friction Reynolds number in the linear operator (Eq.~\ref{eqLinearOperator}). This bulk-averaged Reynolds number corresponds to $\Ret \approx 110$ ($U_B \approx 15 u_\tau$) for the smooth-wall case. The wall is forced by a combination of pressure and turbulent stresses, such that the forcing function on the right-hand side of Eq.~\ref{eqDynamicBCF2008} is
\vspace{-0.2cm}
\begin{equation}\label{eqForcingF2008}
g_\mathbf{k} = \left[ \left( -p_\mathbf{k} + \frac{2}{Re_B} \frac{\partial v_\mathbf{k}}{\partial y} \right) \cos \theta + \frac{1}{Re_B} \left( \frac{\partial u_\mathbf{k}}{\partial y}+ i \kappa_x v_\mathbf{k} \right) \sin \theta \right]_{y=0}.
\end{equation}
Another important feature of this anisotropic wall model is the kinematic constraint imposed by the rotating link, which ensures that $u_\mathbf{k}$ and $v_\mathbf{k}$ are in phase at the wall (see Fig.~\ref{fig:F2008} and Eq.~\ref{eqKinematicBCF2008} above; \cite{Fukagata2008}). As a result, the mean turbulent Reynolds stress, $-re(0.5 u_\mathbf{k}^* v_\mathbf{k})$, where $re()$ denotes the real component and $()^*$ denotes a complex conjugate, is always negative at the wall. Per the Fukagata-Iwamoto-Kasagi identity \cite{Fukagata2002}, this reduction in the Reynolds shear stress is expected to decrease momentum transfer towards the wall, leading to a reduction in skin friction.
The DNS-based evolutionary optimization of wall parameters pursued by F2008 in a small domain of length and width $3h$ suggested that compliant walls with the properties listed in Table~\ref{tab:F2008} were optimal\footnote{Per F2008, the search algorithm had not converged but the available computational time had been exhausted.}, leading to a reduction in drag of $8.3\%$ (case A1 in F2008). However, the same wall led to a near $200\%$ \textit{increase} in drag in a domain of length $6h$. This increase in drag was accompanied by the emergence of energetic two-dimensional wavelike structures at the wall that spanned the length of the domain (i.e.\, modes with $\kappa_z=0$ and $\kappa_x = 2\pi/6$), and substantial changes in the mean velocity profile (see Fig.~\ref{fig:Fukagatakz0}a). While domain-spanning wavelike motions were also observed in the small domain simulations, there was a near $100\%$ increase in the root-mean-square (rms) wall displacement in the large domain DNS.
In \S\ref{sec:results-F2008} below, it is shown that the resolvent framework is able to anticipate some of this deterioration in performance. Specifically, the framework predicts that two-dimensional modes with wave numbers smaller than $\kappa_x = 2\pi/3$ (wavelengths greater than $3h$) are susceptible to significant further amplification over compliant walls with the properties listed in Table~\ref{tab:F2008}.
\begin{table}
\centering
\caption{Properties of one optimal wall (case A1) identified in F2008 \cite{Fukagata2008}. The wall thickness and Poisson' ratio were fixed at $b = 0.01$ and $\nu_p = 0.5$, respectively. Values in the brackets below denote the initial ranges specified for the optimization.}
\begin{tabular}{ccccc}
\hline
$\rho_m$ & $E$ & $C_k$ & $C_d$ & $\theta$ [deg.] \\
$1.23$ & $3.00\times 10^{-3}$ & $4.25\times 10^{-5}$ & $2.21 \times 10^{-5}$ & $62.7$ \\
$[0.1,10]$ & $7.2 \times 10^{[-4,2]}$ & $1.1 \times 10^{[-5,-2]}$ & ${1.0 \times 10^{[-4,-2]}}$ & $[30,90]$ \\
\hline
\end{tabular}%
\label{tab:F2008}%
\end{table}%
\section{Results}
\subsection{Effect of Mass Ratio}\label{sec:results-CM}
\begin{figure}
\begin{center}
\subfigure[]{\resizebox*{4.5cm}{!}{\includegraphics{figures/Fig2a-Cm00.jpg}}}%
\subfigure[]{\resizebox*{4.5cm}{!}{\includegraphics{figures/Fig2b-Cm02.jpg}}}%
\subfigure[]{\resizebox*{4.5cm}{!}{\includegraphics{figures/Fig2c-Cm20.jpg}}}%
\caption{Shaded contours showing the ratio of compliant wall to null-case singular values, $\sigma_{\mathbf{k}c}/\sigma_{\mathbf{k}0}$, for the low $C_m$ (a), base case (b), and high $C_m$ (c) walls listed in Table~\ref{tab:walls}. Blue regions denote mode suppression while red regions indicate further amplification. The solid black lines indicate the resonant frequency. The dashed lines represent isocontours of the magnitude of the admittance $|Y|$ at level $0.01$. All results correspond to $\kappa_z=10$.}
\label{fig:Cm}
\end{center}
\end{figure}
One of the key differences between aerodynamic and hydrodynamic flows over compliant walls is the mass ratio, which is determined by the ratio of the solid density to the fluid density. While $C_m \sim O(1)$ is appropriate for hydrodynamic applications, it is expected that $C_m \sim O(10^3)$ for aerodynamic applications. A high mass ratio translates into a much smaller wall response to fluid pressure perturbations away from resonance, which in turn means that the wall does not significantly influence the flow structures. This is illustrated by the low $C_m$, base case, and high $C_m$ results shown in Fig.~\ref{fig:Cm}. The spectral region over which the compliant wall has a strong influence on the singular values (positive or negative) shrinks significantly as the mass ratio is increased from $C_m = 0.2$ (Fig.~\ref{fig:Cm}a) to $C_m = 20$ (Fig.~\ref{fig:Cm}c). As an example, for $\kappa_x = 10$ the $C_m=0.2$ wall affects modes with speeds up to $c^+ \approx 13$, while the $C_m = 20$ wall only affects modes with speeds up to $c^+ \approx 3$. Note that the region of influence in all cases is centered approximately around the resonant frequency (bold black line), where the magnitude of the admittance $|Y|$ peaks.
In general, the compliant walls seem to have a positive influence (suppression) on modes with frequencies higher than the resonant frequency (i.e.\, above the solid black line) and a negative effect on modes with lower frequencies. Although, this reverses for modes with $\kappa_x < 1$ and $c^+ > 15$ for the low $C_m$ wall (Fig.~\ref{fig:Cm}a). These transitions in performance may be attributed to two factors: (i) changes in the phase relationship between the pressure and velocity fields as the mode speed increases (i.e.\, as the modes move further away from the wall), and (ii) the phase shift in the wall response, $\angle Y$, across the resonant frequency.
Also shown in Fig.~\ref{fig:Cm} are isocontours of the magnitude of the wall admittance $|Y|$ at level $0.01$ (dashed lines). A comparison of Fig.~\ref{fig:Cm}a-c shows that the region enclosed by these isocontours reduces rapidly with increasing mass ratio. More quantitatively, the half-power bandwidth of a spring-damper system is expected to scale as $\zeta \omega_n \sim C_m^{-1}$ for $\zeta \ll 1$ \cite{CUEDMechanics}. So the ten-fold increase in the mass-ratio translates into a roughly ten-fold decrease in the frequency bandwidth of the wall. This bandwidth would decrease even further for $C_m \sim O(10^3)$, suggesting that compliant walls are unlikely to be practical for aerodynamic applications requiring broadband turbulence suppression without the development of novel lightweight materials. On the other hand, the narrow bandwidth at high $C_m$ could enable more effective targeting of specific wavenumber-frequency combinations (i.e.\, to suppress or enhance individual velocity response modes).
Note that the decrease in the spectral influence of the compliant wall is roughly consistent with the decrease in the wall bandwidth. However, there are regions where the wall influences the flow despite low $|Y|$ (e.g.\, for very slow modes with $c^+ < 1$) and where the wall does not have an appreciable effect even at resonance (e.g.\, for faster modes with $c^+ > 18$). This is because the influence of the wall is determined both by the admittance as well as the magnitude of the wall-pressure fluctuations. In general, the magnitude of the wall pressure fields associated with the modes decreases with increasing mode speed $c^+$ \cite[i.e.\, as the modes move further away from the wall, see][]{Luhar2014b}, and so slower modes are likely to interact with compliant walls to a larger extent.
Keep in mind that, even over compliant surfaces, there is unlikely to be significant turbulent activity very close to the wall. As such, paying close attention to response modes with $c^+ < 4$ (i.e.\, corresponding to the viscous sublayer under Taylor's hypothesis) has limited utility except in cases with very high amplification resulting from resonance. Other important exceptions to this rule are the slow moving two-dimensional structures observed in previous experiments and DNS \cite{GadelHak1984,GadelHak1986,Fukagata2008,Kim2014}, which are discussed in greater detail below.
\subsection{Comparing Springs, Tension and Stiffness}\label{sec:results-KTS}
\begin{figure}
\begin{center}
\subfigure[]{\resizebox*{4.5cm}{!}{\includegraphics{figures/Fig3a.jpg}}}%
\subfigure[]{\resizebox*{4.5cm}{!}{\includegraphics{figures/Fig3b.jpg}}}%
\subfigure[]{\resizebox*{4.5cm}{!}{\includegraphics{figures/Fig3c.jpg}}}\\%
\subfigure[]{\resizebox*{4.5cm}{!}{\includegraphics{figures/Fig3d.jpg}}}%
\subfigure[]{\resizebox*{4.5cm}{!}{\includegraphics{figures/Fig3e.jpg}}}%
\subfigure[]{\resizebox*{4.5cm}{!}{\includegraphics{figures/Fig3f.jpg}}
\caption{Shaded contours showing the singular value ratio $\sigma_{\mathbf{k}c}/\sigma_{\mathbf{k}0}$ as a function of streamwise wavenumber and mode speed. Blue regions denote mode suppression and red regions denote amplification. Plots (a,d), (b,e) and (d,f) represent the base case, tensioned wall and stiff wall listed in Table~\ref{tab:walls}, respectively. The dashed contours indicate the magnitude of the singular values $\sigma_{\mathbf{k}c}$ over the compliant walls. The solid lines show the resonant frequency.}
\label{fig:KTS}
\end{center}
\end{figure}
Next we compare the effects of a compliant wall on a spring support with a tensioned membrane and a stiff plate. For the simple spring-damper system, the fluid-structure interactions are dependent solely on frequency. Moreover, the wall does not communicate in the streamwise and spanwise directions, which means that it cannot support wave propagation. In contrast, tensioned membranes and stiff plates have a wavenumber-dependent effective spring constant (Eq. \ref{eqEffectiveSpring}) and can support wave propagation. This means that the three different walls have varying effects across spectral space, despite being optimized to suppress the VLSM-type modes.
The above effects are best understood in terms of the resonant frequency $\omega_r$. Like the results shown in the previous section, for $\kappa_z = 10$, modes with frequencies below the resonant frequency are further amplified by the compliant walls while modes with higher frequencies are generally suppressed (Fig.~\ref{fig:KTS}a-c). However, the resonant frequency (solid black lines) varies substantially across the three different cases. For the basic spring-damper wall, the resonant frequency is constant, and so the effect of the wall is centered around modes with $c^+ \kappa_x = \omega_r$, or $c^+ \sim \kappa_x^{-1}$ (Fig.~\ref{fig:KTS}a). For the tensioned membrane and stiff plate, the response is centered around a similarly decreasing function $c^+ = f(\kappa_x)$ for $\kappa_x \ll \kappa_z(=10)$ (Fig.~\ref{fig:KTS}b,c). This is because the effective spring constant is dominated by the spanwise wavenumber dependence for $\kappa_x \ll \kappa_z$, leading to essentially constant $C_{ke}\approx C_t \kappa_z^2 \approx 505$ and $C_{ke} \approx C_s \kappa_z^4 \approx 500$ for the results shown in Fig.~\ref{fig:KTS}b,c. However, as $\kappa_x \gg \kappa_z$, the effective spring constant for the walls is dominated by the streamwise dependencies $C_{ke}\approx C_t \kappa_x^2$ and $C_{ke}\approx C_s \kappa_x^4$, which translates into resonant frequencies $\omega_r \approx \kappa_x \sqrt{C_t/C_m}$ and $\omega_r \approx \kappa_x^2 \sqrt{C_s/C_m}$. This means that the maximum admittance is found at near-constant $c^+ \approx \sqrt{C_t/C_m} = 1.59$ for the tensioned membrane (Fig.~\ref{fig:KTS}b) and is an increasing function $c^+ = f(\kappa_x)$ for the stiff plate (Fig.~\ref{fig:KTS}c).
Figures~\ref{fig:KTS}d-f show that all three walls also lead to significant amplification of certain classes of two-dimensional ($\kappa_z = 0$) structures, which is consistent with previous experiments and DNS. Interestingly, all three cases exhibit a repeating amplification-suppression pattern across spectral space. As an example, for fixed phase speed $c^+ \approx 10$, long structures with streamwise wavelength $\kappa_x < 1$ are further amplified, while shorter modes with $\kappa_x > 2$ are suppressed over the compliant walls. The wavenumber at which this amplification-suppression transition occurs generally decreases with increasing $c^+$ and there is an additional suppression-amplification transition at higher speeds (see e.g.\, $c^+ \approx 16$ in Fig.~\ref{fig:KTS}d-f), although wall resonance also plays an important role (Fig.~\ref{fig:KTS}d-f, solid black lines). In general, there appear to be two classes of mode that are further amplified over compliant walls. Long, slow-moving modes with $\kappa_x < 5$ and $c^+ < 7$ are amplified regardless of the wall properties, at least for the walls tested. The second class of modes that is further amplified is linked to wall resonance and is generally of smaller wavelength (see e.g.\, $\kappa_x > 10$ in Fig.~\ref{fig:KTS}e,f).
Note once again that the resonant frequency, and hence wave speed, varies significantly across the three different cases. The wave speed corresponding to resonance is a decreasing function of $\kappa_x$ for the spring-damper wall (Fig.~\ref{fig:KTS}d), constant for the tensioned membrane (Fig.~\ref{fig:KTS}e, $c^+ \approx \sqrt{C_t/C_m}$, i.e.\, the free-wave speed of the wall), and an increasing function of $\kappa_x$ for the stiff plate (Fig.~\ref{fig:KTS}f).
\subsection{Anisotropy and Wall-based Instability}\label{sec:results-anisotropy}
\begin{figure}
\begin{center}
\subfigure[]{\resizebox*{5cm}{!}{\includegraphics{figures/Fig4a.jpg}}}%
\subfigure[]{\resizebox*{5cm}{!}{\includegraphics{figures/Fig4b.jpg}}}\\%
\subfigure[]{\resizebox*{10cm}{!}{\includegraphics{figures/Fig4c.jpg}}}\\%
\subfigure[]{\resizebox*{10cm}{!}{\includegraphics{figures/Fig4d.jpg}}}%
\caption{Shaded contours showing the singular value ratio $\sigma_{\mathbf{k}c}/\sigma_{\mathbf{k}0}$ for $\kappa_z = 0$ (a) and $\kappa_z = 10$ (b) over the anisotropic wall listed in Table~\ref{tab:walls}. Blue regions denote mode suppression and red regions denote amplification. Plots (c) and (d) show the structure of the highly amplified two-dimensional modes marked in (a), representing wave number frequency combinations $\mathbf{k} = (\kappa_x,\kappa_z,c^+) = (1,0,1)$ and $\mathbf{k} = (3.8,0,11)$, respectively. The shading on the compliant wall indicates the normalized pressure field. The vectors show the streamwise and wall-normal velocity fields. Wall deflection not to scale.}
\label{fig:Structure}
\end{center}
\end{figure}
In this section, we introduce the effects of anisotropy by testing the effects of a wall with different streamwise and spanwise tension coefficients $C_{tx} = 288$ and $C_{tz} = 2.224$, so that $C_{ke} = C_{tx}\kappa_x^2 + C_{tz}\kappa_z^2$ (Eq.~\ref{eqEffectiveSpring}). This anisotropy changes the resonant frequency of the wall (Fig.~\ref{fig:Structure}a,b) and the free wave speed is now $c^+ \approx \sqrt{C_{tx}/C_m} = 12$. However, the trends observed in the previous section remain. In particular, there is a sharp transition in performance across the resonant frequency for the $\kappa_z = 10$ modes, and spanwise-constant ($\kappa_z=0$) modes are susceptible to significant further amplification. There are two classes of highly-amplified spanwise-constant modes: long, slow-moving structures (e.g.\, $\kappa_x =1, c^+ = 1$, marked $c$ in Fig.~\ref{fig:Structure}a) and shorter, faster structures moving at close to the free wave speed (e.g.\, $\kappa_x \approx 4, c^+ = 11$, marked $d$ in Fig.~\ref{fig:Structure}a).
The above predictions are broadly consistent with the observations of Gad-el-Hak et al., \cite{GadelHak1984,GadelHak1986}, who showed that elastic and viscoelastic layers under turbulent boundary layers gave rise to two distinct classes of surfaces waves: the first, termed static divergence, were very long, slow-moving (nearly static) structures, while the second class of surface waves had shorter wavelengths and faster phase speeds, comparable to the free shear wave speed of the layer. The experiments suggest that the static-divergence waves appear preferentially for viscoelastic coatings while the faster waves appear preferentially for elastic layers. This effect of the viscosity (i.e.\, the damping in our model) remains to be explored.
Figures~\ref{fig:Structure}c,d show the structure associated with the two highly-amplified modes identified in Fig.~\ref{fig:Structure}a. Although the resolvent modes have vastly different wavelengths and speeds, the overall structure is similar. Specifically, the streamwise velocities associated with the modes are confined to a very small layer close to the wall, above which the velocities are primarily in the up-down wall-normal direction. Further, the magnitude of the wall-pressure field is largest over surface troughs and smallest over surface peaks, i.e.\, high pressures coincide with downward deflections and vice versa, as expected physically.
\subsection{Reynolds Number Effects}\label{sec:results-Reynolds}
\begin{figure}
\begin{center}
\subfigure[]{\resizebox*{5cm}{!}{\includegraphics{figures/Fig5a_Re2e3.jpg}}}%
\subfigure[]{\resizebox*{5cm}{!}{\includegraphics{figures/Fig5b_Re2e4.jpg}}}%
\caption{Shaded contours showing the ratio of compliant wall to null-case singular values, $\sigma_{\mathbf{k}c}/\sigma_{\mathbf{k}0}$, at friction Reynolds number $Re_\tau = 2\times 10^3$ (a) and $Re_\tau = 2\times 10^4$ (b) for spring-damper walls optimized to suppress VLSM-type structures. Blue regions denote mode suppression while red regions indicate further amplification. The solid black lines indicate the resonant frequency. All results correspond to $\kappa_z = 10$.}
\label{fig:Re}
\end{center}
\end{figure}
In this section, we consider Reynolds number effects. Specifically, we contrast the effect of walls optimized to suppress VLSM-type modes at $\Ret = 2000$ and $\Ret = 2 \times 10^4$, characterized by wavenumber-frequency combinations $\mathbf{k} = (\kappa_x,\kappa_z,c^+)=(1,10,16)$ and $\mathbf{k} = (1,10,19)$, respectively. The assumed increase in mode speed with Reynolds number is consistent with the $y^+ \sim \sqrt{\Ret}$ scaling for such large-scale modes proposed previously \cite{Marusic2010}.
Similar to the low mass ratio case (Fig.~\ref{fig:Cm}a), Fig.~\ref{fig:Re}b shows that the compliant wall optimized for $\Ret = 2\times 10^4$ has a larger region of influence in spectral space. Physically, this is because, at higher $\Ret$, the optimization targets a faster moving mode with $c^+ = 19$ that is centered further away from the wall. Since slower-moving modes with similar wavenumbers (i.e.\, similar length scales) are likely to be centered closer to the wall and therefore have higher wall pressure signatures, the compliant wall also interacts strongly with them.
Note that the amplification-suppression patterns observed previously (Fig.~\ref{fig:Cm}-Fig.~\ref{fig:KTS}) also persist at higher $\Ret$. For example, at $c^+ = 12$, longer modes with $\kappa_x <2$ are further amplified over the compliant wall while shorter modes with $\kappa_x >3$ are suppressed. For higher mode speeds $c^+ > 18$, this pattern reverses whereby longer modes with $\kappa_x < 1$ are suppressed and shorter modes are amplified. A comparison of Fig.~\ref{fig:Re}a and \ref{fig:Re}b indicates that the general patterns of mode suppression and amplification remain broadly similar, with Reynolds number and wall resonance serving to shift the transition points. Importantly, this observation suggests that it may be possible to generate scaling laws for compliant wall performance that are useful for all Reynolds numbers. Previous work shows that the structure and amplification of smooth-wall resolvent modes exhibit distinct Reynolds-number scaling regimes depending on whether the modes are centered in the near-wall, logarithmic, or outer region of the flow \cite{Moarref2013}. As such, it is perhaps not surprising that the effects of passive or active control also appear to scale predictably with Reynolds number.
\subsection{Optimal Wall from Fukagata et al. 2008}\label{sec:results-F2008}
\begin{figure}
\begin{center}
\subfigure[]{\resizebox*{5cm}{!}{\includegraphics{figures/Fig6a.jpg}}}%
\subfigure[]{\resizebox*{5cm}{!}{\includegraphics{figures/Fig6b.jpg}}}\\%
\subfigure[]{\resizebox*{5cm}{!}{\includegraphics{figures/Fig6c.jpg}}}%
\subfigure[]{\resizebox*{5cm}{!}{\includegraphics{figures/Fig6d.jpg}}}\\%
\subfigure[]{\resizebox*{6.5cm}{!}{\includegraphics{figures/Fig6e.jpg}}}%
\subfigure[]{\resizebox*{6.5cm}{!}{\includegraphics{figures/Fig6f.jpg}}}\\%
\subfigure[]{\resizebox*{6.5cm}{!}{\includegraphics{figures/Fig6g.jpg}}}%
\subfigure[]{\resizebox*{6.5cm}{!}{\includegraphics{figures/Fig6h.jpg}}}%
\caption{(a) The mean velocity profiles obtained in F2008 \cite{Fukagata2008} over a smooth wall, and compliant walls in small and large domains. The inset shows the mean profiles plotted in wall units, with the smooth wall friction velocity $u_{\tau 0}$ used for normalization. (b-d) Shaded contours showing the predicted singular value ratio $\sigma_{\mathbf{k}c}/\sigma_{\mathbf{k}0}$ for spanwise constants modes ($\kappa_z = 0$) over the optimal (A1) compliant wall, but with varying mean velocity profiles used in the resolvent operator. Blue regions denote mode suppression and red regions denote amplification. The dashed black lines indicate the actual singular values $\sigma_{\mathbf{k}c}$, the solid black line shows the natural frequency of the wall, and the vertical gray lines in (c-d) reflect the minimum cell size and maximum domain length. Plots (e-g) show the structure of the highly amplified two-dimensional modes marked in (c) and (d). The shading on the walls shows the normalized fluctuating Reynolds stress field. The vectors show the streamwise and wall-normal velocity fields. Wall deflection not to scale.}
\label{fig:Fukagatakz0}
\end{center}
\end{figure}
Finally, we present resolvent-based predictions for the optimal anisotropic compliant wall from F2008, focusing primarily on the emergence of the two-dimensional ($\kappa_z=0$) wavelike motions observed in the simulations. The wall properties listed in Table~\ref{tab:F2008} lead to an effective spring constant of:
\vspace{-0.5cm}
\begin{multline}
C_{ke} = C_k + E b k^2 \sin^2 \theta + \frac{Eb^3}{12(1-\nu_p^2)} k^4 \cos^2 \theta \\ = 4.25 \times 10^{-5} + 2.39 \times 10^{-5} k^2 + 7.08 \times 10^{-11} k^4.
\end{multline}
{\noindent}Since $b = 0.01$, the third term $\propto Eb^3$ is much smaller than the second term $\propto Eb$ for wavenumbers $k \le O(100)$. Further, the first term is smaller than the second term for $k \ge O(1)$. Thus, for $1 \ll k \ll 100$, the effective spring constant can be approximated as $C_{ke} \approx 2.39 \times 10^{-5} k^2$, leading to $\omega_n = \sqrt{C_{ke}/(\rho_m b)} \approx 0.044 k$. As illustrated by the solid black lines in Fig.~\ref{fig:Fukagatakz0}, this effective wall stiffness results in a near-constant resonant wave speed of $c/2U_B = \omega_n/\kappa_x \approx 0.044$ ($c^+ \approx 1.3$) for modes with $\kappa_z = 0$ and $\kappa_x \approx 3$ to $\kappa_x = 64$. Further, the \textit{soft} nature (i.e.\, low resonant wave speed) of the wall means that it has a marked influence on singular values across most of the spectral space shown in Fig.~\ref{fig:Fukagatakz0}b-d.
Note that panels b-d in Fig.~\ref{fig:Fukagatakz0} show resolvent-based predictions for mode amplification over the compliant wall made using the three different mean velocity profiles in Fig.~\ref{fig:Fukagatakz0}a. In other words, the compliant wall properties are identical for all three panels, but the mean profile used in the resolvent operator varies. Visually, the mean velocity profile over the smooth wall (dashed line) is not substantially different from the mean profile observed in the small domain compliant wall DNS (fine gray line) that led to an $8.3\%$ reduction in drag. In contrast, the mean velocity profile in the large domain compliant wall DNS (bold gray line), which led to a substantial increase in drag, is more rounded and exhibits a much sharper gradient close to the wall. Consistent with these observations, the predicted changes in singular values over the compliant wall are not significantly different when either the smooth wall (Fig.~\ref{fig:Fukagatakz0}b) or small domain $U(y)$ (Fig.~\ref{fig:Fukagatakz0}c) is used in the Eq.~\ref{eqLinearOperator}. This lack of sensitivity to the details of the mean profile suggests that the resolvent framework may be used to generate useful predictions regarding the effects of compliant walls\footnote{Or more broadly, all types of flow control that can be represented via linear boundary conditions, \cite{Luhar2014a}} assuming smooth wall mean velocity profiles. Of course, such predictions should be treated very much as first approximations that provide qualitative insight into the effects of a given compliant wall. As illustrated by Fig.~\ref{fig:Fukagatakz0}d, the changes in predicted mode amplification are more substantial when the large domain compliant wall $U(y)$ is used instead of the smooth wall or small domain mean profiles.
In many ways, the wall characteristics discussed above (e.g.\, constant, low resonant wave speed) are similar to the more traditional tension-only wall considered in Fig.~\ref{fig:KTS}b. However, there are some important differences in how the two walls affect spanwise constant modes. Relative to the tension-only wall considered in \S\ref{sec:results-KTS}, the regions of mode amplification and mode suppression appear to be more distributed in spectral space over the anisotropic compliant wall considered in F2008. For example, there are no sharp transitions in performance across the resonant frequency. In addition, while the tension-only wall led to substantial further amplification of large-scale slow-moving modes, such modes are suppressed over the F2008 wall (see $\kappa_x < 6$ and $c/2U_B < 0.3$ in Fig.~\ref{fig:Fukagatakz0}b,c). We suggest that this mode suppression may be due to the negative Reynolds stress constraint imposed by the kinematic boundary conditions, though a detailed description of the exact mechanism is outside of the scope of the present effort.
Importantly, Fig.~\ref{fig:Fukagatakz0}c shows why the F2008 compliant wall only led to drag reduction in the small domain of length $3h$. Specifically, the resolvent framework predicts that modes with streamwise wavenumbers smaller than $\kappa_x \approx 2\pi/3$ (vertical gray line) and $c/2U_B > 0.3$ are further amplified over the compliant wall. This wavenumber cutoff represents structures with streamwise wavelength greater than $3h$, which can only appear in the larger domain of length $6h$. Of course, there are some regions of increased amplification for $\kappa_x \ge 2\pi/3$ as well. For instance, amplification increases nearly 20-fold for modes with $\kappa_x = 10$ and $c/2U_B \approx 0.22$ (point e in Fig.~\ref{fig:Fukagatakz0}c) and by a more limited $25\%$ for modes with $\kappa_x = 2\pi/3$ and $c/2U_B \approx 0.35$ (point g in Fig.~\ref{fig:Fukagatakz0}c). However, as shown in Fig.~\ref{fig:Fukagatakz0}e, the highly-amplified shorter mode is characterized by strong negative (beneficial) Reynolds stress at the wall and only a limited region of positive (detrimental) Reynolds stress for $y<0.1$. In any case, any detrimental effects associated with these modes that are further amplified is likely offset by the beneficial impact associated with mode suppression across the rest of spectral space.
Figure~\ref{fig:Fukagatakz0}d indicates that structures with $\kappa_x < 2\pi/3$ are amplified even more when the large domain mean velocity profile is used in the resolvent operator. For instance, the domain-spanning large scale mode with $\kappa_x = 2\pi/6$ and $c/2U_B = 0.42$ (point h in Fig.~\ref{fig:Fukagatakz0}d) experiences a near $300\%$ increase in amplification over the compliant wall. Further, unlike the other highly amplified modes shown in Fig.~\ref{fig:Fukagatakz0}e-g, which contribute substantial negative Reynolds stresses at the wall, Fig.~\ref{fig:Fukagatakz0}h shows that this domain-spanning mode primarily contributes positive Reynolds stress in the region $0<y<0.3$. The blue shading at the wall, which denotes negative Reynold stresses, is much less pronounced and there are substantial regions of positive Reynolds stress (red shading) in the fluid domain.
Thus, the larger domain may create a feedback loop whereby two-dimensional wavelike structures with length scales greater than $3h$ arise and generate significant additional detrimental Reynolds stress. In addition to transferring energy from the mean flow to the turbulence, this Reynolds stress also modifies the mean velocity profile such that the wavelike modes are amplified even further. This reinforcement could explain the substantial increase in drag observed in the large domain DNS. However, bear in mind that additional effects such as separation and secondary circulation over the large-amplitude wavelike motion of the compliant wall may also play a role in enhancing drag. Unfortunately, such effects cannot be captured in the resolvent framework without explicit treatment of nonlinear effects in the forcing $\mathbf{f_k}$ (Eq.~\ref{eqNSE}) or boundary conditions. Specifically, \textit{a priori} prediction of changes to the mean profile and prediction of secondary circulations requires an explicit treatment of the nonlinear forcing terms, while accounting for the effects of large wall deflection and separation requires nonlinear boundary conditions.
Despite these limitations, the results presented in this section indicate that, at the very least, the resolvent formulation may be used as a first-order test of material properties prior to testing in more computationally intensive simulations. While increasing the domain size in DNS carries a heavy computational penalty, extending the resolvent analysis to lower wavenumbers is inexpensive, especially when coupled with additional physical insight (e.g.\, focus on 2D modes) to limit the region of spectral space to be explored. As an example, all the results shown in Fig.~\ref{fig:Fukagatakz0} were computed in less than one hour on a single core of a laptop, without any attempt at making the computation efficient. Further, parameter sweep calculations similar to those shown in Fig.~\ref{fig:Fukagatakz0} are easy to parallelize since each wavenumber-frequency combination is independent.
\section{Conclusion}
The results presented in this paper provide some important design and methodology guidelines for future research on the development of compliant surfaces. The transitions in performance across the resonant frequency of traditional compliant walls (involving springs, tension, or stiffness) suggest that such walls must be slightly detuned and resonant at frequencies away from the spectral region of interest. Note that this transition in performance is linked to the phase shift in the relationship between the pressure and wall deflection across the resonant frequency, i.e.\, as $\angle Y$ changes sign, though the phase relationship between the velocity and pressure fields close to the wall also appears to play a role. Importantly, the amplification-suppression transitions persist at higher $\Ret$ as well, suggesting that it may be possible to generate useful scaling guidelines for compliant walls across all Reynolds numbers.
Although the present study and previous research by Luhar et al. \cite{Luhar2015} employ single wavenumber-frequency combinations as models for VLSMs and the NW-cycle, in reality these structures occupy a region in spectral space. As such, designing a compliant wall with a sharp transition in performance within this region is unlikely to be effective. Therefore, instead of optimizing wall performance for a single wavenumber-frequency combination, the optimization must be performed for a range of relevant wavenumbers and frequencies. In other words, there must be a net decrease in amplification across the entire spectral region of interest. This will, of course, increase the computational expense associated with the optimization procedure. However, since the effect of the compliant walls on individual modes (or wavenumber-frequency combinations) can be computed independently, there is significant scope for parallelization. Keep in mind that the framework in its current form does neglect nonlinear interactions across modes, which serve to generate the forcing terms $\mathbf{f_k} = (\mathbf{u} \cdot \nabla \mathbf{u})_\mathbf{k}$. While we cannot provide any definitive insight into the nature of these interactions at this point, this is an area of active research for the authors.
The results presented in \S\ref{sec:results-KTS} show that, in general, compliant walls must minimize the susceptibility to spanwise-constant structures to be effective. This is likely to be difficult given that most natural materials tend to act as low-pass filters. That is, the effective spring constant generally decreases with decreasing $\kappa_z$. One potential solution is to employ walls that are in spanwise compression, $C_{tz} < 0$, which would lead to a larger effective spring constant for $\kappa_z = 0$. Other possibilities include periodic spanwise breaks in the compliant material to disperse the spanwise-constant structures, or the use of mechanical metamaterials which ensure that the curvatures in the streamwise and spanwise directions are coupled, i.e.\, such that surface waves with $\kappa_z = 0$ and $\kappa_x \neq 0$ cannot be generated.
The results presented in \S\ref{sec:results-F2008} suggest that clever anisotropic compliant walls, similar to those proposed by Fukagata et al. \cite{Fukagata2008}, may be used to suppress substantial regions of spectral space. However, the two-dimensional structures discussed above still play a vital role in dictating overall performance, with larger-scale structures being particularly susceptible to further amplification. Thus, any simulation-based design of compliant walls requires the use of large computational domains, which is likely to impose severe restrictions on the Reynolds number or the extent of the parameter space that can be explored.
Resolvent analysis provides a computationally-inexpensive alternative that can be used to test and optimize wall properties prior to more detailed evaluation in DNS. One of the key limitations of the resolvent formulation in its present form is the requirement of a mean velocity profile $U(y)$, which is unknown \textit{a priori} over compliant walls. However, the results presented in this paper indicate that small changes in the mean profile (e.g.\, associated with a $<10\%$ change in drag) do not substantially alter predictions. As a result, smooth-wall mean profiles may still be used to generate useful initial predictions.
\section{Acknowledgments}
The authors gratefully acknowledge financial support from AFOSR grant FA9550-12-1-0469 (Program Manager: Doug Smith) and AFOSR/EOARD grant FA9550-14-1-0042 (Program Manager: Russ Cummings). The authors also thank Professor Koji Fukagata for generously sharing previous DNS results.
\bibliographystyle{tJOT}
|
\section{Introduction}
\label{sec:intro}
Explaining the origin and composition of the matter content of the Universe remains one of the most compelling tasks at the interface of high energy physics, nuclear physics, and cosmology. The identity of the dark matter that comprises 27\% of the cosmic energy density remains undetermined, and little is known about its non-gravitational interactions. The visible matter comprises just under 5\% of the present cosmic energy density and is often characterized by the baryon-to-photon ratio
\begin{eqnarray}
Y_B=\frac{n_B}{s} = (8.59 \pm 0.11) \times 10^{-11}
\end{eqnarray}
where $n_B$ ($s$) is the baryon number (entropy) density and where the value has been obtained from the Planck data\cite{Ade:2013zuv}. This number, though tiny, is clearly decisive for the Universe as we know it, yet the Standard Model (SM) suggests it should be many orders of magnitude smaller. Thus, accounting for the abundance of both the visible and dark matter
provides some of the strongest motivation for physics beyond the Standard Model (BSM).
It is possible that the dynamics associated with the dark matter and the origin of the baryon asymmetry are largely hidden from our view, either because the associated mass scale is too high or the relevant interactions with SM particles too feeble. In light of the discovery of the Higgs-like boson at the LHC, it is interesting to ask whether the properties and interactions of the Higgs boson provide a window, or \lq\lq portal", on the origin and composition of the cosmic matter content. If so, what might one learn from more refined studies of Higgs boson properties and interactions at the LHC or from the search for additional Higgs-like states? The purpose of this document is to summarize the landscape of possibilities as reviewed at the workshop \lq\lq Unlocking the Higgs Portal" held at the Amherst Center for Fundamental Interactions at the University of Massachusetts in May 2014\cite{ACFI14}. In view of Run II of the LHC, it is particularly worthwhile to identify the Higgs boson properties and searches for new states that are most promising from the standpoint of the cosmic matter content problem. In what follows, we provide a snapshot of this landscape as well as a discussion of additional theoretical work needed to delineate the prospective consequences of future LHC studies. Given the prospects for developing the next generation of high energy colliders, we also discuss the prospective opportunities for probing cosmologically relevant Higgs portal scenarios at various new facilities under consideration, including the International Linear Collider (ILC), China Electron Positron Collider (CEPC) and Super Proton Proton Collider (SppC), and the CERN future circular colliders FCC-ee (electron-positron) and FCC-hh (proton-proton).
The collider phenomenology pertaining to dark matter has been extensively investigated, particularly in relation to the WIMP paradigm (see, {\em e.g.}, Refs.~\cite{Morrissey:2009tf,Abdallah:2015ter,Abercrombie:2015wmb}). Dark matter-Higgs boson interactions have also been widely studied, and the high energy community is reasonably well versed in the dark matter problem (for recent discussions, see, {\em e.g.}. Refs.~\cite{Abdallah:2014hon,Craig:2014lda}). The implications of present and future LHC studies for the origin of visible matter, on the other hand, is less widely appreciated. Consequently, in what follows we will place a somewhat heavier weight on the baryogenesis problem, linking it to the Higgs portal and dark matter relic abundance where appropriate. For completeness, we also provide a short summary of the collider phenomenology of Higgs portal dark matter.
In principle, the non-vanishing $Y_B$ could have resulted from initial conditions during the Big Bang or from dynamics of grand unified theories at energy scales above $\sim 10^{16}$ GeV. In practice, the on-going success of the inflationary paradigm implies that any matter-antimatter asymmetry created in either of these ways would have been inflated away and, thus, not able to account for present observation. In this view, the particle physics of the post-inflationary Universe (including the era of preheating) is likely responsible. Over 40 years ago, Sakharov\cite{Sakharov:1967dj} identified three ingredients in the early Universe that must have been present in order to generate a non-vanishing $Y_B$: (a) baryon number (B)-violation; (b) violation of both C- and CP-invariance; and (c) either a departure from equilibrium dynamics or violation of CPT-invariance. The SM contains the first ingredient in the form of (B+L)-violating sphaleron transitions, but fails on the second and third.
The possible BSM scenarios that satisfy all three \lq\lq Sakharov" conditions span the gamut of post-inflationary cosmic history. Among the most theoretically attractive and experimentally testable are those that introduce new particles in the few-hundred GeV to TeV mass range. These scenarios would have generated the matter-antimatter asymmetry during or shortly before the era of electroweak symmetry-breaking (EWSB). The most thoroughly studied (though not exclusive) such scenario is electroweak baryogenesis (EWBG) (for a recent review, see Ref.~\cite{Morrissey:2012db}).
In EWBG, the Universe undergoes a first order phase transition during which electroweak symmetry is broken. The electroweak phase transition (EWPT) proceeds via nucleation of bubbles of broken electroweak symmetry as the Universe cools through a nucleation temperature $T_N$ that lies below the phase transition critical temperature, $T_C$. This transition, which satisfies the Sakharov out-of-equilibrium condition,
is analogous to the condensation of water droplets from vapor with decreasing temperature. Sakharov's second ingredient is provided by C- and CP-violating interactions of new particles at the bubble walls. These interactions ultimately induce the sphalerons to create baryons that diffuse inside the expanding bubbles where they are captured and protected from being washed out by inverse sphaleron processes.
The LHC and prospective future colliders are well-suited to looking for the particle physics ingredients needed for the first order EWPT. Indeed, the possibilities for generating this transition are rich. New particles may modify the Higgs potential through either loop effects or new tree level interactions. In some scenarios, the result may
be new patterns of EWSB that would not arise in the Standard Model, such as the occurrence of a series of transitions that break SM symmetries. In all cases, the dynamics require the existence of new spin-zero particles whose interactions may be more or less analogous to those of the SM Higgs boson. If one (or more) of these new states is stable on cosmological time scales, it (they) may also account for the dark matter relic density. Alternatively, dark matter fields may interact with the new scalars that, in turn, interact with the SM through the Higgs portal.
Generically, we write the Higgs portal interaction as
\begin{eqnarray}
\label{eq:portal1}
\mathcal{L} \supset \frac{a_1}{2} H^\dag \phi H +\frac{a_2}{2} H^\dag H \phi^\dag\phi+\cdots\ \ \ ,
\end{eqnarray}
where $H$ is the SM Higgs doublet, $\phi$ is an additional scalar transforming as either a singlet or non-singlet under the SM, and the \lq\lq $+\cdots$" indicate possible higher dimensional operators. For the neutral component of $\phi$ to contribute to the dark matter relic density, the $Z_2$-odd terms in Eq.~(\ref{eq:portal1}) must be absent ({\em e.g.}, $a_1\to 0$). On the other hand, $\phi$ may not itself be the dark matter candidate but may interact independently with the dark matter as in
\begin{eqnarray}
\label{eq:portal2}
\mathcal{L} \supset {\bar\chi}\left(a+b\gamma_5\right)\phi\chi +\mathrm{h.c.}+\cdots\ \ \ ,
\end{eqnarray}
where $\chi$ in this case is fermionic dark matter. The interactions (\ref{eq:portal1},\ref{eq:portal2}) may provide the Higgs portal into the dark matter sector, while the operators in Eq.~(\ref{eq:portal1}) may give rise for a strong first order EWPT as needed for EWBG.
The LHC and future colliders might discover the new particles ($\phi$, $\chi$) and probe the effects of their interactions in a number of ways:
\begin{itemize}
\item modified SM-like Higgs boson couplings to itself and other SM particles
\item new Higgs boson production mechanisms
\item new Higgs boson decay channels
\item new scalar particles that interact with the Higgs boson and other SM particles
\end{itemize}
Many of these signatures have been discussed elsewhere in the literature. Here we discuss their relation to the possible occurrence of a first order EWPT and connection to dark matter, utilizing representative theoretical scenarios.
\section{Theoretical Scenarios}
\label{sec:theory}
The dynamics of the EWPT are governed by the finite-temperature effective action, $S_\mathrm{EFF}(T)$, that reduces to an integral over the effective potential $V_\mathrm{EFF}(\phi, T)$ for spatial homogenous background fields $\phi$ (here, we generically denote the full set of background fields by $\phi$). For a theory such as the SM wherein only a single field acquires a vacuum expectation value (the background field), one may write the effective potential in the high-temperature limit
\begin{eqnarray}
V_\mathrm{EFF}(\varphi, T) = D(T^2-T_0^2)\varphi^2 - (E T+e) \varphi^3 + {\bar\lambda}\varphi^4 +\cdots\ \ \ ,
\label{eq:veff}
\end{eqnarray}
where $D$, $T_0^2$, $E$, $e$, and ${\bar\lambda}$ are all computable from the zero temperature Lagrangian. In the SM, where $\varphi$ is the Higgs background field (used here interchangeably with the vev), ${\bar\lambda}$ is approximately the Higgs self-coupling. The quantity $D T^2$ corresponds to the square scalar field thermal mass, while $DT_0^2$ is the zero temperature, tachyonic mass parameter, often denoted $\mu^2$. The existence of the cubic term proportional to $ET+e$ is essential for the occurrence of a first order phase transition. In the SM, $e=0$, as there exists no tree-level term that is cubic in the background field. In extended scalar sectors, $e$ may be non-vanishing. The quantity $E$ is generated at loop level in both the SM and its extensions.
The interplay between $(ET+e)$ and ${\bar\lambda}$ governs the character of the EWPT. In the limit that both $E$ and $e$ vanish, the transition becomes second order and no bubble nucleation will occur. The evolution of the potential with temperature for a first order EWPT is shown in Fig.~\ref{fig:veff}. The first order transition is marked by the existence of a barrier between the unbroken ($\varphi=0$) and broken ($\varphi\not=0$) minima that requires $E>0$ and/or $e>0$. In a second order transition the cubic terms in the finite temperature potential do not appear.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.5\linewidth]{VeffT.pdf}
\caption{\label{fig:veff}Evolution of the effective potential with $T$. }
\end{figure}
Assuming the occurrence of a first order transition with a sufficiently large bubble nucleation rate, preservation of any matter-antimatter asymmetry inside the bubbles requires quenching of the sphaleron transitions by making the sphaleron energy sufficiently large relative to the critical temperature, $T_C$ . For the high-$T$ potential of Eq.~(\ref{eq:veff}), the requirement becomes\cite{Quiros:1999jp,Profumo:2007wc,Patel:2011th}
\begin{eqnarray}
\label{eq:sfoewpt}
\frac{(ET_C + e)}{{\bar\lambda} T_C}\buildrel > \over {_\sim} 1\ \ \ .
\end{eqnarray}
This criterion is sometimes referred to as the requirement for a \lq\lq strong first order EWPT" or \lq\lq SFOEWPT".
In the SM, for which $e=0$ and $E$ arises through loops, this criterion is not satisfied, largely because the Higgs self coupling $\lambda$ is too large, thereby suppressing the denominator in Eq.~(\ref{eq:sfoewpt}). Since $m_h^2=\lambda v^2$, where $v$ is the $T=0$ vev, the requirement that the denominator of Eq.~(\ref{eq:sfoewpt}) be sufficiently small is equivalent to an upper bound on the Higgs boson mass. In fact, Monte Carlo studies indicate that the maximum Higgs boson mass for a first order EWPT in the SM is in the vicinity of 80 GeV\cite{Gurtler:1997hr,Laine:1998jb,Csikor:1998eu,Aoki:1999fi}, well below the observed value.
BSM scenarios may remedy the absence of a SFOEWPT by increasing the magnitude of E, reducing the magnitude of ${\bar\lambda}$, or reducing the critical temperature $T_C$. Broadly speaking, BSM scenarios do so {\em via} one or more of the following avenues:
\begin{itemize}
\item new loop effects that effectively increase $E$ (Section \ref{sec:loop})
\item new tree-level interactions that generate non-vanishing $e$ (Section \ref{sec:tree})
\item tree-level interactions that reduce ${\bar\lambda}$ either directly or indirectly (Section \ref{sec:tree})
\item tree-level or loop effects that lower the critical temperature, $T_C$ (Section \ref{sec:tree}).
\item new tree-level interactions that generate an earlier SFOEWPT along a different field direction, enabling EWBG through a multi-step process (Section \ref{sec:tree})
\end{itemize}
\subsection{Loop effects}
\label{sec:loop}
The dynamics of a loop-induced SFOEWPT can be understood by considering the so-called \lq\lq daisy resummation" contribution to $V_\mathrm{EFF}(\varphi, T)$. This contribution yields the following contribution to the difference in energy between the potential in the broken and unbroken phases:
\begin{eqnarray}
\label{eq:deltav}
\Delta V_\mathrm{daisy} = -\frac{T}{12\pi}\sum_k\left\{ \left[m_k^2+ y_k \varphi^2 + \Pi_k(T)\right]^{3/2} - \left[m_k^2+ \Pi_k(T)\right]^{3/2}\right\}\ \ \ ,
\end{eqnarray}
where the sum is over all bosonic degrees of freedom with mass parameters $m_k$, couplings to the Higgs field $y_k$, and thermal masses $\Pi_k(T)$. The presence of $\Delta V_\mathrm{daisy}$ effectively generates a barrier between the two phases. For the transverse components of the $W$ and $Z$ bosons, both $m_k^2$ and $\Pi_k(T)$ vanish, leaving a pure $T\varphi^3$ term in the potential. The barrier can be increased by adding more scalar degrees of freedom and by choosing their mass parameters to roughly cancel the thermal contributions: $m_k^2\sim -\Pi_k(T)$. Since the physical masses are given by $m^2\sim m_k^2+y_k v^2$, a SFOEWPT can lead to relatively light degrees of freedom at $T=0$ under this scenario. The most widely studied example occurs in the MSSM (see, {\em e.g.} Ref.~\cite{Carena:1997ki}), where the right-handed stop mass parameter is chosen to cancel its thermal mass, leading to a stop that is lighter than the top quark. The effect is particularly important for stops, since they have large Yukawa couplings and introduce a factor of $N_C$ when the sum in Eq.~(\ref{eq:deltav}) is performed.
Searches for light stops at the LHC\cite{Aad:2014kra,Chatrchyan:2013xna,CMS-PAS-SUS-13-009,Khachatryan:2015wza,Aad:2014nra} as well as determinations of Higgs boson signal strengths\cite{Curtin:2012aa,Katz:2015uja} appear to have closed most of the window for a SFOEWPT in the MSSM as well as extensions with hard SUSY-breaking terms. Going beyond SUSY, it was realized that loop contributions involving multiple species may also lower $T_C$ in the presence of a barrier between the two phases\cite{Huang:2012wn}. When one of the $y_k<0$, the corresponding contribution to $\Delta V_\mathrm{daisy}$ will be positive, effectively increasing the energy of the broken phase relative to the unbroken phase. As a result, the Universe must cool to a lower temperature than it otherwise would for the two phases to be degenerate in energy, thereby leading to a lower $T_C$ and increasing the left hand side of Eq.~(\ref{eq:deltav})\footnote{Application of this idea to the stop and sbottom contributions to the effective potential and the implications for the EWPT are under investigation.}.
Possible generation of a SFOEWPT through loop contributions has recently been studied in a general way by the authors of Ref.~\cite{Katz:2014bha}. The introduction of colored scalar particles that lead to a SFOEWPT would also lead to significant increases in the $hgg$ and $h\gamma\gamma$ couplings. Color singlets that are charged under the electroweak gauge groups and that give rise to a SFOEWPT would not modify the $hgg$ coupling but could induce observable deviations in the rate for $h\to\gamma\gamma$. For gauge singlets that generate a SFOEWPT solely via loops rather than tree-level interactions (see below), one would expect a change in $\sigma(e^+e^-\to Zh)$. The projected sensitivity of the CEPC and FCC-ee could allow for an observation of these modifications of Higgs boson production and decays\footnote{We also note that loop effects arising from new degrees of freedom may modify the $T=0$ potential in such a way that the SM gauge boson finite-temperature loops induce a SFOEWPT.}.
\subsection{Tree-level interactions}
\label{sec:tree}
The introduction of additional Higgs boson-scalar interactions leads to a number of possibilities for a SFOEWPT, including new patterns of EWSB where the occurrence of multi-step transitions to the present electroweak phase may entail such transitions. At the level of renormalizable operators, a broad range of possibilities are embodied in the Higgs portal interactions of Eq.~(\ref{eq:portal1}).
The simplest scenario arises when $\phi$ is a real singlet, denoted here as $S$. At $T=0$, the presence of the two operators in Eq.~(\ref{eq:portal1}) implies the existence of two mass eigenstates $h_{1,2}$ that are doublet-singlet mixtures\footnote{We will take $h_1$ to be the SM-like Higgs scalar.}. For the dynamics of the transition, taking $a_1<0$, the cubic interaction introduces a tree-level barrier, generating a non-vanishing $e>0$. The presence of a non-vanishing $a_2$ may lead to a reduction in the value of ${\bar\lambda}$ and a reduction in $T_C$.
For $a_2<0$, one finds a direct reduction in ${\bar\lambda}$\cite{Profumo:2007wc}. For positive $a_2$, the effect is indirect, involving the interplay of parameters in the scalar mass-squarked matrix\cite{Profumo:2014opa}. These features have been analyzed in a general fashion in Refs.~\cite{Profumo:2007wc,Espinosa:2011ax,Profumo:2014opa,Damgaard:2013kva,Curtin:2014jma}, while specific model realizations in the NMSSM have been studied in Refs.~\cite{Menon:2004wv,Kozaczuk:2014kva}.
For $\phi$ transforming non-trivially under SU(2$)_L$, the constraints on the electroweak $\rho$-parameter imply that the vev of the neutral component of $\phi$ must be small at $T=0$. As the latter is proportional to $a_1$, the corresponding tree-level barrier induced by non-vanishing $a_1$ between the symmetric and present electroweak vacua is too small to allow for a SFOEWPT at finite-$T$. On the other hand, for suitable choices of parameters, it is possible that electroweak symmetry breaks in two steps: (1) at temperature $T_1$, the neutral component of $\phi$ gets a vev while the doublet vev vanishes; (2) at $T_2< T_1$, a second transition occurs to the vacuum with vanishingly small neutral $\phi$ vev and non-vanishing doublet vev. The first transition may be strongly first order, and in this phase sphaleron transitions are suppressed since $\phi$ transforms non-trivially under SU(2$)_L$. The matter-antimatter asymmetry may be produced during this first step and preserved during the transition at $T_2$. A concrete realization of this possibility for $\phi$ being a real SU(2$)_L$ triplet has been analyzed in Ref.~\cite{Patel:2012pi,Inoue:2015pza}, while general considerations have been discussed in Ref.~\cite{Blinov:2015sna}. The vacuum structure of the potential and the two-step trajectory is illustrated in Fig.~\ref{fig:sigmasm}.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.5\linewidth]{TwoStep.pdf}
\caption{\label{fig:sigmasm} Two-step electroweak symmetry breaking for the real triplet Higgs portal scenario~\cite{Patel:2012pi}. Here, $h$ and $\sigma$ denote the neutral Higgs boson and real triplet background fields, respectively; $O$ gives the location of the symmetric phase, while $\Sigma$ and $H$ denote the locations of the triplet and Higgs phases, respectively. $T$ indicates another extremal point, while the red points are related to the black points by $Z_2$ symmetries. The two step transition proceeds first along the triplet direction at a temperature $T_\sigma$, followed by a transition to the Higgs phase at a temperature $T_h< T_\sigma$. }
\end{figure}
For the special case of the two Higgs doublet model (2HDM) where the $\rho$-parameter constraints do not apply to the vevs, initial studies by the authors of Refs. \cite{Dorsch:2013wja,Dorsch:2014qja} indicate that it is possible to achieve a SFOEWPT in a singlet step transition due to the presence of additional tree-level interactions involving the two doublets. The SFOEWPT-favored region of parameter space suggests that the neutral pseudoscalar $A^0$ should be moderately heavy and that it may be considerably heavier than the non-SM-like neutral scalar $H^0$.
Relaxing the requirement of renormalizability, the existence of higher-dimensional Higgs self-interactions may also enable a SFOEWPT during a singlet-step transition to the present electroweak vacuum\cite{Delaunay:2007wb,Grinstein:2008qi,Grojean:2004xa,Bodeker:2004ws,Henning:2014gca}. Writing
\begin{eqnarray}
\label{eq:higherdim}
V(H) = -\mu^2 H^\dag H + \lambda(H^\dag H)^2 + \eta (H^\dag H)^3+\cdots \ \ \ .
\end{eqnarray}
one observes that when the dimensional parameter $\eta>0$, the coefficient of the quartic self-coupling may be negative (subject to stability requirements), again introducing a tree-level barrier. This possibility may be particularly interesting when the Higgs is a pseudo-Goldstone boson, so that $\eta\sim 1/F_H^2$, with $F_H$ being of order 100 GeV \cite{Grinstein:2008qi}. Alternately, when $\phi$ is a gauge singlet that is sufficiently heavy to be integrated out, a potential of the form in Eq.~(\ref{eq:higherdim}) may arise with $\lambda<0$ and $\eta>0$. For suitable choices of the singlet mass and couplings $a_{1,2}$, a SFOEWPT can be achieved. For related work on the EWPT treating the Higgs boson as a pseudo Goldstone Boson, see Ref.~\cite{Grinstein:2008qi}.
\subsection{Higgs portal dark matter}
\label{sec:hpdm}
The literature on dark matter is vast, and this short overview cannot do justice to the breadth of theoretical work that has been carried out on this subject. From a practical standpoint, we instead focus on a few representative cases that illustrate the features of Higgs portal interactions and that illustrate the phenomenological implications. We consider both scalar dark matter that may arise as part of an extended scalar sector and fermionic dark matter.
\vskip 0.1in
\noindent{\bf Scalar dark matter: gauge singlets}. Extending the SM scalar sector with a real singlet $S$ has been discussed above in the context of the EWPT. To achieve a viable dark matter candidate, the corresponding scalar potential must admit a $Z_2$ symmetry ($S\to -S$) that ensures stability of the $S$, with vanishing $S$ vev at $T=0$. The universe may still undergo a SFOEWPT to the EWSB Higgs vacuum if the singlet potential contains a tachyonic mass-squared term. In this case, the universe may first land in a vacuum with $\langle H^0\rangle =0$ and $\langle S \rangle \not=0$ with decreasing $T$, followed by a transition to the $\langle H^0\rangle \not=0$ and $\langle S \rangle =0$ vacuum at lower temperature \cite{Curtin:2014jma}. For a non-tachyonic mass-squared term, a SFOEWPT is generally possible only when $\langle S \rangle \not=0$ at $T=0$. Avoidance of cosmic domain walls then implies that the potential must not be $Z_2$-symmetric, implying that $a_1\not=0$.
The general case (non-tachyonic mass) has been investigated by a number of authors\cite{Barger:2007im,Cline:2013gha,He:2013suk,Kahlhoefer:2015jma}. In the standard thermal DM paradigm, the relic density is governed by the Higgs portal coupling $a_2$ that sets the strength of the annihilation channels $SS\to hh$ and $SS\to h \to f{\bar f}$ and $VV$. For $m_S$ near $m_h/2$, the annihilation cross section $\sigma_\mathrm{ann}$ is dominated by single Higgs boson exchange (the \lq\lq Higgs pole"), which is resonantly enhanced. The magnitude of $a_2$ must be accordingly reduced in order for the singlet density to saturate the observed DM relic density: $\Omega_S=\Omega_{DM}$. However, if the singlet constitutes one component of a multicomponent dark matter scenario, then one may have a larger magnitude for $a_2$ with $\Omega_S/\Omega_{DM} <1$. Alternately, a non-thermal mechanism may lead to saturation of the observed relic density for larger $a_2$\cite{Feng:2014vea}. Note also that for the tachyonic mass scenario, stability of the Higgs/DM vacuum constrains the DM mass and relic density. The former is given by $-|\mu_s^2| + a_2 v^2$. The coupling $a_2$ must be positive and sufficiently large as to ensure a stable Higgs/DM vacuum. However, increasing $a_2$ leads to a smaller $\sigma_\mathrm{ann}$. For a detailed study, see, {\em e.g.} Ref.~\cite{Gonderinger:2009jp}.
DM direct detection experiments constrain the product $(\Omega_S/\Omega_{DM}) \times \langle \sigma_\mathrm{ann} v \rangle$, $v$ is the DM velocity. Constraints obtained including the recent LUX results are shown in Fig. 9 of Ref.~\cite{Curtin:2014jma} , assuming a thermal DM scenario. For non-thermal DM that yields $\Omega_S/\Omega_{DM} =1$ the constraints are considerably more severe, except for $m_S$ near $m_h/2$ . The corresponding collider signatures are generally quite challenging (for a recent discussion, see, {\em e.g.}, Ref.~\cite{Craig:2014lda}). Since $\langle S \rangle =0$, the singlet cannot be produced directly in $pp$ collisions as it does not mix with the SM Higgs boson and does not couple to any SM fields except pairwise to the Higgs boson. For $m_S \leq m_h/2$, the Higgs portal interaction $a_2$ leads to an invisible decay mode for the Higgs boson. The present LHC upper bound on the invisible branching fraction BR$(h\to\mathrm{invis})$ is 23\%~\cite{Aad:2015pla,Khachatryan:2014jba}. Note that this bound constrains the singlet fraction of the DM relic density from being arbitrarily small, since $|a_2|$ cannot be arbitrarily large. For $m_S > m_h/2$, $S$ is pair produced only through off-shell Higgs boson processes that have a relatively small cross section.
%
Variants on the simplest scenario of the SM plus a real singlet include extending the SM with a complex singlet\cite{Barger:2008jx,Gonderinger:2012rd,Jiang:2015cwa} and extending the 2HDM with a real singlet\cite{He:2013suk,Drozd:2014yla}. As discussed in Ref.~\cite{Barger:2008jx} for the complex singlet $\mathbb{S}$, it is convenient to assign to the singlet field a global U(1) charge. For a U(1) conserving potential, $\mathbb{S}$ is equivalent to two degenerate real singlets that behave much like the real singlet as a dark matter candidate. Introducing U(1)-breaking into the potential then splits the masses of the two components of $\mathbb{S}$, yielding a two-component DM scenario. The relative contributions of each to the relic density depend on the mass splitting and the magnitude of the singlet quartic self coupling. If $\langle \mathbb{S}\rangle\not=0$, then the Goldstone boson of the spontaneously broken U(1) will be stable. Explicitly breaking the U(1) makes this field a massive, viable DM candidate. The remaining degree of freedom associated with $\mathbb{S}$ then behaves like the real singlet that mixes with the SM Higgs boson. The resulting two mass eigenstates then appear as s-channel poles in the DM annihilation amplitude. The DM direct and indirect detection signatures, as well as the collider phenomenology, entails a combination of those associated with the real singlet when it is either a DM candidate or unstable. Under this scenario, one may also encounter a SFOEWPT.
As also noted in Ref.~\cite{Barger:2008jx}, the presence of the complex scalar then allows for complex couplings in the potential as well as a situation in which both components of the scalar receive vevs, a feature that would generally preclude the existence of a viable DM candidate. However, if that situation is associated with the first step of a two-step EWPT, and if it goes to zero during the second transition to the SM vacuum, then one (or more) of the components of $\mathbb{S}$ may remain stable at and below the freeze-out temperature for thermal DM\cite{Jiang:2015cwa}.
Another variant on the simplest scenario of the SM plus a real singlet has been studied recently in Refs.~\cite{He:2013suk,Drozd:2014yla} using a 2HDM plus a real singlet. Generically, the presence of a second neutral CP-even Higgs state opens up the parameter space of Higgs portal couplings that is consistent with the observed relic density. For low values of $m_S$, the annihilation rate to $b{\bar b}$ can be enhanced for large $\tan\beta$ in the Type II 2HDM, further opening the available parameter space, thereby allowing for a value of $a_2$ consistent with direct detection constraints and while achieving the observed relic density\cite{Drozd:2014yla}.
\vskip 0.1in
\noindent{\bf Non-singlet scalar dark matter: minimal dark matter}. When $\phi$ carries SM gauge charges and has integer isospin, the neutral component $\phi^0$ may be a viable DM candidate. A similar situation holds for one of the neutral states in a 2HDM. Stability of $\phi^0$ requires a $\mathbb{Z}_2$ symmetry, which must either be imposed by hand for lower dimensional representations of SU(2$)_L$ or is automatic for higher dimensional representations. The latter situation generally corresponds to \lq\lq minimal dark matter" \cite{Cirelli:2005uq,Cirelli:2007xd}. In order to evade DM direct detection limits, scalar dark matter must either have zero hypercharge or a highly suppressed coupling to the $Z^0$ boson; otherwise, the DM-nucleus neutral current scattering cross section will lie well above present direct search bounds.
A general study of non-singlet scalar DM has been carried out in Ref.~\cite{Hambye:2009pw}. These authors found that the both the $\mathbb{Z}_2$ symmetric 2HDM and both the real or complex scalar multiplets of dimension $n=3$, 5, or 7 may be viable. The lower bounds on the corresponding masses range from roughly 500 GeV to 20 TeV depending on $n$ and assuming saturation of the relic density. A detailed study of the collider phenomenology for the real $n=3$ case has been carried out in Ref.~\cite{FileviezPerez:2008bj}. In general one would search for one or more disappearing charged particle tracks when one or more charged states is produced through electroweak pair production. Results of a CMS search for disappearing charged particle tracks are reported in Ref.~\cite{CMS:2014gxa}. For neutralinos in the MSSM under the anomaly mediated SUSY-breaking scenario, a neutralino with mass less than 260 GeV is excluded. One might anticipate a similar bound for the scalar DM scenarios, though to our knowledge none has yet been reported in the literature. For the real triplet of Ref.~\cite{FileviezPerez:2008bj}, saturation of the full relic density requires a DM mass at or above roughly 2 TeV, well above the present CMS exclusion. Note that in the two-step EWPT scenario described earlier (see Sec.~\ref{sec:tree}), the CMS exclusion would likely preclude a stable neutral triplet in the minimal version of this scenario.
\vskip 0.1in
\noindent{\bf Fermionic dark matter}. In general, a Higgs portal interaction with one or more additional fermionic multiplets $\chi$ may provide for a viable fermionic DM scenario, though it will not substantially affect the EWPT. On the other hand such an interaction may provide for new sources of CPV as needed for EWBG during a SFOEWPT, a possibility that has been analyzed recently in Refs.~\cite{Chao:2014dpa,Chao:2015uoa} but that we will not explore in detail here. For the simplest, pure DM scenario, one introduces a non-renormalizable interaction of the form
\begin{equation}
\label{eq:fermionic}
\mathcal{L} \supset \frac{1}{\Lambda} H^\dag H\left( \cos \xi\, {\bar\chi} \chi + \sin\xi\, {\bar\chi} i\gamma_5 \chi\right)\
\end{equation}
where $\chi$ is a SM gauge singlet and $\Lambda$ defines an effective mass scale that incorporates the values of the scalar and pseudoscalar operator coefficients up to their relative normalization. A detailed study of this scenario has been carried out in Ref.~\cite{Fedderke:2014wda}. Saturation of the relic density requires $\Lambda$ to range from a few hundred GeV to a few TeV, while constraints on the Higgs boson invisible width impose a strong exclusion for $m_\chi<m_h/2$. Unitarity constraints on an extension of the simplest scenario that includes both the SM Higgs boson and a real scalar singlet have been studied in Ref.~\cite{Walker:2013hka}. Generation of the baryon asymmetry in a 2HDM variant of the interaction (\ref{eq:fermionic}) has been studied in Ref.~\cite{Chao:2015uoa}.
\vskip 0.1in
\noindent{\bf Portal to a hidden sector}. Dark matter particles may exist in a hidden sector containing additional degrees of freedom that may couple to the Higgs boson. If these degrees of freedom are sufficiently light, they may significantly affect dark matter dynamics, leading, for example, to an enhanced annihilation cross section. A particularly interesting possibility is that the hidden sector contains a U(1) symmetry that is spontaneously broken. The corresponding massive gauge boson $X$ may mix with the SM hyper charge boson, leading to the existence of an additional \lq\lq dark photon" or \lq\lq dark Z", $Z_D$. Alternately, the hidden sector may contain one or more SM gauge singlet scalars $S$ that couple to the $X$. As discussed in Section \ref{sec:tree}, Higgs portal interactions of the form (\ref{eq:portal1}) will lead to $h$-$S$ mixing. In this context, the Higgs-like state $h_1$ will inherit the couplings of $S$ to $X$, albeit suppressed by the mixing angle $\sin\theta$. For $Z_D$ mass below $m_h/2$, this scenario leads to the exotic Higgs decay $h_1\to Z_D Z_D$. The corresponding phenomenology for this possibility has been extensively analyzed in Ref.~\cite{Curtin:2014cca}. Rather than repeat that discussion here, we refer the reader to that work.
\vskip 0.1in
\noindent A summary of various scenarios is given in Table \ref{tab:scenarios}.
\begin{table}[hb]
\small
\begin{centering}
\begin{tabular}{|c|lc|c|c|c}
\hline\hline
Scenario & BSM & EWPT & DM & Collider Signatures \\
& DOF &&&\\
\hline\hline
Real Singlet & 1 & Tree-level & $X$ & $\ast$ Two neutral mixed states\\
&& single-step && $\ast$ Modified trilinear self-coupling \\
&&or two-step && $\ast$ Reduced SM-like signal strength\\
&&&& $\ast$ Resonant di-Higgs production\\
&&&& w/ final states: $b{\bar b}\tau^+\tau^-$, $b{\bar b}\gamma\gamma$ {\em etc.}\\
&&&& $\ast$ Exotic decays: $f{\bar f} \gamma \gamma$ {\em etc.}\\
\hline
Real Singlet & 1 & $X$ & $\surd$ & $\ast$ Invisible Higgs decays \\
\hline
Real Singlet & 1 & Two-step & $\surd$ & $\ast$ VBF $h^\ast\to$ invisible\\
&&Loop-induced&&$\ast$ Modified trilinear self-coupling \\
&& one-step&&\\
\hline\hline
Complex Singlet & 2 & Tree-level & $\surd$&$\ast$ Two neutral mixed states \\
&& single-step &&plus scalar DM\\
&& &&$\ast$ Modified trilinear self-coupling\\
&&&& $\ast$ Reduced SM-like signal strength\\
&&&& $\ast$ Resonant di-Higgs production\\
&&&& w/ final states: $b{\bar b}\tau^+\tau^-$, $b{\bar b}\gamma\gamma$ {\em etc.}\\
&&&& $\ast$ Exotic decays: $f{\bar f} \gamma \gamma$ {\em etc.}\\
&&&& $\ast$ Invisible decays\\
\hline\hline
Real Triplet & 3 & Loop-induced & $\surd$ & $\ast$ Reduced $\Gamma(h\to\gamma\gamma)$ \\
&& multi-step && $\ast$ Disappearing charged particle tracks \\
\hline
Real Triplet & 3 & Loop-induced & $X$ & $\ast$ Reduced $\Gamma(h\to\gamma\gamma)$ \\
&& multi-step && $\ast$ EW $H^\pm H^\mp$ and $H^\pm H^0$ production\\
&&&& w/ novel final states: $b{\bar b}\tau\nu$, $b{\bar b} W^\pm Z$, {\em etc.}\\
\hline\hline
2HDM & 4 & Loop-induced & $X$ & $A^0\to Z^0 H^0$ \\
&& single step && \\
\hline\hline
Higher-dim & 0 & Tree-level & $X$ & \\
operators&& single-step &&\\
\hline\hline
\end{tabular}
\caption{Higgs portal scenarios. For each scenario, the number of additional BSM degrees of freedom (DOF) is indicated, along with the nature of the EWPT and viability as a contributor to the dark matter (DM) relic density. The quantity $\surd$ ($X$) indicates viable (not viable). The final column summarizes the possible collider signatures for each scenario.\label{tab:scenarios}}
\end{centering}
\end{table}
\section{Signatures and Benchmarks}
\label{sec:signatures}
Because the nature of an EWPT depends critically on the interactions of new particles with the Higgs doublet, measuring the Higgs boson properties, including its self-interactions as well as interactions with other SM particles through production and decays can provide a probe of phase transition dynamics. In addition, it is possible to search for the new degrees of freedom through their direct production. Thus, the scenarios for baryogenesis-favorable phase transitions summarized in Section \ref{sec:theory} may lead to a number of signatures accessible at the LHC and prospective future colliders. For each of the simplest Higgs portal scenarios, the possibilities are indicated in the last column of Table \ref{tab:scenarios}. We also indicate which scenarios may allow for a viable dark matter candidate and indicate some of the corresponding signatures.
In what follows, we summarize work completed to date on some of these prospective signatures, followed by a brief roadmap for future analyses.
\subsection{Modified Higgs self-coupling}
In the real singlet Higgs portal scenario, the Higgs cubic self interaction $\lambda v h^3$ becomes $g_{111} h_1^3$, where $g_{111}$ is a combination of the potential parameters, $v$, and the singlet vev. The value of this coupling can provide a probe of a SFOEWPT\cite{Noble:2007kk,Katz:2014bha,Profumo:2014opa}. As shown in Fig.~\ref{fig:selftc}, the value of $g_{111}$ is strongly correlated with $T_C$, with smaller $g_{111}$ corresponding to lower critical temperature\cite{Profumo:2014opa}. Moreover, its value may be considerably less than its pure SM value $\sim 30$ GeV, thereby allowing for rather low transition temperatures. A measurement of $g_{111}$ can, thus, provide a probe of $T_C$ and the SFOEWPT-viable regions of singlet extensions. Speaking conservatively, one expects a $\sim 30-50\%$ determination of this parameter at the HL-LHC\cite{CMS-PAS-FTR-15-002,ATL-PHYS-PUB-2014-019,ATL-PHYS-PUB-2015-046}. Current projections for the ILC anticipate a 27\% and 16\% (10\%) determination using the full data set at $\sqrt{s}=500$ GeV and 2(5) $ab^{-1}$ at 1 TeV respectively, considering the $4b$ and $bbWW$ final states\cite{Fujii:2015jha}. Expectations next generation $pp$ colliders vary. Refs.~\cite{Barr:2014sga,Azatov:2015oxa} project, respectively a 40\% and 30\% determination with 4 $ab^{-1}$ and 3 $ab^{-1}$ at $\sqrt{s}=100$ TeV, implying on the order of a 10\% determination with 30 $ab^{-1}$.
For center of mass energies below the di-Higgs production threshold, as would be relevant for the future circular $e^+e^-$ colliders, one may determine $g_{111}$ indirectly as it contributes to associated production at one-loop order\cite{McCullough:2013rea,Englert:2013tya}. A 30\% determination may be possible with this approach.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.5\linewidth]{Self_Coupling.pdf}
\caption{\label{fig:selftc} Correlation between the critical temperature and SM-like Higgs scalar self-coupling in the singlet-extended SM, adapted from Ref.~\cite{Profumo:2014opa}. Colors indicate future sensitivities of the HL-LHC (purple), CEPC/FCC-ee (red), ILC (green), and SPPC/FCC-hh (yellow).}
\end{figure}
\subsection{Modified Higgs boson couplings to SM particles}
The aforementioned scenarios may lead to changes in the Higgs boson couplings to other particles through the effect of Higgs mixing and/or new loop contributions. In the case of doublet-singlet mixing, for example, the SM-like state $h_1$ and singlet-like state $h_2$ may be written as
\begin{eqnarray}
\nonumber
h_1& =&h \cos\theta + S \sin\theta \\
h_2 & = &h \sin\theta -S \cos\theta \ \ \ ,
\label{eq:mix}
\end{eqnarray}
where $|\cos\theta|\geq 1/\sqrt{2}$.
In the regime $m_2> m_1/2$, the SM-like Higgs boson has no new decays and its branching ratios are unchanged from the SM. However, the production cross section, and thus, signal strength, will be reduced by $\cos^2\theta$. Present LHC data imply $\cos^2\theta\buildrel > \over {_\sim} 0.71$\cite{Profumo:2014opa}, a bound expected to increase to $\sim 0.95$ with the HL-LHC. Figure~\ref{fig:mixewpt} shows the distribution of parameter space points for a SFOEWPT transition in the $\cos\theta$-$m_2$ plane for $2m_1>m_2> m_1/2$. One sees that for EWPT-viable regions of parameter space, there exist considerable possibilities for future precision Higgs boson studies to observe deviations from signal strengths expected for a purely SM Higgs boson. Some of this parameter space may be probed with the HL-LHC.
A circular $e^+e^-$ Higgs factory, such as the CEPC or FCC-ee, could allow a probe at better than one percent, and we see that there exist many SFOEWPT-viable models that would correspond to mixing of this magnitude or smaller.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.5\linewidth]{Mixing_Angle_v3.pdf}
\caption{\label{fig:mixewpt} Correlation between the singlet-like Higgs boson mass $m_2$ and doublet-singlet mixing angle $\theta$ in the singlet-extended SM for $m_1=125$ GeV, adapted from Ref.~\cite{Profumo:2014opa}. Black points indicate SFOWEPT-viable models that also satisfy electroweak precision constraints. Colors regions indicate allowed regions from various Higgs boson measurements, heavy Higgs searches, and electroweak precision constraints as discussed in Ref.~\cite{Profumo:2014opa}. Vertical lines indicate mixing angle sensitivities of the present LHC combined fit results and prospective future HL-LHC and CEPC sensitivities. The latter have been obtained assuming a similar sensitivity as would be obtained at the FCC-ee (TLEP 240) collider.}
\end{figure}
For the case where the singlet generates the SFOEWPT solely via loop effects, one may expect an observable reduction in the associated production cross section $\sigma(e^+e^-\to Zh)$ of more than 0.4\% . This possibility is illustrated in Fig.~\ref{fig:assoc}, where the dashed lines indicate the relative change in the associated production cross section and the solid lines indicate the value of the RHS of Eq.~(\ref{eq:sfoewpt}). In the regions where the RHS is $\buildrel > \over {_\sim} 1$, the reduction in $\sigma(e^+e^-\to Zh)$ is expected to be large enough to be observed at the CEPC or FCC-ee.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.5\linewidth]{singlet_hZZ.pdf}
\caption{\label{fig:assoc} Correlation between SFOEWPT and $\sigma(e^+e^-\to Zh)$ for the singlet extension of the Standard Model\cite{Katz:2014bha} in the singlet mass-Higgs portal coupling plane. Dashed red lines correspond to relative changes in $\sigma(e^+e^-\to Zh)$. Solid lines give values of the LHS of Eq.~(\ref{eq:sfoewpt}). Purple gives region of non-vanishing singlet vev.}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=0.8\linewidth]{heavysinglet_v2.pdf}
\caption{
\(2\sigma\) Contours of future precision measurements for a heavy singlet-catalyzed EWPT\cite{Henning:2014gca} in the $(a_1/m_S, m_S)$ plane. The magenta contour is the \(2\sigma\) sensitivity to the universal Higgs oblique correction entering associated production at the ILC 500up. Blue contours show the \(2\sigma\) RG-induced constraints from the \(S\) and \(T\) parameters from current measurements (solid) and future sensitivities at ILC GigaZ (dashed) and TLEP TeraZ (dotted). Regions of a viable first order EWPT, from Eq.~(\ref{eq:heavyewpt}), are shown in the gray, hatched regions for \(k\equiv a_2 =1\) and \(4\pi\). Figure courtesy of the authors of Ref.~\cite{Henning:2014gca}. \label{fig:heavysinglet}}
\end{figure}
When the singlet mass $m_S$ parameter is heavy ($m_S >> m_h$), one may integrate out the singlet degrees of freedom, thereby generating the higher dimension operators\cite{Henning:2014gca,Gorbahn:2015gxa}
\begin{eqnarray}
\mathcal{O}_H & = & \frac{1}{2}\left(\partial_\mu H^\dag H\right)^2 \\
\mathcal{O}_6 & = & (H^\dag H)^3\ \ \ .
\end{eqnarray}
The resulting contribution to the Lagrangian will be
\begin{equation}
\mathcal{L}_\mathrm{eff}\supset \frac{a_1^2}{m_S^4}\mathcal{O}_H-\left( \frac{a_1^2 a_2}{m_S^4}-\frac{2 a_2^3 b_3}{m_S^6}\right)\mathcal{O}_6
\end{equation}
where $b_3/3$ is the singlet cubic self coupling. The resulting universal Higgs coupling correction is
\begin{equation}
\delta Z_h = \frac{2 v^2 a_1^2}{m_S^4}
\end{equation}
independent of the quartic Higgs portal coupling $a_2$. On the other hand, both the cubic and quartic portal couplings can enable SFOEWPT if
\begin{equation}
\label{eq:heavyewpt}
\frac{2 v^4}{m_H^2} < \frac{m_S^2}{a_1^2 a_2} < \frac{6 v^4}{m_H^2}\ \ \ .
\end{equation}
The resulting constraints on the $(a_1, m_S)$ parameter space for different representative choices for $a_2$ are shown in Fig.~\ref{fig:heavysinglet}. The gray regions correspond to a SFOEWPT for different representative values of $a_2$. The region below the pink solid lines could be probed with the measurement of $\sigma(e^+e^-\to Zh)$ at the CEPC/FCC-ee. Additional constraints arising from present and possible future electroweak precision tests are induced by the blue lines.
For the Higgs portal scenarios involving SU(2$)_L$ non-singlet representations $\phi$, loop contributions may also lead to modified Higgs boson couplings to SM particles. If $\phi$ is an SU(3$)_C$ singlet, one will expect no modification of the gluon fusion operator $H^\dag H GG$, but new contributions to the rate for $h\to\gamma\gamma$ will appear whose impact will depend on the magnitude and sign of $a_2$ and the $\phi$ mass. For the real triplet illustration, the correlation between the relative change in $\Gamma(h\to\gamma\gamma)$, the values of $a_2$ and the triplet mass, and the occurrence of a two-step transition are shown in Fig.~\ref{fig:triplethgg}. One finds that the existence of the two-step transition generally implies a reduction in the Higgs diphoton decay rate. With the HL-LHC, the CMS collaboration projects a probe of the diphoton signal strength with 2\% precision\cite{CMS:2013xfa}, while ATLAS projects at least a 5\% determination\cite{ATL-PHYS-PUB-2014-016}. Combining the projected LHC determination of the ratio of branching ratios $\mathrm{BR}(h\to\gamma\gamma)/\mathrm{BR}(h\to ZZ)$ with the precise determination of the $ZZh$ coupling at the ILC would allow a one percent determination of the Higgs diphoton coupling\cite{Fujii:2015jha}. The present CEPC projection is for a 4.7\% determination of the coupling with 5 ab$^{-1}$ integrated luminosity\cite{CEPCpCDR}, while for the FCC-ee one anticipates 3\% determination of the signal strength with associated production and 10 ab$^{-1}$ of data\cite{Gomez-Ceballos:2013zzn}. Determinations at this level of precision would provide a decisive test of the simplest realization of the multi-step transition paradigm.
\begin{figure}[ht!]
\centering
\includegraphics[width=0.5\linewidth]{SigmaSM_TwoPhoton.pdf}
\caption{\label{fig:triplethgg} Shift in $\Gamma(h\to\gamma\gamma)$ in the real triplet extended SM~\cite{Patel:2012pi}. The red hatched region corresponds to a two-step phase transition. Dashed curves indicate relative changes in $\Gamma(h\to\gamma\gamma)$ compared to its Standard Model value. Solid green and red curves correspond to relative shifts of $\pm 35\%$ and $\pm 3\%$, respectively. Figure courtesy of H. Patel.}
\end{figure}
\subsection{Exotic Higgs boson Decays}
The discovery of non-standard (exotic) Higgs boson decays as well as of non-SM Higgses could provide important information about Higgs portal interactions. As indicated above, the decay $h\to Z_D Z_D$ would signal the presence of a new dark sector containing a SM gauge singlet scalar and a U(1)$^\prime$ gauge boson\cite{Curtin:2014cca}. For $m_S < m_h/2$, the SM Higgs boson may decay invisibly to a pair of $S$ bosons. In the absence of $h$-$S$ mixing, this decay mode will correspond to an invisibly decaying Higgs boson, whereas in the presence of $h$-$S$ mixing the final states will involve pairs of SM Higgs boson decay final states.
For the non-SM neutral Higgses, which appear in the new physics models with an extended Higgs sector, most of the current searches at the LHC focus on the conventional Higgs boson search channels of $WW$, $ZZ$, $\gamma\gamma$, $\tau\tau$ and $bb$ channel \cite{Aad:2014vgg,Khachatryan:2015qxa,CMS:2014cdp,Chatrchyan:2013qga,ATLAS-CONF-2013-027,Khachatryan:2014jya,Aad:2014kga,Khachatryan:2014wca,ATLAS-CONF-2013-090,CMS-PAS-HIG-13-021}. The production of the extra Higgses is usually suppressed compared to that of the SM Higgs boson, either due to its larger mass or its suppressed couplings to the SM particles. The decay of the heavy neutral Higgses to the $WW$ and $ZZ$ final states is absent for the CP-odd Higgs boson, and could be highly suppressed for the non-SM like CP-even Higgs boson. The decay modes of $\tau\tau$ or $bb$ suffer from either suppressed signal or large SM backgrounds, and are therefore only relevant for regions of the parameter space with an enhanced $bb$ or $\tau\tau$ coupling. The search for the charged Higgs bosons at the LHC is even more difficult. For $m_{H^\pm}>m_t$, the cross section for the dominant production channel of $tbH^\pm$ is typically small. The dominant decay mode $H^\pm \rightarrow tb$ is hard to identify given the large $tt$ and $ttbb$ background, while the subdominant decay of $H^\pm \rightarrow \tau\nu$ has suppressed branching fraction. In the MSSM, even at the end of the LHC running, there is a ``wedge region''~\cite{Dawson:2013bba} in the $m_A-\tan\beta$ plane for $\tan\beta\sim 7$ and $m_A\gtrsim 300$ GeV in which only the SM-like Higgs boson can be covered at the LHC. Similarly, the reach for the non-SM Higgses is limited in models with an extended Higgs sector.
\begin{table}
{\small
\begin{tabular}{l|l|l|l} \hline
&Decay&Final States&Channels \\ \hline
&$HH$ type & $(bb/\tau\tau/WW/ZZ/\gamma\gamma)(bb/\tau\tau/WW/ZZ/\gamma\gamma)$ & $H^0 \rightarrow A A$, ... \\ \cline{2-4}
Neutral Higgs&$HZ$ type & $(\ell\ell/qq/\nu\nu)(bb/\tau\tau/WW/ZZ/\gamma\gamma)$ & $H^0 \rightarrow A Z, A \rightarrow H^0 Z$, ... \\ \cline{2-4}
$H^0$, $A$& $H^+H^-$ type & $(tb/\tau\nu/cs)(tb/\tau\nu/cs)$ & $H^0 \rightarrow H^+H^-$, ... \\ \cline{2-4}
&$H^\pm W^\mp$ type & $(\ell\nu/qq^\prime)(tb/\tau\nu/cs)$ & $H^0/A \rightarrow H^\pm W^\mp$, ... \\ \hline
Charged Higgs&$ H W^\pm$ type & $(\ell\nu/qq^\prime)(bb/\tau\tau/WW/ZZ/\gamma\gamma)$ & $H^{\pm} \rightarrow H^0 W, A W$, ... \\ \hline
\end{tabular}
}
\caption{Exotic decay modes for Higgses in the 2HDM with the heavy Higgs boson decays to two light Higgses or one light Higgs boson with one SM gauge boson. $H$ in column two refers to any of the neutral Higgs bosons $h^0$, $H^0$ or $A$. }
\label{tab:decay_exo}
\end{table}
In addition to their decays to the SM particles, non-SM Higgses can decay via exotic modes, {\em i.e.}, heavier Higgs decays into two light Higgses, or one light Higgs boson plus one SM gauge boson. Five main exotic decay categories for Higgses of the Type II 2HDM are shown in Table~\ref{tab:decay_exo}. These channels typically dominate once they are kinematically open. A recent study on exotic Higgs boson decays can be found in Refs.~\cite{Coleppa:2013xfa,Coleppa:2014hxa,Brownson:2013lka, Coleppa:2014cca, AW_light,Hpm_Tong,Maitra:2014qea,Basso:2012st,Dermisek:2013cxa,Mohn:833753,Assamagan:2000ud,Assamagan:2002ne}.
Note that most of the current experimental searches for the non-SM Higgs boson assume the absence of exotic decay modes. Once there are light Higgs states such that these exotic modes are kinematically open, the current search bounds can be greatly relaxed given the suppressed decay branching fractions into SM final states~\cite{Coleppa:2014hxa,Coleppa:2014cca,Hpm_Tong}. Furthermore, exotic Higgs decays to final states with two light Higgses or one Higgs boson plus one SM gauge boson provide complementary search channels. Here, we list such exotic Higgs decays and consider potential search strategies.
\begin{itemize}
\item{$H^0\rightarrow A A$ or $H^0\rightarrow h^0 h^0$}
\end{itemize}
With the final state Higgs boson decaying via $bb$, $\gamma\gamma$, $\tau\tau$, $WW^*$, final states of $bbbb$, $bb\tau\tau$, $bb\gamma\gamma$ and $\gamma\gamma WW^*$ can be used to search for resonant Higgs boson decay to two light neutral Higgses. Current searches at the LHC 8 TeV with about 20 ${\rm fb}^{-1}$ luminosity observed 95\% C.L. cross section limits of 2.1 pb at 260 GeV and about 0.018 pb at 1000 GeV~\cite{Aad:2015xja} (see also Refs.~\cite{Aad:2015uka,Aad:2014yja,Khachatryan:2015yea}). While $bb\gamma\gamma$ and $bb\tau\tau$ have comparable sensitivities at low mass, the $bbbb$ mode dominates at high mass.
\begin{itemize}
\item{$H^0\rightarrow A Z$ or $A \rightarrow H^0 Z $}
\end{itemize}
With $Z \rightarrow \ell\ell$ and $H^0 /A \rightarrow bb, \tau\tau$, the final states of $bb\ell\ell$, $\tau\tau\ell\ell$ can be obtained with gluon fusion production, or in the $bb$ associated production with two additional $b$ jets~\cite{Coleppa:2013xfa,Coleppa:2014hxa,Brownson:2013lka}. Recent searches from ATLAS and CMS have shown certain sensitivity in this channel~\cite{Khachatryan:2015lba,CMS-PAS-HIG-14-011,Khachatryan:2014jya,Aad:2015wra,CMS-PAS-HIG-15-001}. In parameter regions where ${\rm Br}(A \rightarrow H^0 Z)\times {\rm Br}(H^0 \rightarrow Z Z)$ is not completely suppressed, $ZZZ$ final states with two $Z$ decaying leptonically and one $Z$ decaying hadronically can also be useful~\cite{Coleppa:2014hxa}. Other channels with top final states could be explored as well.
Note that the decay $A\to Z H^0$ has been identified as a particular signature of a SFOEWPT in the 2HDM\cite{Dorsch:2013wja}. As discussed below, the prospects for observing this channel in the $\ell\ell b{\bar b} $ and $\ell\ell W^+ W^-$ model have been analyzed in Ref.~\cite{Dorsch:2014qja}.
\begin{itemize}
\item{$H^0 \rightarrow H^+ H^- $}
\end{itemize}
With both $H^\pm$ decaying via $\tau\nu$ final states, the signal of $\tau\tau \nu \nu$ can be separated from the SM $W^+W^-$ background since the charged tau decay product in the signal typically has a hard spectrum compared to that of the background~\cite{Hpm_Tong}.
\begin{itemize}
\item{$H^0/A \rightarrow H^\pm W^\mp$ }
\end{itemize}
Similar to the $H^+H^-$ case, $H^\pm \rightarrow \tau \nu, tb$ and $W\rightarrow \ell \nu$ with $\ell \tau \nu \bar\nu$ or $tb\ell \nu$ could be used to search for $H^0/A \rightarrow H^\pm W^\mp$. Note that for the CP-even Higgs boson $H^0$, the branching fraction of $H^0 \rightarrow H^\pm W^\mp$ is mostly suppressed comparing to $H^0 \rightarrow H^+H^-$ as long as the latter process is kinematically open and not accidentally suppressed~\cite{Hpm_Tong}. However, for the CP-odd Higgs boson $A$, this is the only decay channel with a charged Higgs boson in the decay products.
\begin{itemize}
\item{$H^{\pm} \rightarrow H^0 W^\pm, A W^\pm $}
\end{itemize}
This is the only exotic decay channel for the charged Higgs boson in the 2HDM. Given the associated production of $tbH^\pm$, and the decay of $H^0$, $A$ into the $bb$ or $\tau\tau$ channel, $\tau\tau bb WW$ or $bbbbWW$ can be used to probe this channel~\cite{Coleppa:2014cca}. $H^0/A \rightarrow t\bar{t}$ could also be used given the boosted top in the high energy environment.
Note that while $H^\pm \rightarrow WZ$ is absent in 2HDM type extension of the SM Higgs sector, it could appear, however, in the real triplet models extension of the SM once the triplet obtains a vev~\cite{FileviezPerez:2008bj}.
\begin{itemize}
\item{$A/H^0\rightarrow tt$, $H^\pm \rightarrow tb$}
\end{itemize}
While $A/H^0\rightarrow tt$ and $H^\pm \rightarrow tb$ are considered to be challenging experimentally, a recent study of BSM Higgs searches with top quarks in the final states~\cite{Hajer:2015gka} showed that a combination of the channels of $pp\to bb H^0/A \rightarrow bb tt, bb\tau\tau$, $pp \to H^0/A \to tt$ as well as $pp \to tb H^{\pm} \rightarrow tbtb, tb \tau\nu$ yields full coverage for $\tan\beta$ and pushes the exclusion limits from the $\mathcal O(1)$ TeV at the LHC to the $\mathcal O(10)$ TeV at a 100 TeV $pp$ collider.
\subsection{New states}
In addition to observing deviations of Higgs boson couplings to SM particles, one may also anticipate direct production of new states associated with SFOEWPT-viable models. Again, the singlet scenario provides the simplest illustration. Its production cross section will be proportional to $\sin^2\theta$, presently constrained to be smaller than $\sim 0.34$. For $m_2< 2 m_1$, the signatures of this state would be identical to those of a SM Higgs boson with reduced signal strength\footnote{Recall that the mixing angle cancels from the branching ratios when no new decay channels are kinematically allowed.}. For $m_2> 2 m_1$, where $m_1$ is the mass of the SM-like Higgs boson, a new decay mode $h_2\to h_1 h_1$ becomes kinematically allowed, introducing the possibility of resonant di-Higgs production. The final states would entail pair-wise combinations of SM Higgs boson decay products, such as $b{\bar b}\gamma\gamma$,
$b{\bar b} \tau^+\tau^-$ {\em etc.}. Recently, the authors of Ref.~\cite{No:2013wsa} showed that observation of resonant di-Higgs production in the
$b{\bar b} \tau^+\tau^-$ channel -- the same one discussed in Ref.~\cite{Dolan:2012rv} for non-resonant production to determine the self-coupling -- could be feasible with 100 fb$^{-1}$ at the LHC, assuming the present maximal value of 0.34 for $\sin^2\theta$. Other studies have analyzed resonant di-Higgs production with $b{\bar b}\gamma\gamma$ and $4b$ final states, though not with an eye toward the EWPT \cite{Chen:2014ask,Barger:2014taa,Dolan:2012ac}.
Looking to the future, should the constraints on the mixing angle become more severe, higher integrated luminosity, a cleaner background environment, or higher parton luminosity as with the SPPC/FCC-hh would be needed to search for such modes.
For non-singlet $\phi$, such as the real triplet ($\Sigma$) or 2HDM discussed above, production and decays provide for a rich phenomenology. The electroweak $\rho$-parameter constraints on the non-doublet vevs imply that single scalar production is typically suppressed, and the dominant production mechanism will be electroweak pair production. For multi-Higgs doublet models, on the other hand, neutral scalars may be produced through gluon fusion, while charged states may be singly-produced through associated production. The final states in regions of interest for cosmology will then depend on the details of each scenario.
\vskip 0.1in
\noindent {\em Two Higgs doublets}. For the 2HDM, the authors of Ref.~\cite{Dorsch:2014qja} have shown that in the SFOEWPT-viable region of parameter space, the decay $A^0\to Z H^0$ could be discovered at the 14 TeV LHC in the $\ell\ell b{\bar b}$ and $\ell\ell W^+W^-$ channels with integrated luminosity in the $\buildrel > \over {_\sim} 20-40$ fb$^{-1}$ range. Analyses of decays involving charged scalars $H^\pm$ remain to be performed.
\vskip 0.1in
\noindent {\em Real triplet}. For vanishing triplet vev, the neutral component may contribute to the DM relic density. For masses in the several hundred GeV range, where the two step phase transition appears to be viable, the triplet contribution will comprise less than 10\% of the relic density due to the sizable cross section for annihilation into $W^+W^-$ pairs. At the LHC, triplet states are produced electroweakly in $\Sigma^+\Sigma^-$ and $\Sigma^\pm \Sigma^0$ pairs. Electroweak radiative corrections raise the $\Sigma^\pm$ masses relative to the $\Sigma^0$ mass by $\sim 160$ MeV, so that the $\Sigma\pm$ may decay to a $\Sigma^0$ plus a soft pion or lepton-neutrino pair. None of the SM final state particles will be detectable. The signature would, thus, be large $\mbox{${\not\! E}_{\rm T}$}$ plus one or more disappearing charged particle tracks. For a 150 GeV $\Sigma^+$ the $c\tau$ is of order 5 cm\cite{FileviezPerez:2008bj}. The prospective implications of the CMS search for disappearing charged particle tracks\cite{CMS:2014gxa} have been discussed above in Section \ref{sec:hpdm}.
For non-vanishing vev, both the neutral and charged triplet can decay, leading to a rich array of possible final states\cite{FileviezPerez:2008bj}. For $\Sigma^\pm$ masses well below the $W^\pm Z$ threshold, the dominant decay mode is $\tau\nu$ when the triplet vev is of order one GeV, while it is $\Sigma^0$ plus a soft pion for very small vev. The largest $\Sigma^0$ branching fraction is to $b{\bar b}$. Thus, one may anticipate $b{\bar b}\tau\nu$ final states for the first case and $b{\bar b}\gamma\gamma$ or $b{\bar b}\tau^+\tau^-$ for the second. In the study of Ref.~\cite{No:2013wsa}, it was found that a discovery with the $b{\bar b}\tau\nu$ channel may be possible with 100 fb$^{-1}$ at the 14 TeV LHC, with a significance depending on the magnitude and sign of the Higgs portal coupling $a_2$ and the triplet mass. Correlating this observation with the SM-like Higgs boson diphoton branching fraction, as discussed above, could indicate the presence of the real triplet with SFOEWPT-favorable parameters.
\section{Future Work}
It is clear that there exists a rich set of experimental signatures of the Higgs portal in cosmologically-relevant regions of parameter space. While detailed studies have been carried out for a subset of these signatures, substantial additional work is called for. In Table \ref{tab:signatures} below, we summarize the landscape of completed and prospective studies. From both this discussion and Table \ref{tab:signatures} we see that there exist considerable opportunities for both new experimental searches and theoretical studies. As the theoretical work develops, we will periodically up-date this document to reflect new results. We also point to a companion document that summarizes the outcome of a recent ACFI workshop \lq\lq Probing the Electroweak Phase Transition with a 100 TeV pp Collider" that was held at the University of Massachusetts Amherst in September 2015.
\begin{table}[hb]
\small
\begin{centering}
\begin{tabular}{|c|lc|c|c|c}
\hline\hline
Scenario & Conditions & Collider & Signatures & Ref. \\
\hline\hline
Real Singlet & EWPT & LHC & Signal reduction & \cite{Profumo:2007wc,Barger:2007im,Profumo:2014opa}\\
&& ILC, FCC-ee, & Signal reduction & \cite{Profumo:2007wc,Profumo:2014opa}\\
&& CEPC, SppC/FCC-hh &&\\
\hline
&EWPT&LHC & Modified self-coupling & \cite{Profumo:2007wc,Noble:2007kk,Katz:2014bha,Profumo:2014opa}\\
&& ILC, TLEP & Modified self-coupling & \cite{Profumo:2014opa}\\
&& CEPC, SppC/FCC-hh &&\\
\hline
&EWPT&LHC& $h_2\to h_1 h_1\to b{\bar b}\tau^+\tau^-$ & \cite{No:2013wsa}\\
&&& $h_2\to h_1 h_1\to b{\bar b}\gamma\gamma$ & \cite{Chen:2014ask,Barger:2014taa}\\
&&& $h_2\to h_1 h_1\to b{\bar b} b{\bar b}$ & \cite{Chen:2014ask} \\
&&& $h_2\to h_1 h_1\to b{\bar b}VV$ & In prog \\
&&& $h_1\to h_2 h_2\to XY$ & In prog \\
\hline
Real Singlet & DM & LHC & Invisible decays & \cite{Barger:2007im,He:2013suk,Curtin:2014jma}\\
Real Singlet & DM$^\ast$ \& EWPT & SppC/FCC-hh & VBF $h_1^\ast\to$ invis & \cite{Curtin:2014jma}\\
&&& invisible decays &\\
&&& modified self-coupling &\\
& & ILC/FCC-ee/CEPC & modified self-coupling &\cite{Curtin:2014jma}\\
&&& Signal reduction & \\
\hline\hline
Real Triplet & EWPT \& DM$^\ast$ & LHC & Disappearing charged & \cite{FileviezPerez:2008bj}\\
& & & particle tracks & \\
&&& Reduced $\Gamma(h\to\gamma\gamma)$ &\cite{Patel:2012pi}\\
\hline\hline
Real Triplet & EWPT only & LHC & Reduced $\Gamma(h\to\gamma\gamma)$& \cite{Patel:2012pi}\\
&&& $H^+ H_2^0\to b{\bar b}\tau\nu$ & \cite{FileviezPerez:2008bj}\\
\hline
&EWPT only&& $H^+ H_2^0\to H_2^0 H_2^0\pi^+$ & \\
&&& $H^+ H^-\to H_2^0 H_2^0\pi^+\pi^-$ & \\
&&& with & \\
&&& $\qquad\qquad H_2^0 H_2^0 \to b{\bar b}\gamma\gamma$ & \cite{FileviezPerez:2008bj}\\
&&& $\qquad\qquad H_2^0 H_2^0 \to b{\bar b}\tau^
+\tau^-$ & \\
\hline\hline
2HDM & EWPT & LHC & $A^0\to Z H^0\to \ell\ell b{\bar b}$ & \cite{Dorsch:2014qja}\\
&&& $A^0\to Z H^0\to \ell\ell W^+ W^-$ &\cite{Dorsch:2014qja} \\
\hline\hline
\end{tabular}
\caption{Higgs portal scenarios signatures. For each scenario, the cosmological relevance is indicated in the second column. The third column gives the collider for which the signature indicated in the fourth column has been analyzed. The final column gives appropriate references. For exotic Higgs boson decays, see Table \ref{tab:decay_exo}. $^\ast$ For this case, we do not assume saturation of the DM relic density.\label{tab:signatures}}
\end{centering}
\end{table}
\newpage
\section{Acknowledgements }
We thank Xiaochuan Lu for providing Figure \ref{fig:heavysinglet} and Michelangelo Mangano for providing detailed comments on this manuscript.
SL is supported through FWO Vlaanderen Odysseus II grant G.0C39.13N. X-G He was supported in part by the MOE Academic Excellent Program (Grant No.~102R891505) and MOST of R.O.C.
(Grant No.~MOST104-2112-M-002-015-MY3), and in part by NSFC (Grant Nos.~11175115 and 11575111) and Shanghai
Science and Technology Commission (Grant No.~11DZ2260700) of P.R.C. The work of A. Kotwal was supported by the Fermi National Accelerator Laboratory and by a Department of Energy grant to Duke University. Fermilab is operated by Fermi Research Alliance, LLC, under Contract No. DE-AC02-07CH11359 with the United States Department of Energy. J. Kozaczuk is supported by the National Sciences and Engineering Research Council of Canada (NSERC). J.M.N. is supported by the People Programme (Marie Curie Actions) of the European Union Seventh
Framework Programme (FP7/2007-2013) under REA grant agreement PIEF-GA-2013-625809. MJRM and PW were supported in part under U.S. Department of Energy contract DE-SC0011095.
\newpage
\vspace*{5mm}
\bibliographystyle{JHEP.bst}
|
\section{Introduction}
\label{sec:intro}
The Standard Model of particle physics has been proven an extremely successful
theory of nature, but it leaves many questions unanswered. It is consequently
widely acknowledged as an effective theory obtained from a more fundamental
theoretical context still to be observed. Supersymmetric extensions of the
Standard Model represent one of the most popular options for new physics and are
motivated by the unification of gauge and space-time symmetries. In addition,
they resolve the hierarchy problem inherent to the Standard Model, feature the
unification of the gauge couplings at high energy scales and naturally provide
an explanation to the presence of dark matter in the universe. By construction,
the Higgs sector of a supersymmetric theory is extended with respect to the
Standard Model case and contains at least two weak doublets of Higgs superfields
(traditionally noted $H_u$ and $H_d$) so that masses for both the up-type and
down-type fermions could be generated. Considering the minimal supersymmetric
extension of the Standard Model, the so-called Minimal Supersymmetric Standard
Model (MSSM)~\cite{Nilles:1983ge,Haber:1984rc}, only the $H_u$ and $H_d$ Higgs
supermultiplets are included and the superpotential contains a supersymmetric
mass term for these superfields $\mu H_u H_d$. While the dimensionful parameter
$\mu$ should in principle be of the order of the only natural scale of the
theory that is either the Planck or the gauge-coupling unification
scale, a working spontaneous breaking of the electroweak symmetry demands this
parameter to be in the ball park of a few hundreds of GeV. This puzzle is called
the `$\mu$-problem' of the MSSM~\cite{Kim:1983dt}. On different lines, the
discovery of a
scalar field with a mass of about 125~GeV and that resembles the Standard Model
Higgs boson~\cite{Aad:2012tfa,Chatrchyan:2012xdj} implies either the existence
of heavy top squarks or large top squark mixing, which raises questions about
the naturalness of the MSSM.
All these issues can be solved elegantly in the framework of the Next-to-Minimal
Supersymmetric Standard Model (NMSSM)~\cite{Ellwanger:2009dp}, where the model
includes an additional superfield $S$ that is singlet under the Standard Model
gauge group. As a result, the Higgs sector of the model features three neutral
scalar states, two neutral pseudoscalar states and one charged state, as well as
one singlino (the fermionic component of $S$) and two higgsinos (the fermionic
components of $H_u$ and $H_d$) fermions that will mix with the gauginos to form
five neutralinos. This enriched particle content yields a phenomenology that
could be largely different from the MSSM case and that could even
accomodate~\cite{Ellwanger:2016qax} the tantalizing hints of an excess of
diphoton events observed in LHC data at a center-of-mass energy of
13~TeV~\cite{ATLAS-CONF-2015-081,CMS:2015dxe}.
In order to satisfy the stringent constraints on the Standard Model Higgs boson
properties derived from LHC measurements~\cite{Khachatryan:2014jba,Aad:2015gba},
and in particular those that are put on the Higgs exotic decay modes,
phenomenologically viable NMSSM scenarios have to contain a
Standard-Model-like Higgs boson with a very small singlet component.
Consequently, one given scalar state and one given pseudoscalar state have to be
almost purely singlet, so that they couple to the Standard Model only through
their small mixing with the $H_u$ and $H_d$ fields. Furthermore, these singlet
fields are weakly constrained by current experimental data and can hence be as
light as a few GeV. This setup with two light singlet-like bosons is
further motivated by the Peccei-Quinn symmetry limit of the NMSSM where one
imposes that the model Lagrangian is invariant under a Peccei-Quinn-like
symmetry. This indeed not only allows the NMSSM to solve the
strong $CP$-problem~\cite{Miller:2003ay}, but also yields a very light
pseudoscalar singlet state $A_1$. Such a prediction has spurred an intense
phenomenological activity over the last years~\cite{Dermisek:2005ar,%
Dermisek:2006wr,Ellwanger:2005uu,Djouadi:2008uw,Cao:2013gba,Bomark:2014gya,%
Bomark:2015fga,Potter:2015wsa}, with a particular focus on processes where light
$A_1$ pairs are produced from the cascade decays of heavier Higgs bosons.
Pioneering works have investigated final state systems made of four jets issued
from the fragmentation of
$b$-quarks \mbox{$H\to A_1 A_1\to b\bar{b}b\bar{b}$}~\cite{Almarashi:2011te},
four leptonically or hadronically decaying tau leptons \mbox{$H\to A_1 A_1\to
\tau^+ \tau^- \tau^+ \tau^-$}~\cite{Forshaw:2007ra,Belyaev:2008gj,%
Cerdeno:2013cz}, four muons \mbox{$H\to A_1 A_1\to \mu^+\mu^-\mu^+\mu^-$}%
~\cite{Belyaev:2010ka} or of one pair of muons and one pair of tau leptons
\mbox{$H\to A_1 A_1\to \mu^+ \mu^- \tau^+ \tau^-$}~\cite{Lisanti:2009uy}. It has
been moreover shown that the discovery of such decay channels would consist of a
no-lose theorem for a direct evidence of the NMSSM~\cite{Ellwanger:2003jt,%
Ellwanger:2013ova}. This has consequently opened the path for dedicated NMSSM
searches in LHC collision data at a center-of-mass energy of 8~TeV~\cite{%
Khachatryan:2015nba,Aad:2015oqa,Khachatryan:2015wka}. Upper limits on the
production cross sections related to the four taus, two taus and two muons, and
four muons decay modes of the heavy Higgs boson of 4.5--10.3~pb, 0.72--2.33~pb
and of about 1~fb have been respectively derived.
In this work, we study Higgs data constraints on the NMSSM and
show that phenomenologically viable scenarios feature configurations in which
the lightest pseudoscalar $A_1$ is dominantly produced from the decays of
neutralino states~\cite{Cheung:2008rh,Cerdeno:2013qta,Han:2015zba}. The LHC
constraints on these scenarios are still both rather weak and very model
dependent. In particular, $\tilde{\chi}^0_i \to A_1 \tilde{\chi}^0_j$ decays
often lead to the production of boosted $A_1$ particles that are difficult to
detect due to the collimation of their decay products into a single object,
regardless of the mass splitting between the two neutralinos $\tilde{\chi}^0_i$
and $\tilde{\chi}^0_j$. We explore in Section~\ref{sec:nmssm} the parameter
space of the NMSSM and investigate specific scenarios compatible with the
above-mentioned Higgs requirements
and featuring a light pseudoscalar state $A_1$, with a focus on cases where its
mass is of at least twice as large as the tau lepton mass and smaller than twice
the mass of the $b$-quark. In our process for constructing such scenarios, we
additonally impose dark matter considerations on the lightest
supersymmetric partner. We then
investigate, in Section~\ref{sec:simu}, the
sensitivity of the current LHC run at 13~TeV to such scenarios and show that
they could be detected through the analysis of a signature comprised of a single
lepton, a ditau-tagged jet and missing transverse energy. To this aim, we make
use of a ditau tagging technique that has been developed in the context of
the Higgs~\cite{Englert:2011iz,Papaefstathiou:2014oja,Katz:2010iq} and that
we have supplemented to a multivariate analysis dedicated to the tagging of the
signal. Our conclusions are summarized in Section~\ref{sec:conl}.
\section{Light scalar and pseudoscalar Higgs bosons in the NMSSM}
\label{sec:nmssm}
\subsection{Theoretical framework}
The NMSSM is constructed by augmenting the MSSM superfield content by one
superfield $S$ that is a singlet under $SU(3)_c\times SU(2)_L\times U(1)_Y$.
After the breaking of supersymmetry (and the consequent breaking of the
electroweak symmetry), the scalar component of $S$ mixes with the Higgs
degrees of freedom, whilst the fermionic component of $S$, dubbed the
singlino, mixes with the two remaining higgsino states and the gauginos. As for
any softly broken supersymmetric theory, the NMSSM is specified by its
superpotential and its supersymmetry-breaking Lagrangian. The superpotential
reads
\begin{equation}
W_{\rm NMSSM} =
- L H_d \mathbf{y_e} E
- Q H_d \mathbf{y_d} D
+ Q H_u \mathbf{y_u} U
+ \lambda S H_u H_d
+ \frac13 \kappa S^3\ ,
\label{eq:wnmssm}\end{equation}
where all indices are omitted for brevity, where $Q$ and $L$ denote the weak
doublets of quark and lepton superfields and where $U$, $D$ and $E$ are the
up-type quark, down-type quark and lepton weak-singlet superfields,
respectively. In addition, we have
introduced the $3\times 3$ Yukawa matrices $\mathbf{y}$ and the $\lambda$
and $\kappa$ parameters that drive the couplings of $S$. In particular, once
the scalar component $s$ of the singlet superfield gets a vacuum expectation
value $\langle s\rangle = v_s/\sqrt{2}$, an effective $\mu$-term is generated,
\begin{equation}
\mu_{\rm eff} = \frac{1}{\sqrt{2}} \lambda v_s \ ,
\end{equation}
which solves the MSSM $\mu$-problem. In the expression of Eq.~\eqref{eq:wnmssm},
we have imposed that the superpotential satisfies a $\mathbb{Z}_3$ symmetry so
that any dimensionful term allowed by the gauge symmetry is forbidden.
The soft supersymmetry breaking Lagrangian contains mass terms for all scalar
(${\bf m^2_{\tilde Q}}$, ${\bf m^2_{\tilde U}}$, ${\bf m^2_{\tilde D}}$,
${\bf m^2_{\tilde L}}$, ${\bf m^2_{\tilde E}}$, $m_{H_u}^2$, $m_{H_d}^2$ and
$m_s^2$) and gaugino ($M_1$, $M_2$ and $M_3$) fields, as well as trilinear
interaction terms (${\bf A^u}$, ${\bf A^d}$, ${\bf A^e}$, $A_\lambda$ and
$A_\kappa$) sharing the form of the superpotential,
\begin{equation} \bsp
\mathcal{L}_{\text{soft}} =
&\ - \frac12 \Big[
M_1\ \tilde B \tilde B +
M_2\ \tilde W \tilde W +
M_3\ \tilde g \tilde g +
{\rm h.c.} \Big]\\
&\ - {\bf m^2_{\tilde Q}}\ \tilde q^\dag \tilde q
- {\bf m^2_{\tilde U}}\ \tilde u^\dag \tilde u
- {\bf m^2_{\tilde D}}\ \tilde d^\dag \tilde d
- {\bf m^2_{\tilde L}}\ \tilde \ell^\dag \tilde \ell
- {\bf m^2_{\tilde E}}\ \tilde e^\dag \tilde e
- m_{H_u}^2\ h_u^\dag h_u
- m_{H_d}^2\ h_d^\dag h_d
- m_s^2\ s^\dag s\\
&\ + \Big[ - {\bf y^u}{\bf A^u} \ \tilde u^\dag \tilde q h_u
+ {\bf y^d}{\bf A^d} \ \tilde d^\dag \tilde q h_d
+ {\bf y^e}{\bf A^e} \ \tilde e^\dag \tilde \ell h_d
- \lambda A_\lambda h_u h_d s
- \frac13 \kappa A_\kappa s^3 + {\rm h.c.} \Big]\ ,
\esp\label{eq:lsoft} \end{equation}
where $q$, $\ell$, $u$, $d$, $e$, $h_u$, $h_d$ and $s$ denote the scalar
components of the $Q$, $L$, $U$, $D$, $E$, $H_u$, $H_d$ and $S$ superfields,
respectively, and $\tilde B$, $\tilde W$ and $\tilde g$ the gauginos associated
with the $U(1)_Y$, $SU(2)_L$ and $SU(3)_C$ gauge groups. All indices are again
understood for clarity.
In order to reduce the number of free parameters, we assume that all parameters
related to the (s)fermion sector are flavor-conserving and universal at the
grand unification scale. Introducing the common scalar mass $m_0$, the common
gaugino mass $m_{1/2}$ and the common trilinear coupling $A_0$, we have
\begin{equation}
{\bf m^2_{\tilde Q}} = {\bf m^2_{\tilde U}} = {\bf m^2_{\tilde D}} =
{\bf m^2_{\tilde L}} = {\bf m^2_{\tilde E}} =
m_0^2 \ \mathbf{1}_{3\times 3} \ ,~~
M_1 = M_2 = M_3 = m_{1/2} \ ,~~
{\bf A^u} = {\bf A^d} = {\bf A^e} = A_0 \ \mathbf{1}_{3\times 3} \ ,
\end{equation}
where $\mathbf{1}_{3\times 3}$ stands for the identity matrix in flavor space.
In this framework, the Higgs sector is defined by the soft parameters
$A_\lambda$ and
$A_\kappa$ that we fix at the grand unification scale, and by the $\lambda$,
$\kappa$, $\tan\beta$ and $\mu_{\rm eff}$ parameters that are provided at the
electroweak scale, $\tan\beta$ being the ratio of the vacuum expectation values
of the neutral components of the two Higgs doublets $h_u$ and $h_d$.
\subsection{Exploration of the NMSSM parameter space}
\begin{table}
\footnotesize
\center
\setlength{\tabcolsep}{2.5mm}
\renewcommand{\arraystretch}{1.4}
\begin{tabular}{c|c|c|c|c}
$m_0$ & $m_{1/2}$ & $A_0$ & $A_\lambda$ & $A_\kappa$\\\hline
[400, 2000]~GeV & [1000, 2000]~GeV & [-5000, -1000]~GeV & [-500, 500]~GeV &
[0, 300]~GeV\\
\end{tabular}\\[.3cm]
\begin{tabular}{c|c|c|c}
$\lambda$ & $\kappa$ & $\tan \beta$ & $\mu$\\ \hline
[0.2, 0.5] & [0.01, 0.2] & [1.5, 15] & [100, 350]~GeV \\
\end{tabular}
\caption{Parameterization of the NMSSM parameter space explored in this work.
We indicate the ranges in which the different parameters have been allowed
to vary. In the first table, the parameters are provided at the grand
unification scale while in the second table, they are given at the
electroweak scale.}
\label{tab:range}
\end{table}
In the previous section, we have defined a parameterization of the NMSSM in
terms of nine free parameters,
\begin{equation}
m_0,\qquad m_{1/2},\qquad A_0,\qquad A_\lambda,\qquad A_\kappa,\qquad
\lambda,\qquad \kappa, \qquad
\tan\beta \quad \text{and}\quad \mu_{\rm eff},
\label{eq:susyprm}\end{equation}
the first five parameters being defined at the grand unification scale and the
last four parameters being defined at the electroweak scale. We supplement to
these the parameters related to the Standard Model sector,
\begin{equation}\bsp
&
\alpha_s(m_Z) = 0.1172,\qquad
G_F = 1.16639\ 10^{-5}~{\rm GeV}^{-2},\qquad
\alpha(m_Z) = 1/127.92,\qquad
m_Z = 91.187~\text{GeV},\\
&\hspace*{2.5cm}
m_t^{\rm pole} = 173.1~\text{GeV},\qquad
m_b(m_b) = 4.214~\text{GeV},\qquad
m_\tau = 1.777~\text{GeV}.
\esp\end{equation}
The QCD interaction strength is computed from the value of the strong coupling
constant at the $Z$-pole $\alpha_s(m_Z)$ and the three independent electroweak
inputs, whose values are taken from the Particle Data Group
review~\cite{Agashe:2014kda}, are chosen to be the Fermi constant $G_F$, the
$Z$-boson mass $m_Z$ and the electromagnetic coupling evaluated at the $Z$-pole
$\alpha(m_Z)$. Finally, the third generation fermion sector is defined by the
pole mass of the top quark $m_t^{\rm pole}$, the running $\overline{\rm MS}$
mass of the bottom quark $m_b(m_b)$ evaluated at the $m_b$ scale and the tau
mass $m_\tau$, all other fermion masses being neglected. All the couplings and
masses appearing in the NMSSM Lagrangian can be subsequently numerically
calculated, using in particular the relations determined by the minimization of
the scalar potential.
For our exploration of the NMSSM parameter space, we use the {\sc NmssmTools}
package~\cite{Ellwanger:2006rn,Das:2011dg,Muhlleitner:2003vg} and perform a scan
over the parameters given in Eq.~\eqref{eq:susyprm}. The ranges in which the
parameters are allowed to vary are given in Table~\ref{tab:range}, and
for scanned each
point, we impose a set of constraints that allows us to accept it or reject it.
Additionally to theoretical considerations such as obtaining a physical spectrum
that does not exhibit any tachyonic state or preventing the appearance of a
Landau pole below the grand unification scale, we impose limits on the Higgs
sector and on the supersymmetric particles derived from collider searches at
LEP, at the Tevatron and at the LHC. We moreover verify the consistency of the
selected points with dark matter data.
In order to accomodate a Higgs boson with properties close to those expected in
the case of the Standard Model, the mixing between the singlet state $s$ and the
two Higgs doublets $h_u$ and $h_d$ has to be small. In this case, the tree-level
masses of the $CP$-even and $CP$-odd singlet-like Higgs bosons are mostly given
by the corresponding entries in the $3 \times 3$ scalar and pseudoscalar squared
mass matrices ${\cal M}^2_S$ and ${\cal M}^2_P$,
\begin{align}
({\cal M}_S^2)_{33} &=
\lambda A_\lambda \frac{ v^2 \sin 2 \beta}{2 v_s}
+ \frac{\kappa v_s}{\sqrt{2}} \Big(A_\kappa+2\sqrt{2}\kappa v_s\Big)~,~
\label{eq:mh1} \\
({\cal M}_P^2)_{33} &=
\lambda A_\lambda \frac{ v^2 \sin 2 \beta}{2 v_s}
+ \frac{\kappa v_s}{\sqrt{2}}
\Big(\frac{\sqrt{2} \lambda v^2\sin 2\beta}{v_s} - 3 A_\kappa\Big)
\label{eq:ma1}~,~
\end{align}
where the vacuum expectation value $v$ is defined, at the tree level, by
\begin{equation}
v^2 = \frac{4 m_Z^2}{g_1^2 + g_2^2}\ ,
\end{equation}
$g_1$ and $g_2$ denoting the hypercharge and weak coupling constants.
In this work, we focus on the phenomenology of scenarios in
which the lightest
pseudoscalar state $A_1$ is mostly singlet-like, so that
$m^2_{A_1}\approx ({\cal M}^2_P)_{33}$, and where its mass is heavier than twice
the tau lepton mass and lighter than twice the bottom quark mass,
$m_{A_1} \in [2 m_\tau, 2 m_b]$. As a consequence, a cancellation between the
different terms of Eq.~\eqref{eq:ma1} should be in place. Since the natural size
of each term is of about the electroweak scale squared
$\mathcal{O}(10^4)$~GeV$^2$, the mass of the lightest singlet-like scalar state,
that is approximately given by Eq.~\eqref{eq:mh1} in the absence of a too large
singlet-doublet mixing, will be smaller than 125~GeV. It turns out that the
second scalar Higgs particle $H_2$ has to be identified with the boson
discovered during the first run of the LHC. Its mass $m_{H_2}$ can moreover
be generally compatible with the observed value of 125~GeV, the presence of
the lighter singlet-like $H_1$ state helping to increase it through mixing
effects~\cite{Kang:2012sy}. We thus impose in our scan that $m_{H_2}$ lies in
the $[122.1, 128.1]$~GeV mass window, such a large range allowing us to account
for the various sources of uncertainties both on the experimental result and on
the theoretical prediction.
Beside its mass, the first run of the LHC has allowed one to measure a lot of
properties of the Higgs boson with high precision. One specific ensemble of such
measurements consists of the so-called Higgs signal strengths $\mu_{X,Y}$ that
are defined as the ratios of predicted rates in a particular production channel
$X$ and decay mode $Y$ of the Higgs boson in a given new physics theory (being
here the NMSSM) to the Standard Model expectation,
\begin{equation}
\mu_{X,Y} = \epsilon_{X,Y}\ \frac{\sigma_X}{\sigma^{\rm SM}_X}\
\frac{{\rm BR}(h\to Y)}{{\rm BR}^{\rm SM}(h\to Y)}\ .
\end{equation}
The $\epsilon_{X,Y}$ factor in the above expression is related to the acceptance
and efficiency of the analyses under consideration, and could be different in a
new physics context and in the Standard Model. Such differences are however
usually assumed to be mild and to largely cancel in the ratio. The signal
strengths moreover depend on the Higgs production cross sections in the
$X$ channel ($\sigma_X^{\rm SM}$ and $\sigma_X$) and branching ratios associated
with the $h\to Y$ decay mode (${\rm BR}(h\to Y)$ and
${\rm BR}^{\rm SM}(h\to Y)$) both in the Standard Model and
in the NMSSM. Making use of the NMSSM signal strengths predicted by {\sc
NmssmTools} and the experimental measurements at the LHC~\cite{Aad:2015gba,%
Khachatryan:2014jba}, we construct a $\chi^2$ quantity~\cite{Bernon:2014vta}
\begin{align}
\chi_Y^2 = & a_Y (\mu_{\text{ggF},Y} - \hat{\mu}_{\text{ggF},Y})^2 + 2 b_Y (\mu_{\text{ggF},Y} - \hat{\mu}_{\text{ggF},Y}) (\mu_{\text{VBF+VH},Y} - \hat{\mu}_{\text{VBF+VH},Y}) \nonumber \\
&+ c_Y (\mu_{\text{VBF+VH},Y} - \hat{\mu}_{\text{VBF+VH},Y})^2~,~
\end{align}
where the `ggF', `VBF' and 'VH' indices respectively refer to the gluon-fusion,
vector-boson fusion and Higgs-Strahlung Higgs production mechanisms. The
$\hat{\mu}_{X,Y}$, $a_Y$, $b_Y$ and $c_Y$ parameters resulting from the fit of
the LHC Higgs measurements are taken from the {\sc Lilith}-1.1.3 program using
the 15.09 database~\cite{Bernon:2015hsa}, and we require $\chi_Y^2 < 6.18$ so
that our NMSSM predictions are compatible at the $2\sigma$ level with data.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{lk}
\includegraphics[width=0.48\textwidth]{bmu} \\
\includegraphics[width=0.48\textwidth]{alak}
\includegraphics[width=0.48\textwidth]{brs}
\caption{\label{fig:scan} Benchmark NMSSM scenarios selected in our scanning
procedure of the NMSSM parameter space, presented in the $(\lambda, \kappa)$
(upper left panel),
$(\tan\beta, \mu_{\rm eff})$ (upper right panel) and $(A_\lambda, A_\kappa)$
(lower left panel) planes. All models for which a physical NMSSM spectrum at
the electroweak scale is found are represented, under the condition that the
second lightest scalar state $H_2$ is Standard-Model-like. We distinguish the
cases where $m_{A_1}<30$~GeV (gray) and where $2 m_\tau < m_{A_1} < 2 m_b$
(green). In the lower right panel, we depict the dependence of the second and
third neutralino branching fractions into a pseudoscalar $A_1$ state on the
compressivity of the spectrum defined as $\mu_{\rm eff}-m_{\tilde\chi_1^0}$,
for all points satisfying $2 m_\tau < m_{A_1} < 2 m_b$.}
\end{figure}
All the points satisfying the requirements introduced so far are represented in
Figure~\ref{fig:scan} as
a function of the parameters defining the NMSSM Higgs sector. More precisely, we
show results in the $(\lambda, \kappa)$, $(\tan\beta, \mu_{\rm eff})$ and
$(A_\lambda, A_\kappa)$ planes in which correlations can be observed. Each
represented point refers to a particle spectrum free from any tachyonic state
and in which the second lightest Higgs boson $H_2$ has a mass comprised in the
[122.1, 128.1]~GeV range and yields a good $\chi^2$ fit of the measured
signal strengths. We indicate the scenarios for which the pseudoscalar mass is
smaller than 30~GeV by gray points, while the green points correspond to cases
in which $2 m_\tau < m_{A_1} < 2 m_b$. It turns out that small $\lambda$ and
$\kappa$ values are generally favored, which results in a light $A_1$ state as
given by
Eq.~\eqref{eq:ma1}. On different lines, the $\tan\beta$ and $\mu_{\rm eff}$
parameters are strongly correlated, and a large (small) value of $\tan\beta$
leads to a small (large) $\mu_{\rm eff}$ value. This feature originates from
imposing that the $H_2$ scalar boson is Standard-Model-like.
Finally, an approximately linear correlation between the $A_\lambda$ and
$A_\kappa$ parameters is shown in the lower left panel of the figure. This
arises from the singlet-doublet mixing that is more conveniently
assessed when the scalar components of the Higgs fields are rotated to the
so-called Higgs basis $(H_{\rm SM}, H', S')$ by means of an appropriate $U(2)$
transformation~\cite{Ginzburg:2004vp,Davidson:2005cw,Haber:2006ue}. This basis
choice has the advantage that only one of the non-singlet fields, $H_{\rm SM}$,
acquires a vacuum expectation value $v$ and features Standard-Model-like
interactions with the Standard Model fermions and gauge bosons. The second
non-singlet field $H'$ has thus a vanishing vacuum expectation value and the
third basis element $S'$ remains a pure singlet state. In the Higgs basis, the
size of the singlet-doublet mixing is given by the \mbox{$H_{\rm SM}-S'$}
element of the squared mass matrix~${\cal M}^{\prime 2}_S$~\cite{Miller:2003ay,%
Ellwanger:2015uaz},
\begin{equation}
({\cal M}_S^{\prime2})_{13} = \frac{\lambda v}{\sqrt{2}}
\Big(2 \mu_{\rm eff}-(A_\lambda + \sqrt{2}\kappa v_s) \sin 2\beta\Big)~.~
\end{equation}
In order for $H_2$ to be Standard-Model-like and thus mostly equivalent to
$H_{\rm SM}$ in the Higgs basis, $({\cal M}_S^{\prime2})_{13}$ must be small,
which leads to
\begin{equation}
A_\lambda \sim \frac{2 \mu_{\rm eff}}{\sin 2 \beta} -
\frac{2 \kappa \mu_{\rm eff}}{\lambda}\ .
\label{eq:alconstraint}\end{equation}
However, in the NMSSM parameter space region of interest, we have
$\sin 2\beta\ll\lambda/\kappa$. Consequently, only the first term of the
right-hand side of Eq.~\eqref{eq:alconstraint} matters. Moreover,
$\mu_{\rm eff}$ and $\tan\beta$ are correlated, as shown in the upper right
panel of Figure~\ref{fig:scan}, and both $A_\lambda$ and $A_\kappa$ are
connected to the Higgs spectrum, as shown, \textit{e.g.}, by Eq.~\eqref{eq:mh1} and
Eq.~\eqref{eq:ma1}. Although deriving the exact relation between the $A_\lambda$
and $A_\kappa$ parameters that are defined in our setup at the grand unification
scale is a complex task due to renormalization group running, an
approximatively linear relation can be derived and is indeed observed in the
figure.
As a result of the above constraints, the typical spectrum exhibited by the
scenarios selected during our scan of the NMSSM parameter space features a
lightest neutralino state $\tilde\chi_1^0$ that is mainly singlino-like, whilst
the next two neutralinos $\tilde\chi_2^0$ and $\tilde\chi_3^0$ are close in mass
and higgsino-like. Once produced, these two heavier neutralinos can decay,
sometimes with a large branching ratio, into a final state system comprised of a
singlino $\tilde\chi_1^0$ and a pseudoscalar Higgs boson $A_1$. In the lower
right panel of Figure~\ref{fig:scan}, we present the dependence of these
branching ratios on the mass difference between the higgsino states (with masses
being approximatively taken as $\mu_{\rm eff}$) and the singlino.
This shows that an important $A_1$ production rate can be achieved in the region
where the spectrum is compressed, \textit{i.e.}, when the mass difference between the
higgsinos and the singlino is small. As soon as
$\mu_{\rm eff}-m_{\tilde\chi_1^0} \gtrsim 90$~GeV, the $\tilde{\chi}^0_i \to Z
\tilde{\chi}^0_1$ channel opens and quickly dominates. We have included in our
results the $\tilde\chi_3^0 \to A_1 \tilde\chi_2^0$
decay contributions. Although the phase space for such a decay process is very
limited as both higgsino states are close in mass (so that the $A_1$ decay
products are soft and very hard to detect), the final-state $\tilde\chi_2^0$
higgsino will further decay into an $A_1$ particle that will be energetic enough
to leave observable tracks in a detector.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{omega}
\includegraphics[width=0.48\textwidth]{ssi}
\caption{\label{fig:dm} Dark matter constraints that have been imposed during
our NMSSM parameter space scan. We show, in the left panel, the dark matter
relic density $\Omega h^2$ as a function of the mass splitting between the
lightest (singlino-dominated) neutralino $\tilde\chi_1^0$ and the higgsinos,
whose mass is of about $\mu_{\rm eff}$. We additionally provide information on the lightest neutralino mass (color code). On the right panel, we present
the spin-independent dark-matter--proton scattering cross section scaled
by $\Omega h^2 /0.119$ as a function of the lightest neutralino mass. We
focus on scenarios for which $2 m_\tau < m_{A_1} < 2 m_b$ (gray) that
additionally feature $\Omega h^2 < 0.131$ (orange) or
$0.107 < \Omega h^2 < 0.131$ (blue). We indicate the LUX bounds from 2015 by
a dark green line.}
\end{figure}
Additionally to the constraints that we have imposed so far, we moreover
restrict the properties of the lightest neutralino $\tilde\chi_1^0$ so that it
could be a good dark matter candidate. To this aim, we first impose that the
associated relic density abundance $\Omega h^2$ is compatible with the latest
Planck results~\cite{Ade:2015xua},
\begin{equation}
0.107 < \Omega h^2 < 0.131\ ,
\label{eq:Planck}\end{equation}
where the allowed range for $\Omega h^2$ includes theoretical uncertainties on
the predictions. Since the $\tilde\chi_1^0$ particle is of a singlino-dominated
nature, a small $\mu - m_{\tilde{\chi}^0_1}$ mass splitting, or equivalently a
large higgsino-singlino mixing, is required for efficient enough dark matter
(co)annihilation processes. This is illustrated on the left panel of
Figure~\ref{fig:dm} in which
we study the relations between the lightest neutralino relic density as computed
with the {\sc MicrOmegas} package~\cite{Belanger:2005kh,Belanger:2013oya} and
the $\mu - m_{\tilde{\chi}^0_1}$ splitting. We observe that the scenarios
selected in our scan are spread among three regions of the parameter
space. In the heavy dark matter region for which $m_{\tilde{\chi}^0_1} \gtrsim
m_{H_2}/2$, dark matter annihilation proceeds via an $s$-channel $Z$-boson
exchange diagram that highly depends on the higgsino fraction of the lightest
neutralino $\tilde\chi_1^0$. As a consequence, a relic density in agreement with
Planck data implies that the mass splitting $\mu - m_{\tilde{\chi}^0_1}$ is at
most of about 70~GeV. In contrast, in the two other regions, the so-called Higgs
funnel region for which $m_{\tilde{\chi}^0_1} \sim m_{H_2}/2$ and $Z$-boson
funnel region for which $m_{\tilde{\chi}^0_1} \sim m_{Z}/2$, larger splittings
are allowed as the dark matter annihilation cross section is enhanced by the
presence of resonant diagrams that allow to recover a relic density in agreement
with Eq.~\eqref{eq:Planck}.
We next focus on the spin-independent dark-matter--proton scattering cross
section,
\begin{equation}
\tilde{\sigma}^{\text{SI}}_p=\sigma^{\text{SI}}_p \frac{\Omega h^2}{0.119}\ ,
\end{equation}
and compare the NMSSM predictions obtained with {\sc MicrOmegas} to LUX
data~\cite{Akerib:2015rjg}. The results are
presented in the right panel of Figure~\ref{fig:dm}, in which we demonstrate
that many models in all three regions can survive all considered dark matter
constraints. The heavy dark matter region is additionally expected to be easier
to probe within future dark matter direct detection experiments that will
further constrain the scattering cross section $\tilde{\sigma}^{\text{SI}}_p$.
\subsection{Identification of benchmark scenarios for Run--II LHC studies}
\label{sec:benchmark}
All the NMSSM scenarios compliant with the constraints imposed so far exhibit
common features. The sfermions are typically lying in the multi-TeV region are
are thus
out of the reach of the LHC from the direct search standpoint, at least with an
assumed luminosity of 50--100~fb$^{-1}$. In contrast, the singlino and higgsino
states are in general light (with masses below 400--500~GeV) and the gauginos
heavy (with masses greater than 1~TeV). As a results, the only superpartners
that could be produced with a sufficiently large rate, and thus observed,
consist of the second and third neutralinos (or equivalently, the higgsino-like
neutralinos) and the lightest chargino. As already discussed in the previous
subsection, the singlet-like Higgs boson $H_1$ has a mass smaller than about
100~GeV, while the second scalar state $H_2$ is identified with the
Standard-Model Higgs boson. The lightest pseudoscalar state $A_1$ is
singlet-like with a mass satisfying $2 m_\tau < m_{A_1} < 2 m_b$ and all other
Higgs particles are beyond 1~TeV.
\begin{table}
\footnotesize
\center
\setlength{\tabcolsep}{2.5mm}
\renewcommand{\arraystretch}{1.4}
\begin{tabular}{c|c|c|c|c||c|c|c|c}
$m_0$ & $m_{1/2}$ & $A_0$ & $A_\lambda$ & $A_\kappa$&
$\lambda$ & $\kappa$ & $\tan \beta$ & $\mu$\\ \hline
1215.3~GeV & 1872.8~GeV & -4112.1~GeV & 301.1~GeV & 204.8~GeV&
0.317 & 0.122 & 12.2 & 121.3~GeV
\end{tabular}\\[.3cm]
\begin{tabular}{c|c|c||c||c|c||c|c|c}
$m_{\tilde{\chi}^0_1}$ & $m_{\tilde{\chi}^0_2}$ & $m_{\tilde{\chi}^0_3}$ &
$m_{\tilde{\chi}^\pm_1}$ & $m_{A_1}$ & $m_{A_2}$ & $m_{H_1}$ & $m_{H_2}$ &
$m_{H_3}$ \\ \hline
75.7~GeV & -135.3~GeV & 149.2~GeV & 124.2~GeV & 5.5~GeV & 1538~GeV & 93.8~GeV
& 125.9~GeV & 1538~GeV
\end{tabular}\\[.3cm]
\begin{tabular}{c||c|c||c||c}
Br$({\tilde{\chi}^0_2\to A_1 \tilde{\chi}^0_1})$ &
Br$({\tilde{\chi}^0_3\to A_1 \tilde{\chi}^0_1})$ &
Br$({\tilde{\chi}^0_3\to A_1 \tilde{\chi}^0_2})$ &
Br$({A_1 \to \tau \tau})$ & Br$({H_2 \to A_1 A_1})$\\
98.9\% & 12.9\% & 87.1\% & 93.6\% & 4.2\%
\end{tabular}\\[.3cm]
\begin{tabular}{c|c||c|c}
$\mu_{gg\to H,\gamma \gamma}$ & $\mu_{{\rm VBF}, VV^*}$ &
$\Omega h^2$ & $\sigma^{\text{SI}}_p$ \\ \hline
1.06 & 1.02 & 0.107 & $2.46 \times 10^{-10}$~pb
\end{tabular}
\caption{\label{tab:bp} Representative NMSSM benchmark scenario satisfying all
Higgs and dark matter constraints imposed in our scan. The scenario is
defined in the upper part of the table, in which we recall that the
parameters of the left part of the table are given at the grand unification
scale while those of its right part are defined at the electroweak scale.
The light state masses are given in the middle panel of the table, while
relevant branching ratios and obsevable results are given in its lower
parts.
}
\end{table}
In order to study the discovery prospects for these scenarios at the LHC, we
focus on a specific benchmark point that is defined in terms of the nine
parameters given in the upper panel of Table~\ref{tab:bp}. The light part of the
resulting mass spectrum is presented in the second panel of the table, while
relevant branching ratios and Higgs and dark matter observables are shown in
its two lower panels. In the context of such a benchmark scenario, the new
physics processes yielding the largest production cross section consist of
the production of a pair of (neutral or charged) higgsinos at the LHC. By a
virtue of their dominant decay modes, current LHC searches are not sensitive to
their associated signature. Neutralino production leads to the further
production
of pseudoscalar Higgs bosons that next decay into pairs of boosted taus, while
the lightest chargino decays into an off-shell $W$-boson and the singlino so
that no LHC bound is expected from the charged higgsino side~\cite{Aad:2014vma}.
Following the procedure developed in Ref.~\cite{Cheng:2013fma,Guo:2013asa}, we
recast all relevant Run--I LHC analyses. We find the CMS search for same-sign
dilepton~\cite{Chatrchyan:2013fea} is the most sensitive to NMSSM benchmark
points such as the one depicted above. However, the relevant production rates
are almost two orders of magnitude lower than the signal cross section that is
excluded at the 95\% confidence level.
In addition, the $H_2 \to A_1 A_1$ branching ratio reaches about 4\%, a result
compatible with Higgs current data. Although this decay mode has been actively
searched for~\cite{Khachatryan:2015nba,Aad:2015oqa,Khachatryan:2015wka} as it
consists of an irrefutable proof of the realization of an NMSSM
scenario~\cite{Ellwanger:2003jt,Ellwanger:2013ova}, we focus instead, in the
next section, on a novel discovery mode for a light NMSSM pseudoscalar based on
boosted ditau tagging.
\section{LHC sensitivity to light NMSSM pseudoscalar Higgs bosons decaying to a
boosted ditau jet}
\label{sec:simu}
\subsection{Event simulation for the NMSSM and boosted ditau tagging}
To determine the LHC sensitivity to the class of NMSSM models introduced in the
previous section, we analyze Monte Carlo simulations of proton-proton collisions
at a center-of-mass energy of 13~TeV as they could occur at the LHC collider at
CERN. Hard-scattering signal and background simulations rely on the
{\sc MadGraph5\_aMC@NLO}
program~\cite{Alwall:2014hca} that contains an NMSSM implementation. The latter,
that has not been described in any earlier publication, is generated by the
{\sc FeynRules} package~\cite{Alloul:2013bka} and its superspace
module~\cite{Duhr:2011se} that automatically produce a UFO
library~\cite{Degrande:2011ua} that can be employed by {\sc MadGraph5\_aMC@NLO}
for event generation, following the strategy of Ref.~\cite{Christensen:2009jx}.
The {\sc FeynRules} model includes a more general version of the superpotential
of Eq.~\eqref{eq:wnmssm} and the soft supersymmetry-breaking Lagrangian of
Eq.~\eqref{eq:lsoft} where the $\mathbb{Z}_3$ symmetry that we have imposed
in this work is not included and where intergenerational sfermion mixings are
allowed. Although such mixings are not handled by {\sc NmssmTools}, they are
compatible with the Supersymmetry Les Houches Accord
conventions~\cite{Skands:2003cj,Allanach:2008qq} in which their implementation
consists of
optional requirements. The translation of the output spectrum files produced by
{\sc NmssmTools} to files compliant with the NMSSM UFO is nevertheless
immediate as this only necessitates to increase the size of $2\times 2$ mixing
matrices to $6\times 6$ matrices with zero entries whenever two different
sfermion generations are concerned. The validation of the NMSSM implementation
of {\sc FeynRules} has been extensively performed during the 2009 Les Houches
workshop on TeV collider physics~\cite{Butterworth:2010ym}, and thousands of
supersymmetric processes have been considered to this aim.
The decays of the produced hard particles and the matching of the
parton-level hard events to a parton shower and hadronization infrastructure
have been performed in the context of the
{\sc Pythia}~6 package~\cite{Sjostrand:2006za},
that is interfaced to {\sc Tauola}~\cite{Jadach:1993hs,Davidson:2010rw}
for handling tau lepton decays. In this framework, any tau lepton polarization
effect is
neglected. Finally, we have simulated the response of the ATLAS detector by
means of the {\sc Delphes 3} program~\cite{deFavereau:2013fsa}, that internally
reconstructs objects on the basis of the anti-$k_T$ jet
algorithm~\cite{Cacciari:2008gp} as implemented in the {\sc FastJet}
software~\cite{Cacciari:2011ma}.
\begin{figure}
\includegraphics[width=0.48\textwidth]{fcent_50}
\includegraphics[width=0.48\textwidth]{fcent_200} \\
\includegraphics[width=0.48\textwidth]{tau21_50}
\includegraphics[width=0.48\textwidth]{tau21_200}
\caption{\label{fig:a1e} $f_{\text{cent}}$ (upper figures) and $\tau_{21}$
(lower figures) distributions for jets originating from the fragmentation of
up quarks (red dashed curves), charmed quarks (blue
dash-dotted curves), bottom quarks (green dotted curves), a single tau lepton
(brown dash-dotted curves) and a pseudoscalar NMSSM Higgs boson (black
plain curves). Two jet energies of 50~GeV (left) and 200~GeV (right) are
considered.}
\end{figure}
Since tau leptons dominantly decay hadronically with a corresponding branching
ratio of about 65\%, the development of efficient related tagging techniques is
a very important task in particular with respect to Higgs precision
measurements~\cite{Aad:2012mea} and new physics searches~\cite{Li:2015sza}. Jets
originating from pure QCD subprocesses and from the hadronic decay of a tau
lepton are mainly distinguished from each other by the number of charged tracks
inside the jet and the jet energy density profile. The properties specific to a
tau jet however turn out
to be preserved in the case of a boosted object comprised of two hadronically
decaying tau leptons that could arise from the decay of a heavier state.
Motivated by such considerations, we have designed an analysis strategy allowing
us to detect the signature of boosted pseudoscalar NMSSM Higgs bosons produced
from the decay of heavier higgsino states at the LHC. This relies on the tagging
of boosted ditau objects via a multivariate method~\cite{Aad:2014rga} that makes
use of the number of tracks inside the ditau jet, the ratio
$f_{\text{cent}}=E^{(0.1)}/E^{(0.2)}$ where $E^{(0.1)}$ ($E^{(0.2)}$) is the
total calorimetric energy deposit in a cone of radius $R=0.1$ (0.2) centered on
the jet direction, the transverse momentum of the hardest track inside a cone of
radius $R=0.2$ centered on the jet direction computed relatively to the jet
$p_T$, the $p_T$-weighted sum of the angular distances of all tracks inside the
jet, the maximum angular distance in the transverse plane between any track
lying inside a cone of radius $R=0.2$ centered on the jet direction and the jet
direction, the track-based jet invariant mass and the ratio of the jet
transverse momentum to the jet invariant mass.
This strategy can furthermore be improved to
gain sensitivity to leptonically decaying taus (within a boosted ditau
object) as well. The produced leptons are indeed unlikely to be isolated,
so that they could be captured by a selection involving the ratio of the
electromagnetic to hadronic calorimetric deposits. Additionally, we also
consider in our multivariate tagging technique the $N$-subjettiness variable
$\tau_{21} = \tau_2/\tau_1$ that allows one order to resolve the substructure of
the ditau jet~\cite{Kim:2010uj,Thaler:2010tr}, with $\tau_N$ being defined by
\begin{equation}
\tau_N =\frac{\sum_k \min \{ \Delta R_{1,k}, \Delta R_{2,k}, \dots, \Delta R_{N,k} \}}{\sum_k p_{T,k} R_0}.
\end{equation}
In this expression, the summations have to be considered upon all jet
constituents, $R_0=0.4$ is the jet cone size parameter in the original jet clustering algorithm and $\Delta R_{I,k}$ denotes the distance in the
transverse plane between the subjet candidate $I$ and the jet constituent $k$.
\begin{figure}
\includegraphics[width=0.48\textwidth]{eff50}
\includegraphics[width=0.48\textwidth]{eff200}
\caption{\label{fig:bdta1} Dependence of the quark and tau lepton rejection
powers on the boosted ditau object tagging efficiency for jets of 50~GeV
(left) and 200~GeV (right).}
\end{figure}
All the variables that we have introduced to tag a ditau boosted object strongly
depend on the object energy, as it is illustrated on Figure~\ref{fig:a1e} for
two representative jet energies of 50~GeV and 200~GeV and the $f_{\text{cent}}$
and $\tau_{21}$ variables. We compare the spectra that would be obtained when
jets solely originate from up quarks (red dashed curves), charmed quarks
(blue dash-dotted curves), bottom quarks (green dotted curves), a single tau
lepton (brown dash-dotted curves) and a pseudoscalar NMSSM Higgs boson
(black plain curves). The properties of ditau boosted objects are
different from the single tau jet and the purely QCD jet cases, so that there
exist handles for discriminating them. To this aim, we use a
boosted decision tree (BDT) technique that uses all the variables presented in
this section. The BDT is trained in the context of jets with specific energies
of 50~GeV and 200~GeV and the correlations between the obtained boosted
ditau object tagging efficiency and the QCD jet or single tau jet mistagging
rates are shown in Figure~\ref{fig:bdta1}. Jets issued from the fragmentation of
light quarks are always harder to distinguish from ditau boosted objects, as
their properties are similar to the ditau case (see Figure~\ref{fig:a1e}).
The corresponding
rejection power is particularly small when the jet energy is smaller. Taking
as a benchmark a tagging efficiency $\epsilon_S=50\%$, a rejection power
$\epsilon_B^{-1}$ of only 50 is found for 50~GeV jets, this number increasing to
500 for 200~GeV jets.
\subsection{LHC sensitivity to NMSSM light higgsinos decaying to
boosted ditau objects}
In order to estimate the sensitivity of the LHC to the class of NMSSM scenarios
under consideration, we study the associated production of a
neutralino and a chargino, followed by a neutralino decay into a pseudoscalar
Higgs boson and a chargino decay into a far off-shell $W$-boson,
\begin{equation}
p p \to \tilde{\chi}^\pm_1 \tilde{\chi}^0_j \to (W^* \tilde{\chi}^0_1)
\ (A_1 \tilde{\chi}^0_1) \ .
\end{equation}
This process features a large cross section, as this was already the case in the
MSSM~\cite{Debove:2008nr,Debove:2009ia,Debove:2011xj,Debove:2010kf,Fuks:2012qx,%
Fuks:2013vua}, and we further impose the off-shell $W$-boson to decay
leptonically so that fully hadronic backgrounds can be suppressed. In our
simulation, we use next-to-leading signal cross sections that are derived with
{\sc Prospino}~\cite{Beenakker:1999xh} and that mostly agree with the most
precise results involving soft-gluon resummation~\cite{Debove:2010kf,%
Fuks:2012qx,Fuks:2013vua}. The dominant associated sources of background
consist of events issued from the production of a (leptonically decaying)
$W$-boson in association with jets, of top-antitop systems and of diboson
systems that all give rise to
final states comprised of a single lepton, missing energy and hard jets. We
generate events exhibiting at least one jet at the
matrix element level and in which the hard-scattering lepton and jets have a
transverse momentum $p_T>10$~GeV and 20~GeV respectively, together with a
pseudorapidity satisfying
$|\eta|<2.5$. We normalize the top-antitop and diboson event samples to the
measured~\cite{xtt} and next-to-leading order~\cite{Campbell:2011bn} cross
section values respectively, and we make use of the leading-order $W$-boson plus
jet fiducial cross section as returned by {\sc MadGraph5\_aMC@NLO} to which a
next-to-leading order $K$-factor of 1.4 is included~\cite{Alwall:2014hca}.
\begin{figure}
\includegraphics[width=0.48\textwidth]{ea1_bench}
\includegraphics[width=0.48\textwidth]{drtata_bench}
\caption{\label{fig:ea} Energy spectrum of the pseudoscalar Higgs bosons $A_1$
issued from the decay of neutralino states (left) and the angular distance in
the transverse plane between the two taus originating from the $A_1$ decays
(right) in the context of signal events. Both distributions have been
calculated for the NMSSM benchmark scenario introduced in
Section~\ref{sec:benchmark} and are normalized to 1.}
\end{figure}
\begin{table}
\center
\setlength{\tabcolsep}{2.5mm}
\renewcommand{\arraystretch}{1.4}
\begin{tabular}{c||c|ccc}
& $\tilde{\chi}^\pm \tilde{\chi}^0$ signal & $W$ plus jets &
Top pair production & Diboson production \\ \hline\hline
$\sigma^{13}$ & 3.38~pb & 8452~pb & 825~pb & 159.3~pb \\ \hline
$\sigma^{\text{pre}}$ & 0.42~pb & 4313~pb & 62.9~pb & 29.2~pb \\
\end{tabular}
\caption{Signal and background cross section before ($\sigma^{13}$) and after
($\sigma^{\text{pre}}$) event preselection at 13 TeV LHC. \label{tab:pre}}
\end{table}
Since signal events feature a final state comprised of a single lepton, a
boosted ditau object issued from the decay of a pseudoscalar $A_1$ particle and
missing energy, we preselect events by requiring that the final state contains
exactly one isolated lepton, at least one jet and we veto the presence of
$b$-tagged jets.
The resulting signal and background cross sections before ($\sigma^{13}$)
and after ($\sigma^{\text{pre}}$) the preselection at 13 TeV LHC are given in
Table~\ref{tab:pre}. At this stage of our analysis, background rates are of
about three orders of
magnitude larger than typical signal cross sections. In the NMSSM scenarios
under consideration, the spectrum generally features a mass splitting $\Delta M=
\mu_{\rm eff}-m_{\tilde\chi_1^0}$ between the lightest singlino-like neutralino
and the heavier neutral and charged higgsino states of at most 100~GeV (see the
lower right panel of Figure~\ref{fig:scan}), so that signal events generally
exhibit final state systems for which the lepton and jet transverse momenta and
the missing transverse energy are of about 50~GeV. It is consequently not
straightforward to design appropriate selections to enhance the signal over
background ratio by only means of the kinematical properties of the signal. The
mass splitting $\Delta M$ also determines the energy $E(A_1)$ that is typically
carried by the pseudoscalar Higgs bosons $A_1$ originating from the decays of
the higgsino-like neutralinos. Since $E(A_1)$ is in general much larger than
the $A_1$ mass, the pseudoscalar decay products (two tau leptons) turn out to
be highly collimated. This is illustrated, for the benchmark scenario introduced
in Section~\ref{sec:benchmark}, on Figure~\ref{fig:ea} where we present the
pseudoscalar Higgs boson $A_1$ energy spectrum (left panel) and the distribution
of the angular distance, in the transverse plane, between the two tau leptons
issued from the $A_1$ decays (right panel). %
\begin{figure}
\includegraphics[width=0.48\textwidth]{nj_bench}
\includegraphics[width=0.48\textwidth]{a1id_bench}
\caption{\label{fig:nj} The distribution in the number of jets (left) and the
ranking of the boosted ditau jet associated with the $A_1$ particle (right)
in the case of signal events and in the context of the benchmark scenario
introduced in Section~\ref{sec:benchmark}. Both distributions are
normalized to 1.}
\end{figure} %
In Figure~\ref{fig:nj}, we study the
details of the hadronic activity in the signal events. In the left panel of the
figure, we present the distribution of the number of jets $N_j$ characterizing
the signal events. Although only one jet (the boosted ditau object)
is expected from the partonic process, initial state radiation allows the $N_j$
spectrum to extent to larger values. The bulk of the events however features at
most two jets, while the leading jet is in general the boosted ditau object (in
70\% of the cases for the benchmark scenario of Section~\ref{sec:benchmark}).
This last property is depicted on the right panel of Figure~\ref{fig:nj}
where we present, on the basis of the Monte Carlo truth, the ranking of the
$A_1$ jet once the $p_T$ ordering of the jets is imposed.
As a consequence of these considerations, we further select signal events on the
basis of a multivariate analysis that uses as inputs the number of jets $N_j$,
the transverse momenta of the two leading jets $p_T^{j_1}$ and $p_T^{j_2}$, the
transverse
momentum of the lepton $p_T^\ell$, the invariant mass of the leading jet, the
amount of missing energy $\slashed{E}_T$, the angular distance in the azimuthal
direction with respect to the beam between the lepton and the missing momentum
$\Delta \phi_{\ell,\slashed{E}_T}$, and the angular distance in the transverse
plane between the lepton and the leading jet $\Delta R(\ell,j_1)$. In addition,
we also include in our analysis the reconstructed $W$-boson transverse mass
$m_T^W$ that would be obtained when considering that all the missing transverse
energy is connected to a $W$-boson decay,
\begin{equation}
M_T^W = \sqrt{2 p_T^\ell \slashed{E}_T \Big[1-\cos \Delta \phi_{\ell,\slashed{E}_T} \Big]} \ ,
\end{equation}
and the final state stranverse mass $m_{T2}$~\cite{Lester:1999tx,Cheng:2008hk}.
Both these latter variables are expected to provide a handle to efficiently
suppress the dominant $W$ background. Moreover, assuming that the boosted
ditau object is identified with the leading jet, the next-to-leading jet is
expected to be softer in the signal case than in the background case,
the lepton and missing energy tend to be not correlated in the signal case as
there are several sources of missing energy, and both the $M_T^W$ and $M_{T2}$
variables are distributed towards smaller values for the signal as the
considered signature is free from any on-shell $W$ boson. The analysis can
finally be improved after imposing that the leading jet is a boosted ditau
object.
\begin{figure}
\includegraphics[width=0.48\textwidth]{bdt}
\includegraphics[width=0.48\textwidth]{sig}
\caption{\label{fig:sig} Left: relations between the signal selection
efficiency and background rejection rate obtained by means of our
multivariate
analysis technique. The effect of the tagging of the leading jet as a boosted
ditau object is either included (dark green) or not (orange). Right: signal
significance obtained with our analysis as a function of the luminosity and
for different treatments of the systematic uncertainties on the background.}
\end{figure}
Applying a selection on the BDT response, we can derive the dependence of the
background rejection rate $\epsilon^{-1}_B$ on the signal selection efficiency
$\epsilon_S$, as shown in the left panel of Figure~\ref{fig:sig}. We
consider two different setups in which the boosted ditau object feature of the
signal is accounted for (dark green) or not (orange), and we observe
that an efficiency of about one percent can be obtained together with a background
rejection rate of $10^4$ ($\gtrsim 10^5$) when the ditau tagging is ignored
(included). Optimizing the selection on the BDT output, the resulting signal fiducial
cross section is of 3.1~fb, while the dominant $W$-boson plus jet and diboson
background component ones are of 16.4~fb and 0.9~fb respectively. Moreover, top-antitop
events turn to contribute negligibly, with a fiducial cross section smaller than
0.5~fb. We calculate the associated signal
significance $\sigma_{\rm signal}$ as a function of the luminosity for different
treatment of the systematic uncertainties on the background $\Delta B$,
\begin{equation}
\sigma_{\rm signal} = \frac{S}{\sqrt{B+(\Delta B)^2}} \ ,
\end{equation}
and show the results on the right panel of the figure. In this expression, $S$ and $B$ denote the
number of selected signal and background events respectively. Assuming a
systematic uncertainty on the background at the percent level, a $3\sigma$ hint for the
class of NMSSM scenarios considered in this work could be observed at the
early stage of the LHC Run--II, while a $5\sigma$ discovery could be expected
with an integrated luminosity of at least 50~fb$^{-1}$.
\section{Conclusion}
\label{sec:conl}
It has been recently shown that the discovery of a light pseudoscalar state
could be a direct evidence for the next-to-minimal realization of supersymmetry
in nature. As a result, many studies have been dedicated to the
investigation of the discovery potential associated with such particles. Most
existing works focus on heavier Higgs decay probes, although light pseudoscalar
states could also be copiously produced from neutralino decays.
In this work, we have explored the NMSSM parameter space and impose
Higgs and dark matter constraints on the construction of viable scenarios. We
have found that many of such scenarios include pseudoscalar Higgs bosons with a
mass comprised in the $[2 m_\tau, 2 m_b]$ window. In this case, the
golden discovery mode consists of the production of an associated
chargino-neutralino pair that further decays into a boosted ditau jet
(connected to a pseudoscalar decay), a single lepton and missing energy. We
study a typical reference scenario an investigate the sensitivity of the LHC
Run--II to the corresponding signal. By means of a multivariate analysis and a
boosted ditau object tagging method, we have found that the background could be
rejected at a very large level so that a $3\sigma$ hint for the signal is
expected within the first 13 TeV data, and that a $5\sigma$ discovery could be
envisaged for a luminosity of more than about 50~fb$^{-1}$.
\section*{Acknowledgements}
This work has ben supported in parts by the {\it Th\'eorie LHC France}
initiative of the CNRS and a joint CNRS-CAS doctoral grant. The work of JL and AGW is supported by the Australian Research Council through the Centre of Excellence for Particle Physics at the Terascale CE110001004.
\bibliographystyle{JHEP}
|
\section{INTRODUCTION}
Recently the Weyl semimetals (WSMs) \cite{Wan2011} attract enormous attention as a new family of topological materials,\cite{qi2011RMP,Hasan:2010ku,Yan2012rpp} which was stimulated by the recent discovery in transition-metal monopnictides.\cite{Weng2015,Huang2015,Xu2015TaAs,Lv2015TaAs,Yang2015TaAs} The WSM exhibits topological surface states \cite{Wan2011} that are characterized by Fermi arcs and exotic topological chiral transport properties such as the chiral anomaly effect \cite{Adler1969,Bell1969} and a large magnetoresistance (MR).\cite{Shekhar2015}
While angle-resolved photoemission spectroscopy (ARPES)\cite{Lv2015TaAsbulk,Liu2015NbPTaP,Xu2015NbAs,Xu2015TaP,Souma2015} and theoretical calculations\cite{Sun2015arc,Soluyanov2015,Sun2015MoTe2} have been further employed to address the surface states,
magneto-transport measurements have been also extensively performed recently.\cite{Huang2015anomaly,Zhang2015ABJ,Shekhar2015TaP,Wang:2015wm,Yang:2015vz,Du:2015TaP, Shekhar2015,Moll2015,Zhang2015quantum,Yang2011QHE,Turner:2013tf,Hosur:2013eb,Vafek:2014hl,Parameswaran2014,Baum2015}
In the band structure of a WSM, conduction and valence bands cross each other at nodal points, called Weyl points, and disperse linearly through these Weyl points in the three-dimensional (3D) $k$-space.\cite{Turner:2013tf,Hosur:2013eb} Thus, a WSM can be viewed as a 3D analogue of graphene. The Weyl points appear in pairs, in which the two Weyl points exhibit opposite chiralities. The transition-metal monopnictide WSMs include four members, TaAs, TaP, NbAs, and NbP. These compounds crystalize in a non-centrosymmetric tetragonal lattice structure, but preserve the time-reversal symmetry. There are twelve-pairs of Weyl points that can be classified into two groups: four pairs lie in the $k_z = 0$ plane (labeled as W1) and eight pairs stay in planes of $k_z \sim \pm \pi/c$ (labeled as W2), where $c$ is the lattice constant of the long axis. The 4 compounds exhibit very similar band structures, in which the W1-type Weyl points are slightly lower in energy than the W2-type Weyl points. However, all Weyl points stay slightly away from the Fermi energy with the coexistence of many trivial electron and hole pockets in the Fermi surface (FS), as revealed by recent FS reconstructions in TaP \cite{Shekhar2015TaP} and NbP,\cite{Klotz2015} for example.
The member with the weakest spin-orbit coupling in above four WSM materials, NbP, was found to present an extremely large MR and an ultrahigh mobility.\cite{Shekhar2015} The large MR originates form the combination of the nearly perfect compensation condition \cite{Ali2014} between electron and hole carriers \cite{Shekhar2015,Klotz2015} and the high mobility which is relevant to the topological band structure.
Weyl electrons will generally coexist with normal electrons, and then small changes of the Fermi energy can in principle modify the topology of the Fermi surface due to the smallness of the carrier density. Applying pressure is known to be a powerful approach to tune the electronic structure of a material, which promises a way to vary the energies of Weyl points. Hence, it is interesting to investigate how the positions of the two sets of Weyl points in NbP shift in energy under pressure, and whether pressure can induce other exotic structures and properties in this topological WSM. Since surface-sensitive probes such as ARPES cannot be used for detecting the topological states under pressure, quantum oscillation studies become an ideal tool to determine the effects of pressure on the Fermi-surface topology here. So far, there have been only a few studies on the effect of pressure on NbAs \cite{Luo2016NbAs,Zhang2015NbAs} and TaAs.\cite{Zhou2015TaAs} Although, in all these studies the electronic structure is shown to be very stable, an accurate determination of the effects of pressure on the Fermi-surface topology of Weyl materials is still missing. Shubnikov-de Haas (SdH) oscillations have been reported, but without direct comparison with band-structure calculations.\cite{Luo2016NbAs}
In this paper, we present a pressure study of the magnetotransport properties of the WSM NbP.
By analyzing SdH oscillations observed in the MR, 4 Fermi-surface pockets are revealed, which correspond to two pairs of electron and hole pockets.\cite{Klotz2015} Although we observe a drastic effect on the amplitudes of the quantum oscillations, the frequencies remain almost unaltered up to 2.8~GPa. With support of band-structure calculations we argue that the observed pressure-induced changes in the quantum oscillations arise from small modifications in the shape of the FS. We further relate the strong variation of the high-field MR for magnetic field $B$ and electrical current $I$ parallel to the crystallographic $a$ direction to a combination of different effects, a change in the balance of electron and hole charge carriers and an alteration of the distance of the W2-type Weyl points to the Fermi level which is evidenced by our band structure calculations. Moreover, our findings demonstrate the stability of the electronic band structure of NbP, and show that quantum oscillation studies in association with band-structure calculations are an effective tool to investigate the evolution of the FS topology in Weyl semimetals under pressure.
\section{METHODS}
High-quality single crystals of NbP were grown via a chemical vapor transport reaction. More details on the sample preparation and characterization can be found in Ref.\ \onlinecite{Shekhar2015}. The electrical-transport experiments were performed on high quality NbP single crystals in magnetic fields up to $B=9$~T in a $^{4}$He cryostat (JANIS) equipped with superconducting magnet to temperatures down to $T=1.4$~K. The electrical resistance was measured in 4-point geometry, where the contacts to the sample were made using silver paint and $25~\mu$m gold wire. We prepared two samples taking crystals from the same batch. In both setups $I$ was applied along the crystallographic $a$-axis and the magnetic field perpendicular to the current in sample $S_{B\parallel c}$ and parallel to it in sample $S_{B\parallel a}$. Hydrostatic pressure was generated using a clamp-type pressure cell utilizing silicon oil as pressure transmitting medium. The pressure inside the cell was determined by measuring the shift of the superconducting critical temperature of a piece of Pb.
The \textit{ab-initio} calculations were performed within the framework of the density-functional theory (DFT), implemented in the Vienna \textit{ab-initio} simulation package.\cite{kresse1996} The core electrons were represented by the projector-augmented-wave potential and the generalized gradient approximation (GGA) \cite{perdew1996} was employed for the exchange-correlation functional. A $k$-grid in the first Brillouin zone with Gaussian smearing ($\sigma=0.05$~eV) was utilized. Instead of using experimental lattice parameters, the Murnaghan's equation of state was used to derive the pressure by fitting the total energy dependence of the volume.\cite{Murnaghan1944} The total energy was extracted by optimizing lattice constants and atomic positions for each volume. Then we interpolated the bulk FS using maximally localized Wannier functions.\cite{Mostofi2008}
\section{RESULTS}
The temperature dependence of the longitudinal resistivity $\rho_{xx}(T)$ is depicted in Fig.~\ref{fig1}. At ambient pressure we find a good agreement with previously reported data.\cite{Shekhar2015} Both samples display a metallic behavior with a resistivity ratio $\rho_{xx}({\rm 300\,K})/\rho_{xx}({\rm 2\,K})\approx112$ indicating the high quality of our single crystals. The resistivity ratio is not affected by application of external pressure. The curves obtained at all investigated pressures collapse onto a single line within the experimental resolution. We note that we do not find any indication for superconductivity down to 1.4~K in the whole pressure range $p\leq 2.8$~GPa. At this point, we want to put out that our data are unaltered by pressure even at low temperatures. This is in contrast to observations reported for NbAs,\cite{Luo2016NbAs} where at low temperatures ($T\lesssim30$~K) a small, but monotonous increase in the resistivity upon increasing pressure is eminent in the data. This hints at a more stable electronic structure in NbP compared with NbAs.
\begin{figure}[t!]
\begin{center}
\includegraphics[clip,width=0.95\columnwidth]{Figure1b.eps}
\caption{Temperature dependent of the normalized electrical resistivity $\rho/\rho_{\rm295K}$ in zero magnetic field.
}\label{fig1}
\end{center}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[clip,width=0.95\columnwidth]{Figure2b.eps}
\caption{(a) and (b) Magnetoresistance at 2~K for selected pressures in the two configurations $B\parallel c$ and $B\parallel a$, respectively.
}\label{fig2}
\end{center}
\end{figure}
In the following we will focus on the MR in NbP. We performed MR measurements under magnetic field up to 9~T and to pressure up to 2.8~GPa. Figures~\ref{fig2}a and~\ref{fig2}b show the MR of samples $S_{B\parallel c}$ and $S_{B\parallel a}$ at 2~K for selected pressures, respectively. Here, we define the MR as the magnetic field dependence of the longitudinal resistivity $\rho_{xx}$: ${\rm MR}(B)=[\rho_{xx}(B)-\rho_{xx}(0)]/{\rho_{xx}(0)}$. Owing to the large charge carrier mobility in NbP,\cite{Shekhar2015} we observe a giant unsaturated MR for both samples. However, the behavior of the MR under pressure is different between the two samples. For sample $S_{B\parallel c}$, the magnitude of the MR remains almost unaltered under pressure, except for oscillatory part, which displays characteristic changes which we will discuss below. On the other hand, for the sample $S_{B\parallel a}$ (see Fig.~\ref{fig2}b) pressure induces dramatic changes in the amplitude of the MR. Furthermore, the MR curves become less symmetric upon increasing pressure. The asymmetry of MR curves originates from a Hall contribution due to a small misalignment of the voltage leads in our experimental setup. An extraordinarily large Hall effect has been previously reported for NbP.\cite{Shekhar2015} The Hall contribution can be removed from the data by a symmetrization of MR curves, ${\rm MR}_{\rm symm}(B)=0.5[{\rm MR}(B)+{\rm MR}(-B)]$. The ${\rm MR}_{\rm symm}$ curves for $S_{B\parallel a}$ are displayed in Fig.~\ref{fig3}a. The pressure dependence of ${\rm MR}_{\rm symm}$ at 9~T (see inset of Fig.~\ref{fig3}a) exhibits a double well structure: upon increasing pressure ${\rm MR}_{\rm symm}({\rm 9T})$ decreases from $8\times10^{4}\%$ at ambient pressure to $2.3\times10^{4}\%$ at 1~GPa, reaches $3.2\times10^{4}\%$ at 1.5~GPa and decreases again to $2\times10^{4}\%$ at 2~GPa. Above 2~GPa ${\rm MR}_{\rm symm}({\rm 9T})$ continues to increase and achieves $10\times10^{4}\%$ at 2.8~GPa which is comparable with the value at 0~GPa. ${\rm MR}_{\rm symm}$ in NbP is directly related to the almost perfect balance between electron- and hole-like charge carriers. The concentration of the charge carriers can be inferred from the Hall coefficient. From our data we can only obtain qualitative information on the Hall effect from the antisymmetric contribution to the MR, ${R_H}(B)\propto0.5[{\rm MR}(B)-{\rm MR}(-B)]$. ${R_H}(B)$ changes sign around 1.5~GPa (see Fig.~\ref{fig3}b). This indicates that the balance between electrons and holes is very sensitive to the application of external pressure. A similar effect has been reported in WTe$_{2}$, which has been identified as a Dirac material.\cite{Cai, Kang2015} In Weyl semimetals the presence of the chiral effect is expected to provide a negative contribution to the longitudinal MR, \textit{i.e.}, when electrical current and magnetic field are parallel.\cite{Adler1969,Bell1969,Nielsen1983} Therefore, a negative contribution to the MR might be expected for sample $S_{B\parallel a}$. However, this is only the case, if the chirality is well-defined, \textit{i.e.}, the Fermi energy is close enough to the Weyl nodes.\cite{Nielsen1983} For NbP at ambient pressure, the two groups of Weyl nodes, W1 and W2, lie -57~meV and 8~meV apart from the Fermi energy, respectively. Thus, a contribution from the chiral anomaly to the MR is unlikely.\cite{Klotz2015} On the other hand, pressure might tune the energy position of the Weyl nodes closer to the Fermi level leading to an increase of the negative contribution from the chiral anomaly to the longitudinal MR. Thus, the pressure variation of the longitudinal MR as observed $S_{B\parallel a}$ could be a combination of different effects.
\begin{figure}[t!]
\begin{center}
\includegraphics[clip,width=0.95\columnwidth]{Figure3b.eps}
\caption{ (a) ${\rm MR_{symm}}$ for $S_{B\parallel a}$ with $B\parallel I$. The inset shows the pressure effects on ${\rm MR_{symm}}$ at 9~T for $S_{B\parallel a}$. (b) Anti-symmetrized MR of sample $S_{B\parallel c}$, which is a measure of the Hall coefficient $R_H$ (see text for details).
}\label{fig3}
\end{center}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[clip,width=0.9\columnwidth]{Figure4b.eps}
\caption{(a) Shubnikov-de Haas oscillations as a function of the inverse magnetic field taken at different pressures for $S_{B\parallel c}$. The oscillatory part was extracted by a subtraction of a third order polynomial background. (b) and (c) temperature dependence of the fast Fourier transform of the SdH oscillations under pressures of 0.02 and 2.47~GPa for $B\parallel c$.}\label{figure4}
\end{center}
\end{figure}
\begin{figure}[t!]
\begin{center}
\includegraphics[clip,width=0.9\columnwidth]{Figure5b.eps}
\caption{(a) Oscillatory part of MR as a function of the inverse magnetic field taken at different pressures for the sample $S_{B\parallel a}$. (b) and (c) Fast Fourier transform of the SdH oscillations as a function of temperature under pressures of 0.02 and 2.47~GPa for $B\parallel a$.}\label{figure5}
\end{center}
\end{figure}
The pressure evolution of the Fermi-surface topology can be extracted from the pronounced SdH oscillations in the MR. For both field orientations, $B\parallel c$ and $B\parallel a$, see Figs.~\ref{figure4}a and \ref{figure5}a, respectively, SdH oscillation starting from 1~T are clearly visible, indicating very small effective masses, resulting in high mobilities. The oscillatory part of the signal was obtained by subtracting a third-order polynomial background from the MR data. As expected we find $\Delta\rho_{xx}$ periodic in $1/B$. The SdH frequencies were then determined by a fast Fourier transform on the oscillatory part of the signal.
The 8 fundamental frequencies obtained at ambient pressure, labeled $F1$ to $F8$, are given in Table~I and are in good agreement with literature \cite{Klotz2015}. The corresponding calculated FS pockets including extremal orbits are illustrated in Fig.~\ref{figure6}c and d. In fact, the theory predicts some additional extremal orbits in agreement with our previous work,\cite{Klotz2015} but those could not be resolved in the present experiment. The adopted Fermi level for calculating the extremal orbits is at the ideal electron-hole compensation point.
The frequencies of the oscillations are proportional to extremal Fermi-surface cross-sections $A_{k}$. $A_{k}$ is perpendicular to $B$ following the Onsager relation $F=(\Phi_{0}/2\pi^{2})A_{k}$, where $\Phi_{0}=h/2e$ is the magnetic flux quantum. The temperature dependence of the spectra are displayed in Figs.~\ref{figure4}b and \ref{figure4}c and in Figs.~\ref{figure5}b and \ref{figure5}c at 0 and 2.47~GPa for $S_{B\parallel c}$ and $S_{B\parallel a}$, respectively. We can identify for both cases 4 fundamental frequencies that remain almost unchanged with pressure. The fundamental frequencies in Figs.~\ref{figure4} and ~\ref{figure5} are indicated by bullets. The width of the bullets is an estimation of the error in the determination of the frequencies. We point out that beside of the 8 fundamental frequencies, a series of their higher harmonics is also present, even more pronounced in the spectra at higher pressures, manifesting the high sample quality. While the frequencies are almost not affected, application of external pressure dramatically reduces the amplitudes of the quantum-oscillations. The amplitudes of the oscillations are directly related to the curvature $|\partial^{2}A/\partial k_{\parallel}^{2}|$ of the FS cross sections. Therefore, we can argue that the reduction of the amplitude of the quantum oscillations is caused by an increase in the curvature close to an extremal cross section of the FS.
The effective charge-carrier masses $m^*$ of the different orbits are obtained by analyzing the temperature dependencies of the SdH amplitudes. The data for 0 and 2.47~GPa are displayed in Figs.~\ref{figure4} and ~\ref{figure5}. According the Lifshitz-Kosevich formula, the temperature dependence is proportional to $X/\sinh(X)$, with $X=\alpha m^*T/B$ and $\alpha=2\pi^{2}\kappa_{B}m_{e}/(\hbar e)$. Here, $\kappa_{B}$ and $m_{e}$ are the Boltzmann constant and the free-electron mass, respectively. For $B\parallel c$ the masses are rather small, between $m^{*}\approx0.034m_{e}$ and $0.047m_{e}$ at ambient pressure, as expected for semimetals with small FS pockets. In the case of $B\parallel a$ the masses are about twice as large. We find $m^{*}\approx0.104m_{e}$ and $0.112m_{e}$ for the $F7$ and $F8$ frequencies. These results are in good agreement with previous studies.\cite{Klotz2015} The large difference between the effective masses for the two directions is expected due to the highly anisotropic FS pockets in NbP, since in a first approximation $m^{*}$ is proportional to the extremal area of an orbit. Within the experimental resolution, we find almost constant effective masses in the investigated pressure range up to 2.8~GPa. This result further indicates the stability of the electronic structure of NbP. The obtained effective masses at 0, 1.04, and 2.47~GPa are summarized in Table~I.
\begin{table}[t!]
\begin{centering}
\caption{Experimental values of the dHvA frequencies and effective masses for 0, 1.04, and 2.47~GPa for different orbits. For $F5$ and $F6$ the effective masses could not be determined reliably. The frequency $F$ is in unit of T, and the effective masse $m^{*}$ is in units of free electron masses, $m_{e}$.}
\end{centering}
\label{freq_mass}
\centering{
\begin{tabular}{ccclccccccccc}
&&&&&&&&\tabularnewline
\hline
& & \multicolumn{4}{c}{\textbf{0.02 GPa}} & \multicolumn{3}{c}{\textbf{1.04 GPa}} & & \multicolumn{3}{c}{\textbf{2.47 GPa}}\tabularnewline
& & & $F$ & $m^{*}$ & & & $F$ & $m^{*}$ & & & $F$ & $m^{*}$\tabularnewline
\hline
\hline
& & \multicolumn{4}{c}{$$} & \multicolumn{3}{c}{$B \parallel c$} & & & \multicolumn{2}{c}{$$}\tabularnewline
$F1$ & $\alpha_2$ & & ~~9.2 & 0.047(5) & & & 10.5 & 0.051(5) & & & ~~8.1 & 0.049(6)\tabularnewline
$F2$ & $\beta$ & & 14.8 & 0.040(6) & & & 16.5 & 0.046(5) & & & 13.6 & 0.034(5)\tabularnewline
$F3$ & $H2$ & & 21.8 & 0.034(8) & & & 19.1 & 0.038(4) & & & 18.9 & 0.044(6)\tabularnewline
$F4$ & & & 33.0 & 0.041(7) & & & 34.8 & 0.040(5) & & & 34.0 & 0.048(4)\tabularnewline
\hline
& & \multicolumn{4}{c}{$$} & \multicolumn{3}{c}{$B \parallel a$} & & & \multicolumn{2}{c}{$$}\tabularnewline
$F5$ & & & 45.1 & - & & & 48.6 & - & & & 45.35 & -\tabularnewline
$F6$ & & & 71.2 & - & & & 68.6 & -& & & 72.1 & -\tabularnewline
$F7$ & & & 128 & 0.112 (10) & & & 127.5 & 0.106(8)~~ & & & 127.5 & 0.104(8)~~~\tabularnewline
$F8$ & & & 146 & 0.104 (12) & & & 140.5 & 0.099(10) & & & 140.5 & 0.094(13)\tabularnewline
\hline
\end{tabular}
\end{table}
\section{DISCUSSION}
\begin{figure*}[t!]
\begin{center}
\includegraphics[clip,width=2\columnwidth]{Figure6.eps}
\caption{(a) and (b) Pressure dependence of the experimental SdH (solid symbols) and calculated quantum-oscillation frequencies (dashed lines) for $B\parallel c$ and $B\parallel a$, respectively. The size of the bullets gives an estimate of the error in the determination of the oscillation frequencies as shown in Figs.~\ref{figure4} and ~\ref{figure5}. (c) and (d) Illustration of the calculated Fermi-surface pockets [left (blue) - electron and right (violet) - hole pockets]. The extremal orbits identified in our experimental data, $\alpha$, $\alpha_2$, $\alpha_2'$, $\beta$, $\gamma$, $E1'$, $H1'$, $H2$, $H2'$ and $H2''$ are indicated. The gray dash lines mark those extremal orbits that are not confirmed experimentally. Band structures along the line connecting a pair of Weyl points for (e) W2 and (f) W1 at different pressures. The dash line marks the Fermi energy set to zero. }\label{figure6}
\end{center}
\end{figure*}
In order to get deeper insights in the pressure evolution of the FS topology, we mapped the experimentally obtained frequencies on our calculated FS. The experimental frequencies were taken from the peaks in the spectra obtained from the fast Fourier transforms of the data at 2~K (see Figs.~\ref{figure4},~\ref{figure5} and Table~I). As depicted in the Fig.~\ref{figure6}a and b, we find a good agreement between the pressure dependencies as well as absolute values of the experimental and calculated frequencies. Our calculated FS orbits indicate that the experimental frequencies $F1$, $F2$, $F4$, and $F5$ are related to electron pockets while $F3$ and $F8$ are orbits from hole pockets. Some of the calculated FS orbits are close to one branch of experimental frequencies, such frequencies are experimentally hardly separable. Most likely the $F7$ and $F8$ frequencies are a mixture of $H1'$, $E1'$ and $H2''$ orbits. Sometimes, even though, we know the frequencies belong to one pocket, there are still two different calculated FS orbits close by. Such as the $F4$ frequency branch is composed by $\alpha$ and $\gamma$ orbits of the large electron pocket.
Independent of the field orientation, most of the calculated frequencies belonging to electron orbits ($\alpha$, $\alpha_2$, $\alpha_2'$, $\beta$, and $E1'$) decrease with increasing pressure, except the $\gamma$ orbit. The opposite behavior is observed for the orbits associated with hole pockets ($H1'$, $H2$, $H2'$, and $H2''$) and the arm part of the electron pocket $\gamma$. As we see in our calculations, the effect of pressure is small, which is hard to unravel in the experimental data. Thus, the calculated frequencies can help us to realize the tendency.
Another information we can obtain from our calculations is the energy difference between the Weyl points and the Fermi energy. We picked three pressures and shifted the Fermi energy of the band structures around W1- and W2-type Weyl points to zero. The valley between Weyl points with opposite chiralities in Fig.~\ref{figure6}e and f is almost unchanged, but both Weyl points get more and more closer to the Fermi energy upon increasing pressure. We note that in Fig.~\ref{figure6}e and f the two Weyl points at 0~GPa are not separate by any barrier, which is different compared with our previous work.\cite{Klotz2015} This is caused by carrying out a full relaxation of the lattice parameters in the present work instead of adopting the experimental ones. Here the lattice parameters obtained 2.65~GPa are similar to that used in our previous work.\cite{Klotz2015} Accordingly, we find a tiny barrier of $\sim1$~meV along the line connecting the pair of W2-type Weyl points at 2.65~GPa. However, the Weyl points are still not independent since they merge into the hole pocket in other directions.\cite{Klotz2015} The energy variations of the W1-type Weyl points are small to almost no change (see Fig.~\ref{figure6}f). In contrast to that the W2-type Weyl points exhibit considerable variations. This implies a possibility to induce the chiral anomaly effect with even higher pressures by shifting the W2-type Weyl points to the Fermi energy.
The pressure independence of the SdH frequencies evidences the robustness of the electronic structure of the WSM NbP in the pressure range up to 2.8~GPa. This observation is different compared with several 3D Dirac semimetals. For example, Cd$_{3}$As$_{2}$ exhibits a pressure-induced breakdown of the 3D Dirac semimetal state with the system becoming a gapped semiconductor at 2.57~GPa.\cite{S_Zhang} For WTe$_{2}$ a drastic change in the Fermi surface topology and a strong suppression of the MR was observed at 1.8~GPa.\cite{Cai} We speculate that the robustness of electronic structure of the WSM NbP may be related to its stable noncentrosymmetric tetragonal structure. This speculation is supported by recent studies on NbAs \cite{Luo2016NbAs,Zhang2015NbAs} and TaAs,\cite{Zhou2015TaAs} two others closely related members of the monopnictide family, with the same crystal structure than NbP. However, for both arsenide compounds, the electrical resistivity displays at low temperatures a monotonic increase upon increasing pressure ($T\lesssim25$~K), in contrast to our finding of a pressure independent resistivity in NbP. This suggests that the electronic structure of NbP is more robust than that of NbAs and TaAs.
\section{SUMMARY}
In conclusion, we presented an investigation of the pressure evolution of the FS topology in the noncentrosymmetric Weyl semimetal NbP by combining experimental studies of SdH oscillations and band-structure calculations. We found a robust electronic structure in the pressure range up to 2.8 GPa. The characteristic topological features of the Fermi-surface with two electron and two hole pockets remain unchanged. The strong change in the SdH amplitudes with pressure can be attributed to subtle changes in the shape of the Fermi-surface in the vicinity of the extremal orbits. Furthermore, our results evidence a strong pressure variation of the magnitude of longitudinal magnetoresistance for $B\parallel a$ which might be a related to a combination of different effects, such as a change in the balance of electron and hole charge carriers and/or an alteration of the distance of the W2-type Weyl points to the Fermi level.
Indeed, the W2-type Weyl points move toward the Fermi-energy upon increasing pressure, while the energy difference between the W1-type Weyl points and the Fermi energy is almost unaffected. We may speculate that the chiral anomaly effect might be established at even higher pressures. The good agreement between the experimental and calculated SdH frequencies confirm that the pressure dependence of the quantum oscillations provide a reliable tool to probe the Fermi-surface topology in Weyl semimetals under pressure.
\section*{ACKNOWLEDGMENTS}
We thank E. Hassinger, F. Arnold, and M. Baenitz for stimulating discussions. This work was financially supported by the ERC Advanced Grant No.\ (291472) ``Idea Heusler". R.\ dos Reis acknowledges financial support from the Brazilian agency CNPq.
|
\section{Introduction}
We are interested in the norm approximation of the Schr\"odinger evolution
\begin{eqnarray} \label{eq:schrodingerdynamics}
\Psi_N(t) = e^{-itH_N}\Psi_{N}(0)
\end{eqnarray}
on the bosonic Hilbert space $\gH^N=\bigotimes_{\text{sym}}^N L^2(\mathbb{R}^3)$. Here
\begin{equation*}
H_N= \sum\limits_{j = 1}^N -\Delta_{x_j} + \frac{1}{N-1} \sum\limits_{1 \le j < k \le N} {w_N(x_j-x_k)}
\end{equation*}
is the Hamiltonian of a system of $N$ identical bosons in $\mathbb{R}^3$. The interaction potential is taken of the delta-type form
\begin{equation*}
w_N(x-y)= N^{3\beta} w(N^\beta (x-y)).
\end{equation*}
The parameter $\beta \ge 0$ is fixed and $w\in C^1_0(\mathbb{R}^3)$ is non-negative and spherically symmetric decreasing.
From the physical point of view, the initial state $\Psi_N(0)$ may be interpreted as a ground state of a trapped system and the time evolution $\Psi_N(t)$ in \eqref{eq:schrodingerdynamics} is observed when the trapping potential is turned off. Thus, motivated by the results on ground states in \cite{LewNamSerSol-15} (see also \cite{Seiringer-11,GreSei-13,DerNap-13,NamSei-15}), we expect that
\begin{eqnarray} \label{eq:PsiN0-intro}
\Psi_N(0) \approx \sum_{n=0}^N u(0)^{\otimes (N-n)} \otimes_s \psi_n(0)
\end{eqnarray}
in norm for $N$ large. Here $u(0)$ is a normalized function in $L^2(\mathbb{R}^3)$ describing the Bose-Einstein condensate and $(\psi_n(0))_{n=0}^\infty$ is a {quasi-free} state describing the fluctuations around the condensate.
We will show that if $\Psi_N(0)$ satisfies \eqref{eq:PsiN0-intro}, then for every $t>0$, we have
\begin{eqnarray} \label{eq:PsiNt-intro}
\Psi_N(t) \approx \sum_{n=0}^N u(t)^{\otimes (N-n)} \otimes_s \psi_n(t)
\end{eqnarray}
in norm for $N$ large. Here $u(t) \in L^2(\mathbb{R}^3)$ is determined by a mean-field (Hartree) equation and $(\psi_n(t))_{n=0}^\infty$ is a quasi-free state governed by a quadratic (Bogoliubov) Hamiltonian on Fock space.
The approximation \eqref{eq:PsiNt-intro} has been first established in \cite{LewNamSch-15} for $\beta=0$, and then extended to $0\le \beta<1/3$ in \cite{NamNap-15}. This range of $\beta$ seems to be optimal if we only assume that $(\psi_n(0))_{n=0}^\infty$ in \eqref{eq:PsiN0-intro} is a quasi-free state. In the present work, we will make an additional assumption (still physically reasonable) that $(\psi_n(0))_{n=0}^\infty$ has finite kinetic energy, and prove \eqref{eq:PsiNt-intro} for all $0\le \beta<1/2$. Note that when $\beta>1/3$, the range of the interaction potential is much smaller than the average distance between the particles, and hence every particle essentially interacts only with itself. This so-called self-interaction regime is physically more relevant and mathematically more challenging than the mean-field regime $\beta <1/3$.
An analogue of \eqref{eq:PsiNt-intro} related to the fluctuations around coherent states in Fock space has been justified in \cite{Hepp-74,GinVel-79, GinVel-79b,GriMacMar-10,GriMacMar-11} for $\beta=0$, in \cite{GriMac-13} for $\beta<1/3$ and in \cite{Kuz-15b} for $\beta<1/2$. In particular, our result is comparable to \cite{Kuz-15b}, but our method is different and it can be used to simplify the proof in \cite{Kuz-15b}. Thanks to a heuristic argument in \cite{Kuz-15b}, we also expect that the range $0\le \beta<1/2$ is optimal for the approximation \eqref{eq:PsiNt-intro} to hold, as soon as $u(t)$ is still decoupled from the equation for $(\psi_n(t))_{n=0}^\infty$.
When $\beta>1/2$, the effective equations for $u(t)$ and $(\psi_n(t))_{n=0}^\infty$ in \eqref{eq:PsiNt-intro} have to be modified to take two-body scattering processes into account. This step has been carried out in the Fock space setting in \cite{BocCenSch-15,GriMac-15}, but it is still open in the $N$-particle setting.
Note that the norm convergence \eqref{eq:PsiNt-intro} is much more precise than the usual convergence of density matrices in the context of the Bose-Einstein condensation. In particular, our result can be interpreted as a second order correction to the leading order result in \cite{ErdSchYau-07}. We refer to \cite{NamNap-15} for a further discussion and an extended list of literature in this direction.
The precise statement of our result is given in the next section.
\medskip
\noindent{\bf Acknowledgment.} The first author thanks Mathieu Lewin for helpful discussions. We thank David Mitrouskas and S\"oren Petrat for finding a gap in a previous version of this paper. We thank the referees for useful comments and remarks. The support of the Austrian Science Fund (FWF) project Nr. P 27533-N27 is gratefully acknowledged.
\section{Main result} \label{sec:main-result}
In our paper, the condensate is governed by the Hartree equation
\begin{equation} \label{eq:Hartree-equation}
\left\{
\begin{aligned}
i\partial_t u(t) &= \big(-\Delta +w_N*|u(t)|^2 -\mu_N(t)\big) u(t), \\
u(t=0)&=u(0).
\end{aligned}
\right.
\end{equation}
Here we can choose the phase
$$
\mu_N(t)=\frac12\iint_{\mathbb{R}^3\times\mathbb{R}^3}|u(t,x)|^2w_N(x-y)|u(t,y)|^2 \,{\rm d} x \,{\rm d} y
$$
to ensure an energy compatibility (see \cite{LewNamSch-15} for further explanations). The well-posedness of the Hartree equation is recalled in Lemma \ref{lem:Hartree-equation}.
To describe the fluctuations around the condensate, it is natural to introduce the Fock space
$$ \mathcal{F}(\mathfrak{H})= \bigoplus_{n=0}^\infty \mathfrak{H}^n= \mathbb{C} \oplus \bigoplus_{n=1}^\infty \bigotimes^n_{\rm sym} \mathfrak{H}, \quad \mathfrak{H}=L^2(\mathbb{R}^3).$$
On this Fock space, we define the creation and annihilation operators $a^*(f)$, $a(f)$, with $f\in \mathfrak{H}$, by
\begin{align*}
(a^* (f) \Psi )(x_1,\dots,x_{n+1})&= \frac{1}{\sqrt{n+1}} \sum_{j=1}^{n+1} f(x_j)\Psi(x_1,\dots,x_{j-1},x_{j+1},\dots, x_{n+1}), \\
(a(f) \Psi )(x_1,\dots,x_{n-1}) &= \sqrt{n} \int \overline{f(x_n)}\Psi(x_1,\dots,x_n) \,{\rm d} x_n, \quad \forall \Psi \in \mathfrak{H}^n,\, \forall n.
\end{align*}
These operators satisfy the canonical commutation relations (CCR)
$$
[a(f),a(g)]=[a^*(f),a^*(g)]=0,\quad [a(f), a^* (g)]= \langle f, g \rangle, \quad \forall f,g \in \mathfrak{H}.
$$
Equivalently, we can define the operator-valued distributions $a_x^*$ and $a_x$, with $x\in\mathbb{R}^3$, by
$$
a^*(f)=\int_{\mathbb{R}^3} f(x) a_x^* \,{\rm d} x, \quad a(f)=\int_{\mathbb{R}^3} \overline{f(x)} a_x \,{\rm d} x, \quad \forall f\in \mathfrak{H}.
$$
They satisfy
$$[a^*_x,a^*_y]=[a_x,a_y]=0, \quad [a_x,a^*_y]=\delta(x-y), \quad \forall x,y\in \mathbb{R}^3.$$
These operators allow us to express operators on Fock space in a convenient way. For example, for every operator $h$ on $L^2(\mathbb{R}^3)$ with kernel $h(x,y)$, we can write
$$
{\ensuremath{\rm d}\Gamma}(h):= 0\oplus \bigoplus_{n=0}^\infty \sum_{j=1}^n h_j = \int_{\mathbb{R}^3} a_x^* h a_x \,{\rm d} x = \iint_{\mathbb{R}^3\times \mathbb{R}^3} h(x,y) a_x^* a_y \,{\rm d} x \,{\rm d} y.
$$
In particular, $\mathcal{N}={\ensuremath{\rm d}\Gamma}(1)$ is called the number operator.
In our approximation \eqref{eq:PsiN0-intro}-\eqref{eq:PsiNt-intro}, the particles outside of the condensate are described by a unit vector $\Phi(t)=(\psi_n(t))_{n=0}^\infty$ in the excited Fock space
$$ \mathcal{F}_+(t)= \bigoplus_{n=0}^\infty \bigotimes^n_{\rm sym} \mathfrak{H}_+(t), \quad \mathfrak{H}_+(t)=\{u(t)\}^\bot=Q(t)\mathfrak{H}, \quad Q(t):=1-|u(t) \rangle \langle u(t)|.$$
This vector is governed by the Bogoliubov equation
\begin{eqnarray} \label{eq:Bogoliubov-equation}
\left\{
\begin{aligned}
i \partial_t \Phi(t) &= \mathbb{H}(t) \Phi(t),\\
\Phi(t=0)&= \Phi(0),
\end{aligned}
\right.
\end{eqnarray}
where
\begin{align*}
&\mathbb{H}(t):= {\ensuremath{\rm d}\Gamma}(h(t)) + \frac12\iint_{\mathbb{R}^3\times\mathbb{R}^3}\Big(K_2(t,x,y)a^*_x a^*_y +\overline{K_2(t,x,y)}a_x a_y\Big)\,{\rm d} x\,\,{\rm d} y, \\
& h(t)=-\Delta+|u(t,\cdot)|^2\ast w_N -\mu_N(t) + Q(t) \widetilde{K}_1(t) Q(t), \\
& K_2(t, \cdot, \cdot)=Q(t)\otimes Q(t)\widetilde{K}_2(t, \cdot, \cdot).
\end{align*}
Here $\widetilde{K}_1(t)$ is the operator on $\mathfrak{H}$ with kernel $\widetilde{K}_1(t,x,y)=u(t,x)w_N(x-y)\overline{u(t,y)}$, and $\widetilde{K}_2(t,x,y)=u(t,x)w_N(x-y)u(t,y)$. A heuristic derivation of \eqref{eq:Bogoliubov-equation} will be revised in Section \ref{sec:Bogoliubov}.
We will restrict our attention to quasi-free states. Recall that a unit vector $\Psi\in \mathcal{F}(\mathfrak{H})$ is called a quasi-free state if it has finite particle number expectation, namely $\langle \Psi, \mathcal{N} \Psi \rangle<\infty$, and satisfies Wick's Theorem:
\begin{align*}
&\langle \Psi, a^{\#}(f_{1}) a^{\#}(f_{2}) \cdots a^{\#}(f_{2n-1}) \Psi \rangle = 0, \\%\label{eq:Wick-1}\\
&\langle \Psi, a^{\#}(f_{1}) a^{\#}(f_{2}) \cdots a^{\#}(f_{2n}) \Psi \rangle = \sum_{\sigma} \prod_{j=1}^n \langle \Psi, a^{\#}(f_{\sigma(2j-1)}) a^{\#}(f_{\sigma(2j)}) \Psi \rangle
\end{align*}
for all $f_1,...,f_n \in \mathfrak{H}$ and for all $n$. Here $a^{\#}$ is either the creation or annihilation operator and the sum is taken over all permutations $\sigma$ satisfying $\sigma(2j-1)<\min\{\sigma(2j),\sigma(2j+1) \}$ for all $j$. By the definition, any quasi-free state is determined uniquely (up to a phase) by its one-body density matrices $\gamma_\Psi: \mathfrak{H}\to \mathfrak{H}$ and $\alpha_\Psi:\overline{\mathfrak{H}} \equiv \mathfrak{H}^* \to {\mathfrak{H}}$ which are defined by
$$
\left\langle {f,{\gamma _\Psi }g} \right\rangle = \left\langle \Psi, {{a^*}(g)a(f)} \Psi \right\rangle,\quad \left\langle {{f}, \alpha _\Psi \overline{g} } \right\rangle = \left\langle \Psi, {a(g)a(f)} \Psi\right\rangle, \quad \forall f,g \in \mathfrak{H}.
$$
In \cite{NamNap-15}, we proved that if $\Phi(0)$ is a quasi-free state, then the solution $\Phi(t)$ to \eqref{eq:Bogoliubov-equation} is a quasi-free state for all $t>0$ and $(\gamma_{\Phi(t)}, \alpha_{\Phi(t)})$ is the unique solution to the system
\begin{eqnarray} \label{eq:linear-Bog-dm}
\left\{
\begin{aligned}
i\partial_t \gamma &= h \gamma - \gamma h + K_2 \alpha - \alpha^* K_2^*, \\
i\partial_t \alpha &= h \alpha + \alpha h^{\rm T} + K_2 + K_2 \gamma^{\rm T} + \gamma K_2,\\
\gamma(t&=0)=\gamma_{\Phi(0)}, \quad \alpha(t=0) = \alpha_{\Phi(0)}.
\end{aligned}
\right.
\end{eqnarray}
Here $K_2$ is interpreted as an operator $\mathfrak{H}^*\to \mathfrak{H}$ with kernel $K_2(t,x,y)$. Note that \eqref{eq:linear-Bog-dm} is similar (but not identical) to the equations studied in \cite{GriMac-13,Kuz-15b,BacBreCheFroSig-15}. The well-posedness of \eqref{eq:Bogoliubov-equation}-\eqref{eq:linear-Bog-dm} is recalled in Lemma \ref{lem:Bogoliubov-equation}.
Now we are ready to state our main result.
\begin{theorem}[Validity of Bogoliubov dynamics] \label{thm:main} Let $0\le \beta<1/2$.
\begin{itemize}
\item Let $u(t)$ satisfy the Hartree equation \eqref{eq:Hartree-equation}, where the (possibly $N$-dependent) initial state $u(0,\cdot)$ satisfies
$$\| u(0,\cdot)\|_{W^{\ell,1}(\mathbb{R}^3)} \le \kappa_0$$
for $\ell$ sufficiently large and for a constant $\kappa_0>0$ independent of $N$.
\medskip
\item Let $\Phi(t)=(\psi_n(t))_{n=0}^\infty \in \mathcal{F}_+(t)$ satisfy the Bogoliubov equation \eqref{eq:Bogoliubov-equation}, where the (possibly $N$-dependent) initial state $\Phi(0)$ is a quasi-free state in $\mathcal{F}_+(0)$ satisfying
\begin{eqnarray} \label{eq:assumption-Phi0}
\big\langle \Phi(0), \mathcal{N} \Phi(0) \big\rangle \le \kappa_\varepsilon N^{\varepsilon}\quad \text{and} \quad \big\langle \Phi(0), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(0) \big\rangle\le \kappa_\varepsilon N^{\beta+\varepsilon}
\end{eqnarray}
for all $\varepsilon>0$, where the constant $\kappa_\varepsilon>0$ is independent of $N$.
\medskip
\item Let $\Psi_N(t)$ satisfy the Schr\"odinger equation \eqref{eq:schrodingerdynamics} with the initial state
\begin{eqnarray} \label{eq:PhiN0-thm}
\Psi_N(0) = \sum_{n=0}^N u(0)^{\otimes (N-n)} \otimes_s \psi_n(0) = \sum_{n=0}^N \frac{(a^*(u(0)))^{N-n}}{\sqrt{(N-n)!}} \psi_n(0).
\end{eqnarray}
\end{itemize}
Then for all $\varepsilon>0$ and for all $t>0$ we have
\begin{align} \label{eq:thm-mainresult}
\Big\| \Psi_N(t) - \sum_{n=0}^N u(t)^{\otimes (N-n)} \otimes_s \psi_n(t) \Big\|_{\mathfrak{H}^N}^2 \le C_\varepsilon (1+t)^{1+\varepsilon} N^{(2\beta-1+\varepsilon)/2}
\end{align}
where the constant $C_\varepsilon>0$ depends only on $\kappa_0$ and $\varepsilon$.
\end{theorem}
Strictly speaking, the initial state $\Psi_N(0)$ is not normalized. However, its norm converges to $1$ very fast when $N\to \infty$ (we will see it from the proof). We ignore this trivial normalization in the statement of Theorem \ref{thm:main} for simplicity.
Since $\Psi_N(0)$ is expected to be the ground state of a trapped system with the interaction potential $w_N(x-y)$, the initial data $u(0,\cdot)$ and $\Phi(0)$ are allowed to depend on $N$. In particular, the assumptions \eqref{eq:assumption-Phi0} on $\Phi(0)$ are motivated by the ground state properties of quadratic Hamiltonians (see Remark \ref{rmk:ground-state}). More generally, we can also assume that \eqref{eq:assumption-Phi0} holds for {\em some} $\varepsilon>0$, and replace the right side of \eqref{eq:thm-mainresult} by $C_\varepsilon (1+t)^{1+\varepsilon} N^{(2\beta-1+9\varepsilon)/2}$ (see the estimate \eqref{eq:thm-quantitative-estimate} in the proof).
Our proof builds on ideas in \cite{LewNamSch-15,NamNap-15}, where the case $0\le \beta<1/3$ was studied. However, the extension to $\beta<1/2$ requires several new tools, most notably a new kinetic estimate for the particles outside of the condensate (see Lemma \ref{lem:HN-kinetic}). Our method can be applied to study the norm approximation in Fock space, for example to simplify significantly the proof in \cite{Kuz-15b}. The range $0\le \beta<1/2$ is expected to be optimal under the assumptions on the initial states in Theorem \ref{thm:main}.
The paper is organized as follows. We will revise the well-posedness of the Hartree equation \eqref{eq:Hartree-equation} and the Bogoliubov equation \eqref{eq:Bogoliubov-equation} in Section \ref{sec:well-posedness}. In section \ref{sec:Bogoliubov}, we reformulate the problem using a unitary transformation from $\mathfrak{H}^N$ to a truncated Fock space, following ideas in \cite{LewNamSerSol-15,LewNamSch-15}. Then we provide several estimates which are useful to implement Bogoliubov's approximation. The proof of Theorem \ref{thm:main} is presented in Section \ref{sec:main-proof}.
\section{Well-posedness of the effective equations}\label{sec:well-posedness}
From \cite[Prop. 3.3 \& Cor. 3.4]{GriMac-13} we have the following well-posedness of the Hartree equation.
\begin{lemma} \label{lem:Hartree-equation} If $u(0,\cdot)\in H^2(\mathbb{R}^3)$, then the Hartree equation \eqref{eq:Hartree-equation} has a unique global solution $u \in C( [0,\infty),H^2(\mathbb{R}^3)) \cap C^1((0,\infty),L^2(\mathbb{R}^3))$. Moreover, if $u(0,\cdot)\in W^{\ell,1}(\mathbb{R}^3)$ with $\ell$ sufficiently large, then $\|u(t,\cdot)\|_{H^2} \le C$, $\|\partial_t u(t,\cdot)\|_{L^2} \le C$ and
$$ \|u(t, \cdot)\|_{L^\infty(\mathbb{R}^3)} + \|\partial_t u(t,\cdot)\|_{L^\infty(\mathbb{R}^3)} \le \frac{C}{(1+t)^{3/2}}$$
for a constant $C$ depending only on $\|u(0)\|_{W^{\ell,1}(\mathbb{R}^3)}$.
\end{lemma}
From now on, we always assume that $u(0,\cdot)\in W^{\ell,1}(\mathbb{R}^3)$ with $\ell$ sufficiently large. We will also denote by $C$ a general constant depending only on $\|u(0,\cdot)\|_{W^{\ell,1}}$ (whose value can be changed from line to line). Indeed, more precisely, $C$ depends only on $\kappa_0$ in the condition $\| u(0,\cdot)\|_{W^{\ell,1}(\mathbb{R}^3)} \le \kappa_0$ (c.f. Theorem \ref{thm:main}).
Next, we recall the well-posedness of the Bogoliubov equation from \cite[Theorem 7]{LewNamSch-15} and \cite[Prop. 4]{NamNap-15}.
\begin{lemma} \label{lem:Bogoliubov-equation} For every initial state $\Phi(0)$ in the quadratic form domain of $\mathcal{Q}({\ensuremath{\rm d}\Gamma} (1-\Delta))$, the Bogoliubov equation \eqref{eq:Bogoliubov-equation} has a unique global solution $\Phi \in C([0,\infty), \mathcal{F}(\mathfrak{H})) \cap L^\infty_{\rm loc} ((0,\infty), \mathcal{Q}({\ensuremath{\rm d}\Gamma} (1-\Delta)))$. Moreover, if $\Phi(0)$ is a quasi-free state in $\mathcal{F}_+(0)$, then $\Phi(t)$ is a quasi-free state in $\mathcal{F}_+(t)$ and
$$
\langle \Phi(t), \mathcal{N} \Phi(t) \rangle \le C \Big( \langle \Phi(0),\mathcal{N} \Phi(0)\rangle^2 + [\log(2+t)]^2\Big).
$$
\end{lemma}
We have two remarks on the Bogoliubov equation \eqref{eq:Bogoliubov-equation}. First, although the Bogoliubov Hamiltonian $\mathbb{H}(t)$ is not necessarily bounded from below and has not been defined as a self-adjoint operator, the solution $\Phi(t)$ to \eqref{eq:Bogoliubov-equation} can be still interpreted as an evolution generated by quadratic forms (see \cite[Theorems 7, 8]{LewNamSch-15} for further discussion). Second, when $\Phi(t)$ is a quasi-free state, then the Bogoliubov equation \eqref{eq:Bogoliubov-equation} becomes equivalent to the system \eqref{eq:linear-Bog-dm} (see \cite[Prop. 4]{NamNap-15} for more details), but we will not need this fact in the rest of paper.
The main new result of this section is the following kinetic estimate.
\begin{lemma} \label{lem:bH-kinetic} Assume that $\Phi(0)$ is a quasi-free state in $\mathcal{F}_+(0)$ satisfying
$$
\big \langle \Phi(0), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(0) \big\rangle \le \kappa_\varepsilon N^{\beta+\varepsilon}
$$
for some $\varepsilon>0$, where the constant $\kappa_\varepsilon$ is independent of $N$. Then
$$
\big \langle \Phi(t), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(t) \big\rangle \le C_\varepsilon N^{\beta+\varepsilon} ,\quad \forall t>0.
$$
\end{lemma}
Hereafter, $C_\varepsilon$ is a general constant depending only on $\|u(0,\cdot)\|_{W^{\ell,1}}$ (more precisely, on $\kappa_0$ in the condition $\| u(0,\cdot)\|_{W^{\ell,1}} \le \kappa_0$) and $\varepsilon$.
To prove Lemma \ref{lem:bH-kinetic}, we will need a general lower bound on the ground state energy of quadratic Hamiltonians.
\begin{lemma} \label{lem:Bog-GSE} Let $H>0$ be a self-adjoint operator on $\mathfrak{H}$. Let $K:\overline{\mathfrak{H}}\equiv \mathfrak{H}^*\to \mathfrak{H}$ be an operator with kernel $K(x,y) \in \mathfrak{H}^2$. Assume that $K H^{-1} K^* \le H$ and that $H^{-1/2}K$ is Hilbert-Schmidt. Then
$$ {\ensuremath{\rm d}\Gamma}(H) + \frac{1}{2} \iint \Big( K(x,y) a_x^* a_y^* + \overline{K(x,y)}a_x a_y \Big) \,{\rm d} x \,{\rm d} y \ge -\frac{1}{2} \| H^{-1/2} K\|_{\rm HS}^2.$$
\end{lemma}
This result is taken from \cite[Lemma 9]{NamNapSol-16} (see also \cite[Theorem 5.4]{BruDer-07}). Note that
$$ \| H^{-1/2} K\|_{\rm HS}^2 = \iint |H_x^{-1/2} K(x,y)|^2 \,{\rm d} x \,{\rm d} y =: \|H_x^{-1/2}K(\cdot, \cdot)\|_{L^2}^2.$$
Here we write $H_x$ to mention that the operator $H$ acts on the $x$-variable.
If we apply Lemma \ref{lem:Bog-GSE} with $H=1+\|K_2\|$ and $K=\pm K_2$, then we get
\begin{eqnarray} \label{eq:bound-paring-dG1}
\pm \frac{1}{2} \iint \Big( K_2(t,x,y) a_x^* a_y^* + \overline{K_2(t,x,y)}a_x a_y \Big) \,{\rm d} x \,{\rm d} y \le C \mathcal{N} + \frac{CN^{3\beta}}{(1+t)^3}.
\end{eqnarray}
Here we have used the bound on $\|K_2\|$ in \eqref{eq:norm-K2} and
\begin{align} \label{eq:L2-K2-N3beta}
\|K_2(t,\cdot,\cdot)\|_{L^2}^2 &\le \|\widetilde K_2(t,\cdot,\cdot)\|_{L^2}^2 = \iint | u(t,x)|^2 |w_N(x-y)|^2 |u(t,y)|^2 \,{\rm d} x \,{\rm d} y \nonumber\\
&\le \|w_N\|_{L^2}^2 \|u(t,\cdot)\|_{L^\infty}^2 \| u(t,\cdot)\|_{L^2}^2 \le \frac{CN^{3\beta}}{(1+t)^3}
\end{align}
by Lemma \ref{lem:Hartree-equation}. In order to improve the factor $N^{3\beta}$ in \eqref{eq:bound-paring-dG1}, we will apply Lemma \ref{lem:Bog-GSE} with $H=1-\Delta$. We will need the following estimate.
\begin{lemma}\label{lem:Sobolev-inverse-K2} For all $\varepsilon>0$ we have
\begin{align*}
\|(1-\Delta_x)^{-1/2} K_2(t, \cdot, \cdot)\|^2_{L^2} + \|(1-\Delta_x)^{-1/2} \partial_t K_2(t, \cdot, \cdot)\|^2_{L^2} \le \frac{C_\varepsilon N^{\beta+\varepsilon}}{(1+t)^3}.
\end{align*}
\end{lemma}
\begin{proof} We will present a detailed proof for $\partial_t K_2(t)$ and $K_2(t)$ can be treated by the same way. Recall that
$$ K_2(t,\cdot ,\cdot)=Q(t)\otimes Q(t) \widetilde K_2(t,\cdot,\cdot), \quad \widetilde K_2(t,x,y)=u(t,x)w_N(x-y) u(t,y).$$
Hence,
$$ \partial_t K_2(t)= \partial_t Q(t) \otimes Q(t) \widetilde K_2(t) + Q(t) \otimes \partial_t Q(t) \widetilde K_2(t)+Q(t)\otimes Q(t) \partial_t \widetilde K_2(t).$$
Since $\partial_t Q(t)=-| \partial_t u(t)\rangle \langle u(t)| -|u(t)\rangle \langle \partial_t u(t)|$, we have
\begin{align*}
&\|\partial_t Q(t) \otimes Q(t) \widetilde K_2(t,\cdot,\cdot)\|_{L^2} \le \| (\partial_t Q(t) \otimes 1) \widetilde K_2(t,\cdot,\cdot)\|_{L^2} \\
& \le \| (| \partial_t u(t)\rangle \langle u(t)| \otimes 1 ) \widetilde K_2(t,\cdot,\cdot)\|_{L^2} + \| (| u(t)\rangle \langle \partial_t u(t)| \otimes 1 ) \widetilde K_2(t,\cdot,\cdot)\|_{L^2}
\end{align*}
Using Lemma \ref{lem:Hartree-equation} and $\|w_N\|_{L^1}=\|w\|_{L^1}$, it is straightforward to see that
\begin{align*}
&\left\| (|\partial_t u\rangle \langle u| \otimes 1) \widetilde K_2(t, \cdot, \cdot) \right\|_{L^2}^2 \nonumber\\
&=\iint \left| \int \overline{u(t,z)} u(t,z) w_N(z-y) u(t,y) \,{\rm d} z \right|^2 |\partial_t u(t,x)|^2 \,{\rm d} x \,{\rm d} y \nonumber\\
& \le \|u(t,\cdot)\|_{L^\infty}^4 \|w_N\|_{L^1}^2 \|u(t,\cdot)\|_{L^2}^2 \|\partial_t u(t,\cdot)\|_{L^2}^2 \le \frac{C}{(1+t)^{3}}.
\end{align*}
combining this with similar estimates, we find that
\begin{align} \label{eq:L2-dtQ-K2}
&\|\partial_t Q(t) \otimes Q(t) \widetilde K_2(t,\cdot,\cdot)\|_{L^2} + \| Q(t) \otimes \partial_t Q(t) \widetilde K_2(t,\cdot,\cdot)\|_{L^2} \nonumber \\
& \le \|(\partial_t Q(t) \otimes 1) \widetilde K_2(t,\cdot,\cdot)\|_{L^2} + \| (1 \otimes \partial_t Q(t)) \widetilde K_2(t,\cdot,\cdot)\|_{L^2} \le \frac{C}{(1+t)^{3/2}}.
\end{align}
By the same argument, we also obtain
$$\| (1-Q(t)\otimes Q(t)) \partial_t \widetilde K_2(t,\cdot,\cdot)\|_{L^2} \le \frac{C}{(1+t)^{3/2}}.$$
Note that $(1-\Delta_x)^{-1/2} \le 1$ on $L^2$, and hence we can insert $(1-\Delta_x)^{-1/2}$ into the above $L^2$ norm estimates for free. It remains to show that
\begin{eqnarray} \label{eq:eq:dt-K2-half}
\|(1-\Delta_x)^{-1/2} \partial_t \widetilde K_2(t, \cdot, \cdot)\|^2_{L^2} \le \frac{C_\varepsilon N^{\beta+\varepsilon}}{(1+t)^3}, \quad \forall \varepsilon>0.
\end{eqnarray}
Similarly to \eqref{eq:L2-K2-N3beta}, we have
$$
\|\partial_t \widetilde K_2(t,\cdot,\cdot) \|_{L^2}^2 \le \frac{CN^{3\beta}}{(1+t)^3}.
$$
Therefore, by interpolation (more precisely, by H\"older's inequality in Fourier space), \eqref{eq:eq:dt-K2-half} follows from the following estimate
\begin{align} \label{eq:1-Delta-K2-075}
\|(1-\Delta_x)^{-3/4-\varepsilon} \partial_t \widetilde K_2(t, \cdot, \cdot)\|^2_{L^2} \le \frac{C_\varepsilon}{(1+t)^3}, \quad \forall \varepsilon>0.
\end{align}
It suffices to show that
\begin{align} \label{eq:1-Delta-075}
\|(1-\Delta_x)^{-3/4-\varepsilon} f(t, \cdot, \cdot)\|^2_{L^2} \le \frac{C_\varepsilon}{(1+t)^3}, \quad \forall \varepsilon>0
\end{align}
with $f(t,x,y)=\partial_t u(t,x) w_N(x-y) u(t,y)$. The bound \eqref{eq:1-Delta-075} can be proved using an argument in \cite{GriMac-13}. Let us compute the Fourier transform:
\begin{align*}
\widehat{f}(t,p,q) &= \iint u(t,x)w_N(x-y)(\partial_t u)(t,y) e^{-2\pi i (p\cdot x + q\cdot y)} \,{\rm d} x \,{\rm d} y \\
&=\iint u(t,y+z) w_N(z) (\partial_t u)(t,y) e^{-2\pi i (p\cdot (y+z) + q\cdot y)} \,{\rm d} z\,{\rm d} y \\
&=\int w_N(z) \widehat{(u_z \partial_t u)}(t,p+q) e^{-2\pi ip\cdot z}\,{\rm d} z
\end{align*}
where $u_z(t,\cdot):=u(t,z+\cdot).$ By the Cauchy-Schwarz inequality,
\begin{align*}
\left|\widehat{f}(t,p,q) \right|^2 \le \|w_N\|_{L^1} \int |w_N(z)| \cdot| \widehat{(u_z \partial_t u)}(t,p+q)|^2 \,{\rm d} z.
\end{align*}
Using Plancherel's Theorem, we can estimate
\begin{align*}
& \|(1-\Delta_x)^{-3/4-\varepsilon} f(t, \cdot, \cdot)\|^2_{L^2} = \iint (1+|2\pi p|^2)^{-3/2-2\varepsilon}\left|\widehat{f}(t,p,q) \right|^2 \,{\rm d} p \,{\rm d} q \\
&\le \|w_N\|_{L^1} \iiint (1+|2\pi p|^2)^{-3/2-2\varepsilon} |w_N(z)| \cdot |\widehat{(u_z \partial_t u)}(t,p+q)|^2 \,{\rm d} p \,{\rm d} q \,{\rm d} z.
\end{align*}
By Lemma \ref{lem:Hartree-equation},
\begin{align*}
\int |\widehat{(u_z \partial_t u)}(t,p+q)|^2 \,{\rm d} q & = \| (u_z \partial_t u)(t,\cdot) \|_{L^2}^2 \\
& \le \|u(t,\cdot)\|_{L^\infty}^2 \|\partial_t u(t,\cdot)\|_{L^2}^2 \le \frac{C}{(1+t)^3}.
\end{align*}
Therefore, \eqref{eq:1-Delta-075} follows from $\|w_N\|_{L^1}=\|w\|_{L^1}$ and the fact that
\begin{equation*}
\int (1+|2\pi p|^2)^{-3/2-2\varepsilon} \,{\rm d} p \le C_\varepsilon <\infty.
\end{equation*}
Thus \eqref{eq:1-Delta-075} holds true. By the same argument, we obtain a similar inequality with $f(t,x,y)$ replaced by $u(t,x)w_N(x-y) \partial_t u(t,y)$. Combining these two estimates, we deduce \eqref{eq:1-Delta-K2-075}. This completes the proof.
\end{proof}
Now we apply Lemmas \ref{lem:Bog-GSE} and \ref{lem:Sobolev-inverse-K2} to bound $\mathbb{H}(t)$.
\begin{lemma} \label{lem:bHt-dbHt} For every $\varepsilon>0$ and $\eta>0$, we have
\begin{align*}
\pm \Big( \mathbb{H}(t) + {\ensuremath{\rm d}\Gamma}(\Delta) \Big) &\le \eta {\ensuremath{\rm d}\Gamma}(1-\Delta) + \frac{C_\varepsilon(\mathcal{N}+N^{\beta+\varepsilon})}{\eta(1+t)^3},\\
\pm \partial_t \mathbb{H}(t) &\le \eta {\ensuremath{\rm d}\Gamma}(1-\Delta) + \frac{C_\varepsilon (\mathcal{N} + N^{\beta+\varepsilon})}{ \eta (1+t)^3},\\
\pm i[\mathbb{H}(t),\mathcal{N}] &\le \eta {\ensuremath{\rm d}\Gamma}(1-\Delta) + \frac{C_\varepsilon (\mathcal{N} + N^{\beta+\varepsilon})}{ \eta (1+t)^3},
\end{align*}
as quadratic forms on $\mathcal{F}(\mathfrak{H})$. The constant $C_\varepsilon$ is independent of $\eta$.
\end{lemma}
\begin{proof} First, we consider
$$
\mathbb{H}(t) + {\ensuremath{\rm d}\Gamma}(\Delta)= {\ensuremath{\rm d}\Gamma}(h+\Delta) + \frac{1}{2} \iint \Big( K_2(t,x,y) a_x^* a_y^* + \overline{K_2(t,x,y)}a_x a_y \Big) \,{\rm d} x \,{\rm d} y.
$$
Recall that
$$h+\Delta = |u(t,\cdot)|^2\ast w_N -\mu_N(t) + Q(t) \widetilde K_1(t) Q(t),$$
where $\widetilde K_1(t)$ is the operator with kernel $\widetilde K_1(t)=u(t,x)w_N(x-y)\overline{u(t,y)}$. Using Lemma \ref{lem:Hartree-equation}, we have
\begin{align} \label{eq:K1-norm-1}
&\| |u(t,\cdot)|^2*w_N \|_{L^\infty} \le \|u(t,\cdot)\|_{L^\infty}^2 \|w_N\|_{L^1} \le \frac{C}{(1+t)^3},\\
&\mu_N(t) = \frac{1}{2} \int |u(t,x)|^2 (|u(t,\cdot)|^2*w_N)(x) \,{\rm d} x \le \frac{C}{(1+t)^3}.\label{eq:K1-norm-2}
\end{align}
Moreover, $\| Q(t) \widetilde K_1(t) Q(t) \| \le \| \widetilde K_1(t) \|$ and
\begin{align} \label{eq:K1-norm}
&\| \widetilde K_1(t) \| = \sup_{ \|f\|_{L^2}=1} \left| \iint \overline{f(x)} u(t,x) w_N(x-y) \overline{u(t,y)}f(y) \,{\rm d} x \,{\rm d} y \right| \\
& \le \sup_{\|f\|_{L^2}=1} \|u(t,\cdot)\|_{L^\infty}^2 \iint \frac{|f(x)|^2+|f(y)|^2}{2} w_N(x-y) \,{\rm d} x \,{\rm d} y \le \frac{C}{(1+t)^3}.\nonumber
\end{align}
From \eqref{eq:K1-norm-1}, \eqref{eq:K1-norm-2} and \eqref{eq:K1-norm} and the triangle inequality, we get
\begin{eqnarray} \label{eq:norm-h}
\| h + \Delta \| \le \frac{C}{(1+t)^3}.
\end{eqnarray}
Similarly to \eqref{eq:K1-norm}, we can bound the operator $K_2(t):\mathfrak{H}^*\to \mathfrak{H}$ as
\begin{eqnarray} \label{eq:norm-K2}
\|K_2(t)\| \le \frac{C}{(1+t)^3}.
\end{eqnarray}
Now we apply Lemma \ref{lem:Bog-GSE} with $H=\eta(1-\Delta)+\eta^{-1}\|K_2\|^2$ and $K=\pm K_2$, where $\eta>0$ is arbitrary. Since $H \ge \|K\|$, we have
$$H \ge \|K\| \ge K \|K\|^{-1}K^* \ge KH^{-1}K .$$
Moreover, using $H \ge \eta(1-\Delta)$ and Lemma \ref{lem:Sobolev-inverse-K2}, we get
$$
\|H^{-1/2}K\|_{\rm HS}^2 \le \eta^{-1}\|(1-\Delta)^{-1/2}K_2\|_{\rm HS}^2 \le \frac{C_\varepsilon N^{\beta+\varepsilon}}{\eta(1+t)^3}.
$$
Therefore, Lemma \ref{lem:Bog-GSE} implies that
\begin{align} \label{eq:pairing-dGD}
&\pm \frac{1}{2} \iint \Big( K_2(t,x,y) a_x^* a_y^* + \overline{K_2(t,x,y)}a_x a_y \Big) \,{\rm d} x \,{\rm d} y \nonumber \\
& \le \eta{\ensuremath{\rm d}\Gamma}(1-\Delta) + \eta^{-1}\|K_2\|^2 \mathcal{N} + \frac{C_\varepsilon N^{\beta+\varepsilon}}{\eta(1+t)^3}.
\end{align}
From \eqref{eq:norm-h}, by using the Cauchy-Schwarz inequality $C(1+t)^{-3} \le \eta + C\eta^{-1}(1+t)^{-6}$ we get
$$
\pm {\ensuremath{\rm d}\Gamma}(h+\Delta) \le \eta \mathcal{N} + \frac{C \mathcal{N}}{\eta(1+t)^{6}}
$$
for all $\eta>0$. Combining this with \eqref{eq:pairing-dGD} (and the obvious bound $\mathcal{N}\le {\ensuremath{\rm d}\Gamma}(1-\Delta)$) we conclude that
$$
\pm \Big( \mathbb{H}(t) + {\ensuremath{\rm d}\Gamma}(\Delta) \Big) \le \eta {\ensuremath{\rm d}\Gamma}(1-\Delta) + \frac{C_\varepsilon (\mathcal{N}+N^{\beta+\varepsilon})}{\eta(1+t)^3}.
$$
The bound on $\partial_t \mathbb{H}(t)$ is obtained by the same way. Indeed, by Lemma \ref{lem:Hartree-equation},
\begin{align} \label{eq:der-K1-1}
& \| \partial_t (|u(t,\cdot)|^2\ast w_N) \|_{L^\infty} \le \| \partial_t |u(t,\cdot)|^2 \|_{L^\infty} \|w_N\|_{L^1} \le \frac{C}{(1+t)^3},\\
&| \partial_t \mu_N(t) | \le 2 \iint |\partial_t (|u(t,x)|^2)| w_N(x-y) |u(t,y)|^2 \,{\rm d} x \,{\rm d} y \le \frac{C}{(1+t)^3}, \label{eq:der-K1-2}\\
\label{eq:der-K1}&\| \partial_t (Q(t) \widetilde K_1(t) Q(t) ) \| \le \|\partial_t \widetilde K_1(t)\| + 2 \|\partial_t Q(t)\| \cdot \|\widetilde K_1(t)\| \le \frac{C}{(1+t)^3}.
\end{align}
Similarly to \eqref{eq:der-K1}, we can bound the operator $\partial_t K_2(t): \mathfrak{H}^*\to \mathfrak{H}$ as
\begin{eqnarray} \label{eq:norm-dt-K2}
\| \partial_t K_2(t)\| \le \frac{C}{(1+t)^3}.
\end{eqnarray}
Then we apply Lemma \ref{lem:Bog-GSE} with $H=\eta(1-\Delta) + \eta^{-1}\|\partial_t K_2\|^2$ and $K= \pm \partial_t K_2$ and obtain
\begin{equation*}
\pm \partial_t \mathbb{H}(t) \le \eta {\ensuremath{\rm d}\Gamma}(1-\Delta) + \frac{C_\varepsilon( \mathcal{N} + N^{\beta+\varepsilon})}{ \eta (1+t)^3}.
\end{equation*}
Finally, since $[a_x^*a^*_y,\mathcal{N}]= -2 a_x^* a_y^*$, we have
\begin{align*}
i[\mathbb{H}(t),\mathcal{N}] &= \frac{i}{2} \iint \Big( K_2(t,x,y) [a_x^* a_y^*,\mathcal{N}] + \overline{K_2(t,x,y)}[a_x a_y,\mathcal{N}] \Big) \,{\rm d} x \,{\rm d} y \\
&= - \iint \Big( iK_2(t,x,y) a_x^* a_y^* + \overline{iK_2(t,x,y)}a_x a_y \Big) \,{\rm d} x \,{\rm d} y
\end{align*}
Using Lemma \ref{lem:Bog-GSE} again, we obtain \eqref{eq:pairing-dGD} with $K_2$ replaced by $-2iK_2$, namely
\begin{equation*}
\pm i[\mathbb{H}(t),\mathcal{N}] \le \eta {\ensuremath{\rm d}\Gamma}(1-\Delta) + \frac{C_\varepsilon (\mathcal{N} + N^{\beta +\varepsilon})}{\eta(1+t)^3}
\end{equation*}
for all $\eta>0$. This completes the proof.
\end{proof}
\begin{remark} \label{rmk:ground-state} Assume that $\Phi$ is the ground state of the quadratic Hamiltonian
$$ \mathbb{H}_V(0) = {\ensuremath{\rm d}\Gamma}(h(0)+V) + \frac{1}{2} \iint \Big( K_2(0,x,y) a_x^* a_y^* + \overline{K_2(0,x,y)}a_x a_y \Big) \,{\rm d} x \,{\rm d} y,$$
where $V(x)$ is an appropriate trapping potential which ensures that $h(0)+V \ge \eta >0$ (in particular, this implies that $h(0)+V \ge C_\eta^{-1} (1-\Delta)$ for some constant $C_\eta>0$). By Lemma \ref{lem:Bog-GSE}, we know that
\begin{align*}
\mathbb{H}_V(0) &\ge \frac{1}{2}{\ensuremath{\rm d}\Gamma}(h(0)+V) + \frac{1}{2} \iint \Big( K_2(0,x,y) a_x^* a_y^* + \overline{K_2(0,x,y)}a_x a_y \Big) \,{\rm d} x \,{\rm d} y \\
&\ge - C \|(1-\Delta_x)^{-1/2} K_2(0,\cdot,\cdot)\|_{L^2}^2 \ge -C_\varepsilon N^{\beta+\varepsilon}.
\end{align*}
In particular, $\mathbb{H}_V(0)$ is bounded from below and it can be diagonalized by a Bogoliubov transformation. Moreover, $\Phi$ is a quasi-free state
and
$$ \langle \Phi, \mathcal{N} \Phi \rangle \le C \|(1-\Delta_x)^{-1} K_2(0,\cdot,\cdot)\|_{L^2}^2 \le C.$$
Here the first bound is a consequence of \cite[Theorem 1 (ii)]{NamNapSol-16}.
Since the ground state energy of $\mathbb{H}_V(0)$ is always non-positive (see \cite[Theorem 2.1 (i)]{LewNamSerSol-15}), we have
$$
0 \ge \langle \Phi, \mathbb{H}_V(0) \Phi\rangle \ge \frac{1}{2}\langle \Phi, {\ensuremath{\rm d}\Gamma}(h(0)+V) \Phi \rangle - C_\varepsilon N^{\beta+\varepsilon}.
$$
Combining with $h(0)+V \ge C_\eta^{-1} (1-\Delta)$, we obtain
$$\langle \Phi, {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi \rangle \le C_\varepsilon N^{\beta+\varepsilon}, \quad \forall \varepsilon>0.$$
This motivates the assumptions on $\Phi(0)$ in Theorem \ref{thm:main}.
\end{remark}
We are ready to give
\begin{proof}[Proof of Lemma \ref{lem:bH-kinetic}] From the Bogoliubov equation \eqref{eq:Bogoliubov-equation}, we have
$$
\partial_t \big\langle \Phi(t), \mathbb{H} (t) \Phi(t) \big\rangle = \big\langle \Phi(t), \partial_t \mathbb{H} (t) \Phi(t) \big\rangle
$$
which implies that
\begin{align} \label{eq:Phit-Grw}
\big\langle \Phi(t), \mathbb{H} (t) \Phi(t) \big\rangle - \big\langle \Phi(0), \mathbb{H} (0) \Phi(0) \big\rangle = \int_0^t \big\langle \Phi(s), \partial_s \mathbb{H} (s) \Phi(s) \big\rangle \,{\rm d} s.
\end{align}
From the first bound in Lemma \ref{lem:bHt-dbHt} with $\eta=1/2$, we get
\begin{align*}
\pm \big\langle \Phi(t), (\mathbb{H} (t) + {\ensuremath{\rm d}\Gamma}(\Delta) ) \Phi(t) \big\rangle &\le \frac{1}{2} \big \langle \Phi(t), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(t) \big\rangle \\
&\qquad \qquad + C_\varepsilon \Big( \big \langle \Phi(t), \mathcal{N} \Phi(t) \big\rangle + N^{\beta+\varepsilon} \Big).
\end{align*}
This implies
\begin{align} \label{eq:kin-Phi-1}
\big\langle \Phi(0), \mathbb{H} (0) \Phi(0) \big\rangle &\le \frac{3}{2}\big \langle \Phi(0), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(0) \big\rangle + \nonumber\\
&\qquad + C_\varepsilon \Big( \big \langle \Phi(0), \mathcal{N} \Phi(0) \big\rangle + N^{\beta+\varepsilon} \Big) \le C_\varepsilon N^{\beta+\varepsilon}
\end{align}
(where we have used the assumption on $\Phi(0)$) and
\begin{align}
\big\langle \Phi(t), \mathbb{H} (t) \Phi(t) \big\rangle \ge \frac{1}{2} \big \langle \Phi(t), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(t) \big\rangle - C_\varepsilon \Big( \big \langle \Phi(t), \mathcal{N} \Phi(t) \big\rangle + N^{\beta+\varepsilon} \Big). \label{eq:kin-Phi-2}
\end{align}
Next, from the second bound in Lemma \ref{lem:bHt-dbHt} with $\eta=(1+t)^{-3/2}$ we have
\begin{align}
\big\langle \Phi(t), \partial_t \mathbb{H} (t)\Phi(t) \big\rangle \le C_\varepsilon \frac{ \big\langle \Phi(t), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(t) \big\rangle+ N^{\beta+\varepsilon}}{(1+t)^{3/2}} . \label{eq:kin-Phi-3}
\end{align}
Inserting \eqref{eq:kin-Phi-1}, \eqref{eq:kin-Phi-2} and \eqref{eq:kin-Phi-3} into \eqref{eq:Phit-Grw}, we obtain
\begin{align} \label{eq:Phit-Grw-1}
\big \langle \Phi(t), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(t) \big\rangle & \le C_\varepsilon \int_0^t \frac{\big \langle \Phi(s), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(s) \big\rangle}{(1+s)^{3/2}} \,{\rm d} s \nonumber\\ &\qquad \qquad + C_\varepsilon \Big( \big \langle \Phi(t), \mathcal{N} \Phi(t) \big\rangle + N^{\beta+\varepsilon} \Big).
\end{align}
Now instead of using the bound on $\big \langle \Phi(t), \mathcal{N} \Phi(t) \big\rangle$ in Lemma \ref{lem:Bogoliubov-equation}, we present another argument which will be used again to deal with the many-body Schr\"odinger evolution in Section \ref{sec:Bogoliubov}. From the Bogoliubov equation \eqref{eq:Bogoliubov-equation} and the third bound in Lemma \ref{lem:bHt-dbHt} with $\eta=(1+t)^{-3/2}$, it follows that
\begin{align*}
\partial_t \langle \Phi(t), \mathcal{N} \Phi(t) \rangle = \langle \Phi(t), i[\mathbb{H}(t),\mathcal{N}] \Phi(t) \rangle \le C_\varepsilon \frac{ \big\langle \Phi(t), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(t) \big\rangle+ N^{\beta+\varepsilon}}{(1+t)^{3/2}}.
\end{align*}
Integrating over $t$ and using the assumption on $\Phi(0)$ we have
\begin{align*}
\langle \Phi(t), \mathcal{N} \Phi(t) \rangle \le C_\varepsilon \int_0^t \frac{ \big\langle \Phi(s), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(s) \big\rangle}{(1+s)^{3/2}} + C_\varepsilon N^{\beta+\varepsilon}.
\end{align*}
Inserting the latter inequality into the right side of \eqref{eq:Phit-Grw-1} we obtain
\begin{align} \label{eq:Phit-Grw-2}
\big \langle \Phi(t), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(t) \big\rangle & \le C_\varepsilon \int_0^t \frac{\big \langle \Phi(s), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(s) \big\rangle}{(1+s)^{3/2}} \,{\rm d} s + C_\varepsilon N^{\beta+\varepsilon}.
\end{align}
Now we define
$$ f(t) := \int_0^t \frac{\big\langle \Phi(s), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(s) \big\rangle}{(1+s)^{3/2}} \,{\rm d} s + N^{\beta+\varepsilon}$$
and rewrite \eqref{eq:Phit-Grw-2} as
$$
\frac{\,{\rm d}}{\,{\rm d} t} \log (f(t)) = \frac{f'(t)}{f(t)} \le \frac{C_\varepsilon}{(1+t)^{3/2} }.
$$
Integrating over $t$ and using the fact that $(1+t)^{-3/2}$ is integrable on $(0,\infty)$, we get $f(t)\le C_\varepsilon N^{\beta+\varepsilon}$. The desired result then follows from \eqref{eq:Phit-Grw-2}.
\end{proof}
\section{Bogoliubov approximation} \label{sec:Bogoliubov}
Recall that any vector $\Psi\in \mathfrak{H}^N$ can be decomposed uniquely as
\begin{equation*}
\Psi=\sum_{n=0}^N u(t)^{\otimes (N-n)} \otimes_s \psi_n = \sum_{n=0}^N \frac{(a^*(u(t)))^{N-n}}{\sqrt{(N-n)!}} \psi_n
\end{equation*}
with $\psi_n \in \mathfrak{H}_+(t)^{n}$, see ~\cite[Sec. 2.3]{LewNamSerSol-15}. This gives rise the unitary operator
\begin{equation*}
\begin{array}{cccl}
U_{N}(t): & \mathfrak{H}^N & \to & \displaystyle \mathcal{F}_+^{\le N}(t):=\bigoplus_{n=0}^N \mathfrak{H}_+(t)^n \\[0.3cm]
& \Psi & \mapsto & \psi_0\oplus \psi_1 \oplus\cdots \oplus \psi_N.
\end{array}
\end{equation*}
Following \cite{LewNamSch-15}, we reformulate the Schr\"odinger equation $\Psi_N(t)=e^{-itH_N} \Psi(0)$ by introducing
$$\Phi_N(t):=U_N(t) \Psi_N(t)$$
which belongs to $\mathcal{F}_+^{\le N}(t)$ and satisfies the equation
\begin{eqnarray} \label{eq:eq-PhiNt}
\left\{
\begin{aligned}
i \partial_t \Phi_N(t) &= \widetilde H_N (t) \Phi_N(t), \\
\Phi_N(0) & = \1^{\le N} \Phi(0).
\end{aligned}
\right.
\end{eqnarray}
Here $\1^{\le N}$ is the projection onto $\mathcal{F}^{\le N}=\mathbb{C} \oplus \mathfrak{H} \oplus \cdots \oplus \mathfrak{H}^N$ and
$$
\widetilde H_N (t)= \1^{\le N} \Big[ \mathbb{H}(t) + \frac{1}{2}\sum_{j=0}^4 ( R_{j} + R_j^*) \Big] \1^{\le N}
$$
with
\begin{align*}
R_{0}&=R_0^*= \,{\rm d}\Gamma(Q(t)[w_N*|u(t)|^2+ \widetilde{K}_1(t) -\mu_N(t)]Q(t))\frac{1-\mathcal{N}}{N-1},\\
R_{1}&=-2\frac{\mathcal{N}\sqrt{N-\mathcal{N}}}{N-1} a(Q(t)[w_N*|u(t)|^2]u(t)),\\
R_{2}&= \iint K_2(t,x,y) a^*_x a^*_y \,{\rm d} x \,{\rm d} y \left(\frac{\sqrt{(N-\mathcal{N})(N-\mathcal{N}-1)}}{N-1}-1\right),\\
R_{3}& =\frac{\sqrt{N-\mathcal{N}}}{N-1}\iiiint( 1 \otimes Q(t) w_N Q(t)\otimes Q(t))(x,y;x',y')\times \\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \times \overline{u(t,x)} a^*_y a_{x'} a_{y'} \,\,{\rm d} x \,{\rm d} y \,{\rm d} x' \,{\rm d} y',\\
R_{4}&=R_4^*= \frac{1}{2(N-1)}\iiiint({Q(t)}\otimes{Q(t)}w_N Q(t)\otimes Q(t))(x,y;x',y')\times \\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \times a^*_x a^*_y a_{x'} a_{y'} \,\,{\rm d} x \,{\rm d} y \,{\rm d} x' \,{\rm d} y'.
\end{align*}
Here, in $R_0$ and $R_1$ we write $w_N$ for the function $w_N(x)$, while in $R_3$ and $R_4$ we write $w_N$ for the two-body multiplication operator $w_N(x-y)$.
In order to compare $\Phi_N(t)$ with the Bogoliubov dynamics $\Phi(t)$, we need to bound all error terms $R_j$'s. In \cite[Prop. 3]{NamNap-15}, we proved that
$$ (R_j+R_j^*) \1^{\le N} (R_j+R_j^*) \le C ( N^{6\beta-2} + N^{3\beta-1}) (1+\mathcal{N})^4.$$
Unfortunately, this bound is only useful when $\beta<1/3$. In the present paper, we will derive several improved estimates. Let us start with
\begin{lemma}\label{lem:Rj} We have the quadratic form estimates on $\mathcal{F}_+^{\le N}$:
\begin{align*}
\pm (R_j+R_j^*) \le \eta \Big( R_4 + \frac{\mathcal{N}^2}{N} \Big)+ \frac{C (1+\mathcal{N})}{\eta(1+t)^3}, \quad \forall \eta>0, \,\forall j=0,1,2,3,
\end{align*}
and
$$ 0\le R_4 \le CN^{3\beta-1} \mathcal{N}^2, \quad R_4\le CN^{\beta-1} {\ensuremath{\rm d}\Gamma}(-\Delta) \mathcal{N}.$$
Here the constant $C$ depends only on $\|u(0,\cdot)\|_{W^{1,\ell}}$ (more precisely, on $\kappa_0$ in the condition $\| u(0,\cdot)\|_{W^{\ell,1}} \le \kappa_0$).
\end{lemma}
\begin{proof} Let us go term by term.
\noindent
$\boxed{j=0}$ Recall that
$$
R_0 = \,{\rm d}\Gamma\Big( Q(t)[w_N*|u(t)|^2+\widetilde{K}_1(t) -\mu_N(t)]Q(t) \Big)\frac{1-\mathcal{N}}{N-1}.
$$
From the operator bounds in \eqref{eq:K1-norm-1}-\eqref{eq:K1-norm-2}-\eqref{eq:K1-norm}, we have
$$
\pm \,{\rm d}\Gamma\Big( Q(t)[w_N*|u(t)|^2+\widetilde{K}_1(t) -\mu_N(t)]Q(t) \Big) \le \frac{C \mathcal{N}}{(1+t)^3}.
$$
Since the left side of the latter inequality commutes with $\mathcal{N}$, we get
\begin{align} \label{eq:R0-final}
\pm R_0 \le \frac{C\mathcal{N}^2}{N(1+t)^3} &\le \eta \frac{\mathcal{N}^2}{N} + \eta^{-1} \frac{C\mathcal{N}^2}{N(1+t)^6} \nonumber\\
&\le \eta \frac{\mathcal{N}^2}{N} + \eta^{-1} \frac{C\mathcal{N}}{(1+t)^6}, \quad \forall \eta>0.
\end{align}
In the last inequality, we have used the fact that $\mathcal{N} \le N$ on $\mathcal{F}_{+}^{\le N}(t)$.
\medskip
\noindent
$\boxed{j=1}$ For every $\Phi\in \mathcal{F}_+^{\le N}(t)$, by the Cauchy-Schwarz inequality we have
\begin{align*}
\left| \langle \Phi, R_{1}\Phi \rangle \right| &= \frac{2}{N-1} \left| \left\langle \Phi, \mathcal{N}\sqrt{N-\mathcal{N}} a\Big( Q(t)[w_N*|u(t)|^2]u(t) \Big) \Phi \right\rangle \right| \\
& \le \frac{2}{N-1} \| \mathcal{N}\sqrt{N-\mathcal{N}} \Phi\| \Big\| a \Big( Q(t)[w_N*|u(t)|^2]u(t) \Big) \Phi \Big\|.
\end{align*}
Now we use the elementary inequality $a^*(v)a(v)\le \|v\|_{L^2}^2 \mathcal{N}$ and
\begin{align*}
\Big\| Q(t)[w_N*|u(t)|^2]u(t) \Big\|_{L^2} \le \| [w_N*|u(t)|^2] u(t)\|_{L^2} \le \frac{C}{(1+t)^{3}}.
\end{align*}
Here the last estimate is \eqref{eq:K1-norm-1}. Thus
\begin{align}\label{eq:R1-final-0}
\left| \langle \Phi, R_{1} \Phi \rangle \right| &\le \frac{C}{N(1+t)^{3/2}} \langle \Phi, \mathcal{N}^2(N-\mathcal{N}) \Phi\rangle^{1/2} \langle \Phi, \mathcal{N} \Phi \rangle^{1/2} \nonumber\\
&\le \frac{\eta}{N} \langle \Phi, \mathcal{N}^2 \Phi \rangle + \frac{C}{\eta(1+t)^3} \langle \Phi, \mathcal{N} \Phi \rangle, \quad \forall \eta>0.
\end{align}
In the last estimate we have used $0\le N-\mathcal{N} \le N$ on $\mathcal{F}_+^{\le N}(t)$. Consequently,
\begin{align*}
\pm \langle \Phi, (R_{1}+R_1^*) \Phi \rangle = \pm 2 \Re \langle \Phi, R_1 \Phi \rangle \le \frac{\eta}{N} \langle \Phi, \mathcal{N}^2 \Phi \rangle + \frac{C}{\eta(1+t)^3} \langle \Phi, \mathcal{N} \Phi \rangle
\end{align*}
for all $\Phi\in \mathcal{F}_+^{\le N}(t)$. Therefore,
\begin{eqnarray} \label{eq:R1-final}
\pm (R_1+R_1^*) \le \eta \frac{\mathcal{N}^2}{N} + \eta^{-1} \frac{C\mathcal{N}}{(1+t)^3}, \quad \forall \eta>0.
\end{eqnarray}
\noindent
$\boxed{j=2}$ For every $\Phi \in \mathcal{F}_+^{\le N}(t)$, we have
\begin{align*}
\langle \Phi, R_2 \Phi \rangle &= \iint K_2(t,x,y) \Big\langle \Phi, a^*_x a^*_y \Big(\frac{\sqrt{(N-\mathcal{N})(N-\mathcal{N}-1)}}{N-1}-1\Big) \Phi \Big\rangle \,{\rm d} x \,{\rm d} y \\
&= \iint \widetilde K_2(t,x,y) \Big\langle \Phi, a^*_x a^*_y \Big(\frac{\sqrt{(N-\mathcal{N})(N-\mathcal{N}-1)}}{N-1}-1\Big) \Phi \Big\rangle \,{\rm d} x \,{\rm d} y.
\end{align*}
Here we can replace $K_2(t)=Q(t)\otimes Q(t) \widetilde K_2(t)$ by $\widetilde K_2(t)$, namely replace the projection $Q(t)$ by the identity, because $\Phi$ belongs to the excited Fock space $\mathcal{F}_+(t)$ (putting differently, this is because $a(u)\Phi=0$). By the Cauchy-Schwarz inequality, we can estimate
\begin{align*}
\left| \langle \Phi, R_2 \Phi \rangle \right| & \le \iint |u(t,x)| w_N(x-y) |u(t,y)| \| a_x a_y \Phi \| \times \nonumber\\
&\qquad \qquad \qquad \times \left\| \Big(\frac{\sqrt{(N-\mathcal{N})(N-\mathcal{N}-1)}}{N-1}-1\Big) \Phi \right\| \,{\rm d} x \,{\rm d} y.
\end{align*}
Using $0\le \mathcal{N} \le N$ on $\Phi\in \mathcal{F}_+^{\le N}(t)$, it is straightforward to see that
\begin{eqnarray} \label{eq:sqrt-N-N-1}
\left\| \Big(\frac{\sqrt{(N-\mathcal{N})(N-\mathcal{N}-1)}}{N-1}-1\Big) \Phi \right\| \le \frac{C}{\sqrt{N}} \langle \Phi, (1+\mathcal{N}) \Phi\rangle^{1/2}.
\end{eqnarray}
Moreover, by the Cauchy-Schwarz inequality again,
\begin{align*}
&\iint |u(t,x)| w_N(x-y) |u(t,y)| \| a_x a_y \Phi \| \,{\rm d} x \,{\rm d} y \nonumber\\
&\le \left( \iint |u(t,x)|^2 w_N(x-y) |u(t,y)|^2 \,{\rm d} x \,{\rm d} y \right)^{1/2} \nonumber \\
&\qquad \times \left( \iint w_N(x-y) \| a_x a_y \Phi \|^2 \,{\rm d} x \,{\rm d} y \right)^{1/2} \le \frac{C\sqrt{N}}{(1+t)^{3/2}} \langle \Phi, R_4 \Phi\rangle^{1/2}.
\end{align*}
In the last estimate we have used \eqref{eq:K1-norm-2} and the definition of $R_4$. Thus
\begin{align}
\left| \langle \Phi, R_2\Phi \rangle \right| &\le \frac{C}{(1+t)^{3/2}} \langle \Phi, R_4 \Phi\rangle^{1/2} \langle \Phi, (1+\mathcal{N}) \Phi\rangle^{1/2} \nonumber \\
& \le \eta \langle \Phi, R_4 \Phi\rangle + \frac{C}{\eta (1+t)^3} \langle \Phi, (1+\mathcal{N}) \Phi\rangle, \quad \forall \eta>0. \label{eq:R2-final-0}
\end{align}
Consequently,
\begin{eqnarray} \label{eq:R2-final}
\pm (R_2+R_2^*) \le \eta R_4 + \frac{C (1+\mathcal{N})}{\eta (1+t)^3} , \quad \forall \eta>0.
\end{eqnarray}
\medskip
\noindent
$\boxed{j=3}$ For all $\Phi\in \mathcal{F}_+^{\le N}(t)$, by using the simplification involving the projection $Q(t)$ as above and the Cauchy-Schwarz inequality we have
\begin{align} \label{eq:R3-final-0}
\left| \langle \Phi, R_3 \Phi \rangle \right| &= \frac{1}{N-1} \left| \iint w_N(x-y) \overline{u(t,x)} \Big \langle \Phi,\sqrt{N-\mathcal{N}} a^*_y a_{y} a_x \Phi \Big\rangle \,\,{\rm d} x \,{\rm d} y \right| \nonumber\\
& \le \frac{1}{N-1} \iint w_N(x-y) |u(t,x)| \cdot \| a_y \sqrt{N-\mathcal{N}} \Phi \| \cdot \| a_{y}a_x \Phi \| \,{\rm d} x \,{\rm d} y \nonumber\\
& \le \frac{2\|u(t,\cdot)\|_{L^\infty}}{N-1} \left( \iint w_N(x-y) \|a_x a_y \Phi\|^2 \,{\rm d} x \,{\rm d} y\right)^{1/2} \nonumber \\
&\qquad\qquad\qquad \times \left( \iint w_N(x-y) \| a_y \sqrt{N-\mathcal{N}} \Phi \|^2 \,{\rm d} x \,{\rm d} y \right)^{1/2} \nonumber\\
&\le \frac{C}{(1+t)^{3/2}} \langle \Phi, R_4 \Phi \rangle^{1/2} \langle \Phi, \mathcal{N} \Phi \rangle^{1/2} \nonumber\\
&\le \eta \langle \Phi, R_4 \Phi \rangle + \frac{C}{\eta (1+t)^3} \langle \Phi, \mathcal{N} \Phi \rangle, \quad \forall \eta>0.
\end{align}
Thus we conclude that
\begin{eqnarray} \label{eq:R3-final}
\pm (R_3+R_3^*) \le \eta R_4 + \frac{C\mathcal{N}}{\eta(1+t)^3} , \quad \forall \eta>0.
\end{eqnarray}
Collecting \eqref{eq:R0-final}, \eqref{eq:R1-final}, \eqref{eq:R2-final} and \eqref{eq:R3-final} gives us the first bound in Lemma \ref{lem:Rj}.
\medskip
\noindent
$\boxed{j=4}$ The simple estimate $0\le R_4\le N^{3\beta-1}\mathcal{N}^2$ follows from the uniform bound $0\le w_N\le CN^{3\beta}$. Moreover, by Sobolev's inequality, we have
$$ w_N(x-y) \le C\|w_N\|_{L^{3/2}} (-\Delta_x) \le CN^{\beta } (-\Delta_x-\Delta_y)$$
as quadratic form on $\mathfrak{H}^2$ (see e.g. \cite[Lemma 3.2]{NamRouSei-15} for a proof). Therefore,
$$
R_4 \le CN^{\beta-1} \iint (-\Delta_x -\Delta_y) a_x^* a_y^* a_x a_y \,{\rm d} x \,{\rm d} y \le CN^{\beta-1} {\ensuremath{\rm d}\Gamma}(-\Delta) \mathcal{N}.
$$
This completes the proof of Lemma \ref{lem:Rj}.
\end{proof}
Heuristically, the first estimate in Lemma \ref{lem:Rj} tells us that $R_4$ is the main error term among all $R_j$'s. The simple bound $0\le R_4\le N^{3\beta-1}\mathcal{N}^2$ can serve as a-priori estimate, but it is not sufficient when $\beta>1/3$. On the other hand, in order to use the bound $R_4\le CN^{\beta-1} {\ensuremath{\rm d}\Gamma}(-\Delta) \mathcal{N}$ we need to control the kinetic energy $\langle \Phi_N(t), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi_N(t)\rangle$. We have
\begin{lemma} \label{lem:HN-kinetic} Under the assumptions in Theorem \ref{thm:main}, we have
$$
\big \langle \Phi_N(t), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi_N(t) \big\rangle \le C_\varepsilon N^{\beta+\varepsilon} , \quad \forall t>0, \,\forall \varepsilon\in (0,1-2\beta].
$$
\end{lemma}
The proof of Lemma \ref{lem:HN-kinetic} is similar to that of Lemma \ref{lem:bH-kinetic}. We will need
\begin{lemma} \label{lem:dt-Rj} We have the quadratic form estimates on $\mathcal{F}_+^{\le N}$:
\begin{align*}
\pm \partial_t (R_j+R_j^*) \le \eta \Big( R_4 + \frac{\mathcal{N}^2}{N} \Big)+ \frac{C (1+\mathcal{N})}{\eta(1+t)^3}, \\
\pm i[(R_j+R_j^*),\mathcal{N}] \le \eta \Big( R_4 + \frac{\mathcal{N}^2}{N} \Big)+ \frac{C (1+\mathcal{N})}{\eta(1+t)^3},
\end{align*}
for all $j=0,1,2,3,4$ and $\eta>0$. The constant $C$ depends only on $\|u(0,\cdot)\|_{W^{1,\ell}}$.
\end{lemma}
\begin{proof} First, we bound $i[(R_j+R_j^*),\mathcal{N}]$. If $j=0$ or $j=4$, the commutator is $0$. Moreover, we have
$$ i[R_1,\mathcal{N}]=iR_1,\quad i[R_2,\mathcal{N}]=-2iR_2, \quad i[R_3,\mathcal{N}]=iR_3$$
because $[a_x,\mathcal{N}]=a_x$, $[a_x^*a_y^*,\mathcal{N}]=-2 a_x^* a_y^*$ and $[a_x^*a_y a_z, \mathcal{N}]=a^*_x a_y a_z$, respectively. Thus the desired inequalities can be obtained in the same way as in Lemma \ref{lem:Rj} (more precisely, they follow from \eqref{eq:R1-final-0}, \eqref{eq:R2-final-0} and \eqref{eq:R3-final-0}).
Next, we bound $\partial_t (R_j+R_j^*)$ by proceeding as in the proof of Lemma \ref{lem:Rj}. Let us explain term by term.
\medskip
\noindent
$\boxed{j=0}$ From \eqref{eq:K1-norm-1}-\eqref{eq:K1-norm-2}-\eqref{eq:K1-norm} and \eqref{eq:der-K1-1}-\eqref{eq:der-K1-2}-\eqref{eq:der-K1}, we find that
$$ \Big\| \partial_t \Big( Q(t)[w_N*|u(t)|^2+\widetilde{K}_1(t) -\mu_N(t)]Q(t) \Big) \Big \| \le \frac{C}{(1+t)^3}.$$
Therefore, similarly to \eqref{eq:R0-final}, we have
\begin{align}\label{eq:dt-R0-final}
\pm \partial_t R_0 &= \pm \,{\rm d}\Gamma\Big(\partial_t \Big( Q(t)[w_N*|u(t)|^2+\widetilde{K}_1(t)-\mu_N(t)]Q(t) \Big) \Big)\frac{1-\mathcal{N}}{N-1} \\
& \le \frac{C\mathcal{N}^2}{N(1+t)^{3}} \le \eta \frac{\mathcal{N}^2}{N} + \eta^{-1} \frac{C\mathcal{N}^2}{N(1+t)^6} \le \eta \frac{\mathcal{N}^2}{N} + \eta^{-1} \frac{C\mathcal{N}}{(1+t)^6}, \quad \forall \eta>0 .\nonumber
\end{align}
\noindent
$\boxed{j=1}$ Using $\|Q(t)\|\le 1$, $\|\partial_t Q(t)\| \le C$, \eqref{eq:K1-norm-1} and \eqref{eq:der-K1-1}, we have
\begin{align*}
&\Big\|\partial_t \Big( Q(t)[w_N*|u(t)|^2]u(t) \Big)\Big\|_{L^2} \\
&\le C \| [w_N*|u(t)|^2] u(t)\|_{L^2} + \| [\partial_t[ (w_N* |u(t)|^2) u(t)]\|_{L^2} \le \frac{C}{(1+t)^{3}}.
\end{align*}
Therefore, we can follow the proof of \eqref{eq:R1-final} and obtain
\begin{eqnarray} \label{eq:dt-R1-final}
\pm \partial_t (R_1 + R_1^*) \le \eta \frac{\mathcal{N}^2}{N} + \eta^{-1} \frac{C\mathcal{N}}{(1+t)^3}, \quad \forall \eta>0.
\end{eqnarray}
\noindent
$\boxed{j=2}$ For every $\Phi\in \mathcal{F}_+^{\le N}(t)$, we have
\begin{align*}
&\langle \Phi, \partial_t R_2 \Phi \rangle =\iint \Big[ (\partial_t Q(t) \otimes 1 + 1\otimes \partial_t Q(t)) \widetilde K_2(t,x,y) + \partial_t \widetilde K_2(t,x,y)\Big] \times \\
&\qquad \qquad \qquad \times \Big \langle \Phi, a^*_x a^*_y \Big(\frac{\sqrt{(N-\mathcal{N})(N-\mathcal{N}-1)}}{N-1}-1\Big) \Phi \Big\rangle \,{\rm d} x \,{\rm d} y.
\end{align*}
Here we have used the decomposition
$$
\partial_t K_2(t)= \partial_t Q(t) \otimes Q(t) \widetilde K_2(t) + Q(t) \otimes \partial_t Q(t) \widetilde K_2(t) + Q(t)\otimes Q(t) \partial_t \widetilde K_2(t)
$$
and omitted the projection $Q(t)$ using $\Phi \in \mathcal{F}_+(t)$. Similarly to \eqref{eq:R2-final-0}, we have
\begin{align*}
& \left| \iint \partial_t \widetilde K_2(t,x,y) \Big \langle \Phi, a^*_x a^*_y \Big (\frac{\sqrt{(N-\mathcal{N})(N-\mathcal{N}-1)}}{N-1}-1 \Big) \Phi \Big\rangle \,{\rm d} x \,{\rm d} y \right| \nonumber\\
&\le \eta \langle \Phi, R_4 \Phi\rangle + \frac{C}{\eta (1+t)^3} \langle \Phi, (1+\mathcal{N}) \Phi\rangle, \quad \forall \eta>0.
\end{align*}
The term involving $(\partial_t Q(t) \otimes 1 + 1 \otimes \partial_t Q(t)) \widetilde K_2(t,x,y)$ is bounded as
\begin{align*}
&\left| \iint \big(\partial_t Q(t) \otimes 1+1 \otimes \partial_t Q(t) \big) \widetilde K_2(t,x,y) \times \right. \\
&\qquad \qquad \times \left. \Big \langle \Phi, a^*_x a^*_y \Big(\frac{\sqrt{(N-\mathcal{N})(N-\mathcal{N}-1)}}{N-1}-1\Big) \Phi \Big\rangle \,{\rm d} x \,{\rm d} y \right| \\
&\le \left( \iint \Big| \big(\partial_t Q(t) \otimes 1 +1 \otimes \partial_t Q(t) \big) \widetilde K_2(t,x,y) \Big|^2 \,{\rm d} x \,{\rm d} y \right)^{1/2}\times \nonumber\\
& \qquad\qquad \times \left( \iint \| a_x a_y \Phi\|^2 \,{\rm d} x \,{\rm d} y \right)^{1/2} \left\| \left(\frac{\sqrt{(N-\mathcal{N})(N-\mathcal{N}-1)}}{N-1}-1\right) \Phi \right\| \nonumber \\
&\le \frac{C}{\sqrt{N}(1+t)^{3/2}} \langle \Phi, \mathcal{N}^2 \Phi \rangle^{1/2}\langle \Phi, (1+\mathcal{N})\Phi \rangle^{1/2}.
\end{align*}
Here we have used \eqref{eq:L2-dtQ-K2} and \eqref{eq:sqrt-N-N-1} in the last estimate. In summary,
$$
\left| \langle \Phi, \partial_t R_2 \Phi \rangle \right| \le \eta \Big( \langle \Phi, R_4 \Phi\rangle + \frac{1}{N}\langle \Phi, \mathcal{N}^2 \Phi\rangle\Big) + \frac{C}{\eta (1+t)^3} \langle \Phi, (1+\mathcal{N}) \Phi\rangle
$$
for all $\Phi\in \mathcal{F}_+^{\le N}(t)$ and $\eta>0$. Therefore,
\begin{eqnarray} \label{eq:dt-R2-final}
\pm \partial_t (R_2 + R_2^*) \le \eta \Big( R_4 +\frac{\mathcal{N}^2}{N} \Big)+ \frac{C(1+\mathcal{N})}{\eta(1+t)^{3}}, \quad \forall \eta>0.
\end{eqnarray}
\medskip
\noindent
$\boxed{j=3}$ For all $\Phi\in \mathcal{F}_+^{\le N}(t)$, we have
\begin{align*}
\langle \Phi, \partial_t R_3 \Phi\rangle & = \frac{1}{N-1} \iiiint \Big[ \big( 1 \otimes Q(t) w_N Q(t)\otimes Q(t)\big)(x,y;x',y') \overline{\partial_t u(t,x)} \\
&\qquad \qquad + \Big( \partial_t \big( 1 \otimes Q(t) w_N Q(t)\otimes Q(t)\big)\Big)(x,y;x',y') \overline{ u(t,x)} \Big] \times \\
& \qquad \qquad \qquad \qquad\qquad \times \langle \Phi, \sqrt{N-\mathcal{N}} a^*_y a_{x'} a_{y'} \Phi \rangle \,\,{\rm d} x \,{\rm d} y \,{\rm d} x' \,{\rm d} y'.
\end{align*}
The term involving $\partial_t u(t,x)$ can be estimated similarly to \eqref{eq:R3-final-0}:
\begin{align*}
&\frac{1}{N-1} \left| \iiiint ( 1 \otimes Q(t) w_N Q(t)\otimes Q(t))(x,y;x',y') \overline{\partial_t u(t,x)} \times \right. \\
& \qquad \qquad \qquad \qquad \left. \times \langle \Phi, \sqrt{N-\mathcal{N}} a^*_y a_{x'} a_{y'} \Phi \rangle \,\,{\rm d} x \,{\rm d} y \,{\rm d} x' \,{\rm d} y' \right| \\
&\le \eta \langle \Phi, R_4 \Phi \rangle + \frac{C}{\eta (1+t)^3} \langle \Phi, \mathcal{N} \Phi \rangle, \quad \forall \eta>0.
\end{align*}
In the following, we will use the kernel estimate
\begin{eqnarray} \label{eq:kernel-dtQ}
|(\partial_t Q(t))(z;z')|= |\partial_t u(t,z) \overline{u(t,z')}+ u(t,z) \overline{\partial_t u(t,z')}| \le q(z) q(z')
\end{eqnarray}
where $q(t,z):=|u(t,z)| + |\partial_t u(t,z)|.$ Recall that by Lemma \ref{lem:Hartree-equation},
$$
\|q(t, \cdot)\|_{L^2} \le C, \quad \|q(t,\cdot)\|_{L^\infty} \le \frac{C}{(1+t)^{3/2}}.
$$
Let us decompose $\partial_t \big( 1 \otimes Q(t) w_N Q(t)\otimes Q(t)\big)$ into three terms. For the first term $1 \otimes \partial_t Q(t) w_N Q(t)\otimes Q(t)$, we can estimate
\begin{align*}
&\frac{1}{N-1} \left| \iiiint (1 \otimes \partial_t Q(t) w_N Q(t)\otimes Q(t)) (x,y;x',y') \overline{u(t,x)} \times \right.\\
& \qquad \qquad \qquad \qquad \qquad \qquad\left. \times \Big\langle \Phi, \sqrt{N-\mathcal{N}} a^*_y a_{x'} a_{y'} \Phi \Big\rangle \,\,{\rm d} x \,{\rm d} y \,{\rm d} x' \,{\rm d} y' \right|\\
&=\frac{1}{N-1} \left| \iiiint (\partial_t Q(t))(y;y') w_N(x-y') \delta(x-x') \overline{ u(t,x)} \times \right.\\
&\qquad \qquad \qquad \qquad \qquad \qquad \left.\times \Big \langle \Phi,\sqrt{N-\mathcal{N}} a^*_y a_x a_{y'} \Phi \Big\rangle \,\,{\rm d} x \,{\rm d} y \,{\rm d} x' \,{\rm d} y' \right| \nonumber \\
&\le \frac{1}{N-1} \iiint q(t,y) q(t,y') w_N(x-y') |u(t,x)| \times \\
&\qquad \qquad \qquad \qquad \qquad \qquad \times \| a_y \sqrt{N-\mathcal{N}} \Phi\| \|a_x a_{y'} \Phi \|\,\,{\rm d} x \,{\rm d} y \,{\rm d} y' \nonumber \\
&\le \frac{\|q(t,\cdot)\|_{L^\infty}}{N-1} \left( \int |q(t,y)|^2 \,{\rm d} y\right)^{1/2} \left( \int \| a_y \sqrt{N-\mathcal{N}} \Phi \|^2 \,{\rm d} y\right)^{1/2} \times \\
& \times \left( \iint w_N(x-y') |u(t,x)|^2 \,{\rm d} x \,{\rm d} y'\right)^{1/2} \left( \iint w_N(x-y') \| a_x a_{y'} \Phi\|^2 \,{\rm d} x \,{\rm d} y'\right)^{1/2} \\
&\le \frac{C}{(1+t)^{3/2}} \langle \Phi, R_4 \Phi \rangle^{1/2} \langle \Phi, \mathcal{N} \Phi\rangle^{1/2} \le \eta \langle \Phi, R_4 \Phi \rangle + \frac{C\langle \Phi, \mathcal{N} \Phi \rangle}{\eta (1+t)^3} , \quad \forall \eta>0.
\end{align*}
For the second term $1 \otimes Q(t) w_N \partial_t Q(t) \otimes Q(t)$, we have
\begin{align*}
&\frac{1}{N-1} \left| \iiiint (1 \otimes Q(t) w_N \partial_t Q(t) \otimes Q(t))(x,y;x',y') \overline{u(t,x)} \times \right. \\
& \qquad \qquad \qquad \qquad \qquad \qquad \left. \times \Big \langle \Phi, \sqrt{N-\mathcal{N}} a^*_y a_{x'} a_{y'} \Phi \Big\rangle \,\,{\rm d} x \,{\rm d} y \,{\rm d} x' \,{\rm d} y' \right|\\
&= \frac{1}{N-1} \left| \iiiint w_N(x-y) (\partial_tQ(t))(x,x') \delta(y-y') \overline{ u(t,x)} \times \right.\\
&\qquad \qquad \qquad \qquad \qquad \qquad \left.\times \Big \langle \Phi,\sqrt{N-\mathcal{N}} a^*_y a_{x'} a_{y'} \Phi \Big\rangle \,\,{\rm d} x \,{\rm d} y \,{\rm d} x' \,{\rm d} y' \right| \nonumber \\
&\le \frac{1}{N-1} \iiint w_N(x-y) q(t,x) q(t,x') |u(t,x)| \times \\
&\qquad \qquad \qquad \qquad\qquad \qquad \times \| a_y \sqrt{N-\mathcal{N}} \Phi\| \|a_{x'} a_{y} \Phi\| \,\,{\rm d} x \,{\rm d} y \,{\rm d} x' \nonumber \\
&\le \frac{1}{N-1} \|w_N\|_{L^1} \| q(t,\cdot)\|_{L^\infty} \|u(t,\cdot)\|_{L^\infty} \times \\
&\quad \times \left( \iint \|a_{x'} a_{y} \Phi\|^2 \,\,{\rm d} x' \,{\rm d} y \right)^{1/2} \left( \iint |q(t,x')|^2 \| a_y \sqrt{N-\mathcal{N}} \Phi\|^2 \,{\rm d} x' \,{\rm d} y \right)^{1/2} \\
&\le \frac{C}{\sqrt{N}(1+t)^3} \langle \Phi, \mathcal{N}^2 \Phi\rangle^{1/2} \langle \Phi, \mathcal{N} \Phi\rangle^{1/2} \le \eta \frac{\langle \Phi, \mathcal{N}^2 \Phi\rangle}{N} + \frac{C\langle \Phi, \mathcal{N} \Phi\rangle}{\eta(1+t)^3}, \,\, \forall \eta>0.
\end{align*}
The third term $1 \otimes Q(t) w_N Q(t) \otimes \partial_t Q(t)$ is bounded similarly. Thus
\begin{align*}
\left| \langle \Phi, \partial_t R_3 \Phi \rangle\right| \le \eta \left( \langle \Phi, R_4 \Phi \rangle + \frac{\langle \Phi, \mathcal{N}^2 \Phi\rangle}{N} \right) + \frac{C\langle \Phi, \mathcal{N} \Phi\rangle}{\eta(1+t)^3}
\end{align*}
for all $\Phi\in \mathcal{F}_+^{\le N}(t)$ and $\eta>0$. Consequently,
\begin{eqnarray} \label{eq:dt-R3-final}
\pm \partial_t (R_3+R_3^*) \le \eta \left( R_4 +\frac{\mathcal{N}^2}{N} \right)+ \frac{C\mathcal{N}}{\eta(1+t)^3}, \quad \forall \eta>0.
\end{eqnarray}
\noindent
$\boxed{j=4}$ For all $\Phi\in \mathcal{F}_+^{\le N}(t)$, we have
\begin{align*}
&\langle \Phi, \partial_t R_4\Phi\rangle = \frac{1}{2(N-1)} \Re \iiiint \partial_t \Big ( Q(t) \otimes Q(t) w_N Q(t)\otimes Q(t)\Big)(x,y;x',y') \\
& \qquad \qquad \qquad \qquad \qquad \qquad\qquad \qquad\times \langle \Phi, a_x^* a^*_y a_{x'} a_{y'} \Phi \rangle \,\,{\rm d} x \,{\rm d} y \,{\rm d} x' \,{\rm d} y'.
\end{align*}
Let us decompose $\partial_t \big ( Q(t) \otimes Q(t) w_N Q(t)\otimes Q(t)\big)$ into four terms, and consider for example $\partial_t Q(t) \otimes Q(t) w_N Q(t)\otimes Q(t)$. Using \eqref{eq:kernel-dtQ} again, we have
\begin{align*}
&\frac{1}{N-1} \left| \iiiint \Big(\partial_t Q(t) \otimes Q(t) w_N Q(t)\otimes Q(t)\Big)(x,y;x',y')\times \right.\\
& \qquad \qquad \qquad \qquad \qquad \qquad \qquad \qquad \left. \times \Big\langle \Phi, a_x^* a^*_y a_{x'} a_{y'} \Phi \Big\rangle \,\,{\rm d} x \,{\rm d} y \,{\rm d} x' \,{\rm d} y' \right|\\
&= \frac{1}{N-1} \left| \iiiint (\partial_t Q(t))(x,x') w_N(x'-y) \delta(y-y')\times \right. \\
& \qquad \qquad \qquad \qquad \qquad \qquad\qquad \qquad \left. \times \Big\langle \Phi, a_x^* a^*_y a_{x'} a_{y'} \Phi \Big\rangle \,\,{\rm d} x \,{\rm d} y \,{\rm d} x' \,{\rm d} y' \right|\\
&\le \frac{1}{N-1} \iiint q(t,x) q(t,x') w_N(x'-y) \| a_x a_y \Phi\| \|a_{x'} a_{y} \Phi\| \,\,{\rm d} x \,{\rm d} y \,{\rm d} x'\\
&\le \frac{\|q(t,\cdot)\|_{L^\infty}}{N-1} \left( \iiint w_N(x'-y) \| a_x a_y \Phi\|^2 \,{\rm d} x \,{\rm d} y \,{\rm d} x' \right)^{1/2} \times \\
&\qquad \qquad \qquad \qquad \times \left( \iiint w_N(x'-y) \| a_{x'} a_y \Phi\|^2 |q(t,x)|^2 \,{\rm d} x \,{\rm d} y \,{\rm d} x' \right)^{1/2} \\
&\le \frac{C}{N^{1/2}(1+t)^{3/2}} \langle \Phi, \mathcal{N}^2 \Phi\rangle^{1/2} \langle \Phi, R_4 \Phi\rangle^{1/2} \\
& \le \eta \langle \Phi, R_4 \Phi\rangle + \frac{C\langle \Phi, \mathcal{N}^2 \Phi\rangle}{\eta N (1+t)^{3} } \le \eta \langle \Phi, R_4 \Phi\rangle + \frac{C\langle \Phi, \mathcal{N} \Phi\rangle}{\eta (1+t)^{3} }, \quad \forall \eta>0.
\end{align*}
By similar estimates, we find that
\begin{align*}
\left| \langle \Phi, \partial_t R_4\Phi \rangle \right| \le \eta \langle \Phi, R_4 \Phi\rangle + \frac{C\langle \Phi, \mathcal{N} \Phi\rangle}{\eta (1+t)^{3} }
\end{align*}
for all $\Phi\in \mathcal{F}_+^{\le N}(t)$ and $\eta>0$. Thus
\begin{eqnarray} \label{eq:dt-R4-final}
\pm \partial_t R_4 \le \eta R_4 + \frac{C\mathcal{N}}{\eta(1+t)^3}, \quad \forall \eta>0.
\end{eqnarray}
This completes the proof.
\end{proof}
Now we are ready to provide
\begin{proof}[Proof of Lemma \ref{lem:HN-kinetic}] We use the proof strategy of Lemma \ref{lem:bH-kinetic}. Using the equation \eqref{eq:eq-PhiNt} we can write
\begin{align} \label{eq:PhiN-kin-0}
\big\langle \Phi_N(t), \widetilde H_N(t) \Phi_N(t) \big\rangle &- \big\langle \Phi_N(0), \widetilde H_N(0) \Phi_N(0) \big\rangle \nonumber\\
&= \int_0^t \big\langle \Phi_N(s), \partial_s \widetilde H_N(s) \Phi_N(s) \big\rangle \,{\rm d} s
\end{align}
and
\begin{align} \label{eq:PhiN-kin-a}
\big\langle \Phi_N(t), \mathcal{N} \Phi_N(t) \big\rangle - \big\langle \Phi_N(0), \mathcal{N} \Phi_N(0) \big\rangle = \int_0^t \big\langle \Phi_N(s), i[\widetilde H_N(s),\mathcal{N}] \Phi_N(s) \big\rangle \,{\rm d} s.
\end{align}
Let us estimate both sides of \eqref{eq:PhiN-kin-0}. Recall that
$$
\widetilde H_N (t)= \1^{\le N} \Big[ \mathbb{H}(t) + \frac{1}{2}\sum_{j=0}^4 ( R_{j} + R_j^*) \Big] \1^{\le N}.
$$
From Lemma \ref{lem:bHt-dbHt} and Lemma \ref{lem:Rj}, we have the form estimates on $\mathcal{F}_+^{\le N}(t)$:
\begin{align} \label{eq:HNt-upper-lower}
&\pm \1^{\le N} \Big( \widetilde H_N(t) + {\ensuremath{\rm d}\Gamma}(\Delta) - R_4 \Big) \1^{\le N} \nonumber\\
&= \pm \1^{\le N} \Big( \mathbb{H}(t) + {\ensuremath{\rm d}\Gamma}(\Delta) + \frac{1}{2}\sum_{j=0}^3 ( R_{j} + R_j^*) \Big) \1^{\le N} \nonumber\\
&\le \eta\Big({\ensuremath{\rm d}\Gamma}(1-\Delta)+R_4+\frac{\mathcal{N}^2}{N} \Big) + \frac{C_\varepsilon (N^{\beta+\varepsilon} + \mathcal{N})}{\eta(1+t)^3}.
\end{align}
Similarly, from Lemma \ref{lem:bHt-dbHt} and Lemma \ref{lem:dt-Rj}, we have
\begin{align}
\pm \partial_t \widetilde H_N(t) &\le \eta\Big({\ensuremath{\rm d}\Gamma}(1-\Delta)+R_4 + \frac{\mathcal{N}^2}{N}\Big) + \frac{C_\varepsilon (N^{\beta+\varepsilon} + \mathcal{N})}{\eta(1+t)^3}, \label{eq:dt-HNt-upper-lower}
\end{align}
for all $\eta>0$.
Applying \eqref{eq:HNt-upper-lower} with $\eta=1/2$ and using $\mathcal{N}^2/N \le \mathcal{N} \le {\ensuremath{\rm d}\Gamma}(1-\Delta)$ on $\mathcal{F}_+^{\le N}(t)$, we find that
\begin{align} \label{eq:PhiN-kin-1}
\big\langle \Phi_N(t), \widetilde H_N(t) \Phi_N(t) \big\rangle & \ge \frac{1}{2} \big\langle \Phi_N(t), ({\ensuremath{\rm d}\Gamma}(1-\Delta) + R_4) \Phi_N(t) \big\rangle \nonumber\\
&\qquad\qquad- C_\varepsilon \Big(N^{\beta+\varepsilon} + \big\langle \Phi_N(t), \mathcal{N} \Phi_N(t) \big\rangle \Big),\\
\label{eq:PhiN-kin-2-00}
\big\langle \Phi_N(0), \widetilde H_N(0) \Phi_N(0) \big\rangle &\le C \big\langle \Phi_N(0), ({\ensuremath{\rm d}\Gamma}(1-\Delta) + R_4) \Phi_N(0) \big\rangle + C_\varepsilon N^{\beta+\varepsilon}.
\end{align}
Using $\Phi_N(0)=\1^{\le N}\Phi(0)$, we get
$$ \big\langle \Phi_N(0), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi_N(0) \big\rangle \le \big\langle \Phi(0), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi(0) \big\rangle \le C_\varepsilon N^{\beta+\varepsilon}. $$
On the other hand, recall that $R_4\le CN^{3\beta-1}\mathcal{N}^2$ by Lemma \ref{lem:Rj}. Moreover, it is well-known that for every quasi-free state $\Phi$ and $s\in \mathbb{N}$, we have
\begin{eqnarray} \label{eq:moment-quasi-free}
\Big\langle \Phi, (1+\mathcal{N})^{s} \Phi \Big\rangle \le C_{s} \Big\langle \Phi, (1+\mathcal{N}) \Phi \Big\rangle^{s}
\end{eqnarray}
where the constant $C_s$ depends only on $s$ (see \cite[Lemma 5]{NamNap-15} for a proof). Combining with the assumptions $\big\langle \Phi(0), \mathcal{N} \Phi(0)\big \rangle\le C_\varepsilon N^\varepsilon$ and $\varepsilon \le 1-2\beta$, we obtain
\begin{align*}
\big\langle \Phi_N(0), R_4 \Phi_N(0) \big\rangle &\le CN^{3\beta-1} \big\langle \Phi(0), \mathcal{N}^2 \Phi(0) \big\rangle \le CN^{3\beta-1} \big\langle \Phi(0), \mathcal{N} \Phi(0)\big \rangle^2 \\
& \le C_\varepsilon N^{3\beta-1} N^{2\varepsilon} \le C_\varepsilon N^{\beta+\varepsilon}.
\end{align*}
Thus \eqref{eq:PhiN-kin-2-00} reduces to
\begin{align} \label{eq:PhiN-kin-2}
\big\langle \Phi_N(0), \widetilde H_N(0) \Phi_N(0) \big\rangle \le C_\varepsilon N^{\beta+\varepsilon}.
\end{align}
Next, we apply \eqref{eq:dt-HNt-upper-lower} with $\eta=(1+t)^{-3/2}$ and use $\mathcal{N}^2/N\le \mathcal{N} \le {\ensuremath{\rm d}\Gamma}(1-\Delta)$ on $\mathcal{F}_+^{\le N}(t)$. This gives
\begin{align} \label{eq:PhiN-kin-3}
\big\langle \Phi_N(t), \partial_t \widetilde H_N(t) \Phi_N(t) \big\rangle & \le C_\varepsilon \frac{\big\langle \Phi_N(t), ({\ensuremath{\rm d}\Gamma}(1-\Delta) + R_4) \Phi_N(t) \big\rangle+N^{\beta+\varepsilon}}{(1+t)^{3/2}}.
\end{align}
Inserting \eqref{eq:PhiN-kin-1}, \eqref{eq:PhiN-kin-2} and \eqref{eq:PhiN-kin-3} into \eqref{eq:PhiN-kin-0} we obtain
\begin{align} \label{eq:PhiN-kin-4}
\big\langle \Phi_N(t), ({\ensuremath{\rm d}\Gamma}(1-\Delta) + R_4) \Phi_N(t) \big\rangle &\le C_\varepsilon \int_0^t \frac{\big\langle \Phi_N(s), ({\ensuremath{\rm d}\Gamma}(1-\Delta) + R_4) \Phi_N(s) \big\rangle}{(1+s)^{3/2}} \,{\rm d} s \nonumber
\\
&+ C_\varepsilon \Big( N^{\beta+\varepsilon} + \big\langle \Phi_N(t), \mathcal{N} \Phi_N(t) \big\rangle \Big).
\end{align}
Now we consider \eqref{eq:PhiN-kin-a}. By using $\Phi(0)=\1^{\le N}\Phi(0)$ and the assumption on $\Phi(0)$, we have
$$ \langle \Phi_N(0), \mathcal{N} \Phi_N(0) \rangle \le \langle \Phi(0), \mathcal{N} \Phi(0) \rangle \le C_\varepsilon N^{\beta+\varepsilon}.$$
Moreover, from Lemma \ref{lem:bHt-dbHt} and Lemma \ref{lem:dt-Rj}, we have
$$
\pm i [\widetilde H_N(t), \mathcal{N}] \le \eta\Big({\ensuremath{\rm d}\Gamma}(1-\Delta)+R_4 + \frac{\mathcal{N}^2}{N}\Big) + \frac{C_\varepsilon (N^{\beta+\varepsilon} + \mathcal{N})}{\eta(1+t)^3}, \quad \forall \eta>0.
$$
We can choose $\eta=(1+t)^{-3/2}$ and use $\mathcal{N}^2/N \le \mathcal{N} \le {\ensuremath{\rm d}\Gamma}(1-\Delta)$ on $\mathcal{F}_+^{\le N}(t)$ to obtain
$$
\pm i [\widetilde H_N(t), \mathcal{N}] \le C_\varepsilon \frac{{\ensuremath{\rm d}\Gamma} (1-\Delta)+R_4 + N^{\beta+\varepsilon}}{(1+t)^{3/2}}.
$$
Inserting the latter estimate into the right side of \eqref{eq:PhiN-kin-a}, we get
\begin{eqnarray} \label{eq:PhiN-kin-b}
\langle \Phi(t), \mathcal{N} \Phi(t) \rangle \le C_\varepsilon \int_0^t \frac{\big\langle \Phi_N(s), ({\ensuremath{\rm d}\Gamma}(1-\Delta) + R_4) \Phi_N(s) \big\rangle}{(1+s)^{3/2}} \,{\rm d} s + C_\varepsilon N^{\beta+\varepsilon}.
\end{eqnarray}
Finally, we substitute \eqref{eq:PhiN-kin-b} into the right side of \eqref{eq:PhiN-kin-4} and find that
\begin{align} \label{eq:PhiN-kin-6}
&\big\langle \Phi_N(t), ({\ensuremath{\rm d}\Gamma}(1-\Delta) + R_4) \Phi_N(t) \big\rangle \nonumber\\
&\le C_\varepsilon \int_0^t \frac{\big\langle \Phi_N(s), ({\ensuremath{\rm d}\Gamma}(1-\Delta) + R_4) \Phi_N(s) \big\rangle}{(1+s)^{3/2}} \,{\rm d} s + C_\varepsilon N^{\beta+\varepsilon} .
\end{align}
This bound is similar to \eqref{eq:Phit-Grw-1} and we can argue as in the proof of Lemma \ref{lem:bH-kinetic} to conclude that
$$
\big\langle \Phi_N(t), ({\ensuremath{\rm d}\Gamma}(1-\Delta) + R_4) \Phi_N(t) \big\rangle \le C_\varepsilon N^{\beta+\varepsilon}.
$$
Since $R_4\ge 0$, the desired kinetic estimate follows.
\end{proof}
\section{Proof of the main theorem} \label{sec:main-proof}
\begin{proof}[Proof of Theorem \ref{thm:main}] It suffices to consider the case when $N$ is large and $\varepsilon$ is small (indeed, if the desired bound holds for some $\varepsilon>0$, then it also holds for any $\varepsilon'>\varepsilon$ because $(1+t)^{1+\varepsilon} N^{(2\beta+\varepsilon-1)/2} \le (1+t)^{1+\varepsilon'} N^{(2\beta+\varepsilon'-1)/2}$). In particular, we will assume $\varepsilon<\min \{1/2,1-2\beta\}$.
Since $U_N(t)$ is a unitary operator from $\mathfrak{H}^N$ to $\mathcal{F}_+^{\le N}(t)\subset \mathcal{F}(\mathfrak{H})$, we have
\begin{align} \label{eq:final-proof-1}
\| \Psi_N(t) - U_N(t)^* \1^{\le N} \Phi(t)\|_{\mathfrak{H}^N} &= \| U_N(t) \Psi_N(t) - \1^{\le N} \Phi(t)\| \\
&= \| \1^{\le N}( \Phi_{N}(t) - \Phi(t) )\| \le \| \Phi_{N}(t)- \Phi(t)\|. \nonumber
\end{align}
Using the equations \eqref{eq:Bogoliubov-equation} and \eqref{eq:eq-PhiNt}, we can compute
\begin{align*}
\partial_t \| \Phi_{N}(t)-\Phi(t)\|^2 &= - 2\Re \, \partial_t \big\langle \Phi_{N}(t), \Phi(t) \big\rangle \nonumber\\
&= - 2\Re \Big( \big\langle \partial_t \Phi_N(t), \Phi(t) \big\rangle +\big \langle \Phi_N(t), \partial_t \Phi(t)\big \rangle \Big) \nonumber \\
& = - 2\Re \Big( \big\langle -i \widetilde H_N(t) \Phi_N(t), \Phi(t)\big \rangle +\big \langle \Phi_N(t), -i \mathbb{H}(t) \Phi(t)\big \rangle \Big) \nonumber \\
& = 2\Re \,\big \langle i \Phi_N(t), (\widetilde H_N(t) - \mathbb{H}(t))\Phi(t)\big\rangle .
\end{align*}
Since
$$
\widetilde H_N (t)= \1^{\le N} \Big[ \mathbb{H}(t) + \frac{1}{2}\sum_{j=0}^4 ( R_{j} + R_j^*) \Big] \1^{\le N}
$$
and $\Phi_N \in \mathcal{F}_+^{\le N}(t)$, we can define $\1^{>N} := \1 - \1^{\le N}$ and write
\begin{align} \label{eq:final-proof-2}
\partial_t \| \Phi_{N}(t)-\Phi(t)\|^2 &= \sum_{j=0}^4 \Re \big\langle i\Phi_N(t), (R_j+R_j^*) \1^{\le N} \Phi(t) \big\rangle\nonumber\\
&\qquad \qquad \qquad - 2\Re\,\big\langle i\Phi_N(t), \mathbb{H} \1^{>N} \Phi(t) \big\rangle.
\end{align}
\medskip
\noindent
{\bf Step 1.} Let us consider the last term of \eqref{eq:final-proof-2}. Since $\Phi_N(t) \in \mathcal{F}_+^{\le N}(t)$ and $\1^{\le N}{\ensuremath{\rm d}\Gamma}(h) \1^{>N}=0$, we have
$$
\big\langle \Phi_N(t), \mathbb{H} \1^{>N} \Phi(t) \big\rangle = \big\langle \Phi_N(t), (\mathbb{H} - {\ensuremath{\rm d}\Gamma}(h)) \1^{>N} \Phi(t) \big\rangle.
$$
Recall that by \eqref{eq:bound-paring-dG1},
\begin{eqnarray} \label{eq:final-proof-simplebH}
\pm (\mathbb{H} - {\ensuremath{\rm d}\Gamma}(h)) \le C( \mathcal{N} + N^{3\beta}).
\end{eqnarray}
Here $C$ is a general constant depending only on $\|u(0,\cdot)\|_{W^{\ell,1}}$ (more precisely, on $\kappa_0$ in the condition $\| u(0,\cdot)\|_{W^{\ell,1}} \le \kappa_0$).
We will use the general fact that if $A$ and $B$ are quadratic forms satisfying $\pm B\le A$, then for all vectors $f,g$ we have the Cauchy-Schwarz type inequality
\begin{align} \label{eq:final-proof-CS}
|\langle f, B g\rangle| & \le | \langle f, (A+B) g \rangle | + |\langle f, A g\rangle | \nonumber\\
&\le \langle f, (A+B) f \rangle^{1/2} \langle g, (A+B) g \rangle^{1/2} + \langle f, A f\rangle^{1/2}\langle g, A g\rangle^{1/2} \nonumber\\
&\le 3 \langle f, A f\rangle^{1/2}\langle g, A g\rangle^{1/2}.
\end{align}
From \eqref{eq:final-proof-simplebH} and \eqref{eq:final-proof-CS}, we obtain
\begin{align} \label{eq:final-proof-bH>N}
&\left| \big\langle \Phi_N(t), \mathbb{H} \1^{>N} \Phi(t) \big\rangle \right| = \left| \big\langle \Phi_N(t), (\mathbb{H} - {\ensuremath{\rm d}\Gamma}(h)) \1^{>N} \Phi(t) \big\rangle \right| \nonumber \\
&\le C \Big\langle \Phi_N (t), ( \mathcal{N} + N^{3\beta}) \Phi_N(t) \Big\rangle^{1/2} \Big\langle \1^{>N}\Phi (t), ( \mathcal{N} + N^{3\beta}) \1^{>N}\Phi(t) \Big\rangle^{1/2} \nonumber\\
& \le C (N+N^{3\beta}) \Big\langle \Phi (t), \mathcal{N}^s N^{-s} \Phi(t) \Big\rangle^{1/2}, \quad \forall s \ge 1.
\end{align}
Here, in the last inequality, we have used $\mathcal{N} \le N$ on $\mathcal{F}_+^{\le N}(t)$ and
$$\1^{>N} (\mathcal{N} + N^{3\beta}) \le (N+N^{3\beta}) \mathcal{N}^{s}N^{-s}, \quad \forall s\ge 1.$$
Now we use the moment estimate \eqref{eq:moment-quasi-free}, the bound on $\langle \Phi(t), \mathcal{N} \Phi(t)\rangle$ in Lemma \ref{lem:Bogoliubov-equation} and the assumption $\langle \Phi(0), \mathcal{N} \Phi(0)\rangle \le C_\varepsilon N^{\varepsilon}$. All this gives
\begin{align} \label{eq:final-proof-moment}
\big\langle \Phi(t), (1+\mathcal{N})^{s} \Phi(t) \big\rangle \le C_s \big\langle \Phi(t), (1+\mathcal{N}) \Phi(t) \big\rangle^{s} \le C_{\varepsilon,s} N^{2s\varepsilon} [\log(2+t)]^{2s}.
\end{align}
Hence, \eqref{eq:final-proof-bH>N} reduces to
\begin{align*}
\left| \big\langle \Phi_N(t), \mathbb{H} \1^{>N} \Phi(t) \big\rangle \right| \le C_{\varepsilon,s} (N+N^{3\beta}) N^{s(\varepsilon-1/2)} [\log(2+t)]^{s}
\end{align*}
for all $s\ge 1$. Since $\varepsilon-1/2<0$, we can choose $s=s(\varepsilon)$ sufficiently large (e.g. $s \ge (2+3\beta)/(1/2-\varepsilon)$) to obtain
\begin{align} \label{eq:final-proof-bH>N-last}
\left| \big\langle \Phi_N(t), \mathbb{H} \1^{>N} \Phi(t) \big\rangle \right| &\le C_{\varepsilon} N^{-1} (1+t)^\varepsilon.
\end{align}
Here we have bound $[\log(2+t)]^{s_\varepsilon}$ by $C_\varepsilon (1+t)^\varepsilon$ for simplicity.
\medskip
\noindent{\bf Step 2.} Now we turn to the first term on the right side of \eqref{eq:final-proof-2}. Recall that by Lemma \ref{lem:Rj}, we have the quadratic form estimates on $\mathcal{F}_+^{\le N}(t)$:
\begin{eqnarray} \label{eq:final-proof-Rj}
\pm (R_j+R_j^*) \le 2(1+ \eta) R_4 + \eta \frac{\mathcal{N}^2}{N} + \frac{C (1+\mathcal{N})}{\eta(1+t)^3}
\end{eqnarray}
for all $j=0,1,2,3,4$ and $\eta> 0$ (the bound for $j=4$ does not follow from Lemma \ref{lem:Rj} but it is trivial).
Since we do not have a good control on $\langle \Phi_N(t), \mathcal{N} \Phi_N(t)\rangle$, we need to introduce a cut-off before applying \eqref{eq:final-proof-Rj}. Note that for every $4<M<N-2$,
$$ \1^{\le M} (R_j+R_j^*)\1^{> M+2} =0 \quad\text{and}\quad \1^{> M} (R_j+R_j^*)\1^{\le M-2} =0$$
because there are at most 2 creation or annihilation operators in the expressions of $R_j$'s. Therefore, we can write
\begin{align*}
\big\langle \Phi_N(t), (R_{j} +R_j^*)\1^{\le N} \Phi(t) \big\rangle &= \big\langle \1^{\le M} \Phi_N(t), (R_{j}+R_j^*) \1^{\le M+2} \Phi(t) \big\rangle \\
&\quad + \big\langle \1^{> M} \Phi_N(t), (R_{j} +R_j^*) \1^{\le N} \1^{> M-2} \Phi(t) \big\rangle\nonumber
\end{align*}
and then apply \eqref{eq:final-proof-Rj} and \eqref{eq:final-proof-CS} to each term on the right side. This gives
\begin{align} \label{eq:final-proof-3}
\left| \big\langle \Phi_N(t), (R_{j} +R_j^*)\1^{\le N} \Phi(t) \big\rangle \right| \le C(E_1 + E_2)
\end{align}
where
\begin{align*}
E_1&= \inf_{\eta>0} \left\langle \1^{\le M} \Phi_N(t), \Big( (1+\eta) R_4 + \eta \frac{\mathcal{N}^2}{N} + \frac{1+\mathcal{N}}{\eta(1+t)^3} \Big) \1^{\le M} \Phi_N(t) \right\rangle^{1/2} \nonumber\\
& \qquad \times \left\langle \1^{\le M+2} \Phi(t), \Big( (1+\eta) R_4 + \eta \frac{\mathcal{N}^2}{N} + \frac{1+\mathcal{N}}{\eta(1+t)^3} \Big) \1^{\le M+2} \Phi(t) \right\rangle^{1/2}, \nonumber \\
E_2& = \inf_{\eta>0} \left\langle \1^{> M} \Phi_N(t), \Big( (1+\eta) R_4 + \eta \frac{\mathcal{N}^2}{N} + \frac{1+\mathcal{N}}{\eta(1+t)^3} \Big) \1^{> M} \Phi_N(t) \right\rangle^{1/2} \\
&\qquad \times \left\langle \1^{> M-2} \Phi(t), \Big((1+\eta) R_4 + \eta \frac{\mathcal{N}^2}{N} + \frac{1+\mathcal{N}}{\eta(1+t)^3} \Big) \1^{> M-2} \Phi(t) \right\rangle^{1/2} .
\end{align*}
To bound $E_1$, we use the last estimate in Lemma \ref{lem:Rj} and $\1^{\le M}\mathcal{N} \le M$:
$$ \1^{\le M} R_4 \le C N^{\beta-1} \1^{\le M} \mathcal{N} {\ensuremath{\rm d}\Gamma}(-\Delta) \le CN^{\beta-1}M {\ensuremath{\rm d}\Gamma}(-\Delta).$$
Moreover, recall that we have the kinetic estimate in Lemma \ref{lem:HN-kinetic}:
$$\langle \Phi_N(t), {\ensuremath{\rm d}\Gamma}(1-\Delta) \Phi_N(t) \rangle \le C_\varepsilon N^{\beta+\varepsilon}, \quad \forall \varepsilon \in (0,1-2\beta],$$
where the constant $C_\varepsilon$ is independent of $N$ and $t$. Therefore,
\begin{align*}
&\left\langle \1^{\le M} \Phi_N(t), \Big( (1+\eta) R_4 + \eta \frac{\mathcal{N}^2}{N} + \frac{1+\mathcal{N}}{\eta(1+t)^3} \Big) \1^{\le M} \Phi_N(t) \right\rangle \nonumber\\
& \le C_\varepsilon \Big( (1+\eta) N^{\beta-1} M N^{\beta+\varepsilon} + \eta M^2 N^{-1} + \frac{M}{\eta(1+t)^3} \Big) .
\end{align*}
Similarly, we have the same bound with $\1^{\le M} \Phi_N(t)$ replaced by $\1^{\le M+2} \Phi(t)$ (the kinetic estimate for $\Phi(t)$ is provided in Lemma \ref{lem:bH-kinetic}). Then by optimizing over $\eta>0$ we find that
\begin{align*}
E_1\le C_\varepsilon \Big( M N^{(2\beta+\varepsilon-1)/2} + M^{3/2} N^{-1/2} \Big) .
\end{align*}
Next, we bound $E_2$ using the argument in Step 1. To be precise, let us choose $\eta=1$ in the variational formula of $E_2$ for simplicity and then use $R_4 \le C N^{3\beta-1}\mathcal{N}^2$ (see Lemma \ref{lem:Rj}). We obtain the rough bound
\begin{align*}
E_2 & \le C N^{3\beta} \left\langle \1^{> M} \Phi_N(t), (\mathcal{N} +1)^2 \1^{> M} \Phi_N(t) \right\rangle^{1/2} \\
&\qquad \qquad \qquad \times \left\langle \1^{> M-2} \Phi(t), (\mathcal{N} +1)^2 \1^{> M-2} \Phi(t) \right\rangle^{1/2} .
\end{align*}
Now for the first term we use $\1^{\le N} (\mathcal{N} +1 )\le N+1$ (recall that $\Phi_N(t) \in \mathcal{F}_+^{\le N}(t)$) and get
\begin{align*}
\left\langle \1^{> M} \Phi_N(t), (\mathcal{N} +1)^2 \1^{> M} \Phi_N(t) \right\rangle \le (N+1)^{2}.
\end{align*}
For the second term, we use $\1^{>M-2} (\mathcal{N}+1)^2 \le (\mathcal{N}+1)^{s} (M-1)^{2-s}$ with $s \ge 2$ and then use the moment estimate \eqref{eq:final-proof-moment}. We find that
\begin{align*}
\left\langle \1^{> M-2} \Phi(t), (\mathcal{N} +1)^2 \1^{> M-2} \Phi(t) \right\rangle &\le (M-1)^{2-s} \left\langle \Phi(t), (\mathcal{N} +1)^s \Phi(t) \right\rangle \\
&\le C_{\varepsilon,s} (M-1)^{2-s} N^{2s\varepsilon} [\log(2+t)]^{2s}.
\end{align*}
All this yields
$$
E_2 \le C_{\varepsilon,s}N^{3\beta+1} M^{1-s/2} N^{s\varepsilon} [\log(2+t)]^{s}.
$$
In summary, from \eqref{eq:final-proof-3} it follows that
\begin{align*}
\left| \big\langle \Phi_N(t), (R_{j} +R_j^*)\1^{\le N} \Phi(t) \big\rangle \right| &\le C_\varepsilon \Big( M N^{(2\beta+\varepsilon-1)/2} + M^{3/2} N^{-1/2} \Big) \\
& + C_{\varepsilon,s}N^{3\beta+1} M^{1-s/2}N^{s\varepsilon} [\log(2+t)]^{s}
\end{align*}
for all $4<M<N-2$ and $s\ge 2$. We can choose $M=N^{3\varepsilon}$ and $s=s(\varepsilon)$ sufficiently large (e.g. $s\ge 6 (1+\beta+\varepsilon)/\varepsilon$) to obtain
\begin{align} \label{eq:final-proof-2a}
\left| \big\langle \Phi_N(t), (R_{j} +R_j^*)\1^{\le N} \Phi(t) \big\rangle \right| \le C_\varepsilon \Big( N^{(2\beta+9\varepsilon-1)/2} + N^{-1}(1+t)^{\varepsilon} \Big) .
\end{align}
\medskip
\noindent
{\bf Step 3.} From \eqref{eq:final-proof-2}, \eqref{eq:final-proof-bH>N-last} and \eqref{eq:final-proof-2a}, we find that
$$
\partial_t \| \Phi_{N}(t)-\Phi(t)\|^2 \le C_\varepsilon \Big( N^{(2\beta+9\varepsilon-1)/2} + N^{-1} (1+t)^{\varepsilon} \Big).
$$
Integrating over $t$ and using
$$
\| \Phi_N(0)-\Phi(0)\|^2 = \langle \Phi(0), \1^{>N} \Phi(0) \rangle \le N^{-1} \langle \Phi(0), \mathcal{N} \Phi(0) \rangle \le C_\varepsilon N^{\varepsilon-1}.
$$
we obtain
\begin{align*}
\| \Phi_{N}(t)-\Phi(t)\|^2 &\le C_\varepsilon N^{\varepsilon-1} + C_\varepsilon \Big( t N^{(2\beta+9\varepsilon-1)/2} + N^{-1} (1+t)^{1+\varepsilon} \Big) \\
& \le C_\varepsilon (1+t)^{1+\varepsilon} N^{(2\beta+9\varepsilon-1)/2}.
\end{align*}
Finally, from \eqref{eq:final-proof-1} we conclude that
\begin{align} \label{eq:thm-quantitative-estimate}
\| \Psi_{N}(t)- U_N(t)^* \1^{\le N} \Phi(t)\|_{\mathfrak{H}^N}^2 &\le \| \Phi_{N}(t)-\Phi(t)\|^2 \nonumber \\
& \le C_\varepsilon (1+t)^{1+\varepsilon} N^{(2\beta+9\varepsilon-1)/2}
\end{align}
for all $0<\varepsilon<\min \{1/2,1-2\beta\}$. In the latter estimate we can thus replace $\varepsilon$ by $\varepsilon/9$ and obtain
\begin{align*}
\| \Psi_{N}(t)- U_N(t)^* \1^{\le N} \Phi(t)\|_{\mathfrak{H}^N}^2 \le C_{\varepsilon} (1+t)^{1+\varepsilon} N^{(2\beta+\varepsilon-1)/2}
\end{align*}
for all $0<\varepsilon<\min \{1/2,1-2\beta\}$ (with the constant $C_\varepsilon$ adjusted appropriately). As we have explained, this estimate holds for all $\varepsilon>0$ because $(1+t)^{1+\varepsilon} N^{(2\beta+\varepsilon-1)/2}\le (1+t)^{1+\varepsilon'} N^{(2\beta+\varepsilon'-1)/2}$ when $\varepsilon'\ge \varepsilon$. This ends the proof of Theorem \ref{thm:main}.
\end{proof}
|
\section{Introduction}
Let $m,n$ be two fixed positive integers with $2\leq m\leq n$. Let $G$ be a subset of $$\big\{0,1,\ldots,n-1\big\}\times\big\{0,1,\ldots,m-1\big\}$$
with $N:=\mbox{card}\left(G\right)\geq2$. We consider a family of affine
mappings on $\mathbb{R}^{2}$:
\begin{equation}
f_{ij}:(x,y)\mapsto\big(n^{-1}x+n^{-1}i,m^{-1}y+m^{-1}j\big),\;\;(i,j)\in G.\label{fi's}
\end{equation}
By \cite{Hut:81}, there exists a unique non-empty
compact set $E$ satisfying
\[
E=\bigcup_{(i,j)\in G}f_{ij}(E).
\]
The set $E$ is the self-affine set determined by $(f_{ij})_{(i,j)\in G}$. We also call it a Bedford-McMullen carpet.
Let $(p_{ij})_{(i,j)\in G}$ be a probability vector with $p_{ij}>0$
for all $(i,j)\in G$, there exists a unique Borel probability measure
$\mu$ satisfying
\begin{equation}
\mu=\sum_{(i,j)\in G}p_{ij}\mu\circ f_{ij}^{-1}.\label{selfaffinemeas}
\end{equation}
The measure $\mu$ is referred to as the self-affine measure associated with $(p_{ij})_{(i,j)\in G}$
and $(f_{ij})_{(i,j)\in G}$. Self-affine sets and measures in the above-mentioned cases and some more general cases have
been intensively studied in the past years; one may see \cite{Bed:84,Fal:10,King:95,LG:92,Mcmullen:84,Peres:94b}
for interesting results in this direction. Write
\begin{eqnarray*}
&&G_{x}:= \left\{ i:(i,j)\in G \mbox{ for some } j \right\};
\;\;G_{y}:=\left\{ j:(i,j)\in G \mbox{ for some } i \right\} ,\\
&&G_{x,j}:=\left\{ i:(i,j)\in G\right\},\;\;q_{j}:=\sum_{i\in G_{x,j}}p_{ij}\;,j\in G_y;\;\;\theta:=\frac{\log m}{\log n}.
\end{eqnarray*}
We say that $E$ has uniform horizontal fibres if ${\rm card}(G_{x,j})$ is constant for $j\in G_y$.
By Peres \cite{Peres:94b}, the Hausdorff measure of $E$ is infinite in its Hausdorff dimension if $E$ does not have uniform horizontal fibres; otherwise its Hausdorff measure is finite and positive.
In the present paper, we further study the quantization problem for self-affine measures as defined in (\ref{selfaffinemeas}). We refer to \cite{KZ:15} for some previous work of the author and Kesseb\"{o}hmer.
The quantization problem for probability measures originated in
information theory and engineering technology (cf. \cite{GN:98,Za:63}).
Mathematically, the problem consists in estimating the asymptotic error in
the approximation of a given probability measure by discrete probability
measures with finite support in terms of $L_{r}$-metrics.
We refer to Graf and Luschgy \cite{GL:00} for rigorous mathematical foundations of
quantization theory. One may see \cite{GL:04,GL:05,Kr:08,LM:02,PK:01} for more related results.
Let $\|\cdot\|$ be a norm on $\mathbb{R}^{q}$ and $d$ the metric
induced by this norm. For each $k\in\mathbb{N}$, we write $\mathcal{D}_{k}:=\{\alpha\subset\mathbb{R}^{q}:1\leq{\rm card}(\alpha)\leq k\}$.
Let $\nu$ be a Borel probability measure on $\mathbb{R}^{q}$. The
$k$th quantization error for $\nu$ of order $r\in(0,\infty)$ is defined
by
\begin{eqnarray}\label{quanerror}
e_{k,r}(\nu):=\bigg(\inf_{\alpha\in\mathcal{D}_{k}}\int d(x,\alpha)^{r}d\nu(x)\bigg)^{\frac{1}{r}}
\end{eqnarray}
By \cite{GL:00}, the $k$th quantization error equals the error when approximating $\nu$ with discrete probability measures supported on at most $k$ points.
If the infimum in (\ref{quanerror}) is attained at some $\alpha\in\mathcal{D}_k$, then we call $\alpha$ an $k$-optimal set for $\nu$ of order $r$. The collection of all $k$-optimal sets for $\nu$ of order $r$ is denoted by $C_{k,r}(\nu)$. By Theorem 4.12 of \cite{GL:00}, $C_{k,r}(\nu)$ is non-empty provided that the moment condition $\int |x|^rd\nu(x)<\infty$ is satisfied. This condition is clearly ensured if the support of the measure $\nu$ is compact. Also, under the moment condition, we have $e_{k,r}(\nu)\to 0$ as $k$ tends to infinity (see Lemma 6.1 of \cite{GL:00}).
As natural characterizations of the asymptotics for the quantization error $e_{k,r}(\nu)$ as $k$ tends to infinity, we consider the $s$-dimensional upper and
lower quantization coefficient of order $r$, which are defined below:
\begin{eqnarray*}
\underline{Q}_{r}^{s}(\nu):=\liminf_{k\to\infty}k^{\frac{r}{s}}e_{k,r}^r(P),\;\;\overline{Q}_{r}^{s}(\nu):=\limsup_{k\to\infty}k^{\frac{r}{s}}e_{k,r}^r(\nu),\;\; s\in(0,\infty).
\end{eqnarray*}
The upper and lower quantization dimension for $\nu$ of order $r$ are defined by
\begin{eqnarray}
\overline{D}_{r}(\nu):=\limsup_{k\to\infty}\frac{\log k}{-\log e_{k,r}(\nu)},\;\underline{D}_{r}(\nu):=\liminf_{k\to\infty}\frac{\log k}{-\log e_{k,r}(\nu)}.\label{quandimdef}
\end{eqnarray}
These two quantities are respectively the critical points at which the upper and lower quantization coefficient jump from infinity to zero (cf. Proposition 11.3 of \cite{GL:00} or \cite{PK:01}).
If $\overline{D}_{r}(\nu)=\underline{D}_{r}(\nu)$, the common value
is called the quantization dimension for $\nu$ of order $r$
and denoted by $D_{r}(\nu)$.
Compared with the upper and lower quantization
dimension, the upper and lower quantization coefficient
provide us with more accurate information on the asymptotic properties
of the quantization error. Accordingly, it is usually much more difficult to examine the finiteness and positivity of the upper and lower quantization coefficient.
Next, we recall our previous work on the quantization for self-affine measures in \cite{KZ:15}.
Let $s_{r}$ be the unique solution of the following equation:
\begin{eqnarray}
\bigg(\sum_{(i,j)\in G}(p_{ij}m^{-r})^{\frac{s_{r}}{s_{r}+r}}\bigg)^{\theta}\bigg(\sum_{j\in G_{y}}(q_{j}m^{-r})^{\frac{s_{r}}{s_{r}+r}}\bigg)^{1-\theta}=1.\label{maineq1}
\end{eqnarray}
In \cite{KZ:15}, Kesseb\"{o}hmer and Zhu proved that, for every $r\in(0,\infty)$, the quantization dimension for $\mu$ of order $r$ exists and equals $s_r$. Moreover, the $s_r$-dimensional upper and lower quantization coefficient are both positive and finite if one of the following conditions is fulfilled:
\begin{enumerate}
\item[\rm (a)]$\sum_{i\in G_{x,j}}(p_{ij}q_{j}^{-1})^{\frac{s_{r}}{s_{r}+r}}$
is identical for all $j\in G_{y}$;
\item[\rm (b)]$q_{j}$ is identical for
all $j\in G_{y}$.
\end{enumerate}
While the quantization dimension is determined for $\mu$ in general, the finiteness and positivity of the upper and lower quantization coefficient are examined only for some rare cases (a) and (b); in these cases we could estimate the asymptotics of the quantization error by means of another self-affine measure. One may see \cite{KZ:15} for more details.
As the upper and lower quantization coefficient indicate the convergence order of the quantization errors, they are of significant importance in quantization theory for probability measures. In view of our previous work in \cite{KZ:15}, a natural question is, what will happen if we drop the conditions in (a) and (b). With Peres'results \cite{Peres:94b} in mind, one might compare the quantization coefficient for $\mu$ with the Hausdorff measure of $E$ and conjecture that the above assumption (a) or (b) is a necessary condition for the upper and lower quantization coefficient to be both positive and finite. However, as our main result of the present paper, we will prove
\begin{theorem}
\label{mthm1} Let $\mu$ be the self-affine measure as defined in (\ref{selfaffinemeas}). Then for every $r\in(0,\infty)$ we have
$0<\underline{Q}_{r}^{s_{r}}(\mu)\leq\overline{Q}_{r}^{s_{r}}(\mu)<\infty$.
\end{theorem}
By Theorem \ref{mthm1}, one can see that the $k$th quantization error for $\mu$ of order $r$ is of the same order as $k^{-\frac{1}{s_r}}$, independently of the horizontal fibres of $E$.
The main obstacle in the way of proving Theorem \ref{mthm1} lies in the fact that, without the assumptions (a) and (b), one can hardly transfer the sums over approximate squares (cf. Section 2) of different orders to those over approximate squares of the same order. Our main idea is to associate approximate squares with subsets of the product space $G^{\mathbb{N}}\times G_y^{\mathbb{N}}$ and vice versa. This will enable us to estimate the asymptotic quantization errors for $\mu$ by means of a natural product measure on $G^{\mathbb{N}}\times G_y^{\mathbb{N}}$. We will also need to take care of the overlapping cases which are induced by such procedures.
\section{Preliminaries}
In order to avoid degenerate cases, in the following, we always assume that
\begin{equation}
2\leq m<n,\;{\rm card}\left(G_{x}\right),{\rm card}\left(G_{y}\right)\geq2.\label{hypo1}
\end{equation}
Since norms on $\mathbb{R}^q$ are pairwise equivalent, we will always work with Euclidean metrics for convenience.
For $x\in\mathbb{R}$, let $[x]$ denote the largest integer not exceeding
$x$. For every $k\in\mathbb{N}$, we set $\ell(k):=[k\theta]$ and
\begin{equation}\label{s1}
\Omega_{k}:=\left\{ \begin{array}{ll}
G_{y}^k,&\mbox{ if }\;k<\theta^{-1}\\
\Omega_{k}:=G^{\ell(k)}\times G_{y}^{k-\ell(k)},&\mbox{ if }\;k\geq\theta^{-1}
\end{array}\right.,\;k\in\mathbb{N};\;\Omega^{*}:=\bigcup_{k\geq1}\Omega_{k}.
\end{equation}
For $\sigma=\big((i_{1},j_{1}),\ldots,(i_{\ell(k)},j_{\ell(k)}),j_{\ell(k)+1},\ldots,j_{k}\big)\in\Omega^{*}$, we define
\begin{eqnarray}\label{g8}
&&|\sigma|:=k,\;\;\mu_{\sigma}:=\mu\left(F_{\sigma}\right)=\prod_{h=1}^{\ell(k)}p_{i_{h}j_{h}}\prod_{h=\ell(k)+1}^{k}q_{j_{h}},\\
&&\sigma_{a}:=\left((i_{1},j_{1}),\ldots,(i_{\ell(k)},j_{\ell(k)})\right),\;\sigma_{b}:=\left(j_{\ell(k)+1},\ldots,j_{k}\right)\nonumber.
\end{eqnarray}
We also write $\sigma=\sigma_a\ast\sigma_b$.
For $\sigma,\tau\in\Omega^{*}$, we write $\sigma\prec\tau$ if $F_{\tau}\subset F_{\sigma}$;
and write $\sigma=\tau^{\flat}$ if $\sigma\prec\tau$ and $|\tau|=|\sigma|+1$. For a word
\begin{equation}\label{g9}
\sigma=\big((i_{1},j_{1}),\ldots,(i_{\ell(k)},j_{\ell(k)}),j_{\ell(k)+1},\ldots,j_{k}\big)\in\Omega^{*},
\end{equation}
$\sigma^{\flat}$ takes the following two possible forms:
\begin{eqnarray}\label{s14}
\left\{ \begin{array}{ll}
((i_{1},j_{1}),\ldots,(i_{\ell(k)},j_{\ell(k)}),j_{\ell(k)+1},\ldots,j_{k-1}),\;{\rm if}\;\ell(k)=\ell(k-1)\\
((i_{1},j_{1}),\ldots,(i_{\ell(k)-1},j_{\ell(k)-1}),j_{\ell(k)},\ldots,j_{k-1}),\;{\rm if}\;\ell(k)=\ell(k-1)+1
\end{array}\right..
\end{eqnarray}
We say that $\sigma,\tau\in\Omega^{*}$ are incomparable if neither
$\sigma\prec\tau$ nor $\tau\prec\sigma$. A finite set $\Gamma\subset\Omega^{*}$
is called a finite antichain if any two words $\sigma,\tau\in\Gamma$
are incomparable; a finite antichain $\Gamma$ is called maximal if
$E\subset\bigcup_{\sigma\in\Gamma}F_{\sigma}$.
To each word $\sigma$ of the form (\ref{g9}),
there correspond two numbers $p,q$:
\[
p:=\sum_{h=1}^{\ell(k)}i_{h}n^{\ell(k)-h},\;\;q:=\sum_{h=1}^{k}j_{h}m^{k-h};
\]
and a unique rectangle which is called an approximate square of order $k$:
\begin{equation}\label{g3}
F_{\sigma}:=\bigg[\frac{p}{n^{\ell(k)}},\frac{p+1}{n^{\ell(k)}}\bigg]\times\bigg[\frac{q}{m^{k}},\frac{q+1}{m^{k}}\bigg].
\end{equation}
We call $\sigma$ the location code for the approximate square $F_\sigma$.
\begin{remark}{\rm (see \cite{KZ:15})
We have the following facts about approximate squares.
\begin{enumerate}
\item[\rm (f1)] Let $|A|$ denote the diameter of a set $A\subset\mathbb{R}^{2}$. One can easily see
\[
m^{-|\sigma|}\leq|F_{\sigma}|\leq m^{-|\sigma|}\sqrt{n^{2}+1}.
\]
\item[\rm (f2)] For $\sigma,\tau\in\Omega^*$, by the definition, we have, either $F_\sigma,F_\tau$ are non-overlapping, or one is a subset of the other.
\item[\rm (f3)] For $\sigma\in\Omega^*$, let $\mu_\sigma$ be as defined in (\ref{g8}). Then by (\ref{s14}), we have
\begin{equation}\label{g5}
\frac{\mu_\sigma}{\mu_{\sigma^\flat}}\leq\max\bigg\{\max_{(i,j)\in G}\frac{p_{ij}}{q_j}\max_{\hat{j}\in G_y}q_{\hat{j}},\max_{\hat{j}\in G_y}q_{\hat{j}}\bigg\}=\max_{\hat{j}\in G_y}q_{\hat{j}}.
\end{equation}
\end{enumerate}
}\end{remark}
For $r>0$ and each $k\geq1$, we define
\begin{eqnarray}\label{newanti}
&\underline{\eta}_{r}:=\min\big\{p_{ij}q_{k}m^{-r}:(i,j)\in G,k\in G_{y}\big\};\nonumber\\
&\Upsilon_{k,r}:=\big\{\sigma\in\Omega^{*}:\mu_{\sigma^{\flat}}m^{-|\sigma^{\flat}|r}\geq \underline{\eta}_{r}^k>\mu_{\sigma}m^{-|\sigma|r}\big\},\;\psi_{k,r}:={\rm card}(\Upsilon_{k,r}).
\end{eqnarray}
For two number sequences $(a_k)_{k=1}^\infty$ and $(b_k)_{k=1}^\infty$, we write $a_k\asymp b_k$ if there exists a constant $C$ independent of $k$ such that $Cb_k\leq a_k\leq C^{-1}b_k$. As the proof of Lemma 4 in \cite{KZ:15} shows, we have
\begin{eqnarray}\label{characterization}
e_{\psi_{k,r},r}^r(\mu)\asymp\sum_{\sigma\in\Upsilon_{k,r}}\mu_\sigma m^{-|\sigma|r}.
\end{eqnarray}
\begin{remark}{\rm
Let us make some remarks about $\Upsilon_{j,r}$ and the mass distribution of $\mu$.
\item[\rm (f4)]The set $\Upsilon_{k,r}$ possesses some kind of uniformity, which allows us estimate the number of points in a $\psi_{j,r}$-optimal set $\alpha$ which are lying in disjoint neighborhoods of the approximate squares $F_\sigma,\sigma\in\Upsilon_{k,r}$. We may think of (\ref{characterization}) roughly as follows. For each $\sigma\in\Upsilon_{j,r}$, $F_\sigma$ "owns" one point $a_\sigma$ of a $\psi_{j,r}$-optimal set $\alpha$ and
\[
\int_{F_\sigma}d(x,\alpha)^rd\mu(x)\asymp\mu_\sigma m^{-|\sigma|r}.
\]
We refer to \cite{KZ:15b} for some more intuitive interpretations on such estimates.
\item[\rm (f5)] The structure of the set $\Upsilon_{k,r}$ is not clear enough for us to estimate the sum on the right side of (\ref{characterization}). Let $\sigma$ be given in (\ref{g9}). Assume that $\ell(k+1)=\ell(k)+1$. For $j\neq j_{\ell(k)+1},(i,j)\in G$ and $\hat{j}\in G_y$, we write
\[
\hat{\sigma}=\big((i_{1},j_{1}),\ldots,(i_{\ell(k)},j_{\ell(k)}),(i,j),j_{\ell(k)+2},\ldots,j_{k},\hat{j}\big).
\]
One can see that $F_{ \hat{\sigma}}$ is not a subset of $F_\sigma$. Roughly speaking, approximate squares do not enjoy enough "freedom" as far as sub-approximate squares are concerned.
\item[\rm (f6)]For distinct words of the form (8), the measure $\mu$ are distributed in different manners among sub-approximate squares of them, since the vectors $(p_{ij})_{i\in G_{x,j}}$ are typically not identical for $j\in G_y$.
}\end{remark}
The facts as stated in (f5) and (f6) seem to prevent us from constructing a suitable auxiliary measure via approximate squares without the assumptions (a) and (b) (see Section 1). In order to show the finiteness of the upper quantization coefficient for $\mu$, we will "embed" the sets $\Upsilon_{j,r}$ into the product space $G^{\mathbb{N}}\times G_y^{\mathbb{N}}$, and then estimate the quantization errors by using a product measure $W$ on $G^{\mathbb{N}}\times G_y^{\mathbb{N}}$ and counting all possible overlapping cases. To establish a lower bound for the lower quantization coefficient for $\mu$, we will construct a new sequence of subsets $\mathcal{L}_{j,r}(2)$ of $\Omega^*$ such that, on one hand, they can play the same role as $\Upsilon_{j,r}$, and on the other hand, they enjoy enough "freedom" so that the corresponding integrals can be well estimated by means of the above-mentioned product measure $W$.
For convenience, in the remaining part of the paper, we write
\[
\mathcal{E}_r(\sigma):=(\mu_\sigma m^{-|\sigma|r})^{\frac{s_r}{s_r+r}},\;\sigma\in\Omega^*.
\]
Note that $\psi_{j,r}\asymp\psi_{j+1,r}$ by the proof of Lemma 1 in \cite{KZ:15}. To study the finiteness and positivity of the upper and lower quantization coefficient for $\mu$, we will show that it suffices to examine the asymptotics of the sequence $(e_{\psi_{j,r},r}(\mu))_{j=1}^\infty$. By H\"{o}lder's inequality with exponent less than one, the problem further reduces to the asymptotics of the following number sequence:
\[
\sum_{\sigma\in\Upsilon_{j,r}}\mathcal{E}_r(\sigma),\;j\geq 1.
\]
For the proof of the main theorem, we will need to go back and forth between words in $\Upsilon_{j,r}$ and subsets of $G^{\mathbb{N}}\times G_y^{\mathbb{N}}$.
\section{The finiteness of the upper quantization coefficient for $\mu$}
We denote by $\vartheta$ the empty word and define
\[
G^0=G_y^0:=\{\vartheta\};\;\;G^*:=\bigcup_{k=0}^\infty G^k,\;\;G_y^*:=\bigcup_{k=0}^\infty G_y^k.
\]
Let $\sigma=\left((i_{1},j_{1}),\ldots,(i_{k},j_{k})\right)\in G^{k}$. We define
\begin{eqnarray}\label{s2}
|\sigma|=k,\;\;\sigma|_h=((i_{1},j_{1}),\ldots,(i_{h},j_{h})),\;1\leq h\leq k;\;\;\sigma^-:=\sigma|_{k-1}.
\end{eqnarray}
For $\sigma,\omega\in G^*$ with $\sigma=\omega|_{|\sigma|}$, we write $\sigma\prec\omega$. We define $\sigma|_h$ similarly for $\sigma\in G^{\mathbb{N}}$ and $h\geq 1$.
If $\omega\in G^*$ and $\sigma\in G^{\mathbb{N}}$ satisfy $\omega=\sigma_{|\omega|}$, then we also write $\omega\prec\sigma$.
For $\sigma=((i_{1},j_{1}),\ldots,(i_{k},j_{k}))$ and
$\omega=((i_{k+1},j_{k+1}),\ldots,(i_{k+h},j_{k+h}))\in G$, we write
\[
\sigma\ast\omega:=((i_{1},j_{1}),\ldots,(i_{k},j_{k}),(i_{k+1},j_{k+1}),\ldots,(i_{k+h},j_{k+h})).
\]
For $\rho,\tau\in G_y^*$, we define $\rho^-,\rho\ast\tau$ and a partial order "$\prec$" in the same manner as we did for words in $G^*$.
For $r\in(0,\infty)$, we write
\[
P_r:=\sum_{(i,j)\in G}(p_{ij}m^{-r})^{\frac{s_r}{s_r+r}},\;Q_r:=\sum_{j\in G_y}(q_jm^{-r})^{\frac{s_r}{s_r+r}}.
\]
It is noted in the proof of Lemma 5 of \cite{KZ:15} that $P_r\geq 1\geq Q_r$.
Set $\overline{q}:=\max_{j\in G_y}q_j$ and $\overline{\eta}_r:=(\overline{q}m^{-r})^{\frac{s_r}{s_r+r}}$. We define
\begin{equation}\label{s13}
H_{1,r}:=\min\{h:\overline{\eta}_r^h<\underline{\eta}_r\}.
\end{equation}
For every $k\geq 1$ and $\sigma=((i_1,j_1),\ldots,(i_k,j_k))\in G^k$ and $\tau=(j_1,\ldots,j_k)\in G_y^k$, we write
\begin{eqnarray*}
&[\sigma]=[(i_1,j_1),\ldots,(i_k,j_k)]:=\{\omega\in G^{\mathbb{N}}:\omega|_k=\sigma\};\\
&[\tau]=[j_1,\ldots,j_k]:=\{\rho\in G_y^{\mathbb{N}}:\rho|_k=\tau\}.
\end{eqnarray*}
Now for every $\sigma\in\Upsilon_{j,r}$, we associate $F_\sigma$ to a subset of $G^{\mathbb{N}}\times G_y^{\mathbb{N}}$ in the following way:
\[
\sigma=\sigma_a\ast\sigma_b\in\Upsilon_{j,r}\mapsto [\sigma_a]\times [\sigma_b]\subset G^{\mathbb{N}}\times G_y^{\mathbb{N}}.
\]
For every $(i,j)\in G$ and $j\in G_y$, we define
\[
\widetilde{p}_{ij}:=P_r^{-1}(p_{ij}m^{-r})^{\frac{s_r}{s_r+r}},\;\;\widetilde{q}_j:=Q_r^{-1}(q_jm^{-r})^{\frac{s_r}{s_r+r}}.
\]
Let $G$ and $G_y$ be endowed with discrete topology and $G^{\mathbb{N}},G_y^{\mathbb{N}}$ be endowed with the corresponding product topology. We denote by $\mathcal{B}_1,\mathcal{B}_2$ the Borel sigma-algebra on $G^{\mathbb{N}},G_y^{\mathbb{N}}$.
By Kolmogrov consistency theorem, there exist a unique Borel probability measure $\lambda$ on $G^{\mathbb{N}}$ and a unique Borel probability measure $\nu$ on $G_y^{\mathbb{N}}$ such that
\begin{eqnarray*}
&\lambda([\sigma])=\prod_{h=1}^k\widetilde{p}_{i_hj_h},\;{\rm for\;every}\;\;\sigma=(i_1,j_1),\ldots,(i_k,j_k)\in G^k\;\;{\rm and}\;\;k\geq 1;
\\&\nu([\tau])=\prod_{h=1}^k\widetilde{q}_{i_hj_h},\;{\rm for\;every}\;\;\tau=(j_1,\ldots,j_k)\in G_y^k\;\;{\rm and}\;\;k\geq 1.
\end{eqnarray*}
Thus, we obtain a unique product measure $W$ on $G^{\mathbb{N}}\times G_y^{\mathbb{N}}$ such that
\begin{eqnarray*}
W(A\times B)=\lambda(A)\nu(B),\;\;A\in\mathcal{B}_1,\; B\in\mathcal{B}_2.
\end{eqnarray*}
We know that words in $\Upsilon_{j,r}$ are pairwise incomparable and $F_\sigma,\sigma\in\Upsilon_{j,r}$, are non-overlapping. However, it can happen that $[\sigma^{(1)}_a]\times[\sigma^{(1)}_b]$ and $[\sigma^{(2)}_a]\times[\sigma^{(2)}_b]$ are overlapping. We will use the following lemma to treat such overlapping cases.
\begin{lemma}\label{pre0}
For every $\sigma\in\Upsilon_{j,r}$, we write
\[
S_1(\sigma):=\{\tau\in\Upsilon_{j,r}:\sigma_a\prec\tau_a,\sigma_b\prec\tau_b\}.
\]
Then we have
\[
\sum_{\tau\in S_1(\sigma)}W([\tau_a]\times[\tau_b])\leq H_{1,r}W([\sigma_a]\times[\sigma_b]).
\]
\end{lemma}
\begin{proof}
For every $h\geq 1$, let $\Gamma_h([\sigma_a]\times[\sigma_b])$ denote the collection of the subsets $[\rho]\times [\omega]$ of $G^{\mathbb{N}}\times G_y^{\mathbb{N}}$ satisfying
\begin{eqnarray*}
[\rho]\times [\omega]\subset[\sigma_a]\times[\sigma_b],\;|\rho|+|\omega|=|\sigma|+h,\;[(|\rho|+\omega)\theta]=|\rho|.
\end{eqnarray*}
Note that the words in $\Gamma_1([\sigma_a]\times[\sigma_b])$ take exactly one of the following two forms:
\begin{equation}\label{z2}
[\sigma_a\ast(i,j)]\times[\sigma_b],\;\;{\rm or}\;\;[\sigma_a]\times[\sigma_b\ast \hat{j}],\;(i,j)\in G,\;\hat{j}\in G_y.
\end{equation}
Using this fact and mathematical induction, for every $h\geq 1$, we obtain
\begin{equation}\label{z10}
\sum_{\rho\times\omega\in\Gamma_h([\sigma_a]\times[\sigma_b])}W(\Gamma_h([\sigma_a]\times[\sigma_b]))=W([\sigma_a]\times[\sigma_b]).
\end{equation}
Also, using (\ref{z2}) and mathematical induction, for every $\rho\times\omega\in\Gamma_h([\sigma_a]\times[\sigma_b])$, we have
\begin{equation}\label{z3}
\underline{\eta}_r^{\frac{hs_r}{s_r+r}}\mathcal{E}_r(\sigma)\leq\mathcal{E}_r(\rho\ast\omega)\leq\overline{\eta}_r^{\frac{hs_r}{s_r+r}}\mathcal{E}_r(\sigma).
\end{equation}
By the definition, one can see that for every $\tau\in S_1(\sigma)$, we have
\begin{eqnarray*}
[\tau_a]\times[\tau_b]\in\Gamma_h([\sigma_a]\times[\sigma_b])\;\;{\rm for\; some}\;\; h.
\end{eqnarray*}
Suppose that for some $\tau\in S_1(\sigma)$, we have $|\tau|\geq |\sigma|+H_{1,r}$. By (\ref{z3}), we would have
\[
\mathcal{E}_r(\tau)\leq \overline{\eta}_r^{H_{1,r}}\mathcal{E}_r(\sigma)<\underline{\eta}_r^{\frac{s_r}{s_r+r}}\mathcal{E}_r(\sigma).
\]
This contradicts (\ref{newanti}), since by (\ref{newanti}), for every $\tau\in\Upsilon_{j,r}$, we have
\[
\underline{\eta}_r^{\frac{s_r}{s_r+r}}\mathcal{E}_r(\sigma)\leq\mathcal{E}_r(\tau)\leq\underline{\eta}_r^{\frac{-s_r}{s_r+r}}\mathcal{E}_r(\sigma).
\]
Thus, for every $\tau\in S_1(\sigma)$, we have $|\tau|\leq|\sigma|+H_{1,r}$. It follows that
\begin{equation}\label{g1}
\bigcup_{\tau\in S_1(\sigma)}[\tau_a]\times[\tau_b]\subset\bigcup_{h=1}^{H_{1,r}}\Gamma_h([\sigma_a]\times[\sigma_b]),
\end{equation}
For distinct words $\sigma^{(1)},\sigma^{(2)}\in\Upsilon_{j,r}$, we have either $\sigma^{(1)}_a\neq\sigma^{(2)}_a$, or $\sigma^{(1)}_b\neq\sigma^{(2)}_b$. So,
\begin{equation}\label{g2}
[\sigma^{(1)}_a]\times[\sigma^{(1)}_b]\neq[\sigma^{(2)}_a]\times[\sigma^{(2)}_b].
\end{equation}
Thus, the lemma follows by (\ref{z10}), (\ref{g1}) and (\ref{g2}).
\end{proof}
Next, we show the finiteness of the upper quantization coefficient for $\mu$, by using Lemma \ref{pre0} and the auxiliary measure $W$.
\begin{proposition}\label{pre1}
Let $\mu$ be a measure as defined in (\ref{selfaffinemeas}). Then $\overline{Q}_r^{s_r}(\mu)<\infty$.
\end{proposition}
\begin{proof}
First, we estimate $\sum_{\sigma\in\Upsilon_{j,r}}\mathcal{E}_r(\sigma)$ from above by means of the measure $W$. For a word $\sigma\in\Upsilon_{j,r}$, by the definition, it takes the form:
\[
\sigma=\big((i_{1},j_{1}),\ldots,(i_{\ell(k)},j_{\ell(k)}),j_{\ell(k)+1},\ldots,j_{k}\big)\in\Omega^{*}.
\]
We associate $\sigma$ with the following subset of $G^{\mathbb{N}}\times G_y^{\mathbb{N}}$:
\[
[\sigma_a]\times[\sigma_b]=[(i_1,j_1),\ldots,(i_{\ell(k)},j_{\ell(k)})]\times[j_{\ell(k)+1},\ldots,j_{k}].
\]
Note that for all $k\geq\theta^{-1}$, we have $P_r^{-1}Q_r\leq P_r^{\ell(k)}Q_r^{(k-\ell(k))}\leq 1$. We deduce
\begin{eqnarray}\label{s3}
W([\sigma_a]\times[\sigma_b])&&=\prod_{h=1}^{\ell(k)}\widetilde{p}_{i_hj_h}\prod_{h=\ell(k)+1}^k\widetilde{q}_{j_h}\nonumber
\\&&=P_r^{-\ell(k)}Q_r^{-(k-\ell(k))}
(\mu_\sigma m^{-|\sigma|r})^{\frac{s_r}{s_r+r}}\nonumber\\&&\left\{ \begin{array}{ll}
\leq P_r Q_r^{-1}\mathcal{E}_r(\sigma)\\
\geq\mathcal{E}_r(\sigma)
\end{array}\right..
\end{eqnarray}
For distinct words $\sigma^{(1)},\sigma^{(2)}\in\Upsilon_{j,r}$, we have either $\sigma^{(1)}_a\neq\sigma^{(2)}_a$ or $\sigma^{(1)}_b\neq\sigma^{(2)}_b$. Thus, they are associated to distinct subsets of $G^{\mathbb{N}}\times G_y^{\mathbb{N}}$. We write
\[
\mathcal{W}_{j,r}:=\{[\sigma_a]\times[\sigma_b]:\;\sigma=\sigma_a\ast\sigma_b\in\Upsilon_{j,r}\}.
\]
We distinguish two cases:
Case (i): either $\sigma^{(1)}_a,\sigma^{(2)}_a$ or, $\sigma^{(1)}_b,\sigma^{(2)}_b$ are incomparable. In this case, we have
\[
([\sigma^{(1)}_a]\times[\sigma^{(1)}_b])\cap([\sigma^{(2)}_a]\times[\sigma^{(2)}_b])=\emptyset.
\]
Case (ii): both $\sigma^{(1)}_a,\sigma^{(2)}_a$ and $\sigma^{(1)}_b,\sigma^{(2)}_b$ are comparable. Note that
\[
[(|\sigma_a|+|\sigma_b|)\theta]=|\sigma_a|,\;{\rm for\;all}\;\;\sigma=\sigma_a\ast\sigma_b\in\Upsilon_{j,r}.
\]
Thus, whenever $|\sigma^{(1)}_a|<|\sigma^{(2)}_a|$, we have $|\sigma^{(1)}_b|\leq|\sigma^{(2)}_b|$. Hence, we may assume that
\[
\sigma^{(1)}_a\prec\sigma^{(2)}_a\;\;{\rm and}\;\;\sigma^{(1)}_b\prec\sigma^{(2)}_b.
\]
In this case we have
\[
([\sigma^{(1)}_a]\times[\sigma^{(1)}_b])\supset([\sigma^{(2)}_a]\times[\sigma^{(2)}_b]).
\]
Let $H_{1,r}$ be as defined in (\ref{s13}). Then by the proof of Lemma \ref{pre0}, we have $|\sigma^{(2)}|\leq|\sigma^{(1)}|+H_{1,r}$. For every $\sigma\in\Upsilon_{j,r}$, we write
\[
\mathcal{F}(\sigma):=\{\omega\in\Upsilon_{j,r}:\sigma_a,\omega_a; \;{\rm and}\; \sigma_b,\omega_b\;{\rm are\;both\;comparable}\}.
\]
Let $\widetilde{\sigma}$ denote the shortest word in $\mathcal{F}(\sigma)$ and $\mathcal{F}_{j,r}$ the set of all such words. Then, For every pair $\widetilde{\sigma},\widetilde{\omega}\in\mathcal{F}_{j,r}$, we have
\begin{equation}\label{s5}
([\widetilde{\sigma}_a]\times[\widetilde{\sigma}_b])\cap([\widetilde{\omega}_a]\times[\widetilde{\omega}_b])=\emptyset.
\end{equation}
By Lemma \ref{pre0}, we have
\begin{equation}\label{s4}
\sum_{\omega\in\mathcal{F}(\sigma)}W([\omega_a]\times[\omega_b])=\sum_{\omega\in S(\widetilde{\sigma})}W([\omega_a]\times[\omega_b])\leq H_{1,r} W([\widetilde{\sigma}_a]\times[\widetilde{\sigma}_b])
\end{equation}
Combining this with (\ref{s3})-(\ref{s5}), we deduce
\begin{eqnarray}\label{g4}
\sum_{\sigma\in\Upsilon_{j,r}}\mathcal{E}_r(\sigma)&\leq&\sum_{\sigma\in\Upsilon_{j,r}}W([\sigma_a]\times[\sigma_b])
\nonumber\\&=&\sum_{\widetilde{\sigma}\in\mathcal{F}_{j,r}}\sum_{\sigma\in\mathcal{F}(\widetilde{\sigma})}W([\sigma_a]\times[\sigma_b])\nonumber\\&\leq&
H_{1,r}\sum_{\widetilde{\sigma}\in\mathcal{F}_{j,r}}W([\widetilde{\sigma}_a]\times[\widetilde{\sigma}_b])\nonumber\\&\leq& H_{1,r}.
\end{eqnarray}
This, together with (\ref{newanti}), implies
\begin{equation}
\psi_{j,r}\underline{\eta}_r^{\frac{(j+1)s_r}{s_r+r}}\leq H_{1,r},\;\;{\rm implying}\;\;\underline{\eta}_r^{\frac{jr}{s_r+r}}\leq H_{1,r}^{\frac{r}{s_r}}\underline{\eta}_r^{\frac{-r}{s_r+r}}\psi_{j,r}^{-\frac{r}{s_r}}.
\end{equation}
Using this, (\ref{characterization}) and (\ref{g4}), we have
\begin{eqnarray}\label{z1}
e_{\psi_{j,r},r}^r(\mu)&\asymp&\sum_{\sigma\in\Upsilon_{j,r}}\mu_\sigma m^{-|\sigma|r}=\sum_{\sigma\in\Upsilon_{j,r}}\mathcal{E}_r(\sigma)(\mu_\sigma m^{-|\sigma|r})^{\frac{r}{s_r+r}}\nonumber\\&\leq&\sum_{\sigma\in\Upsilon_{j,r}}\mathcal{E}_r(\sigma)\underline{\eta}_r^{\frac{jr}{s_r+r}}\leq H_{1,r}^{1+\frac{r}{s_r}}\underline{\eta}_r^{\frac{-r}{s_r+r}}\psi_{j,r}^{-\frac{r}{s_r}}.
\end{eqnarray}
By Lemma 1 in \cite{KZ:15}, we have $\psi_{j,r}\leq\psi_{j+1,r}\leq (mn)^{H_{1,r}}\psi_{j,r}$. For each $k\geq \psi_{1,r}$, there exists some $j$ such that $\psi_{j,r}\leq k<\psi_{j+1,r}$. Thus, by (\ref{z1}) and Theorem 4.12 of \cite{GL:00}, we deduce
\begin{eqnarray}\label{g7}
\overline{Q}_r^{s_r}(\mu)&=&\limsup_{k\to\infty}k^{\frac{r}{s_r}}e_{k,r}^r(\mu)\leq\limsup_{j\to\infty}\psi_{j+1,r}^{\frac{r}{s_r}}
e_{\psi_{j,r},r}^r(\mu)\nonumber\\&\leq& (mn)^{\frac{rH_{1,r}}{s_r}}\limsup_{j\to\infty}\psi_{j,r}^{\frac{r}{s_r}}
e_{\psi_{j,r},r}^r(\mu)\nonumber\\&\leq&(mn)^{\frac{rH_{1,r}}{s_r}}H_{1,r}^{1+\frac{r}{s_r}}\underline{\eta}_r^{\frac{-r}{s_r+r}}.
\end{eqnarray}
The proof of the proposition is now complete.
\end{proof}
\section{The positivity of the lower quantization coefficient for $\mu$}
Let $\Upsilon_{j,r}$ be as defined in (\ref{newanti}). We write
\begin{eqnarray*}
k_{1j}:=\min_{\sigma\in\Upsilon_{j,r}}|\sigma|,\;k_{2j}:=\max_{\sigma\in\Upsilon_{j,r}}|\sigma|;\;\Lambda_{j,r}(k):=\Upsilon_{j,r}\cap\Omega_k.
\end{eqnarray*}
For $\sigma\in G^*$ and $\omega\in G_y^*$, we write $\sigma\times\omega$ for the corresponding word in $G^*\times G_y^*$.
We consider words of $G^*\times G_y^*$ which takes the following form:
\[
\sigma\times\omega,\;|\sigma|+\ell(k_{1j})=\ell(|\sigma|+|\omega|+k_{1j}),\;\sigma\in G^*,\omega\in G_y^*.
\]
Let $\mathcal{H}_{j,r}$ denote the set of all such words. Note that
\begin{eqnarray*}
\ell(|\sigma|+|\omega|+k_{1j}-1)&=&[(|\sigma|+|\omega|+k_{1j}-1)\theta]\geq(|\sigma|+|\omega|+k_{1j}-1)\theta-1\\
&=&(|\sigma|+|\omega|+k_{1j})\theta-(1+\theta)>|\sigma|+\ell(k_{1j})-2.
\end{eqnarray*}
Thus, $\ell(|\sigma|+|\omega|+k_{1j}-1)$ takes two possible values: $|\sigma|+\ell(k_{1j})$, or, $|\sigma|+\ell(k_{1j})-1$.
This allows us to define $(\sigma\times\omega)^\flat\in\mathcal{H}_{j,r}$:
\begin{eqnarray}\label{s7}
(\sigma\times\omega)^\flat:=\left\{ \begin{array}{ll}
\sigma\times\omega^-\;\;{\rm if}\;\;\ell(|\sigma|+|\omega|+k_{1j}-1)=|\sigma|+\ell(k_{1j})\\
\sigma^-\times\omega\;\;{\rm if}\;\;\ell(|\sigma|+|\omega|+k_{1j}-1)=|\sigma|+\ell(k_{1j})-1
\end{array}\right.,
\end{eqnarray}
where $\sigma^-,\omega^-$ are as defined in section 3.
We write $P(\sigma\times\omega):=[\sigma]\times[\omega]$ and
$P((\sigma\times\omega)^\flat):=[\sigma]\times[\omega^-]\;{\rm or}\;[\sigma^-]\times[\omega]$ in accordance with (\ref{s7}). One can easily see
\begin{eqnarray}\label{s8}
P_r^{-1}\underline{\eta}_r^{\frac{s_r}{s_r+r}} W(P(\sigma\times\omega)^\flat)\leq W(P(\sigma\times\omega))<W(P(\sigma\times\omega)^\flat).
\end{eqnarray}
By the definition, for two words $\sigma^{(i)}\times\omega^{(i)}\in\mathcal{H}_{j,r},i=1,2$, if $|\sigma^{(1)}|<|\sigma^{(2)}|$, we have $|\omega^{(1)}|\leq|\omega^{(2)}|$. Thus, whenever $\sigma^{(1)}\prec\sigma^{(2)}$ and $\sigma^{(1)}\neq\sigma^{(2)}$, we have $\omega^{(1)}\prec\omega^{(2)}$.
We write $\sigma^{(1)}\times\omega^{(1)}\prec\sigma^{(2)}\times\omega^{(2)}$, if $\sigma^{(1)}\prec\sigma^{(2)}$ and $\omega^{(1)}\prec\omega^{(2)}$; if neither $\sigma^{(1)}\times\omega^{(1)}\prec\sigma^{(2)}\times\omega^{(2)}$, nor $\sigma^{(1)}\times\omega^{(1)}\prec\sigma^{(2)}\times\omega^{(2)}$, then we say that $\sigma^{(i)}\times\omega^{(i)}\in\mathcal{H}_{j,r},i=1,2$ are incomparable. A finite set $\Gamma\subset\mathcal{H}_{j,r}$ is called a finite maximal antichain, if the words in $\Gamma$ are pairwise incomparable, and for every word $\sigma\times\omega$ in $G^{\mathbb{N}}\times G_y^{\mathbb{N}}$, there exists some word $\sigma'\times\omega'$ such that $\sigma'\prec\sigma$ and $\omega'\prec\omega$; for such a $\Gamma$ in $\mathcal{H}_{j,r}$, we have
\begin{eqnarray}\label{s15}
\bigcup_{\sigma\times\omega\in\Gamma}[\sigma]\times[\omega]=G^{\mathbb{N}}\times G_y^{\mathbb{N}};
\end{eqnarray}
and for every pair of distinct words $\sigma^{(1)}\times\omega^{(1)},\sigma^{(2)}\times\omega^{(2)}\in\Gamma$, we have
\[
([\sigma^{(1)}]\times[\omega^{(1)}])\cap([\sigma^{(2)}]\times[\omega^{(2)}])=\emptyset.
\]
In order to establish a lower bound for the lower quantization coefficient for $\mu$, we will construct a family of subsets of $G^{\mathbb{N}}\times G_y^{\mathbb{N}}$ and associate them with approximate squares. The following lemma will be used to estimate the possible overlapping cases in this process. Recall that for $\sigma,\omega\in\Omega^*$, $\sigma\prec\omega$ means $F_\omega\subset F_\sigma$.
\begin{lemma}\label{pre2}
Let $\sigma\in\Omega^*$ and $H_{2,r}:=P_r^3Q_r^{-2}\underline{\eta}_r^{-\frac{s_r}{s_r+r}}$. We write
\begin{equation}\label{s6}
S_2(\sigma):=\{\omega\in\Omega^*:\;\sigma\prec\omega,\mathcal{E}_r(\omega)\geq H_{2,r}^{-1}\mathcal{E}_r(\sigma)\}.
\end{equation}
Then there exists a constant $H_{3,r}$, which is independent of $\sigma$, such that
\[
\sum_{\omega\in S_2(\sigma)}\mathcal{E}_r(\omega)\leq H_{3,r}\mathcal{E}_r(\sigma).
\]
\end{lemma}
\begin{proof}
Let $\overline{\eta}_r$ be as defined in Section 3.
Write
\begin{eqnarray*}
&\Lambda(\sigma,h):=\{\omega\in\Omega^*:|\omega|=|\sigma|+h,\sigma\prec\omega\},\;h\geq 1;\\
&M_r:=\min\{h\in\mathbb{N}:\overline{\eta}_r^{\frac{hs_r}{s_r+r}}<H_{2,r}^{-1}\}.
\end{eqnarray*}
Then for every $\omega\in\Lambda(\sigma,M_r)$, by (\ref{g5}), we have
\[
\mathcal{E}_r(\omega)\leq\overline{\eta}_r^{\frac{M_rs_r}{s_r+r}}\mathcal{E}_r(\sigma)<H_{2,r}^{-1}\mathcal{E}_r(\sigma).
\]
Hence, for every $\omega\in S_2(\sigma)$, we have $|\omega|\leq|\sigma|+M_r$. It follows that
\[
S_2(\sigma)\subset\bigcup_{h=0}^M\Lambda(\sigma,h).
\]
Note that $0<Q_r\leq1$. By (\ref{s14}), we also have
\begin{eqnarray*}
\sum_{\omega\in\Lambda(\tau,1)}\mathcal{E}_r(\omega)&\leq& Q_r\sum_{i\in G_{x,j_{\ell(k)+1}}}\bigg(\frac{p_{ij_{\ell(k)+1}}}{q_{j_{\ell(k)+1}}}\bigg)^{\frac{s_r}{s_r+r}}\mathcal{E}_r(\tau)
\\&\leq&\max_{j\in G_y}\sum_{i\in G_{x,j}}\bigg(\frac{p_{ij}}{q_j}\bigg)^{\frac{s_r}{s_r+r}}\mathcal{E}_r(\tau)=:\xi_r\mathcal{E}_r(\tau).
\end{eqnarray*}
Using this fact and finite induction, we further deduce
\begin{eqnarray*}
\sum_{\omega\in S_2(\sigma)}\mathcal{E}_r(\omega)\leq \sum_{h=0}^{M_r}\sum_{\omega\in \Lambda(\tau,h)}\mathcal{E}_r(\omega)\leq \sum_{h=0}^{M_r}\xi_r^h\mathcal{E}_r(\sigma).
\end{eqnarray*}
Setting $H_{3,r}:=\sum_{h=0}^{M_r}\xi_r^h$, the lemma follows.
\end{proof}
Using Lemma \ref{pre1} and the product measure $W$, we are now able to prove the positivity of the lower quantization coefficient for $\mu$.
\begin{proposition}\label{pre3}
Let $\mu$ be a measure as defined in (\ref{selfaffinemeas}). Then $\overline{Q}_r^{s_r}(\mu)>0$.
\end{proposition}
\begin{proof}
For every $\tau\in\Omega_{k_{1j}}\setminus\Lambda_{j,r}(k_{1j})$, by (\ref{newanti}), we have $\mathcal{E}_r(\tau)\geq\underline{\eta}_r^{\frac{js_r}{s_r+r}}$. Set
\[
\epsilon(\tau):=\underline{\eta}_r^{\frac{js_r}{s_r+r}}\mathcal{E}_r(\tau)^{-1}.
\]
Then clearly $\epsilon(\tau)\leq 1$ for all $\tau\in\Omega_{k_{1j}}\setminus\Lambda_{j,r}(k_{1j})$. We define
\begin{eqnarray*}
\Gamma(\tau):=\{\sigma\times\omega\in\mathcal{H}_{j,r}:W(P((\sigma\times\omega)^\flat)\geq\epsilon(\tau)>W(P(\sigma\times\omega)\}.
\end{eqnarray*}
Then $\Gamma(\tau)$ is a finite maximal antichain in $\mathcal{H}_{j,r}$. Using (\ref{s15}), we deduce
\begin{eqnarray*}
\sum_{\sigma\times\omega\in\Gamma(\tau)}W([\tau_a\ast\sigma]\times[\tau_b\ast\omega])&=&\sum_{\sigma\times\omega\in\Gamma(\tau)}
\lambda([\tau_a\ast\sigma])\nu([\tau_b\ast\omega])\\&=&\sum_{\sigma\times\omega\in\Gamma(\tau)}
\lambda([\tau_a])\lambda([\sigma])\nu([\tau_b])\nu([\omega])\\&=&W([\tau_a\ast\tau_b])\sum_{\sigma\times\omega\in\Gamma(\tau)}
W([\sigma]\times[\omega])\\&=&W([\tau_a\ast\tau_b]).
\end{eqnarray*}
We need to note the following facts:
\begin{enumerate}
\item[\rm (A)] For every $\tau\in\Omega_{k_{1j}}\setminus\Lambda_{j,r}(k_{1j})$ and $\sigma\times\omega\in\Gamma(\tau)$, by (\ref{g3}), $\tau_a\ast\sigma\ast\tau_b\ast\omega$ is a location code for an approximate square;
\item[\rm (B)]
For distinct $\sigma^{(i)}\times\omega^{(i)}\in\Gamma(\tau),i=1,2$, we have, either $\sigma^{(1)},\sigma^{(2)}$; or $\omega^{(1)},\omega^{(2)}$ are incomparable. Hence,
\begin{eqnarray*}
(\tau_a\ast\sigma^{(1)})\ast(\tau_b\ast\omega^{(1)})\neq(\tau_a\ast\sigma^{(2)})\ast(\tau_b\ast\omega^{(2)});\\
\big[\tau_a\ast\sigma^{(1)}\big]\times\big[\tau_b\ast\omega^{(1)}\big]\neq\big[\tau_a\ast\sigma^{(2)}]\times[\tau_b\ast\omega^{(2)}\big].
\end{eqnarray*}
\item[\rm (C)]For different $\tau^{(i)}\in\Omega_{k_{1j}}\setminus\Lambda_{j,r}(k_{1j}),i=1,2$, we have
\[
|\tau^{(1)}_a|=|\tau^{(2)}_a|,\;\;|\tau^{(1)}_b|=|\tau^{(2)}_b|.
\]
Since $\tau^{(1)}\neq \tau^{(2)}$, we have either $\tau^{(1)}_a,\tau^{(2)}_a$ are incomparable, or $\tau^{(1)}_b,\tau^{(2)}_b$ are incomparable. Thus, for every pair $\sigma^{(i)}\times\omega^{(i)}\in\Gamma(\tau_i),i=1,2$, we have
\begin{eqnarray*}
(\tau^{(1)}_a\ast\sigma^{(1)})\ast(\tau^{(1)}_b\ast\omega^{(1)})\neq(\tau^{(2)}_a\ast\sigma^{(2)})\ast(\tau^{(2)}_b\ast\omega^{(2)});\\
\big[\tau^{(1)}_a\ast\sigma^{(1)}\big]\times\big[\tau^{(1)}_b\ast\omega^{(1)}\big]\neq\big[\tau^{(2)}_a\ast\sigma^{(2)}]\times[\tau^{(2)}_b\ast\omega^{(2)}\big].
\end{eqnarray*}
\item[\rm (D)]
It may happen that
\[
F_{\tau_a\ast\sigma^{(1)})\ast(\tau_b\ast\omega^{(1)})}\subset F_{\tau_a\ast\sigma^{(2)})\ast(\tau_b\ast\omega^{(2)})}.
\]
\end{enumerate}
We denote by $\mathcal{L}_{j,r}(1)$ the set of all the words $(\tau_a\ast\sigma)\ast(\tau_b\ast\omega)$ and words in $\Lambda_{j,r}(k_{1j})$:
\[
\mathcal{L}_{j,r}(1):=\Lambda_{j,r}(k_{1j})\cup\bigg(\bigcup_{\tau\in\Omega_{k_{1j}}\setminus\Lambda_{j,r}(k_{1j})}
\big\{(\tau_a\ast\sigma)\ast(\tau_b\ast\omega):\sigma\times\omega\in\Gamma(\tau)\big\}\bigg).
\]
For every $\tau\in\Omega_{k_{1j}}\setminus\Lambda_{j,r}(k_{1j})$ and $\sigma\times\omega\in\Gamma(\tau)$, using (\ref{s3}) and (\ref{s8}), we have
\begin{eqnarray*}
\mathcal{E}_r((\tau_a\ast\sigma)\ast(\tau_b\ast\omega))&=&(\mu_{\tau_a\ast\sigma)\ast(\tau_b\ast\omega)}m^{-|\tau_a\ast\sigma)\ast(\tau_b\ast\omega)|r})^{\frac{s_r}{s_r+r}}
\\&\geq& P_r^{-1}Q_r W([\tau_a\ast\sigma]\times[\tau_b\ast\omega])\\&=&P_r^{-1}Q_r W([\tau_a]\times[\tau_b])W([\sigma]\times[\omega])
\\&=&P_r^{-1}Q_r W([\tau_a]\times[\tau_b])W(P(\sigma\times\omega))\\&\geq& P_r^{-1}Q_r\mathcal{E}_r(\tau)W(P(\sigma\times\omega))
\\&\geq&P_r^{-1}Q_r\mathcal{E}_r(\tau)P_r^{-1}\underline{\eta}_r^{\frac{s_r}{s_r+r}}W(P((\sigma\times\omega)^\flat))
\\&\geq& P_r^{-1}Q_r\mathcal{E}_r(\tau) P_r^{-1}\underline{\eta}_r^{\frac{s_r}{s_r+r}}
\underline{\eta}_r^{\frac{js_r}{s_r+r}}\mathcal{E}_r(\tau)^{-1}\\&=&P_r^{-2}Q_r\underline{\eta}_r^{\frac{(j+1)s_r}{s_r+r}}.
\end{eqnarray*}
Analogously, one can see that $\mathcal{E}_r((\tau_a\ast\sigma)\ast(\tau_b\ast\omega))\leq P_rQ_r^{-1}\underline{\eta}_r^{\frac{js_r}{s_r+r}}$. In addition, for every $\tau\in\Lambda_{j,r}(k_{1j})$, by (\ref{newanti}), one can see that
\[
\underline{\eta}_r^{\frac{(j+1)s_r}{s_r+r}}\leq\mathcal{E}_r(\tau)<\underline{\eta}_r^{\frac{js_r}{s_r+r}}.
\]
Thus, for all words $\rho\in\mathcal{L}_{j,r}(1)$, we have
\begin{eqnarray}\label{s9}
P_r^{-2}Q_r\underline{\eta}_r^{\frac{(j+1)s_r}{s_r+r}}\leq\mathcal{E}_r(\rho)<P_r Q_r^{-1}\underline{\eta}_r^{\frac{js_r}{s_r+r}}.
\end{eqnarray}
For every $\rho\in\mathcal{L}_{j,r}(1)$, we write
\[
T(\rho):=\{\omega\in\mathcal{L}_{j,r}(1):F_\omega\subset F_\rho,\;{\rm or}\;F_\rho\subset F_\omega\}.
\]
We choose the shortest word of $T(\sigma)$ and denote $\mathcal{L}_{j,r}(2)$ the set of all such words. Then $F_\sigma,\sigma\in\mathcal{L}_{j,r}(2)$ are pairwise non-overlapping. By Lemma \ref{pre2} and (\ref{s3}),
\begin{eqnarray}\label{s10}
\sum_{\rho\in\mathcal{L}_{j,r}(2)}\mathcal{E}_r(\rho)&\geq& H_{3,r}^{-1}\sum_{\rho\in\mathcal{L}_{j,r}(2)}\sum_{\omega\in T(\rho)}\mathcal{E}_r(\omega)=H_{3,r}^{-1}\sum_{\rho\in\mathcal{L}_{j,r}(1)}\mathcal{E}_r(\rho)\nonumber
\\&\geq&H_{3,r}^{-1} P_r^{-1} Q_r\sum_{\tau\in\Lambda_{j,r}(k_{1j})}W([\tau_a]\times[\tau_b])+\nonumber\\&&+H_{3,r}^{-1}P_r^{-1} Q_r
\sum_{\tau\in\Omega_{k_{1j}}\setminus\Lambda_{j,r}(k_{1j})}\sum_{\sigma\times\omega\in\Gamma(\tau)}W([\tau_a\ast\sigma]\times[\tau_b\ast\omega])
\nonumber\\&\geq&H_{3,r}^{-1}P_r^{-1} Q_r\sum_{\tau\in\Lambda_{j,r}(k_{1j})}W([\tau_a]\times[\tau_b])\nonumber\\&&+H_{3,r}^{-1}P_r^{-1} Q_r
\sum_{\tau\in\Omega_{k_{1j}}\setminus\Lambda_{j,r}(k_{1j})}W([\tau_a]\times[\tau_b])
\nonumber\\&=&H_{3,r}^{-1}P_r^{-1} Q_r\sum_{\tau\in\Omega_{k_{1j}}}W([\tau_a]\times[\tau_b])=H_{3,r}^{-1}P_r^{-1} Q_r.
\end{eqnarray}
Analogously, one may show that
\begin{eqnarray}\label{s11}
\sum_{\rho\in\mathcal{L}_{j,r}(2)}\mathcal{E}_r(\rho)\leq 1.
\end{eqnarray}
We denote by $\phi_{j,r}$ the cardinality of $\mathcal{L}_{j,r}(2)$. Then by (\ref{s9})-(\ref{s11}), we deduce
\[
\phi_{j,r}P_r^{-2}Q_r\underline{\eta}_r^{\frac{s_r(j+1)}{s_r+r}}\leq 1;\;\phi_{j,r}P_r Q_r^{-1}\underline{\eta}_r^{\frac{js_r}{s_r+r}}\geq H_{3,r}^{-1}P_r^{-1} Q_r.
\]
Set $H_{4,r}:=P_r^2Q_r^{-1}$ and $H_{5,r}:=H_{3,r}^{-1}P_r^{-2}Q_r^2$. It follows that
\begin{eqnarray}\label{s12}
H_{5,r}\underline{\eta}_r^{\frac{-js_r}{s_r+r}}\leq \phi_{j,r}\leq H_{4,r}\underline{\eta}_r^{\frac{-s_r(j+1)}{s_r+r}}
\end{eqnarray}
Now let $H:=2([\theta^{-1}]+2)$ and $\delta:=(n^2+1)^{-1/2}$. Using the method in the proof of Lemma 2 of \cite{KZ:15}, we may choose a $\widetilde{\rho}$ for every word $\rho\in\mathcal{L}_{j,r}(2)$ such that
\[
\rho\prec\widetilde{\rho},\;\;|\widetilde{\rho}|\leq|\rho|+H
\]
and for every pair of distinct words $\rho,\omega$ of $\mathcal{L}_{j,r}(2)$, we have
\begin{eqnarray*}
d(F_{\widetilde{\rho}},F_{\widetilde{\omega}})\geq\delta\max\{|F_{\widetilde{\rho}}|,|F_{\widetilde{\omega}}|\}.
\end{eqnarray*}
Let $\alpha\in C_{\phi_{j,r},r}(\mu)$. Then by Lemma 3 of \cite{KZ:15}, we can find a constant $D$, which is independent of $j$, such that
\begin{eqnarray}\label{g6}
e_{\phi_{j,r},r}^r(\mu)&\geq&\sum_{\rho\in\mathcal{L}_{j,r}(2)}\int_{F_\rho} d(x,\alpha)^rd\mu(x)\geq
\sum_{\rho\in\mathcal{L}_{j,r}(2)}\int_{F_{\widetilde{\rho}}} d(x,\alpha)^rd\mu(x)
\nonumber\\&\geq&D\sum_{\rho\in\mathcal{L}_{j,r}(2)}\mu_{{\widetilde{\rho}}}m^{-|\widetilde{\rho}|r}
\geq\widetilde{D}\sum_{\rho\in\mathcal{L}_{j,r}(2)}\mu_\rho m^{-|\rho|r},
\end{eqnarray}
where $\widetilde{D}:=D\underline{\eta}_r^H$.
Thus, by (\ref{s10}), (\ref{g6}) and H\"{o}lder's inequality with exponent less than one, we further deduce
\begin{eqnarray*}
e_{\phi_{j,r},r}^r(\mu)&\geq&\widetilde{D}\bigg(\sum_{\rho\in\mathcal{L}_{j,r}(2)}(\mu_\rho m^{-|\rho|r})^{\frac{s_r}{s_r+r}}\bigg)^{\frac{s_r+r}{s_r}}\phi_{j,r}^{-\frac{r}{s_r}}
\\&=&\widetilde{D}\bigg(\sum_{\rho\in\mathcal{L}_{j,r}(2)}\mathcal{E}_r(\rho)\bigg)^{\frac{s_r+r}{s_r}}\phi_{j,r}^{-\frac{r}{s_r}}
\\&\geq&\widetilde{D} (H_{3,r}P_r Q_r^{-1})^{-\frac{s_r+r}{s_r}}\phi_{j,r}^{-\frac{r}{s_r}}.
\end{eqnarray*}
By (\ref{s12}), we may choose a smallest integer $H_{6,r}$ such that for every $j$, we have
\[
\phi_{j+H_{6,r},r}\geq H_{5,r}\underline{\eta}_r^{\frac{-(j+H_{6,r})s_r}{s_r+r}}>H_{4,r}\underline{\eta}_r^{\frac{-s_r(j+1)}{s_r+r}}\geq\phi_{j,r}
\]
For this integer $H_{6,r}$ and $j\geq 1$, we also have
\begin{eqnarray*}
\phi_{j+H_{6,r},r}&\leq& H_{4,r}\underline{\eta}_r^{\frac{-(j+H_{6,r}+1)s_r}{s_r+r}}= H_{4,r}\underline{\eta}_r^{\frac{-(H_{6,r}+1)s_r}{s_r+r}}\underline{\eta}_r^{\frac{-js_r}{s_r+r}}\\&\leq& H_{5,r}^{-1}H_{4,r}\underline{\eta}_r^{\frac{-(H_{6,r}+1)s_r}{s_r+r}}\phi_{j,r}.
\end{eqnarray*}
We set $N_{j,r}:=\phi_{[\theta^{-1}+jH_{6,r}],r}$ and $H_{7,r}:=H_{5,r}^{-1}H_{4,r}\underline{\eta}_r^{\frac{-(H_{6,r}+1)s_r}{s_r+r}}$. Then we have
\[
N_{j,r}<N_{j+1,r}\leq H_{7,r} N_{j,r},\;\;N_{j,r}^{\frac{r}{s_r}}e_{N_{j,r}}^r(\mu)\geq\widetilde{D} (H_{3,r}P_r Q_r^{-1})^{-\frac{s_r+r}{s_r}}.
\]
For each $k\geq\phi_1$, we choose $j$ such that $k\in[ N_{j,r}, N_{j+1,r})$. Then using Theorem 4.12 of \cite{GL:00}, we deduce
\begin{eqnarray*}
\underline{Q}_r^{s_r}(\mu)&=&\liminf_{k\to\infty}k^{\frac{r}{s_r}}e_{k,r}^r(\mu)\geq\liminf_{j\to\infty} N_{j,r}^{\frac{r}{s_r}}
e_{ N_{j+1,r},r}^r(\mu)\\&\geq& (H_{7,r})^{-\frac{r}{s_r}}\liminf_{j\to\infty} N_{j+1,r}^{\frac{r}{s_r}}
e_{ N_{j+1,r},r}^r(\mu)\\&\geq&(H_{7,r})^{-\frac{r}{s_r}}\widetilde{D} (H_{3,r}P_r Q_r^{-1})^{-\frac{s_r+r}{s_r}}.
\end{eqnarray*}
This completes the proof of the proposition.
\end{proof}
\emph{Proof of Theorem \ref{mthm1}}
It is an immediate consequence of Proposition \ref{pre1} and \ref{pre3}.
\section*{References}
|
\section{Introduction}
In the realm of the standard field theories, more general fields than symmetric tensors are rarely used. It is partly because we are living in four-dimensional spacetime where any fundamental particle can be described by symmetric tensor(-spinor) fields.
Hence, mixed-symmetry fields --- fields whose indices have more general permutation symmetry than the fully symmetric ones --- would be indispensable only in describing certain physics of higher dimensions, such as the physics of String Theory.
Actually, the infinite tower of massive excitations in String Theory
carry mixed-symmetry representations in general and
their presence is crucial for the consistency of various stringy dualities.
Moreover, mixed-symmetry fields
might be useful, although not necessary, even in four dimensions,
and their use may open a new avenue to unexplored land of physics.
The two-derivative Lagrangian for a massless mixed-symmetry field in flat spacetime is first given for the simplest case by Curtright \cite{Curtright:1980yk} and for generic case by Labastida
\cite{Labastida:1986gy,Labastida:1986ft,Labastida:1987kw}
and then further investigated in \cite{Metsaev:1995re,Metsaev:1997nj,Brink:2000ag,Alkalaev:2003hc,deMedeiros:2003qel,
Francia:2005bv,Alkalaev:2006hq,Alkalaev:2005kw,Alkalaev:2006rw, Skvortsov:2008vs, Skvortsov:2008sh, Campoleoni:2008jq, Campoleoni:2009gs,Campoleoni:2009je,
Reshetnyak:2010ga,Alkalaev:2010af,Boulanger:2011qt,Burdik:2011cb,
Campoleoni:2012th,Bekaert:2015fwa,Reshetnyak:2016sng}.
Mixed-symmetry fields can be classified according to the
symmetry property of the index permutations,
hence can be associated with Young diagrams.
The Lagrangian has in general a number of gauge symmetries,
depending on the shape of the Young diagram associated with the field under consideration.
One of the interesting and non-trivial properties of mixed-symmetry fields is
that their flat space Lagrangian does not admit a smooth
deformation towards the background with non-vanishing cosmological constant
\cite{Metsaev:1995re,Metsaev:1997nj,Brink:2000ag}:
around (A)dS background,
there is no two-derivative mixed-symmetry Lagrangian
respecting all the gauge symmetries available in flat space.
Instead, one can choose to preserve only one gauge symmetry
and the choice determines the mass squared term to
a specific value in units of cosmological constant.
In this paper, we shall focus on the mixed-symmetry field
associated with the Young diagrams having two columns.
In a sense, the number of columns plays the role of `spin' hence
we are considering here only `spin two' cases.
Among generic two-column cases,
we will analyze the mixed-symmetry field corresponding to the simple hook Young diagram
$\tiny\yng(2,1)$ \cite{Curtright:1980yk} in great details
since it already contains all the essential features of the more general two-columns cases.
The simple hook field is of particular interest because
it appears in the first order formulation of Gravity through the spin connection.
There have been various attempts to describe Gravity by the connection
or related object
(see for instance the recent discussion \cite{Basile:2015jjd}
and references therein).
We shall also discuss
how these attempts can be understood from the perspective of
the physics of mixed-symmetry field.
The main target of the current paper is the identification and the analysis
of the higher-derivative action of mixed-symmetry field having Weyl transformation --- that is, the trace shift --- as its symmetry. In the more familiar case of symmetric spin-two field,
the Weyl action is nothing but the linearization
of the four-derivative Weyl gravity.
The latter can be decomposed around (A)dS
into the massless and partially-massless modes (see \cite{Deser:2012qg} for
the related discussion).
In the case of symmetric spin-$s$ field, the Weyl action involves $2s$ derivatives
which can be split into $s$ different modes each of which
is described by a specific two-derivative action with a certain gauge symmetry
\cite{Joung:2012qy,Metsaev:2014iwa,Nutma:2014pua}:
they are partially-massless spin $s$ modes of depth $0$ to $s-1$\,.
In fact, these are all the known short representations (of isometry group) containing
the helicity $s$ mode as the highest one.
This pattern suggests that the Weyl action of certain type of field ---
fully symmetric or mixed symmetry ---
could describe all the short representations associated with that field
and each of these short representations can be realized as a two-derivative Lagrangian with a certain gauge symmetry (see \cite{Joung:2012qy} for some related discussions).
In this paper, we examine this idea with
the mixed-symmetry fields of
two-column Young diagrams.
The Weyl action of the field associated with a Young diagram having $s$ columns ought to
involve $2s$ derivatives. Hence, we expect it to describe $s$ different short modes.
The pattern of short representations involved might be non-trivial
and the necessary analysis would be lengthy for generic Young diagrams.
We reserve the general analysis for the future work
and focus here only on the two-column cases where,
irrespectively of the height of each column in the Young diagram,
the Weyl action has four derivatives.
Let us provide more details on what has been done in this paper.
We consider the fields having the symmetry of two-column Young diagram,
\be
\parbox{25pt}{
\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2.4);
\draw (0.4,2.4) -- (0.8,2.4) -- (0.8,0.8) -- (0.4,0.8);
\node at (0.2,1.2){${\st p}$};
\node at (0.6,1.6){${\st q}$};
\end{tikzpicture}}\,.
\ee
There are two short representations in (A)dS described by
two distinct Lagrangians with different gauge symmetries:
the gauge parameters have the index symmetry of Young diagram, either
\be
\parbox{45pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2);
\draw (0.4,2) -- (0.8,2) -- (0.8,0.4) -- (0.4,0.4);
\draw [dotted] (0,0) rectangle (0.4,-0.4);
\node at (-0.3,1){${\st p-1}$};
\node at (0.6,1.2){${\st q}$};
\end{tikzpicture}}\qquad
{\rm or}\qquad
\parbox{40pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2.4);
\draw (0.4,2.4) -- (0.8,2.4) -- (0.8,1.2) -- (0.4,1.2);
\draw [dotted] (0.4,1.2) rectangle (0.8,0.8);
\node at (0.2,1.2){${\st p}$};
\node at (1.1,1.8){${\st q-1}$};
\end{tikzpicture}}\,.
\label{hk gg sym}
\ee
In the flat limit, the two Lagrangians coincide and enjoy both of the gauge symmetries.
Now considering the Weyl action of this field in flat space,
we find that there is a unique four-derivative action invariant with respect to both gauge symmetries and a trace shift symmetry with the parameter,
\be
\parbox{60pt}{
\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2);
\draw (0.4,2) -- (0.8,2) -- (0.8,0.8) -- (0.4,0.8);
\draw [dotted] (0,0) rectangle (0.4,-0.4);
\draw [dotted] (0.4,0.8) rectangle (0.8,0.4);
\node at (-0.3,1){${\st p-1}$};
\node at (1.1,1.4){${\st q-1}$};
\end{tikzpicture}}\,.
\ee
We show that this flat space Weyl action
now admits a smooth deformation to the background with non-vanishing cosmological constant,
as opposed to the two-derivative one.
Moreover, this four-derivative action describes
two short-representation modes corresponding to the gauge symmetries
\eqref{hk gg sym}.
In the case where the heights of two columns coincide,
the Weyl action describes,
on top of the usual massless field,
new degrees of freedom (DoF) having a two-derivative gauge symmetry
similar to the partially-massless spin two field.
Therefore, our analysis shows
that the two-column cases precisely fit in the pattern discussed before.
Further, we study two-derivative massive deformations of the Weyl action. In (A)dS there are two distinct mass deformations, which preserve one or the other of the symmetries \eqref{hk gg sym}. We show that the massive action does not admit a smooth (A)dS deformation, similarly to the conventional two-derivative actions. Two different deformations have different spectra around (A)dS background. We observe that the sign of the coefficients in front of the free actions depend on the sign of the cosmological constant. In special dimension \mt{d=p+q+1}, one of the massive deformations is unitary around dS and non-unitary around AdS, while the other deformation follows the opposite pattern --- unitary around AdS and non-unitary around dS. In other words, the unitarity requirement selects one of
the massive deformations, and this choice is different for different signs of cosmological constant.
The plan of the paper is as follows.
We focus first on the simple hook case in Section \ref{sec:2}.
After a detailed review of the two-derivative systems in Section \ref{sec:2.1}
and \ref{sec:2.2},
we construct and discuss the four-derivative Weyl action in Section \ref{sec:2.3}.
The result of the simple hook is generalized in Section \ref{sec:3}
towards more general two-column cases.
We provide separate discussions on
the different height case (Section \ref{sec:3.1}) and
the equal height case (Section \ref{sec:3.2}).
Finally, we provide various discussions, in particular, related to New Massive Gravity
\cite{Bergshoeff:2009hq,Bergshoeff:2012ud,Joung:2012sa,Dalmazi:2012dq},
in Section \ref{sec:3.3}.
\section{Simple Hook}
\label{sec:2}
Let us first consider the simplest mixed-symmetry field
$\phi_{\m\n,\r}$\,, corresponding to the simple hook Young diagram,
\be
\small{\young(\mu \rho,\nu)}\,.
\ee
The shape of Young diagram dictates the symmetries of the field
under index permutations. The symmetries are
\be
\phi_{\m\n,\r}=-\phi_{\n\m,\r}\,,
\qquad \phi_{\m\n,\r}+\phi_{\n\r,\m}+\phi_{\r\m,\n}=0\,.
\ee
Notice that we work in the base where the anti-symmetry of indices in each column is manifest.
One can equally work with the symmetric base, but for the construction of the
Lagrangian we find the antisymmetric base more advantageous.
The Young diagram is fixing not only how fields transform
under index permutations but also under Lorentz transformations.
Hence, it defines a representation under Lorentz group $SO(1,d-1)$
that the off-shell field carries.
\subsection{Einstein Action in Flat space}
\label{sec:2.1}
In flat spacetime, one can find a proper set of on-shell conditions
--- or, an action principle ---
that makes the field carry the same Young diagram representation
but now under the little group $SO(d-1)$ or $SO(d-2)$\,,
depending on whether the field is massive or massless\footnote{Only the representations of compact subgroup $SO(d-2)$ of massless little group $ISO(d-2)$ are relevant for our discussion in this paper.}, respectively.
When given Young diagram representation of Lorentz group does not exist
for the little group, an interesting mechanism may emerge when the propagating modes carry certain Young diagram representations of little group which are different from the Young diagram representation of the Lorentz group that the off-shell field carries. For any mixed-symmetry Young diagram this happens in space-time dimensions lower than certain value.
We shall return to this point later, but work for a while with an arbitrary $d$\,.
In \cite{Curtright:1980yk},
Curtright constructed the free action for the hook field $\phi_{\m\n,\r}$
which describes massless DoF carrying the hook Young diagram representation
of the little group $SO(d-2)$. Note that such Young diagram exists only when $d$ is greater than four.
By introducing the scalar product,
\be
\la A\,|\,B\ra=\frac1{m!\,n!}\,\int d^dx\sqrt{|g|}\,A^{\m_1\cdots \m_m,\n_1\cdots \n_n}(x)\,B_{\m_1\cdots \m_m,\n_1\cdots \n_n}(x)\,,
\label{scalar pr}
\ee
the action for the hook field can be written as
\be
\mathcal S^{\rm\sst flat}_{\rm\sst E}[\phi]=\la \phi\,|\, \cG\,\phi\ra\,.
\label{sh}
\ee
Here the `Einstein tensor' $\mathcal G\,\phi$ is defined
through the `Ricci tensor' $\mathcal F\,\phi$ as
\be
\left(\cG\,\phi\right)_{\m\n,\r}=\left(\cF\,\phi\right)_{\m\n,\r}
-\eta_{\r[\m}\,\left(\cF\,\phi\right)_{\n]\l,}{}^{\l}\,.
\ee
In this paper, the round/square brackets denote full
symmetrization/anti-symmetrization with weight one,
e.g. $A_{[[\m\n]]}=A_{[\m\n]}=-A_{[\n\m]}$.
The `Ricci tensor' itself is given by
\be
\left(\cF\,\phi\right)_{\m\n,\r}=\partial^2\,\phi_{\m\n,\r}
+2\,\partial_{[\mu}\,\partial^\l\, \phi_{\n]\l,\r}
-\partial_\r\,\partial^\l\,\phi_{\m\n,\l}
-2\,\partial_{\r}\,\partial_{[\m}\,\phi_{\n]\l\,,}{}^{\l}\,,
\ee
which extends the Fierz-Pauli massless spin two equation to the hook field.
By making use of the generalized Kronecker delta,
\be
\d_{\m\n\a\b}^{\r\l\g\d}
=4!\,\d^{\r}_{[\m\phantom{]}}
\d^{\l\phantom{]}}_{\n\phantom{]}}
\d^{\g}_{\a\phantom{]}}
\d^{\d\phantom{]}}_{\b]}\,,
\ee
the Einstein tensor can be written in a more compact form \cite{Curtright:1980yk} as
\be
\left(\cG\,\phi\right)_{\m\n,}^{\quad\r}=-\frac12\,
\d_{\m\n\a\b}^{\r\l\g\d}\,\partial^{\a}\,\partial_{\l}\,\phi_{\g\d,}^{\quad\b}\,.
\label{Kr flat}
\ee
The action \eqref{sh}
has two gauge symmetries,
\be
\d_\e\, \phi_{\m\n,\r}=2\,\partial_{[\m}\,\e_{\n]\r},\qquad
\d_\th\,\phi_{\m\n,\r}=\partial_{\r}\,\th_{\m\n}-\partial_{[\m}\,\th_{\n]\r}\,,
\label{flat gs}
\ee
generated by symmetric and anti-symmetric gauge parameters:
\be
\e_{\m\n}=\e_{\n\m}\,,\qquad
\th_{\m\n}=-\th_{\n\m}\,.
\ee
The above symmetry can be understood diagrammatically as
\be
\delta_\e\,\young(\phi\phi,\phi)=\young(\e\e,\partial)\,,
\qquad
\delta_\th\,\young(\phi\phi,\phi)=\young(\th\partial,\th)\,.
\ee
This gauge symmetry is reducible admitting the gauge-for-gauge symmetry,
\be
\e_{\m\n}(\xi)=\partial_{(\m}\,\xi_{\n)}\,,\qquad
\th_{\m\n}(\xi)=\partial_{[\m}\,\xi_{\n]}\,,\label{gfg}
\ee
which reads in terms of Young diagram,
\be
\young(\e\e)=\young(\xi\partial)\,,
\qquad
\young(\th,\th)=\young(\xi,\partial)\,.
\ee
The gauge symmetries generated by $\e_{\m\n}(\xi)$ and $\th_{\m\n}(\xi)$
have the same form,
\be
\young(\xi\partial,\partial)\,,
\ee
and therefore sum up to zero for an appropriate choice of the relative coefficient.
The structure of gauge symmetry tells us the number of DoF of this system.
According to the covariant counting (see e.g \cite{Kaparulin:2012px}), we get
\be
\cS_{\rm\sst E}\quad : \quad \yng(2,1)_{\sst GL(d)} \ominus 2\left(\yng(2)_{\sst GL(d)} \oplus \yng(1,1)_{\sst GL(d)}\right)\oplus 3\,\yng(1)_{\sst GL(d)}\,,
\ee
where we subtract twice the DoF associated to gauge parameters as usual
and put back three times of the DoF of gauge-for-gauge parameter.
The above can be recast into the counting in the Hamiltonian analysis,
\be
\yng(2,1)_{\sst SO(d-1)} \ominus \left(\yng(2)_{\sst SO(d-1)} \oplus \yng(1,1)_{\sst SO(d-1)}\right)
\oplus \yng(1)_{\sst SO(d-1)}\,,
\ee
where we subtract once the DoF for each of traceless gauge parameters
and put back once that of gauge-for-gauge.
This again can be rearranged into
\be
\yng(2,1)_{\sst SO(d-1)} \ominus \left(\yng(2)_{\sst SO(d-2)} \oplus \yng(1,1)_{\sst SO(d-2)}
\oplus \yng(1)_{\sst SO(d-2)}\right) =
\yng(2,1)_{\sst SO(d-2)}\,.
\ee
which is the DoF of massive hook
minus those of massless fields $\tiny\yng(2)\,,\ \yng(1,1)$ and $\Box$\,.
Hence, this gives the interpretation of obtaining massless hook
from the massless limit of massive hook by eliminating other `lower spin' components
the latter involves.
Eventually, we end up with the DoF of the simple hook of the little group $SO(d-2)$\,.
Hence, the DoF of the hook field can be conveniently counted,
in flat space, through the number of components of the Young diagram representations of the little group.
\subsection{Einstein Action in (A)dS}
\label{sec:2.2}
Let us move on to the (A)dS background,\footnote{Explicit analysis making use of Stueckelberg fields
can be found in \cite{Zinoviev:2002ye}.}
and consider the analogous Lagrangian to
the flat space one \eqref{sh}.
One can think of the same expression for the Lagrangian as in \eqref{sh} where all flat partial derivatives $\partial_\m$
are replaced by the (A)dS covariant derivatives $\nabla_\m$\,,
but this definition of Lagrangian
has an ambiguity coming from the commutators of $\nabla_\m$'s
which give terms proportional to (A)dS curvature.
The ambiguous term is hence a non-derivative mass-like term proportional to the cosmological constant,
and it ought to be determined by the invariance of the
action with respect to either of the gauge symmetries,
\be
\d^{\sst \L}_\e\, \phi_{\m\n,\r}=2\,\nabla_{[\m}\,\e_{\n]\r},\qquad
\d^{\sst \L}_\th\,\phi_{\m\n,\r}=\nabla_{\r}\,\th_{\m\n}-\nabla_{[\m}\,\th_{\n]\r}\,.
\label{hk sym}
\ee
By examining the gauge invariance with a possible mass-like term,
one can realize that it is impossible to preserve both of the gauge symmetries.
Depending on which symmetry we want to keep, the mass-like term is determined with a different
mass-squared coefficient \cite{Metsaev:1995re,Metsaev:1997nj}.
Let us explictly determine the mass-term or equivalently the Einstein tensor $\cG_{\m\n,\r}$
of the (A)dS Lagrangian defined in the same manner as in \eqref{sh}.
To proceed, let us first fix our conventions:
\be
[\nabla_\m, \nabla_\n]\,V_{\l}^{\r}=R_{\m\n,\l}{}^{\s}\,V_{\s}^{\r}-
R_{\m\n,\s}{}^{\r}\,V_{\l}^{\s}\,,\qquad
R_{\m\n,\l}{}^{\r}=\frac{2\,\L}{(d-1)(d-2)}\left(g^{\sst\L}_{\m\l}\,\d_{\n}^{\r}-g^{\sst\L}_{\n\l}\,\d_{\m}^{\r}\right),
\ee
where $\L$ is the cosmological constant and $g^{\L}$ is the (A)dS metric.
By requiring the Lagrangian to be invariant under the transformation $\d^{\sst\L}_\e$,
we determine the Einstein tensor as
\be
\left(\cG^{\sst \L}_\syd\,\phi\right)_{\m\n,}{}^{\r}=-\frac12\,
\d_{\m\n\a\b}^{\r\l\g\d}\,\nabla^{\a}\,\nabla_{\l}\,\phi_{\g\d,}^{\quad\b}\,,
\label{Einstein sym}
\ee
which has exactly the same form as the flat space one \eqref{Kr flat} except for
the replacement of $\partial_\m$ by $\nabla_\m$\,. Note that, as opposed to flat space case, the order of derivatives is important in the expression \eqref{Einstein sym}.
This system admits a gauge-for-gauge symmetry
\be
\e_{\m\n}=\left(\nabla_{(\m}\,\nabla_{\n)}+\frac{2\L}{(d-1)(d-2)}\,g^{\sst \L}_{\m\n}\right)\zeta\,,
\ee
which corresponds to the partially-massless spin-two transformation.
On the other hand, for the invariance under the transformation $\d^{\sst\L}_\th$\,,
the Einstein operator should acquire a mass-like term,
\be
\cG^{\sst \L}_{\ayd}=
\cG^{\sst \L}_{\syd}-\,m_\L^2
\,\cI,\label{Einstein antisym}
\ee
where the mass-squared parameter is fixed by
\be
m_\L^2=-\frac{4(d-3)}{(d-1)(d-2)}\,\L\,,
\label{mass term}
\ee
and the mass-term operator by
\be
(\cI\,\phi)_{\m\n,}{}^{\r}=\frac12\,\d_{\m\n\a}^{\r\b\g}\,\phi_{\b\g,}{}^{\a}
=\phi_{\m\n,}^{\quad\r}-2\,\d^{\,\r}_{[\n}\,\phi_{\m]\a,}{}^{\a}\,.
\ee
Both of the actions $\cG^{\sst \L}_{\syd}$ and $\cG^{\sst \L}_{\ayd}$ vanish identically for $d\leq 3$.
Notice that the mass-squared parameter is positive in AdS and negative in dS.
In both cases, the equation with higher mass-squared term is unitary
--- $\cG^{\sst \L}_{\ayd}$ for AdS and $\cG^{\sst\L}_{\syd}$ for dS ---
whereas the other is non-unitary. See \cite{Zinoviev:2002ye} for more details.
To recapitulate, the simple hook field in (A)dS space cannot have a two-derivative
action respecting both of the gauge symmetries \eqref{hk sym},
but either the action $\cS^{\sst \L}_{\sst\rm E\,\syd}$ with only the symmetric parameter gauge symmetry
or the action $\cS^{\sst \L}_{\sst\rm E\,\ayd}$ with only the anti-symmetric parameter gauge symmetry:
\be
\cS^{\sst \L}_{\sst\rm E\,\syd}[\phi]=\la\phi\,|\,\cG^{\sst \L}_{\syd}\,\phi\ra\,,
\qquad
\cS^{\sst \L}_{\sst\rm E\,\ayd}[\phi]=\la\phi\,|\,\cG^{\sst \L}_{\ayd}\,\phi\ra\,.
\label{Einstein hk}
\ee
Let us examine the DoF of the above systems. We first
consider the action $\cS^{\sst \L}_{\sst\rm E\,\syd}$\,.
The DoF of the system can be counted
in terms of the $GL(d)$ Lorentz covariant tensors as
\ba
\cS^{\sst \L}_{\sst\rm E\,\syd}\quad : \quad && \yng(2,1)_{\sst GL(d)} \ominus
2\,\yng(2)_{\sst GL(d)}\oplus 3\,\bullet \nn
&&\hspace{50pt}
\ominus\ \yng(1,1)_{\sst GL(d)}
\oplus\yng(1)_{\sst GL(d)}\,,
\ea
which has a non-trivial pattern due to the mixture of first- and second-class constraints
as well as gauge-for-gauge.
Instead, in the Hamiltonian analysis, we find a
simpler pattern,
\be
\yng(2,1)_{\sst SO(d-1)} \ominus \yng(2)_{\sst SO(d-1)}\oplus \bullet\,,
\ee
where we simply remove once the DoF of the traceless gauge parameter
and put back that of gauge-for-gauge.
We can further decompose these DoF as propagation on light-cone, ending up with
\be
\cS^{\sst \L}_{\sst\rm E\,\syd}\quad:\quad \yng(2,1)_{\sst SO(d-2)}\oplus
\yng(1,1)_{\sst SO(d-2)}\,.
\ee
Similarly, the number of DoF of the action $\cS^{\sst \L}_{\sst\rm E\,\ayd}$
can be counted in terms of $GL(d)$ covariant tensors as
\ba
\cS^{\sst \L}_{\sst\rm E\,\ayd}\quad : \quad && \yng(2,1)_{\sst GL(d)} \ominus
2\,\yng(1,1)_{\sst GL(d)}
\nn
&&\hspace{50pt} \ominus\
\yng(2)_{\sst GL(d)}\oplus \yng(1)_{\sst GL(d)}\,.
\ea
The Hamiltonian analysis gives the same result, in terms of $SO(d-1)$ branching:
\be
\yng(2,1)_{\sst SO(d-1)} \ominus \yng(1,1)_{\sst SO(d-1)}\,.
\ee
In terms of light-cone DoF, the equation \eqref{Einstein antisym} propagates
\be
\cS^{\sst \L}_{\sst\rm E\,\ayd}\quad:\quad \yng(2,1)_{\sst SO(d-2)}\oplus \yng(2)_{\sst SO(d-2)}\,.
\ee
Hence, compared to the flat space case, the hook action
$\cS^{\sst \L}_{\sst\rm E\,\syd}$ describes extra DoF corresponding to the light-cone propagation of a
massless two-form field $\tiny\yng(1,1)$\,,
while $\cS^{\sst \L}_{\sst\rm E\,\ayd}$ describes extra DoF of massless spin-two field $\tiny\yng(2)$\,.
The general pattern of the decomposition of modes in flat limit for
an arbitrary mixed-symmetry field has been conjectured in \cite{Brink:2000ag}
and proved in \cite{Boulanger:2008up,Boulanger:2008kw,Alkalaev:2009vm,Alkalaev:2011zv}.
Let us note that in four dimensions, the hook mode identically vanishes hence
the action $\cS^{\sst\L}_{\sst\rm E\,\ayd}$ describes only the massless spin two mode.
We will comment on this action in the section \ref{sec:3.1}, where a generalization to any dimensions with an off-shell field of type $[d-2,1]$ will be discussed in detail and the relation with the recent work \citep{Basile:2015jjd} will be clarified.
\subsection{Weyl Action}
\label{sec:2.3}
Even though there is no two-derivative action preserving both of the gauge symmetries of \eqref{hk sym} in (A)dS, there may exist a higher-derivative action which is invariant with respect to both of the symmetries.
It turns out that the four-derivative action
invariant under both symmetries has a simple form,
\be
\cS^{\sst \L}_{\rm\sst W}[\phi]
=-\la \phi\,|\,\cG^{\sst\L}_{\syd}\,\cI^{-1}\,\cG^{\sst\L}_{\ayd}\,\phi\ra\,,
\label{S4}
\ee
where the overall sign is chosen for the positive definite Euclidean action.
This action makes use of both Einstein tensors
as well as the inverse mass-term operator $\cI^{-1}$ given by
\be
(\cI^{-1}\,\phi)_{\m\n,}{}^{\r}
=\phi_{\m\n,}{}^{\r}-\frac{2}{d-2}\,\d^{\,\r}_{[\n}\,\phi_{\m]\a,}{}^{\a}\,.
\ee
All the operators $\cG^{\sst\L}_{\syd}$, $\cG^{\sst\L}_{\ayd}$ and $\cI$ are self-adjoint with respect to
the scalar product \eqref{scalar pr}\,:
\be
\la f\,|\,\cO\,g\ra=\la \cO\,f\,|\,g\ra\,,
\qquad
\cO=\cG^{\sst \L}_{\syd},\ \cG^{\sst \L}_{\ayd},\ \cI\,.
\ee
Since the Einstein operators $\cG^{\sst\L}_\syd$ and $\cG^{\sst\L}_\ayd$ differ by
a mass-like term,
it is easy to show that
\ba
\cS^{\sst \L}_{\sst\rm W}[\phi]
\eq - \la\phi\,|\left(\cG^{\sst \L}_{\syd}\,\cI^{-1}\,\cG^{\sst \L}_{\syd}
-m_\L^2\,\cG^{\sst \L}_{\syd}\right)\phi\ra \nn
\eq - \la\phi\,|\left(\cG^{\sst \L}_{\ayd}\,\cI^{-1}\,\cG^{\sst \L}_{\ayd}
+m_\L^2\,\cG^{\sst \L}_{\ayd}\right)\phi\ra\,.
\label{S4m}
\ea
In the first line, the action is manifestly invariant under the gauge symmetry $\d^{\sst\L}_\e$
with the symmetric parameter,
whereas the second line is manifestly invariant under the $\d^{\sst\L}_\th$ one
with anti-symmetric parameter.
In the flat limit, the four-derivative action \eqref{S4} reduces to
\be
\cS^{\rm\sst flat}_{\rm\sst W}[\phi]=-\la\phi\,|\,\cG\,\cI^{-1}\,
\cG\,\phi\ra\,,
\label{S4 flat}
\ee
which is actually the unique action invariant under both gauge symmetries
as well as the Weyl transformation,
\be
\delta_\a\,\phi_{\m\n,\r}=\eta_{\r[\m}\,\a_{\n]}\,.
\ee
This is the analogue of the Weyl gravity action, which is uniquely fixed by diffeomorphism and Weyl symmetries.
Hence, one can regard $\cS^{\sst\rm flat}_{\sst\rm W}$ as
the Weyl action for the simple hook field.
In four dimensions, this action has the conformal invariance,
and its form has been determined in \cite{Metsaev:2015rda}
together with other two-column mixed symmetry fields.
Coming back to the AdS action $\cS^{\sst\L}_{\sst\rm W}$ of \eqref{S4},
one may wonder whether it
also admits a Weyl symmetry:
\be
\delta^{\sst\L}_{\a}\,\phi_{\m\n,\r}=g^{\sst\L}_{\r[\m}\,\a_{\n]}\,.
\label{Weyl}
\ee
One can check the invariance of \eqref{S4} under \eqref{Weyl}
by a brute force computation,
but there is in fact a simpler way to see it. Since any linear combination of gauge symmetries should be a gauge symmetry, taking the following gauge parameters,
\ba
\e_{\m\n}(\a)=\nabla_{(\m}\,\a_{\n)}\,,\qquad
\th_{\m\n}(\a)=-\frac13\,\nabla_{[\m}\,\a_{\n]}\,,
\ea
we immediately get the Weyl transformation,
\ba
(\d_{\e(\a)}+\d_{\th(\a)})\,\phi_{\m\n,\r}
=\frac{\L}{(d-1)(d-2)}\,g^{\sst\L}_{\r[\m}\,\a_{\n]}\,.
\ea
This is due to the cancellation of derivative terms.
Hence, in a sense,
the Weyl symmetry arises
as the (A)dS remnant of the gauge-for-gauge symmetry
in flat space \eqref{gfg}.
We conclude that {\it any theory of hook field in (A)dS, which is invariant with respect to both gauge symmetries is also Weyl invariant}.
This conclusion can be generalized to any mixed symmetry fields in (A)dS.
Therefore, any theory, that is invariant with respect to two distinct gauge transformations of mixed-symmetry field, has to enjoy certain Weyl symmetry.
This argument is true even for symmetric fields.
Let us take the example of spin-two field $\phi_{\m\n}$\,:
if there exists a theory of $\phi_{\m\n}$
invariant with respect to both massless and partially-massless
gauge symmetries,
\be
\d \phi_{\m\n}=\nabla_{(\m}\e_{\n)}+\nabla_{\m}\nabla_{\n}\s+\frac{2\L}{(d-1)(d-2)}\,
g^{\L}_{\m\n}\,\s\,,
\ee
then such a theory will also admit a Weyl symmetry.
For the demonstration,
it is enough to set the gauge parameter $\e_\m=-\nabla_\m\s$
to get $\d \phi_{\m\n}=\frac{2\L}{(d-1)(d-2)}\,g^{\L}_{\m\n}\,\s$. The action having both of massless and partially-massless gauge symmetries can be constructed analogously to \eqref{S4},
using the massless Einstein operator $\cG^{\sst \L}_m$
and the partially-massless one $\cG^{\sst \L}_{pm}$\,, as
\be
\cS_{\rm\sst W}^{\sst \L}[\phi]
=-\la \phi\,|\,\cG^{\sst \L}_{\sst m}\,
\cI_{\rm\sst FP}^{-1}\,\cG^{\sst \L}_{\sst pm}\,\phi\ra\,,
\label{conformal gravity}
\ee
where $\cG^{\sst \L}_{pm}$ differs $\cG^{\sst \L}_m$ by
a particular mass term given through a Fierz-Pauli operator $\cI_{\rm\sst FP}$\,.
This four-derivative action \eqref{conformal gravity}
is nothing but Conformal Gravity linearized around (A)dS background.
As the reader might notice,
this is the inverse logic
of the discussion in \cite{Deser:2012qg}. In general, one can use this argument to support the conjecture of \cite{Joung:2012qy} about the spectrum of higher-derivative Weyl-like actions for symmetric higher spin fields.
\subsubsection{Curvature Formulation}
In order to grasp some geometrical intuitions,
let us revisit our construction in terms of (generalized) curvatures.
First of all, notice that an analogue of the Riemann curvature can be constructed for the hook field.
It is a tensor having the symmetry of the $GL(d)$ Young diagram $\tiny\yng(2,2,1)$ given in flat space by
\be
(\cR\,\phi)_{\m_1\m_2\m_3,}{}^{\n_1\n_2}=6\,\partial^{[\n_1}\,\partial_{[\m_1}\,\phi_{\m_2\m_3],}{}^{\n_2]}\,,
\ee
and is invariant with respect to both gauge symmetries
\eqref{flat gs}. Around (A)dS background,
it is impossible to construct a curvature
invariant under both gauge transformations \eqref{hk sym}.
Instead, we can consider the curvature,
\be
(\cR^{\sst\L}_{\syd}\,\phi)_{\m_1\m_2\m_3,}{}^{\n_1\n_2}
=6\,\nabla^{[\n_1}\,\nabla_{[\m_1}\,\phi_{\m_2\m_3],}{}^{\n_2]}\,,
\label{Curv sym}
\ee
which is invariant under the gauge transformation
$\delta^{\sst \L}_{\e}$ with symmetric parameter.
In order to get a curvature that is invariant
with respect to the gauge symmetry $\delta^{\sst \L}_{\th}$ with anti-symmetric parameter,
one should deform the above curvature into
\be
(\cR^{\sst\L}_{\ayd}\,\phi)_{\m_1\m_2\m_3,}{}^{\n_1\n_2}
=(\cR^{\sst\L}_{\syd}\,\phi)_{\m_1\m_2\m_3,}{}^{\n_1\n_2}
+\frac{24\,\L}{(d-1)(d-2)}\,\d_{[\m_1}^{[\n_1}\,\phi_{\m_2\m_3],}{}^{\n_2]}\,.\label{Curv antisym}
\ee
These two curvatures can be directly related to the Einstein actions \eqref{Einstein hk}:
the Einstein tensors \eqref{Einstein sym} and \eqref{Einstein antisym} are
given through the curvatures
\eqref{Curv sym} and \eqref{Curv antisym} respectively as
\be
(\cG^{\sst\L}_{\syd}\,\phi)_{\m\n,}{}^\r
=-\frac1{12}\,\d_{\m\n\a\b}^{\r\g\d\s}\,(\cR^{\sst \L}_{\syd}\,\phi)_{\g\d\s,}{}^{\a\b}\,,
\qquad
(\cG^{\sst\L}_{\ayd}\,\phi)_{\m\n,}{}^\r
=-\frac1{12}\,\d_{\m\n\a\b}^{\r\g\d\s}\,(\cR^{\sst \L}_{\ayd}\,\phi)_{\g\d\s,}{}^{\a\b}\,,
\ee
in the same manner as Einstein tensor is given through the Riemann curvature: remind, that Einstein tensor can be written as $G_{\m}^{\n}=R_{\m}^{\n}-\tfrac12\, \d_{\m}^{\n}\,R
=-\frac14\, \d_{\m\l\r}^{\n\a\b}\,R_{\a\b,}^{\quad\,\l\r}$, where $R_{\a\b,\l\r}$ is Riemann curvature, $R_{\m\n}$ and $R$ are Ricci curvature and its trace respectively.
Hence, it is clear how each curvature is related to the two-derivative Einstein
action for the simple hook field.
Let us now move on to the four-derivative action $\cS^{\sst\L}_{\sst\rm W}$\,,
which admits Weyl symmetry. In the case of gravity, the Weyl gravity action
is simply given by the square of Weyl tensor.
Hence, the action $\cS^{\sst\L}_{\sst\rm W}$ may also admit such a description.
We can first define the Weyl tensor from the curvature
as the traceless part of the latter. This can be conveniently done
by introducing the traceless projector $\cT$ whose explicit form
is not necessary for now.
In terms of $\cT$, the Weyl tensor can be determined as
\be
\cW=\mathcal T\,\cR_{\syd}^{\sst\L}=\mathcal T \,\cR_{\ayd}^{\sst\L}\,.
\label{WeylTensorForHook}
\ee
Remark that the definition of the Weyl tensor
does not distinguish between curvatures $R^{\sst\L}_\syd$ and $R^{\sst\L}_\ayd$
because the difference between two curvatures is precisely a trace term
which is projected out under the action of $\mathcal T$.
Therefore, the Weyl tensor is invariant under both gauge symmetries.
As the invariance with respect to both gauge symmetries implies the invariance
under Weyl symmetry transformation, the tensor $(\cW\,\phi)_{\m\n\r,\l\k}$
can be rightfully referred to as Weyl tensor.
Again, mimicking the Weyl gravity case,
we can consider a four-derivative action, that is the square of the Weyl tensor.
Since the four-derivative action invariant under both gauge symmetries is unique,
they should be simply related with a proportionality constant as
\be
\cS^{\sst \L}_{\sst\rm W}[\phi]
=-\frac{d-3}{d-4}\,\la \cW\,\phi\,|\,\cW\,\phi\ra\,.
\label{WeylSquaredHook}
\ee
The uniqueness of the four derivative action with both symmetries in (A)dS can be checked directly, but can be also understood by the following simple argument: As we have shown above, any action for hook field with at most four derivatives and invariant with respect to both gauge symmetries in (A)dS enjoys Weyl symmetry, and therefore coincides with the action \eqref{S4 flat} in the flat limit. It follows then, that if there are distinct actions with those symmetries, their difference is encoded in two-derivative terms proportional to (A)dS curvature. Therefore there is a linear combination of them that is a two-derivative action invariant under both symmetries.
Since there is no such two-derivative action, it follows that
the four-derivative action \eqref{S4} is unique.
In order to show that the formula \eqref{WeylSquaredHook}
coincides with \eqref{S4}, one needs to perform several integrations by part ignoring the boundary terms.
In the case of Weyl gravity, this amounts to
\be
W^{\m\n,\r\l}\, W_{\m\n,\r\l} = \cL_{\rm\sst GB}+\frac{4(d-3)}{d-2}\left(
R^{\m\n}\,R_{\m\n}-\frac{d}{4(d-1)}\,R^2\right)\,,
\ee
where the Gauss-Bonet (GB) term given by
\be
\cL_{\rm\sst GB}=\frac14\,\d_{\n_1}^{\m_1}{}_{\n_2}^{\m_2}{}_{\n_3}^{\m_3}{}_{\n_4}^{\m_4}\,
R_{\m_1\m_2,}{}^{\n_1\n_2}\,R^{\n_3\n_4,}{}_{\m_3\m_4}\,.
\ee
vanishes identically in $d\le 3$ and is a topological invariant in $d=4$\,.
In higher dimensions, its linearization always reduces to a boundary term.
Here again, the `Weyl squared' action for the simple hook field
can be related to the form \eqref{S4}
by adding a GB-like term for the simple hook field.
In flat space, such term is given by
\be
\mathcal{L}^{\rm\sst flat}_{\rm\sst GB}(\phi)
=\frac1{2!\,3!}\,\d_{\n_1}^{\m_1}{}_{\n_2}^{\m_2}{}_{\n_3}^{\m_3}{}_{\n_4}^{\m_4}{}_{\n_5}^{\m_5}\,
(\cR\,\phi)_{\m_1\m_2\m_3,}{}^{\n_1\n_2}\,(\cR\,\phi)^{\n_3\n_4\n_5,}{}_{\m_4\m_5}\,.
\label{GBHook}
\ee
This term now vanishes identically in $d\le 4$ and
becomes a boundary term in $d\ge 5$\,.
There is a unique (A)dS generalization of \eqref{GBHook} that is invariant with respect to both gauge symmetries and
reduces to a boundary term in $d\ge5$. It is given by
\be
\mathcal{L}^{\sst \L}_{\rm\sst GB}(\phi)
=\frac1{2!\,3!}\,\d_{\n_1}^{\m_1}{}_{\n_2}^{\m_2}{}_{\n_3}^{\m_3}{}_{\n_4}^{\m_4}{}_{\n_5}^{\m_5}\,
(\cR^{\sst\L}_\syd\,\phi)_{\m_1\m_2\m_3,}{}^{\n_1\n_2}\,(\cR^{\sst\L}_\ayd\,\phi)^{\n_3\n_4\n_5,}{}_{\m_4\m_5}\,.
\label{GBHook AdS}
\ee
The Weyl squared action \eqref{WeylSquaredHook}
differs from the action \eqref{S4}
by the boundary term \eqref{GBHook AdS}:
\be
\la \cW\,\phi\,|\,\cW\,\phi\ra=
\frac{1}{12}\int d^dx\sqrt{|g^{\sst \L}|}\,
\cL^{\sst\L}_{\rm\sst GB}(\phi)
+\frac{d-4}{d-3}\, \la \phi\,|\,\cG^{\sst\L}_{\syd}\,\cI^{-1}\,\cG^{\sst\L}_{\ayd}\,\phi\ra\,.\label{WeylGB}
\ee
Let us notice that both
the Weyl tensor and the GB-like term \eqref{GBHook} vanish identically
for $d\leq 4$\,, while the action \eqref{S4}
does not vanish in $d=4$. The equation \eqref{WeylGB} still holds, since the coefficient in front of the last term vanishes for $d=4$\,. In fact, in four dimensions, the action \eqref{S4} describes special spectra, which is different from the off-shell field. We will come back to this later.
\subsubsection{Factorization and Degrees of Freedom}
The Lagrangian \eqref{S4} can be written in the ordinary-derivative form
by introducing an auxiliary field $f_{\m\n,\r}$ as
\be
\cS^{\sst\L}_{\rm\sst W}[\phi]\simeq m_\L^2
\left(\la\phi\,|\, \cG^{\sst\L}_{\syd}\,\phi\ra
+2\,\la \phi\,|\,\cG^{\sst\L}_\syd\,f\ra
+m^2_\L\,\la f\,|\,\cI\,f\ra\right).\label{AuxWeyl}
\ee
We can solve the equation for $f$ and go back to the Lagrangian \eqref{S4}.
The Lagrangian \eqref{AuxWeyl} can be diagonalized into
\be
\cS^{\sst\L}_{\sst\rm W}[\phi]\simeq m_\L^2\left(\cS^{\sst\L}_{\sst\rm E\,\syd}[h]
-\cS^{\sst\L}_{\sst\rm E\,\ayd}[f]\right).
\label{ODActionWeylAdS}
\ee
using the field redefinition,
\be\label{FieldRedef}
\phi_{\m\n,\r}=h_{\m\n,\r}-f_{\m\n,\r}\,.
\ee
This rewriting procedure is singular in the flat limit (note that the action
\eqref{AuxWeyl} is not well defined in the flat space limit $m_\L^2\to 0$), while for the action \eqref{S4} the flat space limit is well defined and the number of DoFs of the theory is the same in (A)dS and flat spaces.
The action \eqref{ODActionWeylAdS} contains both short representations
of the hook in (A)dS$_d$.
In the flat-space limit, it propagates two relatively ghost hook-helicity $\tiny\yng(2,1)$ modes (which are propagating only for $d\ge5$), supplemented with a spin-two helicity mode $\tiny\yng(2)$ and a two-form helicity mode $\tiny\yng(1,1)$\,.
Even in (A)dS background, the reasoning in terms of
the $SO(d-2)$ helicity modes can serve as a useful guideline,
despite the fact that they are not irreducible anymore when $\L\neq 0$.
Let us now analyze the DoFs of the theory $\cS^{\sst \L}_{\rm\sst W}$ in terms
of the helicity modes.
The result of the analysis can be schematically formulated as
\be
m^2_{\sst \L} \left(\,\yng(2,1)_h + \,\L\,\yng(1,1)_h \right)
-m^2_{\sst \L} \left(\,\yng(2,1)_f -\L\,\yng(2)_f\right).
\label{DoF hf}
\ee
The $h$-field comes with a supplementary propagating two-form mode $\tiny\yng(1,1)_h$\,,
which is unitary in dS space and
non-unitary in AdS \cite{Zinoviev:2002ye}. Since the action for that mode comes with the factor $m_\L^2$ in front, which is positive in AdS and negative in dS, corresponding two-form $\tiny\yng(1,1)_h$ is non-unitary in both cases, while the leading hook $\tiny\yng(2,1)_h$ itself is unitary in AdS and non-unitary in dS. In the same way, the spin-two mode $\tiny\yng(2)_f$ of the $f$-field has the same sign of kinetic term as the leading hook $\tiny\yng(2,1)_f$ in AdS and opposite sign in dS, therefore is non-unitary in both cases, taking into account the factor $-m_\L^2$ in front of the action for $f$. We conclude, that among the four propagating helicity modes, only
one of the hook modes has positive sign of kinetic term, therefore the theory is non-unitary for any $d\geq 5$, when hook modes propagate.
In four dimensions, the helicity group is $SO(2)$, which does not have any hook representation, therefore the hook helicity modes do not propagate,
hence we are left with a pseudo scalar $\tiny\yng(1,1)_h\sim \bullet_h$ and a spin-two
$\tiny\yng(2)_f$ modes both non-unitary.
We can render these two modes into unitary one by simply introducing a negative sign
factor in the original action.
In this way, we can describe a unitary system of spin-two and scalar in four dimensions
using the four-derivative Weyl action $\cS^{\sst \L}_{\sst\rm W}$\,.
This theory is unitary in four dimensions and has the same spectrum
for any value of $\L$ (including zero), and coincides with the massless limit of $4d$ New Massive Gravity \cite{Bergshoeff:2012ud} in flat space.
This action can be equally obtained by a particular dimensional reduction
of massless spin-two field from $5d$ down to $4d$ \cite{Joung:2012sa}.
\subsubsection{Massive Deformation}
Let us consider mass deformations of the Weyl action
with Einstein terms.
Again we have two options:
the first one is to introduce $\cS^{\sst \L}_{\rm\sst E\,\syd}$ term
so that the total action preserves the $\syd$ gauge symmetry:
\be
\cS^{\sst\rm massive}_{\syd}[\phi,m^2]
= \cS^{\sst\L}_{\rm\sst W}[\phi]+m^2\,\cS^{\sst\L}_{\sst\rm E\,\syd}[\phi]\,,
\ee
and the other possibility is keeping the $\ayd$ symmetry by considering,
\be
\cS^{\sst\rm massive}_{\ayd}[\phi,m^2]
= \cS^{\sst\L}_{\rm\sst W}[\phi]+m^2\,\cS^{\sst\L}_{\sst\rm E\,\ayd}[\phi]\,.
\ee
In each cases, we can introduce an auxiliary field $f$ and do a proper redefinition
of the type `$\phi=h-f$' as in the Weyl action case
so that the four-derivative actions turn into two-derivative ones,
\ba
&& \cS^{\rm\sst massive}_{\syd}[\phi,m^2]
\simeq
\left(m^2_\L+m^2\right)\left(\la h\,|\,\cG^{\sst \L}_{\syd}\,h\ra
-\la f\,|\left(\cG^{\sst \L}_{\ayd}-m^2\,\cI\right) f\ra\right),\\
&& \cS^{\rm\sst massive}_{\ayd}[\phi,m^2]\simeq
\left(m^2_\L-m^2\right)\left(\la h\,|\left(\cG^{\sst \L}_{\syd}-m^2\,\cI\right) h\ra
-\la f\,|\,\cG^{\sst \L}_{\ayd}\,f\ra\right).
\ea
One can see that depending on which symmetry we decide to preserve,
the other field --- whose gauge symmetry is spoiled by the Einstein term --- acquires a mass.
Let us also note that
the action $\cS^{\rm\sst massive}_{\syd/\ayd}$ becomes singular
when $\mp m^2$ approaches $m_\L^2$\,.
This corresponds to the coincidence limit of two spectra described by $h$ and $f$\,.
Now let us focus on the four dimensional case. In the previous section,
we have already remarked that the Weyl action $\cS^{\L}_{\sst\rm W}$
describes a massless spin-two and a pseudo scalar in four dimensions.
With the deformation of the Einstein term, the action
$\cS^{\rm massive}_{\syd}$ describes
\be
\cS^{\rm massive}_{\syd}\sim
\left(m^2_{\L}+m^2\right)
\left( \L\,\bullet_h -\, \yng(2)^{\sst SO(3)}_f\right),
\ee
whereas the $\cS^{\rm massive}_{\ayd}$ gives
\be
\cS^{\rm massive}_{\ayd}\sim
\left(m^2_{\L}-m^2\right)
\left(\,\yng(2)^{\sst SO(3)}_h +\L\, \yng(2)^{\sst SO(2)}_f\right),
\ee
where we have dualized all the modes except for the last one $\tiny\yng(2)^{\sst SO(2)}_f$
and used the fact that the hook helicity mode identically vanishes in four dimensions.
Let us remark that
in $ \cS^{\rm massive}_{\syd}$ the two modes ---
scalar $\bullet_h$ and massive spin two $\tiny\yng(2)^{\sst SO(3)}_f$ ---
have the same sign of the kinetic term only for $\L<0$\,, namely in AdS space.
On the contrary,
the two modes of $ \cS^{\rm massive}_{\ayd}$ --- massive and massless spin two fields $\tiny\yng(2)^{\sst SO(3)}_h$ and
$\tiny\yng(2)^{\sst SO(2)}_f$ ---
have the same sign in dS space ($\L>0$).
Hence, again by introducing an overall minus sign in these actions,
we obtain two models of four-derivative theories with unitary propagation.
It will be interesting to explore possible links of these four-derivative formulations
with existing models.
For instance, the field content of $\cS^{\rm massive}_{\ayd}$
coincides with that of bigravity \cite{Hassan:2011zd} proposed by Hasan and Rosen.
\section{Generalizations}
\label{sec:3}
All the discussions about the simple hook field
can be straightforwardly generalized to more general
fields having the symmetry of two-column Young diagrams.
The number of columns is playing the role of `spin' and
the two-column fields show many common features
with spin-two field.
In particular, their Weyl actions contain four derivatives.
This considerably reduces
the technical complexities
and allows us to perform the analysis explicitly.
The only qualitative difference of two-columns fields
with respect to the hook one rises in
the case where the two columns have the same height,
which we refer to as `long window' diagram.
Hence, we shall first consider
the non-window case in below
and then do a separate analysis for the long window case in the succeeding section.
\subsection{Two Columns}
\label{sec:3.1}
In this section, we will generalize the results of Section \ref{sec:2} for a tensor field
$\phi_{\m_1\cdots\m_p,}^{\qquad\quad\n_1\cdots\n_q}$
having the symmetry of two column Young diagram:
\be
\phi_{\m_1\cdots\m_p,}^{\qquad\quad\n_1\cdots\n_q} \sim\,
\parbox{25pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2.4);
\draw (0.4,2.4) -- (0.8,2.4) -- (0.8,0.8) -- (0.4,0.8);
\node at (0.2,1.2){${\st p}$};
\node at (0.6,1.6){${\st q}$};
\end{tikzpicture}}\,,
\label{pq field}
\ee
where we assume $p$ is strictly larger than $q$\,.
\subsubsection{Einstein Action}
The action for this field is \ed{schematically} given making use of the diagram,
\be
S_{\rm\sst E}[\phi]
\sim \parbox{45pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2.4);
\draw (0.4,2.4) -- (0.8,2.4) -- (0.8,0.8) -- (0.4,0.8);
\draw (0.8,0.8) -- (0.8,-2) -- (0,-2) -- (0,0);
\draw (0.4,0) -- (0.4,-2);
\draw (0,-0.4) -- (0.4,-0.4);
\draw (0.4,0.4) -- (0.8,0.4);
\node at (0.2,1.2){${\st p}$};
\node at (0.6,1.6){${\st q}$};
\node at (0.2,-1.2){${\st q}$};
\node at (0.6,-0.8){${\st p}$};
\node at (0.2,-0.2){${\st \partial}$};
\node at (0.6,0.6){${\st \partial}$};\,,
\end{tikzpicture}}\,,
\label{pq Lagrangian}
\ee
which is made of two fields and two derivatives
contracted with
the generalized Kronecker delta $\d_{\m_1\cdots\m_{p+q+1}}^{\n_1\cdots\n_{p+q+1}}$\,. \ed{This form of the action for generic two column fields was introduced in \cite{Bekaert:2004dz}, generalizing the analogous formula for simple hook \cite{Curtright:1980yk} and the window diagram \cite{Boulanger:2004rx}. It is equivalent to Labastida action \cite{Labastida:1986ft} up to total derivatives.}
The corresponding Einstein tensor is hence given by
\be
(\cG\,\phi)_{\m_1\cdots\m_p,}{}^{\n_1\cdots\n_q}
=-\frac{(-1)^{(p+1)(q+1)}}{p!\,q!}\,
\d_{\m_1\cdots\m_{p+q+1}}^{\n_1\cdots\n_{p+q+1}}\,
\partial^{\m_{p+1}}\,\partial_{\n_{q+1}}\,
\phi_{\n_{q+2}\cdots\n_{p+q+1},}{}^{\m_{p+2}\cdots\m_{p+q+1}}\,,
\label{pq einstein}
\ee
which is invariant with respect to two gauge symmetries:
\ba
&\d_\e\,\phi_{\m_1\cdots\m_p\,,}{}^{\n_1\cdots\n_q}
=\partial_{[\m_1}\,\e_{\m_2\cdots\m_p]\,,}{}^{\n_1\cdots\n_q}\,,\\
&\d_\th\,\phi_{\m_1\cdots\m_p\,,}{}^{\n_1\cdots\n_q}
=\partial^{[\n_1}\,\th_{\m_1\cdots\m_p\,,}{}^{\n_2\cdots\n_q]}
+(-1)^{p+1}\,\frac{p}{p-q+1}\,
\partial_{[\m_1}\,\th_{\m_2\cdots\m_p]}{}^{[\n_1\,,\,\n_2\cdots\n_q]},
\label{pq gauge}
\ea
with gauge parameters of the following Young symmetry type:
\be
\e_{\m_1\cdots\m_{p-1}\,,}{}^{\n_1\cdots\n_{q}}\sim\,
\parbox{45pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2);
\draw (0.4,2) -- (0.8,2) -- (0.8,0.4) -- (0.4,0.4);
\draw [dotted] (0,0) rectangle (0.4,-0.4);
\node at (-0.3,1){${\st p-1}$};
\node at (0.6,1.2){${\st q}$};
\end{tikzpicture}}\,,\qquad
\th_{\m_1\cdots\m_{p}\,,}{}^{\n_1\cdots\n_{q-1}}\sim\,
\parbox{40pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2.4);
\draw (0.4,2.4) -- (0.8,2.4) -- (0.8,1.2) -- (0.4,1.2);
\draw [dotted] (0.4,1.2) rectangle (0.8,0.8);
\node at (0.2,1.2){${\st p}$};
\node at (1.1,1.8){${\st q-1}$};
\end{tikzpicture}}\,.
\ee
The Lagrangian \eqref{pq Lagrangian}, or
equivalently the Einstein tensor
\eqref{pq einstein}
can be also given through the gauge invariant curvature tensor $(\cR\,\phi)_{\m_1\cdots\m_{p+1}\,,}{}^{\n_1\cdots\n_{q+1}}$, which again
has a symmetry of two-column Young diagram
but with additional boxes in each column (corresponding to derivatives):
\be
(\cR\,\phi)_{\m_1\cdots\m_{p+1}\,,}{}^{\n_1\cdots\n_{q+1}}\sim\,
\parbox{25pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2.4);
\draw (0.4,2.4) -- (0.8,2.4) -- (0.8,0.8) -- (0.4,0.8);
\draw (0,0) -- (0,-0.4) -- (0.4,-0.4) -- (0.4,0);
\draw (0.4,0.4) -- (0.8,0.4) -- (0.8,0.8);
\node at (0.2,1.2){${\st p}$};
\node at (0.6,1.6){${\st q}$};
\node at (0.2,-0.2){${\st \partial}$};
\node at (0.6,0.6){${\st \partial}$};
\end{tikzpicture}}\,.
\label{curvature}
\ee
Its explicit expression reads
\be
(\cR\,\phi)_{\m_1\cdots\m_{p+1},}{}^{\n_1\cdots\n_{q+1}}
=(p+1)(q+1)\,\partial_{[\m_1}\,\partial^{[\n_1}\,\phi_{\m_2\cdots\m_{p+1}],}
{}^{\n_2\cdots\n_{q+1}]}\,,
\label{pq curvature}
\ee
and it is invariant with respect to both gauge symmetries
\eqref{pq gauge}.
Around (A)dS background, the curvature cannot be gauge invariant with respect to both of the symmetries, and one has to choose a zero-derivative deformation to preserve one of the symmetries:
\ba
&\d^{\sst\L}_\e\,\phi_{\m_1\cdots\m_p\,,}{}^{\n_1\cdots\n_q}
=\nabla_{[\m_1}\,\e_{\m_2\cdots\m_p]\,,}{}^{\n_1\cdots\n_q}\,,\\
&\d^{\sst \L}_\th\,\phi_{\m_1\cdots\m_p\,,}{}^{\n_1\cdots\n_q}
=\nabla^{[\n_1}\,\th_{\m_1\cdots\m_p\,,}{}^{\n_2\cdots\n_q]}
+(-1)^{p+1}\,\frac{p}{p-q+1}\,
\nabla_{[\m_1}\,\th_{\m_2\cdots\m_p]}{}^{[\n_1\,,\,\n_2\cdots\n_q]},
\label{pq gauge AdS}
\ea
The curvature invariant
under the $\delta^{\sst \L}_{\e}$ symmetry
takes the form,
\be
(\cR^{\sst \L}_{\sst [p-1,q]}\,\phi)_{\m_1\cdots\m_{p+1},}{}^{\n_1\cdots\n_{q+1}}
=(p+1)(q+1)\,\nabla^{[\n_1}\,\nabla_{[\m_1}\,
\phi_{\m_2\cdots\m_{p+1}],}{}^{\n_2\cdots\n_{q+1}]}\,,\label{CurvTheta}
\ee
which is again a simple extension
of the flat space curvature \eqref{pq curvature}.
Considering now the
curvature invariant under the $\delta^{\sst \L}_{\th}$
transformation, we find
\ba
&& (\cR^{\sst\L}_{\sst [p,q-1]}\,\phi)_{\m_1\cdots\m_{p+1},}{}^{\n_1\cdots\n_{q+1}}\nn
&&=\,
(\cR^{\sst \L}_{\sst [p-1,q]}\,\phi)_{\m_1\cdots\m_{p+1},}{}^{\n_1\cdots\n_{q+1}}
-\frac{2(p-q+1)\,\L}{(d-1)(d-2)}\,\d_{[\m_1}^{[\n_1}\,\phi_{\m_2\cdots\m_{p+1}],}{}^{\n_2\cdots\n_{q+1}]}\nonumber\\
&&=\,(p+1)(q+1)\,\nabla_{[\m_1}\,\nabla^{[\n_1}\,\phi_{\m_2\cdots\m_{p+1}],}{}^{\n_2\cdots\n_{q+1}]}
-\frac{2\,\L}{(d-1)(d-2)}\,\d_{[\m_1}^{[\n_1}\,\phi_{\m_2\cdots\m_{p+1}],}{}^{\n_2\cdots\n_{q+1}]}\,.\qquad\quad\label{CurvSigma}
\ea
We can see that these two curvatures differ
by a zero-derivative term.
One can then construct Einstein tensors out of these
curvatures as
\ba
(\cG^{\sst\L}_{\sst [p-1,q]}\,\phi)_{\m_1\cdots\m_{p},}{}^{\n_1\cdots\n_{q}}
\eq -\frac{(-1)^{(p+1)(q+1)}}{(p+1)!(q+1)!}\,\delta_{\m_1\cdots \m_{p+q+1}}^{\n_1\cdots \n_{p+q+1}}\,
(\cR^{\sst\L}_{\sst [p-1,q]}\,\phi)_{\n_{q+1}\cdots\n_{p+q+1},}{}^{\m_{p+1}\cdots\m_{p+q+1}}\,,\qquad\,\,\\
(\cG^{\sst\L}_{\sst [p,q-1]}\,\phi)_{\m_1\cdots\m_{p},}{}^{\n_1\cdots\n_{q}}
\eq -\frac{(-1)^{(p+1)(q+1)}}{(p+1)!(q+1)!}\,\delta_{\m_1\cdots \m_{p+q+1}}^{\n_1\cdots \n_{p+q+1}}\,
(\cR^{\sst\L}_{\sst [p,q-1]}\,\phi)_{\n_{q+1}\cdots\n_{p+q+1},}{}^{\m_{p+1}\cdots\m_{p+q+1}}\,.\qquad\,\,
\ea
The Einstein tensors
immediately define the Einstein actions
\ba
\cS^{\sst\L}_{\sst {\rm E}\,[p-1,q]}=\la \phi\,|\,\cG^{\sst\L}_{\sst [p-1,q]}\,\phi\ra\,,\quad
\cS^{\sst\L}_{\sst {\rm E}\,[p,q-1]}=\la \phi\,|\,\cG^{\sst\L}_{\sst [p,q-1]}\,\phi\ra\,
\label{El}
\ea
for the two-column field \eqref{pq field}.
The action $\cS^{\sst\L}_{\sst {\rm E}\,[p-1,q]}$
is (not) unitary around (A)dS
and the opposite for $\cS^{\sst\L}_{\sst {\rm E}\,[p,q-1]}$\,.
When \mt{p+q=d-1}, the $[p,q]$ mode of
the Einstein actions $\cS^{\sst\L}_{\sst {\rm E}\,[p-1,q]}$
or $\cS^{\sst\L}_{\sst {\rm E}\,[p,q-1]}$
vanishes identically leaving only the $[p,q-1]$ or $[p-1,q]$ mode, respectively.
The kinetic terms of these modes come with the factor $\L$ or $-\L$\,.
With an appropriate choice of an overall factor of the Einstein action,
these modes may become unitary.
\subsubsection*{Long hook $[d-2,1]$}
An interesting example of the case $p+q=d-1$ is the \emph{long hook} where $p=d-2$ and $q=1$\,.
The action $\rm{sign}(\L)\,\cS^{\sst\L}_{\sst {\rm E}\,[d-3,1]}$ has a propagating degree of freedom of a scalar,
whereas the other action $-\,\rm{sign}(\L)\,\cS^{\sst\L}_{\sst {\rm E}\,[d-2,0]}$ has a propagating
mode dual to massless graviton.
The long hook field $[d-2,1]$ is used in New Massive Gravity in arbitrary dimensions \cite{Joung:2012sa,Dalmazi:2012dq},
and also in the recent work \cite{Basile:2015jjd}.
Two different ways the long hook field appears in \cite{Basile:2015jjd} and \cite{Joung:2012sa}
have analogous features which deserve a few remarks.
On the one hand in \cite{Basile:2015jjd}, the authors obtained a two-derivative action
from the linearized Einstein-Cartan gravity in AdS by integrating out the vielbein instead
of the spin connection. This determines the linearized vielbein in terms of spin connection as
$h_\mu^a=h_\m^a(\o)$\,, which has the form of Schouten tensor written in terms of spin connection.
On the other hand in \cite{Joung:2012sa},
the simple hook field $\phi_{\m\n,\r}$ appears as a result of solving the constraint,
\be
\partial^\m\,\partial^\n\,h_{\m\n}-\Box\,h^\m{}_\m=0\,,
\label{constraint h}
\ee
arising in the course of a special dimensional reduction of the massless spin two system.
The solution to \eqref{constraint h} is given \cite{DuHe} in terms of a hook field $\varphi_{\m\n,\r}$\,:
$h_{\m\n}=h_{\m\n}(\varphi)$\,.
Interestingly, the form $h_\m^a(\o)$ can be brought to the form of $h_{\m\n}(\varphi)$
using gauge transformations and with $\partial_\mu$ replaced by the AdS covariant derivative.
Moreover, it has been shown \cite{Joung:2012sa} that the action resulting after solving the constraint \eqref{constraint h} has additional symmetry $\d_\s\,\phi_{\m\n,\r}=\eta_{\r[\mu}\,\pr_{\nu]}\s$
under which $h_{\m\n}(\varphi)$ transforms like the Weyl transformation of Schouten tensor,
$\d_\s\,h_{\m\n}(\varphi)=\na_\m\pr_\nu\s$.
This interplay between two constructions in \cite{Basile:2015jjd} and \cite{Joung:2012sa} can be understood from
the fact that the constraint \eqref{constraint h} is the identity
that Schouten tensor satisfies.
Let us notice also that the action of \cite{Basile:2015jjd}
after dualization actually coincides with $\cS^{\sst\L}_{\sst {\rm E}\,[d-2,0]}$\,\ed{, the second action of \eqref{El} for $p=d-2\,,\, q=1$}.
Another interesting feature of the construction \cite{Joung:2012sa} is that
the final action is given by massive Fierz-Pauli action in terms of $h_{\m\n}(\varphi)$\,.
After dualization, in terms of the dual long hook field, $\phi_{\mu_1\cdots\mu_{d-2},\nu}$,
the two-derivative part proportional to $h^{[\m}_{\m}(\varphi)\,h^{\n]}_{\n}(\varphi)$
coincides with the Einstein action $\cS_{\sst {\rm E}}[\phi]$,
while the four-derivative one, proportional to $\la h(\varphi)\,|\,\cG\,h(\varphi)\ra$,
coincides with the Weyl action $\cS_{\sst\rm W}[\phi]$.
\subsubsection{Weyl Action}
Let us now move to the Weyl action.
For that, we consider
the Weyl tensor for two column field \eqref{pq field}
by making use of the trace projector $\cT$ as
\be
\cW=\mathcal T\,\cR_{\sst [p-1,q]}^{\sst\L}=\mathcal T \,\cR_{\sst [p,q-1]}^{\sst\L}\,.
\label{WeylTensorForpq}
\ee
The form of the trace projector
acting on a $[p+1,q+1]$ tensor can be conveniently given by the following expression:
\be
\mathcal T = \frac{(-1)^{(p+1)(q+1)}}{(p+1)!(q+1)!}\left(\d_{p+q+2}-\frac{d-p-q-1}{d-p-q}\d_{p+q+1}(\d_{p+q})^{-1}\d_{p+q+1}\right),
\label{TraceProjector pq}
\ee
where the operator $\d_n$ acts on the field $f_{\m_1\cdots\m_{p}\,,}{}^{\n_1\cdots\n_{q}}$ of type $[p,q]$ in the following way
\be
(\d_n f)_{\r_{1}\cdots\r_{n-q}\,,}{}^{\l_{1}\cdots\l_{n-p}}
=\d^{\l_1\cdots\l_{n-p}\m_1\cdots\mu_p}_{\r_1\cdots\r_{n-q}\n_1\cdots\n_{q}}\,
f_{\m_{1}\cdots\m_{p}\,,}{}^{\n_{1}\cdots\n_{q}}\,,
\ee
contracting all of its indices to the generalized Kronecker-delta. Inverse operator $(\d_{p+q})^{-1}$ is defined as:
\be
(\d_{p+q}\,(\d_{p+q})^{-1}\,f)_{\m_1\cdots\m_{p}\,,}{}^{\n_1\cdots\n_{q}}=
f_{\m_1\cdots\m_{p}\,,}{}^{\n_1\cdots\n_{q}}=
((\d_{p+q})^{-1}\,\d_{p+q}\,f)_{\m_1\cdots\m_{p}\,,}{}^{\n_1\cdots\n_{q}}\,.
\ee
The tracelessness of the expression
\eqref{TraceProjector pq} can be proven using the identity,
\be
\d_{\n_1}^{\m_1}\,\d_{\m_1\dots\m_{n}}^{\n_1\dots\n_{n}}
=(d-n+1)\,\d_{\m_2\dots\m_{n}}^{\n_2\dots\n_{n}}\,.
\ee
It is again clear that Weyl tensor \eqref{WeylTensorForpq} is
invariant under both gauge symmetries \eqref{pq gauge AdS}
hence also under Weyl transformation:
\be
\delta_\a\,\phi_{\mu_1\cdots\mu_p,}{}^{\n_1\cdots\n_q}
=\delta_{[\mu_1}^{[\nu_1}\,
\a_{\mu_2\cdots \mu_p],}{}^{\nu_2\cdots \nu_q]}\,.
\label{Weyl sym}
\ee
Analogously to the hook case,
the Weyl action is given by the square
of Weyl tensors,
\be
\cS^{\sst \L}_{\rm\sst W}[\phi]= -\frac{d-p-q}{d-p-q-1}
\la \cW\,\phi\,|\,\cW\,\phi\ra\,,
\label{Sw W}
\ee
where the coefficient
is fixed such that
the Weyl action takes another representation,
\be
\cS^{\sst \L}_{\rm\sst W}[\phi]=-
\la \phi\,|\,\cG^{\sst\L}_{\sst [p-1,q]}\,\cI^{-1}\,
\cG^{\sst\L}_{\sst [p,q-1]}\,\phi\ra\,,
\label{Sw E}
\ee
up to a GB-like total derivative term,
\be
\cL^{\sst \L}_{\rm\sst GB}(\phi)
= (-1)^{(p+1)(q+1)}\,
\la\, \cR^{\sst\L}_{\sst [p-1,q]}\,\phi\,|\,
\delta_{p+q+2}\,\cR^{\sst\L}_{\sst [p,q-1]}\,\phi\,\ra\,.
\ee
In \eqref{Sw E} the mass operator $\cI$ is defined as
\be
\cI=\frac{(-1)^{p\,q}}{p!\,q!}\,\d_{p+q}\,.
\ee
Hence, one can see that the entire
construction of the hook example
can be generalized to the two column case
in a straightforward manner.
The DoF can be conveniently analyzed by using the
same factorization technique as the simple hook case.
Skipping the straightforward derivation part, let us
directly present the end result,
\be
m^2_{\sst \L} \left(\,
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2.4);
\draw (0.4,2.4) -- (0.8,2.4) -- (0.8,0.8) -- (0.4,0.8);
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-2)}_h
+ \,\L\
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2.4);
\draw (0.4,2.4) -- (0.8,2.4) -- (0.8,1.2) -- (0.4,1.2);
\draw [dotted] (0.4,1.2) rectangle (0.8,0.8);
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-2)}_h \right)
-m^2_{\sst \L} \left(\,
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2.4);
\draw (0.4,2.4) -- (0.8,2.4) -- (0.8,0.8) -- (0.4,0.8);
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-2)}_f -\L\
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2);
\draw (0.4,2) -- (0.8,2) -- (0.8,0.4) -- (0.4,0.4);
\draw [dotted] (0,0) rectangle (0.4,-0.4);
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-2)}_f\right),
\ee
which is the natural generalization of the result \eqref{DoF hf}.
All these DoF vanish in dimensions $d\le p+q$\,.
When $d=p+q+1$\,, the helicity mode $[p,q]$ vanishes,
but there remains propagating DoFs given by
\be
-\parbox{42pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,1.2);
\draw (0.4,0) -- (0.4,1.2);
\draw [dotted] (0,0) rectangle (0.4,-1.2);
\draw [dotted] (0.4,0) rectangle (0.8,-0.4);
\node at (-0.3,0.6){${\st q-1}$};
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-2)}_h
-\,
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,1.6);
\draw (0.4,0) -- (0.4,1.6);
\draw [dotted] (0,0) rectangle (0.4,-0.8);
\node at (0.2,0.8){${\st q}$};
\node at (0.6,0.8){${\st q}$};
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-2)}_f\,,
\ee
where we have dualized both modes.
Note that these modes have the same kinetic term sign,
hence can describe unitary propagation by introducing a negative factor
in the original action.
Notice that when $q=1$ we get in this way
the DoF of
a scalar and a helicity two mode.
The corresponding $[p,q]=[d-2,1]$ Weyl action coincides
in fact with the massless limit of New Massive Gravity action in any $d$
\cite{Joung:2012sa}, discussed in previous subsection.
\subsection{Long Window}
\label{sec:3.2}
In the special case where the height of two columns are equal,
that is $q=p$\,:
\be
\phi_{\m_1\cdots\m_p\,,}^{\qquad\quad\n_1\cdots\n_p} \sim\,
\parbox{25pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,2);
\draw (0.4,0) -- (0.4,2);
\node at (0.2,1){${\st p}$};
\node at (0.6,1){${\st p}$};
\end{tikzpicture}}\,,
\label{pp field}
\ee
the analysis is no more analogous to the hook field case,
but actually more similar to the spin-two case.
Let us see explicitly how this works.
First of all, the field $\phi_{\mu_1\cdots \m_p,}{}^{\n_1\cdots \n_p}$ admits only the gauge symmetry generated by
the parameter of $[p,p-1]$ Young diagram,
because there is no Young diagram $[p-1,p]$.
Therefore, even in flat space,
the long window has only one gauge symmetry.
In (A)dS background,
we can first consider the curvature
$\cR^{\sst\L}_{\sst [p,p-1]}$ \eqref{CurvSigma}
or equivalently
the Einstein tensor $\cG^{\sst\L}_{\sst [p,p-1]}$
having the $[p,p-1]$ gauge symmetry.
In this case, there is nothing different from the generic two column case
and we can obtain the corresponding tensor and action.
Considering now the curvature
$\cR^{\sst\L}_{\sst [p-1,p]}$ \eqref{CurvTheta}
or the Einstein tensor $\cG^{\sst\L}_{\sst [p-1,p]}$,
we first note that they cannot have the $[p-1,p]$ gauge symmetry as it simply does not exist.
Instead, one may wonder whether this action still plays a special role.
It tuns out that with the sacrifice of the $[p-1,p]$ gauge symmetry,
the action $\cS^{\sst \L}_{\sst {\rm E}\, [p-1,p]}$ acquires a new gauge symmetry,
\be
\d^{\sst \L}_\e\,\phi_{\m_1\cdots\m_p\,,}{}^{\n_1\cdots\n_p}
=\left(\nabla_{[\m_1}\,\nabla^{[\n_1}-
\frac{2\,\L}{(d-1)(d-2)}\,\d_{[\m_1}^{[\n_1}\right)
\e_{\m_1\cdots\m_{p-1}\,,}{}^{\n_1\cdots\n_{p-1}}\,,
\label{pp pm sym}
\ee
with the gauge parameter,
\be
\e_{\m_1\cdots\m_{p-1}\,,}^{\qquad\quad\n_1\cdots\n_{p-1}} \sim\,
\parbox{65pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,1.6);
\draw (0.4,0) -- (0.4,1.6);
\draw [dotted] (0,0) rectangle (0.4,-0.4);
\draw [dotted] (0.8,0) -- (0.8,-0.4) -- (0.4,-0.4);
\node at (-0.4,0.8){${\st p-1}$};
\node at (1.2,0.8){${\st p-1}$};
\end{tikzpicture}}\,.
\label{pp pm gauge}
\ee
This is clearly the analogue of the partially-massless gauge symmetry of symmetric second rank field.
Indeed, when $p=1$ the corresponding action
coincides with that of partially-massless spin two.
Now considering the $[p-1,p-1]$ gauge symmetry \eqref{pp pm sym},
we can construct a one-derivative
gauge invariant, or curvature, as
\be
(\cC\,\phi)_{\m_1\cdots\m_{p+1}\,,}{}^{\n_1\cdots\n_p}
=(p+1)\,\nabla_{[\m_1}\,\phi_{\m_2\cdots\m_{p+1}],}{}^{\n_1\cdots\n_p},
\ee
and the action $\cS^{\sst \L}_{\sst {\rm E}\, [p-1,p]}$
can be also expressed as the square of this curvature and its traces.
We can proceed to construct a four-derivative action $\cS^{\sst\L}_{\sst\rm W}$
having
both of $[p,p-1]$ and $[p-1,p-1]$ gauge symmetries:
\ba
\d^{\sst \L}\,\phi_{\m_1\cdots\m_p\,,}{}^{\n_1\cdots\n_p}
\eq \left(\nabla_{[\m_1}\,\nabla^{[\n_1}-\frac{2\,\L}{(d-1)(d-2)}\,\d_{[\m_1}^{[\n_1}\right)
\e_{\m_1\cdots\m_{p-1}\,,}{}^{\n_1\cdots\n_{p-1}}+\nn
&&+\,\nabla^{[\n_1}\,\th_{\m_1\cdots\m_p\,,}{}^{\n_2\cdots\n_p]}
+(-1)^{p+1}\,p\,
\nabla_{[\m_1}\,\th_{\m_2\cdots\m_p]}{}^{[\n_1\,,\,\n_2\cdots\n_p]}.
\label{pp gauge sym}
\ea
Such an action will be automatically invariant under the Weyl transformation \eqref{Weyl sym}
because it can be realized as a particular configuration of \eqref{pp gauge sym}.
The action $\cS^{\sst\L}_{\sst\rm W}$ can be constructed
exactly in the same way as in the generic $[p,q]$ case,
either using Weyl tensor as in \eqref{Sw W}
or using Einstein tensors as in \eqref{Sw E}.
The DoF of the Weyl action can be analysed in terms of $SO(d-2)$ representations in a similar manner as before.
The result reads
\be
m^2_{\sst \L} \left(\,
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,2);
\draw (0.4,0) -- (0.4,2);
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-2)}_h
+ \,\L\
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.4,2);
\draw (0.4,0.4) -- (0.8,0.4) -- (0.8,2) -- (0.4,2);
\draw [dotted] (0.4,0) rectangle (0.8,0.4);
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-2)}_h \right)
-m^2_{\sst \L} \,
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,2);
\draw (0.4,0) -- (0.4,2);
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-2)}_f \,,
\ee
Notice that the $h$- and $f$-fields
are the long-window analogs of partially-massless spin two
and massless spin two. When $2p=d-1$\,,
we end up with only one mode,
\be
\parbox{42pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,1.6);
\draw (0.4,0) -- (0.4,1.6);
\draw [dotted] (0,0) rectangle (0.8,-0.4);
\draw [dotted] (0.4,0) rectangle (0.4,-0.4);
\node at (-0.3,0.8){${\st p-1}$};
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-2)}_h
\ee
which we have dualized.
Notice that when $p=1$, this action describes
a scalar mode in three-dimension.
This is nothing but the propagating content of
the 3d parity-invariant linear Weyl gravity
--- or massless limit of New Massive Gravity \cite{Bergshoeff:2009hq} ---
and the scalar mode corresponds
to the parity-invariant partially-massless spin two.
It would be also interesting to remark
that in $5d$\,, we get in this way a helicity two mode from the action of the $[2,2]$ window field,
which might provide an alternative theory of Gravity.
\subsection{Massive Deformation and New Massive Gravity}
\label{sec:3.3}
The massive deformation of the Weyl action for the generic two-column
fields follows the same pattern as the simple hook case:
\ba
\cS^{\sst\rm massive}_{\sst [p,q-1]}[\phi,m^2]
&=& \cS^{\sst\L}_{\rm\sst W}[\phi]+m^2\,\cS^{\sst\L}_{\sst{\rm E}\,[p,q-1]}[\phi] \nn
&\simeq&
\left(m^2_\L+m^2\right)\left(\la h\,|\,\cG^{\sst \L}_{\sst [p-1,q]}\,h\ra
-\la f\,|\left(\cG^{\sst \L}_{\sst [p,q-1]}-m^2\,\cI\right) f\ra\right),\\
\cS^{\sst\rm massive}_{\sst [p-1,q]}[\phi,m^2]
&=& \cS^{\sst\L}_{\rm\sst W}[\phi]+m^2\,\cS^{\sst\L}_{\sst{\rm E}\,[p-1,q]}[\phi]\nn
&\simeq&
\left(m^2_\L-m^2\right)\left(\la h\,|\left(\cG^{\sst \L}_{\sst [p-1,q]}-m^2\,\cI\right) h\ra
-\la f\,|\,\cG^{\sst \L}_{\sst [p,q-1]}\,f\ra\right),
\ea
which describe one massive $[p,q]$ mode
and one massless $[p,q]\oplus [p,q-1]$ or $[p,q]\oplus[p-1,q]$ mode, where the latter massless mode
becomes partially-massless for $q=p$ case.
Let us consider the dimensions $d=p+q+1$\,,
where the leading $[p,q]$-helicity modes disappear leaving only
lower helicity modes.
In this case, there is a preferred choice of massive deformation
between $\cS^{\sst\rm massive}_{\sst [p,q-1]}$
and $\cS^{\sst\rm massive}_{\sst [p-1,q]}$ depending on the sign
of the cosmological constant as in the simple hook case.
In dS background, only the massive action $\cS^{\sst\rm massive}_{\sst [p,q-1]}$
can describe massive and massless modes
with the same sign of kinetic terms because
\be
\cS^{\sst\rm massive}_{\sst [p,q-1]}\sim (m^2_\L-m^2)
\left(\,
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,1.6);
\draw (0.4,0) -- (0.4,1.6);
\draw [dotted] (0,0) rectangle (0.4,-0.8);
\node at (0.2,0.8){${\st q}$};
\node at (0.6,0.8){${\st q}$};
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-\bm 1)}_h
+\,\L\
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,1.6);
\draw (0.4,0) -- (0.4,1.6);
\draw [dotted] (0,0) rectangle (0.4,-0.8);
\node at (0.2,0.8){${\st q}$};
\node at (0.6,0.8){${\st q}$};
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-\bm 2)}_f
\right).
\ee
Instead in AdS background, the other action
$\cS^{\sst\rm massive}_{\sst [p-1,q]}$ is preferred since
\be
\cS^{\sst\rm massive}_{\sst [p-1,q]}\sim (m^2_\L+m^2)
\left(\,\L\,
\parbox{42pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,1.2);
\draw (0.4,0) -- (0.4,1.2);
\draw [dotted] (0,0) rectangle (0.4,-1.2);
\draw [dotted] (0.4,0) rectangle (0.8,-0.4);
\node at (-0.3,0.6){${\st q-1}$};
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-\bm 2)}_h
-\,
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,1.6);
\draw (0.4,0) -- (0.4,1.6);
\draw [dotted] (0,0) rectangle (0.4,-0.8);
\node at (0.2,0.8){${\st q}$};
\node at (0.6,0.8){${\st q}$};
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-\bm 1)}_f
\right).
\ee
Notice that in the above formulas two different Young diagrams
--- one carrying $SO(d-2)$ representation
and the other carrying $SO(d-1)$ representation --- are used.
Now taking the long-window case ($2p=d-1$), we have two analogue massive theories,
\be
\cS^{\sst\rm massive}_{\sst [p,p-1]}\sim
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,2);
\draw (0.4,0) -- (0.4,2);
\node at (0.2,1){${\st p}$};
\node at (0.6,1){${\st p}$};
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-\bm 1)}_h\,,
\qquad
\cS^{\sst\rm massive}_{\sst [p-1,p]}
\sim
\left(\,\L\,
\parbox{42pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,1.6);
\draw (0.4,0) -- (0.4,1.6);
\draw [dotted] (0,0) rectangle (0.8,-0.4);
\draw [dotted] (0.4,0) rectangle (0.4,-0.4);
\node at (-0.3,0.8){${\st p-1}$};
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-\bm 2)}_h
-
\parbox{22pt}{\begin{tikzpicture}
\draw (0,0) rectangle (0.8,2);
\draw (0.4,0) -- (0.4,2);
\node at (0.2,1){${\st p}$};
\node at (0.6,1){${\st p}$};
\end{tikzpicture}}{\phantom{\Bigg|}}^{\sst SO(d-\bm 1)}_f \right).
\ee
Hence, the massive action $\cS^{\sst\rm massive}_{\sst [p-1,p]}$
having the `partially-massless' gauge symmetry
have one additional $[p-1,p-1]$ helicity mode
compared to the action $\cS^{\sst\rm massive}_{\sst [p,p-1]}$
corresponding to `massless' gauge symmetry.
The action $\cS^{\sst\rm massive}_{\sst [p,p-1]}$
has only single massive irreducible $[p,p]$ mode, hence
unitary in both of AdS$_{2p+1}$ and dS$_{2p+1}$ background.
On the contrary, the other action $\cS^{\sst\rm massive}_{\sst [p-1,p]}$
propagates two modes with the same kinetic term sign
only in AdS background.
Focusing on the $p=1$ case,
the action $\cS^{\sst\rm massive}_{\sst [1,0]}$
describes a massive spin two in three dimensions and actually
coincides with the linearization of the New Massive Gravity \cite{Bergshoeff:2009hq}.
We can actually see here, that in AdS$_3$ there is an alternative ``New Massive Gravity'', that makes use of the partially-massless symmetry instead of the diffeomorphism one, and contains additional scalar in the spectrum.
As already discussed in previous sections, the term ``New Massive Gravity'' in higher dimensions refers to the models with fields of type $[p,1]$ (see \cite{Joung:2012sa,Dalmazi:2012dq} for related discussion) in dimensions $d=p+2$.
Let us note, that the flat limit of (A)dS$_{p+2}$ New Massive Gravity in dimensions higher than three is not smooth. In flat limit we have only a massive spin two mode, while in AdS we have unitary model with (massive spin two + massless scalar), and in dS the unitary model contains (massive spin two + massless spin two).
Only in three dimensions, one can have New Massive Gravity with the same spectrum around flat and constantly curved backgrounds, at least at the linearized level.
\acknowledgments
We thank Dario Francia
for discussions which incited our interest in the problem explored here.
The work of EJ was supported in part by the National Research Foundation of Korea through the grant NRF-2014R1A6A3A04056670 and the Russian Science Foundation grant 14-42-00047 associated with Lebedev Institute.
The work of KM was supported by the BK21 Plus Program funded by the Ministry of Education (MOE, Korea) and National Research Foundation of Korea (NRF).
\bibliographystyle{JHEP}
|
\section{Introduction\label{sec:Introduction}}
We propose a computational method to position a surveillance system with multiple sensors, where each sensor has a limited viewing angle and range. Some of the sensors may have a non-zero failure rate. Our goal is to achieve the maximum surveillance of a complicated environment.
The increasing societal demands for security lead to a growing need for surveillance activities in various environments, such as public places like transportation infrastructures and parking lots, shopping malls and telecommunication relay towers. How to optimally position the sensors is an interesting, important and maybe challenging problem, which can directly impact the public
safety and security. As reported in \cite{Murray2007133}, the optimal position of sensors
may also provide better resources allocation and system performance.
In practice, sensors used in many applications often come with limited sensing ability, for example,
an object cannot be detected if its distance to the sensor is too far (larger than $r>0$) or it is located outside of the sensors viewing angle ( $[v,v+\theta]$, $0<\theta<2\pi$, where $v$ is a viewing direction and $\theta$ is a limited width of viewing angle). There may exist obstacles that block the view of the sensors.
In addition, with certain probability, sensors may fail.
There exist many related studies in optimal sensor positioning.
Using graph-based approaches, this problem can be posed as a combinatorial optimization problem which has been shown to be NP-hard \cite{KSG2008}. Thus, simple enumeration and search techniques, as well as general purpose algorithms may experience extraordinary difficulty
in some cases. The placements of sensors have been undertaken for the well-known \emph{art gallery} problem, a classical problem that aims at placing stationary observers to maximize the surveillance area \cite{Aggarwal:1984:AGT:911725} in a cluttered region. The art gallery problem and its variations have drawn considerable attention in recent decades, for example in sensor placements
\cite{959341,Erdem04optimalplacement,Book:Foresti,959342,SISC2}. Similar studies have
been conducted to other problems such as routes for patrol guards with limited visibility
or limited mobility. We refer readers to a book \cite{MR921437} and papers
\cite{SISC1,POLSA,She92,Urrutia00artgallery} for more references therein.
Problems with possible sensor failures have also been studied. For example,
the use of a Bayesian approach is discussed in \cite{Cameron01101990}, incorporating
environmental data, such as temperatures, from multiple sensors to gather spatial
statistics can be found in \cite{Book_Cressie}.
A Gaussian process model is widely used to predict the effect of placing sensors
at particular locations \cite{DCI2002,KSG2008,Worden2001885,ZS_spatial,Zimmerman}.
See also \cite{FGGH2015} for optimally placing unreliable sensors in a $1$-dimensional environment.
Despite of the existence of extensive literature on related topics, the problem that we are
interested in remains to be challenging. The main difficulties often come from following aspects: (1) The obstacles have arbitrary shapes. (2) Finding a feasible placement with largest coverage area is often costly because it is an infinite dimensional problem. (3) Finding the globally optimal position, if possible, is usually computationally intractable. In addition, we consider the case of each sensor having a certain failure rate.
In this paper, we tackle the problem using the level set formulation together with the intermittent diffusion, a stochastic differential equation (SDE) based global optimization strategy \cite{ID}. The level set framework gives us the freedom to consider various shapes of obstacles and a way to incorporate the limited coverage into the definition of the coverage level set function.
The intermittent diffusion is then used to find one of the optimal viewing directions and locations in all of the feasible positions which gives the largest surveillance area of the environment.
The layout of the paper is as follows. In Section 2, the level set formulation of limited sensing range and viewing angle is introduced. We propose methods to find optimal locations and viewing directions in this setting. To have a possibility of finding a global optimum, we use the intermittent diffusion.
In Section 3, we describe the details of having a nonzero failure rate for the sensors and optimal condition for surveilling an area. Numerical implementation details and different applications are given in Section 4. We conclude by giving a summary and discussion in Section 5.
\section{Optimal positioning of the sensors\label{sec:Coverage-optimization-GD}}
In this section, we introduce the level set formulation for the coverage function of sensors with limited sensing range and limited width of viewing angle. Then we propose a strategy to find the optimal viewing directions of multiple stationary sensors to achieve the largest surveillance of the environment. Finally, we introduce the intermittent diffusion to find the globally optimal placement.
\subsection{Level set formulation for the sensors coverage\label{sub:Level-set-formulation}}
In the level set framework, the environment is described by a level set function, say $\psi(x)$, with positive values outside of the obstacles, negative values inside, and the zero level curves representing boundaries of obstacles.
A related application on visibility problem uses this level set setting \cite{CT:CMS2005,TCOBS:JCP2004}, where light rays from a vantage point travel in straight lines and are obliterated
upon contact with the surface of an obstacle. A point is seen by the observer (or sensor) placed at a given vantage point, if the line segment between the point and the vantage point does not intersect
any of the obstructions. In existing work \cite{POLSA,Tsai_robotic_path,LT:2008,LTC:2006} using such a setting, the observer has either a circular coverage (viewed) region, or an infinite range.
In this paper, we further consider the case of limited width of viewing angle with a finite range of coverage region.
We denote the computational domain as $D$,
a compact subset of $\mathbb{R}^{2}$. Let $\Omega$ be the collection
of obstacles, i.e. a closed set containing finite number of connected
components comprising one or multiple given obstacles in $D$. The
level set function $\psi(x)$ represents the environment: inside the
obstacles $\Omega$ is negative, and outside is positive. In our algorithms,
we define $\psi(x)$ as the typical signed distance function from the
boundaries of $\Omega$. A sensor located at a point $x$ in $D\backslash\Omega$,
has the line-of-sights as straight line segments originated from $x$
and ended at $y$ with either $|x-y|=r$ and $y\in D\backslash\Omega$,
or $|x-y|\le r$ and $y$ is a point on the boundary of the obstacles.
The coverage region of the sensor is the union of all line-of-sights
emanating from the sensor and the angle to the horizontal axis is
in $[v,v+\theta].$ Figure \ref{fig:A-sensor} depicts the coverage
area of the sensor in the setup.
\begin{figure}
\begin{centering}
\includegraphics[width=4cm]{figs/sensor}
\par\end{centering}
\caption{\label{fig:A-sensor} A sensor with a finite sensing range and a limited width viewing angle. The coverage area of the sensor is described by a circular sector. $x$ is the location of the sensor, $r$ is the limited sensing range, $\theta$
is the limited width of viewing angle, and $v$ is the viewing direction.
}
\end{figure}
We use a normalized viewing vector $\vec{n}_{x}(y)=\frac{y-x}{\left|y-x\right|}$
which connects an initial point $x$ with a terminal point $y$. Then,
we define the sensor coverage as a level set function $\phi$ given
by
\begin{equation}
\phi(y;x,r,v)=\underset{z\in\mathcal{L}(x,y)}{\min}\left\{ \psi(z),r-|z-x|,\varphi(x,y)\right\} \label{eq:visibility-level-set-fcn}
\end{equation}
where $\mathcal{L}(x,y)$ is the line segment connecting $x$ and
$y$, and $\varphi(x,y)=1$ if the angle of $\vec{n}_{x}(y)$ from
the horizontal axis lies in $[v,v+\theta]$ and $\varphi(x,y)=-1$,
otherwise. This way the function $\phi(\cdot;x,r)$ represents the
coverage in a bounded domain $D$ by
\begin{equation}
\begin{cases}
\phi(\cdot;x,r,v)>0, & \mbox{covered regions,}\\
\phi(\cdot;x,r,v)\leq0, & \mbox{non-covered regions.}
\end{cases}\label{eq:visible}
\end{equation}
With the coverage level set function $\phi(y;x,r,v)$ we can calculate
the coverage area of the sensor at $x$ with the viewing direction $v$
by
\[
V(x,v)=\int_{D}H(\phi(y;x,r,v))dy
\]
where $H$ denotes the one-dimensional Heaviside function.
To compute the coverage function $\phi(y;x,r,v)$, we present a simple
strategy in Algorithm~\ref{PDE-based-algorithm} which follows
the PDE based strategy presented in \cite{TCOBS:JCP2004}. The main
idea is to use the method of characteristics for the nonlinear first
order PDEs, and we modify to handle the finite sensing
range and the limited width of viewing angle.
\bigskip
\begin{algorithm}
\noindent \textup{\caption{PDE based algorithm to compute coverage}\label{PDE-based-algorithm}}
Input: level set function $\psi$, a sensor $x\in D\backslash\Omega$,
radius $r$, and angles $v,\;\theta$.
1. Set $\phi(y)=\psi(y)$ for $y\in D$.
2. If $y\in B_{r}(x),$ solve for $\phi$:
\begin{equation}
\nabla_{y}\phi(y;x,r,v)\cdot\vec{v}_{x}(y)=0,\label{eq:1st-PDE}
\end{equation}
subject to the boundary condition $\phi=\psi$.
$\quad$If else, set $\phi(y;x,r,v)=-1$.
3. Update $\phi=\min(\phi,\psi,r-|y-x|,\varphi(x,y))$.
\end{algorithm}
We solve (\ref{eq:1st-PDE}) using the upwind finite differencing
scheme. Note the computational workload in each iteration is dominated
by the second step. However, these steps are only needed for computing
$\phi$ inside the sensor's coverage range, so we can save the computations.
Generalization of the above setting for multiple sensors is immediate.
Let $\{x_{1},\cdots,x_{m}\}$ denote the location of $m$ sensors.
We also let $\{r_{1},\cdots,r_{m}\}$ be the sensing range of the individual sensors,
$\{\theta_{1},\cdots,\theta_{m}\}$ be the limited width of viewing angle, and $\{v_{1},\cdots,v_{m}\}$
be the viewing direction of each, respectively. The coverage of multiple
sensors is the union of the coverage of all sensors. Similar to (\ref{eq:visibility-level-set-fcn}),
we define the coverage level set function with respect to multiple
sensors by
\[
\phi(y;x_{1},\cdots,x_{m},r_{1},\cdots,r_{m},v_{1},\cdots,v_{m})=\underset{i=1,\cdots,m}{\max}\phi(y;x_{i},r_{i},v_{i})
\]
where the associated $\varphi(x_{i},y)$ in $\phi$ is 1 if the angle
of $\vec{n}_{x_{i}}(y)$ from the horizontal axis is in $[v_{i},v_{i}+\theta_{i}]$
and $\varphi(x_{i},y)=-1$, otherwise. We apply Algorithm \ref{PDE-based-algorithm}
to compute $\phi(y;x_{i},r_{i},v_{i})$ with respect to each sensor,
and then take the union to obtain the combined coverage. Moreover,
the coverage area of multiple sensors with respect to the coverage
range $\{r_{1},\cdots,r_{m}\}$ and viewing angles $\{v_{1},\cdots,v_{m}\}$
is given by
\begin{equation}
V(x_{1},\cdots,x_{m},v_{1},\cdots,v_{m})=\int_{\Omega}H(\phi(y;x_{1},\cdots,x_{m},r_{1},\cdots,r_{m},v_{1},\cdots,v_{m}))dy.\label{eq:joint_visible_volume_function}
\end{equation}
We note that $V$ is Lipschitz continuous under the condition that
there is no fattening in the level set function $\phi$ near the boundary of the obstacles.
The details can be found in \cite{CT:CMS2005}.
In the following sections, we focus on maximizing this coverage area with multiple sensors.
\subsection{Optimal viewing direction (fixed locations)\label{subsec:Optimal-displacement-stationary}}
We propose a procedure for the optimal position of multiple
stationary sensors to maximize the surveillance area. Our approach incorporates
the previous level set formulation into the optimization methods.
We first consider the case when the location of sensors is fixed and only the viewing direction is allowed to change. The goal is to find the optimal viewing direction of each sensor that maximizes the coverage
area $V$ in (\ref{eq:joint_visible_volume_function}) with fixed
locations $\{x_{1},\cdots,x_{m}\}$. We use the gradient
ascent method, i.e., following the gradient flow to steer the viewing direction:
\begin{equation}
\partial_{t}v_{i}=\nabla_{v_{i}}V(v_{1},\cdots,v_{m}),\; i=1,2,\cdots,m,\label{eq:gd_angle}
\end{equation}
where $\nabla_{v_{i}}$ is the gradient operator with respect to the angle $v_{i}$.
To solve (\ref{eq:gd_angle}) numerically, we use the forward difference
to approximate the time derivative and the central difference to approximate
the gradient operator and, which leads to
\begin{equation}
D_{+}^{k}v_{i}=D_{0}^{h_{v}}V(v_{1},\cdots,v_{m}),\; i=1,2,\cdots,m,\label{eq:gd_central}
\end{equation}
where $k$ is a temporal step size and $h_{v}$ is the step size for
the angle.
The gradient of $V$ requires
to compute $\phi(y;x_{i},r_{i},v_{i}\pm h_{v})$. We use
the Algorithm \ref{PDE-based-algorithm} for computing $\phi$ over
the grid with equally spaced points with a spacing of spatial step
size $h$ in the computational domain $D$.
In the evaluation of V, we numerically
integrate $H(\phi)$ using a trapezoidal rule over $D$ and
regularize the Heaviside function by the method proposed in \cite{ETT:2005}:
\[
H_{\epsilon}(\phi)=\begin{cases}
1, & \phi\geq\epsilon,\\
\frac{1}{2}(1+\frac{\phi}{\epsilon}), & |\phi|<\epsilon,\\
0, & \phi\leq-\epsilon,
\end{cases}
\]
using the point-wise scaling with $l_1$ norm of $\nabla\phi$, $\epsilon=\frac{h}{2}\left|\nabla\phi\right|_{l_{1}}$.
The gradient method (\ref{eq:gd_central}) finds a local rather than
a global optimum. To find the globally optimal solution, we employ
the intermittent diffusion method \cite{ID} which adds random perturbations
to the gradient flow (\ref{eq:gd_central}). This helps the deployment
move out of local traps and have a chance to find the globally optimal
one. The following SDE is proposed to be solved before the gradient
ascent flow,
\begin{equation}
dv_{i}(t)=\nabla_{v_{i}}V(v_{1},\cdots,v_{m})dt+\sigma(t)dW(t),\; i=1,\cdots m,
\label{eq:id_sde}
\end{equation}
where $W(t)$ is the Brownian motion in $\mathbb{R}^{m}$, and $\sigma(t)$
is a piecewise constant function alternating between zero and a random
positive constant. The random component perturbs the initial condition
for the gradient ascent so that the solution can escape from the trap
of a local maximizer and approach other ones.
\begin{algorithm}[H]
\textup{\caption{Optimal viewing direction for globally maximal coverage}\label{alg:Multiple-observer-algorithm}}
\textbf{Input}: sensors $\{x_{1},\cdots,x_{m}\}$, radii
of coverage $\{r_{1},\cdots,r_{m}\}$, viewing directions $\{v_{1},\cdots,v_{m}\}$,
limited widths $\{\theta_{1},\cdots,\theta_{m}\}$, step size for location
$h_x$, step size for angle
$h_{v}$, and temporal step size $k$.
\textbf{Initialize}: Randomly locate the sensors on the allowable area.
\textbf{Repeat}: Set $j=1$ and iterate $N$ times to obtain $N$
different sets of positioning.
$\quad$$\quad$1. Set $\alpha$ as the scale for diffusion strength,
and $\gamma$ the scale for diffusion time. Let $\sigma:=\alpha d$,
$\quad$$\quad$and $T:=\gamma t$ where two positive random numbers
$d$, $t$ within $[0,1]$ by uniform distribution.
$\quad$$\quad$2. Set the optimal deployment $L_{opt}=(v_{1},\cdots,v_{m}).$
$\quad$$\quad$3. Taking $L_{opt}$ as the initial condition, compute
the SDE for $t\in[0,T],$ $i=1,\cdots,m.$
\
dv_{i}(t)=\nabla_{v_{i}}V(v_{1},\cdots,v_{m})dt+\sigma(t)dW(t),\; i=1,\cdots m,\label{eq:sde_id}
\
$\quad$$\quad$and record the final configuration \textup{as $L_{ini}=(v_{1},\cdots,v_{m}).$}
$\quad$$\quad$4. Compute the gradient ascent flow until convergence
with the initial condition $L_{ini}$
\
\partial_{t}v_{i}=\nabla_{v_{i}}V(v_{1},\cdots,v_{m}),\; i=1,2,\cdots,m\label{eq:ga_id}
\
$\quad$$\quad$and record the final configuration as $L_{candidate}.$
$\quad$$\quad$5. If $V(L_{candidate})>V(L_{opt})$, set $L_{opt}=L_{candidate}$.
$\quad$$\quad$6. Repeat with $j=j+1.$
\end{algorithm}
One iteration of the procedure is to solve a SDE \eqref{eq:id_sde}
for $t\in[0,T]$ and then compute the gradient ascent flow \eqref{eq:gd_central} until convergence.
This procedure finds one of the globally optimal solutions with probability arbitrarily close to one.
Indeed, denoting $\alpha\in(0,1)$ by the probability that one iteration attains a close approximation to the globally optimal solution,
the rate of convergence is proven to be $1-(1-\alpha)^N$ with $N$ iterations. See \cite{ID} for further details.
Figure \ref{fig:multi-stationary-id} illustrates a result of the
algorithm applied to 8 stationary sensors to cover the environment
with polygonal obstacles. For simplicity, we assumed that all sensors have the same sensing range $r$ and the width of viewing angle $\theta$. In
subfigure (a), the obstacles are the shaded three polygons, and the
locations of each sensor are fixed at red crosses. The dotted sector
represents the sensor's coverage if any obstruction is not given by
an obstacle. Otherwise, the red (shaded) area inside the sector shows
the coverage. The objective is now translated into finding the viewing directions such that the union of each coverage is largest. Algorithm
\ref{alg:Multiple-observer-algorithm} seeks one of the desired sets
of the viewing direction $\{v_{1},\cdots,v_{m}\}$ which globally maximizes
(\ref{eq:joint_visible_volume_function}). The final result is presented
in subfigure (b) with $N=80$ iterations.
\begin{figure}[h]
\begin{centering}
\subfloat[Initial visible volume is 0.9678.]{\begin{centering}
\includegraphics[width=4cm]{figs/8figure2_p0}
\par\end{centering}
}\subfloat[Final visible volume is 1.4184.]{\begin{centering}
\includegraphics[width=4cm]{figs/8figure2_sta50_50_opt}
\par\end{centering}
}\subfloat[]{\begin{centering}
\includegraphics[width=4cm]{figs/hist_ID_stationary}
\par\end{centering}
}
\par\end{centering}
\caption{\label{fig:multi-stationary-id}Optimal viewing directions for multiple sensors
with fixed locations. The viewing directions of multiple sensors
change from the image (a) to the image (b) by Algorithm \ref{alg:Multiple-observer-algorithm}.
(c) The value of coverage area with respect to the iteration. The
maximum coverage is achieved at the 51th iteration. Parameters are
chosen as $r=0.5$, $h=0.005$, $\theta=\pi/2$, $h_{v}=4h$, and
$h_{x}=h$.}
\end{figure}
\subsection{Optimal Sensor Positioning (OSP)\label{subsec:Optimal-displacement-mobile}}
We now consider the case where one can adjust the location of sensors.
The objective is extended to find optimal locations as well as viewing directions that globally maximize the coverage area $V$
in (\ref{eq:joint_visible_volume_function}). The optimal positioning problem can be reformulated as a $2m$-dimensional problem:
\[
\underset{x_{i},v_{i}\in\mathcal{A}}{\max}V(x_{1},\cdots,x_{m},v_{1},\cdots,v_{m})
\]
where $\mathcal{A}$ is the feasible set of sensors' locations and
viewing directions. Typical locations of the sensors are on the boundary of obstacles $\partial\Omega$.
Among the sensors, we denote a subset of sensors whose locations are adjustable
by $\{\tilde{x}_{1},\tilde{x}_{2},\cdots,\tilde{x}_{\tilde{m}}\}$,
i.e., we take $\{\tilde{x}_{1},\tilde{x}_{2},\cdots,\tilde{x}_{\tilde{m}}\}$
$\subseteq$ $\{x_{1},\cdots,x_{m}\}$, and each $\tilde{x}_{\tilde{i}}$
is uniquely identified with some sensor $x_{i}$. Extending the
previous approach for fixed sensors, we employ the intermittent
diffusion and the gradient ascent to maximize the coverage area. However,
due to the adjustability of some sensors, a parametrization of the
boundary of obstacles is required to conduct the optimization with
respect to the location. This can be
done by using a reinitialization method given in \cite{TO:Acta-Numerica:2005} and parameterizing zero level curves which represent boundaries
of obstacles.
Now, the gradient ascent is used for the modified equation of (\ref{eq:gd_angle}):
\begin{eqnarray}
\partial_{t}v_{i} & = & \nabla_{v_{i}}V(x_{1},\cdots,x_{m},v_{1},\cdots,v_{m}),\; i=1,2,\cdots,m,\nonumber \\
\partial_{t}\tilde{x}_{j} & = & \nabla_{\tilde{x}_{j}}V(x_{1},\cdots,x_{m},v_{1},\cdots,v_{m}),\; j=1,2,\cdots,\tilde{m},\label{eq:gd_mobile}
\end{eqnarray}
where $\nabla_{v_{i}}$ is the gradient operator with respect to the viewing direction $v_{i}$,
and $\nabla_{\tilde{x}_{j}}$ is the gradient operator with respect
to the location $\tilde{x}_{j}$ along the boundary of obstacle where the sensor
$\tilde{x}_{j}$ lies in.
As in (\ref{eq:gd_angle}), when the solution reaches a steady state, this gives one of many local optimal solutions. To find a globally optimal solution, we use the intermittent diffusion. It adds random perturbations intermittently to the gradient flow (\ref{eq:gd_mobile})
which leads to the following SDEs,
\begin{eqnarray}
\partial_{t}v_{i} & = & \nabla_{v_{i}}V(x_{1},\cdots,x_{m},v_{1},\cdots,v_{m})dt+\sigma_{1}(t)dW_{1}(t),\; i=1,2,\cdots,m,\nonumber \\
\partial_{t}\tilde{x}_{j} & = & \nabla_{\tilde{x}_{j}}V(x_{1},\cdots,x_{m},v_{1},\cdots,v_{m})dt+\sigma_{2}(t)dW_{2}(t),\; j=1,2,\cdots,\tilde{m},\label{eq:ID_loca-1}
\end{eqnarray}
where $W_{1}(t)$ and $W_{2}(t)$ are the Brownian motions in $\mathbb{R}^{m}$
and $\mathbb{R}^{\tilde{m}}$, respectively. In addition, $\sigma_{1}(t)$
and $\sigma_{2}(t)$ are piecewise constant functions alternating
between zero and a random positive constant. This set of SDEs is followed
by the gradient ascent which can escape from the trap of a local minimizer
and approach other ones.
Algorithm \ref{alg:Multiple-movable observer-algorithm} presents Position Optimization for Sensors (POS) for maximal coverage area. Here we use the term positioning to describe both locations and viewing directions.
\begin{algorithm}
\textup{\caption{OSP}\label{alg:Multiple-movable observer-algorithm}}
\textbf{Input}: sensors $\{x_{1},\cdots,x_{m}\}$, radii
of coverage $\{r_{1},\cdots,r_{m}\}$, viewing directions $\{v_{1},\cdots,v_{m}\}$,
limited widths $\{\theta_{1},\cdots,\theta_{m}\}$, step size for location
$h_{x}$, step size for angle $h_{v}$, and temporal step size
$k$. Also, adjustable sensors\textup{ $\{\tilde{x}_{1},\tilde{x}_{2},\cdots,\tilde{x}_{\tilde{m}}\}\subseteq\{x_{1},\cdots,x_{m}\}.$}
\textbf{Initialize}: Randomly locate the sensors on the allowable area.
\textbf{Repeat}: Set $j=1$ and iterate $N$ times to obtain $N$
different deployments.
$\quad$$\quad$1. Set $\alpha$ as the scale for diffusion strength,
and $\gamma$ the scale for diffusion time. Let $\sigma:=\alpha d$,
and $T:=\gamma t$ where two positive random numbers
$d$, $t$ within $[0,1]$ by uniform distribution.
$\quad$$\quad$2. Set the optimal deployment $L_{opt}=(x_{1},\cdots,x_{m},v_{1},\cdots,v_{m}).$
$\quad$$\quad$3. Taking $L_{opt}$ as the initial condition, compute
the SDE (\ref{eq:ID_loca-1}) for $t\in[0,T]$ and record the
final configuration \textup{as $L_{ini}=(x_{1},\cdots,x_{m},v_{1},\cdots,v_{m}).$}
$\quad$$\quad$4. Compute the gradient ascent flow (\ref{eq:gd_mobile})
until convergence with the initial condition $L_{ini}$ and record the final configuration as $L_{candidate}.$
$\quad$$\quad$5. If $V(L_{candidate})>V(L_{opt})$, set $L_{opt}=L_{candidate}$.
$\quad$$\quad$6. Repeat with $j=j+1.$
\end{algorithm}
Evaluations of $\nabla_{0}^{h_{v}}V$ and $\nabla_{0}^{v_{x}}V$
require to compute $\phi(y;x_{i},r_{i},v_{i}\pm h_{v})$ for
$i=1,\cdots,m$ and $\phi(y;\tilde{x}_{j}\pm h_{x},r_{i},v_{i})$ for $i=1,\cdots,\tilde{m}$ if $x_{j}$
is allowed to move along the boundary. In Step 4, we use Euler's method
to update $v_{i}$ by $v_{i}+kD_{0}^{h_{v}}V(x_{1},\cdots,x_{m},v_{1},\cdots,v_{i},\cdots,v_{m})$
for $i=1,\cdots,m$ and $\tilde{x}_{j}$ by $\tilde{x}_{j}+kD_{0}^{h_{x}}V(x_{1},\cdots,\tilde{x}_{j},\cdots,x_{m},v_{1},\cdots,v_{m})$
for movable $\tilde{x}_{j}$, respectively. Figure \ref{fig:multi-mobile}
describes a result of Algorithm \ref{alg:Multiple-movable observer-algorithm}
applied to the same environment in Figure \ref{fig:multi-stationary-id},
but all of the sensors
are movable except for the two sensors not on the boundary of any obstacle. An optimal position which locally maximizes (\ref{eq:joint_visible_volume_function})
is presented in subfigure (b). It is evident that the resulting coverage
area is much larger than the one in Figure \ref{fig:multi-stationary-id}
because mobile sensors have more degrees of freedom than that of stationary
sensors. For simplicity, we suppose that all sensors have the same
sensing range $r$ and width of viewing angle $\theta$.
\begin{figure}
\begin{centering}
\subfloat[Initial visible volume is 0.9678.]{\begin{centering}
\includegraphics[width=4cm]{figs/8figure2_p0}
\par\end{centering}
}\subfloat[Final visible volume is 1.5461.]{\begin{centering}
\includegraphics[width=4cm]{figs/8figure2_mobile0\lyxdot 5_50_opt}
\par\end{centering}
}\subfloat[]{\begin{centering}
\includegraphics[width=4cm]{figs/hist_ID_mobile}
\par\end{centering}
}
\par\end{centering}
\caption{\label{fig:multi-mobile}
Optimal locations and viewing directions of multiple sensors.
The locations and viewing directions of multiple sensors change from the
image (a) to the image (b) by Algorithm \ref{alg:Multiple-movable observer-algorithm}. (c) The
value of coverage area with respect to the iteration. The maximum
coverage is achieved at the 91st iteration. Parameters are chosen
as $r=0.5$, $h=0.005$, $\theta=\pi/2$, $h_{v}=4h$, and $h_{x}=h$.}
\end{figure}
\section{Optimal sensor positioning with a failure rate\label{sec:possible_failure}}
The failure of sensors is crucial for
the surveillance system to achieve its intended task, and
this can be added to the stage of sensor positioning.
We extend the previous approach to find optimal positions when the sensors have a non-zero probability of failure.
As a priori estimate, we consider an expected coverage area of the sensors in the objective functional and analyze an optimal solution to maximize it. If there are sufficiently many sensors in the sense that the collective coverage area is much larger than the area of $D\backslash\Omega$, the optimal solution will be shown to uniformly spread the sensors over the monitoring area.
We assume that sensors fail independently, i.e., if two sensors
have a failure probability $p_{1}$ and $p_{2}$, respectively, then
$p_{1}p_{2}$ is the joint failure probability. Given a point $y$
in the environment, we denote $0<p(y)<1$ by a probability that a
set of sensors fails to detect $y$. If $y$ is not in any
of the sensor's coverage, then $p(y)$ is set to be 1. Furthermore,
we denote $p_{k}(y)$ by a probability that the individual $k$-th
sensor fails to detect a point $y$.
The expectation of the coverage area with possible sensor failures is defined as
\begin{equation}
\mbox{\ensuremath{\mathbb{E}}}(V(x_{1},\cdots,x_{m},v_{1},\cdots,v_{m}))=\int_{D}H(\phi(y;x_{1},\cdots,x_{m},r_{1},\cdots,r_{m},v_{1},\cdots,v_{m}))(1-p(y))dy.\label{eq:expected_coverage}
\end{equation}
That is, the functional is the integral of the products of covered area with the probabilities that the sensors to cover it are working.
Note that if there is no sensor failure, then $p(y)=0$ so that the
objective functional (\ref{eq:expected_coverage}) is reduced to (\ref{eq:joint_visible_volume_function}).
The optimization problem is to maximize the expected area of coverage
which is formulated as:
\begin{equation}
\underset{x_i, v_i}{\max} \;\mbox{\ensuremath{\mathbb{E}}}(V(x_{1},\cdots,x_{m},v_{1},\cdots,v_{m}))\label{eq:objective_functional_random}
\end{equation}
subject to the locations of the sensors adjustable on either the boundary of obstacles
$\partial\Omega$ or the boundary of the computational domain.
For simplicity, we discretize a two dimensional environment by uniformly placing a grid $(ih,jh)$
on it and denote $0<p_k (i,j)<1$ by a probability that the $k$-th sensor fails
to detect a grid point $(ih,jh)$.
With an uniformly space
grid, the expected area of sensors' coverage is numerically approximated
by
\[
\underset{i,j}{\sum}h^{2}\left(1-\overset{n}{\underset{k=1}{\prod}}p_{k}(i,j)\right).
\]
The following considers how the multiple sensors should be positioned in order to
maximize the objective functional \eqref{eq:expected_coverage}. We
show that positioning multiple sensors with less overlapping area is closer to the
optimal solution than doing so with more overlapping area.
\begin{prop}
Positioning multiple sensors with minimal overlapping regions attains the greater
expected coverage area than those with
more overlapping regions.
\end{prop}
\begin{proof}
Suppose that the small region $\mathcal{R}$ with an area $B$ is
given, and that this region is fully covered by $m$ sensors. Firstly,
suppose that their coverages do not intersect.
We denote $\mathcal{R}_{nonoverlap}$ by the sensor positions where all sensors do not intersect and cover
entire $\mathcal{R}$. Then the contribution
of this sensor positioning to the total expected coverage area is
\begin{equation}
\mathbb{E}[\mathcal{R}_{nonoverlap}]=B\left(\overset{m}{\underset{k=1}{\sum}}(1-p_{k})\right)
\end{equation}
where the $k$-th sensor has a failure probability $p_{k}$.
Secondly, suppose that several $\tilde{m}$ $(\leq m)$ sensors overlap and fully
cover $\mathcal{R}$. In this case, the contribution to the total
expected coverage area is
\begin{eqnarray}
\mathbb{E}[\mathcal{R}_{overlap}] & = & B\left(1-\overset{\tilde{m}}{\underset{k=1}{\prod}}p_{k}\right).
\end{eqnarray}
Now, we compare $\overset{m}{\underset{k=1}{\sum}}(1-p_{k})$ with
$1-\overset{\tilde{m}}{\underset{k=1}{\prod}}p_{k}$. Since the failure
probability is between 0 and~1,
\[
1-\overset{m}{\underset{k=1}{\prod}}p_{k}>1-\overset{\tilde{m}}{\underset{k=1}{\prod}}p_{k}.
\]
Using the mathematical induction, suppose that
\[
\overset{m}{\underset{k=1}{\sum}}(1-p_{k})\geq1-\overset{m}{\underset{k=1}{\prod}}p_{k}
\]
holds true. Then, multiplying $p_{n+1}$ yields that
\[
p_{n+1}\overset{n}{\underset{k=1}{\sum}}(1-p_{k})\geq p_{n+1}-\overset{n+1}{\underset{k=1}{\prod}}p_{k}.
\]
Now,
\begin{eqnarray*}
\overset{n+1}{\underset{k=1}{\sum}}(1-p_{k}) & \geq & \overset{n+1}{\underset{k=1}{\sum}}(1-p_{k})-(1-p_{n+1})\overset{n}{\underset{k=1}{\sum}}(1-p_{k})\\
& = & 1-p_{n+1}+p_{n+1}\overset{n}{\underset{k=1}{\sum}}(1-p_{k})\\
& \geq & 1-\overset{n+1}{\underset{k=1}{\prod}}p_{k}.
\end{eqnarray*}
Therefore, $\mathbb{E}[\mathcal{R}_{nonoverlap}]\geq\mathbb{E}[\mathcal{R}_{overlap}]$.
\end{proof}
\noindent
In other words, when the sensor fails with non-zero probability, spreading multiple
sensors over the monitoring area as much as possible attains the largest
expectation.
Consequently, if the algorithm maximizes \eqref{eq:expected_coverage},
the optimal position will not only cover all of the free areas,
$D\backslash\Omega$, but also minimize the overlaps among the sensors.
We compare two scenarios in the following section; optimal positioning of the sensors with or without a failure rate
and test the performance of our proposed algorithm.
\section{Numerical applications\label{sec:more_examples}}
In this section, various numerical experiments are given. We use Algorithm~\ref{alg:Multiple-movable observer-algorithm} with the functional \eqref{eq:expected_coverage} for most of the examples, considering a nonzero failure rate. In one example depicted in Figure \ref{fig:no_fail} where the failure rate $p=0$, we use the objective functional (\ref{eq:joint_visible_volume_function}).
\subsection{Comparison: failure rate $p=0$ and $p\neq 0$}
Two examples compare the sensor positioning with or without a failure rate.
We first consider a problem with sensors which work perfectly, i.e., the
probability of failure is $p=0$. With a set of these sensors, a rectangular
environment without obstacles will be monitored. Suppose that the
number of sensors is large so that the environment can be fully covered
with some redundancies. Figure \ref{fig:no_fail}(a) shows an initial
position of 16 sensors which is randomly determined by putting 8
sensors for each side. We display the number of overlaps in the sensors
coverage by color scales using a horizontal color bar. The area of the
region to be covered is $1$. With the finite sensing range $r=0.6$
and the limited width of sensing angle $\theta=\pi/3$, the collective sensors\rq{}s
coverage area, if they are disjoint, is 3.016.
As shown in subfigures (b) and (c), with iterations our algorithm
focuses on covering all of the regions without any empty space to
achieve the optimal solution. The number of coverage overlaps is not
considered.
\begin{figure}
\begin{centering}
\hspace*{-1.0cm}\includegraphics[width=4cm]{figs/colorbar}
\par\end{centering}
\begin{centering}
\hspace*{-0.5cm}\subfloat[Initial visible area is 0.4131.]{\begin{centering}
\includegraphics[width=4cm]{figs/16ID_fail_tube_many_p0_60_0}
\par\end{centering}
}\subfloat[After 10 iterations, resulting visible area is 0.9928.]{\begin{centering}
\includegraphics[width=4cm]{figs/16ID_fail_tube_many_p0_60_1}
\par\end{centering}
}\subfloat[After 50 iterations, resulting visible area is 0.9996.]{\begin{centering}
\includegraphics[width=4cm]{figs/16ID_fail_tube_many_p0_60_2}
\par\end{centering}
}
\par\end{centering}
\caption{\label{fig:no_fail} Position optimization with no possible failure.
The resulting sensor positions are obtained using the parameters
$r=0.6$, $h=0.005$, $\theta=\pi/3$. $N=50$ iterations of our algorithm are used.
The horizontal colorbar displays how many sensors cover the area using the colormap.}
\end{figure}
In real applications, sensors may fail.
Assuming that the sensor fails to work with a probability $p$, our algorithm
maximizing (\ref{eq:expected_coverage}) not only focuses on covering
all of the free areas but also minimizes the overlaps among the sensors
as proven in Proposition~1. We take the same environment of Figure~\ref{fig:no_fail}, but assume that each sensor has a 50\% chance of failure,
i.e., with a probability $p=0.5.$ Figure \ref{fig:can_fail} shows
the results of our algorithm. Since we compute the expected area of
covered region with a probability, the initial value of the objective
functional is different from the one in Figure \ref{fig:no_fail}.
The algorithm automatically finds the optimal deployment of sensors
which leads to the full coverage as well as the minimal area of the
intersections. The subfigure (c) shows evenly distributed coverages
of sensors compared to Figure \ref{fig:no_fail}(c).
\begin{figure}[H]
\begin{centering}
\hspace*{-1.0cm}\includegraphics[width=4cm]{figs/colorbar}
\par\end{centering}
\begin{centering}
\hspace*{-0.5cm}\subfloat[Initial expected area is 0.3130.]{\begin{centering}
\includegraphics[width=4cm]{figs/16ID_fail_tube_many_p0_60_0}
\par\end{centering}
}\subfloat[After 10 iterations, resulting expected visible area is 0.7810.]{\begin{centering}
\includegraphics[width=4cm]{figs/16ID_fail_tube_many_p0\lyxdot 5_60_1}
\par\end{centering}
}\subfloat[After 50 iterations, resulting expected visible area is 0.8079.]{\begin{centering}
\includegraphics[width=4cm]{figs/16ID_fail_tube_many_p0\lyxdot 5_60_38}
\par\end{centering}
}
\par\end{centering}
\caption{\label{fig:can_fail} Position optimization with a possible failure. The
resulting sensor placements are obtained using the parameters $r=0.6$,
$h=0.005$, $\theta=\pi/3$. $N=50$ iterations of our algorithm
are used. Each sensor has a 50\% chance of not working, $p=0.5.$
Compared to Figure \ref{fig:no_fail}, the result shows more uniformly distributed coverages.
}
\end{figure}
\subsection{General examples for $p \neq 0$}\label{subsec:general_examples}
We consider more general shapes of environment. The first is
a narrow alley with a corner. Given a sufficient number of sensors with the same failure rate, we will optimally position
them on the walls to maximize an expectation of the collective coverage. The second is an arbitrary polygon,
and we consider the problem of optimal positioning on the boundary.
Figure \ref{fig:corner} illustrates how multiple sensors are optimally positioned to monitor the alley with a corner.
4 sensors are randomly located on each wall facing the alley, i.e., a total of 16 sensors, as in subfigure (a).
Red crosses indicate
the sensor locations, and sectors with dotted lines
describe the sensors\rq{} coverage.
The optimal positions found by our algorithm with 50 iterations are presented in subfigure (b).
The coverage of sensors is uniformly distributed across the target area.
This environment will be generalized
to 3D in Section \ref{subsec:3d}.
\begin{figure}[H]
\begin{centering}
\hspace*{-1.0cm}\includegraphics[width=4cm]{figs/colorbar}
\par\end{centering}
\begin{centering}
\subfloat[Initial expected coverage area is 0.3313.]{\begin{centering}
\includegraphics[width=4cm]{figs/16ID_fail_corner_p0\lyxdot 5_80_0}\;
\includegraphics[width=4cm]{figs/16ID_fail_corner_many_p0\lyxdot 5_80_0}
\par\end{centering}
}\subfloat[After 50 iterations, resulting expected visible area is 0.9020.]{\begin{centering}
\includegraphics[width=4cm]{figs/16ID_fail_corner_many_p0\lyxdot 5_80_118}
\par\end{centering}
}
\par\end{centering}
\caption{\label{fig:corner} Optimal positioning in an alley with a corner. The resulting
sensor positions are obtained using the parameters $r=0.8$, $h=0.005$,
$\theta=\pi/2$, $p=0.5$. 16 sensors are considered. The result shows uniformly distributed coverages.}
\end{figure}
In Figure \ref{fig:complicated}, optimal sensor positioning in the environment with polygonal obstacles is considered.
This scenario can be interpreted as positioning security surveillance cameras on the outside (e.g., rooftop) of urban buildings.
The polygons are shaded gray, and the failure rate of the sensors is taken as 0.5.
We apply our algorithm to optimally position 15 sensors on the boundary. The result in subfigure (b) is obtained with 200 iterations.
As shown in the figure, our algorithm can achieve a maximum area of coverage while handling complicated shapes of obstacles.
\begin{figure}[h]
\begin{centering}
\hspace*{-1.0cm}\includegraphics[width=4cm]{figs/colorbar}
\par\end{centering}
\begin{centering}
\subfloat[Initial expected coverage area is 0.2675.]{\begin{centering}
\includegraphics[width=4cm]{figs/15ID_city_p0\lyxdot5_50_0}\;
\includegraphics[width=4cm]{figs/15ID_city_many_p0\lyxdot5_50_0}
\par\end{centering}
}\subfloat[After 200 iterations, resulting expected visible area is 0.5445.]{\begin{centering}
\includegraphics[width=4cm]{figs/15ID_city_many_p0\lyxdot5_50_259}
\par\end{centering}
}
\par\end{centering}
\caption{\label{fig:complicated} Optimal positioning in an environment with polygonal obstacles. The resulting
sensor placements are obtained using the parameters $r=0.5$, $h=0.005$,
$\theta=\pi/3$, $p=0.5$. 15 sensors are considered.
This algorithm handles a complicated environment when sensors can move along the boundary.}
\end{figure}
\subsection{Access monitoring}
Access monitoring system plays an important role in securing a building
and its surroundings. Well-designed monitoring systems
are able to detect all of the suspicious activities while considering a possible failure of the sensors.
In this section, we consider a problem of optimally positioning the sensors to monitor the population entering a building.
We position the sensors on the boundary of the important building. The building is considered as the obstacle, and the sensors move along the boundary of the building to find the optimal positions.
To use our approach with a possible failure rate, we construct the functional as
\begin{equation}
\mbox{\ensuremath{\mathbb{E}}}(V(x_{1},\cdots,x_{m},v_{1},\cdots,v_{m}))=\int_{D} \chi_A (y)H(\phi(y;x_{1},\cdots,x_{m},r_{1},\cdots,r_{m},v_{1},\cdots,v_{m}))(1-p(y))dy
\nonumber
\end{equation}
where $\chi_A$ is the indicator function of a set $A$, and $0<p(y)<1$ is a probability that a set of sensors fails to detect $y$.
Here, the set $A$ is a closed region which encloses the building so that the floating population must traverse this region in order to either enter or exit the building. Various shapes of $A$ can be considered, and we use a narrow band around the building for simplicity.
The main difference from the previous examples is that the area of full coverage only needs to be considered on a narrow band around the building as long as the access is fully covered.
In the following experiments, we consider a pentagon shaped building located at the center of the computational domain.
Two red pentagons enclosing the building are the monitored region $A$. Note that one cannot move out or in to the building without passing the strip $A$.
We employ the algorithm which maximizes the above functional to find
the optimal position as well as the viewing direction of sensors while considering the sensor's possible failure.
In the case of symmetric objects, one can consider solving the problem more efficiently
by adding a symmetry constraint to the feasible position.
Figures~\ref{fig:corner-tube-sym}(a-b) show such an example where we position three sensors to each
side symmetrically. One sensor is fixed at the center with a viewing direction normal to the boundary,
and the other two move with symmetry: (1) when one sensor moves closer to the corner, the other moves to the
other corner with the same distance; (2) when the viewing direction of one sensor rotates clockwise, the one of the other rotates
counterclockwise at the same angle.
This problem requires a 2-dimensional optimization (location and viewing direction) to position two sensors away from the center. This case,
from the initial condition in Figure~\ref{fig:corner-tube-sym}(a), the optimal solution is computed as Figure~\ref{fig:corner-tube-sym}(b) with the expected coverage of 0.3366. Under this symmetry constraint, we find that the optimal location of three sensors divides the edge in the ratio 0.29:0.5:0.71. Interestingly, the optimal result does not equally divide the edge into four as one may expect.
\begin{figure}
\begin{centering}
\hspace*{-1.0cm}\includegraphics[width=4cm]{figs/colorbar}
\par\end{centering}
\begin{centering}
\subfloat[Initial expected coverage area is 0.2874.]{\begin{centering}
\includegraphics[width=4cm]{figs/15ID_pentagon_sym_p0\lyxdot5_25_0}\includegraphics[width=4cm]{figs/15ID_pentagon_many_sym_p0\lyxdot5_25_0}
\par\end{centering}
}\subfloat[After 100 iterations, 0.3366.]{\begin{centering}
\includegraphics[width=4cm]{figs/15ID_pentagon_many_sym_p0\lyxdot5_25_result}
\par\end{centering}
}
\par\end{centering}
\begin{centering}
\subfloat[Initial expected coverage area is 0.2422.]{\begin{centering}
\includegraphics[width=4cm]{figs/15ID_pentagon_sym2_p0\lyxdot5_25_0}\includegraphics[width=4cm]{figs/15ID_pentagon_many_sym2_p0\lyxdot5_25_0}
\par\end{centering}
}\subfloat[After 100 iterations, 0.3180.]{\begin{centering}
\includegraphics[width=4cm]{figs/15ID_pentagon_many_sym2_p0\lyxdot5_25_result}
\par\end{centering}
}
\par\end{centering}
\caption{\label{fig:corner-tube-sym} Entrance security monitoring for the pentagon under two symmetry constraints. The resulting
sensor positions are obtained with 15 sensors using the same parameters
$r=0.5$, $h=0.0025$, $\theta=\pi/2$, $p=0.5$.
There are 2 degrees of freedom in optimization.
}
\end{figure}
As a remark, one can also consider different types of symmetry. Figures~\ref{fig:corner-tube-sym}(c-d) show another of such example.
Each sensor is fixed at the corner with a viewing direction exactly opposite to the center. Other two sensors move with symmetric on the edge as before.
The result of our algorithm is given in Figure~\ref{fig:corner-tube-sym}(d). The resulting expected value is not as high as the one in Figure~\ref{fig:corner-tube-sym}(b).
We can solve the problem without any symmetry constraint. This is more suitable for the general shapes.
In this case, however, the degree of freedom is increased from 2 to 30 (position and viewing direction of 15 sensors).
With intermittent diffusion and gradient descent for all variables,
from the initial condition in Figure~\ref{fig:corner-tube}(a), the optimal solution is computed as Figure~\ref{fig:corner-tube}(b) with the expected coverage of 0.3244. The result distributes the sensors and adjusts the viewing directions to uniformly cover the entire area of $A$.
This result provides the better expected coverage area than the one of Figure~\ref{fig:corner-tube-sym}(d) but the worse than the one of Figure~\ref{fig:corner-tube-sym}(b).
Although the expected coverage area is not as sharp as the one with symmetry constraint, this result shows that our algorithm can
handle various cases and approximate the globally optimal solution.
\begin{figure}
\begin{centering}
\subfloat[Initial expected coverage area is 0.2881.]{\begin{centering}
\includegraphics[width=4cm]{figs/15ID_pentagon_p0\lyxdot 5_25_0}\includegraphics[width=4cm]{figs/15ID_pentagon_many_p0\lyxdot 5_25_0}
\par\end{centering}
}\subfloat[After 100 iterations, 0.3244.]{\begin{centering}
\includegraphics[width=4cm]{figs/15ID_pentagon_many_p0\lyxdot 5_25_12}
\par\end{centering}
}
\par\end{centering}
\caption{\label{fig:corner-tube} Entrance security monitoring for the pentagon. The resulting
sensor positions are obtained with 15 sensors using the same parameters as Figure~\ref{fig:corner-tube-sym}.
There are 30 degrees of freedom in optimization.
}
\end{figure}
\subsection{Important regions}
We consider a particular problem which some regions in the monitored area are more important than others.
In other words, the various weights of importance are assigned and given a priori. To apply the previous algorithm, we modify the objective functional as
\begin{equation}
\mbox{\ensuremath{\mathbb{E}}}(V(x_{1},\cdots,x_{m},v_{1},\cdots,v_{m}))=\int_{D}w(y)H(\phi(y;x_{1},\cdots,x_{m},r_{1},\cdots,r_{m},v_{1},\cdots,v_{m}))(1-p(y))dy
\nonumber
\end{equation}
where the weight $w(\cdot)$ is distributed over the environment.
This can take care of the scenario where there are blind spots in the environment we want to monitor.
This setup resembles a realistic case, e.g., sensor positioning for grocery market and book store. To set-up this environment, we first placed a stationary long-range sensor (which can represent the location of the casher), and computed the blind spots using the visibility computation using Algorithm \ref{PDE-based-algorithm}. In Figure \ref{fig:shop0.5} (a), the circle represents the cashier's location, and the red lines behind the gray rectangles (e.g. book shelves) represents the edge of visibility boundary. In addition, near the casher is an important area to monitor, this is represented by the red square box enclosing the red circle.
The objective is to optimally position the sensors to monitor the important regions, i.e., the two separate areas, each enclosed by the red lines. This mission can be easily tackled using our approach.
We assign a weight $1$ only on the important regions and $0$ on the others, then apply Algorithm \ref{alg:Multiple-movable observer-algorithm} to maximize the above functional.
Considering the strategy of sensor's possible failure, the initial positions in the subfigure (b) are changed into the subfigure (c) with 50 iterations. The failure rate is chosen $p=0.5$. As shown in the result, full coverage of the area is not achieved although the maximum of the expected coverage is attained.
Instead of covering the small non-covered regions, sensors make a large overlap in the covered regions for maximum expectation because the contribution of the overlapping areas outweighs that of the non-covered areas. This is also due to the difference between the shape and the topology of the important region and the coverage shape of the sensors.
When the failure rate is lower, one can achieve full coverage.
Figure \ref{fig:shop0.1} shows an example with a smaller failure rate $p=0.1$. We apply Algorithm \ref{alg:Multiple-movable observer-algorithm} to maximize the coverage area, and compared to the case when $p=0.5$, the important regions are both fully covered by the sensors, and the result gives a higher expected coverage.
\begin{figure}
\begin{centering}
\hspace*{-0.6cm}\includegraphics[width=4cm]{figs/colorbar}
\par\end{centering}
\begin{centering}
\subfloat[Important regions.]{\begin{centering}
\includegraphics[width=4cm]{figs/0ID_store_p0\lyxdot5_105_0}
\par\end{centering}
}
\subfloat[Initial expected coverage area on the red regions is 0.5596.]{\begin{centering}
\includegraphics[width=4cm]{figs/9ID_store_many_p0\lyxdot5_105_0}
\par\end{centering}
}
\subfloat[After 50 iterations, 0.8712.]{\begin{centering}
\includegraphics[width=4cm]{figs/9ID_store_many_p0\lyxdot5_105_11}
\par\end{centering}
}
\par\end{centering}
\caption{\label{fig:shop0.5} Important region monitoring with $p=0.5$. The resulting sensor
placements are obtained with 9 sensors using the parameters $r=1.05$,
$h=0.005$, $\theta=\pi/3$. Most of the important regions are covered, but there exist non-covered regions.}
\end{figure}
\begin{figure}
\begin{centering}
\subfloat[Initial expected coverage area on the red regions is 0.8985.]{\begin{centering}
\includegraphics[width=4cm]{figs/9ID_store_many_p0\lyxdot5_105_0}
\par\end{centering}
}
\subfloat[After 50 iterations, 1.1298.]{\begin{centering}
\includegraphics[width=4cm]{figs/9ID_store_many_p0\lyxdot1_105_9}
\par\end{centering}
}
\par\end{centering}
\caption{\label{fig:shop0.1} Important region monitoring with $p=0.1$. The resulting sensor
placements are obtained with 9 sensors using the same parameters as Figure~\ref{fig:shop0.5}.
The important regions are fully covered.}
\end{figure}
\subsection{Higher dimensional example\label{subsec:3d}}
The examples presented so far were restricted to 2-dimensional environments. The proposed method can be easily generalized to the higher dimensions. By modifying the level set formulation in Section~\ref{sec:Coverage-optimization-GD}, the previous approach can be applicable to the higher dimensional setup.
For simplicity, we focus on 3-dimensional which addresses a real world environment.
The higher ($>3$) dimensions can be considered in a similar manner.
A sensor in 3-dimensional space is given
in Figure \ref{fig:3D-sensor}. The coverage of a sensor is modeled
by a spherical sector whose apex point is located at the sensor location.
That is, as described in the right subfigure, we generalize the 2-dimensional
sensor in Figure \ref{fig:A-sensor} to the 3-dimension by rotating it about
the central axis.
\begin{figure}[H]
\begin{centering}
\includegraphics[width=10cm]{figs/3D_icecream1}
\par\end{centering}
\caption{\label{fig:3D-sensor}The coverage area of a 3-dimensional sensor
is described by a spherical sector. The location of the sensor is
at $(0.4,0.5,0.25)$ which is the apex point of a spherical sector.
This sensor has finite sensing distance $r=0.4$ and a base radius
$d=0.8$.}
\end{figure}
Extension to 3D can be done by following \cite{TCOBS:JCP2004}.
First, the level set
formulation introduced in Section \ref{sec:Coverage-optimization-GD}
is converted to the 3D setup. The coverage function $\phi(y;x,r,v)$
is computed by the 3D version of Algorithm \ref{PDE-based-algorithm}.
Then,
the coverage area of multiple sensors is computed by the integration
over the 3D computational domain. Secondly, we maximize the coverage
area with respect to four degrees of freedom for each sensor: two for adjusting
the viewing direction, and the other two for the coordinates on the boundary of the
3D obstacles. The optimization using the intermittent diffusion automatically approximates
one of the globally optimal positions of the sensor in terms of four degrees
of freedom.
A 3D version of the example in Figure \ref{fig:corner} will be constructed on a 3D computational domain whose size is $[0,2]\times[0,2]\times[0,0.5]$.
We place two obstacles: a square box with the dimension $[0,1.4]\times[0,1.4]\times[0,0.5]$ and an upside down letter "L" shaped box with the same height.
In Figure \ref{fig:3D_corner}(a), the left subfigure represents an initial position using a projection onto $x-y$ plane. The numbers besides the red crosses refer to the $z$-coordinate of the position. The middle subfigure shows the non-monitored regions
with respect to the initial position of the sensors when we look at the environment in the front. The obstacles are shaded transparent gray, the blue surface represents the boundary of uncovered regions by the sensors, and the red surface is the cut due to the top $z=0.5$.
While our algorithm changes the initial position as the subfigure (b) with 20 iterations, the
expected coverage area increases from 0.1011 to 0.6259. It is also evident from the results that most of the areas are now covered under the new position of the sensors.
\begin{figure}[h]
\begin{centering}
\subfloat[The initial expected coverage area is 0.1011.
(Left) Projection of the sensor locations onto $x-y$
plane. Red crosses denote the position of the sensor and the numbers shows the height of each sensor.
(Middle) A front side view in 3D. The obstacles are shaded transparent gray, the blue surface represents the boundary of uncovered regions by the sensors, and the red surface is the cut due to the top $z=0.5$.
(Right) A reverse side view. ]{\begin{centering}
\includegraphics[width=3.7cm]{figs/8ID_fail_corner_3D0\lyxdot 5_1_0}
\includegraphics[width=5cm]{figs/8ID_fail_corner_3D0\lyxdot 5_3_0a}\includegraphics[width=5cm]{figs/8ID_fail_corner_3D0\lyxdot 5_3_0b}
\par\end{centering}
}
\par\end{centering}
\begin{centering}
\subfloat[After 20 iterations, the resulting expected coverage area is 0.6259.
(Left) Projection of the sensor locations onto $x-y$
plane. Red crosses denote the position of the sensor and the numbers shows the height of each sensor.
(Middle) A front side view in 3D.
(Right) A reverse side view. ]{\begin{centering}
\includegraphics[width=3.7cm]{figs/8ID_fail_corner_3D0\lyxdot 5_1_19}
\includegraphics[width=5cm]{figs/8ID_fail_corner_3D0\lyxdot 5_3_19a}\includegraphics[width=5cm]{figs/8ID_fail_corner_3D0\lyxdot 5_3_19b}
\par\end{centering}
}
\par\end{centering}
\caption{\label{fig:3D_corner} A 3D alley with a corner. The resulting sensor positions are obtained
with 8 sensors using the parameters: sensing range $r=0.9$, base
radius $d=0.7$, step size $h=0.01$, and failure rate $p=0.5$.
The result shows the movement of the sensors and how most of the areas are covered.
Our algorithm increases
the expected coverage area from 0.1011 to 0.6259.}
\end{figure}
\section{Summary\label{sec:Summary}}
We proposed a novel method for finding the optimal position of sensors
with a limited coverage range and width of viewing angle in a known environment. The
efficient positioning of such equipment is increasingly important
as it directly impacts the efficiency of allocated resources and
system performance. We considered two global optimization problems.
One is to achieve the greatest surveillance area of the region with
the given number of multiple sensors which have characteristics such
as range and angle limits. The other is to optimize the sensor placement
given that each sensor may randomly fails to operate.
In this case, the sensor placement maximizing the expected coverage
area in the environment will be the desired optimal solution. From
the results of several examples, we verified the effectiveness of
our algorithm. The more realistic scenario where sensor coverages are correlated
to each other shall be addressed in a future work.
\selectlanguage{american
\bibliographystyle{plain}
|
\section{Introduction}
The characteristic time $\tau_R$ of the rotational Brownian diffusion for nanoparticles is typically much smaller than the time resolution $t$ in the experiments, $t>>\tau_R$. In this case, nanoparticles can be treated as point-like,
and described by the standard Brownian theory \cite{Kampen}. However, for
Brownian microparticles, the characteristic time
$\tau_R$ is of the order of seconds, and therefore
non-negligible in comparison to the typical time scales $t$ of the measured Brownian motion, $t \lesssim \tau_R$. In this case, the mean translational and rotational displacements of the Brownian particle significantly change with time, and a more general theoretical approach is needed.
On the other hand, nowadays there is a lot of interest in experimental studies of Brownian particles of relatively large sizes \cite{2D,holografic,leptospira,Kraft}.
In Ref. \!\cite{Kraft}, the Brownian motions of non-symmetric microparticles were investigated at the time scales comparable with the characteristic time $\tau_R$ of the rotational Brownian diffusion
The time-dependent cross-correlations of the Brownian translational and orientational displacements of microparticles with different shapes were measured and the initial slopes of these curves were used to experimentally determine the
friction matrices. Based on the details of the particle shape, known from the experiment, these matrices were also evaluated numerically with {\sc hydrosub}, and then used as the input to time-dependent numerical simulations of the Brownian displacement cross-correlations.
Although the results qualitatively agree with each other, but there are significant quantitative differences, which are by the authors explained by the statistical uncertainty of the measurements.
The goal of this comment is to demonstrate that the method to analyze Brownian motion and to extract the friction coefficients, presented in Ref. \cite{Kraft}, can be significantly improved based on the new analytical expressions which have been recently derived from the Smoluchowski equation for
the Brownian motion of a particle with an arbitrary shape \cite{CEW2015JCP,CEW2016}. These expressions allow to fit the the friction (or, equivalently, mobility) coefficients of a particle using its Brownian displacement cross-correlations in the whole range of the measured times. This procedure allows to determine the mobility coefficients (and therefore a more detailed structure of the particle) with a significantly higher precision than in Ref.~\cite{Kraft}.
\section{Goals and theoretical framework}\label{II}
It is worthwhile to consider two generic cases. First, the particle structure and size are known and the goal is to study its translational and rotational Brownian motion. Second, the particle structure is not known, and the Brownian displacement cross-correlations are used to determine its mobility coefficients which in turn provide information abut the particle structure and size. Numerical computations performed in Ref.~\cite{Kraft} correspond to the first case. However, analysis of experimental data is usually related to the second case. The challenging question is if the cross-correlations measurements can be analyzed with a precision high enough to provide information about the particle geometry.
In this Comment, we apply the analytic expressions from Refs.~\cite{CEW2015JCP,CEW2016} to satisfy both goals. In Sec.~\ref{III}, we perform calculations which belong to the first case. In Sec.~\ref{IV}, we generalize this approach to the second case.
The basic theoretical framework used in Sec. \ref{III} is the following.
First, we evaluate the friction matrix \cite{Kim,CEW2015JCP} for the particles at their initial orientations, to compare with the experimental and numerical results from Ref.~\cite{Kraft}.
We follow Ref.~\cite{Kraft}, where the friction matrix was divided by the fluid viscosity $\eta$
and denoted as $\bm{\mathcal H}$. Therefore, the translational-translational, rotational-translational and rotational-rotational elements of $\bm{\mathcal H}$ are given in terms of $\mu$m, $\mu$m$^2$ and $\mu$m$^3$, respectively.
We also evaluate the diffusion tensor $\bm{\mathcal D}$ for the particles at their initial orientations,
\begin{eqnarray}
\bm{\mathcal D} = k_BT \bm{\mu},
\end{eqnarray}
where $k_B$ is the Boltzmann constant, $T$ is the temperature and $\bm{\mu}$ is the mobility matrix ($\eta \bm{\mathcal H}$ is the inverse of $\bm{\mu}$).
Next, we use the elements of the diffusion matrix $\bm{\mathcal D}$ to determine the cross-correlation matrix ${\mathbf C}(t)$ of the time-dependent Brownian translational and orientational displacements of these particles,\cite{CEW2015JCP,CEW2016}
\begin{eqnarray}
{\mathbf C}(t) = \left[\!\! \begin{array}{c} \left\langle \Delta \mathbf{R}(t)\Delta \mathbf{R}
(t)\right\rangle_0 \; \left\langle \Delta \mathbf{R}%
(t)\Delta \mathbf{u}(t)\right\rangle_0\\
\left\langle \Delta \mathbf{u}(t)\Delta \mathbf{R}%
(t)\right\rangle_0 \; \,\left\langle \Delta \mathbf{u}(t)\,\Delta \mathbf{u}%
(t)\right\rangle_0
\end{array} \!\!\right].\label{C(t)}
\end{eqnarray}
with $\Delta \mathbf{R}$ and $\Delta\mathbf{u}(t)$ defined as in Refs. \cite{Kraft,CEW2015JCP,CEW2016},
\begin{eqnarray}
&&\Delta \mathbf{R}(t) = \mathbf{R}(t) - \mathbf{R}(0),\\
&&
\Delta \mathbf{u}(t)=\frac{1}{2}\sum_{p=1}^{3}\mathbf{u}%
^{(p)}(0)
\times \mathbf{u}^{(p)}(t),\label{defDeltau}
\end{eqnarray}
where $\mathbf{R}(t)$ denotes the time-dependent position of a reference center
and $\mathbf{u}%
^{(p)}(t)$, $p=1,2,3$, are
three mutually perpendicular unit vectors which describe the particle orientation at time $t$ \cite{CEW2015JCP,Kraft}.
The averages $\langle ... \rangle_0$ are taken with respect to the particle positions and orientations, using the conditional probability which satisfies the Smoluchowski equation \cite{Kampen,SE2}.
To determine the hydrodynamic friction matrix $\eta \bm{\mathcal H}$ and the diffusion tensor $\bm{\mathcal D}$
we solve the Stokes equations by the multipole method, using the accurate numerical codes {\sc Hydromultipole} with the lubrication correction \cite{CEW}.
We apply the multipole truncation order $L\!=\!20$.
Then, we apply the expressions for the cross-correlations derived from the Smoluchowski equation in Refs. \cite{CEW2015JCP,CEW2016}.
For spheres or some other symmetric particle shapes,
the mobility and friction matrices are diagonal. Therefore, the mobility center coincides with the center-of-mass, we are in the frame in which the rotational-rotational diffusion tensor is diagonal, and we can directly use the
simple analytical expressions
derived
in Ref.~\cite{CEW2015JCP}. For irregular shapes, we first rotate the system of coordinates to the reference frame in which the rotational-rotational diffusion matrix is diagonal. Still, the translational-rotational coupling does not vanish, and therefore, we use more complicated analytical expressions for the cross-correlations from Ref.~\cite{CEW2016}. To compare with the experiments, we rotate back the frame of reference.
\vspace{-0.3cm}
\section
Calculations: case 1}\label{III}
\subsection{Particles}\label{particles}
Following Ref.~\cite{Kraft}, we consider three particles: regular trimer, regular tetramer and irregular trimer, made of spheres (labeled by $i\!=\!1,2,3,4$).
For the
regular trimer and regular tetramer, the beads have equal diameters $d$ and overlap, with equal distances $l$ between the closest bead centers, with $d\!=\!2.1 \mu$m and $l\!=\!1.5\mu$m for the trimer and
$d\!=\!2.4 \mu$m and
$l\!=\!2.3\mu$m for the tetramer.
For the irregular trimer, the beads have diameters
$d_1\!=\!2.1\mu$m, $d_2\!=\!1.3\mu$m, $d_3\!=\!1.7\mu$m, they do not overlap, and the distances between the bead centers are $l_{13}\!=\!2.2\mu$m, $l_{12}\!=\!2.2\mu$m, $l_{23}\!=\!1.7\mu$m.
The particles at their initial orientations with respect to the chosen coordinate system are shown in Fig.~\ref{confi}; the centers of three beads are in the plane $x_1=0$.
From now on we will choose the center of mass position as the reference center position $\mathbf{R}(t)$, and will use the same notation as in Ref.~\cite{Kraft}, to allow for the comparison with the experiments.
\begin{figure}[h!] \psfrag{y}{\small $\;x_2$} \psfrag{z}{\small $x_3$}
\includegraphics[height=2.67cm]{reg_trimer.eps} \includegraphics[height=2.67cm]{reg_tetramer.eps}
\psfrag{y}{\small $\;\;\;x_2$ } \psfrag{z}{\small $x_3$} \includegraphics[height=2.67cm]{irr.eps}\vspace{-0.4cm}
\caption{Bead models of different rigid particles. For irregular trimer, the center of mass (yellow triangle) does not coincide with the mobility center (cyan circle).}\label{confi} \end{figure}%
\vspace{-0.5cm}
\subsection{Regular trimer}
\begin{figure*}
\begin{center}
REGULAR TRIMER\\
\includegraphics[height=3.6cm]{figure2_a.eps} \!\!\!\hspace{0.835cm}REGULAR TETRAMER\hspace{0.835cm}
\!\!\! \includegraphics[height=3.6cm]{figure2_b.eps}\\
\includegraphics[height=3.6cm]{figure2_c.eps} \!\!\!\hspace{1.03cm}IRREGULAR TRIMER\hspace{1.03cm
\!\!\!\includegraphics[height=3.6cm]{figure2_d.eps} \\
\vspace{0.3cm}
\! \hspace{-0.1cm}\includegraphics[height=3.6cm]{figure2_e.eps} \!\!\!\!\!
\includegraphics[height=3.6cm]{figure2_f.eps}\!\!\!
\includegraphics[height=3.6cm]{figure2_g.eps}
\end{center}
\vspace{-0.6cm}
\caption{Cross-correlations $C_{ij}(t)$ for regular trimer (top row), regular tetramer (middle row) and irregular trimer (bottom row). Translational-translational, translational-rotational and rotational-rotational couplings are shown in the left, middle and right columns, respectively. Horizontal line:
$C_{44}(\infty)=C_{55}(\infty)=C_{66}(\infty)\!=\!1/6.$}\label{panels}
\end{figure*}
\vspace{-0.4cm}
For the regular trimer at the chosen orientation shown in Fig.~\ref{confi}, the friction matrix has the form,
\begin{eqnarray}
\bm{\mathcal H} = \left( \begin{array}{cccccc}
30.5 & 0 & 0 & 0 & 0 & 0 \\
0 & 28.4 & 0 & 0 & 0 & 0\\
0 & 0 & 28.4 & 0 & 0 & 0\\
0 & 0 & 0 & 112 & 0 & 0\\
0 & 0 & 0 & 0 & 89.3 & 0 \\
0 & 0 & 0 & 0 & 0 & 89.3
\end{array} \right),\label{rtrimer}
\end{eqnarray}
with the units $\mu$m and $\mu$m$^3$, respectively.
The cross-correlations are given by the same
expressions as for axially symmetric shapes~\cite{CEW2015JCP},
\begin{eqnarray} \hspace{-0.4cm}
C_{11}\!&\!\!\!\!=\!\!\!& 0.126\,t/s-0.0481(1\!-\!e^{-0.123\,t/s}),\label{rtr11}\\
C_{22}\!&\!\!\!\!\!=\!\!\!\!&C_{33}=\! 0.126\,t/s+0.0240(1\!-\!e^{-0.123\,t/s}),\;\;\;\;\\
C_{44}\!&\!\!\!=\!\!\!& \!\! \frac{1}{6}\!+\!\frac{1}{12}e^{-0.123\,t/s}\!\!-\!\frac{1}{2}e^{-0.106\,t/s}\!\!+\!\frac{1}{4}e^{-0.0410\,t/s}\!, \nonumber \\
\\
C_{55}\!&\!\!\!=\!\!\!&C_{66}= \frac{1}{6}-\frac{1}{6}e^{-0.123\,t/s}
\!-\!\frac{1}{4}e^{-0.119\,t/s} \nonumber\\
&\!+\!&\frac{1}{4}e^{-0.0367\,t/s}.\label{rtr55}
\end{eqnarray}
The expressions \eqref{rtr11}-\eqref{rtr55} are plotted versus time in the top row of Fig. \ref{panels}.
The off-diagonal components vanish.
\vspace{-0.5cm}
\subsection{Regular tetramer}
For the regular tetramer at the chosen orientation shown in Fig.~\ref{confi}, and the units $\mu$m and $\mu$m$^3$, respectively,
the friction matrix has the form,
\begin{eqnarray}
\bm{\mathcal H} = \left( \begin{array}{cccccc}
39.7 & 0 & 0 & 0 & 0 & 0 \\
0 & 39.7 & 0 & 0 & 0 & 0\\
0 & 0 & 39.7 & 0 & 0 & 0\\
0 & 0 & 0 & 248 & 0 & 0\\
0 & 0 & 0 & 0& 248 & 0 \\
0 & 0 & 0 & 0& 0 & 248
\end{array} \right).\label{rtetramer}
\end{eqnarray}
\newpage
The cross-correlations are given by the same
expressions as for a spherical particle \cite{CEW2015JCP},
\begin{eqnarray}
C_{11}&\!\!=\!\!&C_{22}=C_{33}= 0.0920\, t/s,\label{rte11}\\
C_{44}&\!\!=\!\!&C_{55}=C_{66}=\nonumber\\
\frac{1}{6}&\!-\!&\frac{5}{12}e^{-0.0443\,t/s}+\frac{1}{4}e^{- 0.0148\,t/s}.\;\;\label{rte55}
\end{eqnarray}
The expressions \eqref{rte11}-\eqref{rte55} are plotted versus time in the middle row of Fig. \ref{panels}.
The off-diagonal components vanish.
\vspace{-0.3cm}
\subsection{Irregular trimer}
For the irregular trimer at the chosen orientation shown in Fig.~\ref{confi}, and the units $\mu$m and $\mu$m$^3$, respectively, the friction matrix has the form,
\begin{eqnarray}
\bm{\mathcal H}\! = \!\!\left( \!\!\begin{array}{cccccc}
29.9 & 0 & 0 & 0 & -4.06 & -2.95\\
0 & 28.0 & 0.318 & 2.92 & 0 & 0\\
0 & 0.318 & 26.6 & 2.07 & 0 & 0\\
0 &2.92& 2.07 & 117 & 0 & 0\\
-4.06 & 0 & 0 & 0 & 101 & 8.51\\
-2.95 & 0 & 0 & 0 & 8.51 & 69.3
\end{array} \!\!\right)\!\!.\;\;\label{itrimer}
\end{eqnarray}
To determine the cross-correlations, we first go to the frame of reference where the rotational-rotational mobility in diagonal, and evaluate the correlation matrix $\mathbf{C}^{diag}$ in this frame, using the explicit expressions from Ref. \cite{CEW2016}. Then,
with the use of the $3\!\times \!3$ transformation matrix ${\bm T}$,
we transform it to the original frame of reference,
\begin{eqnarray}
C_{n+i,m+l}=T^{-1}_{ij}C^{diag}_{n+j,m+k}T_{lk},\label{transf}
\end{eqnarray}
In Eq.~\eqref{transf}, $i,j,k,l\!\!\!=\!\!\!1,2,3$ are the Cartesian components,
and $n,m\!\!\!=\!\!\!0,3$ label the translational and rotational parts of the correlation matrix.
The expressions are
lengthy and therefore not explicitly written in this note. All the non-vanishing
translational-translational, translational-rotational and rotational-rotational correlations are plotted in the bottom row of Fig.~\ref{panels}.
\subsection{Discussion}
In general, the numerical friction tensors and cross-correlations of the Brownian displacements from Ref. \cite{Kraft} agree well with our results presented in this Comment.
No wonder, sine they are obtained for the same sizes and relative positions of the spherical beads which model the particle shape.
The only meaningful differences are observed the rotational-translational couplings of the irregular trimer which are small and difficult to be determined numerically.
Comparing
the
corresponding elements of the hydrodynamic friction matrices in Eqs.~\eqref{rtrimer}, \eqref{rtetramer} and \eqref{itrimer}
with those given in Fig. 1 of Ref.~\cite{Kraft}, we need to take into account
different geometries and accuracies of the models. In the HYDROSUB algorithm and numerical program, used in Ref.~\cite{Kraft},
the surface of the particle is represented by a shell of small elements
(“minibeads”); the results are extrapolated to a zero minibead radius \cite{hydrosub}. In this work, each sphere of the cluster is represented by a single bead, and the accurate HYDROMULTIPOLE numerical codes based on a very precise multipole method corrected for lubrication
are used to evaluate the friction matrix elements \cite{CEW}.
\section{New method: case 2}\label{IV} \vspace{-0.4cm}
We will now apply our analytical expressions from Ref. \cite{CEW2016} to analyze the experimental results for the irregular trimer given in Ref.~\cite{Kraft}. For this particle, the non-linear deviations present in the analytical expressions from Sec.~\ref{III} are very small,
and the theoretical translational-translational correlations
grow with time almost linearly. We will now estimate the corresponding self-diffusion constant $D_{cm}$ (see Eq. (20) in Ref.~\cite{CEW2016}) which characterizes the isotropic mean square displacement at large times. Using the theoretical and experimental friction matrices, we obtain $D_{cm}=0.065 \mu m^2/s$ and $0.073 \mu m^2/s$, respectively. However, the experimental results shown in Fig. 2 in Ref. \cite{Kraft} correspond to value of $D_{cm}$ which is around two times smaller. There is a clear mismatch between the values which characterize essential feature of the Brownian motion: the value deduced from the measured time-dependent translational-translational correlations in a wide range of times and the value determined from the initial slopes of these functions. This difference can be understood taking into account large statistical uncertainty of the experimental results. However, there is no doubt that a method based on fitting in the whole range of times is more accurate than the method based on the initial slope.
Therefore, we propose the following new method to determine the self-diffusion matrix $\bm{\mathcal D}$ from the measured time-dependent cross-correlation matrix $\bf{C}(t)$. First, we determine (and go to) the frame of reference in which the matrix $\,\left\langle \Delta \mathbf{u}(t)\,\Delta \mathbf{u}%
(t)\right\rangle_0$ is diagonal. Then, we use the time-dependent analytic expressions from Ref. \cite{CEW2016} to determine in this frame $D_1,\;D_2,\;D_3$, and the rest of the self-diffusion coefficients. Finally, we can transform $\bm{\mathcal D}$ back to the original frame of reference. This procedure can be applied to
particles of arbitrary shapes.
{\small M.L.E.-J. was supported in part by the Polish National Science Centre (Narodowe Centrum Nauki) under grant No. 2014/15/B/ST8/04359. }
|
\section*{Introduction}
Elections are an important way to make decisions, both in human
and electronic settings. Arguably the most important class
of election systems are the scoring rules. A scoring rule is defined
by, for each number $m$ of candidates, a scoring
vector $\alpha_1 \geq \alpha_2 \geq \cdots \geq \alpha_m$.
In addition, we typically want these vectors to be somehow similar.
This is captured nicely by the notion of pure scoring rules
from~\cite{bet-dor:j:possible-winner-dichotomy} where the
length-($m+1$) vector is obtained by adding a coefficient in
the length-$m$ vector.
Voters have complete tie-free preferences over the candidates, and
a candidate ranked $i$th by a voter receives a score of $\alpha_i$
from that voter. The winners are the candidates with the highest
score.
We are interested in determining, for all scoring rules at once,
which of them give rise to easy computational problems and which
of them lead to hard problems. Theorems of that form are
known as dichotomy theorems.
For weighted scoring rules, in which each voter has a weight $w$ and
counts as $w$ regular voters,
there are dichotomy theorems for
all standard manipulative actions: manipulation~\cite{hem-hem:j:dichotomy},
bribery~\cite{fal-hem-hem:j:bribery}, and
control~\cite{DBLP:journals/jair/FaliszewskiHH15}.
The arguably more natural unweighted case is much harder to analyze
(since in the unweighted case we can only get hardness when the number of
candidates is unbounded, whereas in the weighted case hardness already
occurs with a fixed number of candidates; since weighted dichotomy theorems
typically look at a fixed number of candidates,
the results for the unweighted cases do not at all follow from the
results for the weighted cases).
Despite the prevalence of scoring
rules, there are only two dichotomy theorems for the unweighted case,
namely for the possible winner
problem~\cite{bet-dor:j:possible-winner-dichotomy,bau-rot:j:possible-winner-dichotomy-final-step}
and
for the constructive control by adding voters (CCAV)
problem~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}.
In this paper, we look at bribery and manipulation for unweighted
scoring rules, and, since bribery can be viewed as deleting voters
followed by a manipulation, we also look at
the constructive control by deleting voters (CCDV) problem.
For manipulation, we show that for all scoring rules
with a constant number of different coefficients, manipulation is in P.
This subsumes all known polynomial-time results for unweighted manipulation
for scoring rules. We conjecture that there is no dichotomy theorem for manipulation.
For bribery and CCDV, we obtain a dichotomy theorem for
pure scoring rules. In particular, we show exactly when these
problems are easy (in P) and that they are hard (NP-complete)
in all other cases. Interestingly,
our characterization is the ``dual'' of the CCAV characterization in the following sense:
For every scoring rule $f$,
the complexity of $f$-CCDV (and of $f$-bribery)
is the same as for $\dual{f}$-CCAV, where
$\dual{f}$ is obtained from $f$ by multiplying each entry in a scoring
vector by $-1$, and reversing the order of the vector.
These results are quite surprising:
CCDV has less structure to encode hard problems into it than CCAV, but we still obtain the same complexity characterization (modulo duality). On the other hand, bribery can be seen as the combination of CCDV and manipulation, but the complexity is the same as for CCDV.
However, in another sense bribery behaves very differently from CCDV: The complexity of bribery changes from polynomial-time solvable to NP-complete by small changes in the definition of the problem, while the complexity of the former is much more robust.
The structure of the paper is as follows: In Section~\ref{sect:prelim}, we introduce relevant definitions, including the specific problems we study in this paper. In Section~\ref{sect:results:manipulation}, we state our results on manipulation. Section~\ref{sect:ccdv bribery dichotomy} contains our dichotomy result for CCDV and bribery. Our individual complexity results for CDDV and bribery can be found in Sections~\ref{sect:results:ccdv} and~\ref{sect:bribery}, respectively. We conclude with open questions in Section~\ref{sect:open questions}.
All proofs not contained in the main paper can be found in the appendix.
\section{Preliminaries}\label{sect:prelim}
An \emph{election} consists of a non-empty, finite set of candidates and a finite set of voters. Each voter is identified with her vote, which is simply a linear order on the set of candidates. An \emph{election system} or \emph{voting rule} is a rule that, given an election, determines the set of candidates who are winners of the election according to this rule. A \emph{scoring vector} for $m$ candidates is simply a vector $(\alpha_1,\dots,\alpha_m)$ of integer coefficients, where $\alpha_i\ge\alpha_{i+1}$ for all $1 \leq i < m$. Such a vector defines a voting rule for elections with $m$ candidates by simply awarding, for each vote in the election, $\alpha_i$ points to the candidate ranked in the $i$-th position of this vote, and defining the candidates with the most points to be the winners of the election. A \emph{scoring rule} is an election system that for each number of candidates applies an appropriate scoring vector. Such a system can be described by a \emph{generator}, which is a function $f$ such that for each $m\in\mathbb N$, $f(m)$ is a scoring vector for $m$
candidates. Well-known scoring rules are Borda (using $f(m)=(m-1,m-2,\dots,1,0)$), $k$-approval (using $f(m)=(\underbrace{1,\dots,1}_{k\mathtext{ many}},0,\dots,0)$) and $k$-veto (using $f(m)=(1,\dots,1,\underbrace{0,\dots,0}_{k\mathtext{ many}})$) for natural numbers $k$. For readability, we usually identify a generator with the election system it defines.
To capture that the elections for different numbers of candidates should use ``similar'' generators, we use the following notion \cite{bet-dor:j:possible-winner-dichotomy}: A generator $f$ as above is \emph{pure}, if for all $m\ge1$, the vector $f(m)$ can be obtained from the vector $f(m+1)$ by removing one coefficient from the sequence. In~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}, it is shown that the restriction to the set of pure generators with rational numbers covers all generators in a large and reasonable class.
We define standard manipulative actions:
Manipulation~\cite{bar-tov-tri:j:manipulating,con-lan-san:j:when-hard-to-manipulate},
bribery~\cite{fal-hem-hem:j:bribery}, and
control~\cite{bar-tov-tri:j:control},
for generators.
\begin{restatableDefinition}
For a generator $f$, the \emph{constructive control problem for $f$ by deleting voters}, $f$-CCDV, is the following problem: Given a set of voters $V$ over a set of candidates $C$, a
candidate $p\in C$ and a number $k$, is there a subset $V'\subseteq V$ with
$\card{V'}\leq k$ such that $p$ is a winner of the election if the
votes in $V - V'$ are evaluated using $f$?
\end{restatableDefinition}
A similar problem, the \emph{constructive control problem for $f$ by adding voters}, called $f$-CCAV, asks whether $p$ can be made a winner by adding to $V$ at most $k$ voters from a given set of so-called unregistered voters.
In the \emph{manipulation problem for $f$}, we are given a set $V$ of
nonmanipulative voters and a set of manipulators, and we
ask whether $p$ can be made a winner by setting the votes of the manipulators,
with no restriction on how these votes can be chosen.
Finally, the \emph{bribery problem} for $f$ asks whether $p$ can be made
a winner by replacing up to $k$ votes in $V$ with the same number of arbitrary votes.
Clearly, for every polynomial-time uniform generator $f$, the problems $f$-CCDV, $f$-bribery, and $f$-manipulation are in \complexityclassname{NP}.
Two scoring vectors $(\alpha_1,\dots,\alpha_m)$ and $(\beta_1,\dots,\beta_m)$ are \emph{equivalent} if they describe the same election system, i.e., if for any election, they lead to the same winner set. It is easy to see~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014} that this is the case if and only if there are numbers $\gamma>0$ and $\delta$ such that for each $i$, $\beta_i=\gamma\alpha_i+\delta$. We say that $f_1$ and $f_2$ are \emph{ultimately equivalent} if $f_1(m)$ and $f_2(m)$ are equivalent for all but finitely many $m$. It is easy to see that in this case, CCDV, bribery, and manipulation have the same complexity for $f_1$ and $f_2$.
For algorithms, we need the function $f$ is efficiently computable. A generator $f$ is polynomial-time uniform if $f(m)$ can be computed in polynomial time, given $m$ in unary. (Given $m$ in binary, polynomial time would not suffice to even write down a sequence of $m$ numbers.)
For the remainder of this paper, a generator is always a polynomial-time uniform pure generator with rational coefficients.
\section{Manipulation}\label{sect:results:manipulation}
Our main result on manipulation is the following: Every generator $f$ for which there is a fixed, finite upper bound on the number of coefficients that are used for any number of candidates has a polynomial-time solvable manipulation problem. We mention that this result also holds for generators that are not pure (but still are polynomial-time uniform). We note that the special cases where $f$ generates $k$-approval or $k$-veto were shown in~\cite{pro-ros-zuc:j:borda}.
\begin{restatable}{restatableTheorem}{theoremmanp}\label{theorem:man-p} Let $f$ be a generator such that there is a constant $c$ such that for each number $m$ of candidates, at most $c$ different coefficients appear in the vector $f(m)$. Then $f$-manipulation can be solved in polynomial time.
\end{restatable}
We give a proof sketch for a simple special case of the theorem, namely generators $f$ of the form $f=(0,\dots,0,-\beta,-\alpha)$.
With great care, this proof sketch generalizes to the general case.
\begin{proof}(Sketch)
Consider a preferred candidate $p$, a set of candidates $C=\set{c_1,\dots,c_m,p}$, a surplus $\surpl c$ for each $c\in C$ (i.e., the value $\score{c}-\score{p}$, which can easily be computed from the election instance, as $f$ is polynomial-time uniform), and a set of $k$ manipulators. Clearly, we can assume that
all manipulators will vote $p$ first.
The obvious greedy approach of having a manipulator rank a
candidate with the largest surplus last won't always work:
If $\beta = 2$, $\alpha = 3$, the surplus of $c_1$ is 4,
the surplus of $c_2$ and $c_3$ is 3, and we have two manipulators,
the only successful manipulation is to have the manipulators vote
$\cdots > c_1 > c_2$ and $\cdots > c_1 > c_3$ and so we cannot put
$c_1$ last.
If there is a successful manipulation, then for all $i\in\set{1,\dots,m}$, there are
numbers $x_i$ and $y_i$ (the number of times $c_i$ is ranked next to last / last
by a manipulator) such that:
\begin{enumerate}
\item $x_i + y_i \leq k$,
\item $\sum_{1 \leq i \leq m} x_i = k$,
\item $\sum_{1 \leq i \leq m} y_i = k$, and
\item $\surpl{c_i} - \beta x_i - \alpha y_i \leq 0$.
\end{enumerate}
We derive an
algorithm deciding the condition by dynamic programming. For this, we define the Boolean predicate $M$ such that
$M(k, k_\beta, k_\alpha, s_1, \ldots, s_\ell)$
is true if and only if for all $i$, $1 \leq i \leq \ell$,
there exist natural numbers $x_i$ and $y_i$ such that
\begin{enumerate}
\item $x_i + y_i \leq k$,
\item $\sum_{1 \leq i \leq \ell} x_i = k_\beta$,
\item $\sum_{1 \leq i \leq \ell} y_i = k_\alpha$, and
\item $s_i - \beta x_i - \alpha y_i \leq 0$.
\end{enumerate}
It is immediate that if there is a successful manipulation, then
$M(k, k, k, \surpl{c_1}, \ldots , \surpl{c_m})$ is true.
It is not so easy to see that the converse holds. This is shown
by induction on $k$.
It is easy to come up with a simple ad-hoc proof for the
simple case we are looking at here, but
we will instead
describe an approach that
generalizes to the general case.
The inductive step uses the following argument.
If $M(k+1, k_\beta, k_\alpha, s_1, \ldots, s_\ell)$ is true,
let $X = \{c_i \ | \ x_i > 0\}$ and let
$Y = \{c_i \ | \ y_i > 0\}$. Then the sequence $(X, Y)$ can be
shown to fulfill the ``marriage condition,'' which then,
by Hall's Theorem~\cite{hall},
implies that there
is a ``traversal,'' i.e., a sequence of distinct representatives of
this sequence of sets which then gives us a vote for one of the
manipulators. In this particular case, the traversal consists of
two distinct candidates $(c_i,c_j)$ such that $x_i > 0$ and $y_j > 0$.
Let one manipulator vote $\cdots > c_i > c_j$, subtract 1 from $x_i$ and
$y_j$, subtract $\beta$ from
$\surpl{c_i}$, and subtract $\alpha$ from from $\surpl{c_j}$.
It follows from the induction hypothesis that
the remaining $k$ manipulators can vote to make $p$ a winner.
To conclude the proof sketch, we can show by dynamic programming
that $M$ is computable in polynomial for unary $k, k_\alpha, k_\beta$
(which is sufficient to solve the manipulation problem).
This holds since:
\begin{enumerate}
\item
$M(k,k_\beta, k_\alpha)$ is true if and only if
$k_\beta = k_\alpha = 0$.
\item For $\ell \geq 1$,
$M(k,k_\beta, k_\alpha, s_1, \ldots, s_\ell)$ if and only if
there exist natural numbers $x_\ell$ and $y_\ell$ such that:
\begin{enumerate}
\item
$x_\ell + y_\ell \leq k$,
\item
$x_\ell \leq k_\beta$,
\item
$y_\ell \leq k_\alpha$,
\item
$s_\ell - \beta x_\ell - \alpha y_\ell \leq 0$, and
\item
$M(k,k_\beta - x_\ell, k_\alpha - y_\ell, s_1, \ldots, s_{\ell-1}).$
\end{enumerate}
\end{enumerate}
\end{proof}
Given Theorem~\ref{theorem:man-p} and the fact that manipulation for Borda is NP-complete~\cite{DBLP:conf/ijcai/BetzlerNW11,dav-kat-nar-wal:c:complexity-and-algorithms-for-borda},
it is natural to ask whether the manipulation problem is NP-complete for all remaining generators.
But this is extremely unlikely:
Though our approach does not give
polynomial-time algorithms when the number of coefficients is
unbounded, it will give quasipolynomial algorithms when the
coefficients are small enough and grow slowly enough.
It is also conceivable that additional cases will be in P.
Though a general greedy approach seems unlikely (as manipulation for Borda is NP-complete),
a greedy approach for specific cases is still possible.
We conjecture that there is no dichotomy theorem for manipulation
for pure scoring rules, with
different intermediate (between P and
NP-complete) complexities showing up.
\section{CCDV and Bribery Dichotomy}\label{sect:ccdv bribery dichotomy}
We completely characterize the complexity of $f$-CCDV and $f$-bribery for every generator $f$. For each generator $f$, these two problems are polynomial-time equivalent, and are polynomial-time solvable or NP-complete. For the cases where $f$ generates $k$-approval or $k$-veto, the complexity classification is already stated in~\cite{lin:thesis:elections}.
For CCDV, the special case where $f$ generates $1$-approval was shown
in~\cite{bar-tov-tri:j:control}.
For bribery, the
special cases where $f$ generates $1$-approval or $1$-veto were shown in~\cite{fal-hem-hem:j:bribery}.
We also note the existence of an unpublished manuscript that proves NP-hardness
for bribery for generators of the form $(\alpha, \beta, 0, \ldots, 0)$, where $\alpha$ and $\beta$ are coprimes with
$\alpha > \beta \geq 1$~\cite{car-kak-kar-woe:unpub:bribery}.
\begin{restatable}{restatableTheorem}{theoremccdvandbriberydichotomy}\label{theorem:ccdv dichotomy}
Let $f$ be a pure, polynomial-time uniform generator. If $f$ is ultimately equivalent to one of the following generators, then $f$-CCDV and $f$-bribery can be solved in polynomial time:
\begin{enumerate}
\item $f_1=(1,\dots, 1, 0, 0, 0)$ ($3$-veto),
\item $f_2=(1, 0, \dots,0)$ ($1$-approval),
\item $f_3=(1, 1, 0, \dots, 0)$ ($2$-approval),
\item for some $\alpha\ge\beta \ge 0$, $f_4=(0, \ldots, 0, -\beta, -\alpha)$ (this includes triviality, $1$-veto, and $2$-veto),
\item $f_5=(2,1,\dots,1,0)$.
\end{enumerate}
Otherwise, $f$-CCDV and $f$-bribery are NP-complete.
\end{restatable}
This dichotomy theorem mirrors the one obtained for CCAV in~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014} (modulo duality, see below). For the relationship of CCDV and CCAV, this implies that the difficulty of implementing ``setup votes'' (see below) does not have any influence on the complexity of our decision problems for the class of generators we study. The below proof of Theorem~\ref{theorem:0 dots 0 -gamma -beta -alpha CCDV} is an example of a non-trivial implementation of these setup votes.
In particular, our results imply that the complexity of CCDV is ``robust'' in the following sense: The complexity of CCDV does not depend on whether we add a bit to each voter stating whether she can be deleted or not. This will be made formal below in our discussion of CCDV$^*$\xspace (defined below).
The situation is different for bribery:
Generalizing the bribery problem to allow marking some voters as ``unbribable'' increases the complexity to NP-complete for some generators. As an example of this phenomenon, we state the following result. (The version of bribery defined here may be of independent interest, but is only used here to highlight the differences between CCDV and bribery.)
\begin{restatableTheorem}\label{theorem:bribery with unbribable voters np complete 0 dots 0 -1 -2}
The variation of the bribery problem for $(0, \dots, 0, -1, -2)$ where each voter has a bit that states whether this
voter can be bribed or not is NP-complete.
\end{restatableTheorem}
\section{Control by Deleting Voters}\label{sect:results:ccdv}
In this section we give an overview over our proof of CCDV-part of Theorem~\ref{theorem:ccdv dichotomy}. In Section~\ref{sect:ccav and ccdv}, we show that our CCDV polynomial-time cases easily follow from a relationship to CCAV, whose complexity was studied in~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}. Sections~\ref{sect:ccdv few coefficients general cases} and~\ref{sect:ccdv many coefficients} then contain our hardness results for CCDV.
\subsection{Relationship between CCAV and CCDV and CCDV polynomial time cases}\label{sect:ccav and ccdv}
CCDV and CCAV are closely related as follows: For a generator $f$, let $\dual f$ be the generator obtained from $f$ by multiplying each coefficient with $-1$ and reversing the order of the coefficients (to maintain monotonicity). Removing a vote in an $f$-election has the same effect as adding the same vote in a $\dual f$-election. Using this observation, one can show that $f$-CCDV reduces to $\dual f$-CCAV for all generators $f$. We mention that a similar relation holds
for weighted $k$-approval and $k$-veto
elections~\cite{DBLP:journals/jair/FaliszewskiHH15}.
The other direction of this relationship does not follow so easily, as there is an important difference between CCDV and CCAV: In CCAV, the set of voters is partitioned into a set of \emph{registered} voters and a set of \emph{potential} voters, where the controller's actions can only influence the potential voters. This provides the problem with additional structure, as the controller cannot modify the registered voters. We often call these registered votes, which the controller cannot influence anymore, \emph{setup votes}, as these allow us to set up the scenario of an NP-hard problem in hardness proofs.
The CCDV problem does not have a corresponding structure; here every vote may be (potentially) deleted by the controller. This makes it harder to construct the above-mentioned setup votes: Since we cannot simply ``forbid'' the controller to delete certain votes, hardness proofs for CCDV need to ``setup'' the relevant scenario with votes that are designed to be ``unattractive'' to delete for the controller.
To obtain CCDV hardness results, it is therefore natural to consider the following analog to the CCAV problem: CCDV$^*$\xspace is a version of CCDV providing the additional structure that CCAV has. In CCDV$^*$\xspace, the set of votes is partitioned into a set $R$ of voters that cannot be deleted, and voters $D$ that can be deleted. From the above discussion, it follows that the complexities of CCAV and CCDV$^*$\xspace are related with the following duality, which, together with the polynomial-time results obtained for CCAV in~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014} immediately implies the polynomial-time CCDV cases of Theorem~\ref{theorem:ccdv dichotomy}.
\begin{restatable}{restatableProposition}{propccdvstarpolyequivccav}\label{prop:ccdvstar poly equiv ccav}
For every generator $f$, $f$-CCDV$^*$\xspace and $\dual f$-CCAV are polynomially equivalent.
\end{restatable}
Proposition~\ref{prop:ccdvstar poly equiv ccav} is not completely trivial, since the reductions must convert election instances maintaining the relative points of the candidates. However, this is done using standard constructions.
From Theorem~\ref{theorem:ccdv dichotomy} and the results in~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}, it follows that
$f$-CCDV and $f$-CCDV$^*$\xspace always have the same complexity. In fact, our algorithms for CCDV and CCDV$^*$\xspace do not take the structure of the ``registered'' votes into account, but can work with scores for the candidates that do not come from any set of votes. For bribery, the situation is quite different, see the above Theorem~\ref{theorem:bribery with unbribable voters np complete 0 dots 0 -1 -2}.
All polynomial-time cases for CCDV follow from the above relationship in a straight-forward manner. This is not surprising, since CCDV in the above-discussed sense has less structure than CCAV, and thus easiness results for CCAV translate to (dual) CCDV. The interesting part of our dichotomy is the converse: If CCAV is NP-hard for some generator $f$, then CCDV is hard for $\dual f$ as well.
A natural approach for the proof of the dichotomy theorem, suggested by Proposition~\ref{prop:ccdvstar poly equiv ccav}, is to show that $f$-CCDV$^*$\xspace always reduces to $f$-CCDV. While this does in fact follow, proving a generic reduction from $f$-CCDV$^*$\xspace to $f$-CCDV for all generators $f$ seems to be difficult, due to the additional structure provided by CCDV$^*$\xspace.
Our proof of the CCDV part of Theorem~\ref{theorem:ccdv dichotomy} therefore uses a case distinction to obtain $f$-CCDV hardness for each remaining pure, polynomial-time uniform generator $f$.
We note that due to the relationship between CCDV and CCAV, all CCAV-hardness results in~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014} easily follow from the results obtained in the current paper. However, our proofs make use of the results and proofs from~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}.
\subsection{CCDV hardness: ``few coefficients''}\label{sect:ccdv few coefficients general cases}
We first consider generators with ``few'' different coefficients, i.e., generators of the form $f=(\alpha_1,\alpha_2,\alpha_3,\dots,\alpha_3,\alpha_4,\alpha_5,\alpha_6)$ for rationals $\alpha_1,\dots,\alpha_6$.\footnote{This only uniquely defines $f$ for elections with at least $6$ candidates, however $f$ is uniquely defined up to ultimate equivalence, which is sufficient for the complexity analysis.}
Using equivalence-preserving transformations, we can assume that all $\alpha_i$ are nonnegative integers, and that their greatest common divisor is $1$. Note that a generator of this form is trivially polynomial-time uniform.
\subsubsection{Reductions from CCDV$^*$\xspace}\label{sect:ccdv hardness:generic ccdvstar reductions}
A general reduction from $f$-CCDV to $f$-CCDV$^*$\xspace does not seem feasible, as discussed above. However, there are cases where hardness of $f$-CCDV$^*$\xspace leads to hardness of $f$-CCDV with a direct proof.
The following two results (Theorems~\ref{theorem:alpha 0 dots 0 alpha < beta ccdv} and~\ref{theorem:alpha 0 dots 0 beta < alpha ccdv}) are proven in this way.
\begin{restatable}{restatableTheorem}{theoremccdvhardnessalphazweodotsalphasmallerbeta}
\label{theorem:alpha 0 dots 0 alpha < beta ccdv}
Let $f=(\alpha,0,\dots,0,-\beta)$ be a generator with $1\leq\alpha<\beta$. Then $f$-CCDV is NP-complete.
\end{restatable}
\begin{proof}(Sketch)
From~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}, we know that $\dual f$-CCAV is NP-complete, Proposition~\ref{prop:ccdvstar poly equiv ccav}, then implies that $f$-CCDV$^*$\xspace is NP-complete as well. We show that $f$-CCDV$^*$\xspace reduces to $f$-CCDV. Given an instance of $f$-CCDV$^*$\xspace, we convert it into an equivalent instance of $f$-CCDV by replacing the undeletable votes $R$ with votes that
\begin{enumerate}
\item\label{enum label:setup votes need same relative points} result in the same relative points as the votes in $R$, and
\item are not deleted in a successful CCDV action.
\end{enumerate}
We denote a vote $c_1>c_2>\dots>c_{n-1}>c_n$ simply as $c_1>c_n$ (the rest is irrelevant). With some light preprocessing, we can assume that no deletable vote has $p$ in the first or last position, this allows us to compute the number $\score p$ of points that $p$ will have after the delete action. Similarly, we can assume that $\score p=N_p\alpha$ for some $N_p\ge2$.
Satisfying point~\ref{enum label:setup votes need same relative points} above boils down to add, for an arbitrary candidate $c\neq p$, votes that let $c$ gain $\alpha$ points against $p$, and which will not be deleted. This is done as follows: We add dummy candidates $d_1,\dots,d_\ell$ (for a suitably chosen number $\ell$) and a single vote $c>d_1$, letting $c$ gain $\alpha$ points relative to $p$. To ensure that the vote $c>d_1$ cannot be removed, we add votes setting up the scores as follows:
\begin{itemize}
\item Each $d_i$ for $1\leq i\leq\ell-1$ ties with $p$,
\item the only way to make $d_i$ lose points (relative to $p$) is to remove votes $d_i>d_{i+1}$, which then lets $d_{i+1}$ gain points (relative to $p$).
\end{itemize}
Hence removing the vote $c>d_1$, which lets $d_1$ gain $\beta$ points relative to $p$ requires the controller to remove votes of the form $d_1>d_2$, which each lets $d_2$ gain $\beta$ points. This process continues for $d_i$ with $i\ge 2$. Thus, removing $c>d_1$ triggers a ``chain'' of additional removals---more than the budget allows. The numbers of votes needed to setup grows exponentially in the number of steps. However, since the controller can only remove a polynomial number of votes, we only require logarithmically many steps, yielding a polynomial construction.
Constructing the actual set of votes that results in the above scores and satisfies the two points above is nontrivial, the construction is in fact the main technical difficulty in the proof.
\end{proof}
The following result is shown similarly, the difference is that instead of logarithmically many steps of an exponentially growing construction, here we apply a simpler linear process.
\begin{restatable}{restatableTheorem}{theoremalphazerodotszerobetasmalledalphaccdv}\label{theorem:alpha 0 dots 0 beta < alpha ccdv}
Let $f=(\alpha,0,\dots,0,-\beta)$ be a generator with $\alpha>\beta\ge1$. Then $f$-CCDV is NP-complete.
\end{restatable}
\subsubsection{Reductions by inspection of the CCAV reduction}\label{sect:ccdv hardness by inspecting ccav reduction}
Similarly to the preceding Section~\ref{sect:ccdv hardness:generic ccdvstar reductions}, the results in this section are proved by a reduction from $f$-CCDV$^*$\xspace to $f$-CCDV. However, while the reductions above were ``generic'' (reducing from an arbitrary CCDV$^*$\xspace-instance), we now start with instances of $f$-CCDV$^*$\xspace produced by the hardness proof of $f$-CCDV$^*$\xspace. Therefore, we do not need to construct ``setup votes'' that implement any possible given set of scores, but only need to achieve exactly the points used in the hardness proof of $\dual f$-CCAV in~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}.
In the following theorem, this is a significant advantage, as here the ``setup votes'' grant more points to the preferred candidate than to the remaining candidates. Therefore, it is easy to construct these votes in such a way that the controller has no incentive to delete them.
The proof of the theorem also illustrates the relationship between hardness results for CCAV (as obtained in~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}), and the hardness results for CCDV and bribery we obtain in the current paper: The proof of Theorem~\ref{theorem:0 dots 0 -gamma -beta -alpha CCDV} below uses the reduction of the corresponding hardness result for CCAV in~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014} as a starting point, but is technically more involved. We will later re-use parts of the following construction to obtain the corresponding hardness result for bribery as well, in the later Theorem~\ref{theorem:0 dots 0 -gamma -beta -alpha bribery}.
\begin{restatable}{restatableTheorem}{theoremzerodotszerominusgammaminusbetaminusalphaccdv}\label{theorem:0 dots 0 -gamma -beta -alpha CCDV}
Let $f=(\alpha_3,\dots,\alpha_3,\alpha_4,\alpha_5,\alpha_6)$ with $\alpha_3>\alpha_4>\alpha_6$. Then $f$-CCDV is NP-complete.
\end{restatable}
\begin{proof}
For the proof, we equivalently write $f$ as $f=(0,\dots,0,-\gamma,-\beta,-\alpha)$ with $0<\gamma<\alpha$.
Then, $\dual f=(\alpha,\beta,\gamma,0,\dots,0)$. From~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}, it follows that $\dual f$-CCAV is NP-complete, and hence, due to Proposition~\ref{prop:ccdvstar poly equiv ccav}, it suffices to show that $f$-CCDV$^*$\xspace reduces to $f$-CCDV.
Therefore, let an $f$-CCDV$^*$\xspace instance with undeletable votes $R$, deletable votes $D$, preferred candidate $p$, and budget $\ell$ be given. From the proof of Proposition~\ref{prop:ccdvstar poly equiv ccav}, we can assume that this instance is obtained from the hardness proof of $\dual f$-CCAV as follows:
\begin{itemize}
\item the votes in $D$ are exactly the votes available for addition in the CCAV instance, with the order of candidates reversed,
\item the relative points gained by the candidates from the votes in $R\cup D$ are the same as the points of the candidates in the CCAV instance (before the addition of votes by the controller).
\end{itemize}
The hardness proof of $\dual f$-CCAV uses a reduction from 3DM.
3DM is the following problem: Given a multiset
$M\subseteq X\times Y\times Z$ with $X$, $Y$ and $Z$ pairwise disjoint sets of equal size such that each $s\in X\cup Y\cup Z$ appears in exactly $3$ tuples of $M$, decide whether there is a set $C\subseteq M$ with $\card C=\card X$ such that each $s\in X\cup Y\cup Z$ appears in some tuple of $C$ (we also say that $C$ \emph{covers} $s$).
From the problem definition, it follows that $\card M=3\card X$.
The condition that each $s$ appears in exactly $3$ tuples is not standard;
we prove in the appendix
remains NP-complete.
Let $M\subseteq X\times Y\times Z$ be an instance of 3DM. Following the notation used in~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}, we set $n=\card M$ and $k=\card X$. Since $\card M=3\card X$ in every 3DM instance, it follows that $n=3k$. The hardness proof of $\dual f$-CCAV, translated to the CCDV setting (i.e., we present the votes as in the CCDV instance---as reversals of votes from the CCAV instance), constructs the following situation:
\begin{itemize}
\item the candidate set is $\set p\cup X\cup Y\cup Z\cup\set{S_i,S_i'\ensuremath{\ \vert\ } S_i\in M}$,
\item for each $S_i=(x,y,z)\in M$, the following votes are available for deletion (we only list the candidates gaining non-zero points from the vote):
\begin{itemize}
\item $\dots > S_i > p > x$
\item $\dots > S_i > p > y$
\item $\dots > S_i' > p > z$
\item $\dots > S_i' > p > S_i$
\end{itemize}
\item the relative scores resulting from the registered voters of the CCAV instance (i.e., the undeletable voters of the CCDV$^*$\xspace instance) are as follows:
\begin{itemize}
\item $\scorefinal p=\alpha+2\gamma$,
\item $\scorefinal c=(n+2k)\beta+2\gamma$ for each $c\in X\cup Y\cup Z$,
\item $\scorefinal{S_i}=(n+2k)\beta+\min(\alpha,2\gamma)$,
\item $\scorefinal{S_i'}=(n+2k)\beta+\alpha+\gamma$.
\end{itemize}
\end{itemize}
From the above votes introduced for the 3DM-elements, the candidates gain the following initial points---recall that each $c\in X\cup Y\cup Z$ appears in exactly $3$ triples from $M$, and, since $\card M=n=3k$, and there are $4$ votes introduced for every element in $M$, there are exactly $12k$ deletable votes introduced above:
\begin{itemize}
\item $\scorethreedm{p}=-12k\beta$, since $p$ gains $-\beta$ points in each of the $12k$ votes,
\item $\scorethreedm{S_i}=-2\gamma-\alpha$, since $S_i$ gains $0$ points in all of the votes introduced for other elements $S_j\neq S_i\in M$, and gains $-\gamma$ points in $2$ of the votes above, and $-\alpha$ points in one of the vote,
\item $\scorethreedm{S_i'}=-2\gamma$ analogously,
\item $\scorethreedm{c}=-3\alpha$ for each $c\in X\cup Y\cup Z$, since each $c$ appears in exactly $3$ triples $S_i$ from $M$.
\end{itemize}
For each candidate $x$, the undeletable votes from the CCDV$^*$\xspace instance thus let her gain exactly $\scorefinal x-\scorethreedm x$ points (modulo an offset added to the points of all candidates, since the CCAV reduction relies on an implementation lemma
that only fixes the relative points of each candidate). The scores that the undeletable votes implement are therefore as follows (recall that $n=3k$):
\medskip
\begin{tabular}{ll}
candidate $x$ & $\scorefinal x-\scorethreedm x$ \\ \hline
$p$ & $12k\beta+\alpha+2\gamma$ \\
$c\in X\cup Y\cup Z$ & $5k\beta+3\alpha+2\gamma$ \\
$S_i$ & $5k\beta+\alpha+2\gamma+\min(\alpha,2\gamma)$ \\
$S_i'$ & $5k\beta+\alpha+3\gamma$
\end{tabular}
\medskip
Since we only need to implement the relative scores among the candidates, it is enough to consider the points the candidates have to gain/lose relative to $p$. These are as follows (clearly, $p$ does not gain or lose any points relative to herself):
\medskip
\begin{tabular}{lll}
candidate $x$ & points $x$ needs to lose (result) \\ \hline
$c\in X\cup Y\cup Z$ & $7k\beta-2\alpha$ \\
$S_i$ & $7k\beta-\min(\alpha,2\gamma)$ \\
$S_i'$ & $7k\beta-\gamma$
\end{tabular}
\medskip
To achieve this, we first introduce three dummy candidates $d_1$, $d_2$, and $d_3$ (by placing them in the $0$-point segment of all present votes), and add sufficiently many votes voting all relevant candidates ahead of the dummy candidates. This lets the dummies lose $\alpha$, $\beta$, or $\gamma$ points relatively to the other candidates; we do this often enough to ensure that the dummy candidates cannot win the election. Using these dummies, we can easily let a relevant candidate $c\neq p$ lose $\delta\in\set{\alpha,\beta,\gamma}$ points, by using a vote placing $c$ in the position worth $-\delta$ points, filling the other two positions out of the last three with dummy candidates, and voting the remaining candidates (including $p$) in the positions gaining $0$ points. Such a vote will never be deleted by the controller, since it has $p$ in the first position.
If $1=\delta\in\set{\alpha,\beta,\gamma}$, then the required amount of points can easily be achieved by repeatedly losing $\delta=1$ points as described above. Hence assume that, in particular, $\beta\ge2$. For the candidate $c$, we proceed as follows:
\begin{itemize}
\item We add $\beta-2$ many votes letting $c$ lose $\alpha$ point each.
\item After this, $c$, still needs to lose $(7k-\alpha)\beta$ points, which we can achieve by using $7k-\alpha$ many votes letting $c$ lose $\beta$ points as described above (we can without loss of generality assume that $7k\ge\alpha$).
\end{itemize}
For candidates $S_i$ and $S_i'$, we proceed analogously.
It hence follows that $p$ can be made a winner by removing at most $\ell$ votes in the CCDV$^*$\xspace instance if and only if this is possible in the constructed $f$-CCDV instance.
\end{proof}
The following result uses a similar, but technically more involved argument.
\begin{restatable}{restatableTheorem}{theoremzerominusalphafiveminusalphaonedotsminusalphaoneccdvhardness}
\label{theorem:0 minus alpha5 minus alpha1 dots minus alpha1 CCDV hardness}
Let $f=(\alpha_1,\alpha_2,\alpha_3,\dots,\alpha_3)$ with $\alpha_1>\alpha_2>\alpha_3$. Then $f$-CCDV is NP-complete.
\end{restatable}
\subsubsection{Direct Proofs}\label{sect:ccdv hardness by direct 3DM reduction}
The remaining cases are shown with a direct proof (reducing from a variant of three dimensional matching, 3DM); they are in part similar to the hardness proofs of CCAV in~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}.
\begin{restatable}{restatableTheorem}{theoremcollectionoffewcoefficientsCCDVhardnesscases}\label{theorem:collection of few coefficients CCDV hardness cases}
$f$-CCDV is NP-complete in the following cases:
\begin{enumerate}
\item\label{few coefficients CCDV:ccdv hardness alpha1 alpha2 0 dots 0 minus alpha4 minus alpha5 minus alpha6} $f=(\alpha_1,\alpha_2,\alpha_3,\dots,\alpha_3,\alpha_4,\alpha_5,\alpha_6)$ with $\alpha_1>\alpha_3>\alpha_5$.
\item\label{few coefficients CCDV:ccdv alpha1alpha2>alpha3>alpha6}$f=(\alpha_1,\alpha_2,\alpha_3,\dots,\alpha_3,\alpha_6)$ with $\alpha_1,\alpha_2>\alpha_3>\alpha_6$.
\item\label{few coefficients CCDV:ccdv hardness alpha1 > alpha 3 > alpha 4} $f=(\alpha_1,\alpha_2,\alpha_3,\dots,\alpha_3,\alpha_4,\alpha_5,\alpha_6)$ with $\alpha_1>\alpha_3>\alpha_4$.
\end{enumerate}
\end{restatable}
\subsection{CCDV hardness: ``Many'' coefficients}\label{sect:ccdv many coefficients}
In Section~\ref{sect:ccdv few coefficients general cases}, we have covered all generators of the form $f=(\alpha_1,\alpha_2,\alpha_3,\dots,\alpha_3,\alpha_4,\alpha_5,\alpha_6)$. Generators \emph{not} of this form must satisfy $\alpha^m_3>\alpha^m_{m-3}$ for some value $m$. We now prove that the CCDV problem is NP-hard for all generators satisfying this condition. The proof is similar to the corresponding result in~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}.
An important observation is that when $\alpha^m_3>\alpha^m_{m-3}$ holds for some $m$, then purity of our generators implies that the condition also holds for any $m'\ge m$. Clearly, if $m\ge6$ is a multiple of $3$ and $\alpha^m_3>\alpha^m_{m-3}$, then one of $\alpha^m_3>\alpha^m_{\frac23m}$ and $\alpha^m_{\frac23m}>\alpha^m_{m-3}$ must hold. In the first case, the generator---for this number $m$ of candidates---behaves in a similar way as $3$-approval in the sense that there are three positions in the vote that let the candidates gain more points than a ``large'' number of positions later in the sequence of coefficients. In the latter case, the generator (for this number $m$) behaves similarly to $4$-veto, as there are four ``bad'' positions in the votes.
Therefore, a generator satisfying $\alpha^m_3>\alpha^m_{m-3}$ behaves, for every $m'\ge m$, similarly to $3$-approval or to $4$-veto. However, the behavior can be different for different values for $m$. Therefore, our proof of NP-completeness for generators satisfying this condition uses an ``adaptive'' reduction from the problem 3DM, which, given a 3DM instance, constructs either a reduction exploiting the ``$3$-approval likeness'' or the ``$4$-veto likeness'' of the generator, depending on the size of the instance (which linearly corresponds to the number of candidates in the constructed election instance). This gives the following result:
\begin{restatable}{restatableTheorem}{theoremmanydifferentcoefficients}
\label{theorem:many different coefficients}
Let $f$ be a polynomial-time pure generator such that $\alpha^m_3>\alpha^m_{m-3}$ for some $m$. Then $f$-CCAV is NP-hard.
\end{restatable}
\section{Bribery}\label{sect:bribery}
Bribery is closely related to both CCDV and manipulation:
Bribing $k$ voters can be viewed as deleting $k$ voters followed by
adding $k$ manipulation voters. We thus can use our results obtained
for CCDV and manipulation in Sections~\ref{sect:results:manipulation} and~\ref{sect:results:ccdv}
as a starting point to obtain a complexity classification of the bribery
problem.
However, it is not necessarily the
case that an optimal bribery consists of an optimal deletion followed by
an optimal manipulation (see Example~\ref{e:bribery}),
and so it is possible for bribery to be
hard while the manipulation and deletion of voters problems are easy.
However, we will show that for every pure scoring rule $f$,
$f$-Bribery is polynomial-time solvable if and only if $f$-CCDV is, though
the proofs for Bribery are more (and sometimes much more)
involved.
We also obtain an interesting relationship between the complexities of manipulation and bribery: From Theorem~\ref{theorem:man-p}, it follows that every ``few coefficients'' case leads to a polynomial-time solvable manipulation problem. From Theorem~\ref{theorem:ccdv dichotomy}, we know that, for CCDV, \emph{only} such cases can be solved in polynomial time (unless $\complexityclassname{P}=\complexityclassname{NP}$). Therefore, we obtain the following corollary:
\begin{restatableCorollary}
Let $f$ be a polynomial-time uniform pure generator. Then $f$-manipulation reduces to $f$-CCDV and $f$-manipulation reduces
to $f$-bribery.
\end{restatableCorollary}
\subsection{Bribery Polynomial-Time Cases}\label{sect:polynomial time bribery}
Our first bribery algorithm (for a generator equivalent to $(2,1,\dots,1,0)$)
uses a reduction to network flow.
\begin{restatable}{restatableTheorem}{theorembriberyonezeroesminusoneptime}\label{theorem:100-1 bribery in ptime}
$(1,0,\dots,0,-1)$-bribery is in polynomial time.
\end{restatable}
\begin{proof}(Sketch)
There are three types of voters. Let $V_1$ be the set of voters that rank $p$
last, let $V_2$ be the set of voters that rank $p$ neither first nor last,
and let $V_3$ be the set of voters that rank $p$ first.
In this case, we can assume without loss of generality that we bribe
as many $V_1$ voters as possible, followed by as many $V_2$ voters
as possible. We never have to bribe $V_3$ voters.
All bribed voters will put $p$ first, so we also know $p$'s score after
bribery.
The hardest case is the one where we bribe all $V_1$ voters and some
$V_2$ voters.
We view bribery as deletion followed by manipulation.
Delete all $V_1$ voters. In $V_2$,
deleting a voter $a > \dots > b$ corresponds to transferring a point
from $a$ to $b$. After deleting $k$ voters, the deleted voters will
be bribed to rank $p$ first and to rank some other candidate last.
After deleting $V_1$, for every candidate $c$,
$\score{c} = \scoresub{V_2\cup V_3}{c}$, i.e.,
the score of $c$ in $V_2 \cup V_3$.
For every $V_2$ voter $a > \dots > b$ that is deleted, transfer
one point from $a$ to $b$. For every bribe $p > \dots > d$,
delete a point from $d$. There are exactly $k$ bribes.
After bribery, $\score{p} = \card{V_3} + k$ and the score
of every other candidate needs to be at most
$\score{p} = \card{V_3} + k$.
It is not too hard to see that this problem can be translated in
min-cost network flow problem, in a similar,
though somewhat more complicated, way as in
the CCAV algorithm for the same generator
in~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}.
\end{proof}
We now state our second bribery result.
\begin{restatable}{restatableTheorem}{theorembriberyzeroesbetaalphainptime}\label{theorem:bribery zeroes -beta -alpha in ptime}
Let $\alpha \geq \beta \geq 0$. Then bribery for $(0, \ldots, 0, -\beta, -\alpha)$ is solvable in polynomial time.
\end{restatable}
\begin{proof}(Sketch)
As in the proof of
Theorem~\ref{theorem:100-1 bribery in ptime},
we partition $V$ into $V_1$, $V_2$, and $V_3$.
$V_1$ consists of all voters in $V$ that rank $p$ last,
$V_2$ consists of all voters in $V$ that rank $p$ second-to-last,
and $V_3$ consists of the remaining voters.
In the proof of Theorem~\ref{theorem:100-1 bribery in ptime},
it was important that we never had to bribe $V_3$ voters.
This is not always the case here, as
shown in the example
below. That also means that this case is very different from CCDV, since in
CCDV we never have to delete $V_3$ voters.
It also shows that an optimal bribery is not always an optimal
deletion followed by an optimal manipulation.
\begin{example}
\label{e:bribery}
This example shows that we sometimes need to bribe $V_3$ voters.
We will use the scoring rule $(0, \ldots, 0, -1, -3)$.
Let $C = \{p, a, b, c, d, e, f\}$ and let $V$ consist of
the following votes:
\begin{tabular}{ccc}
$\cdots > p > a$ &
$\cdots > e > f$ &
$\cdots > f > e$\\
$\cdots > p > b$ &
$\cdots > e > f$ &
$\cdots > f > e$\\
$\cdots > p > c$
\end{tabular}
Then $\score{p} = \score{a} = \score{b} = \score{c} = -3$,
$\score{d} = 0$, and
$\score{e} = \score{f} = -8$.
We can make $p$ a winner by bribing one of the $V_3$ voters
to vote $p > \cdots > d$. But
it is easy to see that we can not make $d$ a winner by bribing a $V_2$
voter, wlog, the voter voting $\cdots > p > a$, since
in the bribed election, the score of $p$ will be at most
$-2$, and so both $a$ and $d$ must be in the last position of
the bribed voter.
\end{example}
Though we may need to bribe $V_3$ voters, we can show that we
never need to bribe more than a constant number of $V_3$ voters.
This is crucial in obtaining a a polynomial-time bound on a
dynamic programming
approach similar to, but more complicated than, the one
in the proof
for Theorem~\ref{theorem:man-p}.
\end{proof}
Together with Lin's results on the complexity of bribery for $k$-veto and $k$-approval election systems in~\cite{lin:thesis:elections},
the above results prove all polynomial-time bribery cases of our main result, Theorem~\ref{theorem:ccdv dichotomy}. In the remainder of Section~\ref{sect:bribery}, we therefore discuss our NP-hardness results for bribery.
\subsection{Bribery Hardness Approach}\label{sect:bribery hardness from ccdv hardness}
Our dichotomy results imply that each generator $f$ for which CCDV is \complexityclassname{NP}-hard also has an \complexityclassname{NP}-hard bribery problem. Proving this via a generic reduction from CCDV to bribery does not seem to be easy: Even though bribery can be seen as CCDV followed by manipulation, solving a bribery instance is not the same as first finding an optimal deletion of votes and then performing an optimal manipulation. Therefore, hardness of $f$-bribery does not easily follow from hardness of $f$-CCDV. We briefly discuss a proof strategy to obtain a hardness proof for $f$-bribery from a hardness proof of $f$-CCDV.
A key difficulty in the construction of a bribery hardness proof is that the manipulation action of the controller allows her more freedom than her delete action: For the latter, the reduction controls the available votes, whereas the manipulation action can use arbitrary permutations of the candidates. However, we can always assume that the bribed voters will vote $p$ in the first position, and will place any ``dummy'' candidates in the positions following $p$ in their votes. This often allows us to compute the score of $p$ after the bribery action.
To limit the controller's freedom in the manipulation votes, we proceed as follows: We identify a sufficiently long subsequence of the coefficients that differ by only a ``small'' amount. (Such a sequence exists by definition for the generators treated in Section~\ref{sect:ccdv few coefficients general cases}, and can be found using a pigeon-hole argument for other generators.) We then set up the points such that the ``relevant'' candidates must be placed into this ``low-variation'' sequence of the vote. This is done using ``blocking'' candidates that must be placed in low-score positions, and dummy candidates occupying the high-score positions (except for the first position, in which the bribed voters always vote $p$). This ensures that moving a candidate inside the ``low-variation'' does not make a large difference, and allows the reduction to control the possible manipulation actions very tightly.
A second difficulty is that the controller is more powerful in bribery than in CCDV: In bribery, she can perform a manipulation action in addition to her delete action. Therefore, to make it ``hard'' for the controller to find an optimal bribery action, the scores in the election instance must be ``worse'' for $p$ than in the CCDV setting. This leads to setup votes that are more attractive to delete than in the CCDV case. Similarly as in CCDV, the main technical issue is to define setup votes that both obtain the required scores and still will not be deleted by the controller, however due to the reasons above this is even more difficult for bribery as for CCDV.
The ideas outlined above allow
us to prove the bribery hardness results of Theorem~\ref{theorem:ccdv dichotomy}.
As an example for our approach to bribery hardness, we prove the following theorem, which is the ``bribery version'' of Theorem~\ref{theorem:0 dots 0 -gamma -beta -alpha CCDV}. We therefore apply our recipe to the proof of Theorem~\ref{theorem:0 dots 0 -gamma -beta -alpha CCDV}, which is the CCDV hardness result for the same generator.
\begin{restatable}{restatableTheorem}{theoremzerodotszerominusgammaminusbetaminusalphabribery}\label{theorem:0 dots 0 -gamma -beta -alpha bribery}
Let $f=(\alpha_3,\dots,\alpha_3,\alpha_4,\alpha_5,\alpha_6)$ with $\alpha_3>\alpha_4>\alpha_6$. Then $f$-bribery is NP-complete.
\end{restatable}
\begin{proof}
We follow the above recipe to obtain a hardness proof for $f$-bribery. Applying the recipe requires to make some changes to the construction of the CCDV hardness proof. As in the proof of Theorem~\ref{theorem:0 dots 0 -gamma -beta -alpha CCDV}, we write $f$ as $f=(0,\dots,0,-\gamma,-\beta,-\alpha)$. In particular, since we can assume that all bribed voters will vote $p$ first, the score of $p$ will not change from manipulation votes (but may, of course, change from the deleted votes).
The budget for the controller is, as in the CCAV proof from~\cite{HemaspaandraHemaspaandraSchnoor-CCAV-AAAI-2014}, $n+2k$, where $n=3k$, i.e., the budget is $5k$ (here, $k$ is again the number $\card X$ from the given 3DM instance). We make the following changes to the setup in the proof of Theorem~\ref{theorem:0 dots 0 -gamma -beta -alpha CCDV}:
\begin{itemize}
\item We introduce three new distinct candidates $b_\alpha$, $b_\beta$, and $b_\gamma$ (i.e., even if $\beta=\gamma$, then $b_\beta$ is still a different candidate from $b_\gamma$). These candidates will be placed in the three ``relevant'' positions of each manipulation vote in every successful bribery action.
\item Recall that in the proof of the CCDV hardness result, i.e., Theorem~\ref{theorem:0 dots 0 -gamma -beta -alpha CCDV}, each element $(x,y,z)$ leads to four 3DM votes. We make the following addition: For each 3DM vote introduced in this way, we introduce $G$ additional votes obtained from the 3DM votes by swapping the candidates in the $-\gamma$ and $-\beta$-positions (for an appropriate value of $G$).
More precisely: For each $(x,y,z)\in M$, we add the following votes:
\begin{itemize}
\item a single vote $\dots > S_i > p > x$
\item $G$ many votes $\dots > p > S_i > x$
\item a single vote $\dots > S_i > p > y$
\item $G$ many votes $\dots > p > S_i > y$
\item a single vote $\dots > S_i' > p > z$
\item $G$ many votes $\dots > p > S_i' > z$
\item a single vote $\dots > S_i' > p > S_i$
\item $G$ many votes $\dots > p > S_i' > S_i$
\end{itemize}
The purpose of these added votes (we will call them $G$-votes in the sequel) is to ensure that the candidates $b_\alpha$, $b_\beta$, and $b_\gamma$ gain points relatively to $p$, using votes that are less attractive than the actual 3DM votes from the construction of Theorem~\ref{theorem:0 dots 0 -gamma -beta -alpha CCDV}, such that the controller will delete the latter votes instead of the $G$-votes.
\end{itemize}
Note that the controller will clearly vote $p$ in the first position of all manipulation votes. Therefore, the score of $p$ will not change from the manipulation votes, but only from the CCDV-aspect of the bribery action. Hence, the score of $p$ behaves in exactly the same way as in the CCDV proof. In the bribery instance we construct, the final scores (i.e., from the 3DM votes including the $G$-votes plus the setup votes that we will introduce below) will be as follows (recall that $n=3k$ as above):
\begin{itemize}
\item $\scorefinal p=\alpha+2\gamma$,
\item $\scorefinal c=5k\beta+2\gamma$ for each $c\in X\cup Y\cup Z$,
\item $\scorefinal{S_i}=5k\beta+\min(\alpha,2\gamma)$,
\item $\scorefinal{S_i'}=5k\beta+\alpha+\gamma$.
\item $\scorefinal{b_\heartsuit}=\alpha+2\gamma+5k(\beta+\heartsuit)$ for $\heartsuit\in\set{\alpha,\beta,\gamma}$.
\end{itemize}
The score of $b_\heartsuit$ is such that $b_\heartsuit$ ties with $p$ if $5k$ 3DM votes are removed (letting $p$ gain $5k\beta$ points), and $b_\heartsuit$ is placed in the $-\heartsuit$-position of each of the $5k$ manipulation votes. Since the score of $b_\heartsuit$ cannot be decreased by deleting 3DM votes, this ensures that the $b_\heartsuit$ candidates must in fact be placed in the relevant positions of each manipulation vote, and therefore, with regard to the remainder of the candidates, the construction works as in the CCDV case.
As in the proof of Theorem~\ref{theorem:0 dots 0 -gamma -beta -alpha CCDV}, we first compute the scores the candidates receive from the 3DM- and $G$-votes, for a candidate $x$, we call this value $\scorethreedm{x}$.
\begin{itemize}
\item $\scorethreedm{p}=-12k(\beta+G\gamma)$, since, for each tuple in $M$, $p$ gains $-4\beta-4G\gamma$ points, and $\card M=3k$.
\item $\scorethreedm{S_i}=-2\gamma-(G+1)\alpha-2G\beta$,
\item $\scorethreedm{S_i'}=-2\gamma-2G\beta$,
\item $\scorethreedm{c}=-3\alpha(G+1)$ for each $c\in X\cup Y\cup Z$, since each $c$ appears in exactly $3$ tuples from $M$,
\item $\scorethreedm{b_\heartsuit}=0$ for each $\heartsuit\in\set{\alpha,\beta,\gamma}$.
\end{itemize}
For each candidate $x$, with $\scoresetup{x}$ we denote the points that $x$ needs to receive from the setup votes in order to ensure that $\scorefinal x=\scorethreedm x+\scoresetup x$, i.e., $\scoresetup x=\scorefinal x-\scorethreedm x$. We get the following:
\begin{itemize}
\item $\scoresetup{p}=\alpha+2\gamma+12k(\beta+G\gamma)=(12kG+2)\gamma+12k\beta+\alpha$,
\item $\scoresetup{S_i}=5k\beta+\min(\alpha,2\gamma)+2\gamma+(G+1)\alpha+2G\beta=\min(\alpha,2\gamma)+2\gamma+(5k+2G)\beta+(G+1)\alpha$,
\item $\scoresetup{S_i'}=5k\beta+\alpha+\gamma+2\gamma+2G\beta=3\gamma+\beta(5k+2G)+\alpha$,
\item $\scoresetup{c}=2\gamma+5k\beta+3\alpha(G+1)$ for each $c\in X\cup Y\cup Z$,
\item $\scoresetup{b_\heartsuit}=2\gamma+\alpha+5k(\beta+\heartsuit)$.
\end{itemize}
Clearly, it again suffices to realize the relative scores among the candidates (and clearly, the absolute points of all candidates will be at most $0$, since we wrote $f$ as having no strictly positive coefficient). Hence, it suffices to construct setup votes that for each candidate $x$, let $x$ gain $\scoresetup{p}-\scoresetup{x}$ points \emph{less than the preferred candidate $p$}; this is the number of points that the $x$ must lose against $p$ from the setup votes. For each $x$, we get the following value (clearly for $p$ itself, the value is $0$):
\begin{itemize}
\item $\scorelosep{S_i}=(12kG+2)\gamma+12k\beta+\alpha-(\min(\alpha,2\gamma)+2\gamma+(5k+2G)\beta+(G+1)\alpha) = -\min(\alpha,2\gamma)+12kG\gamma+\beta(7k-2G)-G\alpha$,
note that this value can be made arbitrarily large if $k$ is chosen sufficiently large, since $G$ is a constant chosen depending on $\alpha,\beta,\gamma$, but independent of the instance and therefore of $k$.
\item $\scorelosep{S_i'}=(12kG+2)\gamma+12k\beta+\alpha-(3\gamma+\beta(5k+2G)+\alpha)
=(12kG+2)\gamma+12k\beta+\alpha-3\gamma-\beta(5k+2G)-\alpha
=\gamma(12kG-1)+\beta(7k-2G)$, again this value grows arbitrarily large in $k$.
\item $\scorelosep{c}=(12kG+2)\gamma+12k\beta+\alpha - (2\gamma+5k\beta+3\alpha(G+1))
= (12kG+2)\gamma+12k\beta+\alpha - 2\gamma-5k\beta-3\alpha(G+1)
= 12kG\gamma+7\beta+\alpha(-3G-2)$,
again the value grows arbitrarily large in $k$.
\item $\scorelosep{b_\heartsuit}=(12kG+2)\gamma+12k\beta+\alpha-(2\gamma+\alpha+5k(\beta+\heartsuit))
=(12kG+2)\gamma+12k\beta+\alpha-2\gamma-\alpha-5k(\beta+\heartsuit)
=12kG\gamma+(7k)\beta-5k\heartsuit
=k(12G\gamma+7\beta-5\heartsuit)$.
Note that, for a suitable choice of $G$, this value also grows arbitrarily for increasing $k$.
\end{itemize}
Hence, all candidates must lose points against $p$, and the number of points they must lose grows arbitrarily in $k$. Therefore, the points can be implemented using setup votes letting a candidate $x$ lose $\alpha$, $\beta$, or $\gamma$ points against all other relevant candidates. The controller will not delete these votes, since they have $p$ in one of the top positions.
Note that, since $b_\alpha$, $b_\beta$ and $b_\gamma$ must lose $5k(\alpha+\beta+\gamma)$ points against $p$ even if $p$ gains $5k\beta$ points, it follows that these candidates must take all relevant positions in the manipulation votes, and, if $\beta>\gamma$, then $p$ can only win if $p$ indeed gains $5k\beta$ points from the deletions, i.e., only if only non-$G$ 3DM votes are deleted. Therefore, for the remainder of the proof, assume that $\gamma=\beta$.
It remains to show that the controller will in fact delete only the non-$G$-3DM votes. Therefore, assume that there is a successful bribery action in which at least one of the $G$-votes is also deleted. If for every $G$-vote $v_G$ that is removed in the bribery action, the corresponding non-$G$ vote $v_{3DM}$ (obtained from the $G$-vote by swapping the $-\gamma$ and $-\beta$ positions) is not removed, then a bribery action deleting only non-$G$-3DM votes can be obtained by deleting $v_{3DM}$ instead of $v_G$ for every relevant vote (the effect for $p$ is at least as good when deleting $v_{3DM}$). Therefore, we can without loss of generality assume that there is a successful bribery action in which there is a non-$G$ 3DM vote $v_{3DM}$ such that both $v_{3DM}$ and the corresponding $G$-vote $v_G$ are deleted. We show that in this case, $p$ cannot win the election. To see this, first note that $p$ gains at most $5k\beta$ points from the delete actions. Therefore, after the bribery action, $p$'s points are at most (recall that we can assume $\beta=\gamma$)
$$\scoresub{max}p\leq\alpha+2\gamma+(5k-1)\beta+\gamma=\alpha+(5k+2)\beta.$$
We now make a case distinction depending on which type of $G$-vote is deleted.
\begin{itemize}
\item First assume that $v_G$ has a candidate $c\in X\cup Y\cup Z$ in the last position. Then the same is true for $v_{3DM}$. Since both votes are deleted, $c$ gains at least $2\alpha$ points (and note that $c$ cannot lose points from deleting other votes). Therefore, $c$ has at least $5k\beta+2\gamma+2\alpha=(5k+2)\beta+2\alpha$ points, which is strictly more than $\scoresub{max}p$.
\item Now assume that $v_G$ has a candidate $S_i$ in the last position. Analogously to the above, $S_i$ then gains at least $2\alpha$ points, and hence ends up with at least (recall that $\beta=\gamma$) $5k\beta+\min(\alpha,2\gamma) + 2\alpha \ge 5k\beta+2\gamma + 2\alpha = (5k+2)\beta+2\alpha$ points, which again is strictly more than $\scoresub{max}p$.
\end{itemize}
Therefore, in both cases $p$ does not win the election and we have a contradiction. Therefore, if the bribery instance is positive, then there exists a successful bribe in which only non-$G$ 3DM votes are deleted, as required.
\end{proof}
\section{Open Questions}\label{sect:open questions}
The main open question is to completely characterize the complexity
of $f$-manipulation, for all generators $f$. As discussed at the
end of Section~\ref{sect:results:manipulation},
we conjecture that this will not be a dichotomy theorem.
As a first step, we would like to prove that the
cases listed in Theorem~\ref{theorem:man-p} are exactly the polynomial-time
cases (under some reasonable complexity-theoretic assumptions) or to
construct an explicit counterexample to this statement.
Other interesting avenues to pursue are
going beyond NP-completeness, by looking at such issues as
fixed-parameter tractability (see, e.g., \cite{fal-nie:b:FPT}),
approximability (see e.g., \cite{DBLP:journals/jair/FaliszewskiHH15}),
and experimental results (see, e.g.,
\cite{wal:j:where-hard-veto} and \cite{rot-sch:c:empirical-control}).
|
\subsection{Necessary and sufficient condition for stability (Without approximation)}
For the system \eqref{eq:modelb}, we will perform a local stability analysis to derive a necessary and sufficient condition for stability. Suppose the equilibrium of the system is $(w_1^\ast,w_2^\ast)$. Let $u_{1}(t) = w_{1}(t)-w_{1}^{*}$ and $u_{2}(t) = w_{2}(t)-w_{2}^{*}$ be small perturbations about $w_{1}^{*}$ and $w_{2}^{*}$ respectively. Linearising the system of differential equations about its equilibrium $(w_{1}^{*},w_{2}^{*})$, we get
\begin{align}
\label{eq:linearb}
&\frac{\mathrm{d}u_{1}(t)}{dt} = -\mathcal{M}_{1}u_{1}(t)-\mathcal{N}_{1}u_{1}(t-\tau_{1})-\mathcal{P}_{1}u_{2}(t-\tau_{2}),\notag\\
&\frac{\mathrm{d}u_{2}(t)}{dt} = -\mathcal{M}_{2}u_{2}(t)-\mathcal{N}_{2}u_{2}(t-\tau_{2})-\mathcal{P}_{2}u_{1}(t-\tau_{1}),
\end{align}
where, for Compound TCP, the increase and decrease functions \eqref{eq:Compound} yield the following coefficients
\begin{align*}
\mathcal{M}_{j} &=-\frac{\alpha}{\tau_{j}}\left(k-2\right)\ \left(w_{j}^{*}\right)^{k-1},\\
\mathcal{N}_{j} &=\frac{\beta b_{j}\left(w_{j}^{*}\right)^{b_{j}+1}}{\left(\tau_{j}\right)^{b_{j}+1}\left(C_{j}\right)^{b_{j}}}+\frac{\beta B \left(w_{j}^{*}\right)^{2}}{\tau_{j}^{2}\left(C\right)^{B}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B-1},\\
\mathcal{P}_{j}&=\frac{\beta B \left(w_{j}^{*}\right)^{2}}{\tau_{1}\tau_{2}\left(C\right)^{B}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B-1},\ \ j=1,2.
\end{align*}
At equilibrium, the following equations are satisfied
\begin{align*}
\frac{\alpha}{\tau_{j}}\left(w_{j}^{*}\right)^{k-2}-\frac{\beta}{\tau_{j}}\left(\frac{w_{j}^{*}}{\tau_{j}C_{j}}\right)^{B_{j}}-\frac{\beta}{\tau_{j}C^{B}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B}=0,
\end{align*}
$ j=1,2.$ For tractability, we assume that $B_{1}=B_{2}=B, C_{1}=C_{2}=C, \tau_{1}=\tau_{2}=\tau$. Then, $w_{1}^{*}=w_{2}^{*}=w^{*}$ will be an equilibrium of the system, and satisfies the following equation:
\begin{align*}
\alpha\left(w^{*}\right)^{k-2}=\beta\left(1+2^{B}\right)\left(\frac{w^{*}}{\tau C}\right)^{B},
\end{align*}
Let $\mathcal{M}=\frac{\beta B\left(w^{*}\right)^{B+1}}{\left(\tau\right)^{B+1}\left(C\right)^{B}}$, then the coefficients $\mathcal{M}_{1}$, $\mathcal{M}_{2}$, $\mathcal{N}_{1}$, $\mathcal{N}_{2}$, $\mathcal{P}_{1}$, $\mathcal{P}_{2}$ reduce to
\begin{align}
\label{eq:reduced}
\mathcal{M}_{1}&=\mathcal{M}_{2}=-\frac{\mathcal{M}}{B}\left(1+2^{B}\right)\left(k-2\right)=a,\notag\\
\mathcal{N}_{1}&=\mathcal{N}_{2}=\mathcal{M}\left(1+2^{B-1}\right)=b,\notag\\
\mathcal{P}_{1}&=\mathcal{P}_{2}=\mathcal{M}2^{B-1}=c.
\end{align}
Note that $a$, $b$, $c$ $>$ 0. Looking for exponential solutions, we get the characteristic equation for the linearised system \eqref{eq:linearb} as
\begin{align}
\label{eq:characb}
\left(\lambda +a+b e^{-\lambda \tau}\right)^2 -c^2 e^{-2\lambda \tau}=0,
\end{align}
which can be written as
\begin{align*}
g_{1}\left(\lambda\right)g_{2}\left(\lambda\right)=0,
\end{align*}
where,
\begin{align}
g_{1}\left(\lambda\right)&= \lambda +a+\left(b+c\right)e^{-\lambda \tau}, \hspace{1ex} \text{and}\notag\\
g_{2}\left(\lambda\right)&= \lambda +a+\left(b-c\right)e^{-\lambda \tau}.
\end{align}
For stability, the roots of both $g_1(\lambda)$ and $g_2(\lambda)$ should have negative real parts. The system becomes unstable if one pair of complex conjugate roots of either $g_1(\lambda)$ or $g_2(\lambda)$ or both crosses over the imaginary axis due to increase in the average delay $\tau$ and hence have positive real parts. We find the points at which the roots of $g_1(\lambda)$ and $g_2(\lambda)$ cross over the imaginary axis. Substituting $\lambda=j\omega_1$ in $g_1(\lambda)$ and separating real and imaginary parts we get
\begin{align}
\left(b+c\right)\sin \omega_1\tau &= \omega_1, \hspace{1ex} \text{and} \label{eq:equation_1}\\
\left(b+c\right)\cos \omega_1 \tau &= -a.\label{eq:equation_2}
\end{align}
Solving \eqref{eq:equation_1} and \eqref{eq:equation_2} for $\omega_1$ we get
\begin{align*}
\omega_1 = \sqrt{\left(b+c\right)^2-a^2}
\end{align*}
and under the condition $b+c>a$, a positive value of $\omega_1^2$ exists which implies that a value of $\omega_1$ exists at which the roots of the system having the characteristic equation $g_1(\lambda)$ cross over and hence have positive real parts. Solving \eqref{eq:equation_1} and \eqref{eq:equation_2} for $\tau$ we get the critical value of delay at which the system transits from stability to instability as
\begin{align}
\label{eq:delay_critical}
\tau_{0}=\frac{1}{\omega_1}\cos^{-1}\left(\frac{-a}{b+c}\right).
\end{align}
Simiarly, substituting $\lambda=j\omega_2$ in $g_2(\lambda)$ we get the crossover frequency as
\begin{align}
\label{eq:omega_2}
\omega_2=\sqrt{\left(b-c\right)^2-a^2}
\end{align}
We substitute the Coefficients $a, b$ and $c$ in \eqref{eq:omega_2} to get
\begin{align}
\label{eq:omega_2final}
\omega_2=\mathcal{M}\sqrt{1-\frac{\left(1+2^B\right)^2\left(k-2\right)^2}{B^2}}
\end{align}
It can be shown that, for sufficiently large value of $B$, $\omega_2$ does not exist. Hence, the system having the characteristic equation $g_2(\lambda)$ is stable for all values of the delay $\tau$. Hence, the system \eqref{eq:linearb} is asymptotically stable for all $\tau<\tau_0$ and unstable for $\tau>\tau_0$. Therefore, the necessary and sufficient condition for stability of \eqref{eq:linearb} is
\begin{align}
\label{eq:ns}
\tau<\frac{1}{\omega_1}\cos^{-1}\left(\frac{-a}{b+c}\right).
\end{align}
Substituting values of $\omega_1$, $a$, $b$ and $c$ in \eqref{eq:ns}, we get the necessary and sufficient condition for local stability of \eqref{eq:linearb} as
\begin{align}
\label{eq:condition}
\alpha\left(w^{*}\right)^{k-1}\sqrt{B^{2}-\left(k-2\right)^{2}}<\cos^{-1}\left(\frac{k-2}{B}\right).
\end{align}
\subsection{Hopf Condition}
We have seen that an increase in delay prompts the system to transit from stability to instability. Instability in the system can be induced by any of the system parameters. To see the cumulative effect of all the parameters on the system stability, we introduce a non dimensional exogenous parameter $\kappa$ which can act as the bifurcation parameter. If the model parameters themselves drive the system to the edge of stability, then $\kappa=1$. To show that the roots of the system \eqref{eq:modelb} cross the imaginary axis with positive velocity as $\kappa$ is varied and the system undergoes a \emph{Hopf Bifurcation}, we proceed to prove the transversality condition of the Hopf spectrum.\\
\indent Recall that $u_1(t)=w_1(t)-w_1^{\ast}$ and $u_2(t)=w_2(t)-w_2^{\ast}$. The linearised system, with the non dimensional parameter $\kappa$ and the assumptions that $B_{1}=B_{2}=B, C_{1}=C_{2}=C, \tau_{1}=\tau_{2}=\tau$, now becomes
\begin{align}
\label{eq:linearb_kappa}
&\frac{\mathrm{d}u_{1}(t)}{dt} = \kappa \Big(-a u_{1}(t)-b u_{1}(t-\tau)-c u_{2}(t-\tau)\Big),\notag\\
&\frac{\mathrm{d}u_{2}(t)}{dt} = \kappa \Big(-a u_{2}(t)-b u_{2}(t-\tau)-c u_{1}(t-\tau)\Big).
\end{align}
Looking for exponential solutions, we get the characteristic equation of \eqref{eq:linearb_kappa} as
\begin{align}
\label{eq:characb_kappa}
\left(\lambda +\kappa a +\kappa \left(b+c\right)e^{-\lambda \tau}\right)\left(\lambda +\kappa a +\kappa \left(b-c\right)e^{-\lambda \tau}\right)=0
\end{align}
Differentiating \eqref{eq:characb_kappa} with respect to $\kappa$, we get
\scriptsize
\begin{align}
\label{eq:dl_dk}
\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}&=\notag\\
&\frac{-\kappa a^2-\lambda a-\lambda b e^{-\lambda \tau}-2\kappa abe^{-\lambda \tau}-\kappa \left(b^2-c^2\right)e^{-2 \lambda \tau}}{\lambda + \kappa a + \kappa b e^{-\lambda \tau}-\lambda \kappa b \tau e^{-\lambda \tau}-\kappa ^2ab\tau e^{-\lambda \tau}-\kappa ^2 \tau \left(b^2-c^2\right)e^{-2 \lambda \tau}}.
\end{align}
\normalsize
From \eqref{eq:characb_kappa} we get,
\begin{align}
\label{eq:exp}
e^{-\lambda \tau}= -\frac{\lambda+\kappa a}{\kappa \left(b+c\right)}
\end{align}
Substituting \eqref{eq:exp} in \eqref{eq:dl_dk} we get
\begin{align}
\label{eq:prove_real}
\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}=\frac{ \lambda}{\kappa\left(1+\lambda \tau+\kappa a \tau\right)}
\end{align}
Note that, at $\tau=\tau_0$, $\kappa=\kappa_c$. Substituting $\lambda=j\omega_1$ in \eqref{eq:prove_real} we get
\begin{align*}
\mathrm{Re}\left(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\right)_{\lambda=j \omega_1}= \frac{\omega_1^2 \tau_0}{\kappa\left(\left(1+\kappa a \tau\right)^2+\left(\omega_1 \tau\right)^2\right)}>0.
\end{align*}
Thus the system undergoes a \emph{Hopf Bifurcation} at $\kappa=\kappa_c$.
\subsection{Local stability and Hopf bifurcation analysis}
We now state conditions for local stability and Hopf bifurcation for Compound, Reno and High Speed TCP. To derive these conditions, we focus on long-lived flows. We consider the case where the core router has small buffer size and Drop-Tail queuing policy. As argued in \cite{Wischik}, for a large number of flows, the loss probability of a small-buffer Drop-Tail queue can be approximated by the blocking probability of an $M/M/1$ queue. Then, the loss probability of the bottleneck queue at equilibrium is \cite{Raja}
\begin{align}
\label{eq:loss}
p(w^{*})=\left( \frac{w^{*}}{C\tau}\right)^{B},
\end{align}
where $C$ is the capacity of the bottleneck link, $B$ is the buffer size and $w^{*}=x^{*}\tau$.
\begin{itemize}
\item{\emph{Compound TCP}.}
A necessary and sufficient condition for local stability with Compound TCP flows is \cite{Raja}
\begin{align}
\alpha \left(w^{*}\right)^{k-1}\sqrt{B^{2}-\left(k-2\right)^{2}}<\cos^{-1}(k-2/B),
\end{align}
and a Hopf bifurcation would occur at
\begin{align}
\alpha \left(w^{*}\right)^{k-1}\sqrt{B^{2}-\left(k-2\right)^{2}}=\cos^{-1}(k-2/B).
\end{align}
\item{\emph{TCP Reno}.}
Using the increase and decrease functions given by \eqref{eq:Reno} and the fluid approximation of small buffer Drop-Tail queues given by \eqref{eq:loss}, we get a necessary and sufficient condition for local stability as
\begin{align}
\label{eq:sreno}
\frac{1}{w^{*}}\sqrt{B^{2}-4}<\cos^{-1}(-2/B).
\end{align}
From \eqref{eq:hopf}, we note that a system with TCP Reno flows will undergo a Hopf bifurcation at
\begin{align}
\frac{1}{w^{*}}\sqrt{B^{2}-4}=\cos^{-1}(-2/B).
\end{align}
\item{\emph{High Speed TCP}.}
A necessary and sufficient condition for local stability for a system with High Speed TCP is
\end{itemize}
\begin{align}
\label{eq:shstcp}
&\frac{f_{1}\left(w^{*}\right)}{w^{*}}\sqrt{B^{2}-\left(\frac{w^{*}f_{1}'(w^{*})}{f_{1}(w^{*})}-2-\frac{f_{2}'(w^{*})\left(w^{*}\right)^{2}p(w^{*})}{
f_{1}\left(w^{*}\right)}\right)^{2}}\notag\\
&<\cos^{-1}\left(\frac{\frac{w^{*}f_{1}'(w^{*})}{f_{1}(w^{*})}-2-\frac{f_{2}'(w^{*})\left(w^{*}\right)^{2}p(w^{*})}{
f_{1}\left(w^{*}\right)}}{B}\right).
\end{align}
Using condition \eqref{eq:hopf}, the associated Hopf condition can be stated. Clearly, the functional forms of $i(w^{*})$, $d(w^{*})$ and the model for the queue all greatly influence stability. Additionally, protocol parameters like $\alpha$ and $k$, and network parameters like queue thresholds or buffer sizes have to be chosen rather carefully if stability is to be ensured.
\subsection{Non-oscillatory convergence with long-lived flows}
Given these various conditions of stability, we now proceed to state a condition for non-oscillatory convergence of equation \eqref{eq:lineara}. For non-oscillatory convergence, we seek conditions on model parameters $a$, $b$ and the delay $\tau$ for which the characteristic equation \eqref{eq:chareq} has negative real solution.
\begin{theorem}
The solution of the system shows non-oscillatory convergence if and only if the parameters $a$, $b$ and $\tau$ satisfy the condition $\ln\left(b\tau\right)+a\tau+1< 0$.
\end{theorem}
\begin{proof}
The boundary condition for the solution of \eqref{eq:lineara} to be non-oscillatory is the point at which the curve $f(\lambda)=\lambda+a+be^{-\lambda\tau}$ touches the real axis. If this point is $\sigma$, then
\begin{eqnarray}
f(\sigma) =& \sigma+a+be^{-\sigma\tau} &= 0, \label{conv1} \hspace{1ex} \mbox{and}\\
f^{'}(\sigma) =& 1-b\tau e^{-\sigma\tau} &= 0 \label{conv2}.
\end{eqnarray}
From \eqref{conv2}, we get
\begin{align}
be^{-\sigma\tau}=\frac{1}{\tau} \hspace{1ex} \mbox{and} \hspace{1ex} \sigma= \frac{\ln\left(b\tau\right)}{\tau}.
\end{align}
Substituting values of $\sigma$ and $be^{-\sigma\tau}$ in \eqref{conv1} gives
\begin{align}
\ln\left(b\tau\right)+a\tau +1=0.
\end{align}
Then, a condition for non-oscillatory convergence of the equilibrium point is
\begin{align}
\ln\left(b\tau\right)+a\tau+1< 0.
\end{align}
\end{proof}
\begin{itemize}
\item{\emph{Compound TCP}.}
The condition for non-oscillatory convergence with Compound TCP is
\begin{align}
\label{eq:osccompound}
\frac{\alpha B \left(w^{*}\right)^{k-1}}{\exp \left(\alpha\left(k-2\right)\left(w^{*}\right)^{k-2}-1\right)}<1.
\end{align}
We can immediately observe the relation between the protocol parameters $\alpha$ and $k$, and the network parameters like the buffer size $B$ in ensuring that convergence of the equilibrium point is non-oscillatory.
\item{\emph{TCP Reno}.}
The condition for TCP Reno is
\begin{align}
\label{eq:oscreno}
\frac{B}{w^{*}\exp \left(\frac{2}{w^{*}}-1\right)}<1.
\end{align}
This condition highlights the central role played by the parameter $B$ to ensure non-oscillatory convergence.\\
\begin{figure}
\begin{subfigure}{\columnwidth}
\hspace*{5mm}
\includegraphics[height=2in,width=3in]{plot_15_0.pdf}
\caption{Buffer size = 15}
\label{fig:long15}
\end{subfigure}%
\begin{subfigure}{\columnwidth}
\hspace*{5mm}
\includegraphics[height=3in,width=2in,angle=-90]{plot_270_0.eps}
\caption{Buffer size = 270}
\label{fig:long270}
\end{subfigure}
\caption{ \emph{Long-lived flows}. 60 long-lived Compound TCP flows over a 2 Mbps link feeding into a single bottleneck queue with link capacity 100 Mbps.}
\label{fig:long1}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{\columnwidth}
\hspace*{5mm}
\includegraphics[height=3in,width=2in,angle=-90]{plot_15_200.eps}
\caption{Buffer size = 15}
\label{fig:short15}
\end{subfigure}
\begin{subfigure}{\columnwidth}
\hspace*{5mm}
\includegraphics[height=3in,width=2in,angle=-90]{plot_270_200.eps}
\caption{Buffer size = 270}
\label{fig:short270}
\end{subfigure}
\caption{\emph{Long-lived and short-lived flows}. 60 long-lived Compound TCP flows over a 2 Mbps link, and \emph{exponentially distributed short files}, feeding into a single bottleneck queue with link capacity 100 Mbps.}
\label{fig:short1}
\end{figure}
\item{\emph{High Speed TCP}.}
With High-Speed TCP flows, the corresponding condition is
\end{itemize}
\begin{scriptsize}
\begin{align}
\label{oschstcp}
\frac{f_{1}\left(w^{*}\right)B}{w^{*}\exp \left(\frac{f_{1}(w^{*})}{w^{*}}\left(\frac{w^{*}f_{1}'(w^{*})}{f_{1}(w^{*})}-2-\frac{f_{2}'(w^{*})\left(w^{*}\right)^{2}p(w^{*})}{
f_{1}\left(w^{*}\right)}\right)-1\right)}<1.
\end{align}
\end{scriptsize}
Clearly functional forms of $f_{1}(w^{*})$ and $f_{2}(w^{*})$, and the choice of the parameter $B$, all have a role to play. It is thus clear that both the design of transport protocols and queue management policies influence the dynamics of the system.
\subsection{Long-lived and short-lived flows}
We now deviate from the assumption that the system has only long-lived flows and consider the scenario where the traffic consists of both long and short flows. On a short time scale, short TCP connections may act as an uncontrolled and random background load on the network. Suppose the workload arriving at the bottleneck queue over a time period $T$ is modelled as Gaussian with mean $x^{*}T$ and variance $x^{*}\sigma_{1}^{2}T$ and the background load due to the short transfers over the time period $T$ is also modelled as Gaussian with mean $vT$ and variance $v\sigma_{2}^{2}T$. Then the loss probability at the bottleneck queue can be expressed as \cite{Kelly}
\begin{align}
p(w^{*})=\exp \left(\frac{-2B\left(C\tau-w^{*}-v\tau\right)}{w^{*}\sigma_{1}^{2}+v\sigma_{2}^{2}\tau}\right).
\end{align}
Recall that \eqref{eq:sufficient} gives a sufficient condition for local stability and \eqref{eq:hopf} gives the condition for which the system undergoes a Hopf-type bifurcation.
\begin{itemize}
\item{\emph{Compound TCP}.}
For Compound TCP, a sufficient condition for local stability of the system is
\begin{align}
\label{eq:shortcompound}
2B \alpha \left(w^{*}\right)^{k}\tau\frac{v\sigma_{2}^{2}+\left(C-v\right)\sigma_{1}^{2}}{\left(w^{*}\sigma_{1}^{2}+v\sigma_{2}^{2}\tau\right)^2}<\frac{\pi}{2}.
\end{align}
\item{\emph{TCP Reno}.}
For TCP Reno, using the increase and decrease functions given by \eqref{eq:Reno}, we get a sufficient condition for stability as
\begin{align}
\label{eq:shortreno}
2B\tau\frac{v\sigma_{2}^{2}+\left(C-v\right)\sigma_{1}^{2}}{\left(w^{*}\sigma_{1}^{2}+v\sigma_{2}^{2}\tau\right)^2}<\frac{\pi}{2}.
\end{align}
\item{\emph{High Speed TCP}.}
With HSTCP flows, a sufficient condition for the local stability is
\begin{align}
\label{eq:shorthstcp}
2f_1(w^{*})B\tau\frac{v\sigma_{2}^{2}+\left(C-v\right)\sigma_{1}^{2}}{\left(w^{*}\sigma_{1}^{2}+v\sigma_{2}^{2}\tau\right)^2}<\frac{\pi}{2}.
\end{align}
\end{itemize}
We are also in a position to state the Hopf bifurcation conditions, which are left out due to space constraints.
The conditions \eqref{eq:shortcompound}, \eqref{eq:shortreno} and \eqref{eq:shorthstcp} capture the relationships between the various protocol and network parameters. It is interesting to note that in general, larger the value of parameter $B$, greater the possibility of driving the system to an unstable state. In Compound TCP, there appears to be an intrinsic trade off in the choice of the parameter $\alpha$ and the queue threshold parameter $B$. The presence of short-lived flows, which are modelled here as random uncontrolled traffic, does not change the requirement of choosing smaller values of $B$ to ensure stability.
\subsection{Simulations}
We now conduct packet-level simulations, using NS2 \cite{ns2}, for the single bottleneck topology over an intermediate and a small buffer sizing regime. With small buffers, we employ 15 packets. With intermediate buffers, the buffer dimensioning rule is $B=C\cdot RTT/\sqrt{N}$, where $C$ is the bottleneck link capacity, $RTT$ is the average round trip time of TCP flows and $N$ is the number of long-lived flows in the system. Using this dimensioning rule, the bottleneck queue will have a buffer size of 270 packets. We consider two scenarios (i) only long-lived flows, and (ii) a combination of long and short flows. The bottleneck link has a capacity of 100 Mbps. The packet size is fixed at 1500 bytes. In either scenario, there are 60 long-lived flows where each flow has an access link speed of 2 Mbps. The file size of each of the short flows is exponentially distributed with a mean file size of 1 Mb.
Fig.~\ref{fig:long1} depicts the simulations where the system only has long-lived flows. With small buffers, as expected, the queuing delay is negligible and the system is stable in the sense that there are no limit cycles in the queue size. With intermediate buffers, with smaller round trip times, the queues are full which yields full link utilization but at the cost of extra latency. With larger delays, limit cycles will emerge in the queue size which also start to hurt link utilisation. Fig.~\ref{fig:short1} depicts the simulation results where the system has a combination of long and short flows. Qualitatively, the results are very similar to those shown in Fig.~\ref{fig:long1}. This is expected as the models did indeed predict that despite the presence of short flows the system could readily lose stability if key system parameters were not properly dimensioned.
In the next section, we consider a multiple bottleneck topology where we will again consider the impact of long-lived, and a combination of long and short flows.
\subsection{Local stability and Hopf bifurcation analysis with long-lived flows}
We let $u(t)=w(t)-w^{*}$, and linearise \eqref{eq:modela2} about its non-trivial equilibrium point $w^{*}$ to get
\begin{align}
\dot{u}(t) = -au(t) - bu(t-\tau),\label{eq:lineara}
\end{align}
where
\begin{eqnarray}
&a& = -\frac{w^{*}}{\tau}\left(i'(w^{*})(1-p(w^{*}))-d'(w^{*})p(w^{*})\right), \hspace{1ex} \mbox{and}\notag\\
&b& = \frac{w^{*}}{\tau}p'(w^{*})d(w^{*}).
\end{eqnarray}
Looking for exponential solutions of \eqref{eq:lineara}, we get
\begin{align}
\lambda+a+be^{-\lambda\tau}=0.\label{eq:chareq}
\end{align}
From \cite{Gaurav}, if $a\geq0$, $b>0$, $b>a$ and $\tau>0$, a sufficient condition for stability of \eqref{eq:lineara} is
\begin{align}
\label{eq:sufficient}
b\tau<\frac{\pi}{2},
\end{align}
a necessary and sufficient condition for stability of \eqref{eq:lineara} is
\begin{align}
\label{eq:necessary}
\tau\sqrt{b^{2}-a^{2}}<\cos^{-1}(-a/b),
\end{align}
and the system undergoes the first Hopf bifurcation at
\begin{align}
\label{eq:hopf}
\tau\sqrt{b^{2}-a^{2}}=\cos^{-1}(-a/b).
\end{align}
We now state conditions for local stability and Hopf bifurcation for Compound, Reno and HighSpeed TCP, with long-lived flows. We consider the case where the core router has small buffer size and deploy Drop-Tail queue policy. As argued in \cite{Wischik}, for a large number of flows, the packet drop probability of a small buffer Drop-Tail queue can be approximated by the blocking probability of an $M/M/1/B$ queue. Then, the drop probability of the bottleneck queue at equilibrium is \cite{Raja}
\begin{align}
\label{eq:loss}
p(w^{*})=\left( \frac{w^{*}}{C\tau}\right)^{B},
\end{align}
where $C$ is the capacity per flow of the bottleneck router, $B$ is the buffer size and $w^{*}=x^{*}\tau$.
\begin{itemize}
\item{\emph{Compound TCP}.}
A necessary and sufficient condition for local stability with Compound TCP flows is \cite{Raja}
\begin{align*}
\alpha \left(w^{*}\right)^{k-1}\sqrt{B^{2}-\left((k-2)(1-p(w^{*}))\right)^{2}}<\cos^{-1}\left(\frac{(k-2)(1-p(w^*))}{B}\right),
\end{align*}
and a Hopf bifurcation would occur at
\begin{align*}
\alpha \left(w^{*}\right)^{k-1}\sqrt{B^{2}-\left((k-2)(1-p(w^{*}))\right)^{2}}=\cos^{-1}\left(\frac{(k-2)(1-p(w^*))}{B}\right).
\end{align*}
\item{\emph{TCP Reno}.}
Using the increase and decrease functions given by \eqref{eq:Reno} and the fluid approximation of small buffer Drop-Tail queues given by \eqref{eq:loss}, we get a necessary and sufficient condition for local stability as
\begin{align*}
\label{eq:sreno}
\frac{1}{w^{*}}\sqrt{B^{2}-4(1-p(w^*))^{2}}<\cos^{-1}\left(\frac{-2(1-p(w^{*}))}{B}\right).
\end{align*}
From \eqref{eq:hopf}, we note that a system with TCP Reno flows will undergo a Hopf bifurcation at
\begin{align}
\frac{1}{w^{*}}\sqrt{B^{2}-4(1-p(w^*))^{2}}=\cos^{-1}\left(\frac{-2(1-p(w^{*}))}{B}\right).
\end{align}
\item{\emph{HighSpeed TCP}.}
A necessary and sufficient condition for local stability for a system with High Speed TCP is
\end{itemize}
\begin{align*}
\label{eq:shstcp}
&\frac{f_{1}\left(w^{*}\right)}{w^{*}}\sqrt{B^{2}-\left(\frac{w^{*}f_{1}'(w^{*})}{f_{1}(w^{*})}-2-\frac{\left(w^{*}\right)^{3}f_{2}'(w^{*})p(w^{*})}{
f_{1}\left(w^{*}\right)}\right)^{2}\left(1-p(w^{*})\right)^{2}}\notag\\
&<\cos^{-1}\left(\frac{\left(\frac{w^{*}f_{1}'(w^{*})}{f_{1}(w^{*})}-2-\frac{\left(w^{*}\right)^{3}f_{2}'(w^{*})p(w^{*})}{
f_{1}\left(w^{*}\right)}\right)\big(1-p(w^{*})\big)}{B}\right).
\end{align*}
Using condition \eqref{eq:hopf}, the associated Hopf condition can be stated. Clearly, the functional forms of $i(w^{*})$, $d(w^{*})$ and the model for the queue all greatly influence stability. Additionally, protocol parameters like $\alpha$ and $k$, and network parameters like queue thresholds or buffer sizes have to be chosen rather carefully if stability is to be ensured.
\subsection{Non-oscillatory convergence with long-lived flows}
Given these various conditions of stability, we now proceed to state a condition for non-oscillatory convergence of equation \eqref{eq:lineara}. For non-oscillatory convergence, we seek conditions on model parameters $a$, $b$ and the delay $\tau$ for which the characteristic equation \eqref{eq:chareq} has negative real solution.
\begin{theorem}
The solution of the system shows non-oscillatory convergence if and only if the parameters $a$, $b$ and $\tau$ satisfy the condition $\ln\left(b\tau\right)+a\tau+1< 0$.
\end{theorem}
\begin{figure}
\centering
\begin{subfigure}{0.5\columnwidth}
\psfrag{aaaa}{\hspace{9mm}Time (seconds)}
\psfrag{cccc}{\hspace{1mm}Queue size (pkts)}
\psfrag{dddd}{\hspace{6mm}Round trip time = 200 ms}
\psfrag{eeee}{\hspace{6mm}Round trip time = 10 ms}
\psfrag{bbbb}{\hspace{3mm}Utilisation (\%)}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\psfrag{125}{\begin{scriptsize}$125$\end{scriptsize}}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{60}{\begin{scriptsize}$60$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\includegraphics[height=3.75in,width=2.75in,angle=-90]{plot_15_0.eps}
\caption*{\hspace{15mm}\large{(a) Buffer size = 15}}
\label{fig:long15}
\end{subfigure}%
\vspace{1mm}
\begin{subfigure}{0.5\columnwidth}
\psfrag{aaaa}{\hspace{9mm}Time (seconds)}
\psfrag{cccc}{\hspace{1mm}Queue size (pkts)}
\psfrag{dddd}{\hspace{6mm}Round trip time = 200 ms}
\psfrag{eeee}{\hspace{6mm}Round trip time = 10 ms}
\psfrag{bbbb}{\hspace{3mm}Utilisation (\%)}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\psfrag{125}{\begin{scriptsize}$125$\end{scriptsize}}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{60}{\begin{scriptsize}$60$\end{scriptsize}}
\psfrag{270}{\begin{scriptsize}$270$\end{scriptsize}}
\includegraphics[height=3.75in,width=2.75in,angle=-90]{plot_270_0.eps}
\caption*{\hspace{15mm}\large{(b) Buffer size = 270}}
\label{fig:long270}
\end{subfigure}
\caption{ \emph{Long-lived flows}. 60 long-lived Compound TCP flows over a 2 Mbps link feeding into a single bottleneck queue with link capacity 100 Mbps.}
\label{fig:long1}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{0.5\columnwidth}
\psfrag{aaaa}{\hspace{9mm}Time (seconds)}
\psfrag{cccc}{\hspace{1mm}Queue size (pkts)}
\psfrag{dddd}{\hspace{6mm}Round trip time = 200 ms}
\psfrag{eeee}{\hspace{6mm}Round trip time = 10 ms}
\psfrag{bbbb}{\hspace{3mm}Utilisation (\%)}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\psfrag{125}{\begin{scriptsize}$125$\end{scriptsize}}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{60}{\begin{scriptsize}$60$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\includegraphics[height=3.75in,width=2.75in,angle=-90]{plot_15_200.eps}
\caption*{\hspace{15mm} \large{(a) Buffer size = 15}}
\label{fig:short15}
\end{subfigure}
\vspace{1mm}
\begin{subfigure}{0.5\columnwidth}
\psfrag{aaaa}{\hspace{9mm}Time (seconds)}
\psfrag{cccc}{\hspace{1mm}Queue size (pkts)}
\psfrag{dddd}{\hspace{6mm}Round trip time = 200 ms}
\psfrag{eeee}{\hspace{6mm}Round trip time = 10 ms}
\psfrag{bbbb}{\hspace{3mm}Utilisation (\%)}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\psfrag{125}{\begin{scriptsize}$125$\end{scriptsize}}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{60}{\begin{scriptsize}$60$\end{scriptsize}}
\psfrag{270}{\begin{scriptsize}$270$\end{scriptsize}}
\includegraphics[height=3.75in,width=2.75in,angle=-90]{plot_270_200.eps}
\caption*{ \hspace{15mm}\large{(b) Buffer size = 270}}
\label{fig:short270}
\end{subfigure}
\caption{\emph{Long-lived and short-lived flows}. 60 long-lived Compound TCP flows over a 2 Mbps link, and \emph{exponentially distributed short files}, feeding into a single bottleneck queue with link capacity 100 Mbps.}
\label{fig:short1}
\end{figure}
\begin{proof}
The boundary condition for the solution of \eqref{eq:lineara} to be non-oscillatory is the point at which the curve $f(\lambda)=\lambda+a+be^{-\lambda\tau}$ touches the real axis. If this point is $\sigma$, then
\begin{eqnarray}
f(\sigma) =& \sigma+a+be^{-\sigma\tau} &= 0, \label{conv1} \hspace{1ex} \mbox{and}\\
f^{'}(\sigma) =& 1-b\tau e^{-\sigma\tau} &= 0 \label{conv2}.
\end{eqnarray}
From \eqref{conv2}, we get
\begin{align}
be^{-\sigma\tau}=\frac{1}{\tau} \hspace{1ex} \mbox{and} \hspace{1ex} \sigma= \frac{\ln\left(b\tau\right)}{\tau}.
\end{align}
Substituting values of $\sigma$ and $be^{-\sigma\tau}$ in \eqref{conv1} gives
\begin{align}
\ln\left(b\tau\right)+a\tau +1=0.
\end{align}
We now claim that the necessary and sufficient condition for non-oscillatory convergence of the equilibrium point is
\begin{align}
\label{eq:oscillation_condition}
\ln\left(b\tau\right)+a\tau+1< 0.
\end{align}
Suppose the solution of \eqref{eq:lineara} exhibits non-oscillatory convergence to its equilibrium \emph{i.e.} all roots of \eqref{eq:chareq} are real and lie on the left half of the complex plane. Then, we prove that the region of non-oscillatory convergence is characterised by \eqref{eq:oscillation_condition}. We prove this claim by contradiction. We assume that the condition for non-oscillatory convergence is
\begin{align}
\ln\left(a\tau\right)+b\tau +1>0.\label{eq:condition_contradiction}
\end{align}
Let $\sigma=-\alpha,$ where $\alpha>0$ is a root of \eqref{eq:chareq}. Then, substituting $\sigma=-\alpha$ in \eqref{eq:chareq}, we obtain
\begin{align}
\alpha = a+be^{\alpha\tau}.\label{eq:eqn_beg}
\end{align}
Multiplying both sides of \eqref{eq:eqn_beg} by $\tau e^{a\tau}$ yields
\begin{align*}
\alpha \tau e^{a\tau} =a \tau e^{a\tau}+b\tau e^{a\tau}e^{\alpha \tau}> a\tau e^{a\tau}+\frac{e^{\alpha\tau}}{e}.
\end{align*}
where, the inequality follows from \eqref{eq:condition_contradiction}. This leads to
\begin{align}
\ln\left(\alpha \tau -a \tau\right)>-1+\left(\alpha\tau-a\tau\right).\label{eq:contradiction}
\end{align}
Now, it can be easily observed form \eqref{eq:eqn_beg} that $\alpha>a.$ Hence, we obtain a contradiction \eqref{eq:contradiction}, since $\ln x\leq x-1, \hspace{2mm}\forall x>0.$ Thus, if the solution of system \eqref{eq:lineara} exhibits non-oscillatory convergence, then the parameters $a$ and $b$ satisfy the following condition:
\begin{align*}
\ln\left(a\tau\right)+b\tau +1<0.
\end{align*}
Now, we prove the converse statement \emph{i.e.} if the system parameters satisfy \eqref{eq:oscillation_condition} and all roots of \eqref{eq:chareq} have negative real parts, then all roots are real. To prove this by contradiction, we assume that all roots of \eqref{eq:chareq} have non-zero imaginary part and are of the form $\lambda=-\sigma-j\omega,$ where $\sigma>0.$ Substituting $\lambda$ in \eqref{eq:chareq} and separating real and imaginary parts, we obtain
\begin{align}
\sigma &=a+be^{\sigma \tau}\cos \omega \tau, \hspace{2ex} \text{and}\label{eq:eqn_1}\\
\omega &=be^{\sigma \tau}\sin \omega \tau. \label{eq:eqn_2}
\end{align}
From \eqref{eq:eqn_1} and \eqref{eq:eqn_2}, we get
\begin{align}
\frac{\tan \omega \tau}{\omega \tau}=\frac{1}{\left(a-\sigma\right)\tau}.\label{eq:eqn_3}
\end{align}
Now, the condition given by \eqref{eq:oscillation_condition} implies that $\left(\sigma-a\right)\tau\geq 1.$ This in turn implies that $\omega=0$ is the unique solution to the equation \eqref{eq:eqn_3}. Hence, the necessary and sufficient condition for non-oscillatory convergence of the solution of \eqref{eq:lineara} is
\begin{align*}
\ln\left(a\tau\right)+b\tau +1<0.
\end{align*}
\end{proof}
\begin{itemize}
\item{\emph{Compound TCP}.}
The condition for non-oscillatory convergence with Compound TCP is
\begin{align*}
\alpha B \left(w^{*}\right)^{k-1}<\exp \left(\alpha\left(k-2\right)\left(w^{*}\right)^{k-1}\left(1-p(w^{*})\right)-1\right).
\end{align*}
We can immediately observe the relation between the protocol parameters $\alpha$ and $k$, and the network parameters like the buffer size $B$ in ensuring that convergence of the equilibrium point is non-oscillatory.
\item{\emph{TCP Reno}.}
The condition for TCP Reno is
\begin{align*}
B<w^{*}\exp \left(-\frac{2}{w^{*}}\left(1-p(w^{*})\right)-1\right).
\end{align*}
This condition highlights the central role played by the parameter $B$ to ensure non-oscillatory convergence.\\
\item{\emph{HighSpeed TCP}.}
With HSTCP flows, the corresponding condition is
\end{itemize}
\begin{align*}
\label{oschstcp}
f_{1}\left(w^{*}\right)B<w^{*}\exp \left(A(w^{*})-1\right),
\end{align*}
\begin{figure*}
\hspace{5mm}
\begin{minipage}[b]{0.45\linewidth}
\psfrag{T}{\hspace{-1.65cm}Complementary Cumulative Distribution Function }
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\psfrag{1}{\begin{scriptsize}$1$\end{scriptsize}}
\psfrag{x}{\hspace{-1cm}Number of flows = 60}
\psfrag{b}{\hspace{5.5mm}Queue threshold}
\psfrag{aaaaaaaaaaaa}{\hspace*{-1mm}\begin{tiny}$M/M/1$\end{tiny}}
\psfrag{xyzxyz}{\hspace*{-1mm}\begin{tiny}$M/M/1/B$\end{tiny}}
\psfrag{cccccccccccc}{\hspace*{-1mm}\begin{tiny}Empirical\end{tiny}}
\psfrag{aaaaaaaaaa}{\hspace*{-1mm}\begin{tiny}$M/D/1$\end{tiny}}
\psfrag{xyz}{\hspace*{-1mm}\begin{tiny}$M/D/1/B$\end{tiny}}
\psfrag{cccccccccc}{\hspace*{-1mm}\begin{tiny}Empirical\end{tiny}}
\includegraphics[height=2.68in,width=1.75in,angle=-90]{plot1.eps}
\label{fig:minipage1}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\psfrag{1}{\begin{scriptsize}$1$\end{scriptsize}}
\psfrag{x}{\hspace{-1cm}Number of flows = 120}
\psfrag{aaaaaaaaaaaa}{\hspace*{-1mm}\begin{tiny}$M/M/1$\end{tiny}}
\psfrag{xyzxyz}{\hspace*{-1mm}\begin{tiny}$M/M/1/B$\end{tiny}}
\psfrag{cccccccccccc}{\hspace*{-1mm}\begin{tiny}Empirical\end{tiny}}
\psfrag{aaaaaaaaaa}{\hspace*{-1mm}\begin{tiny}$M/D/1$\end{tiny}}
\psfrag{xyz}{\hspace*{-1mm}\begin{tiny}$M/D/1/B$\end{tiny}}
\psfrag{cccccccccc}{\hspace*{-1mm}\begin{tiny}Empirical\end{tiny}}
\includegraphics[height=2.68in,width=1.75in,angle=-90]{plot2.eps}
\end{minipage}\vspace{-6mm}
\caption*{\hspace{-4mm}(a) Round trip time = $100$ ms}
\quad
\hspace{1mm}
\begin{minipage}[b]{0.45\linewidth}
\psfrag{T}{\hspace{-1.65cm}Complementary Cumulative Distribution Function}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\psfrag{1}{\begin{scriptsize}$1$\end{scriptsize}}
\psfrag{x}{\hspace{-1cm}Number of flows = 60}
\psfrag{b}{\hspace{5.5mm}Queue threshold}
\psfrag{aaaaaaaa}{\hspace*{-1mm}\begin{tiny}$M/M/1$\end{tiny}}
\psfrag{xyzxyz}{\hspace*{-1mm}\begin{tiny}$M/M/1/B$\end{tiny}}
\psfrag{cccccccc}{\hspace*{-1mm}\begin{tiny}Empirical\end{tiny}}
\psfrag{xx}{\hspace*{-1mm}\begin{tiny}$M/D/1$\end{tiny}}
\psfrag{xyz}{\hspace*{-1mm}\begin{tiny}$M/D/1/B$\end{tiny}}
\psfrag{dddddddd}{\hspace*{-1mm}\begin{tiny}Empirical\end{tiny}}
\includegraphics[height=2.68in,width=1.75in,angle=-90]{plot3.eps}
\label{fig:minipage2}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\psfrag{1}{\begin{scriptsize}$1$\end{scriptsize}}
\psfrag{x}{\hspace{-1cm}Number of flows = 120}
\psfrag{aaaaaaaa}{\hspace*{-1mm}\begin{tiny}$M/M/1$\end{tiny}}
\psfrag{xyzxyz}{\hspace*{-1mm}\begin{tiny}$M/M/1/B$\end{tiny}}
\psfrag{cccccccc}{\hspace*{-1mm}\begin{tiny}Empirical\end{tiny}}
\psfrag{zzz}{\hspace*{-1mm}\begin{tiny}$M/D/1$\end{tiny}}
\psfrag{xyz}{\hspace*{-1mm}\begin{tiny}$M/D/1/B$\end{tiny}}
\psfrag{dddddddd}{\hspace*{-1mm}\begin{tiny}Empirical\end{tiny}}
\includegraphics[height=2.68in,width=1.75in,angle=-90]{plot4.eps}
\end{minipage}\vspace{-6mm}
\caption*{\hspace{-4mm}(b) Round trip time = $150$ ms}
\quad
\hspace{1mm}
\begin{minipage}[b]{0.45\linewidth}
\psfrag{T}{\hspace{-1.65cm}Complementary Cumulative Distribution Function}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\psfrag{1}{\begin{scriptsize}$1$\end{scriptsize}}
\psfrag{x}{\hspace{-1cm}Number of flows = 60}
\psfrag{b}{\hspace{5.5mm}Queue threshold}
\psfrag{aaa}{\hspace*{-1mm}\begin{tiny}$M/M/1$\end{tiny}}
\psfrag{zzz}{\hspace*{-1mm}\begin{tiny}$M/D/1$\end{tiny}}
\psfrag{xxx}{\hspace*{-1mm}\begin{tiny}$M/M/1/B$\end{tiny}}
\psfrag{ccc}{\hspace*{-1mm}\begin{tiny}Empirical\end{tiny}}
\psfrag{xyz}{\hspace*{-1mm}\begin{tiny}$M/D/1/B$\end{tiny}}
\psfrag{ddd}{\hspace*{-1mm}\begin{tiny}Empirical\end{tiny}}
\includegraphics[height=2.68in,width=1.75in,angle=-90]{plot5.eps}
\label{fig:minipage3}
\end{minipage}
\begin{minipage}[b]{0.45\linewidth}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\psfrag{1}{\begin{scriptsize}$1$\end{scriptsize}}
\psfrag{x}{\hspace{-1cm}Number of flows = 120}
\psfrag{aaaaaaaa}{\hspace*{-1mm}\begin{tiny}$M/M/1$\end{tiny}}
\psfrag{xyzxyz}{\hspace*{-1mm}\begin{tiny}$M/M/1/B$\end{tiny}}
\psfrag{cccccccc}{\hspace*{-1mm}\begin{tiny}Empirical\end{tiny}}
\psfrag{zzz}{\hspace*{-1mm}\begin{tiny}$M/D/1$\end{tiny}}
\psfrag{xyz}{\hspace*{-1mm}\begin{tiny}$M/D/1/B$\end{tiny}}
\psfrag{dddddddd}{\hspace*{-1mm}\begin{tiny}Empirical\end{tiny}}
\includegraphics[height=2.68in,width=1.75in,angle=-90]{plot6.eps}
\end{minipage}\vspace{-6mm}
\caption*{\hspace{-4mm}(c) Round trip time = $200$ ms}
\caption{ \emph{Statistics of the queue size.} Empirical queue length distribution for single bottleneck topology with $60$ and $120$ long-lived flows each having an access speed of $2$ Mbps, and $1$ Mbps respectively for round trip times (a) $100$ ms, (b) $150$ ms and (c) $200$ ms. We compare the empirical queue statistics with the queue length distributions of $M/M/1$, $M/D/1$, $M/M/1/B$, and $M/D/1/B$ for three different round trip times.}
\label{fig:queuedist}
\end{figure*}
\noindent where, $A(w^{*})=\frac{f_{1}(w^{*})}{w^{*}}\left(\frac{w^{*}f_{1}'(w^{*})}{f_{1}(w^{*})}-2-\frac{\left(w^{*}\right)^{3}f_{2}'(w^{*})p(w^{*})}{
f_{1}\left(w^{*}\right)}\right)$.
Clearly functional forms of $f_{1}(w^{*})$ and $f_{2}(w^{*})$, and the choice of the parameter $B$, all have a role to play. It is thus clear that both design of transport protocols and queue management policies influence the dynamics of the system. To facilitate a better understanding of the analytical results derived above, we present a stability chart as shown in Fig.~\ref{fig:stability}. Since, in this paper, we primarily focus on the performance of Compound TCP, we illustrate the stability chart in terms of its protocol parameter $\alpha$ and buffer threshold $B$. This characterises the Hopf Condition, sufficient condition for local stability and condition for non-oscillatory convergence for Compound TCP with a large number of long-lived flows in the single bottleneck scenario. The stability chart complements our analytical insight in the sense that as buffer size is increased, the protocol parameter $\alpha$ has to be decreased to ensure both stability and non-oscillatory convergence of the system.
\subsection{Local stability and Hopf bifurcation analysis with long-lived and short-lived flows}
We now deviate from the assumption that the system has only long-lived flows and consider the scenario where the traffic consists of both long and short flows. On a short time scale, short TCP connections may act as an uncontrolled and random background load on the network. Suppose the workload arriving at the bottleneck queue over a time period $T$ is modelled as Gaussian with mean $x^{*}T$ and variance $x^{*}\sigma_{1}^{2}T$ and the background load due to the short transfers over the time period $T$ is also modelled as Gaussian with mean $vT$ and variance $v\sigma_{2}^{2}T$. Then the loss probability at the bottleneck queue can be expressed as \cite{Kelly}
\begin{align}
p(w^{*})=\exp \left(\frac{-2B\left(C\tau-w^{*}-v\tau\right)}{w^{*}\sigma_{1}^{2}+v\sigma_{2}^{2}\tau}\right).
\end{align}
Recall that \eqref{eq:sufficient} gives a sufficient condition for local stability and \eqref{eq:hopf} gives the condition for which the system undergoes a Hopf-type bifurcation. We now particularise the sufficient condition for Compound TCP, TCP Reno and HighSpeed TCP. We are also in a position to state the Hopf bifurcation conditions, which are left out due to space constraints.
\begin{figure}
\begin{center}
\label{fig:meanqueue}
\psfrag{eeeee}{\hspace{-9mm}Round trip time, $\tau$}
\psfrag{fffff}{Mean queue size}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\psfrag{10}{\begin{scriptsize}$10$\end{scriptsize}}
\psfrag{200}{\begin{scriptsize}$200$\end{scriptsize}}
\psfrag{aaaaaaaaaaaa}{\begin{scriptsize}No. of flows = $60$\end{scriptsize}}
\psfrag{xyzxyz}{\begin{scriptsize}No. of flows = $120$\end{scriptsize}}
\includegraphics[width=2.25in,height=3in,angle=270]{queuemean.eps}
\caption{\emph{Mean queue size.} Evolution of mean queue size as a function of round trip time. We consider the number of long-lived flows to be $60$ and $120$ with link access speeds being $2$ Mbps and $1$ Mbps respectively.}
\label{fig:charts}
\end{center}
\end{figure}
\begin{itemize}
\item{\emph{Compound TCP}.}
For Compound TCP, a sufficient condition for local stability of the system is
\begin{align}
\label{eq:shortcompound}
2B \alpha \left(w^{*}\right)^{k}\tau\frac{v\sigma_{2}^{2}+\left(C-v\right)\sigma_{1}^{2}}{\left(w^{*}\sigma_{1}^{2}+v\sigma_{2}^{2}\tau\right)^2}<\frac{\pi}{2}.
\end{align}
\item{\emph{TCP Reno}.}
For TCP Reno, using the increase and decrease functions given by \eqref{eq:Reno}, we get a sufficient condition for stability as
\begin{align}
\label{eq:shortreno}
2B\tau\frac{v\sigma_{2}^{2}+\left(C-v\right)\sigma_{1}^{2}}{\left(w^{*}\sigma_{1}^{2}+v\sigma_{2}^{2}\tau\right)^2}<\frac{\pi}{2}.
\end{align}
\item{\emph{HighSpeed TCP}.}
With HSTCP flows, a sufficient condition for the local stability is
\begin{align}
\label{eq:shorthstcp}
2Bf_1(w^{*})\tau\frac{v\sigma_{2}^{2}+\left(C-v\right)\sigma_{1}^{2}}{\left(w^{*}\sigma_{1}^{2}+v\sigma_{2}^{2}\tau\right)^2}<\frac{\pi}{2}.
\end{align}
\end{itemize}
The conditions \eqref{eq:shortcompound}, \eqref{eq:shortreno} and \eqref{eq:shorthstcp} capture the relationships between the various protocol and network parameters. It is interesting to note that in general, larger the value of parameter $B$, greater the possibility of driving the system to an unstable state. In Compound TCP, there appears to be an intrinsic trade off in the choice of the parameter $\alpha$ and the queue threshold parameter $B$. The presence of short-lived flows, which are modelled here as random uncontrolled traffic, does not change the requirement of choosing smaller values of $B$ to ensure stability. We now present some packet-level simulations, which will enable us to comment on the dynamical and statistical properties of the system.
\subsection{Simulations}
\subsubsection*{Dynamical Properties}
We now conduct packet-level simulations, using NS2 \cite{ns2}, for the single bottleneck topology in an intermediate and a small buffer sizing regime. With small buffers, we employ $15$ packets. With intermediate buffers, the buffer dimensioning rule is $B=C\cdot RTT/\sqrt{N}$, where $C$ is the bottleneck link capacity, $RTT$ is the average round trip time of TCP flows and $N$ is the number of long-lived flows in the system. Using this dimensioning rule, the bottleneck queue will have a buffer size of $270$ packets. We consider two scenarios (i) only long-lived flows, and (ii) a combination of long and short flows. The bottleneck link has a capacity of $100$ Mbps. The packet size is fixed at $1500$ bytes. In either scenario, we consider $60$ and $120$ long-lived flows where each flow has an access link speed of 2 Mbps and 1 Mbps respectively. The file size of each of the short flows is exponentially distributed with a mean file size of $1$ Mb. Since the dynamical properties of the underlying system for both $60$ and $120$ long-lived flows are qualitatively similar, we demonstrate simulations only for the case of 60 long-lived flows.
Fig.~\ref{fig:long1} depicts the simulations where the system only has long-lived flows. With small buffers, as expected, the queuing delay is negligible and the system is stable in the sense that there are no limit cycles in the queue size. With intermediate buffers, with smaller round trip times, the queues are full which yields full link utilisation but at the cost of extra latency. With larger delays, limit cycles will emerge in the queue size which also start to hurt link utilisation. Fig.~\ref{fig:short1} depicts the simulation results where the system has a combination of long and short-lived flows. Qualitatively, the results are very similar to those shown in Fig.~\ref{fig:long1}. This is expected as the models did indeed predict that despite the presence of short flows the system could readily lose stability if key system parameters were not properly dimensioned.
\subsubsection*{Statistical Properties}
First, in the small-buffer regime, \emph{i.e.}, for buffer sizes of 15 and 100 packets, and small round trip times, the queue is almost full. Thus, the bottleneck queue acts as an integrator. Second, for a buffer size of $100$ packets and larger round trip times, the queue dynamics exhibit stable limit cycles. Clearly, in these two cases, the queue size does not have a stationary distribution. Further, note that, for a buffer size of $15$ packets and larger round trip times ($200$ ms), there are random fluctuations in the bottleneck queue. Hence, we focus on the statistical properties of the queue in the third regime.
In Fig.~\ref{fig:queuedist}, we demonstrate the empirical complementary cumulative distribution of the queue size for a buffer size of $15$ packets and $60$ long-lived flows, each with an access speed of $2$ Mbps, and with $120$ long-lived flows each with an access speed of $1$ Mbps. The round trip times which we consider are $100$ ms, $150$ ms and $200$ ms. For each of these cases, we perform a comparative study of the empirical queue length distribution with the theoretical queue distributions of $M/M/1$, $M/D/1$, $M/M/1/B$ and $M/D/1/B$ queues. From the simulations we can infer that, with $60$ long-lived flows, the empirical queue distribution can be reasonably approximated by the corresponding queue distribution of either an $M/M/1/B$ or an $M/D/1/B$ queue when the round trip time is large ($100$ ms). Notably, as the number of flows is increased to $120$, this approximation holds true for larger round trip times perhaps due to increased statistical multiplexing. Indeed, it can be verified that, as the number of long-lived flows is increased further, this approximation holds true for even larger round trip time values. This serves to validate the packet drop probability at the core router and the fluid model in the regime with a large number of long-lived flows and high bandwidth-delay product. Fig.~\ref{fig:charts} illustrates the evolution of mean queue size as a function of round trip time for two scenarios. The first scenario consists of $60$ long-lived Compound TCP flows, each with an access speed of $2$ Mbps, while the second consists of $120$ long-lived flows, each with an access speed of $1$ Mbps. It can be immediately observed that, the mean queue sizes in the second scenario are larger than in the first scenario, for all round trip times. Due to increased statistical multiplexing, the average packet loss incurred by each user in the second scenario is less as compared to the first scenario. Consequently, the average arrival rate of packets at the bottleneck queue is larger in the second scenario. This is intuitive, since each TCP sender backs off less aggressively, owing to low packet loss. This, in turn, results in higher mean queue size in the second scenario, for all round tip times.
In the next section, we consider a multiple bottleneck topology which depicts a more realistic network scenario as opposed to the simple single bottleneck topology. For this topology, we will again consider the impact of long-lived, and a combination of long and short flows.
\subsection{Necessary and sufficient condition for stability}
For system \eqref{eq:modelb}, we will perform a local stability analysis to derive a necessary and sufficient condition for stability. Suppose the equilibrium of the system is $(w_1^\ast,w_2^\ast)$. Let $u_{1}(t) = w_{1}(t)-w_{1}^{*}$ and $u_{2}(t) = w_{2}(t)-w_{2}^{*}$ be small perturbations about $w_{1}^{*}$ and $w_{2}^{*}$ respectively. Linearising system \eqref{eq:modelb} about its equilibrium $(w_{1}^{*},w_{2}^{*})$, we get
\begin{align}
\label{eq:linearb}
&\dot{u}_1(t) = -\mathcal{M}_{1}u_{1}(t)-\mathcal{N}_{1}u_{1}(t-\tau_{1})-\mathcal{P}_{1}u_{2}(t-\tau_{2}),\notag\\
&\dot{u}_2(t) = -\mathcal{M}_{2}u_{2}(t)-\mathcal{N}_{2}u_{2}(t-\tau_{2})-\mathcal{P}_{2}u_{1}(t-\tau_{1}),
\end{align}
where, for Compound TCP, the increase and decrease functions \eqref{eq:Compound} yield the following coefficients
\begin{align}
\label{eq:coefficients}
\mathcal{M}_{j} =&-\frac{\alpha}{\tau_{j}}\left(k-2\right) \left(w_{j}^{*}\right)^{k-1}\Bigg(1-\bigg(\frac{w_j^*}{C_j\tau_j}\bigg)^{B_j}-\frac{1}{(2C)^B}\bigg(\frac{w_1^*}{\tau_1}+\frac{w_2^*}{\tau_2}\bigg)^{B}\Bigg),\notag\\
\mathcal{N}_{j} =&\left(\alpha\left(w_{j}^{*}\right)^{k-1}+\beta w_{j}^{*}\right)\Bigg(\frac{B_{j}}{\tau_j}\left(\frac{w_{j}^{*}}{C_j \tau_j}\right)^{B_j}+\frac{B w_{j}^{*}}{(2C)^{B}\tau_{j}^{2}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B-1}\Bigg),\notag\\
\mathcal{P}_{j}=&\left(\alpha\left(w_{j}^{*}\right)^{k-1}+\beta w_{j}^{*}\right)\frac{B w_{j}^{*}}{\tau_{1}\tau_{2}\left(2C\right)^{B}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B-1},\hspace{6mm}j=1,2.
\end{align}
At equilibrium, the following equations are satisfied
\begin{align*}
\alpha \left(w_{j}^{*}\right)&^{k-1}\left(1-\left(\frac{w_{j}^{*}}{\tau_{j}C_{j}}\right)^{B_{j}}+\frac{1}{(2C)^{B}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B}\right)=\beta w_{j}^{*}\left(\left(\frac{w_{j}^{*}}{\tau_{j}C_{j}}\right)^{B_{j}}+\frac{1}{(2C)^{B}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B}\right),\hspace{1ex} j=1,2.
\end{align*}
For analytical tractability, we consider two different scenarios with simple assumptions. It is worthwhile to note that both scenarios are valid in the context of a real network.
\subsection*{Case I}
In this scenario, we assume that the network parameters for all routers are the same \emph{i.e.} $B_{1}=B_{2}=B, C_{1}=C_{2}=C.$ We further assume that the round trip times of both sets of TCP flows are identical \emph{i.e.} $\tau_{1}=\tau_{2}=\tau$. Then, $w_{1}^{*}=w_{2}^{*}=w^{*}$ will be an equilibrium of the system, and satisfies the following equation:
\begin{align*}
\alpha\left(w^{*}\right)^{k-2}=2\left(\alpha\left(w^{*}\right)^{k-2}+\beta\right)\left(\frac{w^{*}}{\tau C}\right)^{B}.
\end{align*}
\noindent Let $\mathcal{M}=\frac{\left(\alpha\left(w^{*}\right)^{k-2}+\beta\right)B\left(w^{*}\right)^{B+1}}{\tau^{B+1}C^{B}}$, then the coefficients $\mathcal{M}_{1}$, $\mathcal{M}_{2}$, $\mathcal{N}_{1}$, $\mathcal{N}_{2}$, $\mathcal{P}_{1}$, $\mathcal{P}_{2}$ reduce to
\begin{align}
\label{eq:reduced}
\mathcal{M}_{1}&=\mathcal{M}_{2}=\frac{2\mathcal{M}\beta w^{*}\left(2-k\right)}{\left(\alpha\left(w_{j}^{*}\right)^{k-1}+\beta w_{j}^{*}\right)B}=a,\notag\\
\mathcal{N}_{1}&=\mathcal{N}_{2}=\frac{3}{2}\mathcal{M}=b,\notag\\
\mathcal{P}_{1}&=\mathcal{P}_{2}=\frac{1}{2}\mathcal{M}=c.
\end{align}
Note that $a$, $b$, $c$ $>$ 0, and $b>c$. Then, the linearised system \eqref{eq:linearb} becomes
\begin{align}
\label{eq:linearb_wkappa}
&\dot{u}_{1}(t) = -a u_{1}(t)-b u_{1}(t-\tau)-c u_{2}(t-\tau),\notag\\
&\dot{u}_{2}(t) = -a u_{2}(t)-b u_{2}(t-\tau)-c u_{1}(t-\tau).
\end{align}
\begin{theorem}
System \eqref{eq:linearb_wkappa} is stable if and only if the parameters $a$, $b$, $c$ and $\tau$ satisfy the condition $\tau<\frac{1}{\omega_1}\cos^{-1}\left(\frac{-a}{b+c}\right)$ with crossover frequency $\omega_1 = \sqrt{\left(b+c\right)^2-a^2}$.
\end{theorem}
\begin{proof}
Looking for exponential solutions, we get the characteristic equation for the linearised system \eqref{eq:linearb_wkappa} as
\begin{align}
\label{eq:characb}
\left(\lambda +a+b e^{-\lambda \tau}\right)^2 -c^2 e^{-2\lambda \tau}=0,
\end{align}
which can be written as
\begin{align*}
g_{1}\left(\lambda\right)g_{2}\left(\lambda\right)=0,
\end{align*}
where,
\begin{align}
g_{1}\left(\lambda\right)&= \lambda +a+\left(b+c\right)e^{-\lambda \tau}, \hspace{1ex} \text{and} \notag\\
g_{2}\left(\lambda\right)&= \lambda +a+\left(b-c\right)e^{-\lambda \tau}.
\end{align}
For stability, all roots of \eqref{eq:characb} should have negative real parts. For negligible values of delay $\tau$, system \eqref{eq:linearb_wkappa} is stable, \emph{i.e.} all roots of the characteristic equation lie of the left half of the complex plane. As the delay is increased, the system becomes unstable if one pair of complex conjugate roots of either $g_1(\lambda)$ or $g_2(\lambda)$ or both crosses over the imaginary axis. We aim to determine the values of delay $\tau$ at which one pair of complex conjugate roots of $g_1(\lambda)$ and $g_2(\lambda)$ cross over the imaginary axis. Let $\tau_{1,c}$ and $\tau_{2,c}$ denote the values of $\tau$ at which $g_1(\lambda)$ and $g_2(\lambda)$ have exactly one pair of purely imaginary roots. Then, the critical value of $\tau$, denoted by $\tau_c$, at which \eqref{eq:characb} has one pair of purely imaginary roots is $\tau_c=\min(\tau_{1,c},\tau_{2,c})$~\cite{Campbell}. Substituting $\lambda=j\omega_1$ in $g_1(\lambda)$ and separating real and imaginary parts we get
\begin{align}
\left(b+c\right)\sin \omega_1\tau &= \omega_1, \hspace{1ex} \text{and} \label{eq:equation_1}\\
\left(b+c\right)\cos \omega_1 \tau &= -a.\label{eq:equation_2}
\end{align}
Solving \eqref{eq:equation_1} and \eqref{eq:equation_2} for $\omega_1$ we get
\begin{align}
\omega_1 = \sqrt{\left(b+c\right)^2-a^2},
\end{align}
and under the condition $b+c>a$, $\omega_1^2$ is strictly positive. This implies that there exists a cross over frequency $\omega_1$ at which one pair of complex conjugate roots of $g_1(\lambda)$ crosses over to the right half of the complex plane. Solving \eqref{eq:equation_1} and \eqref{eq:equation_2} for $\tau$, we get the critical value of delay at which the system transits from stability to instability as
\begin{align}
\label{eq:delay_critical}
\tau_{1,c}=\frac{1}{\omega_1}\cos^{-1}\left(\frac{-a}{b+c}\right).
\end{align}
Similarly, substituting $\lambda=j\omega_2$ in $g_2(\lambda)$ we get the cross over frequency as
\begin{align}
\label{eq:omega_2}
\omega_2=\sqrt{\left(b-c\right)^2-a^2}.
\end{align}
and under the condition $b-c>a$, $\omega_2^2$ is strictly positive. Hence, there exists a cross over frequency $\omega_2$ at which
one pair of complex conjugate roots of $g_2(\lambda)$ crosses over to the right half of the complex plane. Solving for $\tau$, we obtain the critical value of the delay at which the system having the characteristic equation $g_2(\lambda)$ has exactly one pair of purely imaginary roots as
\begin{align}
\label{eq:delay_critical_2}
\tau_{2,c}=\frac{1}{\omega_2}\cos^{-1}\left(\frac{-a}{b-c}\right).
\end{align}
Observe that $\omega_1>\omega_2.$ Since $\cos^{-1}(x)$ is monotonically decreasing for $x\in [-1,1]$, it can be shown that $\tau_1<\tau_2.$ This implies that $\tau_c=\tau_{1,c}$. Consequently, all roots of the characteristic equation \eqref{eq:characb} lie on the left half of the complex plane for all $\tau<\tau_{c}$. Hence, system \eqref{eq:linearb_wkappa} is asymptotically stable for $\tau<\tau_c$, and unstable for $\tau>\tau_c$. Further, it can be analytically shown that this loss of stability occurs via a Hopf bifurcation when one pair of complex conjugate roots of \eqref{eq:characb} crosses over the imaginary axis with non-zero velocity at $\tau=\tau_c$. Therefore, the necessary and sufficient condition for local stability of \eqref{eq:linearb_wkappa} is
\begin{align}
\label{eq:ns}
\tau<\frac{1}{\omega_1}\cos^{-1}\left(\frac{-a}{b+c}\right).
\end{align}
\end{proof}
Substituting values of $\omega_1$, $a$, $b$ and $c$ in \eqref{eq:ns}, we get the necessary and sufficient condition for local stability of \eqref{eq:modelb}, with Compound TCP as
\begin{align}
\label{eq:condition}
\alpha \left(w^{*}\right)^{k-1}\sqrt{B^{2}
-(k-2)^2\left(1-2p(w^{*})\right)^{2}}<\cos^{-1}
\left(\frac{(k-2)\left(1-2p(w^*)\right)}{B}\right),
\end{align}
\noindent where $p(w^{*})=\left(\frac{w^{*}}{C\tau}\right)^B$. This condition captures the relationship between the equilibrium window size, protocol parameters $k$ and $\alpha$, and buffer size $B$ of the core router to ensure stability of the system. If we increase the buffer size of the core router, keeping other parameters fixed, condition \eqref{eq:condition} would get violated.
\subsection*{Case II}
In this scenario, we assume that the network parameters for all routers are distinct, and the average round trip time of the first set of TCP flows is much larger than the other. Further, we consider that the average round trip time of the second set of TCP flows is negligible. This implies that $\tau_1>>\tau_2$ and $\tau_2\approx 0$. As a consequence of this assumption, the dynamics of the second set of TCP flows will appear to be almost instantaneous. This leads to the following non-linear, time-delayed fluid model of the system:
\begin{align}
\dot{w}_1(t) =& \frac{w_{1}(t-\tau_{1})}{\tau_{1}}\bigg(i\left(w_{1}(t)\right)\Big(1-p_{1}(t-\tau_{1})-q(t,\tau_{1},\tau_{2})\Big)- d\left((w_{1}(t)\right)\Big(p_{1}(t-\tau_{1})+q(t,\tau_{1},\tau_{2})\Big)\bigg),\notag \\
\dot{w}_2(t) =& \frac{w_{2}(t)}{\tau_{2}}\bigg(i\left(w_{2}(t)\right)\Big(1-p_{2}(t)-q(t,\tau_{1},\tau_{2})\Big)- d\left((w_{2}(t)\right)\Big(p_{2}(t)+q(t,\tau_{1},\tau_{2})\Big)\bigg).
\label{eq:modelb_1}
\end{align}
The loss probabilities at the three routers are approximated as
$$
p_{1}(t)=\left(\frac{w_{1}(t)}{C_{1}\tau_{1}}\right)^{B_{1}} \hspace{1ex},\hspace{1ex} p_{2}(t)=\left(\frac{w_{2}(t)}{C_{2}\tau_{2}}\right)^{B_{2}},\ {\rm and}
$$$$q(t,\tau_{1},\tau_{2})= \left(\frac{w_{1}(t-\tau_{1})/\tau_{1}+w_{2}(t)/\tau_{2}}{\widetilde{C}}\right)^{B}.$$
Here, $\widetilde{C}=2C.$ We now proceed to perform a local stability analysis for system \eqref{eq:modelb_1}. This will enable us to characterise the stability of the system in the presence of heterogeneity in network parameters. Suppose $(w_{1}^{*},w_{2}^{*})$ be a non-trivial equilibrium of system \eqref{eq:modelb_1}. Let $u_{1}(t)=w_{1}(t)-w_{1}^{*}$ and $u_{2}(t)=w_{2}(t)-w_{2}^{*}$ represent small perturbations about $w_1^{*}$ and $w_{2}^{*}$ respectively. Linearising system \eqref{eq:modelb_1} about its equilibrium, we get the following:
\begin{align}
\label{eq:linearb1}
&\dot{u}_1(t) = -\mathcal{M}_{1}u_{1}(t)-\mathcal{N}_{1}u_{1}(t-\tau_{1})-\mathcal{P}_{1}u_{2}(t),\notag\\
&\dot{u}_2(t) = -\big(\mathcal{M}_{2}+\mathcal{N}_{2}\big)u_{2}(t)-\mathcal{P}_{2}u_{1}(t-\tau_{1}),
\end{align}
where, the coefficients are given by \eqref{eq:coefficients}. Looking for exponential solutions, we get the characteristic equation for the linearised system \eqref{eq:linearb1} as
\begin{align}
\label{eq:charac_assum2}
\lambda^2 + a\lambda +b\lambda e^{-\lambda \tau_1} + c e^{-\lambda \tau_1} + d=0,
\end{align}
where,
\begin{align}
\label{eq: abcd}
&a = \mathcal{M}_{1}+\mathcal{M}_{2}+\mathcal{N}_{2} \hspace{1ex}, \hspace{1ex} b=\mathcal{N}_{2},\notag\\
&c=\mathcal{M}_{1}\left(\mathcal{M}_{2}+\mathcal{N}_{2}\right), \hspace{2ex}d=\mathcal{N}_{1}\left(\mathcal{M}_{2}+\mathcal{N}_{2}\right)-
\mathcal{P}_{1}\mathcal{P}_{2}.
\end{align}
\begin{theorem}
System \eqref{eq:linearb1}, having the characteristic equation \eqref{eq:charac_assum2} is stable if and only if the parameters satisfy the condition
$\tau_1<\frac{1}{\omega}\cos^{-1}\Big(\frac{\omega^2(c-ab)-cd}{b^2\omega^2+c^2}\Big)$ with the cross over frequency as
\begin{align*}
\omega = \sqrt{\frac{(2d-a^2+b^2)}{2}+ \frac{\sqrt{(2d-a^2+b^2)^2-4(d^2-c^2)}}{2}}.
\end{align*}
\end{theorem}
\begin{proof}
For system \eqref{eq:linearb1} to be stable, and hence system \eqref{eq:modelb_1} to be locally stable about its equilibrium, all roots of the characteristic equation should lie on the left half of the complex plane. For negligible delays, system \eqref{eq:linearb1} can be shown to be stable. As $\tau_1$ increases beyond a critical value, the system will transit into the unstable region if one pair of complex conjugate roots crosses over the imaginary axis. Following a similar kind of analysis as done in Case I, we can conclude that the system has a pair of purely imaginary complex conjugate roots with cross over frequency satisfying
\begin{align*}
\omega^2 = \frac{(2d-a^2+b^2)}{2}\pm \frac{\sqrt{(2d-a^2+b^2)^2-4(d^2-c^2)}}{2}.
\end{align*}
\emph{Condition 1}: There exists only one positive value of $\omega^2$ if either of the following conditions holds
\begin{itemize}
\item[(i)]$(2d-a^2+b^2)>0,\text{and}\hspace{1mm} (2d-a^2+b^2)^2=4(d^2-c^2).$
\item[(ii)]$d^2-c^2<0.$
\end{itemize}
\emph{Condition 2}: There exist two positive values of $\omega^2$ if the following condition holds
\begin{align*}
(2d-a^2+b^2)>0,\hspace{1mm} (2d-a^2+b^2)>4(d^2-c^2)\hspace{1mm}\text{and}\hspace{1mm} (d^2-c^2)>0.
\end{align*}
When either condition is satisfied, the system transits from the locally stable regime to instability as $\tau_1$ increases beyond a critical value. However, when condition $2$ is satisfied, further increase in delay could lead to restabilisation of the system. This implies that the system could undergo stability switches when condition $2$ is satisfied~\cite{Cooke}. Since extensive simulations suggest the nonoccurrence of stability switches in the system, we focus only on the case when condition $1$ is satisfied and only one positive value of $\omega^2$ exists. This implies that there exists a cross over frequency at which one pair of complex conjugate roots crosses over the imaginary axis and the system transits from a locally stable to an unstable regime. The critical value of $\tau_1$, denoted by $\tau_{1,c}$, at which this transition occurs is given by
\begin{align*}
\tau_{1,c}=\frac{1}{\omega}\cos^{-1}\left(\frac{\omega^2(c-ab)-cd}{b^2\omega^2+c^2}\right).
\end{align*}
It can be analytically shown that this loss of stability occurs via a Hopf bifurcation when one pair of complex conjugate roots of \eqref{eq:characb} crosses over the imaginary axis with non-zero velocity at $\tau_1=\tau_{1,c}$. Hence, we can conclude that the necessary and sufficient condition for the system \eqref{eq:modelb_1} to be locally stable about its equilibrium is
\begin{align*}
\tau_1<\frac{1}{\omega}\cos^{-1}\left(\frac{\omega^2(c-ab)-cd}{b^2\omega^2+c^2}\right).
\end{align*}
\end{proof}
This condition essentially captures the interdependence among different network parameters and Compound TCP parameters to ensure stability of the system.
\subsection{Hopf Condition}
We have seen that an increase in delay above the critical delay value prompts the system to transit from stability to instability. Varying any of the system parameters beyond the critical value can also drive the system to instability. Thus, instead of treating delay or any of the system parameters as the bifurcation parameter, we introduce an exogenous non-dimensional parameter $\kappa$ which can act as the bifurcation parameter. If $\kappa$ is varied keeping the values of the system parameters constant at their critical values, the system loses stability at $\kappa_c=1$. To show that this loss of stability occurs via a Hopf bifurcation, we proceed to verify the transversality condition of the Hopf spectrum
\cite[Chapter $11$, Theorem $1.1.$ ]{Hale} for both scenarios. To verify the transversality condition, we need to show that $\mathrm{Re}(\mathrm{d}\lambda/\mathrm{d}\kappa)\neq 0$ at $\kappa=\kappa_c$.
\begin{figure}
\centering
\begin{subfigure}{0.5\columnwidth}
\psfrag{aaaa}{\hspace{9mm}Time (seconds)}
\psfrag{cccc}{\hspace{2mm}Queue size (pkts)}
\psfrag{dddd}{\hspace{-3mm}Round trip times = 10 ms and 200 ms}
\psfrag{eeee}{\hspace{-3mm}Round trip times = 10 ms and 10 ms}
\psfrag{bbbb}{\hspace{3mm}Utilisation (\%)}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\psfrag{125}{\begin{scriptsize}$125$\end{scriptsize}}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{60}{\begin{scriptsize}$60$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\includegraphics[height=3.75in,width=2.45in,angle=-90]{plo_15_0.eps}
\caption*{\hspace{15mm}\large{(a) Buffer size = 15}}
\label{fig:longb15}
\end{subfigure}%
\vspace{1mm}
\begin{subfigure}{0.5\columnwidth}
\psfrag{aaaa}{\hspace{9mm}Time (seconds)}
\psfrag{cccc}{\hspace{2mm}Queue size (pkts)}
\psfrag{dddd}{\hspace{-3mm}Round trip times = 10 ms and 200 ms}
\psfrag{eeee}{\hspace{-3mm}Round trip times = 10 ms and 10 ms}
\psfrag{bbbb}{\hspace{3mm}Utilisation (\%)}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\psfrag{125}{\begin{scriptsize}$125$\end{scriptsize}}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{60}{\begin{scriptsize}$60$\end{scriptsize}}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\includegraphics[height=3.75in,width=2.45in,angle=-90]{plo_100_0.eps}
\caption*{\hspace{15mm}\large{(b) Buffer size = 100}}
\label{fig:longb100}
\end{subfigure}
\vspace{1mm}
\begin{subfigure}{0.5\columnwidth}
\psfrag{aaaa}{\hspace{9mm}Time (seconds)}
\psfrag{cccc}{\hspace{2mm}Queue size (pkts)}
\psfrag{dddd}{\hspace{-3mm}Round trip times = 10 ms and 200 ms}
\psfrag{eeee}{\hspace{-3mm}Round trip times = 10 ms and 10 ms}
\psfrag{bbbb}{\hspace{3mm}Utilisation (\%)}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\psfrag{125}{\begin{scriptsize}$125$\end{scriptsize}}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{60}{\begin{scriptsize}$60$\end{scriptsize}}
\psfrag{360}{\begin{scriptsize}$360$\end{scriptsize}}
\includegraphics[height=3.75in,width=2.45in,angle=-90]{plo_360_0.eps}
\caption*{\hspace{15mm}\large{(c) Buffer size = 360}}
\label{fig:longb360}
\end{subfigure}
\caption{\emph{Long-lived flows.} Two sets of 60 long-lived Compound flows over a 2 Mbps link, regulated by two edge routers, feeding into a core router with link capacity 180 Mbps.}\label{fig:long}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{0.5\columnwidth}
\psfrag{aaaa}{\hspace{9mm}Time (seconds)}
\psfrag{cccc}{\hspace{2mm}Queue size (pkts)}
\psfrag{dddd}{\hspace{-3mm}Round trip times = 10 ms and 200 ms}
\psfrag{eeee}{\hspace{-3mm}Round trip times = 10 ms and 10 ms}
\psfrag{bbbb}{\hspace{3mm}Utilisation (\%)}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\psfrag{125}{\begin{scriptsize}$125$\end{scriptsize}}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{60}{\begin{scriptsize}$60$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\includegraphics[height=3.75in,width=2.45in,angle=-90]{plo_15_200.eps}
\caption*{\hspace{15mm}\large{(a) Buffer size = 15}}
\label{fig:shortb15}
\end{subfigure}
\vspace{1mm}
\begin{subfigure}{0.5\columnwidth}
\psfrag{aaaa}{\hspace{9mm}Time (seconds)}
\psfrag{cccc}{\hspace{2mm}Queue size (pkts)}
\psfrag{dddd}{\hspace{-3mm}Round trip times = 10 ms and 200 ms}
\psfrag{eeee}{\hspace{-3mm}Round trip times = 10 ms and 10 ms}
\psfrag{bbbb}{\hspace{3mm}Utilisation (\%)}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\psfrag{125}{\begin{scriptsize}$125$\end{scriptsize}}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{60}{\begin{scriptsize}$60$\end{scriptsize}}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\includegraphics[height=3.75in,width=2.45in,angle=-90]{plo_100_200.eps}
\caption*{\hspace{15mm}\large{(b) Buffer size = 100}}
\label{fig:shortb100}
\end{subfigure}
\vspace{1mm}
\begin{subfigure}{0.5\columnwidth}
\psfrag{aaaa}{\hspace{9mm}Time (seconds)}
\psfrag{cccc}{\hspace{2mm}Queue size (pkts)}
\psfrag{dddd}{\hspace{-3mm}Round trip times = 10 ms and 200 ms}
\psfrag{eeee}{\hspace{-3mm}Round trip times = 10 ms and 10 ms}
\psfrag{bbbb}{\hspace{3mm}Utilisation (\%)}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\psfrag{125}{\begin{scriptsize}$125$\end{scriptsize}}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{60}{\begin{scriptsize}$60$\end{scriptsize}}
\psfrag{360}{\begin{scriptsize}$360$\end{scriptsize}}
\includegraphics[height=3.75in,width=2.45in,angle=-90]{plo_360_200.eps}
\caption*{\hspace{15mm}\large{(c) Buffer size = 360}}
\label{fig:shortb360}
\end{subfigure}
\caption{\emph{Long-lived and short-lived flows.} Two sets of 60 long-lived Compound flows over a 2 Mbps link, along with \emph{exponentially distributed short files}, regulated by two edge routers and feeding into a core router with capacity 180 Mbps.}\label{fig:short}
\end{figure}
\subsection*{Case I}
\indent In this scenario, the linearised system, with the non-dimensional parameter $\kappa$, becomes
\begin{align}
\label{eq:linearb_kappa}
&\dot{u}_1(t) = \kappa \Big(-a u_{1}(t)-b u_{1}(t-\tau)-c u_{2}(t-\tau)\Big),\notag\\
&\dot{u}_2(t) = \kappa \Big(-a u_{2}(t)-b u_{2}(t-\tau)-c u_{1}(t-\tau)\Big).
\end{align}
Looking for exponential solutions of \eqref{eq:linearb_kappa} we get
\begin{align}
\label{eq:characb_kappa}
\left(\lambda +\kappa a +\kappa \left(b+c\right)e^{-\lambda \tau}\right)\left(\lambda +\kappa a +\kappa \left(b-c\right)e^{-\lambda \tau}\right)=0.
\end{align}
Differentiating \eqref{eq:characb_kappa} with respect to $\kappa$, we get
\begin{align}
\label{eq:dl_dk}
\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}=\frac{-\kappa a^2-\lambda a-\lambda b e^{-\lambda \tau}-2\kappa abe^{-\lambda \tau}-\kappa \left(b^2-c^2\right)e^{-2 \lambda \tau}}{\lambda + \kappa a + \kappa b e^{-\lambda \tau}-\lambda \kappa b \tau e^{-\lambda \tau}-\kappa ^2ab\tau e^{-\lambda \tau}-\kappa ^2 \tau \left(b^2-c^2\right)e^{-2 \lambda \tau}}.
\end{align}
From \eqref{eq:characb_kappa} we get,
\begin{align}
\label{eq:exp}
e^{-\lambda \tau}= -\frac{\lambda+\kappa a}{\kappa \left(b+c\right)}.
\end{align}
Next, substituting \eqref{eq:exp} in \eqref{eq:dl_dk}, we get
\begin{align}
\label{eq:prove_real}
\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}=\frac{ \lambda}{\kappa\left(1+\lambda \tau+\kappa a \tau\right)}.
\end{align}
At $\tau=\tau_0$, $\kappa=\kappa_c$. Substituting $\lambda=j\omega_1$ in \eqref{eq:prove_real} we get
\begin{align*}
\mathrm{Re}\left(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\right)_{\lambda=j \omega_1}= \frac{\omega_1^2 \tau_0}{\kappa_c\left(\left(1+\kappa_c a \tau_0\right)^2+\left(\omega_1 \tau_0\right)^2\right)}>0.
\end{align*}
In particular, we have proved that, $\mathrm{Re}(\mathrm{d}\lambda/\mathrm{d}\kappa)>0$, which implies that the roots cross over the imaginary axis with positive velocity at $\kappa=\kappa_c$.
\subsection*{Case II}
For the second scenario, we observe that the linearised system, with the non-dimensional exogenous parameter $\kappa$ is given as
\begin{align}
\label{eq:linearb1_kappa}
&\dot{u}_1(t) = \kappa \Big(\-\mathcal{M}_{1}u_{1}(t)-\mathcal{N}_{1}u_{1}(t-\tau_{1})-\mathcal{P}_{1}u_{2}(t-\tau_{2})\Big),\notag\\
&\dot{u}_2(t) = \kappa \Big(-\big(\mathcal{M}_{2}+\mathcal{N}_{2}\big)u_{2}(t)-\mathcal{P}_{2}u_{1}(t-\tau_{1})\Big).
\end{align}
To show that system \eqref{eq:linearb1_kappa} loses stability via a Hopf bifurcation as the non-dimensional parameter $\kappa$ is increased, we need to verify the transversality of the Hopf spectrum. Note that, for any complex number $z$, $\mathrm{Re}(z)\neq 0$ if and only if $\mathrm{Re}(z^{-1})\neq 0$. Hence, for ease of analysis, we proceed to verify that $\mathrm{Re}(\mathrm{d}\lambda/\mathrm{d}\kappa)\neq 0$ at $\kappa=\kappa_c$. Looking for exponential solution of \eqref{eq:linearb1_kappa} leads us to the following characteristic equation:
\begin{align}
\label{eq:case2_kappa}
\lambda^2 + \kappa a\lambda +\kappa b\lambda e^{-\lambda \tau_1} + \kappa^2 c e^{-\lambda \tau_1} + \kappa^2 d=0.
\end{align}
Differentiating \eqref{eq:case2_kappa} with respect to $\kappa$, we get
\begin{align}
\label{eq:dl_dk_2}
\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa} = \frac{-a \lambda - b\lambda e^{-\lambda \tau_1}-2\kappa c e^{-\lambda \tau_1}-2kd}{2\lambda+ \kappa a +\kappa be^{-\lambda\tau_1}-\kappa b \lambda \tau_1 e^{-\lambda \tau_1}-\kappa^2 c \tau_1 e^{-\lambda\tau_1}}
\end{align}
From the characteristic equation \eqref{eq:case2_kappa}, we get
\begin{align}
e^{-\lambda \tau_1}=-\frac{\lambda^2+ \kappa a\lambda +\kappa^2 d}{\kappa b \lambda +\kappa^2 c}.
\end{align}
Now, substituting the value of $e^{-\lambda\tau_1}$ in \eqref{eq:dl_dk_2} and performing some algebraic manipulations, we obtain
\begin{align*}
\Bigg(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\Bigg)^{-1}=\Bigg(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\Bigg)_{1}^{-1}+\Bigg(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\Bigg)_{2}^{-1}+\Bigg(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\Bigg)_{3}^{-1},
\end{align*}
where,
\begin{align}
\label{eq:dl_dk_3}
&\Bigg(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\Bigg)_{1}^{-1} = \frac{\kappa}{\lambda}, \hspace{4ex} \Bigg(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\Bigg)_{2}^{-1} = \kappa \tau_1, \notag\\
&\Bigg(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\Bigg)_{3}^{-1} = \frac{\kappa^2 \big(\lambda^2 ab\tau_1 - \lambda^2 c\tau_1 +2\kappa \lambda bd\tau_1+\kappa^2 cd \tau_1\big)}{\lambda \Big(\lambda^2 b+\kappa^2 ac+ 2\kappa \lambda c -\kappa^2 bd\Big)}.
\end{align}
\begin{comment}
\begin{align}
\label{eq:dl_dk_3}
& \Bigg(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\Bigg)_{1}^{-1} = \frac{\kappa}{\lambda}, \notag\\
& \Bigg(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\Bigg)_{2}^{-1} = \kappa \tau_1, \notag\\
& \Bigg(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\Bigg)_{3}^{-1} = \frac{\kappa^2 \big(\lambda^2 ab\tau_1 - \lambda^2 c\tau_1 +2\kappa \lambda bd\tau_1+\kappa^2 cd \tau_1\big)}{\lambda \Big(\lambda^2 b+\kappa^2 ac+ 2\kappa \lambda c -\kappa^2 bd\Big)}.
\end{align}
\end{comment}
Recall that, at the crossover point, the system has one pair of complex conjugate roots on the imaginary axis. Hence, substituting $\lambda=j\omega$ in \eqref{eq:dl_dk_3}, we obtain $\left(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\right)_{1,\lambda=j\omega}^{-1} = \frac{\kappa}{j\omega}$, which is purely imaginary. Similarly, we see that $\left(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\right)_{2,\lambda=j\omega}^{-1} = \kappa \tau_1$ which is strictly positive. Thus, to verify that $\mathrm{Re}\left(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\right)_{\lambda=j\omega}^{-1}>0,$ verifying $\mathrm{Re}\left(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\right)_{3,\lambda=j\omega}^{-1}>0$ suffices. Now,
\begin{align}
\label{eq:dl_dk_3real}
\mathrm{Re}\Bigg(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\Bigg)_{3,\lambda=j\omega}^{-1}= \frac{2\omega^2 \kappa^3\tau_1\big(abc-c^2-bd^2\big)\big(\omega^2+\kappa^2d\big)}{4\omega^4\kappa^2 c +\big(\kappa^2 \omega ac-\omega^3 b -\kappa^2 \omega bd\big)^2}.
\end{align}
Recall that $d$ is positive. Hence, the expression $\omega^2+\kappa^2 d$ is positive. Thus, it suffices to verify that $(abc-c^2-bd^2)>0$. Substituting the values of $a,d,c$ and $d$ from \eqref{eq: abcd}, we ge
\begin{align*}
abc-c^2-bd^2 &= 2N_1N_2P_1P_2-P_1^2P_2^2-N_1M_1P_1P_2\\
&= P_1P_2\left(N_1N_2-P_1P_2\right)+N_1P_1P_2\left(N_2-M_1\right).
\end{align*}
Recall that, $M_j$, $N_j$ and $P_j$ are strictly positive for $j=1,2$. Now, it can be easily concluded that $N_1N_2>P_1P_2$. Hence, the first term in the above expression is positive. Further, we note that, with the assumption $\tau_1>>\tau_2$ and $\tau_2 \approx 0$, $N_2$ becomes unbounded but $M_1$ remains bounded. Thus, we can assume that, in this regime, $N_2>>M_1$ which ensures that $abc-c^2-bd^2>0$. Hence, we can conclude that $\mathrm{Re}\left(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\right)_{3,\lambda=j\omega}^{-1}>0$, which in turn ensures that
\begin{align*}
\mathrm{Re}\Bigg(\frac{\mathrm{d}\lambda}{\mathrm{d}\kappa}\Bigg)_{\lambda=j\omega}^{-1}>0.
\end{align*}
Thus, we observe that, the system undergoes a \emph{Hopf Bifurcation} at $\kappa=\kappa_c$ for both scenarios. This implies that the system loses stability, as the system parameters vary, leading to the emergence of limit cycles in the system dynamics. These limit cycles could in turn induce synchronisation among the Compound TCP flows which leads to periodic packet losses and makes the downstream traffic bursty.
\subsection{Simulations}
To validate our analytical insights, we simulate two scenarios in the multiple bottleneck topology: only long-lived flows, and a combination of long and short flows.
The system consists of two distinct sets of 60 Compound TCP flows, regulated by two edge routers, feeding into a common core router. The round trip time of one set of TCP flows is fixed at 10 ms, and that of the other set of flows is varied from 10 ms to 200 ms. The short-lived flows are exponentially distributed with a mean file size of 1Mb. Each edge router has a link capacity of 100 Mbps and the core router has a link capacity of 180 Mbps. In the small buffer regime, we fix the edge router buffers to 15 packets and the core router buffer size is varied from 15 packets to 100 packets. We recapitulate that in the intermediate buffer regime, the proposed buffer dimensioning rule is $C \cdot RTT/\sqrt{N}$. Using this buffer dimensioning rule, we fix the edge router buffers to be 270 packets and the core router buffer to be 360 packets.
Fig.~\ref{fig:long} shows the simulations when there are only long-lived flows in the system. It is clear that larger queue threshold leads to non-linear oscillations, in the form of limit cycles, in the queue size. Such limit cycles can have a detrimental effect on link utilisation. Fig.~\ref{fig:short} shows the simulations when there is a combination of long and short flows. Clearly, the time-delayed feedback effects of the long flows dominates the system stability and dynamics.
\section{Introduction}
\input{intro_v6.tex}
\section{Compound TCP}
\label{compound}
\input{compoundTCP.tex}
\section{ Single Bottleneck}
\label{model_a}
\input{modela_v2.tex}
\section{Multiple Bottlenecks}
\label{model_b}
\input{modelb_v2.tex}
\section{Concluding Remarks}
\label{conclusions}
\input{conclusions_v2.tex}
\input{bibliography.tex}
\newpage
\section*{Appendix I}
\input{appendix1.tex}
\end{document}
|
\section*{Appendix}
$s_0=\delta g/\delta n|_0=3n_0^{-1/3}/5+\delta v_{xc}/\delta n|_0$, $s_1=\ln(n_0)'$, $s_2=n_0''/n_0^2$, $a_4=H^2/4$, $a_3(z)=-H^2s_1/2$ and:
\begin{widetext}
\begin{eqnarray}\label{eq_11}
a_2(z,q)=-n_0s_0+\frac{H^2}{2}(-q^2+s_1^2/2-2s_1')\\
a_1(z,q)=H^2(\frac{q^2s_1^2}{2}-\frac{9s_1^3}{4}+3s_1s_2-\frac{3s_3}{4})-2(n_0s_0)'+n_0's_0\\
a_0(z,q)=n_0+q^2n_0s_0-(n_0s_0')'+\frac{H^2}{4}(q^4+9s_1^4-2q^2s_1^2+2q^2s_2-17s_1^2s_2+4s_2^2+5s_1s_3-s_4)
\end{eqnarray}
\end{widetext}
%
\bibliographystyle{apsrev4-1}
|
\section{Results}
We consider a three-dimensional system of spherical particles of diameter
$\sigma$ with three patches along the equator with an opening angle of
$2\pi/3$. The core-core interaction is pairwise repulsive, described by a
Yukawa-like potential
$V_Y=\frac{A}{\alpha}\exp{\left(-\alpha\left[r-\sigma\right]\right)}$, where
$A$ is the strength of the potential, $r$ is the distance between the particles
centers, and $\alpha^{-1}$ is the range of the interaction. The short-range
patch-patch attraction is described by an inverse Gaussian potential,
$V_G=-\varepsilon\exp(-r_p^2/\xi^2)$, where $\varepsilon$ is the strength of
the attraction, $\xi=0.1\sigma$ the width of the Gaussian, and $r_p$ is the
patch-patch distance. For simplicity, patch-particle interactions are neglected.
In order to follow the collective dynamics, we integrate the corresponding Langevin
equations for the translational and rotational motions (see
Methods for further details).
Figure~\ref{fig.diagram}(a) shows the temperature dependence of the
bonding fraction, $p_b$, defined as the fraction of patches that are bonded.
The squares are obtained by extrapolating the time-dependence of $p_b$
to its asymptotic value from the numerical data for three different system sizes, up to $L=32$,
which is constant within the error bars (see Supplementary Fig.~S2). The solid
(red) line corresponds to the equilibrium value as predicted by Wertheim's
first-order perturbation theory (details in the Methods section). While at high temperatures the numerical equilibrium
results overlap, at low temperatures the bonding fraction is
systematically lower than the equilibrium value and does not converge to unity
at zero temperature. This signals the presence of an arrested gel in the
coexistence region (see snapshots in Fig.~\ref{fig.diagram}(a) and the equilibrium
phase diagram in Fig.~\ref{fig.diagram}(b)). Note that the numerical
data are the asymptotic values, suggesting that the thermodynamic structures
are not accessible.
Far-from equilibrium the asymptotic structures depend on the dynamics. We have
considered two limiting initial configurations with the same number of
particles: A tree-like gel (single aggregate) and a fully unbonded
suspension. The former was grown by ballistic aggregation and loopless
irreversible bond formation \cite{Dias2013b} and corresponds to relaxing a
structure grown at very low temperature \cite{Sciortino2009}. The latter is equivalent to
quenching a colloidal solution from a much higher temperature. Figure~\ref{fig.valence}(a)
shows the time evolution of the fraction of particles with one ($n_1$), two
($n_2$), and three bonds ($n_3$) for both initial configurations at
$k_BT/\varepsilon=0.025$. When starting from a tree-like gel, $n_3$ increases
monotonically, while $n_1$ and $n_2$ decrease, suggesting that structural
relaxation is driven by maximizing the bonded
patches. When starting from a completely unbonded system, the kinetic
pathway is significantly different, resembling a typical aggregation process.
$n_1$ and $n_2$ initially increase due to the formation of small, chain-like,
aggregates but eventually peaks at an intermediate time, when branching and
loop closing dominate. Note that, for the two drastically different initial configurations and pathways, the distributions
of particles with one, two, and three bonds, are surprisingly similar, sufficiently long times, albeit different from the
thermodynamic (equilibrium) ones (fully bonded, $n_3=1$). This result suggests the existence of high energy barriers to relax to
the equilibrium configuration and reveals that the geometrical kinetic structures are statistically robust.
To characterize the dynamics, in the low temperature limit, we follow the convergence of $p_b$ to its asymptotic value $p_b(\infty)$
(see Fig.~\ref{fig.valence}(b)). In Fig.\ref{fig.valence}(b), we observe the size dependence of $p_b(\infty)$ for systems
ranging from $L=8$ to $L=16$. In this regime, within the coexistence region (see Fig.~\ref{fig.diagram}(b)),
$p_b$ follows a power law in time, with an exponent of $0.6\pm0.1$.
The finite-size study shows that the power law regime increases with the size of the system. The
numerical data is consistent with the same exponent for different initial
conditions (tree-like gel and fully unbonded) and temperatures in the
coexistence region (see Supplementary Fig.~S1).
\begin{figure}[t]
\begin{center} \includegraphics[width=0.9\columnwidth]{fig4.png} \\
\caption{(a) The fraction of particles in the largest cluster $P(\infty)$ and
wrapping probability $\Pi$ of the aggregate of liquid particles (with three-bonds) as
a function of time for a system of
linear size $L=16$, at $k_BT/\varepsilon=0.025$. Time is
rescaled by the Brownian time $\tau_B=\sigma^2/D_t$, where $\sigma$ is the
particle diameter and $D_t$ its diffusion coefficient.
(b) Size of the largest cluster at the percolation threshold, represented by a vertical solid
line in (a), for system sizes $L=\{8,12,16,24,32\}$. The solid line represents the theoretical slope
for random percolation where $s_{max}\sim L^{d_f}$ with $d_f=2.52$ the largest aggregate fractal
dimension at threshold.
(c) Order parameter variance $\chi$, at the percolation threshold,
for system sizes $L=\{8,12,16,24,32\}$. The solid line represents the theoretical slope
for random percolation where $\chi_{max}\sim L^{\gamma/\nu}$ with $\gamma/\nu=2.07$.
All simulations were performed at a density $0.2\sigma^{-3}$.} \label{fig.percolation}
\end{center}
\end{figure}
Previous studies for isotropic particles suggest that, within the gas-liquid coexistence region, the
relaxation dynamics is driven by the formation of individual clusters that coarsen to minimize the
surface tension \cite{Bray2003}. By contrast here, due to the strong particle-particle interaction, a
single cluster spanning the entire system is formed well before the scale-free behavior is observed
(see Supplementary Fig.~S3). This is similar to what has been observed previously in
other network fluids, namely Laponite \cite{Ruzicka2011b}. There, the heterogeneity of the
observed cluster suggests a coarsening process, where highly bonded dense
zones of the aggregate represent the liquid phase. In order to discard that possibility, we assume
that only particles with three bonds are in the liquid phase and identify the clusters of bonded
particles. We proceed to analyze their relaxation dynamics
(see a schematic representation in Fig.~\ref{fig.schematic}). Since all particles belong to the gel-like
network the growth of the liquid phase evolves through the formation of new bonds of particles with
one or two
bonds. Snapshots of the system at the bottom of Fig.~\ref{fig.schematic} illustrate this behavior. For
surface tension driven relaxation (coarsening) the characteristic size of the
clusters $l^*$ should scale as $l^*\sim t^{1/3}$. However, we observe a clearly different exponent,
as $<s>\sim t^{0.3\pm0.1}$ implies $l^*\sim t^{0.10\pm0.04}$ (considering $l\sim <s>^{1/3}$),
where $<s>$ is the mean cluster size, and suggests a different mechanism
(see Supplementary Fig.~S3 for more details). We should note that the surface tension
of network fluids is ultra-low \cite{Bernardino2012} and this is likely to render coarsening in
these systems rather ineffective.
In Fig.~\ref{fig.percolation} we report the analysis of the kinetics of
the clusters of liquid particles. Figure~\ref{fig.percolation}(a) shows
the time dependence of the fraction of particles in the largest cluster
and the wrapping probability, defined as the probability of finding a cluster that
touches opposite sides of the box in the vertical direction. We find a
percolation transition at $t/\tau_\mathrm{B}\approx 10^2$. Figures~\ref{fig.percolation}(b) and (c)
show the system-size dependence of the number of particles in the largest cluster and its variance
at the percolation threshold.
The results are consistent with power-law scaling with the exponents of the
Random Percolation universality class \cite{Xu2014,Stauffer1994}, suggesting that the relaxation
dynamics of the liquid clusters goes through a percolation transition. This could explain why the
late time relaxation dynamics of the liquid is independent of the initial condition.
In summary, we investigated the relaxation dynamics of particles with limited valence, and found a new dynamical
regime within the coexistence region. The slow thermal reversibility of the bonds induces structural
disorder that is quenched on the experimental timescale effectively hindering
the relaxation of the structure towards thermodynamic equilibrium.
The numerical data for gel aging within the coexistence region is consistent with a
universal exponent, $0.6\pm0.1(\approx2/3)$, for the evolution of the bonding
fraction, which is significantly lower than what is known for surface tension driven coarsening.
\section{Discussion}
Our study reveals a new relaxation mechanism of the gel network, where the liquid does not
coarsen by interfacial fluctuations as a result of very long lived bonds (practically irreversible)
and ultra-low interfacial tension.
The relaxation mechanism proceeds through the relaxation of the network and not through
local rearrangements of a small number of neighboring particles.
The liquid phase does not grow by the addition of clusters of nearby particles but through
geometrical relaxation events occurring over the whole network. At least initially these
geometrical relaxation events are random. The power-law relaxation is also a hint of
these collective rearrangements.
Previous studies for Lennard-Jones particles, within the solid-gas coexistence, revealed a power law time dependence
for the size of the largest cluster at very low temperatures, distinct from (surface tension driven) coarsening \cite{Midya2017}.
Their numerical data suggests that the exponent
decreases with the temperature. Although particle diffusion is also strongly suppressed by the gel-like structure,
our numerical data is consistent with a robust exponent over the range of temperatures considered.
A crucial difference from the Lennard-Jones fluid is that in the network fluid the liquid clusters are also immobile.
Finally, a comment on other low-valence systems is in order.
Numerical studies of network fluids of particles with (two) oppositely oriented dipole
moments also suggest a power-law decay of the bond correlation
function~\cite{Schmidle2013,Klapp2016}. These results, obtained by Monte
Carlo (equilibrium) simulations hint at a dynamic slowing down
driven by gel formation. It would be interesting
to investigate the dynamics of these dipolar systems and to compare their aging
to that reported here.
There is also the question of the dependence of the dynamics on the particles valence.
At low valence, crowding effects are absent. However, the network density increases
with valence~\cite{Kraft2011,Sabapathy2015,Kern2003,Dias2015} and as the valence
increases, crowding effects are expected to play a role.
Khalil~\textit{et al.} investigated structural arrest transitions in polymer
suspensions with cross linkers~\cite{Khalil2014}, a problem that may be related to mixtures of
limited valence particles. An interplay between the glass and the gel transitions
was observed depending on the polymer and linker concentrations. Future work
could address similar questions in network fluids and/or their mixtures in order to
elucidate the role of the valence.
\section{Methods}
\textbf{Model and simulations.} We consider spherical particles, all of the same size, with three identical patches distributed along the equator.
The core-core interaction is repulsive, described by a
Yukawa-like potential,
\begin{equation}
V_Y(r)=\frac{A}{\alpha}\exp{\left(-\alpha\left[r-\sigma\right]\right)}, \label{eq.Yukawa}
\end{equation}
where $\sigma$ is the diameter of the particles, $A/(\alpha k_BT)=0.25$ is the
interaction strength and $\alpha/\sigma=0.4$ the inverse screening length. The core-core interaction
is truncated at a cutoff distance $r_c=1.5\sigma$ (at $r=r_c$ the potential is $10^{-9}A/\sigma$).
The patch-patch interaction is described by an attractive inverted Gaussian potential \cite{Dias2016,Vasilyev2013},
\begin{equation}
V_G(r_p)=-\varepsilon\exp(-r_p^2/\xi^2),
\end{equation}
where $\varepsilon$ is the strength of the attraction, $\xi=0.1\sigma$ the width of the Gaussian,
and $r_p$ the patch-patch distance. This interaction is truncated also at a cutoff distance $r_{pc}=\sigma$.
For this set of parameters, we expect at most one bond per patch.
To resolve the stochastic trajectories of the particles, we integrate the Langevin
equations of motion for the translational degrees of freedom,
\begin{equation}
m\dot{\vec{v}}(t)=-\nabla_{\vec{r}} V(\vec{r})-\frac{m}{\tau_t}\vec{v}(t)+\sqrt{\frac{2mk_BT}{\tau_t}}\vec{\xi}(t), \label{eq.trans_Langevin_dynamics}
\end{equation}
and the rotational ones,
\begin{equation}
I\dot{\vec{\omega}}(t)=-\nabla_{\vec{\theta}} V(\vec{\theta})-\frac{I}{\tau_r}\vec{\omega}(t)+\sqrt{\frac{2Ik_BT}{\tau_r}}\vec{\xi}(t).\label{eq.rot_Langevin_dynamics}
\end{equation}
We use the velocity
Verlet scheme with a time step $\Delta t=1\times 10^{-5}$, in units of the Brownian time (time to diffuse one
particle diameter), and the Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS)
for efficient simulations \cite{Plimpton1995}.
$\vec{v}$ and $\vec{\omega}$
are the translational and angular velocities, $m$ and $I$ are the
mass and moment of inertia of the particles, $V$ is the pairwise potential, and $\vec{\xi}(t)$ is the stochastic term
drawn from a random distribution with zero mean.
We consider the damping time for the translational motion,
\begin{equation}
\tau_t=\frac{m}{6\pi\eta R}. \label{eq.damping}
\end{equation}
From the Stokes-Einstein-Debye
relation \cite{Mazza2007},
\begin{equation}
\frac{D_r}{D_t}=\frac{3}{4R^2}, \label{eq.coeff_rel}
\end{equation}
and thus the rotational damping time is $\tau_r=10\tau_t/3$.
\textbf{Wertheim's theory.} The thermodynamic equilibrium properties of the model were calculated using Wertheim's first order perturbation theory
(see, e.g.~\cite{Sciortino2007}). The Helmholtz free energy per particle $f$ is the sum of a reference free energy $f_{ref}$ and a
perturbation (arising from bonding) $f_b$, $\beta f = \beta f_{ref}+\beta f_b$ where $\beta=1/k_BT$. The reference potential was taken to be
the repulsive Yukawa, $V_Y(r)$, and the perturbation (that promotes bonding) the attractive patch-patch interaction, $V_G(r_p)$.
The bonding term is given by,
\begin{equation}
\label{fb}
\beta f_b=n\ln (1-p_b)-\frac{n}{2}(1-p_b),
\end{equation}
where $n$ is the number of (identical) patches ($n=3$) and $p_b$ is the bonding fraction, which in the thermodynamic limit
corresponds to the fraction of bonded patches. We approximate $f_{ref}$ by the hard sphere free energy (in the Carnahan-Starling
approximation) of a system of particles with an effective diameter $D$ calculated within the Barker-Henderson approximation \cite{Barker1967},
\begin{equation}
\label{DBH}
D=\int_0^{+\infty}\left(1-\exp\left[-\beta V_Y(r)\right]\right)dr.
\end{equation}
The probability $p_b$ was calculated using the law of mass action,
\begin{equation}
\label{lma}
p_b=n\rho\Delta(1-p_b)^2,
\end{equation}
where $\rho$ is the number density defined as the number of particles per unit volume.
$\Delta$ is the integral of the Mayer function of the patch-patch interaction,
\begin{equation}
\label{Delta}
\Delta=\frac{1}{(4\pi)^2}\int d\vec r \int d\hat r_1 \int d\hat r_2 \left[\exp\left(-\beta V_G(r_p)\right]-1\right) g_{ref}(r),
\end{equation}
where $\vec r$ is the vector between the centers of particles 1 and 2, and $\hat r_i$ is the unit vector that defines the position of a patch on particle $i$ relative to the center of that particle. $g_{ref}(r)$ is the pair correlation function of the reference system, which is approximated by the contact value of the pair correlation function of a system of hard spheres of diameter $D$: $g_{ref}(r)=g_{HS}(r=D)=(1-\eta/2)/(1-\eta)^3$,
with $\eta=\pi/6D^3\rho$.
The multiple integral (\ref{Delta}) is then reduced to a sum of simple integrals through the following steps (a simple extension of Ref.~\cite{Wertheim1986a}):
\begin{itemize}
\item [1] {Define $z=r_p\equiv |\vec r +\sigma/2(\hat r_2-\hat r_1)|$ and $x=|\vec r + \sigma/2\hat r_2|$.}
\item [2] {Simplify (\ref{Delta}) using the change of variables:
(i) from $\theta_1$ to $z^2$, in the integration over $\hat r_1$, with $z^2=
x^2+\sigma^2/4-x\sigma\cos\theta_1$; (ii) from $\theta_2$ to $x^2$, in the integration over $\hat r_2$, with
$x^2=r^2+\sigma^2/4-r\sigma\cos\theta_2$. One obtains,
\begin{widetext}
\begin{equation}
\label{Deltachv}
\Delta=\frac{\pi g_{HS}(\eta)}{\sigma^2}\int_D^{+\infty}r dr \int_{(r-\sigma/2)^2}^{(r+\sigma/2)^2} \frac{dx^2}{x}
\int_{(x-\sigma/2)^2}^{(x+\sigma/2)^2} f(z^2/\xi^2) dz^2,
\end{equation}
\end{widetext}
where $f(z^2/\xi^2)=\exp\left[-\beta V_G(z)\right]-1$.
}
\item[3] {Change the order of integration in (\ref{Deltachv}), introducing a cut-off $z<\sigma$ (consistent with the procedure adopted in the
simulations to define a bond) and distinguish the cases $D\ge \sigma$ and $D<\sigma$. After a simple but long calculation, $\Delta$ may
be expressed as a sum of simple integrals. For $D\ge \sigma$,
\begin{equation}
\label{DeltaDge0}
\Delta=4\pi \sigma^3 g_{HS}(\eta)\alpha^2\int_{-\frac{\delta}{\alpha}}^{\frac{1}{\alpha}} y f(y) F_1(y,\alpha,\delta) dy,
\end{equation}
where $\alpha=\xi/\sigma$, $\delta=(\sigma-D)/\sigma$, and
\begin{equation}
\label{ADelta}
F_1(y,\alpha,\delta)=\frac{\alpha^3 y^3}{6}+\frac{\alpha^2 y^2}{2}+\alpha y\delta\left(1-\frac{\delta}{2}\right)+\delta^2\left(\frac{1}{2}-\frac{\delta}{3}\right).
\end{equation}
For $D<\sigma$,
\begin{equation}
\label{DeltaDlt0}
\Delta=4\pi \sigma^3 g_{HS}(\eta)\alpha^2\left(\Delta_1+\Delta_2+\Delta_3\right),
\end{equation}
where,
\begin{equation}
\label{Delta1}
\Delta_1=\int_0^{\frac{1}{\alpha}}y f(y) F_1(y,\alpha,0)dy,
\end{equation}
\begin{equation}
\label{Delta2}
\Delta_2=\frac{\alpha\delta(2-\delta)}{2}\int_0^{\frac{1-\delta}{\alpha}}y^2 f(y) dy +
\int_{\frac{1-\delta}{\alpha}}^\frac{1}{\alpha} y f(y) F_2(y,\alpha,\delta) dy,
\end{equation}
with,
\begin{equation}
\label{BDelta2}
F_2(y,\alpha,\delta)=\frac{\alpha y}{2}-\frac{\alpha^3y^3}{6}-\frac{(1-\delta)^3}{3},
\end{equation}
and,
\begin{widetext}
\begin{equation}
\Delta_3=\frac{\alpha}{2}\int_0^\frac{\delta}{\alpha}y^2f(y) F_3(y,\alpha,\delta)dy+
\frac{\delta^2}{2}\left(1-\frac{2\delta}{3}\right)\int_\frac{\delta}{\alpha}^{\frac{1-\delta}{\alpha}}yf(y)dy
+
\int_{\frac{1-\delta}{\alpha}}^\frac{1}{\alpha}y f(y)F_4(y,\alpha,\delta)dy,
\label{Delta3}
\end{equation}
\end{widetext}
with,
\begin{equation}
\label{CDelta3}
F_3(y,\alpha,\delta)=1-\alpha y -(1-\delta)^2,
\end{equation}
and,
\begin{equation}
\label{EDelta3}
F_4(y,\alpha,\delta)=\frac{1}{6}(1-\alpha^3y^3)+\frac{1}{2}(1-\alpha y)(1-\delta)^2.
\end{equation}
}
\end{itemize}
The percolation line in the $\rho,T$ diagram (see the main text) is obtained by setting $p_b=\frac{1}{n-1}=\frac{1}{2}$ in Eq.~(\ref{lma}).
The coexistence line is
calculated from the equality of the pressures and chemical potentials obtained from $\beta f$. The bonding fraction $p_b$
is obtained by solving Eq.~(\ref{lma}), in single phase thermodynamic states.
For state points in the coexistence region ($k_BT/\epsilon<0.0678$ in Fig.~1 at $\rho=0.2$) $p_b$ is calculated
using the so-called lever rule: $p_b(\rho,T)=p_{b,l}+x(p_{b,v}-p_{b,l})$, with $x=(\rho_l-\rho)/(\rho_l-\rho_v)$, where $\rho_v$ and $\rho_l$ are
the coexistence gas and liquid densities at $T$, and $p_{b,v}$ and $p_{b,l}$ are the bonding fractions in the coexisting gas and liquid phases at $T$.
\begin{acknowledgments}
\textbf{Acknowledgments.} We acknowledge fruitful discussions with Emanuela Del Gado and financial
support from the Portuguese Foundation for Science and Technology (FCT) under
Contracts nos. EXCL/FIS-NAN/0083/2012, UID/FIS/00618/2013, and IF/00255/2013.
\end{acknowledgments}
|
\section{Introduction}
Extremal decomposition problems are due to well-known Lavrentiev's inequality
\begin{equation}\label{lavr}
r(a_1,D_1)\cdot r(a_2,D_2)\leq |a_1-a_2|^2
\end{equation}
where $r(a_1,D_1)$ and $r(a_2,D_2)$ are conformal (inner) radii of planar disjoint domains, $a_i\in D_i$, $i\in 1,2$.
The conformal radius plays an important role in geometric function theory. Its generalization to higher dimensional domains is known as $p$-harmonic radius introduced by Levitski\u{i} in \cite{Lev}. For $p=2$ we deal with harmonic radius which has various applications to partial differential equations (see, for example, \cite{Ban}). The applications of the extremal decomposition problems to analytic functions are numerous including distortion theorem, coefficient inequalities, polynomial inequalities and other similar problems.
Therefore it is desirable to
extend extremal decomposition problems to higher-dimensional domains. In this paper we are going to obtain inequalities of a similar type as~\ref{lavr}) for the $p$-harmonic radius.
Note that, for planar domains, the inequality (\ref{lavr}) was generalized in several directions. The notion of Robin radius (the inner radius is a particular case of it) was introduced in papers by V.N. Dubinin and his students (see \cite{Dub1} and references therein). Some extremal decomposition problems for Robin radius were considered there as well. The approach used there was essentially based on the fact that the sum of harmonic functions is again a harmonic function. In $\mathbb{R}^n$, the sum of $p$-harmonic functions is not a $p$-harmonic function in general. So it was possible only to obtain similar results for harmonic radius in \cite{DP} and for 2-harmonic Robin radius in \cite{GKP}. In \cite{W}, W.~Wang considered the $n$-harmonic radius and extended the method of the harmonic transplantation to that of the $n$-harmonic transplantation.
G.V. Kuz'mina, A.Yu. Solynin, E.G. Emel'yanov, A. Vasiliev, Ch.~Pommerenke used a different approach to study extremal decomposition problems, namely, their technique was based on the method of extremal metric (see for example \cite{Kuz,Emel,Vas,VP,Sol} and references therein).
In this paper, we also use the technique of moduli of curve families to prove theorems on extremal decomposition for the $p$-harmonic radius. To formulate our results, we need some definitions and notation.
Throughout the paper $\mathbb{R}^n$ denotes the $n$-dimensional
Euclidean space consisting of points $x=(x_1,\dots,\,x_n)$, $n\geq2$, $y=(y_1,\dots,\ y_n)\in \mathbb{R}^n$, $<x,y>=\sum_{i=1}^n x_iy_i$ is the inner product of $x$ and $y$,
and $|x|=\sqrt{x_1^2+\dots+x_n^2}$ is the length of a
vector $x\in\mathbb{R}^{n}$. For a ball and hypersphere, we introduce the following notation:
$B(a,r)=\{x\in\mathbb{R}^{n}:\,|a-x|<r\},$
$S(a,r)=\{x\in\mathbb{R}^{n}:\,|a-x|=r\}$, ~~~$a\in\mathbb{R}^{n}$, respectively. For $\tau\in \mathbb{R}$ and $a\in \mathbb{R}^n \setminus \{0\}$, we denote by
$L(a,\tau)=\{x\in \mathbb{R}^{n}:<x,a>=\tau\}$ a hyperplane perpendicular to the vector $a.$
In what follows, we need the cylindrical coordinates $[\rho,\theta, x']$ of a point $x=(x_1,...,x_n) \in \mathbb{R}^n$, $n\geq 2,$ connected with the initial coordinates by the formulas: $x_1=\rho \cos \theta$, $x_2=\rho \sin \theta$, $x'=(x_3, x_4, \dots, x_n)$.
By the rotation by an angle $\beta$, we understand the transformation: $[\rho,\theta,x']\mapsto [\rho, \theta+\beta,x']$.
If $p>1$ then the {\it $p$-Laplacian} is defined as
$$
\Delta_p u=-{\rm div} (|\nabla u|^{p-2}\nabla u),
$$
on $\mathbb{R}^n$ for $u\in C^2(\mathbb{R}^n)$. For the potential theory for the $p$-Laplacian, we refer to~\cite{HKM}
and references therein. Let $D$ be a domain in $\mathbb{R}^n$, $x_0\in D$ ($x_0\neq \infty$), $\delta(x_0)$ be Dirac delta measure or function at the point $x_0$, $\omega_n$ be the volume of the $n$-dimensional unit ball in $\mathbb{R}^n$. It is known that, in the domain $D$ with a regular boundary, there exists a generalized solution $u_D(x,x_0)\in C^{1,\alpha} (D\setminus \{x_0\})$, $\alpha>0$, of the following Dirichlet problem
$$
\left\{ \begin{array}{l}
-\Delta_p u =n \omega_n \delta(x_0)\\
u=0 \ \ \ \text{on} \ \partial D.
\end{array}\right.
$$
The function $u_D(x,x_0)$ is called {\it $p$-harmonic Green function} of the domain $D$ with a pole at the point $x_0$.
If we introduce the notation
$$
\mu_p(t)=\begin{cases}
-\log(t),\ p=n,\\ \frac{1}{\gamma}t^{-\gamma}, \ \text{where} \ \gamma=\frac{n-p}{p-1}, \ p\ne n,
\end{cases}
$$
for $t>0$ then we get from the results of \cite{KV} that the difference
$$
h_p(x,x_0)=u_D(x,x_0)-\mu_p(|x-x_0|)
$$
belongs to the class $L^{\infty}(D)$. The quantity $R_p(x_0,D)\geq 0$, such that
$$
\lim_{x\to x_0} h_p(x,x_0)=-\mu_p(R_p(x_0,D)),
$$
is called the {\it inner $p$-harmonic radius} of the domain $D$ at the point $x_0$ \cite{Lev,KV}. For a domain $D$ with not smooth boundary by the {\it inner $p$-harmonic radius} at a point $x_0\in D$, we will call the quantity
$$
R_p(x_0,D) =\sup R_p(x_0,D'),
$$
where the supremum is taken over all domains with smooth boundaries and such that $D'\subset D$. In what follows, we will call the quantity $R_p(x_0,D)$ simply {\it $p$-harmonic radius} if $p\ne 2$ and {\it harmonic radius} if $p=2$.
Non-linearity of $p$-harmonic functions and the fact that the set of conformal mappings in spaces of dimension greater than 2 is much more restricted than in the planar case make calculations of $p$-harmonic radii of extremal domains complicated. In the case when $p=n$, we may apply M\"{o}bius transformations. It is easy to see that the mapping
$$
f(x)=-\frac{a}{|a|}+\frac{2|a|(x+a)}{|x+a|^2}
$$
maps a point $a\in \mathbb{R}^n$ to the origin and the hyperplane $L(a,0)$ onto the hypershpere $S(0,1)$. According to \cite[formula (2.21)]{W} we get
\begin{equation}\label{rad1}
R_n(a,B^*)=R_n(0,B(0,1))/|f'(a)|=\left|a-\tilde{a}\right|,
\end{equation}
where $|f'(a)| := | \det Df(a)|^{1/n},$ $B^*$ is the half-space containing the point $a$ and with the boundary $L(a,0)$. By $\tilde{a}$ we denote a point symmetric to $a$ with respect to the mentioned hyperplane $\partial B^*.$
Note that if a function $y = f(x)$ conformally maps a domain $B\subset \mathbb{R}^n$ onto a domain $\tilde B\subset \mathbb{R}^n$
and $y_0 = f(x_0),$ $x_0\in B,$ then we have \cite[p. 196]{Ban}
\begin{equation}\label{harm}
R_2(\tilde B,y_0) =|f'(x_0)| R_2(B,x_0).
\end{equation}
Using the symmetry principle for harmonic
functions, it is not difficult to see \cite{DP} that the harmonic radius of the dihedral angle $B_{2k}^*=\left\{x=[\rho,\theta,x']:|\theta|<\frac{\pi}{2k}\right\}$ at the point $x_0=[t,0,0]$ is
\begin{equation}\label{rad2}
R_2(B^*_{2k},x_0)=\left(\sum\limits_{l=1}^{2k-1}(-1)^{l-1}|x_0-x_l|^{2-n}\right)^{\frac{1}{2-n}},
\end{equation}
where $x_l=[t,\pi l/k, 0], \ l=1,\ldots, 2k-1, \ k=1,2,\ldots.$
Other special cases of calculation of p-harmonic radii that we do not use here can be found in the following papers \cite{Ban}, \cite{W}.
\section{Statements of main results}
If we consider the function $\mu_2(t)=-\log(t)$ as the fundamental solution of the Laplace equation then the inequality (\ref{lavr}) can be written in the following form
$$
\log r(a_1,D_1) + \log r(a_2,D_2) \leq 2 |a_1-a_2|,
$$
or equivalently
\begin{equation}
\mu_2(r(a_1,D_1))+\mu_2(r(a_2,D_2))\geq \mu_2(r(a_1,D_1^*))+\mu_2(r(a_2,D_2^*)),
\end{equation}
where $D_1^*$ and $D_2^*$ are half-planes with a common boundary $L^*$ such that the points $a_1$, $a_2$ are symmetric with respect to $L^*$. As a corollary of the following theorem we show that Lavrentiev's inequality remains true for $p$-harmonic radii of non-overlapping domains in the Eucledean space.
\begin{thm}\label{lavr1}
Let $G$ be a domain symmetric with respect to a hyperplane $L$, $a_1\in G$, $a_2\in G$, and let points $a_1$, $a_2$ $(a_1\ne a_2)$ be symmetric with respect to $L$ as well. Then for any non-overlapping domains $D_1\subset G$, $D_2\subset G$ such that $a_1\in D_1$, $a_2\in D_2$, we have
$$
\mu_p(R_p(a_1,D_1))+\mu_p(R_p(a_2,D_2))
$$
$$
\geq \mu_p(R_p(a_1,D_1^*))+\mu_p(R_p(a_2,D_2^*)),
$$
where
$D_1^*$ and $D_2^*$ are domains
obtained by division of the G by hyperplane L, i.e. $D_1^*$ and $D_2^*$ are non-overlapping symmetric to each other with respect to $L$ domains and such that $D_1^*\cup D_2^*=G\setminus L$.
\end{thm}
\begin{cor}
Let $a_1$, $a_2$ be arbitrary points in $\mathbb{R}^n$, $D_1$, $D_2$ be non-overlapping domains in $\mathbb{R}^n$, $a_i\in D_i$, $i=1,2$. Then
$$
\mu_p(R_p(a_1,D_1))+\mu_p(R_p(a_2,D_2))
$$
$$
\geq \mu_p(R_p(a_1,D_1^*))+\mu_p(R_p(a_2,D_2^*))=2\mu_p(R_p(a_1,D_1^*)),
$$
where $D_1^*$ and $D_2^*$ are half-spaces with common boundary $L^*$ such that the points $a_1$ and $a_2$ are symmetric with respect to $L^*$. In particular, for $p=n$ by~$(\ref{rad1})$ we obtain Lavrentiev's inequality
$$
R_n(a_1,D_1)\cdot R_n(a_2,D_2)\leq |a_1-a_2|^2.
$$
\end{cor}
Theorem \ref{lavr1} also allows us to extend well-known Kufarev's inequality concerning the product of inner radii of subdomains of the unit disk \cite[Section 6]{Dub1} to the case of $n$-harmonic radius. Let $a_1$, $a_2$ be arbitrary points of the ball $B(0,1)$, $a_1\ne 0$. The inversion $y=f_1(x)=a+r^2(x-a)/|x-a|^2$ with parameters $a=a_1/|a_1|^2$, $r^2=|a|^2-1$, preserves the ball $B(0,1)$ and maps the point $a_1$ to the origin. The second inversion $z=f_2(y)=b+\rho^2(y-b)/|y-b|^2$ with parameters
$$
b=\frac{1+\sqrt{1-|f_1(a_2)|^2}}{|f_1(a_2)|^2}f_1(a_2), \ \ \rho^2=|b|^2-1,
$$
preserves $B(0,1)$ and maps the points $0$ and $f_1(a_2)$ to a pair of symmetric points with respect to the origin. Therefore, the composition
\begin{equation}\label{phi}
\psi_{a_1,a_2}(x)=f_2(f_1(x))
\end{equation}
preserves the
unit ball and maps the points $a_1$ and $a_2$ to some symmetric points $c$ and $-c$
respectively. If $a_1 = 0$ then $f_1(x) = x$. Denote by $C(a_1, a_2)$ the image of
hyperplane $< c, x >= 0$ under the mapping $\psi_{a_1,a_2}^{-1}$. It is
clear that $C(a_1, a_2)$ is a "hypersphere" (either a hyperplane or hypersphere) which is orthogonal to
the sphere $S(0, 1)$ and lies between the points $a_1$ and $a_2$.
\begin{cor}\label{cor}
Let $a_1$, $a_2$ be arbitrary points in $B(0,1)$, $D_1$, $D_2$ be non-overlapping domains in $B(0,1)$, $a_i\in D_i$, $i=1,2$. Then
$$
R_n(a_1,D_1)R_n(a_2,D_2)\leq R_n(a_1,D_1^*)R_n(a_2,D_2^*))
$$
where $D_1^*$ and $D_2^*$ are domains
obtained by division of the ball $B(0,1)$ by "hypersphere" $C(a_1,a_2)$. \end{cor}
An analogue of Kufarev's inequality for $p=2$, $n\geq 3$ was obtained in \cite[Theorem~3]{DP}. According to this result, the quantity
$$
-R_2(a_1,D_1)^{2-n}-R_2(a_2,D_2)^{2-n}
$$
attains its maximum when $D_1$ and $D_2$ are subdomains of the unit ball that are described by the following inequalities
$$
D_l=\{x\in B(0,1):\sum_{k=1}^2 (-1)^{k+l}\left(|x-a_k|^{2-n}-||a_k|x-a_k/|a_k||^{2-n}\right)>0\}, \ \ l=1,2.
$$
Figure 1 depicts the section of the domains $D_l$, $l=1,2$, by the plane $(x_1,x_2,0,0)$ for the case of $n=4$, $p=2$ with $a_1=(1/2,0,0,0)$ and $a_2=(1/3, 0,0,0)$. Also this figure shows the extremal decomposition from Corollary~\ref{cor} for $p=n=4$. The problem of obtaining an analogue of Kufarev's theorem for arbitrary $p>1$ is interesting and still open.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{pic0_2.pdf}
\caption{Example of exteremal decompositions}
\end{figure}
By applying (\ref{harm}) to the harmonic radius, we conclude that the extreme configuration of Corollary \ref{cor} is preserved for the
quantity
$$\left(|\psi'_{a_1,a_2}(a_1)|R_2(a_1,D_1)\right)^{2-n}+\left(|\psi'_{a_1,a_2}(a_2)|R_2(a_2,D_2)\right)^{2-n},$$ where
$\psi_{a_1,a_2}(x)$ is defined by (\ref{phi}).
\begin{cor}\label{cor2}
Let $a_1$, $a_2$ be arbitrary points in $B(0,1)$, $D_1$, $D_2$ be non-overlapping domains in $B(0,1)$, $a_i\in D_i$, $i=1,2$. Then
$$
\left(|\psi'_{a_1,a_2}(a_1)|R_2(a_1,D_1)\right)^{2-n}+\left(|\psi'_{a_1,a_2}(a_2)|R_2(a_2,D_2)\right)^{2-n}\geq$$
$$
\left(|\psi'_{a_1,a_2}(a_1)|R_2(a_1,D_1^*)\right)^{2-n}+\left(|\psi'_{a_1,a_2}(a_2)|R_2(a_2,D_2^*)\right)^{2-n},
$$
where $D_1^*$ and $D_2^*$ are domains
obtained by division of the ball $B(0,1)$ by "hypersphere" $C(a_1,a_2)$. \end{cor}
In the following theorem we solve the problem on extremal decomposition of a ring or cylinder with free poles belonging to a circle. For harmonic radius, this theorem is proved in \cite[Theorem 4]{DP}. Common features of the proof also remain, although we had to extend the technique of p-modules of curves families.
\begin{thm}\label{dis1}
Let $G$ be either a ring $K(\rho_1,\rho_2)=\{x\in \mathbb{R}^n:\rho_1<|x|<\rho_2\}$ or a cylinder $Z(\rho_1,\rho_2)=\{[\rho,\theta,x']\in \mathbb{R}^n:\rho_1<\rho<\rho_2\}$, $m\geq 1,$ $0\leq\rho_1\leq\rho_2\leq\infty.$ Then, for any points $a_l$ lying on the circle $O(\rho_0)=\{[\rho,\theta,x']: \rho=\rho_0\}$, $\rho_1<\rho_0<\rho_2$ and any non-overlapping domains $D_l$, $D_l\subset G$, $a_l\in D_l$, $l=0,1,\dots,m-1$, we have
$$
\sum_{l=0}^{m-1} \mu_p(R_p(a_l,D_l))\geq \sum_{l=0}^{m-1}\mu_p(R_p(a_l^*,D_l^*))=m\mu_p(R_p(a_0^*,D_0^*)).
$$
Here
$$
a_l^*=\left[\rho_0,\frac{2\pi l}{m}, 0\right] \ \ \text{and} \ \
D_l^*=G\cap \left\{[\rho,\theta, x']: \frac{\pi(2l-1)}{m}<\theta<\frac{\pi(2l+1)}{m}\right\}.
$$
\end{thm}
In particular, if $m=2k,$ $\rho_1=0,$ $\rho_2=\infty$ by (\ref{rad2}) we have the inequality \cite{DP}
$$\sum_{l=0}^{2k-1} R_2(a_l, D_l)^{2-n}\geq m \sum_{l=1}^{2k-1}
(-1)^{l+1}|a_{0}^*-a_l^*|^{2-n}.$$
In the case $m=2,$ $\rho_1=0,$ $\rho_2=\infty$ we obtain Theorem \ref{lavr1} with the additional condition that the midpoint of the segment $[a_1, a_2]$ does not belong to the union $D_1\cup D_2.$
Note that Theorem 2 is formulated for a ring or cylinder but the proof presented below can be easily extended to any domain $G$ invariant under arbitrary rotations.
\section{Background results}
Here we mean by a "curve" a Borel set $\gamma\in \mathbb{R}^n$ with $s(\gamma)>0,$ where $s(\gamma)$ is one-dimensional Hausdorf measure. We need it, for example, to apply later a dissymmetrization transformation that in a general case breaks a curve. If we want to emphasize that a curve is understood in a standard sense as a homeomorphic image of a segment or circle we will call it {\it continuous curve}. Also, when we say that a curve joins two sets $A$ and $B$ in $G$ we mean that this curve is continuous and has a representation $\gamma: [a,b] \rightarrow \mathbb{R}^n$ such that one of the of the end-points $\gamma(a)$, $\gamma(b)$ belongs to $A$ and the other to $B$, and $\gamma(t)\in G$ for $a<t<b$.
Let $\Gamma$ be a family of curves in $\mathbb{R}^n.$ Then the {\it $p$-modulus} of the curve family is the following quantity
$$
M_p(\Gamma)=\inf \int_{\mathbb{R}^n}\rho^p dx,
$$
where the inf is taken over all Borel functions $\rho: \mathbb{R}^n\rightarrow[0,\infty]$, such that the inequality $\int_{\gamma}\rho d s\geq 1$ holds for every curve $\gamma\in \Gamma$. Functions $\rho$ that satisfy the mentioned above conditions are called {\it admissible} for the curve family $\Gamma$ and the set of all such functions is denoted by ${\rm Adm} \Gamma$. In the case of a family of continuous curves above the notion of the $p$-modulus coincides with the traditional notion of $p$-modulus of families of curves.
B. Levitski\u{i} in \cite[Theorem 1]{Lev} showed a connection of the $p$-harmonic radius with the $p$-capacity.
Taking into account the equality the $p$-capacity between the $p$-modulus of a corresponding curve family (see \cite{Shlyk}) we can also define {\it $p$-harmonic radius} with the help of the following identity
\begin{equation}\label{radmod}
-\mu_p(R_p(a,D))=\lambda_n M_p(t,a,D)^{\frac{1}{1-p}}-\mu_p(t)+o(1), \ \ t\rightarrow 0,
\end{equation}
where $M_p(t,a,D)$ is the $p$-modulus of the curve family $\Gamma(t,a,D)$, which consists of all curves joining the hypersphere $S(a,t)$ and $\partial D$ in $D$, $\lambda_n=(n\omega_n)^{\frac{1}{p-1}}$, $\omega_n$-is the volume of the ball $B(0,1)$.
List here some basic properties of the $p$-modulus (see for example \cite{zbMATH03130373}):
\smallskip
1) If $\Gamma_1\subset \Gamma_2$ then $M_p(\Gamma_1)\leq M_p(\Gamma_2)$.
\smallskip
2) $M_p\left(\cup_{i=1}^{\infty}\Gamma_i\right)\leq \sum_{i=1}^{\infty}M_p(\Gamma_i).$
\smallskip
3) If $\Gamma_2$ is {\it longer} than $\Gamma_1$ which means that each curve $\gamma\in\Gamma_2$ has a subcurve belonging to $\Gamma_1$, then $M_p(\Gamma_1)\geq M_p(\Gamma_2)$.
\smallskip
4) If $\Gamma_1, \Gamma_2,\dots$ are separated and $\Gamma_i$ is longer than $\Gamma$, $i=1,2,\dots$ then $M_p(\Gamma)\geq \sum_{i=1}^{\infty}M_p(\Gamma_i)$. (Curves families $\Gamma_1, \Gamma_2,\dots$ are called {\it separated} if there exist disjoint Borel sets $E_i$ in $\mathbb{R}^n$ such that if $\gamma\in\Gamma_i$ then $\int_{\gamma}\chi_i ds=0$, where $\chi_i$ is the characteristic function of $\mathbb{R}^n\setminus E_i$).
\smallskip
5) If $\Gamma_1, \Gamma_2,\dots$ are separated curves families and $\Gamma$ is longer than $\Gamma_i$, $i=1,2,\dots$, then $M_p(\Gamma)^{1/(1-p)}\geq \sum_{i=1}^{\infty}M_p(\Gamma_i)^{1/(1-p)}$.
\begin{lem}\label{szero} Let $L$ be an arbitrary hyperplane and $\Gamma$ consist of curves $\gamma$ such that the intersection $\gamma\cap L$ has positive one-dimensional Hausdorf measure $\left(s(\gamma\cap L)>0\right)$. Then
$$
M_p(\Gamma)=0.
$$
\end{lem}
\begin{proof}
Let $\Gamma_k=\{\gamma\in \Gamma:s(\gamma\cap L)>1/k,\ k\ \text{is\ positive \ integer}\}$. In this case, $\Gamma=\cup_{k=1}^{\infty}\Gamma_k$. It is easy to verify that the function
$$
\rho(x)=\left\{\begin{array}{l}\ 1/k,\ x\in\ L,\\ \
0,\ x\notin\ L,
\end{array}\right.
$$
is admissible for $\Gamma_k$ and $\int_{\mathbb{R}^n}\rho^p dx=0.$ Then by the definition of the $p$-modulus $M_p(\Gamma_k)= 0$. By property 2, $M_p(\Gamma)=0$. The lemma is proved.
\end{proof}
Now for $m\geq 1$, we put
$$
N_k^*=\left\{[\rho,\theta,x']\in \mathbb{R}^n, \ \frac{\pi k}{m}\leq \theta \leq \frac{\pi (k+1)}{m} \right\}, \ k=0,1,\dots, 2m-1,
$$
and
$$
L_k^*=\left\{[\rho, \theta, x']\in \mathbb{R}^n:\theta=\frac{\pi k}{m}\right\}, \ \ k=0,\dots, 2m-1.
$$
We denote by $\Phi$ the group of symmetries in $\mathbb{R}^n$ consisting of the superpositions
of the reflections in hyperplanes containing $L_l^*$, $l=0,\dots,2m-1$.
\begin{lem}\label{refl}
Let $m\geq 1,$ $\Gamma_0^*$ be a family of curves $\gamma_0^*$ lying in $N_0^*$ and let $\phi_k(x)$ denote the reflection in a hyperplane containing $L_k^*$. Let $\gamma_k^*=\phi_k(\gamma_{k-1}^*)$, $k=1,\dots, 2m-1$, and $\gamma^*=\cup_{k=0}^{2m-1}\gamma_k^*$ be a curve symmetric with respect to the group $\Phi$ and consisting of $2m$ reflections of $\gamma_0^*$, $\Gamma^*$ be the family of curves $\gamma^*$. Then
$$
M_p(\Gamma^*)=(2m)^{1-p}M_p(\Gamma_0^*).
$$
\end{lem}
\begin{proof}
According to lemma \ref{szero} we may assume that the family $\Gamma_0^*$ consists of the curves $\gamma_0^*$ such that
$$
s(\gamma_0^*\cap L_0^*\cap L_1^*)=0,
$$
where $s$ is one-dimensional Hausdorf measure. Let
$$
{\rm Int} N_k^*=\left\{[\rho, \theta,x']: \frac{\pi k}{m}<\theta<\frac{\pi(k+1)}{m}\right\}, \ k=0,1,\dots, 2m-1,
$$
be the interior of the angle $N_k^*$. Denote by $f_k(z)$ the mapping from ${\rm Int} N_0^*$ onto ${\rm Int} N_k^*$ constructed by the following formula
$$
f_0(z)=z\ {\mathrm{and}}\ f_k(z)=\varphi_k (f_{k-1}(z)), \ k=1,\dots, 2m-1.
$$
Let $\rho_0^*\in {\rm Adm} \Gamma_0^*$. Then, for the function
$$
\rho(z)=\left\{\begin{array}{cl}
\rho_0^*(f_k^{-1}(z)), & z\in {\rm Int} N_k^*, \\
0, & z\in \cup_{k=0}^{2m-1}L_k^*
\end{array}\right.
$$
we have
$$
\int_{\gamma^*}\rho ds = \sum_{k=0}^{2m-1}\int_{\gamma_k^*}\rho ds = 2m\int_{\gamma_0^*}\rho_0^* ds\geq 2m,
$$
hence $\dfrac{\rho}{2m}\in {\rm Adm} \Gamma^*.$ In view of
$$
\int_{\mathbb{R}^n}\rho^p d\mu = 2m \int_{N_0^*}(\rho_0^*)^p d\mu,
$$
we get
$$
(2m)^{1-p}\int_{N_0^*}(\rho_0^*)^p d\mu = \int_{\mathbb{R}^n}\frac{\rho^p}{(2m)^p}d\mu\geq M_p(\Gamma^*).
$$
If we take an infimum then we get
$$
(2m)^{1-p}M_p(\Gamma_0^*)\geq M_p(\Gamma^*).
$$
Now we are going to show the reverse inequality. Let $\rho\in {\rm Adm} \Gamma^*$. We construct a function $\rho_0^*(z)$ by the formula
$$
\rho^*_0(z)=\sum_{k=0}^{2m-1}\rho(f_k(z)), \ \ z\in {\rm Int} N_0^*.
$$
Then
$$
\int_{\gamma_0^*}\rho_0^*ds = \sum_{k=0}^{2m-1}\int_{\gamma_0^*}\rho(f_k(z))ds = \sum_{k=0}^{2m-1}\int_{\gamma_k^*} \rho ds = \int_{\gamma^*}\rho ds\geq 1.
$$
Hence
$$
M_p(\Gamma_0^*)\leq \int _{N_0^*}(\rho_0^*)^p d\mu = \int_{N_0^*}\left(\sum_{k=0}^{2m-1}\rho(f_k(z))\right)^p d\mu
$$
$$
\leq (2m)^{p-1}\sum_{k=0}^{2m-1}\int_{N_0^*}(\rho(f_k(z)))^p d\mu
$$
$$
=(2m)^{p-1}\sum_{k=0}^{2m-1}\int_{N_k^*} \rho^p d\mu =(2m)^{p-1}\int_{\mathbb{R}^n}\rho^p d\mu.
$$
To get the second line here we have applied the following inequality for the mean values
$$
\left(\frac{t_0^p+t_1^p+\dots +t_{2m-1}^p}{2m}\right)^{\frac{1}{p}}\geq \frac{t_0+t_1+\dots+t_{2m-1}}{2m},
$$
where $p>1$ and $t_j\geq 0$, $j=0,\dots,2m-1$, are non-negative numbers.
Taking an infimum over $\rho$ we get
$$
M_p(\Gamma_0^*)\leq (2m)^{p-1}M_p(\Gamma^*).
$$
Lemma is proved.
\end{proof}
Now we are going to use the dissymmertrization (see \cite[p. 32] {zbMATH00797804}). We introduce a symmetric structure $\{P_l\}_{l=1}^N$ in $\overline{\mathbb{R}}^n$ as a collection of closed angles $P_l=\{[\rho, \theta, x'] : \theta_{l_1}\leq \theta \leq \theta_{l_2}\}$, $l=1,\dots, N$, satisfying the conditions:
$aP) \bigcup\limits_{l=1}^N P_l = \overline{\mathbb{R}}^n, \ \ \sum\limits_{l=1}^N (\theta_{l_2}-\theta_{l_1})=2\pi,$
$bP)\ \{\phi(P_l)\}_{l=1}^N=\{P_l\}_{l=1}^N$ for any isometry $\phi\in \Phi$.
Recall that the rotation by an angle $\beta$ is the transformation: $[\rho,\theta,x']\mapsto [\rho, \theta+\beta,x']$. We call a collection of rotations $\{\alpha_l\}_{l=1}^{N}$ the {\it dissymmetrization} of the symmetric structure $\{P_l\}_{l=1}^{N}$ if the images $S_l=\alpha_l(P_l)$ satisfy the following conditions:
$aS) \bigcup\limits_{l=1}^{N} S_l = \overline{\mathbb{R}}^n$,
$bS)$ for every non-empty intersection $S_l\cap S_p$, $l,p = 1,\dots, N$, there exists
an isometry $\phi \in \Phi$ such that $\phi(\alpha_l^{-1}(S_l\cap S_p))=\alpha_p^{-1}(S_l\cap S_p)$.
Let $A$ be an arbitrary subset of $\overline{\mathbb{R}}^n$. We introduce the notation
$$
{\rm Dis} A = \bigcup_{k=1}^N \alpha_l (A\cap P_l).
$$
We also need the following lemma originally proved by Dubinin in the planar case (see for example \cite[Lemma 4.2]{Dub1}).
\begin{lem}\label{Dub}
Let $m\geq 1,$ $0\leq\theta_0<\theta_1<\ldots<\theta_{m-1}<2\pi, \theta_m=\theta_0+2\pi,$ $\Lambda_l=
\{[\rho,\theta,x']\in\mathbb{R}^n:\theta=\theta_l\}$ and $\Lambda^*_l=
\{[\rho,\theta,x']\in\mathbb{R}^n:\theta=2\pi l/m\},\ l=0,...,m.$ Then there exists a symmetric structure $\{P_k\}_{k=1}^N$, $N\geq m$, and a dissymmetrization $\{\alpha_k\}_{k=1}^N$ such that ${\rm Dis} \Lambda_l^*=\Lambda_l,$ $l=0,\ldots,m-1$.
\end{lem}
The proof of this lemma practically does not differ from the one in the planar case, so we omit it. In the following lemma we show that dissymmetrization preserves the module of a curve family.
\begin{lem}\label{dis}
If $\Gamma$ is a curve family in $\mathbb{R}^n$ and ${\rm Dis} \Gamma=\{{\rm Dis} \gamma:\gamma\in \Gamma\}$ is the result of the dissymmetrization of the family $\Gamma$ then
$$
M_p(\Gamma)=M_p({\rm Dis}\Gamma).
$$
\end{lem}
\begin{proof}
Let $\{P_l\}_{l=1}^{N}$ be a symmetric structure and $\{\alpha_l\}_{l=1}^N$ be its dissymmetrization, $\alpha_l(P_l)=S_l$. According to Lemma \ref{szero}, we can assume that the curves family $\Gamma$ consists of curves $\gamma$ satisfying the condition $s\left(\cup_{l=1}^N(\partial P_l \cap\gamma)\right)=0.$ Hence ${\rm Dis}\Gamma$ consists of curves ${\rm Dis}\gamma$ satisfying the similar condition $s\left(\cup_{l=1}^N(\partial S_l \cap {\rm Dis}\gamma)\right)=0.$ Denote by ${\rm Int} P_l$ the set $P_l\setminus \partial P_l$ and by ${\rm Int} S_l$ the set $S_l\setminus \partial S_l$.
If $\rho$ is an admissible function for the curve family $\Gamma$ then
$$
\tilde{\rho}(x)=\left\{\begin{array}{l}\rho(\alpha_l^{-1}x), \ x\in {\rm Int} S_l, \ l=1,\dots,N,\\ 0, \ x\in \cup_{l=1}^N(\partial S_l ),
\end{array}\right.
$$
is admissible for ${\rm Dis}{\Gamma}$. Indeed, for ${\rm Dis}{\gamma}\in {\rm Dis}{\Gamma}$ we get
$$ \int_{{\rm Dis}\gamma }\tilde{\rho} ds=\sum_{l=1}^N\int_{{\rm Dis}{\gamma}\cap{\rm Int}{S}_l} \tilde \rho ds=\sum_{l=1}^{N}\int_{\gamma\cap {\rm Int}{P}_l}\tilde {\rho}ds\geq 1.
$$
It is easy to see that
$$
\int_{\mathbb{R}^n}\tilde{\rho}^p dx =\sum_{l=1}^{N}\int_{{\rm Int}{S}_l}\tilde {\rho}^p dx = \sum_{l=1}^{N}\int_{{\rm Int}{P}_l}\rho^p dx =\int_{\mathbb{R}^n}\rho^p dx.
$$
By the definition of the $p$-modulus, we get
$$
M_p({\rm Dis}{\Gamma})\leq \int_{\mathbb{R}^n}\rho^p dx.
$$
Taking an infimum over all admissible functions $\rho$ we get $M_p({\rm Dis}{\Gamma})\leq M_p({\Gamma})$.
Similarly, any admissible for the curve family ${\rm Dis}{\Gamma}$ function $\tilde{\rho}(x)$ induces an admissible for ${\Gamma}$ function $\rho(x)$, moreover, again $\int_{\mathbb{R}^n}\tilde{\rho}^p dx=\int_{\mathbb{R}^n}\rho^p dx$. Therefore $M_p({\Gamma})\leq M_p({\rm Dis}{\Gamma})$. Lemma is proved.
\end{proof}
\section{Proofs}
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{pic1_4.pdf}
\caption{Families $\Gamma_1$ and $\Gamma^*$}
\end{figure}
\begin{proof}[\bf Proof of theorem 1.] As above, $\Gamma(t,a,D)$ denotes the family of all curves joining $S(a,t)$ and $\partial D$ in $D$. Let $\Gamma_1=\Gamma(t,a_1,D_1^*)$, $\Gamma^*=\{\gamma\cup\gamma^*: \gamma\in \Gamma_1, \gamma^* \ \text{is a reflection of}\ \gamma \ \text{in} \ L\}$ (see Figure 2). Consider a curve $\tilde{\gamma} \in \Gamma^*$ generated by a curve $\gamma\in \Gamma_1$. Assume that ${\gamma}$ joins a point $a\in S(a_1,t)$ with a point $b\in \partial D_1^*.$ There are only two possibilities: either $b\in G\cap L$ or $b\in \partial G.$ In the first case $\tilde{\gamma}$ is a continuous curve that joins $S(a_1,t)$ with $S(a_2,t)$ in the set $G$. Hence, there is a subcurve of $\tilde{\gamma}$ that joins $S(a_1,t)$ and $\partial D_1$. In the second case (when $\gamma$ joins $S(a_1,t)$ and $\partial G$), we have either the curve $\gamma$ itself joins $S(a_1,t)$ with $\partial D_1$ or there is a point of intersection $\gamma\cap\partial D_1$. In both these cases there exists a subcurve of $\gamma\subset\tilde{\gamma}$ that joins $S(a_1,t)$ and $\partial D_1$. It means that in any case there exists a subcurve of $\tilde{\gamma}$ joining $S(a_1,t)$ and $\partial D_1$. Therefore $\Gamma^*$ is longer than $\Gamma(t,a_1,D_1)$. Similarly $\Gamma^*$ is longer than $\Gamma(t,a_2,D_2)$. By Lemma \ref{refl}, $M_p(\Gamma^*)=2^{1-p}M_p(\Gamma_1)$ ($m=1$), and by property 5,
$$
M_p(\Gamma^*)^{\frac{1}{1-p}}\geq M_p(\Gamma(t,a_1,D_1))^{\frac{1}{1-p}}+M_p(\Gamma(t,a_2,D_2))^{\frac{1}{1-p}}
$$
or, equivalently,
$$
2M_p(\Gamma_1)^{\frac{1}{1-p}}\geq M_p(\Gamma(t,a_1,D_1))^{\frac{1}{1-p}}+M_p(\Gamma(t,a_2,D_2))^{\frac{1}{1-p}}.
$$
Multiplying by $\lambda_n$ and subtract $2\mu_p(t)$, we get
$$
2(\lambda_nM_p(\Gamma_1)-\mu_p(t))\geq \sum_{i=1}^2 (\lambda_nM_p(t,a_i,D_i)-\mu_p(t)),
$$
taking a limit as $t\rightarrow 0$, we obtain by (\ref{radmod})
$$
-2\mu_p(R_p(a_1,D_1^*))\geq -\mu_p(R_p(a_1,D_1))-\mu_p(R_p(a_2,D_2)).
$$
The theorem is proved.
\end{proof}
\begin{proof}[\bf Proof of theorem 2.]
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{pic2_2.pdf}
\label{fig:fam1} \caption{Arbitrary and extremal configurations ($m=3$)}
\end{figure}
Introduce the following notation
$$
a_l=[\rho_0,\theta_l,0], \ 0\leq \theta_l<2\pi,\
\Lambda_l=\{[\rho,\theta, x']: \theta=\theta_l\}, \ l=0,\dots, m-1.
$$
Without loss of generality we will assume that $\theta_0<\theta_1<\theta_2<\dots<\theta_{m-1}$. Let
$$
T_l^+=\{[\rho,\theta,x']:\theta_l\leq\theta\leq\theta_{l+1}\},\
T_l^-=\{[\rho,\theta,x']:\theta_{l-1}\leq\theta\leq\theta_l\}
$$
for $l=1,\dots, m-1$, and
$$
T_0^+=\{[\rho,\theta,x']:\theta_0\leq\theta \leq\theta_1\}, \
T_0^-=\{[\rho,\theta,x']:\theta_{m-1}\leq\theta\leq\theta_0+2\pi\}.
$$
$$
S^+(a_l,t)=S(a_l,t)\cap T_l^+, \ \ S^-(a_l,t)=S(a_l,t)\cap T_l^-,
$$
$l=0,\dots,m-1$.
Moreover, let
$$
\Lambda^*_l=
\{[\rho,\theta,x']\in\mathbb{R}^n:\theta=2\pi l/m\},\ l=0,...,m,
$$
and, as above,
$$
L_k^*=\left\{[\rho, \theta, x']\in \mathbb{R}^n:\theta=\frac{\pi k}{m}\right\}, \ \ k=0,\dots, 2m-1,
$$
$$
N_k^*=\left\{[\rho,\theta,x']\in \mathbb{R}^n: \ \frac{\pi k}{m}\leq \theta \leq\frac{\pi (k+1)}{m} \right\}, \ k=0,1,\dots, 2m-1.
$$
Arbitrary and extremal configurations are depicted on Figure~3 for $m=3$. Denote by $\Gamma_0^*$ the family of all continuous curves $\gamma_0^*$ from $\Gamma(t,a_0^*,D_0^*)$ that join $S(a_0^*,t)\cap N_0^*$ with the boundary $\partial D_0^*$ in the set $N_0^*$ and $s(\gamma_0^*\cap L_0^*)=0$. Let $\Gamma_l^+$ ($\Gamma_l^-$) be families of all continuous curves $\gamma_l^+$ ($\gamma_l^-$) from $\Gamma(t,a_l,D_l)$ joining $S^+(a_l,t)$ ($ S^-(a_l,t)$) with $\partial D_l$ in $T_l^+$ ($T_l^-$) and $s(\gamma_l^+\cap \Lambda_l)=0$, ($s(\gamma_l^-\cap \Lambda_l)=0$). By lemma~\ref{Dub}, if $t$ is small enough, there is a dissymmetrization "moving" $\Lambda_l^*$ ($a_l^*\in \Lambda_{l}^*$) to $\Lambda_l$ such that ${\rm Dis}\, S(a_l^*,t)=S(a_l,t),$ $l=0,\dots m-1$. Such dissymmetrization is depicted on Figure~4.
As in Lemma \ref{refl}, we construct a family $\Gamma^*$ by $2m$ reflections of each curve $\gamma_0^*\in \Gamma_0^*$.
Show that ${\rm Dis} \Gamma^*$ is longer than $\Gamma_0^+$.
Let $\gamma^*\in \Gamma^*$ be a curve generated by a continuous curve $\gamma_0^*\in \Gamma_0^*$. If $\gamma_0^*$ joins $S(a_0^*,t)$ with $\partial G$ we supply it by a curve $\gamma_1\subset \partial G$ such that $\gamma_0^*\cup \gamma_1$ joins $S(a_0^*,t)$ and $L_1^*$ in $P_0^*$. Note, that if $\gamma_0^*$ joins $S(a_0^*,t)$ with $L_1^*$ then this construction is superfluous. Then we connect a point $z\in \gamma_0^*\cap S(a_0^*,t)$ with $L_0^*$ by a curve $\gamma_2\subset S(a_0^*,t).$
Let $\tilde\gamma_0^*=\gamma_1\cup\gamma^*_0\cup\gamma_2.$ Similarly as in Lemma \ref{refl}, we construct $\tilde{\gamma}^*$ for the curve $\tilde{\gamma}_0^*$. The continuous curve $\tilde{\gamma}^*$ joins successively the hyperspheres $S(a_l^*,t)$, $l=0,\dots,m-1$ so that the upper hemihypersphere $S^+(a_l^*,t)$ is connected with the lower hemihypersphere $S^-(a^*_{l+1},t)$ ($a_{m}^*=a_0^*$). Since the symmetry of $\tilde{\gamma}^*$ taking into account the property $bS$ of dissymmetrization we get that the curve ${\rm Dis} \tilde{\gamma}^*$ contains a continuous curve joining the hemihyperspheres $S^+(a_0,t)$ and $S^-(a_1,t)$ in $T_0^+$. By the conditions of the theorem, $D_0$ contains the ball $B(a_0,t)$, $D_1$ contains $B(a_1,t)$ and $D_0\cap D_1=\emptyset$. Therefore there is a subcurve $\gamma\subset {\rm Dis} \tilde{\gamma}^*$ joining $S^+(a_0,t)$ and $\partial D_0$ in the set $T_0^+$. Let $b_0$ be a point from $S^+(a_0,t)\cap \gamma$ and $b_1\in \gamma\cap \partial D_0$ be the closest point to $b_0$ of $\partial D_0$ on $\gamma$. Then the part of $\gamma$ between $b_0$ and $b_1$ contains a continuous curve joining $\partial D_0$ and $S^+(a_0,t)$ in $D_0\setminus \overline{B(a_0,t)}$. This curve is a subcurve of $\rm{Dis}\gamma^*$ and belongs to $\Gamma_0^+.$ Therefore, ${\rm Dis} \Gamma^*$ is longer than $\Gamma_0^+.$
\begin{figure}
\centering
\includegraphics[width=1\textwidth]{pic3_1.pdf}
\label{fig:diss} \caption{Dissymmetrization}
\end{figure}
Similarly, ${\rm Dis} \Gamma^*$ is longer than $\Gamma_l^+$ and $\Gamma_l^-$ for all $l=0,\dots, m-1$. The families $\Gamma_l^+$ and $\Gamma_l^-$ are separated. By property 5, Lemma \ref{dis} and Lemma \ref{refl}, we get the following inequality
$$
\left((2m)^{1-p}M_p(\Gamma_0^*)\right)^{\frac{1}{1-p}}=M_p(\Gamma^*)^{\frac{1}{1-p}}= M_p({\rm Dis} \Gamma^*)^{\frac{1}{1-p}}
$$
$$
\geq \sum_{l=0}^{m-1}\left(M_p(\Gamma_l^+)^{\frac{1}{1-p}}+M_p(\Gamma_l^-)^{\frac{1}{1-p}}\right).
$$
On the other hand, $\Gamma_l^+$ and $\Gamma_l^-$ both are longer than $\Gamma(t,x_l,D_l)$. By property~4,
$$
M_p(\Gamma(t,a_l,D_l))\geq M_p(\Gamma_l^+)+M_p(\Gamma_l^-).
$$
If $d<0,\ u\geq 0,\ v\geq 0,$ then the following inequality for means
$$
\left(\frac{u^d+v^d}{2}\right)^{\frac{1}{d}}\leq \frac{u+v}{2}
$$
holds, or equivalently
$$
u^d+v^d\geq 2^{1-d}(u+v)^d.
$$
Applying this inequality for $d=\frac{1}{1-p}$ we get
$$
2mM_p(\Gamma_0^*)^{\frac{1}{1-p}}\geq \sum_{l=0}^{m-1}\left(M_p(\Gamma_l^+)+M_p(\Gamma_l^-)\right)^{\frac{1}{1-p}}2^{1-\frac{1}{1-p}}\geq
$$
$$
2^{1-\frac{1}{1-p}}\sum_{l=0}^{m-1}M_p(\Gamma(t, a_l,D_l))^{\frac{1}{1-p}}.
$$
It can be rewritten in the following form
$$
m(2M_p(\Gamma_0^*))^{\frac{1}{1-p}}\geq\sum_{l=0}^{m-1}M_p(\Gamma(t,a_l,D_l))^{\frac{1}{1-p}}.
$$
By the principle of symmetry \cite[Lemma 5.20, p. 55]{Vu} taking into account Lemma~\ref{szero}, we get
$$
2M_p(\Gamma_0^*)=M_p(\Gamma(t,a_0^*,D_0^*)).
$$
Therefore,
$$
mM_p(\Gamma(t, a_0^*,D_0^*))^{\frac{1}{1-p}}\geq \sum_{l=0}^{m-1}M_p(\Gamma(t,a_l,D_l)).
$$
We multiply this inequality by $\lambda_n$ and subtract $m \mu_p(t)$
$$
m\left(\lambda_n M_p(\Gamma_t(a_0^*,D_0^*)^{\frac{1}{1-p}})-\mu_p(t)\right)\geq \sum_{l=0}^{m-1} \left(\lambda_n M_p(\Gamma(t,a_l,D_l))^{\frac{1}{1-p}}-\mu_p(t)\right).
$$
Taking a limit as $t\rightarrow 0$ we obtain
$$
-m \mu_p(R_p(a_0^*,D_0^*))\geq -\sum_{l=0}^{m-1}\mu_p(R_p(a_l,D_l)).
$$
Theorem is proved.
\end{proof}
This work has been supported by the Russian Science Foundation under project 14-11-00022
|
\section{Introduction}
The rarity, deeply embedded nature and large distances to massive young stellar objects (MYSOs) renders them difficult to observe and as a result their formation process, and especially the distinction from low-mass star formation, has yet to achieve consensus among astronomers.
As the number of observational studies increases we begin to find MYSOs which exhibit many of the features commonly associated with low-mass star formation, such as circumstellar disks \citep{Beltran04,Hirota14}, collimated outflows \citep{Beu02,Davis04} and jet rotation \citep{Burns15a}. Distilling the distinction between low and high mass star formation is therefore of high priority in the search for a theory of massive star formation.
\citet{Corc98} show that, in low-mass YSOs, the L$_{\rm IR}$ excess, a tracer of accretion activity, scales linearly with outflow luminosity - a relationship indicative of the typical disk-jet-outflow process ubiquitous to low-mass star formation. This relationship was found to extend unbroken from L$_{\rm bol} = 10^{-1}$ to $10^5$ L$_{\odot}$ \citep{Garatti15}, suggesting that stars in this luminosity range may be formed by a common process; that low mass star formation mechanisms can be `scaled up' to a few tens of solar masses.
In the case of low-mass stars, outflows are thought to be predominantly driven by disk-launched collimated jets, with some contribution form a disk wind \citep{Arce07}. In this scenario linear momentum from the high velocity jet seeps into the ambient gas, entraining it and producing an extended, low-velocity molecular outflow.
Whether or not the outflows of low- and high-mass stars are produced in the same way remains to be firmly established - similarities are seen \citep{Zhang02} however the often cited lower degree of collimation \citep{Wu04} in outflows from massive stars may indicate a different origin, such as a radiation driven outflow brought about by the flashlight effect in wind-blown cavities \citep{Zinnecker07}.
Uncovering the driving mechanism of outflows in MYSOs is therefore an accessible point of comparison between high-mass and low-mass star formation.
Episodic accretion is also emerging as a vital component of star formation, both for low mass \citep{Zu09,Stamatellos11} and very high mass primordial stars \citep{Hosokawa15}. The accretion history of a YSO can be inferred from its ejection history - as such, investigating outflows also exposes the accretion behaviour of MYSOs.
S255IR-SMA1 is the brightest source of molecular line emission in the S255IR star forming region (\emph{see} \citealt{Zin15}), which is sandwiched between two evolved H$\mathrm{II}$ regions, S255 and S257. There also exists a compact source of continuum emission in both millimeter \citep{Wang11,Zin15} and centimeter \citep{Zin15,Reng96}. \citet{Zin15} estimate the mass of the embedded star as M $\simeq 20$ M$_{\odot}$ based on a spectral energy distribution compiled from \citet{Oj11} and \citet{Zinchenko09}.
VLBI maser observations of S255IR-SMA1 were carried out by \citet{Goddi07} in 2005 using the Very Long Baseline Array (VLBA) which revealed the presence of clusters of 22 GHz H$_{2}$O masers, at close proximity ($<100$ AU) to the MYSO - suggesting the possible presence of a new jet ejection event aligned geometrically with bipolar molecular outflows reported by \citet{Wang11} and \citet{Zin15}. As such, S255IR-SMA1 is an ideal target to study the relationship between outflows and jets - in addition to episodic behaviour - in an MYSO.
Five years after the observations of \citet{Goddi07} we pursue new VLBI observations aimed at measuring the annual parallax, distribution and proper motions of water masers in S255IR-SMA1. We use our data to investigate the nature and driving mechanism of the ejections from S255IR-SMA1.
\section{Observations and Data Reduction}
\label{obs}
All observations were carried out using VLBI Exploration of Radio Astrometry (VERA).
Observations of S255IR-SMA1, and the continuum positional reference source J0613+1708 (separation $= 0.87^{\circ}$, PA $= 169^{\circ}$) were made simultaneously by utilising the dual-beam capabilities of VERA \citep{dual}. This removes the need to slew antennae and interpolate atmospheric phase solutions between sources - as is required for \emph{fast-switching} VLBI phase referencing. As such, dual-beam observations achieve excellent suppression of the dynamic troposphere phase contribution, which is known to be the dominant source of astrometric phase error at 22 GHz \citep{Asaki07}.
Typical observing sessions were $\sim$8 hrs long, providing $\sim$2.5 hrs on-source time and good \emph{uv}-coverage. The synthesised beam was typically $1.4\times0.9$ mas, PA $=-49^{\circ}$.
Intermittent observations of BL Lac, DA55 or 3C84 were made every 1.5 hrs for bandpass and group delay calibration.
The continuum source J0613+1708 is found in the VLBA calibrator list \citep{VLBA2} and had an unresolved $K$-band flux of $\sim$40 mJy in our observations with VERA.
Left-handed circular polarisation signals were recorded to magnetic tapes at each VERA station, sampled at 2-bit quantisation, and filtered with the VERA digital filter unit \citep{Iguchi05}.
Signal correlation was carried out using the Mitaka FX correlator \citep{Chikada}.
Individual source phase tracking centers were set to
$(\alpha, \delta)_{\mathrm{J}2000.0}=(06^{\mathrm{h}}12^{\mathrm{m}}54^{\mathrm{s}}.00640929$,
+17$^{\circ}$59'22".95890) and
$(\alpha, \delta)_{\mathrm{J}2000.0}=(06^{\mathrm{h}}13^{\mathrm{m}}36^{\mathrm{s}}.360073$,
+17$^{\circ}$08'24".94542) for S255IR-SMA1 and J0613+1708, respectively.
The total post-correlation bandwidth of 240 MHz was divided into 16 intermediate frequency (IF) channels, with 15 IFs being allocated to continuum sources and 1 IF allocated to the maser emission.
The maser IF, assuming a rest frequency of 22.235080 GHz, was correlated in `zoom band' mode giving 8 MHz bandwidth and a channel spacing of 15.63 kHz, corresponding to a velocity resolution of 0.21 km s$^{-1}$. Fourteen of the 15 IFs allocated to continuum sources had bandwidths of 16 MHz and channel spacings of 125 kHz. The 15th IF allocated to continuum sources was correlated at the same bandwidth and resolution as the maser zoom band. In order to make uniformity in the continuum source IFs the 16 MHz bandwidth IFs were halved and combined with the zoomed 8 MHz bandwidth IF, allowing all 15 IFs to be merged.
\begin{table}
\scriptsize
\caption{Summary of observations\label{obs}}
\begin{center}
\begin{tabular}{cccccc}
\hline
Epoch& Observation &Modified&Number of \\
number& date &Julian date& features\\ \hline
1& 23rd Nov 2008 & 54793 &15\\
2& ~1st Feb 2009 & 54863 &15\\
3& 18th May 2009 & 54969 &10\\
4& 28th Aug 2009 & 55072 &12\\
5& ~15th Sep 2009 $\dag$& 55089 &12&\\
6& ~27th Sep 2009 $\dag$& 55101 &14&\\
7& ~24th Oct 2009 $\dag$& 55128 &13&\\
8& 13th Dec 2009 & 55178 &14\\
9& ~28th Jan 2010 $\dag$& 55224 & -\\
10&10th Feb 2010 & 55237 &12\\
11&~4th Apr 2010 & 55290 &8\\
12&11th Aug 2010 & 55419 &9\\
\hline
\end{tabular}
\begin{tablenotes}
\item $\dag$ These epochs were not used in parallax determination.
\end{tablenotes}
\end{center}
\end{table}
\begin{table*}
\scriptsize
\vspace{-0.2cm}
\begin{center}
\caption{The general properties of H$_{2}$O masers in S255IR-SMA1 detected with VERA. \label{TAB}}
\begin{tabular}{cclcccccccc}
\hline
Maser&$V_{\rm LSR}$&Detected &$\Delta \alpha \cos \delta$&$\Delta \delta$ &$\mu_{\alpha}\cos\delta$&$\mu_{\delta}$&$\pi$\\
ID &(km s$^{-1}$)&epochs&(mas)&(mas)&(mas yr$^{-1}$)&(mas yr$^{-1}$)& mas\\
\hline
\textbf{\underline{A}} & $ 4.798 $ & 12345678*10 11 12 & 26.851 & -6.846 & $ 1.03 \pm 0.16 $ & $ -1.01 \pm 0.18 $ & $0.585 \pm 0.071$ \\
B & $ 4.166 $ & 12345678*10** & 8.894 & -15.367 & $ -1.31 \pm 0.18 $ & $ -2.83 \pm 0.16 $ & \\
C & $ 4.374 $ & 12345678*10 11 12 & 5.322 & -12.98 & $ -1.93 \pm 0.21 $ & $ -2.10 \pm 0.20 $ & \\
D & $ 5.428 $ & 12345678*10 11 12 & 4.018 & -11.924 & $ -1.27 \pm 0.20 $ & $ -0.09 \pm 0.88 $ & \\
E & $ 3.745 $ & *****678*10** & 0.604 & -11.377 & $ -2.30 \pm 0.17 $ & $ -1.45 \pm 0.15 $ & \\
F & $ 5.851 $ & ***456*8*10** & -0.87 & -6.76 & $ -1.79 \pm 0.20 $ & $ -1.70 \pm 0.20 $ & \\
\textbf{\underline{G}} & $ 0.161 $ & 12345678*10 11 12 & 0.046 & -4.794 & $ -2.95 \pm 0.16 $ & $ -0.13 \pm 0.15 $ &$0.505 \pm 0.100$\\
\textbf{\underline{H}} & $ 3.953 $ & 12345678*10 11 * & -1.019 & -3.62 & $ -1.97 \pm 0.18 $ & $ -1.55 \pm 0.19 $ &$0.502 \pm 0.125$\\
I & $ 1.846 $ & 1234******** & -4.352 & 2.585 & $ -2.73 \pm 1.38 $ & $ -0.90 \pm 0.71 $ & \\
\textbf{\underline{J}} & $ 11.322 $ & *2345678*10 11 12 & 73.987 & 131.344 & $ -1.20 \pm 0.17 $ & $ 0.49 \pm 0.17 $ &$0.603 \pm 0.077$\\
K & $ 10.275 $ & ***45678*10 * 12 & 85.739 & 152.066 & $ -1.42 \pm 0.16 $ & $ 0.05 \pm 0.15 $ & \\
\textbf{\underline{L}} & $ 12.169 $ & 12345678*10 11 12 & 88.909 & 152.196 & $ -0.95 \pm 0.17 $ & $ 0.27 \pm 0.16 $ &$0.561 \pm 0.060$\\
M & $ 11.539 $ & *****678*10 11 12 & 150.041 & 253.838 & $ -2.95 \pm 0.16 $ & $ 1.22 \pm 0.29 $ & \\
N & $ 6.476 $ & **345*78*** 12 & 275.557 & 245.008 & $ 1.83 \pm 0.19 $ & $ 1.47 \pm 0.25 $ & \\
O & $ 5.849 $ & 12********** & 249.885 & 235.729 & $ - $ & $ - $ & \\
P & $ 3.11 $ & 12********** & 43.809 & 8.626 & $ - $ & $ - $ & \\
Q & $ 3.321 $ & 12**5678**** & 40.716 & 5.044 & $ 0.51 \pm 0.16 $ & $ -2.83 \pm 0.45 $ & \\
R & $ 2.689 $ & 1*********** & 83.969 & 29.766 & $ - $ & $ - $ & \\
S & $ 3.11 $ & 1*********** & 76.449 & 22.751 & $ - $ & $ - $ & \\
T & $ 0.998 $ & *2********** & 1.52 & 39.99 & $ - $ & $ - $ & \\
U & $ 5.217 $ & 1*********** & 22.615 & -11.963 & $ - $ & $ - $ & \\
V & $ 5.212 $ & *2********** & 18.842 & -15.06 & $ - $ & $ - $ & \\
W & $ 4.585 $ & 1*********** & 16.933 & -15.294 & $ - $ & $ - $ & \\
X & $ 3.324 $ & *****6****** & -2.696 & -5.001 & $ - $ & $ - $ & \\
Y & $ 13.008 $ & *2********** & 173.378 & 264.51 & $ - $ & $ - $ & \\
\hline
Systemic &&&&& $-0.22\pm0.19$ & $0.03\pm0.25$ & $0.563 \pm 0.036$\\
\hline
\end{tabular}
\begin{tablenotes}
\item{\footnotesize{Column (2)}: Line of sight velocities are quoted as the value measured at the first detection.}
\item{\footnotesize{Column (3)}: Numbers indicate detection in the corresponding epoch, while asterisk represents non-detection.}
\end{tablenotes}
\end{center}
\end{table*}
The a-priori delay tracking models used in correlation were improved upon after basic inspection of fringe map analysis to correct for the maser offset from the delay tracking center. In this step we also refined delay tracking solutions using more accurate antenna positions utilising global positioning system (GPS) measurements which also measured atmospheric water vapour zenith delays at each station \citep{Honma08b} and include fine corrections for the Earth rotation parameters.
In total there were 12 observation epochs with intended spacing of around 3 months. In three epochs, toward the middle of the observing calendar, we experienced antenna trouble at at least one station - rendering phase referencing between the quasar and maser impossible. Data from these epochs were not used for parallax measurement, but their self-calibrated maps and fluxes contributed to proper motion and spectrum analyses.
Epoch 9 sustained a fatal error and subsequently could not be imaged, only its spectrum was useful. The observation calendar is summarised in Table~\ref{obs}.
All data were reduced using the Astronomical Image Processing System (AIPS) developed by the National Radio Astronomy Observatory (NRAO).
Data reduction utilised the \emph{inverse phase-referencing} method for VERA data, which was introduced in \citet{Imai12}. A detailed guide is given in \citet{Burns15a}. Flux calibration was performed using system temperatures and gain information recorded at each station. Next we applied the aforementioned delay tracking solutions which are source specific.
A characterised noise signal was injected into the two VERA beams to allow correction of the delay introduced by differences of hardware in the dual beam system \citep{dual}, and group delay was calibrated using the intermittent scans of BLLAC, DA55 or 3C84. What remains at this point are time-varying phase residuals attributed to atmospheric fluctuations which, at 22 GHz, are dominated by the dynamic troposphere \citep{Asaki07}. Atmospheric calibration was carried out by fringe fitting phase residuals using the emission from a strong reference water maser at solution intervals of 1 or 2 minutes. These solutions were then applied to the visibility data of J0613+1708, phase-referencing it. Imaging of J0613+1708 then gives the relative astrometric position of water masers in S255IR-SMA1 with respect to J0613+1708.
The phase solutions obtained from the reference maser channel were also applied to all other maser channels, achieving rms noise values of typically 100 - 300 mJy beam$^{-1}$.
Maser maps were produced by applying the CLEAN procedure, based on \citet{Hog74}, to emission peaks registered at a signal-to-noise cutoff of 7. Following common nomenclature, a maser `spot' refers to an individual maser brightness peak, imaged in one spectral channel, and a maser `feature' refers to a group of spots which are considered to emanate from the same physical maser cloud, thus a maser feature typically comprises of several maser spots. Maser spots are categorised into features when they are part of the same spectral feature and found within 1 mas of another spot in that feature. We define the nominal astrometric position of a maser feature by determining the flux weighted average of the brightest three spots in the feature.
\begin{figure}
\begin{center}
\includegraphics[scale=1.15]{SpectraFull1.eps}
\caption{Scalar averaged cross-power spectra of left-hand circularly polarised maser emission in S255IR-SMA1 as a function of time. Colours are arbitrary.
\label{spectra}}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=1.08]{IC2162.dat.eps}
\caption{Distributions, internal proper motions and line-of-sight velocities of H$_2$O maser features in S255IR-SMA1. The peak position of the centimeter source from \citet{Reng96} is indicated with a cross whose size indicates to the positional error in the measurement.
\label{jet}}
\end{center}
\end{figure}
\section{Results}
\subsection{General properties of masers}
In total, 25 individual maser features were detected with VERA, each comprising of multiple maser spots. The LSR velocities, detection frequency, positions and proper motions of all maser features are summarised in Table~\ref{TAB} (positions are quoted as offsets from the phase tracking center in the maser beam, \emph{see} Section 2). The time-evolution of the maser spectrum is shown in Figure~\ref{spectra}, and shows some velocity features to be stable while others are more variable.
Feature J exhibited a steady increase in flux from 4 to 82 Jy which can be seen clearly in Figure~\ref{spectra} at $v_{\rm LSR} = 11.3$ km s$^{-1}$, while conversely, Feature G at $v_{\rm LSR} = 0.2$ km s$^{-1}$ exhibited a gradual weakening. These changes in flux were not accompanied by any change in the structure of the masers in our images. As such the slow flux changes may be attributed to some physical change in the environment near S255IR-SMA1; a change in the intensity of seed photons or a change in the density or path length of the maser cloud. Presently we cannot speculate beyond this.
We observed a chance alignment where the spectral peaks of two bright maser features (Features C and G) drifted into a common velocity channel. This observational- rather than physical effect fully accounts for the flux increase seen at about $MJD=55230$ days in Figure~\ref{spectra}.
The distribution of water maser features in S255IR-SMA1 is shown in Figures~\ref{jet} and~\ref{micro}. Masers form four main groups; those located to the NE of the system which are slightly redshifted with respect to the star ($V_{\rm LSR} =+5.25$ km s$^{-1}$, based on the central velocity of the rotating core observed by \citealt{Zin15}), those in the SW which are slightly blueshifted with respect to the star, and two groups of redshifted water masers ($\sim$ 12 km s$^{-1}$) North and West of the source of the centimeter emission marked as a plus sign in Figure~\ref{jet}. We refer to these four maser groups as the `NE masers', the `SW masers', the `red-N masers' and the `red-W masers', respectively.
\subsection{Annual parallax}
\label{pie}
The annual parallax of a maser source, obtained via astrometry, can be used to measure its trigonometric distance.
Astrometric motions of maser features trace curlicue paths across the sky-plane. This motion can be separated into a linear component which arises from the proper motion of the maser with respect to the Sun, and a sinusoidal component caused by the annual parallax. Astrometric maser motions are deconstructed into these separate components by simultaneous fitting of data with a linear and a sinusoidal function, assuming a common distance. We only performed this fitting for maser features which were observable in at least 7 epochs, spanning at least one year, and that were not spatially resolved. Maser features suitable for fitting are indicated by underlined, boldface feature ID's in Table~\ref{TAB}.
We avoided using features C and D, eventhough they were present in 11 epochs, because spatial interaction of these two features, with each other, distorted the maser images - compromising astrometric accuracy.
\begin{figure}
\begin{center}
\includegraphics[scale=0.77]{2Blue_IC2162_3.dat.eps}
\caption{Zoom of the SW maser jet in S255IR-SMA1. The distributions and motions of masers delineate a U-shaped bowshock.
\label{micro}}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=1.9]{FFFeaturesBlue.eps}
\caption{Annual parallax motions of 5 maser features in S255IR-SMA1 which are coloured arbitrarily. The upper, black line shows the modelled parallax in the Right ascension direction while the grey line below shows the same for the declination direction. Small horizontal offsets were introduced for readability.
\label{squiggle}}
\end{center}
\end{figure}
Simultaneously fitting 5 maser features we measured an annual parallax of $\pi = 0.563 \pm 0.036$ mas (Figure~\ref{squiggle}), corresponding to a source distance of D $= 1.78^{+0.12}_{-0.11}$ kpc. The parallaxes of individual maser features are shown in column 8 of Table~\ref{TAB} - their self-consistency support the validity of the simultaneous fit.
\citet{Rygl10} measured the annual parallax of S255 to be $\pi = 0.628 \pm 0.027$ mas, using 6.7 GHz methanol masers and the European VLBI Network (EVN); ours and their estimates differ by more than one sigma. Though their estimate reports a higher precision it is likely that our estimate is the more accurate for several reasons: water masers are brighter and more compact, thus providing better astrometric accuracy, than 6.7 GHz methanol masers; our parallax fitting uses 8 epochs, compared to the 4 epochs in their estimate; and our estimate utilises 5 maser features, compared to the 1 in their estimate - thus we can expect a better reduction of random errors.
\subsection{Systemic and internal proper motion}
\label{intMots}
We measured the proper motions of maser features that were detected in at least 3 epochs by measuring motions relative to Feature G (the reference maser) in the self-calibrated maps. The absolute proper motion of Feature G, obtained from the parallax fitting stage outlined in Section~\ref{pie}, was then added to all other masers, thus converting proper motions relative to Feature G into absolute proper motions relative to the reference source, J0613+1708.
Errors were calculated in quadrature to include the uncertainty in the motion of Feature G.
The observed proper motions of maser features are a combination of two components; the `source systemic motion' which is a group motion common to all masers arising from the relative motion of the target and the observer, and `internal motions' which are motions of individual maser features with respect to the driving source.
In S255IR-SMA1 we were able to measure proper motions in both the NE and the SW masers.
First we averaged the motions of maser spots associated with each outflow lobe to get the individual lobe proper motions. Then we determined the systemic motion from the residual of summing the proper motions of the lobes, assuming symmetry and with equal weighting.
We did not include the red masers in this calculation as they do no appear to be associated with the symmetric outflow system.
We derive a (heliocentric) systemic motion of ($\mu_{\alpha}\cos\delta$, $\mu_{\delta}$) = ($ -0.13\pm0.20$, $-0.06\pm0.27$) mas yr$^{-1}$, where errors are the quadrature sum of the average error for each lobe, divided by $\sqrt{N}$ where $N$ is the number of features in the lobe.
Our result is consistent with the estimate of \citet{Rygl10} of ($\mu_{\alpha}\cos\delta$, $\mu_{\delta}$) = ($ -0.14\pm0.54$, $-0.84\pm1.76$) mas yr$^{-1}$, using 6.7 GHz methanol masers which are considered to be good tracers of systemic motions since they are excited radiatively at close proximity to the star. Our higher precision likely comes from the symmetry of the detected water maser kinematics and our use of many maser features. Comparing S255IR-SMA1 with nearby Perseus Arm maser sources: S252 ($\mu_{\alpha}\cos\delta$, $\mu_{\delta}$) = ($+0.02\pm0.30$, $-2.02\pm0.30$) mas yr$^{-1}$ \citep{Reid09a,Oh10} and G192.16-3.84 ($\mu_{\alpha}\cos\delta$, $\mu_{\delta}$) = ($+0.69\pm0.05$, $-1.57\pm0.15$) mas yr$^{-1}$ \citep{Shio11}, the difference is small enough to be attributed to peculiar motion with regards to Galactic rotation.
Subtracting our systemic proper motion from the measured proper motions gives the internal motions of water masers with respect to the source. These are shown in Figures~\ref{jet} and~\ref{micro}, and reveal a highly collimated bipolar jet orientated in the NE-SW direction ($\mathrm{PA} = 49^{\circ}$).
At our distance, we calculate physical sky-plane maser velocities of $v_{\rm sky} =18.74$ km s$^{-1}$. Comparing masers in the NE and SW lobes we note that $\Delta v_{\rm sky} > \Delta v_{\rm LOS}$ meaning that the 3D motion vectors are dominated by proper motions i.e. the jet moves primarily in the sky-plane (we calculate the jet's inclination as $i=86.25^{\circ}$ to the observer). The 3D velocity of each jet lobe is $v_{\rm 3D} =18.78$ km s$^{-1}$ with respect to the protostar, with a position angle of $\mathrm{PA} = 49^{\circ}$ which matches reasonably well with those of the larger molecular outflows ($\mathrm{PA} = 67^{\circ}$, \citealt{Zin15}); perpendicular to the rotating core/disk.
Masers appear to trace a shock front propagating away from the centimeter source in the SW direction. Maser-traced bowshock structures of similar appearance have been reported in the literature including the low mass YSO S106-FIR \citep{Furuya00} and also in high mass YSOs AFGL2591 \citep{Sanna12,Trinidad13} and G31.41+0.31 \citep{Moscadelli13}. Such formations are thought to arise in the shocked gas at the interface between a protostellar jet and its surroundings. We return to this topic in Section~\ref{JETMODEL}.
\begin{figure}
\begin{center}
\includegraphics[scale=0.44]{JetNoJet.eps}
\caption{J0613+1708 images from (\emph{left}) epoch 1, in which good observing conditions were seen at all VLBI stations, and in (\emph{right}) epoch 4, in which conditions were not so good. The contours correspond to -3,3,5,9,13,17,21 times the image \emph{rms} noise which are 1.8 mJy and 2.8 mJy for epochs 1 and 4 respectively.
\label{QuasarJet}}
\end{center}
\end{figure}
\begin{figure*}
\begin{center}
\includegraphics[scale=2.2]{IC2162_Goddi222.dat.eps}
\caption{Distributions of water masers observed in 2005 using the VLBA \citep{Goddi07}, and in 2010 using VERA (this work). Masers from the aforementioned works are shown as open and filled circles, respectively. The VLBA masers were shifted into the frame of the MYSO by correcting for the systemic motion over the 5 years elapsed between observations. The peak position of the centimeter source from \citet{Reng96} is indicated with a cross whose size indicates to the positional error in the measurement.
\label{GODDIROSS}}
\end{center}
\end{figure*}
\subsection{Quasar jet in J0613+1708}
The reference source J0613+1708 has a quasar jet in the North direction which seen at in K-band continuum (Figure~\ref{QuasarJet}) extending to about 2 mas from the position of the core. This jet structure is seen more clearly in X-band images from the Second VLBA Calibrator Survey: VCS2 \citep{VLBA2}. In the X-band image the jet extends about 10 mas from the core.
Non-point-like structures in the positional reference source, such as those arising from a quasar jet, can influence the 2D Gaussian fitting used in determining its position from the interferometric images - in turn degrading maser astrometry. We checked to see if such an issue affected our data.
The extent of the jet in our images depended on the rms noise suppression achieved after inverse phase referencing - which differs for each observing epoch depending on observing conditions; in some epochs the jet is clearly resolved while in others the jet is buried in noise (Figure~\ref{QuasarJet}).
An influence on the astrometric accuracy, if present, would manifest as a larger parallax fitting residuals in the Dec. direction; the direction of the jet. Yet, since the fitting residuals in R.A. and Dec. were similar (about 0.1 mas in both cases) it seems that the jet structure did not significantly degrade the astrometric accuracy of our observations.
\newpage
\section{Discussion}
\subsection{Combined VERA and VLBA view of S255IR-SMA1}
A more complete view of the maser activity in S255IR-SMA1 can be obtained by comparison with the previous observational results of \citet{Goddi07} who observed the same maser transition using the VLBA. Their observations were carried out in 2005, and ours in 2010. We combine our maser data with the positions and line of sight velocities of the 56 masers in their Table 1, corrected for the offset caused by the source motion during the 5 years between observations. The comparison is shown in Figure~\ref{GODDIROSS}.
Although individual maser features are unlikely to have survived the time between VLBA and VERA observations it is clear that the physical structures traced by the different maser groups (NE, SW and red-W masers) have persisted. Compared to the VERA only map (Figure~\ref{jet}) the combined maser distribution better samples the physical structure and reinforces the NE and SW lobes as dominant sources of maser emission. A lengthening of the jet is evident in the SW lobe, the case is not so clear in the NE lobe as no masers sample the shock front.
An expanded view of the SW jet head is given in Figure~\ref{GODDIROSSMOD} which shows jet lengthening and tentative indication of expansion in a jet-widening sense. This is difficult to quantify since the VLBA data contains two incomplete possible shock fronts, making cross identification with the VERA results difficult. We return to this topic, and that of jet widening in Section~\ref{JETMODEL}, via comparison to jet models.
From the combined VERA and VLBA view we infer a jet shell width of about 50 mas. The angular separation between the SW and NE lobes is about 400 mas, leading to a collimation degree of about 8. Taking in to account the system inclination of $i=86.25^{\circ}$ and distance of 1.78 kpc, we estimate a physical jet length of $\sim 335$ au.
Using our 3D maser velocity we calculate a dynamic age of $t_{\rm jet} \leq 130$ years. Since the jet bowshock moves faster than the gas it entrains, the maser motions represent a lower limit to the jet velocity and thus the dynamic age is an upper limit. Given that protostellar ejection are thought to follow accretion events the occurrence of a young bipolar jet implies that S255IR-SMA1 is still actively accreting mass.
\subsection{Comparison with jet- and wind-driven models}
\label{JETMODEL}
\begin{figure*}
\begin{center}
\includegraphics[scale=0.93]{IC2162_Model-and-data3combined22.eps}
\caption{\emph{Left} shows masers associated with the SW bow shock from the 2005 VLBA observations of \citet{Goddi07} (empty circles), and the 2010 VERA observations from this work (filled circles) where arrows indicate proper motions (from this worm only). Overlain is the locus of jet shell model from \citet{Ostriker01} calculated for a primary jet radius of $R_j = 6$ AU. \emph{Middle} and \emph{right} respectively show the transverse and longitudinal proper motion velocities expected for gas at distances along the jet, z, where the bowshock head is at z $=0$ AU. Functions represent material newly swept up by the propagating jet (solid line) and the average motion of the jet's outer shell (dashed line) from the model. Overlain to these are the measured transverse and longitudinal proper motions measured by VERA.
\label{GODDIROSSMOD}}
\end{center}
\end{figure*}
The elongated NE-SW morphology and bipolar outward motions of water masers in
S255IR-SMA1 suggest that they are associated with an ejection event from a MYSO.
Bipolar outflow- and jet-tracing maser observations are frequently reported in the literature (\emph{ex.} \citealt{Imai07,NY08,Moscadelli11,Torrelles14} and G236.81+1.98 in \citealt{Choi14}), such results reveal the physical shape, scale and kinematics (if measured) of the shocked gas shell at the interface between the outflowing and ambient gasses - allowing a comparison with physical models.
We compared maser distributions and proper motions in the SW maser lobe to the models of \citet{Ostriker01} and \citet{Lee01} who, in a pair of papers, describe and compare the jet- and wind-driven outflows of protostars. Figure~\ref{GODDIROSSMOD} (\emph{left}) shows that the distribution of water masers observed by VERA is remarkably well matched by the locus of the jet driven bow shock described by Equation 22 of \citet{Ostriker01}, using a jet radius of 6 AU and position angle of $49^{\circ}$ based on the maser data.
Figure~\ref{GODDIROSSMOD} (\emph{middle}) shows the predicted proper motions of masers in the direction transverse to the jet. Functions are those predicted by the models of \citet{Ostriker01} (their Equations 19 and 21), in which we adopt their values for: the isothermal sound speed $c_s = 8$ km s$^{-1}$, and the ratio of the velocity of material being ejected from the jet surface to the velocity of the propagating bowshock $\beta c_s / v_s = 0.5$. Our singular input parameter was jet radius which we set to $R_j=6$ AU based on the match to our maser data.
In the case of a bow shock, gas is forced edgeways at the head of the forward-propagating jet resulting in large transverse velocities in addition to small transverse velocities associated with the forward-propagating gas. Generally, Figure~\ref{GODDIROSSMOD} (\emph{middle}) can be understood as a plot of these transverse velocities at various distances from the head of the bow shock located at z $=0$ AU. Further from the influence of the bow shock, gas only moves transverse to the jet direction by thermal expansion - we thus see a sharp fall in the transverse velocity away from the jet head.
Our maser data show a large dispersion of transverse velocities near the jet head (Figure~\ref{GODDIROSSMOD}, \emph{middle}) as is expected from the jet-driven outflow model. In contrast, a wind-driven outflow model predicts zero transverse velocity at the jet head while the largest transverse velocities are expected about half-way along the body of the outflow (Figure 9 of \citealt{Lee01}).
Two masers show unexpectedly large transverse velocities considering their distances from the jet head (at $z = 35$ and 64 AU). Consequently an additional source of expansion, such as contribution from a disk wind, must be invoked to explain how these masers reached transverse velocities of up to 10 km s$^{-1}$. This can be reconciled if the outflow is powered by a combination of jet- and wind-driven mechanisms as opposed to a purely jet driven scenario. A similar situation was reported by \citet{Sanna12}, observing water masers in AFGL 2591.
A transverse expansion velocity of 10 km s$^{-1}$ over 5 years would produce an offset of 6 mas which is comparable to the separation between the jet-shell model and the positions of the outermost (from the jet axis) masers observed by \citet{Goddi07} (see Figure~\ref{GODDIROSSMOD}). Thus the outermost VLBA masers likely trace the same (expanding) surface as the VERA masers. The innermost masers (closer to the jet axis) represent a secondary surface. The separation of this surface from the shock front observed by VERA is 20 mas. An internal motion of 32 km s$^{-1}$ over 5 years would be required to explain these as a common shock front - far larger than the fastest measured internal motion in the jet direction ($v_{\rm z max} = 25$ km s$^{-1}$). Thus we deem the inner VLBA maser surface to be independent of the VERA maser surface.
The two surfaces may represent internal and external surfaces in a single bowshock; the inner surface may represent the jet shock while the outer surface may represent the bow shock created in front of the jet as it sweeps up ambient gas. In such a case we calculate the `thickness' of the bow shock as the current separation of the surfaces (36 AU), minus the internal motion traversed in 5 years (26 AU), giving 10 AU, which is comparable to the jet width.
Alternatively the second surface may represent a second shock front from a repeating jet. In this case, the time between ejections would be about 3 yrs, assuming constant jet velocities of $v_s = 66$ km s$^{-1}$, based on the model of \citet{Ostriker01}. The required interval between ejections is shorter than the tens to a few hundred years for which accretion bursts themselves are thought to last, making a 3 yr repeating jet scenario unfeasible. This leads us to favour the former interpretation - that masers trace two surfaces in a single jet bowshock.
Figure~\ref{GODDIROSSMOD} (\emph{right}) represents the predicted proper motions in the jet direction. Again, the water masers observed with VERA are well described by the jet-driven outflow model of \citet{Ostriker01} (Equations 18 and 20). And again data contrasts that expected for a wind-driven outflow (a linear relation between velocity and z, \emph{see} Figure 9 of \citealt{Lee01}). Hydrodynamic simulations of \citet{Lee01} show that the highest velocity gas associates with the jet shell (dashed line in our figure) while the lower velocity gas is attributed to ambient material that has been entrained. In agreement, it can be seen that masers with the highest measured longitudinal velocity lie closely to the jet outer shell. These masers are therefore shocked on the boundary between the jet and the ambient material - corroborating the spatial association of maser positions to the jet shell model seen in Figure~\ref{GODDIROSSMOD} (\emph{left}).
One maser has a far larger longitudinal velocity than that predicted by the jet model (at z = 64 AU). As before, this behaviour may be explained by invoking a wind component accompanying the jet.
\citet{Goddi07} conclude that the water masers in S255IR-SMA1 likely associate with a disk-wind. Their interpretation was based, similarly to ours, on the distribution and proper motions of water masers. Comparing our proper motions with theirs, consistency is seen in the red masers (the `Aw' cluster in their Figure 1), however they measure fewer proper motions associated with the NE-SW jet components. Our jet-tracing masers show a more systematic expansion motion and give a more complete sampling of the bow shock surface in the SW lobe.
Our analysis of the proper motions of water masers in S255IR-SMA1 support a predominantly jet-driven outflow origin, possibly including contribution from a wind. Our jet radius of $R_j = 6$ AU is similar to the 6 AU used by \citet{Sanna12} for a similar analysis of jet-tracing masers in AFGL 2591.
\subsection{Episodic ejection}
Another important feature of low mass star formation is episodic accretion in which short intense bursts of accretion onto the central star perforate periods of inactivity lasting a few thousand years - accretion bursts are known from observations \citep{Herbig77,Hartmann96}, and simulations \citep{Zu09,Stamatellos11}.
In the case of massive primordial star formation episodic accretion bursts may play an important role in regulating the ionising radiation emitted from the embedded massive star - providing a mechanism of prolonging accretion into the UCHII region phase of massive star birth \citep{Hosokawa15}. The timescales involved in episodic accretion in very massive stars are simlilar to those of low mass stars, though the proposed purpose of episodic accretion in both cases is different.
With the exception of FU Ori stars, episodic accretion is impractical to observe due to the long timescales involved. However, the episodic nature of accretion in young stellar objects can be inferred from evidence of episodic ejection observable as symmetric pairs of ejection bow shocks at ever increasing distances from the central object (for example, in HH111 \citep{Bo89}).
At least three episodes of ejection are known to have taken place in S255IR-SMA1. The larger, and therefore older outflow was observed in $^{12}$CO$(2-1)$ and has dynamic age of about 7000 years \citep{Wang11}, and a second younger ejection indicated by bow shocks of Fe$\mathrm{II}$ line emission \citep{Wang11} which is also seen in the HCO$^+$ maps of \citet{Zin15} who estimate its dynamic age to be 1000 years. The third being our maser jet bow-shock of dynamic age $t_{dyn}\leq 130$ years.
We conclude that the 20 M$_{\odot}$ MYSO in S255IR-SMA1 has formed accompanied by three separate episodes of ejection during the last few thousand years.
\section*{Acknowledgments}
R.B. would like to acknowledge the Ministry of Education, Culture, Sports, Science and Technology (MEXT) Japan for support as part of the Monbukagakusho scholarship.
\input{mn-bib.tex}
\bibliographystyle{mn2e}
|
\section{Acknowledgements}
Most of the results presented in this paper were originally obtained in the author's master's thesis \cite{W13} in a, however, substantially less concise format.
The author is deeply grateful and indepted to his then supervisor and now PhD advisor Emo Welzl.
Special thanks go to Raimund Seidel for presenting his beautiful algorithm for counting triangulations at ETH Z\"urich in fall 2012, which undoubtedly marked the beginning of the success of that thesis.
\section{Three Simple Applications}\label{sec:applications}
We present three generic kinds of combination problems.
They directly correspond to the three types of crossing-free combinations depicted earlier in \cref{fig:combinations}.
\begin{comment}
Here, a set $\units$ of units on $\pts$ is called \emph{simple} if it is a
subset of $\frags$.
If $\units$ is simple then for each $\unit$ in $\units$ the underlying
fragment is defined naturally as $\fragof{\unit} \coloneqq \unit$.
\end{comment}
\subsection{All Crossing-free Combinations}\label{subsec:crossingfree}
The main aim of this subsection is to give a proof of \cref{thm:pg}.
Still, some of the following insights are fairly general and can be used in many different settings.
Let us fix a set of units $\units \subseteq \frags$.
We make the explicit assumption that $\cmbscf{\units}$, the set of crossing-free combinations of $\units$, is serializable.
Then, as follows, a combination $\cmb$ in $\cmbscf{\units}$ can be described by a coloring of the point set $\pts$ with three colors $\inlinepoint{free}{unmarked},\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked}$ and a special marking, e.g., $\inlinepoint{alive}{marked}$, on one of the points.
A point $\pt \in \pts$ is given the color $\inlinepoint{dead}{unmarked}$ if $\pt \in \clow{\cmb}$, it is given the color $\inlinepoint{alive}{unmarked}$ if $\pt \in \cpts{\cmb} \setminus \clow{\cmb}$, and it is given the color $\inlinepoint{free}{unmarked}$ in all other cases.
The special marking is put on the left-most point of $\umaxin{\cmb}$.
\Cref{fig:encoding_crossingfree} shows, for the special case $\units = \segs$, that different crossing-free combinations can have identical such descriptions.
Whenever that is the case, we consider two combinations equivalent.
More formally, we put $\cmb \parteq \cmb'$ if and only if\footnote{Of course, the third condition only makes sense if $\cmb$ and $\cmb'$ are non-empty.}
\label{def:cf:eq}
\begin{itemize}
\item $\clow{\cmb} = \clow{\cmb'}$,
\item $\cpts{\cmb} \setminus \clow{\cmb} = \cpts{\cmb'} \setminus \clow{\cmb'}$,
\item $\ulft{\umaxin{\cmb}} = \ulft{\umaxin{\cmb'}}$.
\end{itemize}
\begin{figure}[b]
\begin{center}
\input{fig/encoding_crossingfree}
\end{center}
\caption{
Three elements of $\cmbscf{\segs}$ that are considered equivalent.
}
\label{fig:encoding_crossingfree}
\end{figure}
The goal now is to prove that $\parteq$ is coherent, as in \cref{def:coherent}.
Unfortunately, this endeavor is doomed to fail because for some contrived choices of $\units$ we can have $\ctarrow{\cmb}{\cmb'}{\unit}$ and, at the same time, $\cmb \parteq \cmb'$.
In the language of combination graphs this means that we would have to introduce loops, which leads to a potentially infinite number of source-sink paths.
To avoid this problem, we require for any $\cmb,\cmb' \in \cmbscf{\units}$ and any $\unit \in \units$ that $\ctarrow{\cmb}{\cmb'}{\unit}$ implies $\cmb \not\parteq \cmb'$.
If this additional requirement is met, then we say that $\cmbscf{\units}$ is \emph{progressive}.
\begin{lemma}\label{lem:cf:coherent}
Let $\units \subseteq \frags$ with $\cmbscf{\units}$ both serializable and progressive.
Then, the equivalence relation $\parteq$ on $\cmbscf{\units}$, as defined above, is coherent.
\end{lemma}
\begin{proof}
Let $\cmb_1,\cmb_2 \in \cmbscf{\units}$ be non-empty (otherwise, the proof is trivial) with $\cmb_1 \parteq \cmb_2$ and assume that $\ctarrow{\cmb_1}{\cmb_1'}{\unit}$ holds for $\cmb_1' \in \cmbscf{\units}$ and $\unit \in \units$.
Consider $\cmb_2' \coloneqq \cmb_2 \cup \{\unit\}$.
We show $\cmb_2' \in \cmbscf{\units}$, $\cmb_1' \parteq \cmb_2'$ and $\ctarrow{\cmb_2}{\cmb_2'}{\unit}$, which implies the lemma.
By assumption we have $\clow{\cmb_1} = \clow{\cmb_2}$ and $\cpts{\cmb_1} \setminus \clow{\cmb_1} = \cpts{\cmb_2} \setminus \clow{\cmb_2}$.
From this and from the fact that $\unit$ is extreme in $\cmb_1'$, we first derive that $\unit$ is also extreme in $\cmb_2'$.
So, assume the opposite, that is, either $\upts{\unit} \cap \clow{\cmb_2} \neq \emptyset$ or $\uupp{\unit} \cap \cpts{\cmb_2} \neq \emptyset$ holds.
In the first case, we get $\upts{\unit} \cap \clow{\cmb_1} = \upts{\unit} \cap \clow{\cmb_2} \neq \emptyset$, contradicting that $\unit$ is extreme in $\cmb_1'$.
In the second case, we get at least one of $\uupp{\unit} \cap \cpts{\cmb_1} \supseteq \uupp{\unit} \cap (\cpts{\cmb_1} \setminus \clow{\cmb_1}) = \uupp{\unit} \cap (\cpts{\cmb_2} \setminus \clow{\cmb_2}) \neq \emptyset$ and $\uupp{\unit} \cap \clow{\cmb_1} = \uupp{\unit} \cap \clow{\cmb_2} \supseteq \uupp{\unit} \cap (\cpts{\cmb_2} \cap \clow{\cmb_2}) \neq \emptyset$.
Both possibilities again lead to $\unit$ not being extreme in $\cmb_1'$.
We conclude that, indeed, $\unit$ is extreme in $\cmb_2'$.
Since $\unit$ is extreme in $\cmb_2'$, no other element of $\cmb_2'$ depends on $\unit$.
By \cref{obs:dependence}, since $\cmb_2$ is crossing-free, also $\cmb_2'$ must be crossing-free.
Hence, $\cmb_2' \in \cmbscf{\units}$.
Next, by definition of $\cmb_2'$ it is easily seen that $\clow{\cmb_1'} = \clow{\cmb_2'}$ and $\cpts{\cmb_1'} \setminus \clow{\cmb_1'} = \cpts{\cmb_2'} \setminus \clow{\cmb_2'}$ both hold.
To prove also $\ulft{\umaxin{\cmb_1'}} = \ulft{\umaxin{\cmb_2'}}$, and consequently $\cmb_1' \parteq \cmb_2'$, it suffices to show $\umaxin{\cmb_2'} = \unit$.
So, for $\unit' \coloneqq \umaxin{\cmb_2'}$ and assuming that $\unit' \neq \unit$, we get $\unit \utorgt \unit'$ and thus also $\umaxin{\cmb_2} = \unit'$.
Invoking $\cmb_1 \parteq \cmb_2$ yields $\unit \utorgt \umaxin{\cmb_1}$, which contradicts the assumption $\unit = \umaxin{\cmb_1'}$.
It remains to prove $\ctarrow{\cmb_2}{\cmb_2'}{\unit}$.
The crucial observation here is that $\unit = \umaxin{\cmb_2'}$ and $\cmb_2 \cup \{\unit\} = \cmb_2'$ alone are not sufficient.
What we need instead is $\cmb_2 = \cmb_2' \setminus \{\unit\}$, that is, $\unit$ must not be contained in $\cmb_2$.
However, assuming that $\unit$ is contained in $\cmb_2$, we easily derive $\cmb_1 \parteq \cmb_2 = \cmb_2' \parteq \cmb_1'$, which, when combined with $\ctarrow{\cmb_1}{\cmb_1'}{\unit}$, contradicts the fact that $\cmbscf{\units}$ is progressive.
\end{proof}
We define $\target \coloneqq \quotient{\cmbscf{\units}}{\parteq}$.
With \cref{lem:cf:coherent} we see that $(\units,\cmbscf{\units},\parteq,\target)$
is a combination problem as long as $\cmbscf{\units}$ is serializable and progressive.
The corresponding combination graph over $\units$ is referred to by
$\agraphcf{\units}$.
\begin{corollary}\label{cor:crossingfree}
If $\units \subseteq \frags$ is such that $\cmbscf{\units}$ is both serializable and progressive, then $\agraphcf{\units}$ represents $\cmbscf{\units}$ and the size of $\agraphcf{\units}$ is at most $O(|\quotient{\cmbscf{\units}}{\parteq}| \cdot |\units|)$.
\end{corollary}
\Cref{thm:pg} follows by invoking \cref{cor:crossingfree} with $\units = \segs$ and by making use of the next
two lemmas.
The first one essentially shows that $\cmbscf{\segs}$ is a well-behaved set.
The second one gives us a better upper bound on the size of \mbox{$\agraphcf{\segs}$.
Observe,} however, that a bound of $O(3^nn^3)$ is immediate because we can encode equivalence classes with 3 colors and with one marking, and because of $|\segs| = O(\ptsnum^2)$.
\begin{restatable}{lemma}{cfserializable}
\label{lem:cf:serializable}
For any point set $\pts$, $\cmbscf{\segs}$ is both serializable and progressive.
\end{restatable}
\begin{proof}[Proof (serializable)]
Let $\cmb \in \cmbscf{\segs}$ be arbitrary but non-empty.
We prove serializability by exhibiting a right-most extreme element $\unit$ in $\cmb$.
For this, consider the relative interiors of all segments in $\cmb$.
These segments without endpoints are convex and pairwise non-intersecting.
For a set of convex and non-intersecting shapes in the plane it is well known that at least one of them can be translated in the $y$-direction to infinity in a continuous motion, without intersecting any other segment in the process \cite{T85}.
Any segment with this property corresponds to an extreme element in $\cmb$.
The extreme elements in $\cmb$ now can be ordered from left to right according to their relative positions when projected orthogonally onto the $x$-axis.
The segment $\unit$ on the far right is a right-most extreme element in $\cmb$.
\end{proof}
\hspace*{-\parindent}%
\begin{minipage}{0.8\textwidth}
\begin{proof}[Proof (progressive)]
For the sake of contradiction, assume that we have $\ctarrow{\cmb'}{\cmb}{\unit}$ and $\cmb \parteq \cmb'$ for some $\cmb,\cmb' \in \cmbscf{\segs}$ and $\unit \in \segs$.
Let $\unit' \coloneqq \umaxin{\cmb'}$ (note that the case where $\cmb'$ is empty is trivial).
We have $\ulft{\unit} = \ulft{\unit'}$ and we have to distinguish the two cases illustrated on the right.
In case 1, we see that $\urgt{\unit} \in \cpts{\cmb} \setminus \clow{\cmb}$
but also $\urgt{\unit} \not\in \cpts{\cmb'} \setminus \clow{\cmb'}$ since
otherwise $\unit'$ would not be extreme in $\cmb'$.
In case 2, we see that $\urgt{\unit'} \in \clow{\cmb}$ but also
$\urgt{\unit'} \not\in \clow{\cmb'}$ since otherwise $\unit'$ would not be
extreme in $\cmb'$.
Hence, in both cases we get a contradiction to the assumption $\cmb \parteq \cmb'$.
\end{proof}
\end{minipage}
\begin{minipage}{0.2\textwidth}
\centering
\input{fig/progressive}
\end{minipage}
\vspace{\belowdisplayskip}
\begin{restatable}{lemma}{cfbound}
\label{lem:cf:bound}
For any set $\pts$ of $\ptsnum$ points, the relation $\parteq$ partitions $\cmbscf{\segs}$ into at most $O(\alpha^nn)$ equivalence classes, that is, $|\quotient{\cmbscf{\segs}}{\parteq}| = O(\alpha^nn)$, where $\alpha \lessapprox 2.83929$.\footnote{We write $\alpha \lessapprox \beta$ if a parameter $\alpha$ is approximately equal to $\beta$ and also strictly smaller.}
\end{restatable}
The proof of \cref{lem:cf:bound} is a bit tedious.
We only sketch the main idea here, and postpone a more careful analysis to \cref{sec:bounds}.
Note that for any three consecutive points $\pt_i,\pt_{i+1},\pt_{i+2}$ in $\pts$, the point $\pt_{i+1}$ is either below or above the straight line through $\pt_i$ and $\pt_{i+2}$, as depicted in \cref{fig:upperbound_crossingfree}.
\begin{figure}[b]
\begin{center}
\input{fig/upperbound_crossingfree}
\end{center}
\caption{
Three consecutive points can always be colored in a way that does not correspond to an element of $\cmbscf{\segs}$.
}
\label{fig:upperbound_crossingfree}
\end{figure}
In both cases we can show that at least one of the $3^3 = 27$ different ways of assigning colors to $\pt_i,\pt_{i+1},\pt_{i+2}$ does not describe an actual element of $\cmbscf{\segs}$.
A bound of $O(26^{n/3}n) = O(2.963^nn)$ on the size of $\quotient{\cmbscf{\segs}}{\parteq}$ then follows after partitioning $\pts$ into $n/3$ consecutive triples.
In the first case, if we assign colors $\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked}$, there must be two distinct segments which pass over the points $\pt_i$ and $\pt_{i+2}$, respectively, and which pass under $\pt_{i+1}$.
Clearly, any two such segments are crossing.
In the second case, if we assign colors $\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked}$, there must be a segment that passes over $\pt_{i+1}$ and under $\pt_i$ and $\pt_{i+2}$.
Clearly, only a bent ``segment'' can achieve that.
Note also that if the combination graph $\agraphcf{\segs}$ is constructed bottom-up, as will be explained in more detail in \cref{sec:construction}, then all these impossible colorings are avoided automatically.
\subsection{Crossing-free Partitions}\label{subsec:partitions}
\begin{figure}[b]
\begin{center}
\input{fig/encoding_partition}
\end{center}
\caption{
Three elements of $\protect\cmbslpt{\cparts}$ that are considered equivalent.
}
\label{fig:encoding_partition}
\end{figure}
Let us again fix a set of units $\units \subseteq \frags$.
We define the set $\cmbslpt{\units}$ of all crossing-free combinations $\cmb$ of $\units$
for which the sets $\upts{\unit}$ of all $\unit \in \cmb$ are pairwise disjoint and
for which $\clow{\cmb} \subseteq \cpts{\cmb}$ holds.
\Cref{fig:encoding_partition} depicts three of these combinations for the special case $\units = \cparts$.
Observe that $\cmbspt{\units}$ and $\cmbslpt{\units}$ are different sets, and also observe that we have in fact $\cmbspt{\units} \subseteq \cmbslpt{\units}$.
Assume that $\cmbslpt{\units}$ is serializable.
Similar to the previous subsection, we use two colors $\inlinepoint{free}{unmarked},\inlinepoint{alive}{unmarked}$ and a special marking $\inlinepoint{alive}{marked}$ on the points in $\pts$ to describe an element $\cmb$ of $\cmbslpt{\units}$.
A point $\pt$ receives the color $\inlinepoint{alive}{unmarked}$ if $\pt \in \cpts{\cmb}$, and $\inlinepoint{free}{unmarked}$ otherwise.
The marking is again put on the left-most point of $\umaxin{\cmb}$.
If two combinations have identical such descriptions, we consider them equivalent.
Formally, we put $\cmb \parteq \cmb'$ if and only if
\begin{itemize}
\item $\cpts{\cmb} = \cpts{\cmb'}$,
\item $\ulft{\umaxin{\cmb}} = \ulft{\umaxin{\cmb'}}$.
\end{itemize}
One peculiarity in \cref{fig:encoding_partition} is that some points have been given the color $\inlinepoint{alive}{unmarked}$ even though there are no incident segments.
This is because the set $\cparts$ also contains all isolated points, that is, all $\unit \in \frags$ with $|\upts{\unit}| = 1$.
Also observe that $\ctarrow{\cmb}{\cmb'}{\unit}$ implies $\cmb \not\parteq \cmb'$ because $\cpts{\cmb}$ is always a proper subset of $\cpts{\cmb'}$.
Consequently, an explicit notion of ``progressive'', as in the previous subsection, is not needed here.
\begin{lemma}\label{lem:pt:coherent}
Let $\units \subseteq \frags$ with $\cmbslpt{\units}$ serializable.
Then, the equivalence relation $\parteq$ on $\cmbslpt{\units}$, as defined above, is coherent.
\end{lemma}
\begin{proof}
Let $\cmb_1,\cmb_2 \in \cmbslpt{\units}$ be non-empty (otherwise, the proof is trivial) with $\cmb_1 \parteq \cmb_2$ and assume that $\ctarrow{\cmb_1}{\cmb_1'}{\unit}$ holds for $\cmb_1' \in \cmbslpt{\units}$ and $\unit \in \units$.
Consider $\cmb_2' \coloneqq \cmb_2 \cup \{\unit\}$.
We show $\cmb_2' \in \cmbslpt{\units}$, $\cmb_1' \parteq \cmb_2'$ and $\ctarrow{\cmb_2}{\cmb_2'}{\unit}$, which implies the lemma.
First, observe that since $\unit$ is extreme in $\cmb_1'$, it is also extreme in $\cmb_2'$.
Indeed, $\upts{\unit} \cap \clow{\cmb_2} \subseteq \upts{\unit} \cap \cpts{\cmb_2} = \upts{\unit} \cap \cpts{\cmb_1} = \emptyset$ and $\uupp{\unit} \cap \cpts{\cmb_2} = \uupp{\unit} \cap \cpts{\cmb_1} = \emptyset$ can be derived immediately.
By \cref{obs:dependence}, $\cmb_2'$ is crossing-free.
Moreover, deriving $\clow{\cmb_2'} = \clow{\cmb_2} \cup \ulow{\unit} \subseteq \cpts{\cmb_2} \cup \clow{\cmb_1'} \subseteq \cpts{\cmb_2'} \cup \cpts{\cmb_1'} = \cpts{\cmb_2'}$ proves that, indeed, $\cmb_2' \in \cmbslpt{\units}$.
By making use of the marking in the same way as in the proof of \cref{lem:cf:coherent} we get $\umaxin{\cmb_2'} = \unit$, and it follows that $\cmb_1' \parteq \cmb_2'$ and $\ctarrow{\cmb_2}{\cmb_2'}{\unit}$.
\end{proof}
We define the set $\target \subseteq \quotient{\cmbslpt{\units}}{\parteq}$
which contains all equivalence classes $\partclass{\cmb}$ for which
$\cpts{\cmb} = \pts$ holds.
Observe that then $\bigcup \target = \cmbspt{\units}$.
From \Cref{lem:pt:coherent} it follows that $(\units,\cmbslpt{\units},\parteq,\target)$ is a combination
problem provided that $\cmbslpt{\units}$ is serializable.
We denote by $\agraphpt{\units}$ the corresponding combination graph.
\begin{corollary}\label{cor:partitions}
If $\units \subseteq \frags$ is such that $\cmbslpt{\units}$ is serializable, then $\agraphpt{\units}$ represents $\cmbspt{\units}$ and the size of $\agraphpt{\units}$ is at most $O(|\quotient{\cmbslpt{\units}}{\parteq}| \cdot |\units|)$.
\end{corollary}
\Cref{thm:cp,thm:pm} now follow from \cref{cor:partitions} and the following
two lemmas.
\begin{restatable}{lemma}{ptserializable}
For any point set $\pts$ and any subset $\units$ of $\cparts$, it holds that $\cmbslpt{\units}$ is serializable.
\end{restatable}
\begin{proof}
The proof for the existence of right-most extreme elements is analogous to the first part of the proof of \Cref{lem:cf:serializable}.
However, proving that $\cmb' \coloneqq \cmb \setminus \{\unit\}$, where $\unit = \umaxin{\cmb}$, is also an element of $\cmbslpt{\units}$ is not completely trivial since we have to verify that $\clow{\cmb'} \subseteq \cpts{\cmb'}$ holds.
For the sake of contradiction, let us assume that there exists a point $\pt \in \clow{\cmb'} \setminus \cpts{\cmb'}$.
From $\cmb' \subseteq \cmb$ we get $\clow{\cmb'} \subseteq \clow{\cmb}$ and hence also $\pt \in \clow{\cmb}$.
We now use $\clow{\cmb} \subseteq \cpts{\cmb}$, which holds by definition of $\cmbslpt{\units}$, to obtain $\pt \in \cpts{\cmb}$.
By combining this with the assumption $\pt \not\in \cpts{\cmb'}$ we obtain $\pt \in \upts{\unit}$ because $\cpts{\cmb} \setminus \cpts{\cmb'} = \upts{\unit}$.
It follows that $\clow{\cmb'} \cap \upts{\unit}$ is non-empty since it contains at least the point $\pt$.
This, however, contradicts the fact that $\unit$ is extreme in $\cmb$.
\end{proof}
Note that in the case of convex partitions, the general bound on the size of the resulting combination graph from \cref{cor:partitions} is insufficient to prove \Cref{thm:cp} because $\units = \cparts$ alone can be of size $\Omega(2^\ptsnum)$.
The following lemma is therefore really needed.
\begin{restatable}{lemma}{ptbound}
\label{lem:pt:bound}
For any point set $\pts$ of size $\ptsnum$ and any subset $\units$ of $\cparts$, the size of $\agraphpt{\units}$ is at most $O(2^nn^3)$.
\end{restatable}
\begin{proof}
We prove that the number of labeled edges in $\agraphpt{\units}$ can be bounded by $O(2^nn^3)$ from above, which then implies the lemma.
Fix $\pt_l, \pt_r \in \pts$ with $\pt_l \ptspreceq \pt_r$, and let $\pts_l^r \coloneqq \{\pt_l,\pt_{l+1},\dots,\pt_{r-1},\pt_r\} \subseteq \pts$ be the set of points between $\pt_l$ and $\pt_r$.
Observe that there are at most $2^{r-l+1}$ convex parts $\unit \in \cparts$ for which $\pt_l = \ulft{\unit}$ and $\pt_r = \urgt{\unit}$ holds.
Let us also fix such a convex part $\unit$.
Next, we will give a bound on the number of edges in $\agraphpt{\units}$ with label $\unit$.
This bound will only depend on the indices $l$ and $r$.
Suppose that $\cgarrow{\cmb}{\cmb'}{\unit}$ is an edge in $\agraphpt{\units}$ and assume further that $\ctarrow{\cmb}{\cmb'}{\unit}$ holds.
Then, we easily see that $\uslow{\unit} = \cpts{\cmb} \cap \pts_l^r$, where $\uslow{\unit} \coloneqq \ulow{\unit} \setminus \upts{\unit}$.
This means that for each point $\pt \in \pts_l^r$ it is determined by $\unit$ whether $\pt \in \cpts{\cmb}$ holds or not.
For a point $\pt \in \pts \setminus \pts_l^r$ there are at most two choices, either $\pt \in \cpts{\cmb}$ or $\pt \not\in \cpts{\cmb}$.
It follows that there are at most $2^{n-(r-l+1)}n$ many vertices $\partclass{\cmb}$ with an outgoing edge that is labeled by $\unit$.
As usual, the additional factor $n$ comes from the special marking.
The total number of labeled edges in $\agraphpt{\units}$ can therefore be bounded by
\begin{align*}
\sum_{l=1}^{n} \sum_{r=l}^{n} 2^{r-l+1} \cdot 2^{n-(r-l+1)}n = \sum_{l=1}^{n} \sum_{r=l}^{n} 2^nn = O(2^nn^3)\text{.}
\end{align*}
\end{proof}
\subsection{Subdivisions}
Let $\units \subseteq \cpartstp$, which means in particular that the \emph{shape} of each unit $\unit$ is defined.
We define the set $\cmbslsd{\units}$ which contains all combinations $\cmb$ of $\units$ for which the following holds.
There exists an $x$-monotone polygonal chain, denoted by $\cchn{\cmb}$, which starts in $\pt_1$, ends in $\pt_\ptsnum$, has only points from $\pts$ as vertices, and satisfies the following with regard to $\cmb$.
The shapes of all $\unit$ in $\cmb$ form a subdivision of the region between $\cchn{\cmb}$ and the lower convex hull of $\pts$, by which we mean that the shapes are pairwise interior-disjoint and each point of the plane in the interior of that region is contained in the shape of at least one element $\unit$ of $\cmb$.
As depicted in \cref{fig:encoding_subdivision}, we describe such a combination $\cmb$ by giving the vertex points of $\cchn{\cmb}$ the color $\inlinepoint{alive}{unmarked}$, by giving all other points the color $\inlinepoint{free}{unmarked}$, and by adding the usual marking.
Guided by this description, we put $\cmb \parteq \cmb'$ if and only if
\begin{itemize}
\item $\cchn{\cmb} = \cchn{\cmb'}$,
\item $\ulft{\umaxin{\cmb}} = \ulft{\umaxin{\cmb'}}$.
\end{itemize}
\begin{figure}
\begin{center}
\input{fig/encoding_subdivision}
\end{center}
\caption{
Three elements of $\protect\cmbslsd{\cpartstp}$ that are considered equivalent.
}
\label{fig:encoding_subdivision}
\end{figure}
\begin{lemma}
\label{lem:sd:coherent}
Let $\units \subseteq \cpartstp$ with $\cmbslsd{\units}$ serializable.
Then, the equivalence relation $\parteq$ on $\cmbslsd{\units}$, as defined above, is coherent.
\end{lemma}
\begin{proof}
Let $\cmb_1,\cmb_2 \in \cmbslsd{\units}$ be non-empty (otherwise, the proof is trivial) with $\cmb_1 \parteq \cmb_2$ and assume that $\ctarrow{\cmb_1}{\cmb_1'}{\unit}$ holds for $\cmb_1' \in \cmbslsd{\units}$ and $\unit \in \units$.
Consider $\cmb_2' \coloneqq \cmb_2 \cup \{\unit\}$.
We show $\cmb_2' \in \cmbslsd{\units}$, $\cmb_1' \parteq \cmb_2'$ and $\ctarrow{\cmb_2}{\cmb_2'}{\unit}$, which implies the lemma.
From the definition of $\cmbslsd{\units}$, it is immediate that also $\cmb_2' \in \cmbslsd{\units}$ and $\cchn{\cmb_2'} = \cchn{\cmb_1'}$.
Since $\unit$ is extreme in $\cmb_1'$, it is also extreme in $\cmb_2'$.
Using the usual argument involving the marking we see that $\umaxin{\cmb_2'} = \unit$, it follows that $\cmb_1' \parteq \cmb_2'$ and $\ctarrow{\cmb_2}{\cmb_2'}{\unit}$.
\end{proof}
Let $\target \subseteq \quotient{\cmbslsd{\units}}{\parteq}$ be the set that contains $\partclass{\cmb}$ if and only if $\cchn{\cmb}$ is equal to the upper convex hull of $\pts$.
Observe that $\bigcup \target = \cmbssd{\units}$.
If $\units$ is such that $\cmbslsd{\units}$ is serializable, then $(\units,\cmbslsd{\units},\parteq,\target)$ is a combination problem and we denote by $\agraphsd{\units}$ the corresponding combination graph.
\begin{corollary}\label{cor:subdivisions}
If $\units \subseteq \cpartstp$ is such that $\cmbslsd{\units}$ is serializable, then $\agraphsd{\units}$ represents $\cmbssd{\units}$ and the size of $\agraphsd{\units}$ is at most $O(|\quotient{\cmbslsd{\units}}{\parteq}| \cdot |\units|)$.
\end{corollary}
\Cref{thm:cs,thm:tr} now follow from \cref{cor:subdivisions} and the following two lemmas.
\begin{restatable}{lemma}{sdserializable}
For any point set $\pts$ and any subset $\units$ of $\cpartstp$, it holds that $\cmbslsd{\units}$ is serializable.
\end{restatable}
\begin{proof}
The proof for the existence of right-most extreme elements is again analogous to \Cref{lem:cf:serializable}.
Moreover, it is not hard to see that $\cmb' \coloneqq \cmb \setminus \{\unit\}$, where $\unit = \umaxin{\cmb}$, is also an element of $\cmbslsd{\units}$.
Simply observe that the upper convex hull of the shape of $\unit$ must be contained in $\cchn{\cmb}$, which means that $\cchn{\cmb'}$ is obtained from $\cchn{\cmb}$ by replacing the upper hull of $\unit$ with its lower hull.
\end{proof}
Similar to the previous subsection, the next lemma is not implied by the general bound from \cref{cor:subdivisions}.
\begin{restatable}{lemma}{sdbound}
\label{lem:sd:bound}
For any point set $\pts$ of size $\ptsnum$ and any subset $\units$ of $\cpartstp$, the size of $\agraphsd{\units}$ is at most $O(2^nn^3)$.
\end{restatable}
\begin{proof}
Analogous to the proof of \cref{lem:pt:bound}.
\end{proof}
\section{Definitions and Results}\label{sec:basics}
Let $\frags$ be the set of all crossing-free geometric graphs on all non-empty subsets of $\pts$.
Elements of $\frags$ are called \emph{units}, and they can be thought of as the simple ``pieces'' from the informal introduction.
For every unit $\unit \in \frags$ we denote by $\upts{\unit} \subseteq \pts$ the set of vertex points of $\unit$.
That is, if $\unit$ is a geometric graph on a particular subset $\pts'$ of $\pts$, then $\upts{\unit} = \pts'$.
Let us now define a number of useful subsets of $\frags$.
\begin{itemize}
\item Let $\segs$ be the set of \emph{segments} with both endpoints in $\pts$.
That is, each $\unit \in \segs$ is a geometric graph on exactly two points of $\pts$ with the edge between them.
\item Let $\cparts$ be the set of \emph{convex parts} with vertex points in $\pts$.
That is, for each $\unit \in \cparts$ the convex hull of $\upts{\unit}$ does not contain any points of $\pts \setminus \upts{\unit}$ (in words, interior points are also vertex points of $\unit$).
Moreover, $\unit$ contains all the edges along the boundary of the convex hull of the set $\upts{\unit}$.
Observe that all isolated points and segments with endpoints in $\pts$ are also elements of $\cparts$.
\item Let $\cpartstp$ be the set of \emph{convex faces} with vertex points in $\pts$. That is, $\cpartstp$ contains all $\unit \in \cparts$ with $|\upts{\unit}| \geq 3$ and no interior vertices.
The \emph{shape} of any such $\unit$ is the bounded and closed region delimited by its edges.
\item Let $\trgs$ be the set of \emph{empty triangles} with vertex points in $\pts$.
That is, $\trgs$ contains all $\unit \in \cpartstp$ with $|\upts{\unit}| = 3$.
\end{itemize}
Let us fix a subset $\units$ of $\frags$.
For any units $\unit_1,\unit_2 \in \units$, if $\unit_1$ contains an edge $e_1$ and $\unit_2$ contains a different edge $e_2$ such that $e_1$ and $e_2$ are crossing, then we also say that $\unit_1$ and $\unit_2$ are \emph{crossing}.
Otherwise, $\unit_1$ and $\unit_2$ are \emph{non-crossing}.
A \emph{combination} of $\units$ is a subset $\cmb$ of $\units$, and $\cmb$ is \emph{crossing-free} if the elements of $\cmb$ are pairwise non-crossing.
We denote by $\cmbscf{\units}$ the set of all crossing-free combinations of $\units$.
Furthermore, a combination $\cmb \in \cmbscf{\units}$ is called a \emph{partition} of $\pts$ if the sets $\upts{\unit}$ of all $\unit \in \cmb$ are pairwise disjoint and if their union is equal to $\pts$.
We denote by $\cmbspt{\units}$ the set of all such partitions of $\pts$.
Assuming $\units \subseteq \cpartstp$, a combination $\cmb \in \cmbscf{\units}$ is called a \emph{subdivision} of the convex hull of $\pts$ if the shapes of all $\unit \in \cmb$ are pairwise interior-disjoint and if their union is equal to the convex hull of $\pts$.
We denote by $\cmbssd{\units}$ the set of all such subdivisions of the convex hull of $\pts$.
\Cref{fig:combinations} shows one representative of each type of crossing-free combinations for the special case $\units = \trgs$.
\begin{figure}
\begin{center}
\input{fig/combinations}
\end{center}
\caption{
The three defined types of crossing-free combinations of $\trgs$.
The shaded regions represent the shapes of individual elements of $\trgs$.
}
\label{fig:combinations}
\end{figure}
From now on we will no longer consider the sets $\pgs$, $\cps$, and so on, as defined in the introduction.
Instead, we will talk about crossing-free combinations of specific sets of units.
For example, since there is an obvious bijection between the sets $\cmbscf{\segs}$ and $\pgs$, any counting or enumeration algorithm for one set can be adapted easily for the other.
Similarly, there are bijections between the sets $\cmbspt{\cparts}$ and $\cps$, $\cmbspt{\segs}$ and $\pms$, $\cmbssd{\cpartstp}$ and $\css$, as well as $\cmbssd{\trgs}$ and $\trs$.
In the same spirit, we define the sets $\cmbsst,\cmbssc \subseteq \cmbscf{\segs}$ of all crossing-free combinations of $\segs$ whose segments form spanning trees and spanning cycles on $\pts$, respectively.
\begin{definition}\label{def:cgraph}
Let $\units \subseteq \frags$.
A \emph{combination graph (over $\units$)} is a directed and acyclic multigraph $\cgraph$ with two distinguished vertices $\cgsource$ and $\cgsink$, called the \emph{source} and \emph{sink} of $\cgraph$, respectively.
All edges in $\cgraph$, except for those ending in $\cgsink$, are labeled with an element of $\units$.
Moreover, the sink $\cgsink$ has no outgoing edges.
The \emph{size} of $\cgraph$ is the number of vertices and edges in $\cgraph$.
\end{definition}
For any combination graph $\cgraph$ and any set $\cmbs$ of combinations, we say that $\cgraph$ \emph{represents} $\cmbs$ if there is a bijection between the set of directed $\cgsource$-$\cgsink$ paths in $\cgraph$ and the set $\cmbs$ in the following sense.
Taking any $\cgsource$-$\cgsink$ path in $\cgraph$ and building the set of labels on that path yields the corresponding combination in $\cmbs$.
The following are comparatively simple applications of an abstract framework developed in \cref{sec:framework}.
The corresponding proofs can be found in \cref{sec:applications}.
Many more, in some cases obvious, applications are possible.
\begin{theorem}[All geometric graphs]\label{thm:pg}
There exists a combination graph over $\segs$ of size $O(c^nn^3)$ with $c < 2.83929$ that represents $\cmbscf{\segs}$.
\end{theorem}
\begin{theorem}[Convex partitions]\label{thm:cp}
There exists a combination graph over $\cparts$ of size $O(2^nn^3)$ that represents $\cmbspt{\cparts}$.
\end{theorem}
\begin{theorem}[Perfect matchings]\label{thm:pm}
There exists a combination graph over $\segs$ of size $O(2^nn^3)$ that represents $\cmbspt{\segs}$.
\end{theorem}
\begin{theorem}[Convex subdivisions]\label{thm:cs}
There exists a combination graph over $\cpartstp$ of size $O(2^nn^3)$ that represents $\cmbssd{\cpartstp}$.
\end{theorem}
\begin{theorem}[Triangulations; c.f.\ Theorem 3 in \cite{SA13}]\label{thm:tr}
There exists a combination graph over $\trgs$ of size $O(2^nn^3)$ that represents $\cmbssd{\trgs}$.
\end{theorem}
Within our framework we can explain similar results for spanning trees and spanning cycles.
However, these two classes are substantially harder to deal with.
\Cref{sec:spanning} is devoted to the corresponding proofs.
\begin{theorem}[Spanning trees]\label{thm:st}
There exists a combination graph over $\segs$ of size $O(c^nn^3)$ with $c < 7.04313$ that represents $\cmbsst$.
\end{theorem}
\begin{theorem}[Spanning cycles]\label{thm:sc}
There exists a combination graph over $\segs$ of size $O(c^nn^3)$ with $c < 5.61804$ that represents $\cmbssc$.
\end{theorem}
To get a bound on the running time for computing an explicit representation of any one of the combination graphs $\cgraph$ in the above theorems, it suffices to add another factor $\ptsnum$ to the bound on the size of $\cgraph$.
See \cref{sec:construction} for details.
Given such a representation of $\cgraph$, the corresponding counting problem can be solved in time linear in the size of $\cgraph$ by counting directed $\cgsource$-$\cgsink$ paths using standard graph algorithms.
After removing all dead ends in $\cgraph$, which is also possible in time linear in the size of $\cgraph$, enumeration of the corresponding set $\cmbs$ requires time at most linear in the length of the longest $\cgsource$-$\cgsink$ path per enumerated object.
We will abstain from describing the required algorithms in detail, and instead refer to \cite{SA13} for an example.
Observe that the exponential bases in \Cref{thm:pg,thm:cp,thm:pm,thm:cs,thm:tr} are not larger than the exponential bases of the corresponding lower bounds given in the third row of \Cref{tab:extremalbounds}.
As a corollary we therefore get enumeration algorithms for the sets $\pgs$, $\cps$, $\pms$, $\css$ and $\trs$ whose overall running times are bounded by the length of the output times a polynomial in $n$.
For $\cps$ and $\pms$, a small adaptation, which is described in \cref{sec:enumeration}, results in enumeration algorithms with polynomial delay for each (in particular, the first) output.
\begin{restatable}{theorem}{thmenum}
The sets $\cps$ and $\pms$ can be enumerated such that the time delay for any output is bounded by a polynomial in $n$.
\end{restatable}
With the exception of $\pms$, the lower bounds are even strictly larger, which means that our algorithms compute the numbers $\pgsnum$, $\cpsnum$, $\cssnum$ and $\trsnum$ with exponential speed-up over any procedures that count by enumerating the respective sets.
For spanning trees it might as well be that the constant $c$ in \cref{thm:st} is smaller than $6.75$, but we were unable to prove it.
For spanning cycles we cannot always hope for such an exponential speed-up because for a set $\pts$ of $\ptsnum$ points in convex position we have $\scsnum = 1$.
\section{Bounding the Size of Combination Graphs}\label{sec:bounds}
In this final section, we show how to prove the bounds on the constants $c$ in \cref{thm:pg,thm:st,thm:sc}.
Recall that vertices of combination graphs can be interpreted as colorings of $\pts$ with a finite number of colors.
The following proofs make use of the fact that certain patterns cannot occur in these colorings because of geometric constraints.
Additionally, in the case of spanning trees and spanning cycles, we can improve the bounds further by identifying and discarding vertices from which the sink $\cgsink$ can no longer be reached.
\subsection{All geometric graphs}
Recall the definition of the set $\cmbscf{\segs}$ and the corresponding equivalence relation $\parteq$ from \cref{subsec:crossingfree}.
We are left to prove the following lemma.
{\renewcommand\footnote[1]{}\cfbound*}
\begin{proof}
We may encode equivalence classes $\partclass{\cmb}$ in $\quotient{\cmbscf{\segs}}{\parteq}$ by a string $s_\cmb$ of length $\ptsnum$ over the alphabet $\{\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked},\inlinepoint{free}{unmarked}\}$ and by an index $k_\cmb$.
The $i$-th entry in $s_\cmb$ is $\inlinepoint{dead}{unmarked}$ if $\pt_i \in \clow{\cmb}$, it is $\inlinepoint{alive}{unmarked}$ if $\pt_i \in \cpts{\cmb} \setminus \clow{\cmb}$, and it is $\inlinepoint{free}{unmarked}$ otherwise.
The index $k_\cmb$ is equal to the number $i$ which satisfies $\pt_i = \ulft{\umaxin{\cmb}}$.
From this encoding we immediately get a bound of $O(3^nn)$ on the size of $\quotient{\cmbscf{\segs}}{\parteq}$.
Our proof strategy now is as follows.
We ignore the indices $k_\cmb$ and give an upper bound of $O(\alpha^n)$ on the number of strings $s_\cmb$ which correspond to a combination $\cmb$ in $\cmbscf{\segs}$.
From this, the desired upper bound easily follows after adding an additional factor $n$.
We do so by defining an injective function from a set $A$ to a set $B$.
Set $A$ contains all strings $s_\cmb$.
Set $B$ contains all strings from $\{\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked},\inlinepoint{free}{unmarked}\}^n$ which do not contain any subsequences of the form $(\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\dots,\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked})$, that is, one or more consecutive symbols $\inlinepoint{free}{unmarked}$ enclosed by two symbols $\inlinepoint{dead}{unmarked}$.
Such subsequences are called \emph{forbidden} henceforth.
The bound on $|A|$ then follows from $|B| = \Theta(\alpha^n)$, an elementary counting problem whose proof we omit.
Let us define such an injective function $f \colon A \rightarrow B$.
For any $a \in A$ we construct $f(a)$ by the following process.
We iterate over $a$ from left to right, and whenever we find a forbidden subsequence $(\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\dots,\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked})$ we replace it by a sequence $(\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\dots,\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked})$ of the same length.
For example, if $a = (\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked})$ then $f(a) = (\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked})$.
However, if $a = (\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked})$ then $f(a) = (\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked})$ because the second forbidden subsequence in $a$ is no longer a forbidden subsequence after the first one has been replaced.
Clearly, $f$ is a function $A \rightarrow B$.
It only remains to prove injectivity.
Towards a contradiction, assume thus that $a \neq a'$ satisfy $f(a) = f(a') \eqqcolon b$.
Let $i$ be the smallest index with $a_i \neq a_i'$.
We distinguish the three cases $b_i = \inlinepoint{free}{unmarked}$, $b_i = \inlinepoint{alive}{unmarked}$ and $b_i = \inlinepoint{dead}{unmarked}$.
For $b_i = \inlinepoint{free}{unmarked}$, observe that the function $f$ never uses the symbol $\inlinepoint{free}{unmarked}$ to replace an entry in $a$ or $a'$.
Hence, $a_i = a_i' = \inlinepoint{free}{unmarked}$, a contradiction.
For $b_i = \inlinepoint{alive}{unmarked}$, we may assume without loss of generality that $a_i = \inlinepoint{dead}{unmarked}$ and $a_i' = \inlinepoint{alive}{unmarked}$.
Furthermore, $a_i$ must be either the first or last letter in a forbidden subsequence in $a$ that is replaced under $f$.
From minimality of $i$ it follows that it is the first letter.
Since the following argument generalizes to larger forbidden subsequences, we can now assume for reasons of simplicity that we have $(a_i,a_{i+1},a_{i+2}) = (\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked})$ and $(b_i,b_{i+1},b_{i+2}) = (\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked})$.
There are two possibilities for the corresponding letters in $a'$.
\begin{itemize}
\item
If $(a_i',a_{i+1}',a_{i+2}') = (\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked})$, then observe that the points $\pt_i$, $\pt_{i+1}$, $\pt_{i+2}$ from the set $\pts$ must form a left turn, that is, $\pt_{i+1}$ is strictly below the segment with endpoints $\pt_i$ and $\pt_{i+2}$.
The reason is that $a' = s_{\cmb'}$ corresponds to a combination $\cmb'$ in $\cmbscf{\segs}$, and there must be a segment in $\cmb'$ which has $\pt_{i+1}$, but neither $\pt_{i}$ nor $\pt_{i+2}$, in its lower shadow.
On the other hand, $(a_i,a_{i+1},a_{i+2}) = (\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked})$ implies that the points $\pt_i$, $\pt_{i+1}$, $\pt_{i+2}$ form a right turn, that is, $\pt_{i+1}$ is strictly above the segment with endpoints $\pt_i$ and $\pt_{i+2}$.
Again, this holds because also $a = s_\cmb$ corresponds to a combination $\cmb$, and there must be a segment in $\cmb$ which has both $\pt_{i}$ and $\pt_{i+2}$, but not $\pt_{i+1}$, in its lower shadow.
We have derived a contradiction because these two arrangements of $\pt_i$, $\pt_{i+1}$ and $\pt_{i+2}$ are mutually exclusive.
\item
If $(a_i',\dots,a_{i+3}') = (\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked})$ and $(b_i,\dots,b_{i+3}) = (\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked})$, that is, index $i+2$ is the beginning of a forbidden subsequence in $a'$, then we further obtain $(a_i,\dots,a_{i+3}) = (\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{dead}{unmarked})$.
A contradiction can be derived in the same manner as in the previous case.
The symbols in $a'$ imply that the points $\pt_i$, $\pt_{i+1}$ and $\pt_{i+3} $ form a left turn because there exists a segment which has $\pt_{i+1}$ and $\pt_{i+2}$, but neither $\pt_i$ nor $\pt_{i+3}$, in its lower shadow.
The symbols in $a$ imply that $\pt_i$, $\pt_{i+1}$ and $\pt_{i+3}$ form a right turn because there exists a segment which has $\pt_i$, $\pt_{i+2}$ and $\pt_{i+3}$, but not $\pt_{i+1}$, in its lower shadow.
\end{itemize}
For $b_i = \inlinepoint{dead}{unmarked}$, we may assume that $a_i = \inlinepoint{free}{unmarked}$, $a_i' = \inlinepoint{dead}{unmarked}$, and that in $a$ there exists a forbidden subsequence that starts before index $i$ and ends after index $i$ and that is replaced under $f$.
Again for simplicity we assume $(a_{i-1},a_i,a_{i+1}) = (\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked})$ and $(b_{i-1},b_i,b_{i+1}) = (\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked})$, longer sequences can be treated similarly.
By minimality of $i$ we get $a_{i-1}' = a_{i-1} = \inlinepoint{dead}{unmarked}$.
The only possible way for $a_{i-1}' = \inlinepoint{dead}{unmarked}$ to be replaced by $b_{i-1} = \inlinepoint{alive}{unmarked}$ under $f$ is if $a_{i-1}'$ is the last letter in a forbidden subsequence in $a'$.
Therefore, $(a_{i-3}',a_{i-2}',a_{i-1}') = (\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked})$ and $(b_{i-3},b_{i-2},b_{i-1}) = (\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked})$, again without loss of generality.
By minimality of $i$ we further get $(a_{i-3},a_{i-2},a_{i-1}) = (\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked})$.
In summary, we have derived $(a_{i-3},\dots,a_{i+1}) = (\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{free}{unmarked},\inlinepoint{dead}{unmarked})$ and $(b_{i-3},\dots,b_{i+1}) = (\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked},\inlinepoint{alive}{unmarked})$.
This is a contradiction because at most one of these two forbidden subsequences in $a$ is replaced under $f$.
\end{proof}
\subsection{Spanning Trees}
Recall the definition of $\agraphst{\pts}$ from \cref{subsec:st}.
The following lemma is all that is left to complete the proof of \cref{thm:st}.
\begin{lemma}\label{lem:stbound}
For any $\pts$, there exists a subgraph of $\agraphst{\pts}$ induced by a vertex subset $V \subseteq \quotient{\cmbsst}{\parteq}$ satisfying $|V| = O(\alpha^nn)$, where $\alpha \lessapprox 7.04313$, and such that $V$ contains all vertices that appear on at least one $\cgsource$-$\cgsink$ path in $\agraphst{\pts}$.
Moreover, given a vertex of $\agraphst{\pts}$, we can decide in time $O(n)$ whether it belongs to $V$.
\end{lemma}
Note that since we are able to recognize vertices from the set $V$ efficiently, the smaller subgraph of $\agraphst{\pts}$ can also be constructed efficiently bottom-up, simply by discarding any encountered vertices that do not belong to $V$.
\begin{proof}
We begin by specifying the eight colors $\inlinepoint{colone}{unmarked},\inlinepoint{coltwo}{unmarked},\dots,\inlinepoint{coleight}{unmarked}$ that were introduced only informally in \cref{subsec:st}.
Given a combination $\cmb$ in $\cmbslst$, these colors are assigned to the points in $\pts$ as follows.
If $\pt \in \clow{\cmb}$ then $\pt$ has color $\inlinepoint{colone}{unmarked}$, that is, $\inlinepoint{colone}{unmarked} = \inlinepoint{dead}{unmarked}$ in the original color scheme.
Otherwise, and if additionally $\pt \not\in \upts{\cmb}$, then $\pt$ has color $\inlinepoint{coleight}{unmarked}$, that is, $\inlinepoint{coleight}{unmarked} = \inlinepoint{free}{unmarked}$.
As follows, all remaining points are assigned one of the colors $\inlinepoint{coltwo}{unmarked},\dots,\inlinepoint{colseven}{unmarked}$, which thus correspond to the original $\inlinepoint{alive}{unmarked}$.
Color $\inlinepoint{coltwo}{unmarked}$ means that $\pt$ has out-degree 0 and exposes no drain.
Color $\inlinepoint{colthree}{unmarked}$ means that $\pt$ has out-degree 0 and exposes a drain to the left.
Color $\inlinepoint{colfour}{unmarked}$ means that $\pt$ has out-degree 0 and exposes a drain to the right.
Color $\inlinepoint{colfive}{unmarked}$ means that $\pt$ has out-degree 1 and exposes no drain.
Color $\inlinepoint{colsix}{unmarked}$ means that $\pt$ has out-degree 1 and exposes a drain to the left.
Color $\inlinepoint{colseven}{unmarked}$ means that $\pt$ has out-degree 1 and exposes a drain to the right.
Similar to the proof of \cref{lem:cf:bound}, let $s_\cmb$ be the string from $\{\inlinepoint{colone}{unmarked},\dots,\inlinepoint{coleight}{unmarked}\}^n$ that corresponds to a given combination $\cmb$ in $\cmbslst$.
The present lemma is a consequence of the following three observations, which all impose restrictions on certain patterns in $s_\cmb$.\\
\hspace*{-\parindent}%
\begin{minipage}{0.8\textwidth}
\begin{itemize}
\item For two consecutive points $\pt_i$ and $\pt_{i+1}$, it cannot be that both $\pt_i$ and $\pt_{i+1}$ have out-degree 1 and, at the same time, $\pt_i$ exposes a drain to the left and $\pt_{i+1}$ exposes a drain to the right.
As illustrated on the right, if that were the case then there would be either a crossing or a finite face of out-degree 0, both in contradiction with the definition of $\cmbslst$.
Hence, in $s_\cmb$ we will never observe the pattern $(\inlinepoint{colsix}{unmarked},\inlinepoint{colseven}{unmarked})$, that is, color $\inlinepoint{colsix}{unmarked}$ directly followed by color $\inlinepoint{colseven}{unmarked}$.
Additionally, if $\pt_i$ and $\pt_{i+k}$ (taking the role of $\pt_{i+1}$) are separated by any number of points of degree 0, the same argument still applies.
Hence, we can further rule out the pattern $(\inlinepoint{colsix}{unmarked},\inlinepoint{coleight}{unmarked},\dots,\inlinepoint{coleight}{unmarked},\inlinepoint{colseven}{unmarked})$.
\end{itemize}
\end{minipage}
\begin{minipage}{0.2\textwidth}
\centering
\input{fig/stpattern1}
\end{minipage}
\vspace{\belowdisplayskip}
\hspace*{-\parindent}%
\begin{minipage}{0.8\textwidth}
\begin{itemize}
\item Similarly, for two consecutive points $\pt_i$ and $\pt_{i+1}$, it cannot be that both $\pt_i$ and $\pt_{i+1}$ have out-degree 0 and, at the same time, $\pt_i$ exposes a drain to the left and $\pt_{i+1}$ exposes a drain to the right.
Since both points have out-degree 0 they cannot be connected by an edge, and thus we necessarily get a crossing, as illustrated on the right.
Hence, the pattern $(\inlinepoint{colthree}{unmarked},\inlinepoint{colfour}{unmarked})$ cannot occur in $s_\cmb$.
Using the same argument we can further rule out the pattern $(\inlinepoint{colthree}{unmarked},\inlinepoint{coltwo}{unmarked},\dots,\inlinepoint{coltwo}{unmarked},\inlinepoint{colfour}{unmarked})$, since in such a configuration there will always be two consecutive points $\pt_i$ and $\pt_{i+1}$ of out-degree 0 such that $\pt_i$ is the left endpoint of an edge and $\pt_{i+1}$ is the right endpoint of another edge.
Similar to the first observation, any additional points with degree 0 do not interfere with this argument.
The same can be said for any points that are in the lower shadow of $\cmb$.
Hence, we can rule out the pattern $(\inlinepoint{colthree}{unmarked},\inlinepoint{colone}{unmarked}|\inlinepoint{coltwo}{unmarked}|\inlinepoint{coleight}{unmarked},\dots,\inlinepoint{colone}{unmarked}|\inlinepoint{coltwo}{unmarked}|\inlinepoint{coleight}{unmarked},\inlinepoint{colfour}{unmarked})$, where ``$|$'' indicates an arbitrary choice between several options.
\end{itemize}
\end{minipage}
\begin{minipage}{0.2\textwidth}
\centering
\input{fig/stpattern2}
\end{minipage}
\vspace{\belowdisplayskip}
Lastly, we describe a pattern which might actually occur in $s_\cmb$.
However, for any such occurrence we will prove that the corresponding vertex $\partclass{\cmb}$ is not contained in any $\cgsource$-$\cgsink$ path in $\agraphst{\pts}$.
Therefore, the set $V$ from the lemma may safely be defined as the subset of all vertices of $\agraphst{\pts}$ that do not contain this final pattern.\\
\hspace*{-\parindent}%
\begin{minipage}{0.8\textwidth}
\begin{itemize}
\item Let $\pt_i$ and $\pt_{i+1}$ be two consecutive points such that $\pt_i$ exposes a drain to the right and $\pt_{i+1}$ exposes a drain to the left.
Then, clearly, the out-degree of the infinite face in $\cmb$ is different from 0 and thus there is no direct connection from $\partclass{\cmb}$ to $\cgsink$ in $\agraphst{\pts}$.
Moreover, any new segment $\unit$ that consumes the drain exposed by $\pt_i$ (that is, the border below $\pt_i$ becomes an out-border of the new face directly below $\unit$) must also consume the drain exposed by $\pt_{i+1}$, as illustrated on the right.
However, the face below $\unit$ having out-degree at least 2 contradicts the definition of $\cmbslst$.
Therefore, the two drains exposed by $\pt_i$ and $\pt_{i+1}$ can never be consumed and the sink $\cgsink$ can thus never be reached.
Hence, we can safely discard any vertices with the pattern $(\inlinepoint{colfour}{unmarked}|\inlinepoint{colseven}{unmarked},\inlinepoint{colthree}{unmarked}|\inlinepoint{colsix}{unmarked})$.
Additionally, suppose that $\pt_i$ and $\pt_{i+k}$ (taking the role of $\pt_{i+1}$) are separated by any number of points of degree 0 or points in the lower shadow of $\cmb$.
Then, no matter how a new segment $\unit$ is added, we again get a pair of drains facing each other and all points in between are either of degree 0 or in the lower shadow.
Hence, we can further discard any vertices with the pattern $(\inlinepoint{colfour}{unmarked}|\inlinepoint{colseven}{unmarked},\inlinepoint{colone}{unmarked}|\inlinepoint{coleight}{unmarked},\dots,\inlinepoint{colone}{unmarked}|\inlinepoint{coleight}{unmarked},\inlinepoint{colthree}{unmarked}|\inlinepoint{colsix}{unmarked})$.
\end{itemize}
\end{minipage}
\begin{minipage}{0.2\textwidth}
\centering
\input{fig/stpattern3}
\end{minipage}
\vspace{\belowdisplayskip}
Let $A$ be the subset of $\{\inlinepoint{colone}{unmarked},\dots,\inlinepoint{coleight}{unmarked}\}^n$ containing only strings without subsequences belonging to the three patterns from the above observations.
Then, it can be shown that $|A| = \Theta(\alpha^n)$ using standard techniques, and the lemma follows.
\end{proof}
\subsection{Spanning Cycles}
Finally, recall the definition of $\agraphsc{\pts}$ from \cref{subsec:sc}.
The following lemma concludes the proof of \cref{thm:sc}.
\begin{lemma}
For any $\pts$, there exists a subgraph of $\agraphsc{\pts}$ induced by a vertex subset $V \subseteq \quotient{\cmbslsc}{\parteq}$ satisfying $|V| = O(\alpha^nn)$, where $\alpha \lessapprox 5.61804$, and such that $V$ contains all vertices that appear on at least one $\cgsource$-$\cgsink$ path in $\agraphsc{\pts}$.
Moreover, given a vertex of $\agraphsc{\pts}$, we can decide in time $O(n)$ whether it belongs to $V$.
\end{lemma}
\begin{proof}
We again start by giving specifications for the colors $\inlinepoint{colone}{unmarked},\inlinepoint{coltwo}{unmarked},\dots,\inlinepoint{colsix}{unmarked}$ that are assigned to the points in $\pts$ for any given combination $\cmb$ in $\cmbslsc$.
Color $\inlinepoint{colone}{unmarked}$ means that the corresponding point $\pt$ has degree 2 and exposes no drain.
Color $\inlinepoint{coltwo}{unmarked}$ means that $\pt$ has degree 2 and exposes a drain.
Color $\inlinepoint{colthree}{unmarked}$ means that $\pt$ has degree 1 and exposes no drain.
Color $\inlinepoint{colfour}{unmarked}$ means that $\pt$ has degree 1 and exposes a drain to the left.
Color $\inlinepoint{colfive}{unmarked}$ means that $\pt$ has degree 1 and exposes a drain to the right.
Color $\inlinepoint{colsix}{unmarked}$ means that $\pt$ has degree 0.
The following observation is similar to the third observation in the proof of \cref{lem:stbound}.
Let $\cmb$ be a combination in $\cmbslsc$ and let $\pt_i$ and $\pt_{i+1}$ be two consecutive points.
If $\pt_i$ exposes a drain to the right and $\pt_{i+1}$ exposes a drain to the left in $\cmb$, then $\cmb$ cannot be augmented in such a way that the infinite face has out-degree $0$ without creating finite faces with out-degree larger than $1$.
In other words, the vertex $\partclass{\cmb}$ does not appear on any $\cgsource$-$\cgsink$ path in $\agraphsc{\pts}$ and may safely be discarded.
The same holds for two not necessarily consecutive points $\pt_i$ and $\pt_{i+k}$ such that again $\pt_i$ exposes a drain to the right and $\pt_{i+k}$ exposes a drain to the left, and such that all points in between $\pt_i$ and $\pt_{i+k}$ have either degree 0 or 2.
No matter how a new segment $\unit$ is added to $\cmb$, we will again end up with a combination with two exposed drains facing each other and only points of degree 0 or 2 in between.
The subset $V$ of vertices of $\agraphsc{\pts}$ can thus be defined as follows.
We simply exclude all vertices which contain the pattern $(\inlinepoint{colfive}{unmarked},\inlinepoint{colone}{unmarked}|\inlinepoint{coltwo}{unmarked}|\inlinepoint{colsix}{unmarked},\dots,\inlinepoint{colone}{unmarked}|\inlinepoint{coltwo}{unmarked}|\inlinepoint{colsix}{unmarked},\inlinepoint{colfour}{unmarked})$.
Let now $A$ be the subset of $\{\inlinepoint{colone}{unmarked},\dots,\inlinepoint{colsix}{unmarked}\}^n$ containing only strings without subsequences belonging to the above pattern.
Then, $|A| = \Theta(\alpha^n)$, and the lemma follows.
\end{proof}
\section{Representations of Combination Graphs}\label{sec:construction}
In the preceding sections, vertices of a combination graph $\cgraph$ were always treated as equivalence classes over a set of combinations.
Of course, this is a very inefficient way to represent them in an actual algorithm.
However, in all cases we have seen how to describe these equivalence classes by an assignment of a constant number of colors to the points in $\pts$ and an index to the point with the special marking.
It is thus easy to encode vertices of $\cgraph$ with a linear number of bits for the coloring and a logarithmic number of bits for the index.
As for the construction of $\cgraph$, this is most easily done bottom-up.
That is, we start with the source $\cgsource$, we enumerate all outgoing edges and add the corresponding new vertices to our representation of $\cgraph$.
We continue this process, that is, pick a vertex and enumerate all outgoing edges, until we have done so for all vertices in $\cgraph$.
In all cases except for convex partitions and convex subdivisions, the enumeration of outgoing edges for a given vertex can be done in the most wasteful way while still staying within the required time bounds.
For example, in the case of \cref{thm:pg}, for every vertex in $\agraphcf{\segs}$ we can simply enumerate the whole set $\segs$ and check for each segment whether it corresponds to an outgoing edge in linear time.
In the case of \cref{thm:cp,thm:cs} we have to be more careful since we do not have the time to enumerate the whole set $\cparts$, say, for every vertex in $\agraphpt{\cparts}$.
A simple solution is to enumerate the set $\cparts$ only once at the beginning and to find for each enumerated $\unit \in \cparts$ all vertices in $\agraphpt{\cparts}$ for which $\unit$ corresponds to an outgoing edge.
The running time of this approach can be bounded by using the same arguments as in the proof of \cref{lem:pt:bound}.
\section{Enumeration with Polynomial-Time Delay}\label{sec:enumeration}
In this section we sketch an additional trick\footnote{This trick was suggested to the author by Emo Welzl.}, which allows us to enumerate the sets of crossing-free convex partitions and perfect matchings with polynomial-time delay.
That is, we give algorithms which output the elements of $\cps$ or $\pms$ in some order and without repetitions, and such that the time we have to wait for any new output is not larger than a polynomial.
As already discussed in \cref{sec:basics}, from \cref{thm:cp,thm:pm} we get enumeration algorithms for these two sets that run in time $O^\ast(\cpsnum)$ and $O^\ast(\pmsnum)$, respectively.
However, while the time delay \emph{between} any two outputs is bounded by a polyomial, the time delay \emph{before} the first output is exponential.
Precisely, there is a preprocessing phase that might take time $\Theta^\ast(2^n)$ during which we construct representations of the respective combination graphs $\agraphpt{\cparts}$ and $\agraphpt{\segs}$ and during which we do not produce any outputs.
The trick is to hide this preprocessing phase by outputting $\Theta^\ast(2^n)$ objects obtained by other means.
\thmenum*
\begin{proof}
We begin by defining a sufficiently large subset $\pmss$ of $\cps$ and $\pms$.
Elements of this set are called \emph{easy perfect matchings}, and they are constructed recursively.
If $\pts$ is the empty set, then $\pmss = \pms$.
Otherwise, let $\pt_1$ be the left-most point and let $\pt_i$ be any other point.
Let $\pts_i$ be the set of points that are to the left of the directed line through $\pt_1$ and $\pt_i$, and let $\pts_i'$ be the set of points that are to the right of that line.
The set $\pmss$ contains all perfect matchings that, for any choice of $i$, are composed of the edge $\pt_1\pt_i$ and two easy perfect matchings on $P_i$ and $P_i'$, respectively.
It is clear that easy perfect matchings are crossing-free.
Furthermore, note that efficient enumeration and recognition algorithms for the set $\pmss$ are easy to obtain from the definition.
Lastly, the number of easy perfect matchings on even-sized point sets satisfies the Catalan recurrence, and thus $|\pmss| = \Theta(2^n/n^{3/2})$.
Refer to \cite{GNT00} for more details, where these objects are used to prove a corresponding lower bound on the number of perfect matchings.
We now have everything that we need.
In order to enumerate the set $\pms$, say, with polynomial-time delay, we start the construction of $\agraphpt{\segs}$.
During this preprocessing phase, we output elements of $\pmss$ in appropriate time intervals.
Once we have an explicit representation of $\agraphpt{\segs}$, we continue the enumeration by outputting arbitrary elements of $\pms$.
Of course, whenever we have a new potential output, we have to check first whether it is an easy perfect matching, which means that it has been output before.
If that is the case, then we simply discard it.
One final caveat now is that there might be a long period where we have to discard all potential outputs, which might again lead to a delay that is no longer polynomially bounded.
However, this is easily fixed for example by only using up half of the set $\pmss$ during the preprocessing phase, and by using the other half as a substitute for every other discarded output during the second phase.
\end{proof}
\section{The Abstract Framework}\label{sec:framework}
As before, let $\pts$ be a set of $\ptsnum$ points in general position.
Furthermore, assume that no two points share the same $x$-coordinate, which means the points can be ordered as $\pt_1,\dots,\pt_\ptsnum$ from left to right in a unique way.
If a point $\pt_i$ is to the left of another point $\pt_j$, that is, if $i \leq j$, then we write $\pt_i \ptspreceq \pt_j$.
Recall that $\upts{\unit} \subseteq \pts$ is the set of vertex points in any $\unit \in \frags$.
We define $\ulft{\unit} \coloneqq \min_{\ptspreceq}(\upts{\unit})$ and $\urgt{\unit} \coloneqq \max_{\ptspreceq}(\upts{\unit})$, the \emph{left-most} and \emph{right-most point} of $\unit$, respectively.
For any $\unit_1,\unit_2 \in \frags$, if $\urgt{\unit_1} \ptspreceq \ulft{\unit_2}$ holds then we say that $\unit_1$ is \emph{to the left} of $\unit_2$ and we also write $\unit_1 \utorgt \unit_2$.
For each $\unit$ we define $\ulow{\unit} \subseteq \pts$ and $\uupp{\unit} \subseteq \pts$, the \emph{lower} and \emph{upper shadow} of $\unit$, respectively.
The set $\ulow{\unit}$ contains all points in $\pts$ from which a vertical ray shooting upwards intersects the relative interior of some segment of $\unit$.
The set $\uupp{\unit}$ is defined in an analogous way.
Whenever we have either $\upts{\unit_1} \cap \ulow{\unit_2} \neq \emptyset$ or $\uupp{\unit_1} \cap \upts{\unit_2} \neq \emptyset$ for any $\unit_1,\unit_2$, then we say that \emph{$\unit_2$ depends on $\unit_1$} and we write $\unit_1 \udep \unit_2$.
The following lemma aims at making this cryptic definition more intuitive.
Here, and only here, by a \emph{point on $\unit$} we mean either a point in $\upts{\unit}$ or a point in the relative interior of some edge of $\unit$.
\begin{lemma}\label{obs:dependence}
Let $\unit_1,\unit_2 \in \frags$ be arbitrary.
Then, $\unit_2$ depends on $\unit_1$ if and only if there exists a point on $\unit_2$ directly (that is, same $x$-coordinate) and strictly above a point on $\unit_1$.
In particular, if $\unit_1$ and $\unit_2$ are crossing then they are mutually dependent, that is, $\unit_1 \udep \unit_2$ and $\unit_2 \udep \unit_1$.
\end{lemma}
\hspace*{-\parindent}%
\begin{minipage}{0.8\textwidth}
\begin{proof}
The ``only if'' is immediate by definition of $\unit_1 \udep \unit_2$.
For the ``if'', let us fix two points on $\unit_1$ and $\unit_2$, respectively, the one on $\unit_2$ directly and strictly above the one on $\unit_1$.
If either of those points is contained in $\upts{\unit_1}$ or $\upts{\unit_2}$, respectively, the conclusion $\unit_1 \udep \unit_2$ is again immediate.
Otherwise, let $e_1$ and $e_2$ be the edges of $\unit_1$ and $\unit_2$ that contain the two respective fixed points in their relative interiors.
Without loss of generality we assume that $e_1$ and $e_2$ diverge and thus do not intersect towards the left, as illustrated on the right.
In case 1, the left endpoint of $e_1$ is below $e_2$, which means $\upts{\unit_1} \cap \ulow{\unit_2} \neq \emptyset$ and hence $\unit_1 \udep \unit_2$.
In case 2, the left endpoint of $e_2$ is above $e_1$, which means $\uupp{\unit_1} \cap \upts{\unit_2} \neq \emptyset$ and hence again $\unit_1 \udep \unit_2$.
\end{proof}
\end{minipage}
\begin{minipage}{0.2\textwidth}
\centering
\input{fig/dependingsegments}
\end{minipage}
\vspace{\belowdisplayskip}
To ease notation throughout, we use $\cpts{\cmb}$ and $\clow{\cmb}$ to denote the sets $\bigcup_{\unit \in \cmb} \upts{\unit}$ and $\bigcup_{\unit \in \cmb} \ulow{\unit}$, respectively, where $\cmb$ is any combination.
Besides giving an intuition for the dependence relation, the above lemma turns out to be absolutely crucial for everything that follows.
It suggests a safe and practical way of adding a new segment $\unit \in \segs$, say, to a crossing-free combination $\cmb$ of $\segs$.
Safe means that we do not introduce any crossings, that is, $\cmb \cup \{\unit\}$ is itself crossing-free.
Practical means that we may, to a great extent, remain ignorant of the exact composition of $\cmb$.
Indeed, as long as we know the sets $\cpts{\cmb}$ and $\clow{\cmb}$, and provided that we choose $\unit$ such that $\upts{\unit} \cap \clow{\cmb} = \emptyset$ and $\uupp{\unit} \cap \cpts{\cmb} = \emptyset$, then no element of $\cmb$ can possibly depend on $\unit$ and hence, by \cref{obs:dependence}, $\cmb \cup \{\unit\}$ is crossing-free.
By extension, this suggests a way of constructing a combination $\cmb$ by adding all elements in a succession where earlier occurrences do not depend on later occurrences.
An apparent disadvantage is that this will not work for every conceivable subset $\units$ of $\frags$ and every crossing-free combination $\cmb$ of $\units$.
Most importantly, for it to work, there must be no circular dependencies among elements of $\cmb$.
In the following we formalize this requirement.
Let $\units \subseteq \frags$ and let $\cmb \subseteq \units$ be arbitrary.
An element $\unit$ of $\cmb$ is \emph{extreme (in $\cmb$)} if $\unit \undep \unit'$ holds for all other elements $\unit'$ of $\cmb$.
If it exists, the \emph{right-most extreme element} in $\cmb$ is the unique extreme element $\unit$ in $\cmb$ which satisfies $\unit' \utorgt \unit$ for all other extreme elements $\unit'$ in $\cmb$.
\begin{definition}\label{def:serializable2}
Let $\units \subseteq \frags$ and let $\cmbs$ be a set of combinations of $\units$.
We call $\cmbs$ \emph{serializable} if it is non-empty and if every non-empty $\cmb \in \cmbs$ contains a right-most extreme element, which we then denote by $\umaxin{\cmb}$.
Additionally, $\cmb \setminus \{\umaxin{\cmb}\}$ must itself be an element of $\cmbs$.
\end{definition}
Let $\cmbs$ be a serializable set of combinations of $\units$, and let $\cmb,\cmb' \in \cmbs$.
We will often write $\ctarrow{\cmb}{\cmb'}{\unit}$, which stands for $\cmb = \cmb' \setminus \{\unit\}$ and $\unit = \umaxin{\cmb'}$.
Observe that $\cmbs$ naturally induces a directed and acyclic graph (actually, a tree) with vertex set $\cmbs$ and edges with labels from the set $\units$.
Indeed, whenever $\ctarrow{\cmb}{\cmb'}{\unit}$ holds we simply add an edge from vertex $\cmb$ to vertex $\cmb'$ with label $\unit$.
A combination graph over $\units$ that represents an arbitrary subset of $\cmbs$ is obtained by defining $\cgsource \coloneqq \emptyset$ and by adding appropriate unlabeled edges which end in an additional vertex $\cgsink$.
However, the resulting combination graph is too large as its size is $\Theta(|\cmbs|)$.
In the following we show how to compress it without losing too many of its nice properties.
\begin{definition}\label{def:coherent}
Let $\cmbs$ be serializable.
An equivalence relation $\parteq$ on $\cmbs$ is called \emph{coherent} if $\cmb_1 \parteq \cmb_2$ implies the following.
If $\cmb_1 \xrightarrow{\unit} \cmb_1'$ holds then also $\cmb_2 \xrightarrow{\unit} \cmb_2'$ for some $\cmb_2' \parteq \cmb_1'$.
\end{definition}
Intuitively, to make our combination graph smaller we would like to merge two vertices $\cmb_1$ and $\cmb_2$.
This makes sense only if the subtrees rooted at $\cmb_1$ and $\cmb_2$ are identical when looking at edge labels.
As will be proved later, if $\cmb_1 \parteq \cmb_2$ holds, coherency of $\parteq$ enforces precisely what we want.
In the remainder of the paper we will always deal with a serializable set $\cmbs$ of combinations of some $\units \subseteq \frags$, and an equivalence relation $\parteq$ on $\cmbs$.
For any $\cmb \in \cmbs$ we then define the equivalence class $\partclass{\cmb} \coloneqq \{\cmb' \in \cmbs \colon \cmb' \parteq \cmb\}$, where the relation $\parteq$ will be obvious from the context.
We also define the set $\quotient{\cmbs}{\parteq} \coloneqq \{\partclass{\cmb} \colon \cmb \in \cmbs\}$ of all equivalence classes.
\begin{definition}
A \emph{combination problem (on $\pts$)} is a tuple $(\units,\cmbs,\parteq,\target)$ where $\units \subseteq \frags$, $\cmbs$ is a serializable set of combinations of $\units$, $\parteq$ is a coherent equivalence relation on $\cmbs$, and $\target$ is a subset of $\quotient{\cmbs}{\parteq}$.
\end{definition}
As will be described below, the set $\target$ allows us to specify the subset of $\cmbs$ we are actually interested in.
Every combination problem $(\units,\cmbs,\parteq,\target)$ induces a corresponding combination graph $\cgraph = \cgraph(\units,\cmbs,\parteq,\target)$ over $\units$ as follows.
The vertices of $\cgraph$ are all equivalence classes in $\quotient{\cmbs}{\parteq}$ plus one extra vertex denoted by $\cgsink$.
The vertex $\partclass{\emptyset}$ is also referred to by $\cgsource$.
Existence of $\cgsource$ follows from $\emptyset \in \cmbs$, an easy consequence of serializability of $\cmbs$.
Whenever $\ctarrow{\cmb}{\cmb'}{\unit}$ holds, then we add an edge from vertex $\partclass{\cmb}$ to vertex $\partclass{\cmb'}$ with label $\unit$.
We do not, however, add the same labeled edge more than once.
Lastly, for every vertex $\partclass{\cmb}$ in $\target$ we add an unlabeled edge which starts in $\partclass{\cmb}$ and ends in $\cgsink$.
Observe that $\cgraph$ does not contain any directed cycles because given any such cycle, it would be possible to construct an infinite sequence $(\ctarrow{\cmb}{\cmb'}{\unit},\ctarrow{\cmb'}{\cmb''}{\unit'},\dots)$, which cannot exist.
Recall that $\ctarrow{\cmb}{\cmb'}{\unit}$ stands for $\cmb = \cmb' \setminus \{\unit\}$ and $\unit = \umaxin{\cmb'}$.
Additionally, we write $\cgarrow{\cmb}{\cmb'}{\unit}$ if there exists an edge in $\cgraph$ from vertex $\partclass{\cmb}$ to vertex $\partclass{\cmb'}$ with label $\unit$.
The following observations are straight-forward consequences of coherency of $\parteq$ and of the way $\cgraph$ is constructed.
\begin{observation}\label{obs:cgproperties}
The combination graph $\cgraph = \cgraph(\units,\cmbs,\parteq,\target)$, defined as above, is
\begin{itemize}
\item
Complete: if $\ctarrow{\cmb}{\cmb'}{\unit}$, then
$\cgarrow{\cmb}{\cmb'}{\unit}$,
\item
Sound: if $\cgarrow{\cmb}{\cmb'}{\unit}$, then $\ctarrow{\cmb}{\cmb''}{\unit}$
for some $\cmb'' \parteq \cmb'$,
\item
Deterministic: if $\cgarrow{\cmb}{\cmb'}{\unit}$ and $\cgarrow{\cmb}{\cmb''}{\unit}$,
then $\partclass{\cmb'} = \partclass{\cmb''}$.
\end{itemize}
\end{observation}
By induction it can now be shown that for every vertex $\partclass{\cmb}$ there is a natural bijection from $\partclass{\cmb}$ to the set of directed $\cgsource$-$\partclass{\cmb}$ paths in $\cgraph$, which implies the following lemma.
\begin{restatable}{lemma}{represent}
\label{lem:represent}
Let $(\units,\cmbs,\parteq,\target)$ be a combination problem and let $\cgraph$ be the corresponding combination graph.
Then, $\cgraph$ represents $\bigcup\target$ and the size of $\cgraph$ is at most $O(|\quotient{\cmbs}{\parteq}| \cdot |\units|)$.
\end{restatable}
\begin{proof}
The upper bound on the size of $\cgraph$ holds because it has exactly $|\quotient{\cmbs}{\parteq}| + 1$ vertices and, since it is deterministic (as in \Cref{obs:cgproperties}) and has no duplicate labeled edges, each vertex has at most $|\units| + 1$ outgoing edges.
It only remains to show that $\cgraph$ represents $\bigcup\target$.
For each directed path $\cgpath$ in $\cgraph$ let $\cmbntr{\cgpath}$ denote the set of all labels on $\cgpath$, that is, $\cmbntr{\cgpath}$ is the combination of $\units$ corresponding to the path $\cgpath$.
Moreover, for any $\cmb \in \cmbs$, the \emph{canonical order} of $\cmb$ is the unique sequence over all elements in $\cmb$ which is obtained by successively removing right-most extreme elements from $\cmb$ and then putting the removed elements in reverse order.
For every vertex $\partclass{\cmb}$ in $\cgraph$ we now claim the following.
Firstly, there exists a directed $\cgsource$-$\partclass{\cmb}$ path $\cgpath$ in $\cgraph$ with $\cmbntr{\cgpath} = \cmb$, irrespective of the choice of the representative $\cmb$.
Secondly, for every directed $\cgsource$-$\partclass{\cmb}$ path $\cgpath$ we have that $\cmbntr{\cgpath} \in \partclass{\cmb}$.
Thirdly, the labels on any directed $\cgsource$-$\partclass{\cmb}$ path $\cgpath$ with $\cmbntr{\cgpath} = \cmb$ appear in canonical order of $\cmb$.
The proof of these claims is by induction over an arbitrary topological ordering of the vertices in $\cgraph$.
After that, the lemma follows by combining the claims with the fact that $\cgraph$ is deterministic.
For the first part of the claim, if $\cmb = \emptyset$, then $\partclass{\cmb} = \cgsource$ and the $\cgsource$-$\cgsource$ path without edges works.
Otherwise, if $\cmb \neq \emptyset$, then we know that $\ctarrow{\cmb'}{\cmb}{\unit}$ for some $\cmb'$ and $\unit$.
By completeness, there is an edge $\cgarrow{\cmb'}{\cmb}{\unit}$ in $\cgraph$ and a $\cgsource$-$\partclass{\cmb}$ path $\cgpath$ with $\cmbntr{\cgpath} = \cmb$ can be constructed from the $\cgsource$-$\partclass{\cmb'}$ path $\cgpath'$ with $\cmbntr{\cgpath'} = \cmb'$, which exists by induction.
For the second part of the claim, let $\cgpath$ be a directed $\cgsource$-$\partclass{\cmb}$ path in $\cgraph$.
If $\cgpath$ is of length zero then $\partclass{\cmb} = \cgsource$ and $\cmbntr{\cgpath} = \emptyset \in \partclass{\cmb}$.
Otherwise, let $\cgpath'$ be the $\cgsource$-$\partclass{\cmb'}$ path that is obtained by removing the last edge $\cgarrow{\cmb'}{\cmb}{\unit}$ from $\cgpath$.
By induction, and without loss of generality, $\cmbntr{\cgpath'} = \cmb'$.
By soundness, we have $\ctarrow{\cmb'}{\cmb''}{\unit}$ for some $\cmb'' \parteq \cmb$ and hence also $\cmbntr{\cgpath} = \cmb'' \in \partclass{\cmb}$.
For the third part of the claim, simply observe that in the previous paragraph, the last label on the path $\cgpath$ with $\cmbntr{\cgpath} = \cmb''$ is the right-most extreme element in $\cmb''$.
Hence, this last claim also follows by induction.
\end{proof}
\section{Introduction}\label{sec:introduction}
Let $\pts$ be a set of $\ptsnum$ points in the plane.
We assume $\pts$ to be in \emph{general position}, which means that no three points in $\pts$ are collinear.
A \emph{geometric graph} on $\pts$ is a simple graph with vertex set $\pts$, combined with an embedding into the plane where edges are drawn as straight segments between the corresponding endpoints.
Two distinct edges are \emph{crossing} if their drawings intersect in their respective relative interiors, otherwise they are \emph{non-crossing}.
A geometric graph on $\pts$ is \emph{crossing-free} if its edges are pairwise non-crossing.
\begin{figure}
\begin{center}
\input{fig/graphexamples}
\end{center}
\caption{
The seven defined classes of crossing-free geometric graphs.
}
\label{fig:graphexamples}
\end{figure}
We define the set $\pgs$ of all crossing-free geometric graphs on $\pts$, the set $\cps$ of \emph{crossing-free convex partitions}, the set $\pms$ of \emph{crossing-free perfect matchings}, the set $\css$ of \emph{convex subdivisions}, the set $\trs$ of \emph{triangulations}, the set $\sts$ of \emph{crossing-free spanning trees}, and the set $\scs$ of \emph{crossing-free spanning cycles}.
\Cref{fig:graphexamples} shows one representative of each defined class on a fixed point set.
Note that by a crossing-free convex partition we mean a partition of $\pts$ in such a way that the convex hulls of individual parts are pairwise disjoint.
Convex subdivisions are subdivisions of the convex hull of $\pts$ into convex faces with points from $\pts$ as vertices.
Note that every crossing-free convex partition and every convex subdivision is represented uniquely by a crossing-free geometric graph which contains all the edges on the boundaries of individual parts and faces.
Finally, a triangulation is an edge-maximal crossing-free geometric graph, which is a special case of a convex subdivision.
\begin{table*}[b]
\begin{center}
\input{tab/extremalbounds}
\end{center}
\caption{
Extremal bounds, where cells display the respective exponential bases
}
\label{tab:extremalbounds}
\end{table*}
\footnotetext{Apply the lower bound for perfect matchings on all subsets of $\pts$ and then use the binomial theorem.
Following a common notation, we also define the numbers $\pgsnum \coloneqq |\pgs|$, $\cpsnum \coloneqq |\cps|$, $\pmsnum \coloneqq |\pms|$, $\cssnum \coloneqq |\css|$, $\trsnum \coloneqq |\trs|$, $\stsnum \coloneqq |\sts|$, and $\scsnum \coloneqq |\scs|$.
Some of these numbers have received considerable attention.
In particular, exponential upper and lower bounds have been established and gradually improved over the past decades.
Instead of listing all successive improvements, which are too numerous to count, we summarize the current state of affairs in \cref{tab:extremalbounds}.
As a reading example, the top-left entry says that $\pgsnum = O^*(187.53^ n)$ holds for all point sets $\pts$, see \cite{SS12} (the $*$ indicates that any subexponential factors are ignored).
For an always up-to-date list, which includes bounds for many other types of geometric graphs, we refer the interested reader to Sheffer's webpage \cite{ShWeb}.
The defined classes of geometric graphs have also been studied from an algorithmic point of view.
The problem of enumeration has been solved for $\pgs$, $\trs$ and $\sts$, see \cite{AF96,B02,AAHV07,KT08}.
By solved we mean that these sets can be enumerated in such a way that the time delay for each enumerated object is bounded by a polynomial in $\ptsnum$.
In terms of counting, especially triangulations have been studied extensively \cite{A99,ABRS15,KLS15}.
Furthermore, for counting many other types of geometric graphs there already exists a general framework \cite{ABCR12}, which is based on the onion layer structure of a point set.
However, for a long time no counting algorithm was always provably faster than enumerating the set whose size was to be determined.
For triangulations this changed with a remarkable paper by Alvarez and Seidel \cite{SA13}, who showed how to compute the number $\trsnum$ in time $O(2^nn^ 2)$.
This is always exponentially faster than enumeration because $\trsnum = \Omega^*(2.43^\ptsnum)$ holds for all sets $\pts$, see \cref{tab:extremalbounds}.
An unrelated approach led to a similar result for the class of all crossing-free geometric graphs.
Razen and Welzl \cite{RW11} showed how to compute $\pgsnum$ in time $O^\ast(\trsnum)$.
Since they proved that $\pgsnum = \Omega^*(2.82^n) \cdot \trsnum$ holds for all sets $\pts$, their algorithm also achieves an exponential speed-up over any procedure that counts by enumerating the whole set $\pgs$.
Furthermore, after the first publication of the present paper \cite{W14}, Marx and Miltzow discovered a new class of algorithms which make use of the so called cactus layer structure of a triangulation \cite{MM16}.
These algorithms run in time $n^{O(\sqrt n)}$ and are able to compute the number of triangulations and related structures exactly.
In the remainder of this paper we develop and make use of an abstract framework that generalizes ideas originally used by Alvarez and Seidel \cite{SA13} for counting triangulations.
Loosely speaking, the technique boils down to the following steps.
Fix any set of crossing-free geometric graphs whose elements can be decomposed into reasonably small or simple pieces.
For instance, every triangulation can be decomposed into a set of interior-disjoint triangles and, similarly, each crossing-free perfect matching can be decomposed into a set of non-intersecting segments.
The aim then is to construct a directed acyclic graph $\cgraph$ with the following properties.
Firstly, each edge in $\cgraph$ is labeled with one of the aforementioned pieces.
Secondly, there exist distinguished source and sink vertices in $\cgraph$.
Thirdly, there is a natural bijection between source-sink paths in $\cgraph$ and the fixed set of geometric graphs.
By this we mean that given any source-sink path in $\cgraph$, we can collect all the labels appearing on that path and combine them to obtain the corresponding geometric graph.
Clearly, given such a graph $\cgraph$ for one particular class of geometric graphs, the corresponding counting and enumeration problems can be reduced to counting and enumerating source-sink paths in $\cgraph$.
\section{Spanning Trees and Spanning Cycles}\label{sec:spanning}
In this section we show that it is possible to construct non-trivial combination graphs for the sets of crossing-free spanning trees and spanning cycles.
Spanning trees and spanning cycles are harder to deal with than anything that we have encountered before.
The reason is that these graphs have properties which hold globally.
For example, the construction of $\agraphpt{\segs}$ in \cref{subsec:partitions} can be adapted in such a way that source-sink paths correspond to 2-regular (instead of 1-regular) crossing-free geometric graphs.
We simply would have to keep track of the degree of each vertex (whether it is currently 0, 1, or 2, which means we would need three instead of two colors) and in the end require that every vertex has degree 2.
However, if we want that source-sink paths correspond only to crossing-free spanning cycles, then we also need that each such path corresponds to a connected geometric graph.
Being connected is such a property that holds globally, and there seems to be no obvious and efficient way to deal with it.
To get rid of this problem, at least in the case of spanning trees and spanning cycles, we next state an auxiliary lemma.
It will allow us to translate connectivity into simpler features which can be enforced on a local level.
Let $G$ be a directed multigraph\footnote{To avoid confusion, let us stress that we introduce completely new entities here. That is, $G$ is neither a geometric graph nor a combination graph.} and let $v$ be a vertex in $G$.
$G$ is \emph{root-oriented towards $v$} if all vertices in $G$ have exactly one outgoing edge, except for $v$, which has no outgoing edges.
If $G$ is root-oriented towards $v$, then $v$ is called the \emph{root} of $G$.
Observe that being root-oriented implies that $G$ has exactly $n-1$ edges, where $n$ is the number of vertices in $G$.
It does however not imply that $G$ is connected or, in other words, a tree.
The reason is that there might be a connected component with a directed cycle.
Such components are always disconnected from the root.
A \emph{plane drawing} of $G$ is a drawing which maps all vertices of $G$ to distinct points in the plane and which draws all edges as simple curves such that no two edges intersect except possibly in a common endpoint.
Given two respective plane drawings of directed multigraphs $G_1$ and $G_2$, we say the drawings are \emph{disjoint} if they do not use any common points in the plane.
Moreover, the drawings are \emph{entangled} if for each cycle in either drawing, both its interior and exterior contain a point used by the respective other drawing.
Finally, for fixed vertices $v_1$ in $G_1$ and $v_2$ in $G_2$ we say the two drawings are \emph{tangent} in $v_1$ and $v_2$ if the points corresponding to $v_1$ and $v_2$ can be connected by an additional curve without intersecting any points already used in either drawing.
\begin{restatable}{lemma}{entangledtrees}
\label{lem:entangledtrees}
Let $G_1$ and $G_2$ be finite, directed multigraphs that are root-oriented towards $v_1$ and $v_2$, respectively.
Then, there exist plane drawings of $G_1$ and $G_2$ that are disjoint, entangled and tangent in $v_1$ and $v_2$ if and only if both $G_1$ and $G_2$ are trees.
\end{restatable}
\begin{proof}
If $G_1$ and $G_2$ are trees then the desired drawings clearly exist.
To prove the other direction of the lemma, it suffices to show that both $G_1$ and $G_2$ do not contain any cycles.
Let us thus fix plane drawings of $G_1$ and $G_2$ with the desired properties and assume that $G_1$ contains a cycle $C$.
Without loss of generality, the root $v_2$ of $G_2$ is contained in the exterior of $C$.
Since the drawings of $G_1$ and $G_2$ are tangent in $v_1$ and $v_2$, also the root $v_1$ of $G_1$ is contained in the exterior of $C$.
Now, from all cycles of either $G_1$ or $G_2$ that are contained in the interior of $C$, let us select a minimal cycle $C'$.
Minimal means that $C'$ does itself not contain any other cycles in its interior.
Such a cycle exists since $G_1$ and $G_2$ are finite.
We assume that $C'$ again belongs to $G_1$, the other case being analogous.
Since the drawings of $G_1$ and $G_2$ are entangled, we get a vertex of $G_2$ in the interior of $C'$.
Starting from this vertex we now follow directed edges in $G_2$.
Since the drawings of $G_1$ and $G_2$ are disjoint, we never leave the interior of $C'$, which in particular means that we never reach $v_2$.
However, since $G_2$ only has a finite number of vertices and all except for $v_2$ have an outgoing edge, we are bound to get into a cycle eventually.
Clearly, this new cycle of $G_2$ is still contained in the interior of $C'$, in contradiction to minimality.
\end{proof}
We make two adaptations to the abstract framework from \cref{sec:framework}.
Note that all definitions and lemmas from that section extend naturally to the following setting.
A set of units $\units$ is no longer understood as a simple subset of $\frags$.
Firstly, in this section we restrict units to be segments from the set $\segs$.
Secondly, a unit can have additional information attached to it.
As an example, $\units$ could be defined as the set of directed segments.
That is, each $\unit$ in $\units$ would correspond to an element of $\segs$, but it would also have a direction.
In particular, this means that multiple elements of $\units$ can correspond to the same geometric graph.
Moreover, a combination $\cmb$ of $\units$ is still understood as a subset of $\units$.
However, we do not allow the same geometric graph to appear twice in $\cmb$.
That is, in the above example, the elements of $\cmb$ must be pairwise distinct as segments.
It is not sufficient if only their directions differ.
We conclude by giving some definitions and conventions that will be used in the following two subsections.
We refer to \Cref{fig:drains} for illustrations.
We assume there are unique points $\hat{\pt},\check{\pt} \in \pts$ with largest and smallest $y$-coordinates, respectively.
The horizontal line through $\check{\pt}$ is called the \emph{bottom}.
\begin{figure}[b]
\begin{center}
\input{fig/drains}
\end{center}
\caption{
The thick line is the bottom.
Dotted lines are borders.
The shaded region is a face with out-degree 1.
Points which expose a drain to the left are marked with $\protect\inlinepoint{drlft}{unmarked}$.
Points which expose a drain to the right are marked with $\protect\inlinepoint{drrgt}{unmarked}$.
}
\label{fig:drains}
\end{figure}
For every crossing-free combination $\cmb$ of some set of units $\units$ we define a set of faces as follows.
From the endpoints of each segment $\unit$ in $\cmb$ we draw vertical rays (called \emph{borders}) downwards until we hit either the bottom or the relative interior of another segment in $\cmb$.
Then, a \emph{face} in $\cmb$ is a maximal connected region in the plane.
There is one unbounded region above the bottom, which is called the \emph{infinite face}.
The unbounded region below the bottom is not a face and will be ignored.
Furthermore, we say that two faces in $\cmb$ are \emph{adjacent} if they share a (vertical) border.
Borders are always directed either left-to-right or right-to-left.
In a combination $\cmb$, the \emph{out-degree} of a face is the number of borders directed away from that face.
We further say that a point $\pt$ \emph{exposes a drain to the left} if the border below $\pt$ is directed left-to-right and the region directly to the left of that border belongs to the infinite face.
An analogous definition is given for \emph{exposing a drain to the right}.
If a point exposes a drain either to the left or to the right, we simply say that it \emph{exposes a drain}.
\subsection{Spanning Trees}\label{subsec:st}
We define a very special set $\segsst$ of units.
Each $\unit$ in $\segsst$ is a segment from the set $\segs$ with a direction.
Additionally, below either endpoint of $\unit$ a border might be attached that is directed either left-to-right or right-to-left.
We also define the set $\cmbslst$, which contains all crossing-free combinations $\cmb$ of $\segsst$ with the following additional properties.\footnote{There is one technicality we gloss over, which however can be made precise: We also require that for each $\pt \in \cpts{\cmb}$ only the first segment $\unit$ in $\cmb$ with $\pt$ as an endpoint has a border attached below $\pt$, where first refers to the usual order induced by serializability of $\cmbscf{\segs}$. In this way, for each picture in \cref{fig:stincoherent} (a), (b) and (c), there exists a unique corresponding object in $\cmbslst$.
}
In what follows, the \emph{out-degree} of a point $\pt \in \pts$ in $\cmb$ denotes the number of segments in $\cmb$ that have $\pt$ as an endpoint and are directed away from $\pt$.
\begin{itemize}
\item The point $\hat{\pt}$ has out-degree 0 in $\cmb$.
\item Every point $\pt \in \pts$ has out-degree at most 1 in $\cmb$.
\item Every point $\pt \in \clow{\cmb}$ has out-degree 1 in $\cmb$.
\item Every finite face in $\cmb$ has out-degree 1.
\end{itemize}
Examples can be seen in \Cref{fig:stincoherent} (a), (b) and (c).
The combination in (d) violates the last three properties.
\begin{figure}
\begin{center}
\input{fig/stincoherent}
\end{center}
\caption{
The elements of $\protect\cmbslst$ in (a) and (c) cannot be considered equivalent.
}
\label{fig:stincoherent}
\end{figure}
Note that all of the above properties are maintained when removing the right-most extreme element from a combination from the set $\cmbslst$.
We thus get the following lemma.
\begin{restatable}{lemma}{stserializable}
\label{lem:st:serializable}
For any point set $\pts$, $\cmbslst$ is serializable.
\end{restatable}
We reuse the three colors $\inlinepoint{free}{unmarked},\inlinepoint{alive}{unmarked},\inlinepoint{dead}{unmarked}$ and the special marking from \cref{subsec:crossingfree} with their original meaning to describe elements of $\cmbslst$, as already depicted in \cref{fig:stincoherent}.
However, that same figure illustrates that we cannot reuse the old equivalence relation because it is not coherent.
There are three problems we have to deal with.
Firstly, the out-degree of a point can become larger than 1.
Secondly, a point with out-degree 0 can disappear in the lower shadow of a segment.
Thirdly, a finite face with out-degree not equal to 1 can be created.
To make the equivalence relation coherent, it suffices to partition the points with color $\inlinepoint{alive}{unmarked}$ into six smaller categories.
That is, we have to replace the color $\inlinepoint{alive}{unmarked}$ with 6 new colors, resulting in a total of 8 colors, and then consider two combinations equivalent if they agree in that new coloring and also in the special marking.
We will not define the colors explicitly here, but only explain what information we have to keep track of.
For each point with color $\inlinepoint{alive}{unmarked}$ we keep track of its out-degree, that is, whether it is currently 0 or 1.
This allows us to avoid the first two problems mentioned earlier.
Furthermore, for each point with color $\inlinepoint{alive}{unmarked}$ we keep track of whether and where it exposes a drain.
This allows us to avoid the third problem because whenever a new segment $\unit$ is added to a combination $\cmb$, a new finite face is created below $\unit$, and the out-degree of that face is determined by the number of exposed drains in the lower shadow and at the endpoints of $\unit$.
Indeed, observe that borders corresponding to exposed drains in the lower shadow of $\unit$ become out-borders of the new finite face.
Also, an exposed drain at the left endpoint, say, of $\unit$ becomes an out-border of the new face if and only if it is exposed to the right.
This is why we also have to know the side a drain is exposed to.
We define an equivalence relation $\parteq$ on $\cmbslst$ based on the $2 + 2 \cdot 3 = 8$ colors from the preceding discussion, and with the usual marking.
The above intuition can be made precise, and the following can be proved.
\begin{restatable}{lemma}{stcoherent}
\label{lem:st:coherent}
The equivalence relation $\parteq$ on $\cmbslst$, as defined above, is coherent.
\end{restatable}
We define the set $\target \subseteq \quotient{\cmbslst}{\parteq}$ which contains $\partclass{\cmb}$ if and only if every point (except for $\hat{\pt}$) has out-degree 1 in $\cmb$ and the infinite face has out-degree 0 (equivalently, there are no exposed drains in $\cmb$).
Let $\agraphst{\pts}$ be the combination graph corresponding to the combination problem $(\segsst,\cmbslst,\parteq,\target)$.
\Cref{thm:st} with $c=8$ now follows from \cref{lem:represent} and from the following insight.
\begin{figure}
\begin{center}
\input{fig/trees}
\end{center}
\caption{
Illustrations for the proof of \cref{lem:st:bijection}.
}
\label{fig:trees}
\end{figure}
\begin{lemma}
\label{lem:st:bijection}
There is a natural bijection between the sets $\bigcup \target$ and $\sts$ in the following sense.
For any combination $\cmb$ in $\bigcup \target$, building the geometric graph on $\pts$ with edges that correspond to the segments in $\cmb$ yields the corresponding crossing-free spanning tree in $\sts$.
\end{lemma}
\begin{proof}
Any $\cmb$ in $\bigcup \target$ induces two directed multigraphs $G_1$ and $G_2$ with corresponding plane drawings, as follows.
$G_1$ is the graph with vertex set $\pts$ and edges that correspond to the directed segments in $\cmb$.
By definition of $\cmbslst$ and $\target$, we at least know that $G_1$ is root-oriented towards $\hat{\pt}$, as exemplified in \cref{fig:trees} (a).
For the vertices of $G_2$ we choose one arbitrary point in the interior of each face in $\cmb$.
Two vertices in $G_2$ are connected if their corresponding faces in $\cmb$ are adjacent.
The direction of that edge is chosen in accordance with the direction of the corresponding border in $\cmb$.
Again by definition of $\cmbslst$ and $\target$, $G_2$ is root-oriented towards the vertex corresponding to the infinite face in $\cmb$, as exemplified in \cref{fig:trees} (b).
It is clear that the drawings of $G_1$ and $G_2$ can be chosen such that they are disjoint, entangled, and tangent in $\hat{\pt}$ and the infinite face.
The proof is concluded by applying \cref{lem:entangledtrees} and by observing that any spanning tree on $\pts$ can be root-oriented towards $\hat{\pt}$ in a unique way.
\end{proof}
One can prove that $c<8$ by adapting the arguments from \cref{subsec:crossingfree}.
With some more work, we get $c<7.04313$, as will be shown in \cref{sec:bounds}.
\subsection{Spanning Cycles}\label{subsec:sc}
We define a slightly different set of units $\segssc$.
In the same way as in the previous subsection, below the endpoints of any segment $\unit$ in $\segssc$ directed borders can be attached.
Here, however, the segment $\unit$ itself does not have a direction.
We also define the set $\cmbslsc$, which contains all crossing-free combinations $\cmb$ of $\segssc$ with the following additional properties.
In what follows, the \emph{degree} of a point $\pt \in \pts$ in $\cmb$ stands for the number of segments in $\cmb$ that have $\pt$ as an endpoint.
Also, if the size of $\cmb$ is $\ptsnum$, the \emph{last finite face} in $\cmb$ is defined as the face directly below the right-most extreme segment in $\cmb$.
All other finite faces are called \emph{normal}.
\begin{itemize}
\item If $|\cmb| = \ptsnum$, the last finite face in $\cmb$ has out-degree 0.
\item Every point $\pt \in \pts$ has degree at most 2 in $\cmb$.
\item Every point $\pt \in \clow{\cmb}$ has degree 2 in $\cmb$.
\item Every normal finite face in $\cmb$ has out-degree 1.
\end{itemize}
Note again that the above properties are maintained when removing the right-most extreme element from a combination from the set $\cmbslsc$.
\begin{restatable}{lemma}{scserializable}
\label{lem:sc:serializable}
For any point set $\pts$, $\cmbslsc$ is serializable.
\end{restatable}
It is clear that we have to keep track of the degrees of all points.
For one last time, we change the meaning of the colors $\inlinepoint{free}{unmarked}$, $\inlinepoint{alive}{unmarked}$, $\inlinepoint{dead}{unmarked}$ and use them to identify points of degree 0, 1, and 2, respectively.
\begin{figure}
\begin{center}
\input{fig/scincoherent}
\end{center}
\caption{
The elements of $\protect\cmbslsc$ in (a) and (c) cannot be considered equivalent.
}
\label{fig:scincoherent}
\end{figure}
Unsurprisingly, and similar to the previous subsection, an equivalence relation based only on these three colors and the usual marking is not coherent, as illustrated in \Cref{fig:scincoherent}.
The only problem, however, is that finite faces which do not have the right out-degree can be created.
A by now routine proof shows that these three colors are already sufficient to avoid crossings.
To avoid finite faces with wrong out-degrees, we split $\inlinepoint{alive}{unmarked}$ into 3 subcolors, and we split $\inlinepoint{dead}{unmarked}$ into 2 subcolors, giving us a total of 6 different colors which are then used to define an equivalence relation $\parteq$ on $\cmbslsc$.
For each point with color $\inlinepoint{alive}{unmarked}$ we keep track of whether it exposes a drain.
If it does, then we also keep track of whether it is to the left or to the right.
Remember, this extra information is relevant if and only if the point in question is one of the endpoints of a new segment.
For a point with color $\inlinepoint{dead}{unmarked}$ we only keep track of whether it exposes a drain or not.
The reason why this is sufficient is that such a point has, by definition, degree 2 already and cannot be an endpoint of a new segment.
\begin{restatable}{lemma}{sccoherent}
\label{lem:sc:coherent}
The equivalence relation $\parteq$ on $\cmbslsc$, as defined above, is coherent.
\end{restatable}
We define the set $\target \subseteq \quotient{\cmbslsc}{\parteq}$ which contains $\partclass{\cmb}$ if and only if every point has degree 2 in $\cmb$ and the infinite face has out-degree 0.
Let $\agraphsc{\pts}$ be the combination graph corresponding to $(\segssc,\cmbslsc,\parteq,\target)$.
\Cref{thm:sc} with $c=6$ now follows from \cref{lem:represent} and from the following insight.
The better bound $c < 5.61804$ will be proved in \cref{sec:bounds}.
\begin{lemma}
\label{lem:sc:bijection}
There is a natural bijection between the sets $\bigcup \target$ and $\scs$ in the following sense.
For any combination $\cmb$ in $\bigcup \target$, building the geometric graph on $\pts$ with edges that correspond to the segments in $\cmb$ yields the corresponding crossing-free spanning cycle in $\scs$.
\end{lemma}
\begin{proof}
For any $\cmb$ in $\bigcup \target$ we know that each point is of degree 2, which means that $\cmb$ is a set of disjoint cycles, as exemplified in \cref{fig:cycles} (a).
Similar to the proof of \cref{lem:st:bijection}, $\cmb$ induces two directed multigraphs $G_E$ and $G_O$.
\begin{figure}[b]
\begin{center}
\input{fig/cycles}
\end{center}
\caption{
White faces are vertices in $G_E$.
Shaded faces are vertices in $G_O$.
}
\label{fig:cycles}
\end{figure}
The vertex set of $G_E$ is the set of faces in $\cmb$ contained in an even number of cycles.
The vertex set of $G_O$ is the set of faces in $\cmb$ contained in an odd number of cycles.
In both $G_E$ and $G_O$, two vertices are connected by an edge if the corresponding faces in $\cmb$ are adjacent.
The direction of each edge reflects the direction of the corresponding border in $\cmb$.
By definition of $\cmbslsc$ and $\target$, $G_E$ and $G_O$ are root-oriented towards the infinite face and the last finite face in $\cmb$, respectively, as exemplified in \cref{fig:cycles} (b).
Clearly, there exist plane drawings of $G_E$ and $G_O$ which are disjoint, entangled, and also tangent in the respective roots.
Applying \cref{lem:entangledtrees} hence concludes the proof.
\end{proof}
|
\section{Introduction}
Fast radio bursts (FRBs) are millisecond-duration radio pulses with
mysterious physical origin.
With the exception of one FRB detected by the Arecibo telescope
\citep{Spitler2014} and one by the Green Bank Telescope (GBT), all the
observed FRBs were detected by the Parkes telescope
\citep[e.g.,][]{Lorimer2007,Keane2012,Thornton2013,Keane2016}.
The observed large values of dispersion measure indicate that FRBs
are probably extragalactic origin \citep[e.g.,][]{Katz2016}.
Several models have been introduced to interpret FRBs,
including giant flares from a magnetar \citep{Popov2010,Kulkarni2014},
giant pulses from a young pulsar or a magnetar
\citep{Cordes2016,Pen2015,Lyutikov2016},
the merger of neutron stars (NSs) or white dwarfs (WDs)
\citep{Totani2013,Kashiyama2013,WangJ2016},
the collapse of a massive NS into a black hole (BH)
\citep{Falcke2014,Zhang2014},
the collision between an NS and an asteroid or a comet \citep{Geng2015},
and the collapse of charged BHs \citep{Zhang2016a,Zhang2016b,Liu2016}.
Recently \citet{Spitler2016} showed ten additional bursts from FRB 121102
by Arecibo from May to June, 2015. Moreover, \citet{Scholz2016} showed six
additional bursts from November to December, 2015, among which five bursts
were detected by GBT and one by Arecibo. Thus, for this unique known
repeating FRB, there are totally 17 observed bursts as shown in Table~2 of
\citet{Scholz2016}. Obviously, the repeating bursts can rule out the models
involving cataclysmic events such as the merger of compact stars
or the collapse of a massive NS into a BH.
As mentioned in \citet{Scholz2016}, the probable extragalactic distance
and repeating behavior may favor giant pulses from a young pulsar
or a magnetar \citep{Cordes2016,Pen2015}, or radio counterparts to
magnetar X-ray bursts \citep{Lyutikov2002,Popov2013,Katz2015}.
In addition, \citet{Dai2016} suggested an interesting model that
a repeating FRB originates from a highly magnetized pulsar encountering
with lots of asteroids in an asteroid belt.
Very recently, \citet{WangF2016} studied the frequency distributions
of peak flux, fluence, duration and waiting time for this repeating source.
In this $Letter$, we propose a different model to understand the repeating
behavior of FRB 121102. Our binary system consists of a magnetized WD and an NS
with strong bipolar magnetic fields. We show that the intermittent Roche lobe
overflow mechanism can be responsible for the repeating behavior.
The remainder of this $Letter$ is organized as follows.
A compact binary model is illustrated in Section~2.
Analyses of the time interval between two adjacent bursts are studied
in Section~3.
Application of such a model to the repeating FRB 121102 is presented
in Section~4.
Conclusions and discussion are made in Section~5.
\section{Compact binary model}
In this section we propose a compact binary model for FRB repeaters.
The binary system consists of a magnetic WD and an NS with strong bipolar
magnetic fields. As illustrated in Figure~1, when the WD fills its
Roche lobe and the system is therefore called semi-detached, mass transfer
will occur from the WD to the NS. The materials in the atmosphere
(the yellow region) of the WD can pass through the inner Lagrange point
($L_1$ point) and then be accreted by the NS, as shown in Figure~1(a).
When the accreted magnetized
materials approach the NS surface, magnetic reconnection may be triggered
and therefore the electrons can be instantaneously accelerated to
an ultra-relativistic speed \citep[e.g.,][]{Zhang2011}.
Consequently, strong electromagnetic radiation can be released
by the curvature radiation as the electrons move along the NS
magnetic field lines.
On the other hand, as studied by \citet{King2007} based on the conservation
of angular momentum, the mass transfer will be a continuous behavior
for $q > 2/3$, where $q$ is defined as the mass ratio of the WD to the NS.
On the contrary, for $q < 2/3$, the WD may be kicked away
and the system may become detached after a Roche lobe overflow process,
as shown in Figure~1(b).
The system can become semi-detached again through the gravitational radiation
and therefore the next transfer process and corresponding
magnetic reconnection can restart.
Thus, for the case of $q < 2/3$, the Roche lobe overflow may be an
intermittent type.
Since the mass of an NS is normally larger than $1.4M_{\sun}$ and the
mass distribution for WDs is peaked at $0.6M_{\sun}$, the intermittent
type of Roche lobe overflow may be a common phenomenon
for the semi-detached NS-WD systems.
We would point out that the cataclysmic variable (CV) is an analogous system.
However, accretion in CVs is continuous rather than
episodic. In our opinion, the main difference is that, in our model
the accreted materials are from a WD instead of a main sequence star.
We may therefore expect a much more violent mass transfer than in CVs.
As a consequence, the outward moving due to the conservation of angular
momentum may dominate over the inward moving owing to the gravitational
radiation, and therefore the WD may be kicked away after a mass transfer
process. We would stress that the violent mass transfer and the corresponding
kick-away behavior is a fundamental assumption in our model.
The characteristic frequency of the curvature radiation
of relativistic electrons is expressed as
\begin{equation}
\nu_{\rm c} = \frac{3c \gamma^3}{4\pi R_{\rm c}}
= 1.5 \left( \frac{\gamma}{60} \right)^3
\left( \frac{R_{\rm c}}{10^6 {\rm cm}} \right)^{-1}~{\rm GHz} \ ,
\end{equation}
where $\gamma$ is the Lorentz factor of electrons, and $R_{\rm c}$ is
the curvature radius. The above equation means that, for a typical radius
$R_{\rm c} = 10^6 {\rm cm}$, an FRB with $\nu_{\rm c}$ of order
$\sim$ GHz requires $50 \la \gamma \la 100$, which is a reasonable
range according to the acceleration by magnetic reconnection
\citep[e.g.,][]{Zhang2011,Kowal2012}.
The duration of an FRB may be regarded as around the timescale
of a magnetic reconnection.
According to our model, the energy-released region is near the NS surface,
and the duration $t_{\rm w}$
can be estimated by the ratio of
the NS radius $R_{\rm NS}$ to the Alfv\'{e}n speed
$v_{\rm A}(=B_{\rm NS}/\sqrt{4\pi \bar\rho})$:
\begin{equation}
t_{\rm w} = \frac{R_{\rm NS}}{v_{\rm A}}
= 1.1 \left( \frac{R_{\rm NS}}{10^6 {\rm cm}} \right)
\left( \frac{B_{\rm NS}}{10^{11} {\rm G}} \right)^{-1}
\left( \frac{\bar\rho}{10^3 {\rm g~cm^{-3}}} \right)^{\frac{1}{2}}~{\rm ms} \ ,
\end{equation}
where $B_{\rm NS}$ is the magnetic flux density, and $\bar\rho$ is the
averaged mass density of accreted materials.
Equation~(2) shows that, for typical values
$R_{\rm NS}=10^6 {\rm cm}$, $B_{\rm NS} = 10^{11} {\rm G}$, and
$\bar\rho = 10^3 {\rm g~cm^{-3}}$, an FRB will be a millisecond duration.
Here, we would stress that the mass density in the atmosphere of a WD varies
continuously from large interior values to essentially zero.
In our analyses, for simplicity, we adopt an averaged mass density
$\bar\rho = 10^3 {\rm g~cm^{-3}}$ for the accreted materials.
In Section~5, we will discuss the degeneracy of $\bar\rho$ with other
physical quantities in our model.
\section{Time interval analyses}
In this section we investigate the analytic variation of the time interval
between two bursts with the transferred mass during the former burst.
We assume a circular orbit for the analyses, i.e., the eccentricity $e=0$.
The orbital angular momentum $J$ of a binary system takes the form
\begin{equation}
J = M_1 M_2 \left( \frac{Ga}{M} \right)^{\frac{1}{2}} \ ,
\end{equation}
where $M_1$ and $M_2$ are respectively the NS and WD mass,
$M = M_1 + M_2$ is the total mass, and $a$ is the binary separation.
When the WD fills its Roche lobe, mass transfer will occur from the WD
to the NS. We assume $\Delta M_2$ as the transferred mass during one
burst, where $\Delta M_2$ is negative. Then, the orbital angular
momentum carried by the accreted materials can be expressed as
\begin{equation}
\Delta J = -\lambda \Delta M_2 \Omega (b_1 - a_1)^2 \ ,
\end{equation}
where $\Omega$ is the orbital angular velocity, $a_1$ is the distance
between the NS and the center of the mass, $b_1$ is the distance
between the NS and the $L_1$ point, and $\lambda$ is a parameter
probably in the range $0 \leqslant \lambda \leqslant 1$.
The expression of $b_1$ takes the form \citep[e.g., Chap.~4.4 of][]{Frank2002}
\begin{equation}
\frac{b_1}{a} = 0.5 - 0.227\log{q} \ ,
\end{equation}
where $q \equiv M_2/M_1$.
For $0.1 \la q \la 0.8$ it is often convenient to adopt a simple form
to describe the Roche lobe radius of the secondary $M_2$:
\begin{equation}
\frac{R_2}{a} = 0.462 \left( \frac{M_2}{M} \right)^{\frac{1}{3}} \ .
\end{equation}
The radius of a WD is expressed as
\begin{equation}
R_{\rm WD} = 10^9 \left( \frac{1.402\times 10^{33}~{\rm g}}{M_2} \right)
^{\frac{1}{3}} \ {\rm cm} \ .
\end{equation}
In our scenario, the mass transfer will occur for $R_{\rm WD} \ga R_2$.
After a mass transfer process, the WD may be kicked away and the
next transfer can appear through the gravitational radiation.
For the simple case with $e=0$, the variation of $a$ due to the
gravitational radiation can be written as \citep{Peters1964}
\begin{equation}
\frac{da}{dt} = -\frac{64}{5} \frac{G^3 M_1 M_2 M}{c^5 a^3} \ .
\end{equation}
Based on Equations (3-8), we can derive an analytic relation between
the time interval $\Delta t$ between two adjacent bursts and
the transferred mass $\Delta M_2$ during the former burst,
\begin{equation}
\Delta t = \frac{5 (1+q) c^5 a^4}{32 q G^3 M^3}
\left\{ \lambda \left[ 0.5(1-q) - 0.227(1+q) \log{q} \right]^2
- (1+q)(\frac{2}{3} - q) \right\} \frac{\Delta M_2}{M_2} \ .
\end{equation}
The above equation enables us to obtain $\Delta t$ once $\Delta M_2$ is given.
\section{Application to FRB 121102}
As mentioned in Section~1, for the unique known repeating FRB 121102,
there are totally 17 observed bursts, as shown in Table~2 of
\citet{Scholz2016}.
In this section, we focus on the time intervals between two adjacent
bursts measured during periods of continuous observation.
Based on the data from \citet{Spitler2016} and \citet{Scholz2016},
we can derive totally ten such intervals among the 17 bursts, as shown
by our Table~1. It is seen that the ten intervals are from tens to hundreds
of seconds. According to our scenario, an FRB is related to a magnetic
reconnection event.
The physical mechanism of such reconnection may be
attributed to a self-organized criticality process \citep{Wang2013}.
Now we can compare our analyses with the observations of FRB 121102,
mainly on the relation between the time interval $\Delta t$ and
the transferred mass $\Delta M_2$.
In our model we adopt typical masses for the NS and the WD, i.e.,
$M_1 = 1.4 M_{\sun}$ and $M_2 = 0.6 M_{\sun}$.
The corresponding separation for the critical situation
$R_{\rm WD} = R_2$ is $a = 3.41\times 10^9$~cm.
There are two extreme cases that the accreted materials do not carry
the orbital angular momentum ($\lambda = 0$), and the materials
carry the Keplerian angular momentum at the $L_1$ point ($\lambda = 1$).
Previous studies on CVs indicate that $\lambda$ is
around unity. We therefore adopt $\lambda = 1$ in the following analyses.
Then, Equation~(9) can be simplified as
\begin{equation}
\Delta t = - 2.68 \times 10^{10} \frac{\Delta M_2}{M_{\sun}} \ {\rm s} \ .
\end{equation}
We try to show a comparison between the analyses and
the observations based on two observational quantities, i.e.,
the fluence $F$ and the time interval $\Delta t$.
A strong magnetic field with $B_{\rm WD} \approx 3\times 10^8$G is required
for the WD (magnetic fields have been discovered in over 100 WDs,
ranging from $10^3$G to $10^9$G).
The averaged mass density of the atmosphere is adopted as
$\bar\rho = 10^3$g~cm$^{-3}$.
Here, we choose $z=0.1$ according to the DM measurement,
and the luminosity distance is $D_{\rm L} = 463.4~{\rm Mpc}$
for $H_0=69.6$, $\Omega_{\rm M}=0.286$, and $\Omega_{\rm vac}=0.714$.
In our model, it is difficult to estimate how many electrons have been
accelerated to an ultra-relativistic speed. Alternatively, we try to evaluate
the released energy by the magnetic energy carried by the accreted materials.
In a real case, the magnetic diffusion and the magnetorotational instability
(MRI) may occur in the accretion flow and therefore change the magnetic energy.
Here, for simplicity, we just assume that the carried magnetic energy
keeps unchanged during the accretion process.
Thus, we can evaluate the fluence $F$ by the total magnetic
energy carried by the accreted materials:
\begin{equation}
F \Delta\nu D_{\rm L}^2 \Delta\theta
= - \eta \frac{B_{\rm WD}^2}{8\pi} \frac{\Delta M_2}{\bar\rho} \ ,
\end{equation}
where $\Delta\nu$ is the width of radio frequency,
$\Delta\theta$ is the solid angle for an FRB,
and $\eta$ is the efficiency of the released energy through
the curvature radiation compared with the total magnetic energy
carried by the accreted materials.
We adopt $\Delta\nu = 1$GHz and $\Delta\theta = 0.04\pi$
(corresponding to 1\% of the whole space) for the calculations.
A comparison of analyses with observations in the $\Delta t-F$ diagram
is shown in Figure~2, where $\Delta t$ is the time interval in the rest frame.
The three solid lines represent the analytic
relation for the efficiency $\eta = 0.001$, 0.01, and 0.1,
which are derived by combining Equations (10-11).
The ten color filled circles denote the
observational results shown in Table~1.
It is seen from Figure~2 that all the bursts are well
located among the three theoretical lines, which implies that the analyses
are roughly in agreement with the observations.
Thus, the NS-WD binary model is likely to be responsible for the FRB repeaters.
\section{Conclusions and discussion}
In this $Letter$ we have proposed an NS-WD binary model for the FRB repeaters.
The system consists of a magnetic WD and an NS with strong bipolar
magnetic fields.
Mass transfer will occur through the $L_1$ point
when the WD fills its Roche lobe.
Consequently, the accreted magnetized materials may trigger
magnetic reconnection when they approach the NS surface, and therefore
the electrons can be accelerated to an ultra-relativistic speed.
In this scenario, an FRB can be powered by the curvature radiation
of the relativistic electrons moving along the NS magnetic field lines.
Owing to the conservation of angular momentum, the WD may be
kicked away after a burst, and the next burst may appear when
the system becomes semi-detached again through the gravitational radiation.
Our analyses have shown that such an intermittent Roche lobe overflow
mechanism can be responsible for the repeating behavior of FRB 121102.
For the application of our model to FRB 121102, we would point out that
the required values for physical quantities are degenerate, as indicated
by Equations (2) and (11). For instance, if the averaged mass density is
$\bar\rho = 10 {\rm g~cm^{-3}}$ instead of $10^3 {\rm g~cm^{-3}}$,
these two equations can be again satisfied with $B_{\rm NS} = 10^{10}$G
and $B_{\rm WD} = 3\times 10^7$G. In other words, our model can work
for the repeating FRB with another group of parameters. Actually,
as shown by Equation~(11), either a smaller $\eta$ or a larger
distance $D_{\rm L}$ may be equivalent to a smaller $B_{\rm WD}$.
Thus, Section~4 just shows an example group of parameters
which can work for FRB 121102.
In addition, we would again stress that a fundamental assumption
in our model is that, once the Roche lobe is filled, the mass transfer will
be so violent that the WD can be kicked away after a mass transfer process.
In this work we have focused on the curvature radiation which
mainly contribute to the radio emission. However,
the synchrotron radiation may also be of importance for relativistic
electrons moving in magnetic fields. In the case of $B_{\rm NS}=10^{11}$G
for the NS,
the characteristic frequency for the synchrotron radiation is
$\nu_{\rm s} = 3\gamma^3 eB_{\rm NS}/4\pi m_{\rm e}c \sim 1{\rm MeV}$ for
$\gamma \sim 60$.
Since the energy of curvature (synchrotron) radiation is related to
the component of kinetic energy of electrons parallel (perpendicular)
to the magnetic field lines, we can regard the energy budget for
these two radiation mechanisms is comparable.
As a consequence, it is easy to estimate that
the fluence of 1Jy~ms with 1GHz width is equivalent to $\sim 0.06$eV
for the total energy with a square meter high-energy detector, which means
that the probability of receiving one MeV photon by such a detector
is less than $10^{-7}$.
Thus, the possible gamma-ray emission by the synchrotron radiation
can hardly be detected except that the FRB occurs in the Local Group.
As mentioned in Section~1, \citet{Dai2016} introduced a model related
to asteroids. Their radiation mechanism is also the curvature radiation.
The difference is that, the energy source for FRBs in their model
is the kinetic energy of asteroids, whereas the source in our model
is the magnetic energy of accreted materials. In our scenario, it is
still possible for a part of the kinetic energy of accreted materials to
be transferred to the electrons through the strong magnetic fields,
and therefore have contribution to the radio emission.
In such case, the required magnetic fields of the WD
may be significantly weaker than the present assumption.
\acknowledgments
We thank Zhongxiang Wang, Bing Zhang, and Zi-Gao Dai for beneficial
discussions, and the referee for constructive suggestions that
improved this $Letter$.
This work was supported by the National Basic Research Program of China
(973 Program) under grant 2014CB845800,
the National Natural Science Foundation of China under grants 11573023,
11522323, 11473022, 11473021, 11333004, 11222328, U1531130, and U1331101,
and the Fundamental Research Funds for the Central Universities
under grants 20720140532, 20720150024, 20720160023, and 20720160024.
|
\section{Introduction}
Determination of the physical properties of low mass stars and brown dwarfs,
most importantly luminosities, depends upon having accurate
distances. However, these late-type objects were generally too faint for
inclusion in the all-sky Hipparcos survey. Through the efforts of several
ground-based astrometric surveys, there are now hundreds of low-mass stars
with parallaxes
\citep[e.g.][]{Dahn:2002,Jao:2005,Costa:2006,Andrei:2011,Dupuy:2012,Faherty:2012,Dieterich:2014,Sahlmann:2014,ZapateroOsorio:2014}.
Nevertheless, there are still many objects within 30 pc without
well-measured distances. These nearby, bright objects would be the best
templates for studies of radii, atmospheric composition, metalicity, and
other spectroscopic properties. In addition, low mass stars with excellent
distances provide the templates for spectrophotometric distances to more
distant stars.
In 2007, we began a long-term astrometric search for gas giant planets and
brown dwarfs orbiting nearby low mass dwarf stars \citep{Boss:2009}. The
search employs a specialized astrometric camera, the Carnegie Astrometric
Planet Search Camera (CAPSCam), with a design optimized for high accuracy
astrometry of M dwarf stars. Here we report our trigonometric parallaxes for
134 low mass stars. Of these, 38 have no previously reported measured parallax.
\section{Observations}
CAPSCam operates on the 2.5-m du Pont telescope at the Las Campanas
Observatory in Chile and is described in detail by \citet{Boss:2009}; its main
features for astrometry of low mass stars are briefly described here. CAPSCam
has no internal moving parts and employs an astrometric quality filter as its window
that is approximately z-band (865 nm with a bandpass of 100 nm). The field of
view is 6.7 arcmin on a side, with 2048$\times$2048 pixels each subtending
0$\farcs$196. A subarray, also known as the ``guide window,'' is arbitrarily
sizable and locatable and may be read out independently from the rest of the
field. A bright target star is placed in the guide window, which is then read
out fast enough so the star does not saturate while the rest of the pixels
integrate on the reference stars; a mechanical shutter in front of the
entrance window ensures that the exposure time on the bright star remains, as
much as possible, commensurate with that on the full field. Thus, the camera
can achieve high dynamic range without excessive overhead.
Target selection concentrated on southern M, L and T dwarfs closer than 20 pc
as known from either parallaxes or spectrophotometric distances. At the time
of the initial target selection ten years ago, distances and spectral
sub-types for many late type stars were lacking, so high proper motion stars
were included as well. The earliest spectral type included was M3, and the
majority of targets are spectral type M5.5 and later. In 2011, the target list
was updated to include all objects with spectral type later than M4, closer
than 12 pc, and south of declination $+$16$^\circ$. Stars must have I
magnitudes greater than $\sim$9 so as not to saturate the detector in the
minimum guide window exposure time of 0.2 s. The faintest objects we target
have I$\sim$18 so as to provide S/N$\sim$500 in a 120~s integration.
Our typical observing strategy is to place target stars brighter than
I$\sim$15 in the guide window. Full field integration times are chosen to get
at least 6, and typically more like 25, well-exposed reference stars; the
number of reference stars for each field is given in Table \ref{tab:obs}. The
typical astrometric reference star for our fields has I$\sim$17 and can be as
faint as I$\sim$22. The usual integration times are also given in Table
\ref{tab:obs}, although in some epochs, they were adjusted for seeing and
clouds. At each epoch, we typically observe for an hour and thus obtain 20-40
images of the full field. Targets are almost always observed within an hour
of transit, and given the long wavelength filter of the camera there is little
differential atmospheric refraction as a function of stellar spectral type.
The data for our parallaxes were collected from 2007-2014. The number of
epochs per source varies from 4, the minimum to obtain a parallax with
uncertainty estimates, to more than 20 for a few well-studied targets. The
number of epochs, and the start and end dates for the data, and time baseline
of the observations included in the parallaxes are given in
Table \ref{tab:obs}. We typically observe each star at least twice per
calendar year. The stars range in spectral type from M3 through T7, with the
bulk of the targets being late M-type.
\section{Data Reduction}
Details of CAPSCam astrometric data reduction may be found in
\citet{Boss:2009} and \citet{Anglada:2012} and are briefly summarized here following the
description in \citet{Weinberger:2013}. For each epoch, the x and y pixel
positions of the brightest $\sim$100 stars (more in crowded fields) in the
field are found with a centroiding algorithm. Data from all epochs are
combined in an astrometric solution to derive the positions, proper motions,
and parallaxes of all the cross-matched stars in each target field. The
astrometric solution is an iterative process. An initial catalog of positions
starts with the centroids from a chosen epoch transformed to sky coordinates
based on the coordinates of the target star and the known pixel scale. Next, a
transformation is applied to every other epoch's catalog to match the initial
catalog, and the apparent trajectory of each star is fit to a basic
astrometric model. The parallaxes for all objects are initialized to zero. The
initial catalog is updated with new positions, proper motions, and parallaxes,
and a subset of well-behaved stars is selected to be used as the reference
frame. The reference stars must be successfully extracted in every epoch and a
subset of at least 15, and more typically 30, is chosen that shows the
smallest epoch-to-epoch variation in their solutions. This process is then
iterated a small number of times.
In each iteration the individual parallax and proper motions of every star are
adjusted, so the mean parallax should stay at approximately zero. However, the
subset of reference stars do not necessarily have mean parallax of zero. At
any epoch, the position of a star has centroiding uncertainties, and for
distant stars, proper motion will take out all apparent motion of the star,
leaving positional residuals that are both positive and negative. Therefore,
although the true parallax to every star must be positive, we allow the fit
parallaxes to take on positive and negative values.
To assess the uncertainties on the measured parallax, we perform a Monte Carlo
where we fit the starting position, parallax, and proper motion in each
trial. Each trial draws random positions for each epoch based on the nominal
position determined from the iterative solution and its positional
uncertainty. If the $\chi^2$ of the parallax fit is $>$1, we add to every epoch's
uncertainties and re-fit until $\chi^2$ equals one. This additional uncertainty, or
positional jitter, may arise from any sources of systematic uncertainty. The
final parallax uncertainty is the standard deviation in the parallaxes of each
trial.
The final astrometric solution gives the motion of all the stars in the
field. However, these stars have parallactic motions that are all in the same
direction, since they are generated by Earth's motion. This introduces a small
bias, also known as a zero-point parallax offset, that must be removed to find
the absolute parallax.
To find the zero point for each field, we estimate a photometric distance to
the brightest reference stars by fitting a Kurucz stellar model to cataloged
USNO-B1 magnitudes at B2, R2, and I \citep{Monet:2003} and 2MASS magnitudes
at J, H, and Ks \citep{Skrutskie:2006} and assuming each star is a
dwarf. Dwarf stars with fit Teff $<$ 4000 K are excluded. We average the difference
between our astrometrically determined (even if they are not statistically
significant) and photometric parallaxes to find the average bias and its
uncertainty and subtract it from our relative parallaxes and propagate the
uncertainty. We cannot make a comparable zero-point proper motion correction
because so few stars as faint as our reference stars have measured absolute
proper motions. For 18 of our fields, we were unable to compute a zero point
correction due to a combination of reference stars that were too cool and/or
faint to be fit well. However, inspection of Table \ref{tab:results} shows that
our typical zero point correction is small ($<$ 1 mas) and that the average
correction across the stars for which they were computed is -0.09 $\pm$ 0.43
mas. Therefore, for these 18 objects, we assumed no zero point correction and an
additional uncertainty of 0.4 mas.
\section{Results}
Table \ref{tab:results} lists the relative parallaxes, relative proper motions,
zero-point parallax corrections, and final absolute parallaxes for all our
targets as well as previously published trigonometric parallax values from the
literature. Figure \ref{fig:compare} compares our absolute parallaxes with
published parallaxes from other work.
For 79 of the 96 stars with previously published parallaxes, our measurements have
lower uncertainty. In general there is very good agreement between ours and previous
measurements; only 12 of the 96 disagree by more than 3 $\sigma$ (of the less
accurate measurement), and for 8 of these 12, the difference in parallax is $<$5\%.
The remaining four discrepant sources are explained in more detail, below.
\begin{figure}[ht]
\centering\includegraphics[width=4.5in]{fig_comparelit.ps}
\caption{\small A comparison of parallaxes for 95 of the CAPS targets for which literature values
exist. The diagonal line is drawn as a guide and is not a fit. The obviously discrepant
point is 2MASS J1259-4336, and is discussed in Section \ref{s:J1259}. Not shown is GJ
406, the closest star in our sample, whose parallax is 413 mas.\label{fig:compare}}
\end{figure}
A formal least-squares fit to the published trigonometric parallaxes versus ours gives a
slope of 0.988 $\pm$ 0.003, i.e., the CAPSCam parallaxes are slightly low compared to
published values -- an average of 2.9 mas low. However, the $\chi^2$ of this fit is poor,
which suggests that either the literature uncertainties, our uncertainties, or both, are
underestimated. Note also that this comparison includes the poor matches addressed below.
There are also 38 targets in Table \ref{tab:results} with no previous
trigonometric parallax including 7 stars with spectral types later than M8. A
color-magnitude diagram for all the stars in our sample is shown in Figure
\ref{fig:colormag}. As expected, most of the new nearby objects have the expected
brightnesses and colors of old, field objects. Exceptions are discussed below.
\begin{figure}[htb]
\centering\includegraphics[angle=90,width=5.95in]{figcolormag_revised.ps}
\caption{\small The J-W1 (WISE Band 1) versus $M_J$ color-magnitude diagram including
all the stars for which we obtained parallaxes. Objects without previously
published parallaxes are shown in blue. The M0--M6 field star sequence from
\citet{Pecaut:2013} is shown in a red dashed line while the M6--L8 field
sequence from \citet{Faherty:2016} is shown in a red dotted line. Young objects (aqua circles) lie
above the sequence; their names are not on the plot to avoid crowding, but
they are described in Section \ref{sec:young}. Known
binaries (gray/yellow cirlces) generally lie above the field sequence and are
listed in Section \ref{sec:outliers}. Five other sources that are too
red for their spectral types and/or overluminous are shown with pink
circles and are also discussed in Section \ref{sec:outliers}. \label{fig:colormag}}
\end{figure}
\subsection{Discrepant Sources}
{\bf GJ 3198:} The literature value from \citet{Riedel:2010} is 67.3 $\pm$ 1.2
and our value is 57.2 $\pm$ 1.4. However, our (relative) proper motions agree
well: theirs is (483, -486) mas~yr$^{-1}$ and ours is (480, -474)
mas~yr$^{-1}$ . They have a baseline of 5.3 yr and we have a baseline of 4.1
yr. Our parallax fits are shown in Figure \ref{fig:GJ3198}. So, the source of the
parallax discrepancy is unclear, but our parallax factor coverage is very
good, particularly in right ascension. With either parallax, the star's
position on the color-magnitude diagram is slightly too red for its absolute magnitude
and suggests probably binarity, but the star does not quite make the cuts we
impose to find such objects in Section \ref{sec:outliers}.
\begin{figure}[htb]
\includegraphics[height=2.4in,clip=1]{GJ3198-ra.ps}
\includegraphics[height=2.4in,clip=1]{GJ3198-dec.ps}
\caption{Measurements of the motion of GJ 3198 in right ascension and
declination after removing the proper motion, which otherwise dominates the
scale of the plot. In all the plots shown, the position of the star in the
first epoch of observation is taken to be (0,0). Typical per epoch
uncertainties are $<$1 mas and are plotted but not usually visible within
the symbols. The best fit parallax is shown with the solid
line. \label{fig:GJ3198}}
\end{figure}
{\bf 2MASS J11553952$−$3727350:} Our parallax of 84.4 $\pm$ 0.8 is 20\% smaller
than that of \citet{Faherty:2012}'s 104.4 $\pm$ 4.7. Again, our relative
proper motions agree well: ours is (53.7, -784.49) and theirs is (66.8,
-777.9). They had a baseline of 2.5 yr and we have baseline of 7.1 yr.
Our parallax fits are shown in Figure \ref{fig:2M1155}.
\begin{figure}
\includegraphics[height=2.4in,clip=1]{2M1155-ra.ps}
\includegraphics[height=2.4in,clip=1]{2M1155-dec.ps}
\caption{Measurements of the motion of 2MASS J11553952-3727350 in right
ascension and declination after removing the proper motion, which otherwise
dominates the scale of the plot. The best fit
parallax is shown with the solid line. \label{fig:2M1155}}
\end{figure}
{\bf Ruiz (ESO) 207-61:} In the table, we gave the average of three literature
parallaxes, i.e. 54.7 mas \citep{Ianna:1995,Tinney:1996,vanaltena:1995}, but
the measured values range from 50.4 to 66.1 mas. We get 41.0 $\pm$ 1.6 mas. We
have dropped this source from our program, so we only have 6 epochs, but they
are spread over 5.2 yr with good coverage of the parallax factor. Our parallax
fits are shown in Figure \ref{fig:Ruiz207}.
\begin{figure}[htb]
\includegraphics[height=2.4in,clip=1]{Ruiz207-ra.ps}
\includegraphics[height=2.4in,clip=1]{Ruiz207-dec.ps}
\caption{Measurements of the motion of ESO 207-61 in
ascension and declination after removing the proper motion, which otherwise
dominates the scale of the plot. The best fit
parallax is shown with the solid line. \label{fig:Ruiz207}}
\end{figure}
{\bf 2MASS J12590470$-$4336243: \label{s:J1259}} \citet{Deacon:2005} found a
parallax of 276 $\pm$ 41 mas for this object, which they refer to as
SIPS1259-4336, based on scanned UKST and ESO plates. They noted that their
derived distance (3.6 pc) made the object have an absolute magnitude too
bright for a single dwarf, and suggested it could be a binary. However, we
find a parallax of 129 mas, which puts the object twice as far away, at 7.8
pc, so it need not be a binary, and its absolute magnitude M$_J$=11.09
$\pm$0.05 is consistent with its color of J-W1=1.30 $\pm$ 0.03 for a single
M8. Our proper motion (not adjusted from the apparent value) of 1101.5 $\pm$
1.1 mas~yr$^{-1}$ in RA and $-$253.28 $\pm$ 0.30 mas~yr$^{-1}$ in DEC agrees
quite well with that of Deacon et al.: 1105 $\pm$ 4 mas~yr$^{-1}$ and $-$262
$\pm$ 4 mas~yr$^{-1}$ in RA and DEC respectively. The parallax solution is shown
in Figure \ref{fig:J1259}.
\begin{figure}[htb]
\includegraphics[height=2.4in,clip=1]{2M1259-ra.ps}
\includegraphics[height=2.4in,clip=1]{2M1259-dec.ps}
\caption{Measurements of the motion of 2MASS J12590470-4336243 in right
ascension and declination after removing the proper motion, which otherwise
dominates the scale of the plot. The best fit
parallax is shown with the solid line. \label{fig:J1259}}
\end{figure}
\subsection{Notes on Interesting Individual Sources \label{sec:individual}}
{\bf 2MASS J01365662+0933473:} This nearby brown dwarf, type T2.5, is a
benchmark for the study of atmospheric variability and clouds in cool objects
\citep{Artigau:2009}. It had no previously published
parallax. \citet{Artigau:2006} found a photometric distance of 6.4 $\pm$ 0.3
pc, and our parallax gives a distance consistent with this, namely 6.14 $\pm$ 0.04
pc.
{\bf 2MASS J01392170$-$3936088}: The photometric distance to this source
computed in \citet{Deacon:2007} is 14.99$^{+8.96}_{-5.61}$ pc. Our
trigonometric parallactic distance is 8.80$\pm$0.04 pc, and the location of
the star in the M$_J - \rm (J-W1)$ color-magnitude diagram (Figure
\ref{fig:colormag}) does not look unusual. This is now added to the list of
stars within 10 pc.
{\bf LP 944-20:} This is a low-gravity, i.e. likely young, brown dwarf that is
not co-moving with a known young association \citep{Faherty:2016}. Our
parallax of 154.4 $\pm$ 0.60 mas confirms the parallax measurement 155.9 $\pm$
1.0 mas of \citet{Dieterich:2014} that is markedly different from that of
\citet{Tinney:1996} (201.4 $\pm$ 4.2 mas).
{\bf GJ 3470:} This nearby M dwarf has a Neptune mass planet detected by
radial velocity and transit observations \citep{Bonfils:2012}. It has no
previously published trigonometric parallax; we get 34.15 $\pm$
0.66 mas or 29.28$^{+0.58}_{-0.56}$ pc.
The inferred planetary mass and radius depend sensitively on the stellar
properties. \citet{Demory:2013} measured a stellar density $\rho_\star =
2.91^{+0.37}_{-0.33} \rho_\odot$ and inferred M$_\star$=0.539$^{+0.047}_{-0.043}$ M$_\odot$,
R$_\star$=0.568$^{+0.037}_{-0.031}$ R$_\odot$, and distance=30.7$^{+2.1}_{-1.7}$ pc.
Our new distance is within their uncertainties, but we recompute the best
stellar mass and radius with a Monte Carlo that uses our distance and the
published photometry. Because of the density measurement, there are two nearly
independent methods to find R$_\star$. First, the physical size can be
determined from combining our distance with the K-band magnitude, via the
angular size relation of \citet{Kervella:2004}. Second, the stellar mass can be
determined from the V, J, H and K-band relations of \citet{Delfosse:2000} and
combined with the measured $\rho_\star$ of \citet{Demory:2013} to determine
R$_\star$. We use a Monte Carlo to find the probability densities for both
independent estimates and then multiply the probability densities to get the
combined best estimate and its uncertainty: R$_\star$=0.550 $\pm$ 0.012 R$_\odot$.
Our best stellar radius is again within the uncertainties of the estimate of
\citet{Demory:2013}, but 3.2\% smaller on the mean and with smaller
uncertainty. This also reduces the inferred radius of the planet by the same
amount, and increases the planetary density by 10\% to 0.79 g~cm$^{-3}$.
{\bf 2MASS J20282035+0052265:} This is an L-dwarf binary system that was not
resolved in HST/NICMOS observations analyzed in \citet{Reid:2008} but was
resolved using new analysis techniques of the same data in \citet{Pope:2013}. The latter
work found it to be a nearly equal spectral type binary (L3+L4) and estimated
a new spectrophotometric distance of 26.1 $\pm$ 3.9 pc. Our parallax places
the binary at 30.1 $\pm$ 1.2 pc. We only have four epochs of data so we cannot
say if we observe orbital motion in the astrometric signal; it was dropped
from the planet search program for being too far away.
\subsubsection{Young Sources \label{sec:young}}
Stars can appear overluminous in the color-magnitude diagram
(e.g. Fig. \ref{fig:colormag}) because of youth. The companion to Fomalhaut,
LP 876-10 \citep{Mamajek:2013}, and AP Col \citep{Riedel:2011} are two examples
in Figure \ref{fig:colormag}. We also find two others.
{\bf G 161-71}: The spectrophotometric distance to this source is typically
given as $\sim$6.7 pc \citep{Reid:2002a,Scholz:2005a,Riaz:2006}, but our
parallactic distance is 13.26 $\pm$ 0.14 pc. \citet{Malo:2014} measured an RV
of 13.5 $\pm$ 0.4 km~s$^{-1}$ and listed it as a possible Argus association
member. Using our parallax and proper motions combined with this RV, we
confirm a 99.99\% probability of membership in the 30-50 Myr old Argus
association using the BANYAN I tool \citep{Malo:2013}. An overluminosity of
1.5 mag is possible for such a young star \citep{Gagne:2015}. In addition, the
enhanced X-ray luminosity of this star \citep{Riaz:2006} is also consistent
with that of other young stars \citep{Shkolnik:2009}.
{\bf LP 870-65}: This is an M4 or M4.5 star with a spectrophotometric distance
in \citet{Scholz:2005a} of 8.7 pc, and our parallactic distance is 18.22 $\pm$
0.19 pc. Indeed, \citet{Bowler:2015} identified this star, also known as NLTT
48651, as young based on its X-ray and UV luminosity. That paper also gives a
radial velocity of -7.5 $\pm$ 0.7 (E. Shkolnik, personal communication) and
suggests a tentative association with the AB Dor moving group. Using our
parallax and proper motions combined with this RV, the BANYAN I tool confirms
a 100\% probability of membership in the $\sim$100 Myr old AB Dor Association.
\subsubsection{Overluminous and/or Red Sources \label{sec:outliers}}
{\bf Binaries:} Several known binaries are in our sample; those that are equal
brightness appear overluminous in Figure \ref{fig:colormag}: GJ 2005
\citep{Leinert:1994}, Kelu-1 \citep{Gelino:2006}, G 124-62B \citet{Bouy:2003},
GJ 3900 \citep{Bonfils:2013}, GJ 4074 \citep{Bonfils:2013}, LP 869-19
\citep{Malo:2014}, 2MASS
J20282035+0052265 \citep{Pope:2013}, and $\epsilon$ Indi B
\citep{McCaughrean:2004}. The companions to 2MASS J04234858-0414035
\citep{Burgasser:2005} and 2MASS J13153094-2649513 \citep{Burgasser:2011} are T
dwarfs and do not cause noticable overluminosity. Suprisingly, 2MASS
J02052940-1159296 \citep{Koerner:1999}, which is an equal flux ratio binary,
does not look overluminous.
In addition to these known binaries, we search for stars that appear
overluminous or redder than expected based on their spectral types. For M0 - M6
spectral types, we search for stars that lie redder than the field sequence as
given in \citet{Pecaut:2013}\footnote[1]{Updated at
\url{http://www.pas.rochester.edu/$\sim$emamajek/EEM\_dwarf\_UBVIJHK\_colors\_Teff.txt}}
by more than the combined 1$\sigma$ uncertainties in the dwarf sequence and the
stars' individual color uncertainties. Since Pecaut \& Mamajek do not provide
uncertainties on the colors, we computed J$-$W1 for $\sim$20 stars in each spectral type
bin taken from DwarfArchives.org\footnote[2]{List of M dwarfs at
http://spider.ipac.caltech.edu/staff/davy/ARCHIVE/index.shtml}. For M4, M5 and
M6 stars, we find a color and dispersion of 1.00 $\pm$ 0.06, 1.12 $\pm$ 0.08,
and 1.16 $\pm$ 0.08 mag respetively. We assume a 0.08 mag uncertainty for M0-M3
also. For M7 and later spectral types, we search for stars that lie above the
field sequence given in \citet{Faherty:2016}. Combined, we find four stars that
appear overluminous or redder than expected: DY Psc, GJ 1123, GJ 1129, and 2MASS
J16184503-1321297.
These stars are peculiar. In principle, they could be candidate young
stars. All of these stars have absolute magnitudes that are more than 0.75 mag
from their expected values based on their J$-$W1, so are not just obviously
equal brightness binaries. DY Psc is particularly red (J$-$W1 = 1.82 $\pm$
0.04) for its optically determined spectral type of M9.5 (J$-$W1 =
1.5). However, none of these stars has X-ray emission detected in the ROSAT
all-sky survey \citep{Boller:2016} or were strong UV emitters, at the level of
the known young stars, in the GALEX survey.
\section{Discussion and Summary}
Parallaxes combined with infrared colors can identify interesting low mass stars
and brown dwarfs that are young and/or in multiple systems. Only four of the
targets in our sample are in the Tycho-2 catalog and would therefore be expected
to have full astrometric solutions including parallax in the first GAIA data
release in 2016. These are GJ 3379 (M4), G 108-21 (M3.5), GL 452.1 (M4.5), and
LTT 7434 (M4).
Thirty-two of the stars here are not part of our long-term
monitoring program for any of a number of reasons including: being too far away
($\pi < $50 mas), being a close visual binary or stellar spectroscopic binary,
or having a bad astrometric reference frame. We are continuing to observe all
the stars that have more than 10 epochs in Table \ref{tab:results}.
\acknowledgements
We thank the staff of Las Campanas Observatory for their ongoing support of
this long-term program. Jackie Faherty and Jonathan Gagne provided helpful
input. This work has been supported in part by NSF grant AST-0352912, NASA
Origins of Solar Systems grant NNX09AF62G, and NASA Astrobiology Institute
grant NNA09DA81A. This research has made use of the SIMBAD and Vizier
databases, operated at CDS, Strasbourg, France. This publication makes use of
data products from the Two Micron All Sky Survey, which is a joint project of
the University of Massachusetts and the Infrared Processing and Analysis
Center/California Institute of Technology, funded by the National Aeronautics
and Space Administration and the National Science Foundation. This
publication makes use of data products from the Wide-field Infrared Survey
Explorer, which is a joint project of the University of California, Los
Angeles, and the Jet Propulsion Laboratory/California Institute of Technology,
funded by the National Aeronautics and Space Administration. This publication
made use of the Mikulski Archive for Space Telescopes (MAST). STScI is
operated by the Association of Universities for Research in Astronomy, Inc.,
under NASA contract NAS5-26555. Support for MAST for non-HST data is provided
by the NASA Office of Space Science via grant NNX09AF08G and by other grants
and contracts.
\bibliographystyle{apj}
|
\section{Introduction and Motivation}
Spectroscopic observations provide a very powerful tool to study atmospheric properties and dynamics of solar flares. The long history of these studies includes observations from numerous satellites and rocket missions \citep{Fletcher11,Milligan15}. The currently operating NASA's Interface Region Imaging Spectrograph \citep[IRIS satellite,][]{DePontieu14} observes the chromosphere and chromosphere-corona transition region with high spatial, temporal and spectral resolutions. The IRIS spectral coverage includes several strong lines formed in the upper chromosphere: Mg\,II\,h\&k~2796~and~2803\,{\AA} ($T=8-10\times{}10^{3}K$) and in the lower transition region: C\,II\,1334/1335\,{\AA} ($T=10-20\times{}10^{3}K$) and Si\,IV\,1403\,{\AA} ($T=50-100\times{}10^{3}K$). In the hot plasma of solar flares IRIS can observe the Fe\,XXI\,1354.1\,{\AA} line which corresponds to a forbidden transition and is formed at $1.1\times 10^{7}$\,K. This line appears during flares in the IRIS O\,I spectral window.
Among various physical processes occurring during solar flares, one of the most important is chromospheric evaporation. According to the standard flare model \citep{Carmichael64,Sturrock68,Hirayama74,Kostiuk75,Kopp76,Priest02,Shibata11}, this process is initiated by heating of dense layers of the solar atmosphere and creation of an overpressure region. The dynamical expansion of this region is accompanied by upflows of the hot plasma into the corona, and often by downward motions of relatively cold plasma and shocks. A recent overview of the chromospheric evaporation processes can be found in the paper of \citet{Milligan15}.
IRIS provides a unique opportunity for the chromospheric evaporation studies \citep[see e.g.][]{Battaglia15,Brosius15,Graham15,Li15a,Li15b,Polito15,Tian14,Tian15a,Young15,Sadykov15}. In particular, the Fe\,XXI line appearing only during flares detects the hot upward-moving plasma flows as a Fe\,XXI blueshift. The chromospheric evaporation is also observed in the IRIS UV chromospheric and transition region lines. However, interpretation of the Doppler shift is less straightforward and depends on the energy transfer mechanism and heating rates resulting in ``gentle'' and ``explosive'' types of evaporation (see papers of \citet{Antiochos78,Zarro88} and simulations of \citet{Fisher85a,Fisher85b,Fisher85c} for the details).
The chromospheric evaporation processes are still not well understood. Despite many numerical simulations \citep[e.g.][]{Kostiuk75,Fisher85a,Kosovichev86,Liu09,Rubiodacosta15a,Rubiodacosta15b,Reep15,Reep16}, some details of the process could not be reproduced. One of the most disputed effects is a significant time delay of the coronal evaporation flow relative to the chromospheric response observed as redshift of relatively cold UV lines corresponding to downflowing plasma. \citet{Graham15,Battaglia15,Young15} found the delays of about the 60\,s using IRIS spectral data. However, the numerical simulations of the standard ``thick-target'' flare model predict that both phenomena should occur simultaneously.
There are some attempts to explain this discrepancy. Emission of the Fe\,XXI line might be very weak at the initial and supposedly blueshifted stages of the evaporation, and then became stronger but less blueshifted. This situation is clearly illustrated in the paper of \citet{Graham15}. The weak emission in Fe\,XXI line may happen due to non-equilibrium ionization effects \citep{Battaglia15}. In particular, Fig.~6-8 of~\citet{Bradshaw09} demonstrate that for the number density of $10^8-10^9$\,cm$^{-3}$ the characteristic ionization time can reach $\approx$60\,s for the Fe\,XIX and higher ionization degree ions, which may cause the blueshift delays for about one minute. However, the theory cannot explain the observed delays for a couple of minutes or longer.
In this paper we focus on a detailed spatio-temporal analysis of the chromospheric evaporation during an M\,1.0 class flare occurred on 12 June, 2014 from 21:01\,UT to 21:19\,UT in active region NOAA 12087. At this time the active region was located south-east (heliocentric coordinates S22E49) on the solar disc, and the flare event was well-covered by the IRIS observations in the coarse-raster mode (the IRIS observational set started long before the flare beginning and continued long after the flare decay). The eight slit positions run in a cyclic order with a high cadence ($\approx$\,20\,s for the full cycle) allowed us to study the flare spectra in most of the flare region. Some general properties of the chromospheric evaporation during this flare have already been studied in our previous paper \citep{Sadykov15}. Dynamical and magnetic processes in the vicinity of the magnetic polarity inversion line (PIL) have been studied by~\citet{Kumar15,Sharykin16}. In this paper we study the process of chromospheric evaporation and its relations to the flare magnetic geometry in more detail.
In addition to the spectroscopic data, the knowledge of the magnetic field topology is very important for understanding of the flare dynamics. The magnetic field and corresponding electric current systems are the primary sources of energy of solar flares. They can store the flare energy \citep[about $10^{30}$-$10^{32}$\,erg,][]{Emslie12} and convert it to the kinetic energy of moving plasma and accelerated particles via magnetic reconnection, Joule heating and other mechanisms. Thus, it is especially important to know the magnetic field configuration. Nowadays it is possible to obtain photospheric vector magnetograms from the SDO/HMI telescope \citep{Scherrer12} and reconstruct magnetic field in the solar atmosphere under certain assumptions. One of the key characteristics of the magnetic field structure is the Quasi-Separatrix Layer \citep[QSL,][]{Demoulin96,Demoulin97}. From the physical point of view, the QSL is a relatively thin surface where the magnetic field connectivity exhibits strong gradients~\citep{Aulanier06}, which can work as a channel of magnetic energy dissipation.
Nowadays, it is also possible to analyze flares with high-resolution using observations with large ground-based telescopes. One of the most breakthrough ground-based facilities is the New Solar Telescope \citep[NST,][]{Goode12} at Big Bear Solar Observatory. The 1.6\,m primary mirror and implemented adaptive optics provide diffraction-limited images and resolve features that are smaller than 0.1$^{\prime{}\prime{}}$. The studied flare was observed by the NST, and in this work we utilize the NST observations obtained in the H$\alpha$ line core.
\section{Methodology}
The IRIS observation covered temporarily the entire event for more than one hour from appearance of the first signs of flaring activity until the end of the decay phase. The instrument obtained spectra in several wavelength windows in each point of the region with $\approx$20\,s temporal and 0.33$^{\prime{}\prime{}}\times$2$^{\prime{}\prime{}}$ spatial resolution. To analyze the large amount of spectroscopic data we implemented the following techniques of the line profile analysis.
For each line formed in the chromosphere and chromosphere-corona transition region (i.e. Mg\,II\,k\&h\,2796\,{\AA}\,\&\,2803\,{\AA}, C\,II\,1334\,{\AA}\,\&\,1335\,{\AA}, Si\,IV\,1403\,{\AA}) the center-of-gravity approach used in our previous paper~\citep{Sadykov15} was implemented. For each line profile the following characteristics are calculated: 1)~the line peak intensity and 2)~the Doppler shift defined as a difference between the center of gravity of the line and the reference wavelength for this line $ \left<{\lambda}\right> - \lambda{}_{ref}= {\int{{\lambda}Id{\lambda}}}/{\int{Id{\lambda}}} - \lambda{}_{ref} $. Obviously, the implemented technique cannot be applied for blended lines. An example of such kind of line is, in fact, the IRIS Fe\,XXI\,1354.1\,{\AA} line which is formed in $1.1\times{}10^{7}$\,K hot plasma, and is very important for our study. The blends of this line are discussed by \citet[][Figure~2]{Tian14} and \citet[][Appendix A]{Young15}. We decided to take into account only the strongest blend, the C\,I\,1354.3\,{\AA} line. Our previous study~\citep{Sadykov15} did not reveal significant Doppler shifts of this line during the flare. Thus, for the Fe\,XXI line we performed a double-Gaussian fitting with a fixed peak wavelength of the second Gaussian profile corresponding to the $\lambda{}_{ref}=1354.34$\,{\AA}~--- the reference wavelength of the C\,I line. Using these procedures we determined the temporal and spatial behavior of the Doppler shift of the chromospheric, lower transition region and coronal lines that all are essential for studying the chromospheric evaporation.
As mentioned before, it is especially important to study the delay of the evaporated hot plasma flow observed as blueshift of the hot coronal lines relative to the chromospheric response (observed as redshift or blueshift of the cooler chromospheric or transition region lines). The dynamics of the hot evaporated plasma is studied using observations of the Fe\,XXI\,1354.1\,{\AA} line. We also use the C\,II\,1334.5\,{\AA} line as a representative of the colder chromospheric layer response to the flare heating. The C\,II line is formed at $T=10-20\times{}10^{3}K$. It is not overexposed in this flare unlike the Si\,IV line, and its shape is simpler than that of the Mg\,II lines.
The IRIS raster scans provide an opportunity to study the spatial configuration of the delays across the flare region. For this analysis the following procedure was performed. First, the Doppler shift of the C\,II\,1334.5\,{\AA} line was estimated at every point for each time moment of the IRIS scans in the region, and the same was done for the Fe\,XXI\,1354.1\,{\AA} line. After this, the temporal evolutions of the redshift and blueshift in each point were plotted and smoothed with a 50\,s running window for better estimation of their peak times. The peak times of the redshift and blueshift maxima were determined visually from the plotted curves. In places where the redshifts or blueshifts did not show any peak or even were equal to zero we set the delay to zero. Also, the delay was determined only in the flare ``bright points'', where the averaged over time magnitude of the C\,II\,1334.5 line was greater than the one eighth of the mean magnitude of this line across the flare region.
To reconstruct the magnetic field for the studied event, we followed the approach of~\citet{Wheatland00} implemented in the \textit{NLFFF} package of the Solar Software (SSW) for Interactive Data Language (IDL). The algorithm finds the solution for the Nonlinear Force-Free Field (NLFFF) approximation assuming that all electric currents flow along the field lines. For the boundary conditions, the 12-minute full-Sun vector magnetograms obtained by the HMI/SDO instrument~\citep{Scherrer12} were used. We reconstructed the magnetic field for eight time moments covering the flare period from 20:22:25\,UT to 21:46:25\,UT with 12\,min cadence. For the magnetic force line tracing, a tri-linear interpolation technique implemented in the SSW NLFFF package was used. To estimate topological peculiarities of the magnetic field in the flare region we applied a method of quasi-separatrix layer (QSL) calculation \citep{Demoulin97}. The QSLs mark regions with sharp variations of magnetic field connectivity. To make a quantitative estimate of the connectivity changes at a point $P(x,y,z)$ we use parameter called Squashing factor $N(x,y,z)$ calculated as:
\vskip-.6cm
\begin{eqnarray}
N(x,y,z) = \sqrt{\sum_{i=1}^{2} \left(\frac{\partial X_{i}}{\partial x} \right)^2+\left(\frac{\partial X_{i}}{\partial y}\right)^2+\left(\frac{\partial X_{i}}{\partial z}\right)^2},
\label{eq_qsl}
\end{eqnarray}
where $X_{1}$ and $X_{2}$ are components of the vector connecting the starting point of the magnetic field line crossing $P(x,y,z)$ with its end at the photospheric level. The coordinate derivatives $\partial_x$, $\partial_y$ and $\partial_z$ characterize variations of magnetic connectivity from point to point.
In addition, we analyzed the flare X-ray data from the Reuven Ramaty High-Energy Solar Spectroscopic Imager \citep[RHESSI,][]{Lin02} and compared the 12-25\,keV X-ray sources reconstructed by using the CLEAN algorithm with the magnetic field topology. The observed flux above 25\,keV was very weak and insufficient for the source reconstruction.
\section{Results}
\subsection{Integrated Behavior of Redshifts}
The integrated over the flare region intensities and Doppler shifts of the C\,II\,1334.5\,{\AA} and Fe\,XXI\,1354.1\,{\AA} lines are displayed in Figure~\ref{figure1}. The Doppler shifts and line intensities are estimated by techniques discussed in Sec.~2, and plotted with different colors (see caption of Fig.~\ref{figure1} for the color code).
The upper panel of Fig.~\ref{figure1} represents the Fe\,XXI\,1354.1\,{\AA} line integrated activity. A delay of the Fe\,XXI line intensity relative to its Doppler shift is very obvious, and, probably, occurs because of filling the magnetic loops by the hot evaporated plasma. The lower panel of Figure~\ref{figure1} displays the mean intensity and Doppler shift of the C\,II line. One can notice an increase of the C\,II redshift during the flare, and its correlations with the X-ray 12-25\,keV light curve from RHESSI. Previously~\citep[see][for details]{Sadykov15}, it was mentioned that the slowly varying redshifts mainly represent some background activity in the region. Fig.~\ref{figure1} shows that we observe a superposition of the relatively steady downflows and the fast varying downflows due to the flare energy release.
Two dotted vertical lines in Figure~\ref{figure1} correspond to the first peaks of the C\,II line redshift and Fe\,XXI line blueshift. One can see that the peak of the Fe\,XXI blueshift is delayed with respect to the C\,II line redshift for about one minute. Such delays pose a significant problem for the understanding of the flare dynamics. According to many 1D simulations of the chromospheric evaporation in the framework of the ``thick-target'' model, in which the chromosphere is heated by a beam of accelerated electrons~\citep{Fisher85a,Fisher85b,Fisher85c,Livshits83,Kosovichev86,Liu09,Rubiodacosta15a,Rubiodacosta15b,Reep15,Reep16}, the redshifts and blueshifts should be observed almost simultaneously at the start of the evaporation process.
In the previous paper \citep{Sadykov15} we mentioned that the evaporation process in this flare can be characterized as of the ``gentle'' type because of the subsonic velocities of the evaporated plasma according to~\citet{Antiochos78}. However, the integrated redshift of the C\,II line (see Fig.\,\ref{figure1} for details) obviously increases during the flare activity, which may be a sign of the explosive evaporation according to~\citet{Fisher85a}. Fig.~\ref{figure1} also reveals significant background steady plasma downflows obvious before and after the flare. Possibly, the evaporation in this region is very complex and has a fine structure, and cannot be classified as pure explosive or gentle one, according to the models.
\subsection{Spatial Structure of Chromospheric Evaporation}
The distribution of the Fe\,XXI\,1354.1{\AA} blueshift delay relative to the C\,II\,1334.5\,{\AA} redshift across the flare region is demonstrated in Fig.~\ref{figure2}. The procedure which we have performed to measure the delays is described in Sec.~2. The result is presented in the form of the contour lines corresponding to the delays of 30\,s, 60\,s, 120\,s and 240\,s. The IRIS 1330\,{\AA} SJ image is shown in the bottom for better representation of the chromospheric activity.
As one can see, the delays distributed across the flare region can be longer than two minutes that is longer than the previously reported 1-minute delays \citep{Graham15,Battaglia15,Young15}. The delays are distributed along the flare ribbon visible in the background IRIS\,1330\,{\AA} SJ image, and are not uniform especially in the region in the top-left corner of the white box. Flare ribbons are thought to be closely connected to the magnetic field configuration in the region. In the standard flare model it is assumed that because of the deposit of energy and accelerated particles along the flare loops, the plasma emission becomes stronger near the loop footpoints that becomes visible as the flare ribbons. Thus, we have decided to study the magnetic field properties in the region in order to better understand their relationship to the observed delay distribution.
\subsection{Flare Process and Field Topology}
For the magnetic field reconstruction, we use the NLFFF method of~\cite{Wheatland00} and vector magnetograms from HMI/SDO as the boundary conditions. Figure~\ref{figure3} represents the reconstructed magnetic field structure. In panel (a) this structure resembles the flux-rope which was observed in the NST images and reported by \citet{Sadykov14}~and~\citet{Kumar15}. The bottom grey-scale image represents the radial magnetic field (white for the positive and black for the negative polarity regions). As one can see, the field lines of the flux rope are twisted, reflecting a nonpotential nature of the magnetic field in the studied region with the currents embedded. This configuration is located exactly at the polarity inversion line (PIL). The detailed structure and dynamics of this region, which is likely to be the primary energy source for the flare, is discussed in a separate paper by~\citet{Sharykin16}.
Panels (b) and (c) of Fig.~\ref{figure3} illustrate the reconstructed magnetic field structure and the flare ribbons observed in the IRIS SJ 1330\,{\AA} image. For better understanding the structure, only the magnetic field lines reaching a certain range of heights (2$^{\prime{}\prime{}}$-6$^{\prime{}\prime{}}$, or 1.5-4.5\,Mm) are presented. The higher magnetic field lines have their footpoints far away from the flare ribbons, and thus do not participate in the energy transfer during the flare. The bottom panel is the IRIS 1330\,{\AA} slit-jaw image for 21:04:43\,UT. The field lines corresponding to the flux rope mentioned above are shown in green in this figure. One can see that almost all the lines starting from the flare ribbons have their other footpoint near the flux rope region at the PIL.
One of the possibilities to understand changes of the magnetic field topology and its connection with the observed delays is to reconstruct the so-called Quasi-Separatrix Layer~\citep[QSL,][]{Demoulin96,Demoulin97}. We have already described the computational procedure in Sec.~2. It was found that the QSL evolves with height very smoothly. Thus, we decided to utilize the QSL at height of $\approx{}1000$\,km above the photosphere, and calculated the squashing factor for the comparison.
The QSL structure presented in Figure~\ref{figure4} is mostly stable before (from 20:22:25\,UT to 20:58:25\,UT) and after (from 21:22:25\,UT to 21:46:25\,UT) the flare. However, during the flare impulsive phase the QSL undergoes significant changes in the region marked by the red dashed ellipse. The magnetic field neutral line also undergoes significant changes restricted to the marked region. Because of the 12\,min integration time of the SDO/HMI vector magnetogram data, we cannot determine when exactly during the period from 21:04:25\,UT to 21:16:25\,UT the QSL evolved.
We compare the QSL chromospheric structure with the flare ribbons visible in the IRIS\,1330\,{\AA} SJ images and the NST H$\alpha$ line core images. The result is presented in Figure~\ref{figure5}. The observing times are shown for each panel. One can notice a correspondence between the flare ribbons and the QSL cross-section. Also, the evolution of both the QSL and the flare ribbons (for both NST and IRIS observations) demonstrate similar patterns, confirming the idea of the flare energy transport along the QSL forming the flare ribbons \citep{Schmieder97,Masson09,Chandra11}.
To understand when exactly the evolution of the flare ribbons occurred, we studied the behavior of the H$\alpha$ flare ribbon in more details. We found that the motions of the flare ribbon occurred during the period from 21:12\,UT to 21:15\,UT, i.e after the impulsive phase of the flare. This time interval is within the uncertainty interval determined for the QSL change (from 21:04:25\,UT to 21:16:25\,UT). Also, only the north-eastern part of the ribbon changes, the other parts are mostly stable (see Fig.~\ref{figure5}\,g-h).
Figure~\ref{figure6} demonstrates the distribution of the delays across the flare region with the QSL chromospheric cross-section and the field lines plotted from the delay regions. The NST H$\alpha$ line core image is displayed in the background. Additionally, we have plotted the RHESSI 12-25\,keV contours in the same Figure. For the field lines displayed in this Figure we can say the same as for ones plotted in Fig.~\ref{figure3}: the height of most of the field lines does not exceed 4.5\,Mm (or 6$^{\prime{}\prime{}}$). So, the lines connecting the flux rope site and the delay regions do not extend high into the solar corona. The fact that the RHESSI 12-25\,keV sources plotted in Fig.~\ref{figure6} coincide in general with the footpoints of the large loop arcade supports the idea of transfer of the energetic particles along these loops. However, the acceleration site could not be determined from these data.
\section{Discussion and Conclusion}
In this paper we studied the chromospheric evaporation event during the M\,1.0 GOES class flare occurred on June 12, 2014 from 21:01\,UT till 21:19\,UT. The evaporated plasma flows were detected in the hot Fe\,XXI\,1354.1\,{\AA} line, and the response of the ``colder'' layers was studied with the help of the lower transition region C\,II\,1334.5\,{\AA} line. The main focus was on the distribution of the chromospheric evaporation delays (time between the C\,II Doppler shift maximum and the Fe\,XXI blueshift maximum). In addition, the magnetic field lines were reconstructed from the photospheric vector magnetograms, and the QSL was computed and compared with the flare ribbons. Let us remind the main observational findings mentioned in this study:
\begin{enumerate}
\itemsep0em
\item The averaged over the region C\,II redshift presented in Fig.~\ref{figure1} is correlated with the flare activity observed in the RHESSI 12-25\,keV; Fe\,XXI blueshift maximum is delayed relative to the C\,II redshift maximum for about 1\,min.
\item The detailed spatially-resolved study of the delays demonstrate their presence in many points along the flare ribbon, with the possible delays for longer than 2\,min (see Fig.~\ref{figure2}). The distribution of the delays across the initially-observed flare ribbon (in both IRIS 1330\,{\AA} and NST H$\alpha$ line core observations) is not uniform.
\item The reconstructed magnetic field lines from the delay regions mainly connect the flare ribbon with the flux rope structure. Their height very rare exceeds 4.5\,Mm, revealing their low-lying nature.
\item The RHESSI 12-25\,keV sources reasonably correspond to the footpoints of the main bundle of the reconstructed magnetic field lines.
\item The evolution of the QSL and flare ribbons detected in the IRIS\,1330\,{\AA} and NST H$\alpha$ line core images demonstrate the same patterns: mostly stable configuration with the motion in the North-East part of the region. This region is the only one along the initial QSL where the delays were not detected due to low Fe\,XXI signal.
\end{enumerate}
The spatio-temporal properties of the chromospheric evaporation reveal very strong delays of the blueshift of the hot evaporating plasma relative to the redshifts of the cold chromospheric plasma across the flare region. Despite the integrated blueshift (see Fig.~\ref{figure1}) demonstrates the delay for about 1\,min, the spatially-resolved delays are found to be even more than 2\,min in several zones along the flare ribbon. Thus, the integrated delay of the region represents itself the superposition of many spatially-distributed delays occurred in different points and caused by the excitation of the chromospheric evaporation process in different loops. In some sense, the observed situation corresponds to the ``multi-thread'' model \citep{Warren06} exactly proposing a sequence of independently heated threads occurred in different loops. As it is clearly seen from Fig.~7 in our previous work~\citep{Sadykov15}, the chromospheric excitation took place in different points across the region at different times. And thus it is not surprising that we have received the same kind of behavior for the delays. The ``multi-thread'' model was considered for the chromospheric evaporation studies in the work of~\citet{Rubiodacosta16}, where the authors used the RADYN code and superposition of evaporation events occurred in several loops at different times to adequately model the observed signals. Nevertheless, none of these models can explain the observed blueshift delay.
The reconstructed magnetic field geometry also corresponds to the multi-thread model but reveals an interesting complex configuration. As was observed from Fig.~\ref{figure3}, the magnetic configuration of the region represents a twisted small-scale loops constructing a magnetic flux rope located at the polarity inversion line, and the bundles of more large-scale magnetic field lines with one footpoint located near the flux rope and the other footpoint located in the flare ribbons, i.e. connecting the flare ribbons and evaporating regions with the flux rope. This magnetic flux rope was studied in more details in the paper of~\citet{Sharykin16}. Their study revealed strong current dissipation and large gradients of the magnetic field associated with the flux rope, near the polarity inversion line for this region. One of the conclusions was that the dissipation processes in this region can be the primary energy source for this flare. It is obvious from the reconstructed magnetic field configuration that accelerated particles and heat flux can spread from the flux rope region to the observed flare ribbons along the field lines. Injections of the particles and heat flux into different loops produce the chromospheric evaporation in different spatial zones as we find in the observations. Thus, the flux rope region at the polarity inversion line may play a role of the ``energy source'' for the event. The footpoints of the large-scale magnetic field lines (coming from right top to left bottom in Fig.~\ref{figure6}) adequately coincide with the RHESSI 12-25\,keV sources plotted in the same Figure.
It was found that almost all the magnetic field lines connecting the blueshift delay regions with the flux rope are low-lying (see Fig.~\ref{figure6}). Their height rarely exceeds 4.5\,Mm, thus, these loops mainly do not expand high into the corona. This means that all the delays were observed in the low-lying loops. The delays are non-uniformly distributed along the flare ribbon (upper panel of Fig.~\ref{figure6}), but without any obvious patterns. One of the possible explanations of the delays based on the non-equilibrium Fe ionization \citep{Battaglia15,Bradshaw09,Graham15} was discussed in the introduction. The results presented in Fig.6-8 of~\citet{Bradshaw09} show that the Fe\,XIX ion population reaches equilibrium for the considered durations of the heating phase (up to 60\,s), but the Fe\,XXIV ions are out of equilibrium with low population. There are no results presented for the Fe\,XXI, and it is hard to understand how does the Fe\,XXI ion population behaves during the heating phase. However, the highly ionized Fe fractions (including Fe\,XIX and Fe\,XXIV) are in equilibrium conditions during the thermal conductive cooling phase. The non-equilibrium ionization explanation of delays becomes suitable only if very long continuous heating (for more than 2\,min) is presented. The first assumption may contradicts the impulsive nature of solar flares. The strong growth of the C\,II intensity light curve in Fig.~\ref{figure1} and results presented in Fig.~7 in \citet{Sadykov15} support the idea that the chromosphere heating was impulsive. Thus, the non-equilibrium ionization mechanism seems to be partly, but not fully responsible for the observed delays. The fact that the evaporation takes place in the low-lying loop geometry is the only one we can lean on.
The only region where the delays are not present or not possible to calculate is the upper left corner of Figure~\ref{figure6}. Figures~\ref{figure4}~and~\ref{figure5} clearly show that this region is the only one where the flare ribbon motion and the QSL chromospheric cross-section evolution was observed. We looked at the spectra of this region in detail and revealed the following: despite the C\,II redshift was significant, the weak signal in the Fe\,XXI line led to the impossibility to calculate the delay. As shown in panels a, d, e and h of Fig.~\ref{figure5}, the computed QSL cross-section fits the observed flare ribbons quite accurately before and after the impulsive phase of the flare. Thus, one can assume that the QSL evolved at the same time as flare ribbons did~--- i.e. from 21:12\,UT till 21:16\,UT. The first 12-25\,keV X-ray pulse occurred at $\sim$21:06\,UT (the first RHESSI 12-25\,keV peak corresponds to the first peak in the C\,II integrated light curve in Fig.~\ref{figure1}). However, at the time when the flare ribbon motion was observed, the RHESSI 12-25\,keV curve, as well as the C\,II integrated light curves, experience the decay phase. The motion of the flare ribbons might correspond to the process called slipping magnetic reconnection~\citep{Janvier13,Aulanier12}. This model is quite new but already found observational evidences~\citep{Li15,Janvier14}. However, it seems that the studied flare was not driven by the slipping magnetic reconnection mechanism. Despite the ribbon motion was observed, it occurred definitely after the impulsive phase of the flare. Even if the slipping mechanism is responsible for this motion, it happened after the impulsive phase and could not support the ideas that the flare energy is released in the QSL.
Of course, the found relationship between the chromospheric evaporation delays and the magnetic field configuration is based only on one studied event. Further statistical study is needed to confirm the proposed dependences.
\acknowledgments
The authors acknowledge the BBSO, IRIS and SDO mission teams for their contribution and support. The work was partially supported by NASA grants NNX14AB68G, NNX14AB70G, and NNX11AO736; NSF grant AGS-1250818; RFBR grants 15-32-21078 and 16-32-00462; and an NJIT grant.
|
\section{Introduction}
Compressed sensing techniques have proven to be of great interests
for detecting, estimating and denoising sparse signals lying on discrete
spaces. On the practical side, the applications of sparse modeling
are many: single molecule imaging via fluorescence, blind source separation
in speech processing, precise separation of multiple celestial bodies
in astronomy, or super-resolution radaring, are among those. However,
the discrete gridding required by the compressed sensing framework
weaken the recovery performances, and more precisely the \emph{resolution}:
the required minimal separation between two components of the sparse
signal to be efficiently distinguished by an observation process.
In the recent years, a particular enthusiasm has been placed on solving
sparse linear inverse problems over continuous spaces. This paradigm
aims to recover the finite subset of components generating a signal,
and lying in a continuous space, by discrete observations of this
signal, distorted by a kernel function. Considering such approach
raises new concerns, in particular, those problems are commonly infinitely
ill-posed. This primordial issue has been addressed for the spikes
model \cite{Candes2014a,Candes2012,Tang2013} via the mean of total-variation
(or atomic) convex relaxation techniques, reducing the dimensionality
on a dramatic manner. Later on, similar results have been derived
for sparse signals lying on some known subspaces in \cite{Yang2016},
using particular kernel functions \cite{Schiebinger2015}, or via
incoherent multiple measurements in \cite{Yang2014}. Generic performance
in noise have been provided \cite{Bhaskar2013} and specific gradient
search algorithms proposed in \cite{Boyd2015} to efficiently solve
this category of problems.
For the spectral case, a complex time signal $x$ is said to follow
the $s$-spikes model if and only if it reads,
\begin{equation}
x(t)=\sum_{l=1}^{s}\alpha_{l}e^{i2\pi\xi_{l}t},\quad\forall t\in\mathbb{R},\label{eq:SpikeModel}
\end{equation}
where $\xi=\left[\xi_{1},\dots,\xi_{s}\right]^{\mathrm{{\scriptscriptstyle T}}}\in\mathbb{R}^{s}$
is the vector containing the $s$ spectral components generating the
signal $x$, and $\alpha\in\mathbb{\mathbb{C}}^{s}$ the vector of
their associated complex amplitudes. The frequency estimation problem
is naturally defined as building a consistent estimator $\left(\hat{s},\hat{\xi},\hat{\alpha}\right)$
of the parameters $\left(s,\xi,\alpha\right)$, that are supposed
to be unknown, by $N\in\mathbb{N}$ discrete observations $y\in\mathbb{C}^{N}$
of the time signal $x$.
This problem is obviously ill-posed, and since no assumption is a
priori made on the number of frequencies $\hat{s}$ to estimate, there
are infinitely many triplet $\left(\hat{s},\hat{\xi},\hat{\alpha}\right)$
that are coherent with the observation vector $y$. In particular,
the discrete Fourier transform of $y$ forms a consistent spectral
representation of $x$ by $N$ spectral spikes. Among all those estimators,
the one considered to be optimal, in this context, will be the one
returning the sparsest spectral distribution, i.e., the one achieving
the smallest $\hat{s}_{0}$. The optimal spectral distribution $\hat{x}_{0}$
can be written as the output of an optimization program taking the
form,\vspace{-0.5cm}
\begin{align}
\hat{x}_{0} & =\arg\min_{\hat{x}\in D_{1}}\left\Vert \hat{x}\right\Vert _{0}\label{eq:RegularL0Problem}\\
\text{subject to} & \phantom{\;=\;}y=\mathcal{F}\left(\hat{x}\right),\nonumber
\end{align}
where $\hat{x}$ is the spectral distribution of $x$, $\left\Vert \cdot\right\Vert _{0}$
represents the limit of the $p$-norm towards $0$, counting the cardinality
of the support. $D_{1}$ denotes the space of absolutely integrable
spectral distributions, and $\mathcal{F}$ denotes a linear operator
fully determined by the sampling process and linking the spectral
domain to the measurements. Since this program is an NP-hard combinatorial
problem, a common approach consists in relaxing the cardinality cost
function into a minimization of the total-variation norm over the
spectral domain, leading to the convex program,
\begin{align}
\hat{x}_{{\rm TV}} & =\arg\min_{\hat{x}\in D_{1}}\left\Vert \hat{x}\right\Vert _{\mathrm{TV}}\label{eq:RegularL1Problem}\\
\text{subject to} & \phantom{\;=\;}y=\mathcal{F}\left(\hat{x}\right).\nonumber
\end{align}
The previous works in the literature were mostly studying the regularly
spaced observation model, $y_{k}=x\left(\frac{k}{f}\right)$ for $k\in\left\llbracket 0,N-1\right\rrbracket $.
Under such observations, it has been shown that the relaxation proposed
in (\ref{eq:RegularL1Problem}) is exact in the sense that, under
the minimal separation criterion over the normalized frequencies $\Delta\nu=\min\left\{ \mathrm{frac}\left(\nu_{i}-\nu_{j}\right),i\neq j\right\} \geq\frac{4}{N-1}$
of the sparse spectral distribution $\hat{x}$ to recover, the output
of Programs (\ref{eq:RegularL0Problem}) and (\ref{eq:RegularL1Problem})
are identical. Additionally, Program (\ref{eq:RegularL1Problem})
can be reformulated into a semidefinite program (SDP) of dimension
$N+1$, where coefficients of the optimum define a trigonometric polynomial
$Q_{*}$ locating the frequencies of the original signal over the
unit circle. $Q_{*}$ takes modulus $\left|Q_{*}\left(e^{i2\pi\nu}\right)\right|=1$
whenever $2\pi f\nu=\xi_{l}$ and satisfies $\left|Q_{*}\left(e^{i2\pi\nu}\right)\right|<1$
otherwise. It has been shown in \cite{Tang2013} that this optimality
still holds with high probability when extracting at random a small
number of observations from $y$ and discarding the rest of it. Other
studies revealed that the spectral separation condition can be reduced
\cite{Fernandez-granda2015}, and that this model partially extends
to multidimensional signals \cite{Chi2015,Yang2015}.
In this work, our contribution is focused on extending the previous
results on sparse frequency estimation to the framework of multi-rate
sampling systems (MRSS): the observations $y$ are gathered as the
output of $m$ different uniform samplers, working at different sampling
rates, and potentially desynchronized (the samplers process the time
signal $x$ with some arbitrary delays). According to our knowledge,
this approach is the first to extend super-resolution to such a generic
measurement process. This model is of crucial importance, for instance,
when seeking to achieve joint estimation of sparse signals in distributed
sensor networks. Each node, with limited processing capabilities,
samples at its own rate, a delayed version of a complex signal. Collected
data are then sent and merged at a higher level processing unit, performing
a global estimation of the spectral distribution on a joint manner.
MRSS estimation is also a meaningful step towards a super-resolution
theory from non-uniform sampling.
In Section \ref{sec:Super-Resolving-MRS}, we show in Proposition
\ref{prop:PolynomialAlignment} that, under certain conditions on
the rates and the delays between the samplers, the ``total-variation''
relaxation of the sparse recovery problem can take a polynomial form
similar to the one described in the original paper \cite{Candes2014a}.
We argue that the model benefits from the same performance guarantees,
and from the optimality. We point out that this direct relaxation
has an arbitrary high complexity, making it unsolvable by standard
convex solvers. In Section \ref{sec:Exact-Dimensionality-Reduction},
a novel exact dimensionality reduction of the semidefinite form (\ref{eq:FullSDP})
is presented in Theorem \ref{thm:BRLSparsePol} by extending the theory
of Gram representation of trigonometric polynomials presented in \cite{Dumitrescu2010}
into the sparse case. We conclude that the dual of the main problem
(\ref{eq:PrimalProblem}) can be reformulated in the compact SDP (\ref{eq:ReducedSDP})
whose dimension is equal to the number of observations.
\section{Super-Resolving Multi-Rate Sampling Systems\label{sec:Super-Resolving-MRS}}
\subsection{Observation model}
An MRSS process on a continuous signal $x$ is parametrized by a set
$\mathbb{A}$ of $m$ distinct grids (or samplers) $\mathcal{A}_{j}$,
$j\in\left\llbracket 1,m\right\rrbracket $. Each grid is identified
with a triplet $\mathcal{A}_{j}=\left(f_{j},\gamma_{j},n_{j}\right)$,
where $f_{j}\in\mathbb{R}^{+}$ is its sampling frequency, $\gamma_{j}\in\mathbb{R}$
its delay (in sample unit), and $n_{j}\in\mathbb{N}$ the number of
measurements acquired by the grid. We assume those intrinsic characteristics
to be known. The output $y_{j}$ of the grid $\mathcal{A}_{j}$ sampling
a signal $x$ following the sparse model described in (\ref{eq:SpikeModel})
reads,
\begin{equation}
y_{j}\left[k\right]=\sum_{l=1}^{s}\alpha_{l}e^{i2\pi\frac{\xi_{l}}{f_{j}}\left(k-\gamma_{j}\right)},\quad k\in\left\llbracket 0,n_{j}-1\right\rrbracket .\label{eq:ObservationModel}
\end{equation}
As explained above, the frequency estimation problem is formulated
as finding the sparsest spectral density jointly matching the observation
vectors $y_{j}$, for all $j\in\left\llbracket 1,m\right\rrbracket $.
This problem takes the same form than the combinatorial minimization
program (\ref{eq:RegularL0Problem}), by specifying the equality constraint
$y=\mathcal{F}\left(\hat{x}\right)$ as follows,
\[
y_{j}=\mathcal{F}_{j}\left(\hat{x}\right),\quad\forall j\in\left\llbracket 1,m\right\rrbracket ,
\]
where $\mathcal{F}_{j}$ is a linear operator denoting the effect
of the spectral density on the samples uniformly acquired by the grid
$\mathcal{A}_{j}$, and is characterized by,\vspace{-0.45cm}
\begin{align*}
\mathcal{F}_{j}:D_{1} & \to\mathbb{C}^{n_{j}}\\
\hat{x} & \mapsto y_{j}:y_{j}\left[k\right]=\\
& \phantom{====}\int_{\mathbb{R}}\hat{x}\left(\xi\right)e^{i2\pi\frac{\xi}{f_{j}}\left(k-\gamma_{j}\right)}d\xi,\;\forall k\in\left\llbracket 0,n_{j}-1\right\rrbracket .
\end{align*}
\subsection{Convex relaxation}
We recall that Program (\ref{eq:RegularL0Problem}) is NP-hard in
the general case, due to its combinatorial aspects. The relaxation
described in (\ref{eq:RegularL1Problem}) is introduced and takes
the form,
\begin{align}
\hat{x}_{{\rm TV}} & =\arg\min_{\hat{x}\in D_{1}}\left\Vert \hat{x}\right\Vert _{\mathrm{TV}}\label{eq:PrimalProblem}\\
\text{subject to} & \phantom{\;=\;}y_{j}=\mathcal{F}_{j}\left(\hat{x}\right),\quad\forall j\in\left\llbracket 1,m\right\rrbracket .\nonumber
\end{align}
Such transform has the advantage to turn the original infinite-combinatorial
problem into a convex problem. However, for practical computation,
convexity often is not enough in order to guarantee a successful resolution
of a program. Indeed, the cost function of (\ref{eq:PrimalProblem})
takes values in $D_{1}$, a space having an uncountable dimension.
Convex optimization theory ensures that this category of programs
can be reformulated into semi-infinite programs \cite{Shapiro}: a
convex optimization program of a finite-dimensional cost function
over an infinite-dimensional set of constraints, using the classic
Lagrangian duality. In our settings, the Lagrange dual problem is,
\begin{align}
c_{*} & =\arg\max\sum_{j=1}^{m}\Re\left(\left\langle y_{j},c_{j}\right\rangle \right)\label{eq:DualProblem}\\
\text{subject to} & \phantom{=\;\;}\left\Vert \sum_{j=1}^{m}\mathcal{F}_{j}^{*}\left(c_{j}\right)\right\Vert _{\infty}\leq1,\nonumber
\end{align}
where $c=\left[c_{1}^{\mathrm{{\scriptscriptstyle T}}},\dots,c_{m}^{\mathrm{{\scriptscriptstyle T}}}\right]^{\mathrm{{\scriptscriptstyle T}}}$
is the dual variable, and $\mathcal{F}_{j}^{*}$ denotes the adjoint
of the operator $\mathcal{F}_{j}$ for the Euclidean inner products.
Since the original problem is only equally constrained, Slatter's
condition is automatically met, and strong duality holds. This implies
that the optima of the primal problem (\ref{eq:PrimalProblem}) and
its dual (\ref{eq:DualProblem}) are equal. Moreover this equality
appends if and only if $\hat{x}_{\mathrm{TV}}$ is primal optimal,
and $c_{*}$ dual optimal \cite{Boyd2004}.
Letting by $\omega_{j}=\frac{2\pi\xi}{f_{j}}$ the normalized pulsation
of array $\mathcal{A}_{j}$, the expression of the adjoint operator
$\mathcal{F}_{j}^{*}$ allows to reformulate the dual constraint into
a boundedness constraint of a sum of exponential polynomials of the
form,
\begin{align*}
\mathcal{F}_{j}^{*}\left(c_{j}\right) & =\sum_{k=0}^{n_{j}-1}c_{j}\left[k\right]e^{-i\left(k-\gamma_{j}\right)\omega_{j}}\\
& =e^{i\gamma_{j}\omega_{j}}P_{j}\left(e^{-i\omega_{j}}\right),
\end{align*}
where $P_{j}\in\mathbb{C}^{n_{j}-1}\left[X\right]$ is the dual complex
polynomial related to array $\mathcal{A}_{j}$, and is defined by
$P_{j}\left(z\right)=\sum_{k=0}^{n_{j}-1}c_{j}\left[k\right]z^{k}$.
\subsection{Common grid expansion}
It has been shown in \cite{Candes2014a} that the sparse frequency
recovery problem can take the form of a simple SDP when dealing with
regularly spaced samples. However, those results cannot be transposed
in the MRSS framework, since the dual constrained operator $\sum_{j=1}^{m}\mathcal{F}_{j}^{*}\left(c_{j}\right)$
does not take a polynomial form. As an assumption to bridge this concern,
the sampling process $\mathbb{A}$ is supposed to admit a common supporting
grid, ensuring that the observation samples can be uniformly aligned
at a higher virtual rate. The notion of common supporting grid is
defined bellow. Necessary and sufficient conditions in terms of the
parameters of $\mathbb{A}$ for its existence to hold are stated in
Proposition \ref{prop:ExistenceOfCommonSupportingGrid}.
\begin{defn}
\label{def:CommonSupportingGrid}A grid $\mathcal{A}_{\#}=\left(f_{\#},\gamma_{\#},n_{\#}\right)$
is said to be a \emph{common supporting grid} for a set of sampling
grids $\mathbb{A}=\left\{ \mathcal{A}_{j}\right\} _{j\in\left\llbracket 1,m\right\rrbracket }$
if and only if the set of samples acquired by the MRSS induced by
$\mathbb{A}$ is a subset of the one acquired by $\mathcal{A}_{\#}$.
In formal terms, the definition is equivalent to $\left\{ \frac{1}{f_{j}}\left(k_{j}-\gamma_{j}\right),\,j\in\left\llbracket 1,m\right\rrbracket ,\,k_{j}\in\left\llbracket 0,n_{j}-1\right\rrbracket \right\} \subseteq\left\{ \frac{1}{f_{\#}}\left(k_{\#}-\gamma_{\#}\right),\,k_{\#}\in\left\llbracket 0,n_{\#}-1\right\rrbracket \right\} $.
The set of common supporting grids of $\mathbb{A}$ is denoted by
$\mathcal{C}\left(\mathbb{A}\right)$. Moreover, a common supporting
grid $\mathcal{A}_{*}=\left(f_{*},\gamma_{*},n_{*}\right)$ for $\mathbb{A}$
is said to be \emph{minimal} if and only it satisfies the minimality
condition, $\forall\mathcal{A}_{\#}\in\mathcal{C}\left(\mathbb{A}\right),\quad n_{*}\leq n_{\#}.$
\end{defn}
\begin{prop}
\label{prop:ExistenceOfCommonSupportingGrid}Given a set of $m$ observation
grids $\mathbb{A}=\left\{ \mathcal{A}_{j}=\left(f_{j},\gamma_{j},n_{j}\right)\right\} _{j\in\left\llbracket 1,m\right\rrbracket }$,
a common supporting grid $\mathcal{A}_{\#}$ exists if and only if
there exist $f_{\#}\in\mathbb{R}^{+}$, $\gamma_{\#}\in\mathbb{R}$,
a set of $m$ positive integers $\left\{ l_{j}\right\} \in\mathbb{N}^{m}$,
and a set of $m$ integers $\left\{ a_{j}\right\} \in\mathbb{\mathbb{Z}}^{m}$
satisfying $f_{\#}=l_{j}f_{j}$ and $\gamma_{\#}=l_{j}\gamma_{j}-a_{j}$
for all $j\in\left\llbracket 1,m\right\rrbracket $. Moreover a common
grid $\mathcal{A}_{*}=\left(f_{*},\gamma_{*},n_{*}\right)$ is minimal,
if and only if, $\gcd\left(\left\{ a_{j}\right\} _{j\in\left\llbracket 1,m\right\rrbracket }\cup\left\{ l_{j}\right\} _{j\in\left\llbracket 1,m\right\rrbracket }\right)=1$,
$\gamma_{*}=\max_{j\in\left\llbracket 1,m\right\rrbracket }\left\{ l_{j}\gamma_{j}\right\} $
and $n_{*}=\max_{j\in\left\llbracket 1,m\right\rrbracket }\left\{ l_{j}\left(n_{j}-1\right)-a_{j}\right\} .$
\vspace{-0.1cm}
\end{prop}
The proof of the above proposition is presented in \cite{FerreiraDaCosta}.
In the following, we assume that $\mathbb{A}$ satisfies the conditions
of Proposition \ref{prop:ExistenceOfCommonSupportingGrid}, and we
denote its minimal common supporting grid by $\mathcal{A}_{*}=\left(f_{*},\gamma_{*},n_{*}\right)$.
The next result shows that, under those circumstances, the dual inequality
constraint in (\ref{eq:DualProblem}) takes a polynomial form.
\begin{prop}
Consider the multi-rate sampling system induced by $\mathbb{A}=\left\{ \mathcal{A}_{j}\right\} _{j\in\left\llbracket 1,m\right\rrbracket }$,
if $\mathcal{C}\left(\mathbb{A}\right)\neq\emptyset$ there exists
a complex polynomial $Q\in\mathbb{C}^{n_{*}-1}\left[X\right]$ such
that $\left\Vert \sum_{j=1}^{m}\mathcal{F}_{j}^{*}(c_{j})\right\Vert _{\infty}=\left\Vert Q(e^{i\omega_{*}})\right\Vert _{\infty}$.\label{prop:PolynomialAlignment}
\end{prop}
\begin{IEEEproof}
The proof of this proposition is direct,
\begin{align*}
\sum_{j=1}^{m}\mathcal{F}_{j}^{*}\left(c_{j}\right) & =\sum_{j=1}^{m}e^{i\gamma_{j}\omega_{j}}P_{j}\left(e^{-i\omega_{j}}\right)\\
& =e^{i\gamma_{*}\omega_{*}}\sum_{j=1}^{m}e^{ia_{j}\omega_{*}}P_{j}\left(e^{-il_{j}\omega_{*}}\right),
\end{align*}
by replacing $\omega_{j}$ by $l_{j}\omega_{*}$ and $l_{j}\gamma_{j}$
by $\gamma_{*}+a_{j}$ in the second equality, where $\left\{ l_{j}\right\} \in\mathbb{N}^{m}$
and $\left\{ a_{j}\right\} \in\mathbb{\mathbb{Z}}^{m}$ qualify the
minimal common supporting grid $\mathcal{A}_{*}$ of $\mathbb{A}$.
It comes that,
\begin{align*}
\sum_{j=1}^{m}\mathcal{F}_{j}^{*}(c_{j}) & =e^{i\gamma_{*}\omega_{*}}Q\left(e^{-i\omega_{*}}\right),
\end{align*}
where $Q(z)=\sum_{j=1}^{m}z^{-a_{j}}P_{j}\left(z^{bl_{j}}\right)$
is a well defined complex polynomial, since $a_{j}\leq0$ by assumption
on the minimality of $\mathcal{A}_{*}$. Taking the infinite norm
on both sides and noticing its invariance by $\omega_{*}\leftarrow-\omega_{*}$
lead to the desired result.
\end{IEEEproof}
Due to the upscaling effect created by the expansion of $\mathbb{A}$
on a common grid $\mathcal{A}_{*}$, the resulting dual polynomial
$Q$ has a degree $n_{*}-1$ that can be potentially much higher than
the initial degrees of the individual dual polynomials $\left\{ P_{j}\right\} _{j\in\left\llbracket 1,m\right\rrbracket }$.
This fact is illustrated by an example in the end of this section.
However, it is easy to notice that $Q$ is sparse, and that it can
be decomposed into a sum over $N_{*}\leq N=\sum_{j=1}^{m}n_{j}$ monomials.
Let us denote by $q\in\mathbb{C}^{n_{*}}$ the vector containing the
coefficients of $Q(z)=\sum_{k=0}^{n_{*}-1}q_{k}z^{k}$ and call by
$\mathcal{I}\subseteq\left\llbracket 0,n_{*}-1\right\rrbracket $,
the subset of cardinality $N_{*}$ containing the powers of the supporting
monomials. One can write the relation $q=C_{\mathcal{I}}c$, where
$c$ is the dual variable of Problem (\ref{eq:DualProblem}), for
an orthogonal selection matrix $C_{\mathcal{I}}\in\left[0,1\right]^{n_{*}\times N_{*}}$
for the subset $\mathcal{I}$. The matrix $C_{\mathcal{I}}$ can be
directly inferred from the settings of $\mathbb{A}$.
Proposition \ref{prop:PolynomialAlignment} ensures that the dual
constraint of the dual problem described in (\ref{eq:DualProblem})
is equivalent to restrict a complex polynomial to be bounded in modulus
by one around the unit circle $\mathbb{T}$. We recall a result presented
in \cite{Dumitrescu2010} (Corollary 4.25) emerging from the Gram
parametrization theory of complex polynomials which yields,\vspace{-0.4cm}
{\small{}
\begin{equation}
\left\Vert Q(e^{i\omega})\right\Vert _{\infty}\leq1\Leftrightarrow\exists H\,\text{Hermitian s.t. }\begin{cases}
\begin{bmatrix}H & q\\
q^{\mathrm{{\scriptscriptstyle H}}} & 1
\end{bmatrix}\succeq0\\
\mathcal{T}_{n}^{*}\left(H\right)=e_{1,}
\end{cases}\label{eq:EquivalenceBoundedPolHermitianMatrix}
\end{equation}
}for any $Q\in\mathbb{C}^{n-1}\left[X\right]$, where $\mathcal{T}_{n}^{*}$
is the adjoint to the canonical decomposition of Hermitian Toeplitz
matrices of dimension $n$ $\mathcal{T}_{n}$, and is given by $\mathcal{T}_{n}^{*}\left(H\right)[k]=\mathrm{tr}\left(\Theta_{k}H\right)$,
for $k\in\left\llbracket 0,n\right\rrbracket $, where $\Theta_{k}$
is the elementary Toeplitz matrix equals to $1$ on the $k^{th}$
lower diagonal and zero elsewhere, and where $e_{1}$ is the first
vector of the canonical basis of $\mathbb{C}^{n}$
The semi-algebraic duality (\ref{eq:EquivalenceBoundedPolHermitianMatrix}),
combined with Proposition \ref{prop:PolynomialAlignment}, allows
to rewrite the infinite dimensional constraint of Program (\ref{eq:DualProblem})
into a positivity condition of an Hermitian matrix of dimension $n_{*}+1$
given by,\vspace{-0.33cm}
\begin{align}
c_{*} & =\arg\max\Re\left(\left\langle y,c\right\rangle \right)\label{eq:FullSDP}\\
\text{subject to} & \phantom{\phantom{\;=\;}}\begin{bmatrix}H & C_{\mathcal{I}}c\\
\left(C_{\mathcal{I}}c\right)^{\mathrm{{\scriptscriptstyle H}}} & 1
\end{bmatrix}\succeq0\nonumber \\
\mathcal{} & \phantom{\;=\;}\mathcal{T}_{n_{*}}^{*}\left(H\right)=e_{1}.\nonumber
\end{align}
The above problem is nothing but a particular case of the convex relations
studied in \cite{Tang2013}. This ensures that the optimum $q_{*}=C_{\mathcal{I}}c_{*}$
induces a sparse complex polynomial $Q_{*}$ that exactly locates
the frequencies of $x$ by solving $\left|Q_{*}\left(e^{i\omega}\right)\right|=1$
around the unit circle $\omega\in\mathbb{T}$, as long a sufficient
minimal spectral separation discussed in \cite{FerreiraDaCosta} is
respected.
Although semidefinite programs are theoretically solvable and certifiable,
practical attempts to recover the frequencies of the time signal $x$
via Program (\ref{eq:FullSDP}) might fail or return inaccurate results
due to the high dimensionality of the constraints. This is the case
in our settings, the square block matrix in (\ref{eq:FullSDP}) has
a size of $n_{*}+1$, which can be considerably higher than the effective
dimension of the observations $N_{*}$, depending of the settings
of the MRSS defined by $\mathbb{A}$. As for illustration purposes,
suppose a delay-only MRSS, where $\mathbb{A}$ is constituted of $m$
grids given by $\mathcal{A}_{1}=\left(f,0,n\right)$, $\mathcal{A}_{j}=\left(f,-\frac{1}{b_{j}},n\right)$
for all $j\in\left\llbracket 2,m\right\rrbracket $, and where the
$\left\{ b_{j}\right\} _{j\in\left\llbracket 2,m\right\rrbracket }$are
jointly coprime. One has $\mathcal{A}_{*}=\left(\left(\prod b_{j}\right)f,0,\left(\prod b_{j}\right)n\right)$,
leading to a matrix constraint of asymptotic dimension $\Omega\left(b^{m}n\right)$
for some constant $b\in\mathbb{R}^{+}$, while the essential dimension
of the problem remains of order $\mathcal{O}\left(mn\right)$.
\section{Exact Dimensionality Reduction\label{sec:Exact-Dimensionality-Reduction}}
In this section, we show that the original dual problem described
in (\ref{eq:DualProblem}) is equivalent to a similar SDP of size
exactly equal to $N_{*}+1$, which is optimal in those settings. To
this end, we first need to recall some results about Gram parametrization
of trigonometric polynomials.
\subsection{Gram parametrization of trigonometric polynomials}
For every non-zero complex number $z\in\mathbb{C}^{*}$, its $n^{th}$
power vector is defined by $\psi_{n}(z)=\left[1,z,\dots,z^{n}\right]^{\mathrm{{\scriptscriptstyle T}}}$.
A complex trigonometric polynomial $R\in\bar{\mathbb{C}}^{n}\left[X\right]$
of order $\bar{n}=2n+1$ is a linear combination of complex monomials
with positive and negative exponents absolutely bounded by $n$. Such
polynomial $R$ reads,\vspace{-0.3cm}
\[
R\left(z\right)=\sum_{k=-n}^{n}r_{k}z^{-k},\quad\forall z\in\mathbb{C}^{*}.
\]
Each of such entities can be associated with a Gram set, as defined
in Definition \ref{def:GramSet}. Proposition \ref{prop:GramParametrizationTheorem}
states that this duality holds via a simple linear relation with complex
matrices.
\begin{defn}
A complex matrix $G\in\mathbb{C}^{\left(n+1\right)\times\left(n+1\right)}$
is a \emph{Gram matrix} associated with the trigonometric polynomial
$R$ if and only if,\vspace{-0.1cm}
\[
R\left(z\right)=\psi\left(z^{-1}\right)^{\mathrm{{\scriptscriptstyle T}}}G\psi\left(z\right),\quad\forall z\in\mathbb{C}^{*}.
\]
Such parametrization is, in general, not unique and we denote by $\mathcal{G}\left(R\right)$
the set of matrices satisfying the above relation. $\mathcal{G}\left(R\right)$
is called \emph{Gram set} of $R$.\label{def:GramSet}
\end{defn}
\begin{prop}
For any complex trigonometric polynomial $R$ of order $\bar{n}=2n+1$,
$G\in\mathcal{G}\left(R\right)$ if and only if the relation,
\[
\mathcal{T}_{\bar{n}}^{*}\left(G\right)=r
\]
holds, where $r\in\mathbb{C}^{\bar{n}}$ is the vector containing
the coefficients of $R$ indexed in $\left\llbracket -n,n\right\rrbracket $.\label{prop:GramParametrizationTheorem}
\end{prop}
\subsection{Compact representation of sparse polynomials}
Up to here, the concept of Gram sets adapts to every complex trigonometric
polynomial. If $R$ is of order $\bar{n}$, it defines a set $\mathcal{G}\left(R\right)$
of matrices in $\mathbb{C}^{n\times n}$. In our context, $R$ has
a sparse monomial support, and Gram representations with compact low-dimensional
structures, reflecting this sparsity, are of crucial interest for
the dimensionality reduction approach. Definition \ref{def:CompactGramRepresentation}
introduces the notion of compact representations.
\begin{defn}
A complex trigonometric polynomial $R$ of order $\bar{n}$ is said
to admit a \emph{compact Gram representation} on a matrix $M\in\mathbb{C}^{n\times m}$,
$m\leq n$, if and only if there exists a matrix $G_{M}\in\mathbb{C}^{m\times m}$
such that the relation,
\begin{align*}
R\left(z\right) & =\psi\left(z^{-1}\right)^{\mathrm{{\scriptscriptstyle T}}}MG_{M}M^{\mathrm{{\scriptscriptstyle T}}}\psi\left(z\right)\\
& =\phi_{M}\left(z^{-1}\right)^{T}G_{M}\phi_{M}\left(z\right),\qquad\forall z\in\mathbb{C}^{*}
\end{align*}
holds, where $\phi_{M}(z)=M^{\mathrm{{\scriptscriptstyle T}}}\psi(z)$. We denote by $\mathcal{G}_{M}\left(R\right)$
the subset of complex matrices satisfying this property.\label{def:CompactGramRepresentation}
\end{defn}
Although it can be difficult to characterize the set of polynomials
admitting a compact representation on a given matrix $M\in\mathbb{C}^{n\times m}$,
a simple criterion exists for the special case of selection matrices
$C_{\mathcal{I}}$. This criterion is recalled from \cite{Dumitrescu2010}
in Proposition \ref{prop:ExistenceOfCompactRepresentation}.
\begin{prop}
A sparse trigonometric polynomial $R\in\mathbb{\bar{C}}^{n}\left[X\right]$,
supported on $\mathcal{J\subseteq}\left[-n,\dots,n\right]$, admits
a projected representation on a selection matrix $C_{\mathcal{I}}$,
$\mathcal{I}\subseteq\left\llbracket 0,n\right\rrbracket $ if and
only if $\mathcal{J\subseteq\mathcal{I}}-\mathcal{I}$.\label{prop:ExistenceOfCompactRepresentation}
\end{prop}
\subsection{Real bounded lemma for sparse polynomials}
This part aims to demonstrate the novel Theorem \ref{thm:BRLSparsePol},
certifying that, when the polynomial $Q$ is sparse, the condition
$\left\Vert Q\left(e^{i\omega}\right)\right\Vert _{\infty}\leq1$
is equivalent to the existence of a positive Hermitian matrix $S$
(in a similar way as (\ref{eq:EquivalenceBoundedPolHermitianMatrix})),
whose dimension is equal to $N_{*}+1$, the essential dimension of
Problem (\ref{eq:PrimalProblem}). We latter conclude on the existence
of a compact SDP locating the spikes in $\hat{x}$ with exact precision.
The lemma bellow is first required for the demonstration of the main
theorem.
\begin{lem}
\label{lem:JointSparseGramRep}Let $R\in\mathbb{\bar{C}}^{n}\left[X\right]$
and $R'\in\mathbb{\bar{C}}^{n}\left[X\right]$ be two trigonometric
polynomials with common monomial support on $\mathcal{J\subseteq\mathcal{I}}-\mathcal{I}\subseteq\left[-n,\dots,n\right]$.
Let $\mathcal{G}_{\mathcal{I}}\left(R\right)$ and $\mathcal{G}_{\mathcal{I}}\left(R'\right)$
be respectively the Gram compact sets of $R$ and $R'$ on the selection
matrix $C_{\mathcal{I}}$. The inequality $R'\left(e^{i\omega}\right)\leq R\left(e^{i\omega}\right)$
holds for all $\omega\in\mathbb{T}$ if and only if for every two
Hermitian matrices $G\in\mathcal{G}_{\mathcal{I}}\left(R\right)$
and $G'\in\mathcal{G}_{\mathcal{I}}\left(R'\right)$, one has $G'\preceq G$.
\end{lem}
\begin{IEEEproof}
By Proposition \ref{prop:ExistenceOfCompactRepresentation}, the sets
$\mathcal{G}_{\mathcal{I}}\left(R\right)$ and $\mathcal{G}_{\mathcal{I}}\left(R'\right)$
are not empty. Thus, one can find two matrices $G\in\mathcal{G}_{\mathcal{I}}\left(R\right)$
and $G'\in\mathcal{G}_{\mathcal{I}}\left(R'\right)$. The inequality
$R'\left(e^{i\omega}\right)\leq R\left(e^{i\omega}\right)$ holds
for all $\omega\in\mathbb{T}$ if and only if, $0\leq\phi_{C_{\mathcal{I}}}\left(e^{-i\omega}\right)^{\mathrm{{\scriptscriptstyle T}}}\left(G-G'\right)\phi_{C_{\mathcal{I}}}\left(e^{i\omega}\right)$
for all $\omega\in\mathbb{T}$. Since $\phi_{C_{\mathcal{I}}}\left(e^{-i\omega}\right)=\overline{\phi_{C_{\mathcal{I}}}\left(e^{i\omega}\right)}$
and by noticing that $\left\{ \phi_{C_{\mathcal{I}}}\left(\omega\right),\;\omega\in\mathbb{T}\right\} $
spans the whole space $\mathbb{C}^{N_{*}}$, we conclude that $R'\left(e^{i\omega}\right)\leq R\left(e^{i\omega}\right)$
for all $\omega\in\mathbb{T}$ if and only if $G'\preceq G$.
\end{IEEEproof}
\begin{thm}
\label{thm:BRLSparsePol}Let $P$ and $Q$ be two polynomials from
$\mathbb{C}^{n}\left[X\right]$ with common monomial support on $\mathcal{I}$.
Define the trigonometric polynomial $R\left(z\right)=P\left(z\right)P^{*}\left(z^{-1}\right)$
for all $z\in\mathbb{C}$, and call by $r\in\mathbb{C}^{n+1}$ the
vector of its negative monomial coefficients such that $R$ can be
written under the form $R\left(z\right)=r{}_{0}+\sum_{k=1}^{n}\left(r_{k}z^{-k}+\overline{r_{k}}z^{k}\right)$,
for all $z\in\mathbb{C}^{*}$. Let by $q\in\mathbb{C}^{n+1}$ the
coefficients of $Q$ and define by $u\in\mathbb{C}^{\left|\mathcal{I}\right|}$
the vector satisfying $q=C_{\mathcal{I}}u$. Then the inequality,
\[
\left|Q\left(e^{i\omega}\right)\right|\leq\left|P\left(e^{i\omega}\right)\right|,\quad\forall\omega\in\mathbb{T},
\]
holds if and only if there exists a matrix $S\in\mathbb{\mathbb{\mathbb{C}^{\left|\mathcal{I}\right|\times\left|\mathcal{I}\right|}}}$
satisfying the conditions,
\begin{equation}
\begin{cases}
\begin{bmatrix}S & u\\
u^{H} & 1
\end{bmatrix}\succeq0\\
\mathcal{T}_{n}^{*}\left(C_{\mathcal{I}}SC_{\mathcal{I}}^{\mathrm{{\scriptscriptstyle H}}}\right)=r.
\end{cases}\label{eq:CompactConstraints}
\end{equation}
\end{thm}
\begin{IEEEproof}
The inequality $\left|Q\left(e^{i\omega}\right)\right|\leq\left|P\left(e^{i\omega}\right)\right|$
is equivalent to $\left|Q\left(e^{i\omega}\right)\right|^{2}\leq\left|P\left(e^{i\omega}\right)\right|^{2}$
for all $\omega\in\mathbb{T}$. Denote by $R$ and $R'$ the two trigonometric
polynomials $R\left(e^{i\omega}\right)=\left|P\left(e^{i\omega}\right)\right|^{2}$
and $R'\left(e^{i\omega}\right)=\left|Q\left(e^{i\omega}\right)\right|^{2}$.
It comes the equivalence with the inequality $R'\left(e^{i\omega}\right)\leq R\left(e^{i\omega}\right)$,
while $R$ and $R'$ are commonly supported by some subset $\mathcal{J}$
satisfying $\mathcal{J}\subseteq\mathcal{I}-\mathcal{I}$.
Let $q\in\mathbb{C}^{n+1}$ be the coefficients of the polynomial
$Q$. Since $R'$ is the square of $Q$, the rank one matrix $qq^{\mathrm{{\scriptscriptstyle H}}}$
belongs to $\mathcal{G}\left(R'\right)$. Moreover, $q$ is supported
by the subset $\mathcal{I}$, if and only if there exists a $u\in\mathbb{C}^{\left|\mathcal{I}\right|}$
such that $q=C_{\mathcal{I}}u$, and thus if and only if there exists
a matrix $uu^{\mathrm{{\scriptscriptstyle H}}}\in\mathcal{G}_{\mathcal{I}}\left(R'\right)$.
By application of Lemma \ref{lem:JointSparseGramRep}, an Hermitian
matrix $S\in\mathcal{G}_{\mathcal{I}}\left(R\right)$ satisfying $S\succeq uu^{\mathrm{{\scriptscriptstyle H}}}$
exists if and only if $R\left(e^{i\omega}\right)\leq R'\left(e^{i\omega}\right)$
for all $\omega\in\mathbb{T}$. We conclude by identification with
a Schur complement that the block matrix inequality in (\ref{eq:CompactConstraints})
holds if and only if $\left|Q\left(e^{i\omega}\right)\right|\leq\left|P\left(e^{i\omega}\right)\right|$,
for all $\omega\in\mathbb{T}$.
In addition, by Proposition \ref{prop:GramParametrizationTheorem},
$S\in\mathcal{G}_{\mathcal{I}}\left(R\right)$ is equivalent to $\mathcal{T}_{n}^{*}\left(C_{\mathcal{I}}SC_{\mathcal{I}}^{\mathrm{{\scriptscriptstyle H}}}\right)=r$,
which concludes the proof.
\end{IEEEproof}
\medskip{}
Applying Theorem \ref{thm:BRLSparsePol} in the specific case where
$T\left(e^{i\omega}\right)=1$, for all $\omega\in\mathbb{T}$, the
bounded polynomial constraint of Problem (\ref{eq:DualProblem}) verifies
the semidefinite equivalence,{\small{}
\[
\begin{cases}
\left\Vert Q(e^{i\omega})\right\Vert _{\infty}\leq1\\
q=C_{\mathcal{I}}c
\end{cases}\Leftrightarrow\exists S\,\text{Hermitian s.t. }\begin{cases}
\begin{bmatrix}S & c\\
c^{\mathrm{{\scriptscriptstyle H}}} & 1
\end{bmatrix}\succeq0\\
\mathcal{T}_{n_{*}}^{*}\left(C_{\mathcal{I}}SC_{\mathcal{I}}^{\mathrm{{\scriptscriptstyle H}}}\right)=e_{1}.
\end{cases}
\]
}where $e_{1}$ is the first vector of the canonical basis of $\mathbb{C}^{n_{*}}$.
Finally, we conclude on our main result, stating that Problem (\ref{eq:DualProblem})
is equivalent to the following reduced SDP,
\begin{align}
c_{*} & =\arg\max\,\Re\left(\left\langle y,c\right\rangle \right)\label{eq:ReducedSDP}\\
\text{subject to} & \phantom{\phantom{\;=\;}}\begin{bmatrix}S & c\\
c^{\mathrm{{\scriptscriptstyle H}}} & 1
\end{bmatrix}\succeq0\nonumber \\
& \phantom{\;=\;}\mathcal{T}_{n_{*}}^{*}\left(C_{\mathcal{I}}SC_{\mathcal{I}}^{\mathrm{{\scriptscriptstyle H}}}\right)=e_{1}.\nonumber
\end{align}
It is shown in \cite{FerreiraDaCosta} that due to the sparse structure
of $C_{\mathcal{I}}$, the equality constraint in (\ref{eq:ReducedSDP}),
involving vectors in $\mathbb{C}^{n_{*}}$, can be composed in $o\left(N_{*}^{2}\right)$
independent linear forms, involving a total of $\frac{N_{*}\left(N_{*}+1\right)}{2}$
variables, which do not degrade the computational complexity of Program
(\ref{eq:ReducedSDP}). By equivalence, the dual optima $c_{*}$ returned
by Problems (\ref{eq:FullSDP}) and (\ref{eq:ReducedSDP}) are similar.
Consequently, the optimal polynomial $Q_{*}\left(e^{i\omega}\right)$,
locating the spikes in $\hat{x}$, can directly be recovered from
the optimum $c_{*}$ of the compact SDP (\ref{eq:ReducedSDP}) via
the simple linear transform $q_{*}=C_{\mathcal{I}}c_{*}$.
\section{Conclusion}
In this work, we extended the theory of super-resolution from discrete
uniform samples to fit in the more generic framework of multi-rate
sampling systems. We have shown that, under the existence of a virtual
common supporting grid, one can build a dual polynomial locating with
exact precision the frequencies, as long as a minimal separation criterion
is met. The numerical complexity arising from this direct extension
can be arbitrary high. We addressed this issue in the novel Theorem
\ref{thm:BRLSparsePol} by developing an equivalence between Hermitian
matrices and bounded sparse polynomials over the unit circle. We have
derived an equivalent SDP (\ref{eq:ReducedSDP}) of optimal dimension
recovering the signal frequencies.
We reserve for a latter work a deeper exploration of the performances
of this model, including a characterization of the resolution and
spectral range benefits of MRSS, as well as an extension of this theory
to non-uniform sampling systems, by removing the common grid hypothesis,
that we believe to be artificial and unnecessary.\vfill{}
\bibliographystyle{IEEEtran}
|
\section{Introduction}
The analysis of large datasets with complicated spatio-temporal dependence structures is among the major challenges in modern statistics. With respect to this, sophisticated models are used in order to capture such dependencies. In \cite{opac-b1120661, panoramicaGenton}, an overview of these models is provided. In addition, computational efficiency is crucial, since the size of the datasets available to the scientific community has been growing steadily over the years.
A broad class of the models currently in use in spatial statistics are Gaussian latent models, discussed in full detail in \cite{rue2009}. The case of observations on a lattice is of particular interest. Several case studies are discussed, for instance in \cite{GMRFbook}, and some related modeling and computational challenges are covered in \cite{paperFlorian}. In the case of Gaussian observations $\bb{y}$ over a (regular) lattice, a prototypical model is
\begin{subequations}\label{modellogenerale}
\begin{align}
\bb{y}\mid\bb{\beta},\bb{z},\sigma & \sim \mathcal{N}(\bb{x}^\top\bb{\beta}+\bb{z}, \sigma^2\bb{I}),
\label{primo layer}
\\
\bb{z}\mid\bb{\theta} & \sim \mathcal{N}(\bb{0},\bb{Q}^{-1}(\bb{\theta})),
\label{secondo layer}
\end{align}
\end{subequations}
where $\bb{x}$ are known covariates. In the Bayesian setting (see, e.g., \citep{gelman2013bayesian}), prior distributions on the (hyper) parameters $\bb{\beta},\bb{\theta}$, and $\sigma$ are to be specified. In the previous model, the layer (\ref{primo layer}) accounts for the fact that the data $\bb{y}$ is assumed to be independent at different locations on the lattice, conditionally on a linear term $\bb{x}^\top\bb{\beta}$ and a latent field $\bb{z}$. The layer (\ref{secondo layer}) models complicated spatial dependencies through the parameter $\bb{\theta}$. In this context, the spatial correlation structure can be specified through the off-diagonal non-zero elements of the precision matrix $\bb{Q}(\bb{\theta})$ of the random vector $\bb{z}$ (see, e.g., \cite{GMRFbook} for details). Equivalently, the model (\ref{modellogenerale}) can be specified in terms of the expected value of the Gaussian full-conditional distribution at each location, in the framework of a conditional autoregressive (CAR) model. The CAR models and their importance in spatial statistics are well-known and discussed in \cite{Mardia1988265}.
The ease of interpretability of the model (\ref{modellogenerale}) comes at a price. In the applications, at least three crucial aspects are to be considered. Depending on the structure and the size of the precision matrix $\bb{Q}(\bb{\theta})$, it can be computationally intensive to evaluate quadratic forms and log-determinants involving such matrices. Of course, this can lead to high computational costs when, for instance, implementing a Markov Chain Monte Carlo (MCMC) sampler for large datasets (see, e.g., \citep{gamerman1997markov}). It is then necessary to develop efficient computational strategies in order to evaluate the above-mentioned quantities. In the literature, the case of a univariate field over a regular lattice of size $n = n_1\times n_2$ with a block-Toeplitz precision is thoroughly covered. The main idea, discussed in \cite{Mardia1988265}, is to consider a lattice with toroidal boundary conditions. Under this approximation, the precision matrix becomes block-circulant and very efficient algorithms for the latter class of matrices can be used. They are based on the (multidimensional) fast Fourier transform (FFT), discussed in full detail in \cite{libroFFT}. The third aspect that has to be considered is to determine the valid parameter space for the model~(\ref{modellogenerale}), namely the set of parameters for which the precision matrix $\bb{Q}(\bb{\theta})$ is (semi) positive-definite. The importance of this task is pointed out, for instance in \cite{spam, GMRFbook}. Once again, the solution of this problem is essential if we want to implement a MCMC sampler or an optimization algorithm for the maximum likelihood estimation for a model of the form (\ref{modellogenerale}). In the former case, we need to know the set over which we can impose a prior distribution on $\bb{\theta}$, whereas in the latter case the domain of the function to be optimized must be determined. To our knowledge, in the literature, only specific precision matrix structures (block-Toeplitz matrices, discussed in \cite{GMRFbook}, and block-circulant, e.g., in \cite{Xu12abayesian}), and univariate fields (e.g., in~\cite{paperunivariato}) are covered. The case of bivariate fields, on the contrary, was not analytically explored. So far, only approaches aimed at determining analytically tractable subsets of the domain, such as the diagonal dominance criterion, were considered, for example in \cite{GMRFbook}. The obtained results were, however, unsatisfactory in terms of coverage of the whole valid parameter space. Apart from this, the determination of the parameter space relied on brute-force \lq\lq trial-and-error\rq\rq\ approaches, based on the fact that a Hermitian matrix is positive-definite if and only if it admits a unique Cholesky factorization (see \cite{spam} for the details). However, these techniques are computationally expensive.
In this paper, we consider the model introduced in \cite{Sain:Furr:Cres:11} to analyze bivariate data coming from several regional climate models (RCMs). In particular, we focus on the associated bivariate GMRF, which is in the form of the layer (\ref{secondo layer}).
It depends on seven real parameters, namely $\phi, \rho_{11}, \rho_{12}, \rho_{21}, \rho_{22}, \tau_1$, and $\tau_2$. These in turn stem from the layer decomposition introduced in \cite{Sain:Furr:Cres:11}, which is an approach to interpret a multivariate GMRF as an instance of univariate GMRF. The main idea behind this concept is to work with one layer for each variable of interest. In this context, $\phi$ corresponds to within-location variability, whereas $\rho_{11}$ and $\rho_{22}$ describe within-variable variability. Furthermore, $\rho_{12}$ and $\rho_{21}$ account for the cross-location variability. These parameters correspond to the coefficients of a linear combination through which the expected values of the aforementioned full-conditional distributions are defined. Finally, $\tau_1^2$ and $\tau_2^2$ are the marginal variances of these bivariate distributions. Due to the different nature of these parameters, it natural to set the following:
\begin{gather*}
\bb{\theta} = (\phi,\rho_{11},\rho_{12},\rho_{21},\rho_{22})^\top,\\
\bb{\tau} = (\tau_1,\tau_2)^\top,
\end{gather*}
which induce the precision matrix $\bb{Q}_{n_1, n_2}(\bb{\theta},\bb{\tau})\in\mathbb{R}^{2n\times 2n}$.
The main goal of this paper is to provide asymptotically closed-form expressions of the valid parameter space, namely
\begin{equation}
\label{insiemevalidita}
\bb{\Theta}_{n_1, n_2} = \left\{(\bb{\theta},\bb{\tau})\in\mathbb{R}^5\times\mathbb{R}^2\mid\bb{Q}_{n_1, n_2}(\bb{\theta},\bb{\tau})\succ \bb{0}\right\}.
\end{equation}
In the previous definition, we assumed strictly positive-definite precision matrices. Although in this paper we will mainly focus on a specific bivariate GMRF, one important goal will be to provide a methodology that can be extended to multivariate GMRFs, whose precision matrix has a generalized Toeplitz matrix for which no analytic results for the eigenvalues can be obtained. In other words, the aim of our methodology is to overcome the aforementioned lack of theoretical results for multivariate GMRFs.
The structure of the paper follows. In Section \ref{The Model and Preliminary Results}, we will briefly describe the model outlined in \cite{Sain:Furr:Cres:11} and enumerate some preliminary results, which will be necessary later. In addition, extending the work presented in \cite{CIT-006}, we will develop a toroidal boundary condition approximation: $\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta},\bb{\tau})$ of $\bb{Q}_{n_1,n_2}(\bb{\theta},\bb{\tau})$ in order to characterize the domain (\ref{insiemevalidita}) through
\begin{equation}
\label{insiemevaliditaperturbato}
\widetilde{\bb{\Theta}}_{n_1, n_2} = \big\{(\bb{\theta},\bb{\tau})\in\mathbb{R}^5\times\mathbb{R}^2\mid\widetilde{\bb{Q}}_{n_1, n_2}(\bb{\theta},\bb{\tau})\succ \bb{0}\big\}.
\end{equation}
The asymmetry in the parameters $\rho_{12}$ and $\rho_{21}$, which was introduced in \cite{Sain2007226, Sain:Furr:Cres:11} in order to better describe spatial dependence, for instance in the aforementioned RCM data, will play a central role. The importance and feasibility of our approach in the applications will be further discussed. In Section~\ref{The Main Result}, we will state and formally prove our main result, namely the convergence—in a suitable sense—of $\widetilde{\bb{\Theta}}_{n_1, n_2}$ to $\bb{\Theta}_{n_1, n_2}$. This result will highlight the fact that considering a regular grid with toroidal boundary conditions is not only useful to evaluate log-determinants and quadratic forms but is also useful, for instance, to efficiently sample a prior distribution defined over the set~(\ref{insiemevalidita}). This in turn improves what is discussed in \cite{GMRFbook}, where a regular grid without boundary conditions is embedded in a bigger grid with toroidal boundary conditions. The limitations of other known convergence results for this problem (e.g., the weak convergence of sequences of matrices discussed in \cite{CIT-006}, the Szeg\"{o} theorem covered in \cite{opac-b1092435}, etc.) will also be discussed. In Section \ref{Assessment of the Rate of Convergence}, we will implement a thorough simulation study aimed at numerically assessing the rate of convergence of our approximation. We will additionally discuss some related aspects, which are important in the applications. Finally, in Section \ref{Discussion and Outlook}, we will provide some conclusive remarks.
\section{The Model and Preliminary Results}
\label{The Model and Preliminary Results}
We briefly outline the model introduced in \cite{Sain:Furr:Cres:11} in order to analyze bivariate data over a regular lattice of size $n = n_1\times n_2$, with particular emphasis on the structure of the precision matrix of the latent field $\bb{z}$ of equation (\ref{secondo layer}), which will play a crucial role in what follows. Up to a permutation of the rows and columns, the precision matrix of size $2n\times 2n$ can be rewritten as follows:
\begin{equation}
\label{matriceprecisione}
\bb{Q}_{n_1, n_2}(\bb{\theta},\bb{\tau}) =
\begin{pmatrix}
\bb{\tau}_1^{-1} & \\
& \bb{\tau}_2^{-1}
\end{pmatrix}
\begin{pmatrix}
\bb{T}_{n_1,n_2}(\rho_{11}, 1, \rho_{11}) & \bb{T}_{n_1,n_2}(\rho_{21}, \phi, \rho_{12}) \\
\bb{T}_{n_1,n_2}^\top(\rho_{21}, \phi, \rho_{12}) & \bb{T}_{n_1,n_2}(\rho_{22}, 1, \rho_{22})
\end{pmatrix}
\begin{pmatrix}
\bb{\tau}_1^{-1} & \\
& \bb{\tau}_2^{-1}
\end{pmatrix},
\end{equation}
where $\bb{\tau}_1 = \operatorname{diag}(\tau_1)\in\mathbb{R}^{n\times n}$, $\bb{\tau}_2 = \operatorname{diag}(\tau_2)\in\mathbb{R}^{n\times n}$ and $\operatorname{diag}(x)=x\,\bb{I}$.
The four blocks of the form $\bb{T}_{n_1,n_2}(\cdot)$ in the previous equation have size $n\times n$ and are block-Toeplitz, with the following structure:
\begin{equation}
\label{funzioneT}
\begin{aligned}
\bb{T}_{n_1,n_2}(x,y,z) & =
\begin{pmatrix}
\operatorname{tridiag}(x,y,z) & \operatorname{diag}(z) & & & \\
\operatorname{diag}(x) & \operatorname{tridiag}(x,y,z) & \operatorname{diag}(z) & & \\
& \operatorname{diag}(x) & \ddots & \ddots & \\
& & \ddots & \ddots & \operatorname{diag}(z)\\
& & & \operatorname{diag}(x) & \operatorname{tridiag}(x,y,z)
\end{pmatrix},
\end{aligned}
\end{equation}
where
\begin{equation*}
\begin{aligned}
\operatorname{tridiag}(x,y,z) & =
\begin{pmatrix}
y & z & & & \\
x & y & z & & \\
& x & \ddots & \ddots & \\
& & \ddots & \ddots & z \\
& & & x & y
\end{pmatrix}\in\mathbb{R}^{n_1\times n_1}.
\end{aligned}
\end{equation*}
Then, $\bb{T}_{n_1,n_2}(\rho_{11}, 1, \rho_{11})=\bb{T}_{n_1,n_2}(\rho_{22}, 1, \rho_{22})$ if and only if $\rho_{11}= \rho_{22}$. This means that $\bb{Q}_{n_1, n_2}(\bb{\theta},\bb{\tau})$ is in general not block-Toeplitz. This structure generalizes what is discussed in \cite{Kent1996379, GMRFbook}, where the precision (or, in turn, the variance-covariance) matrix is block-Toeplitz. Our precision matrix is also sparse, and the non-zero entries pattern is shown in the left panel of Figure~\ref{precisionepiccolaesempio}. It can be shown that $\bb{Q}_{n_1,n_2}(\bb{\theta},\bb{\tau})$ has at most $20n_1 n_2 - 8n_1 - 8n_2\in\Theta(n)$ non-zero entries. In this paper, we use the big theta notation, namely $f(n)\in\Theta(g(n))$ if and only if $f(n)$ is eventually bounded from both above and below by $g(n)$.
\begin{figure}
\centering
\includegraphics[scale = 0.44, trim={0 2.6cm 0 3.cm},clip]{Rplot03ter.pdf}
\caption{The left and right panel show the sparsity pattern of the precision matrix (\ref{matriceprecisione}) and of the perturbation matrix $\Delta\bb{Q}_{n_1,n_2}(\bb{\theta})$, respectively. The blue points correspond to the non-zero entries of these matrices. In the display, $n_1 = 4$ and $n_2 = 6$.}
\label{precisionepiccolaesempio}
\end{figure}
Recall that the first goal of this study is to characterize the valid parameter space~(\ref{insiemevalidita}). In the literature (e.g., in \cite{GMRFbook}) the diagonal dominance criterion is used in order to determine analytically tractable subsets. More precisely, it is well-known that an Hermitian matrix $\bb{A}\in\mathbb{C}^{n\times n}$ such that $\big\vert(\bb{A})_i^i\big\vert\geq\sum_{j\neq i} \big\vert(\bb{A})_i^j\big\vert$, for all $i = 1,2,\dots,n$, is semipositive-definite (a formal proof of this statement is provided in \cite{libroalgebralineare}). However, the converse does not hold in general. In order to highlight the importance of the convergence result that we will discuss in the next section, we will preliminarily deal with the limitations of the diagonal dominance criterion for the precision matrix (\ref{matriceprecisione}). With respect to this, we uniformly drew 500,000 values of $\bb{\theta}$ belonging to the valid parameter space (\ref{insiemevalidita}) with $n_1 = n_2 = 100$. Among these, we determined the ones also satisfying the diagonal dominance criterion. The ratio between the latter and the former was 0.1288. This rate of coverage is clearly unsatisfactory and agrees with \cite{GMRFbook} for univariate GMRFs. In addition, in order to attain a partial understanding of the geometrical structure of the set (\ref{insiemevalidita}), we additionally drew uniformly four values of $(\phi,\rho_{11},\rho_{22})^\top$ from the set (\ref{insiemevalidita}). For each of them, we uniformly sampled 10,000 values of $(\rho_{12},\rho_{21})^\top$ such that $(\phi,\rho_{11},\rho_{22}, \rho_{12},\rho_{21})^\top$ belong to (\ref{insiemevalidita}). We chose this approach since it is relatively easy to devise closed-form bounds for each of the parameters $\phi,\rho_{11}$, and $\rho_{22}$, whereas the same does not hold for $\rho_{12}$ and $\rho_{21}$. The obtained results are displayed in Figure \ref{figuradd}. The blue points correspond to values of the parameters belonging to the set (\ref{insiemevalidita}). The ones also satisfying the diagonal dominance criterion are highlighted in orange.
Several features are striking. First, there are values of $(\phi,\rho_{11},\rho_{22})^\top$ for which the diagonal dominance criterion provides no coverage. Even when it does, the coverage rate is clearly unsatisfactory. Second, the region determined by this criterion is symmetric around the origin, while the region determined by the valid points is generally not symmetric around this point. In other words, the diagonal dominance criterion is not able to capture the geometrical structure of the valid parameter space (\ref{insiemevalidita}).
\begin{figure}
\centering
\includegraphics[scale = 0.58]{plotDD}
\caption{in each panel, 10,000 values of $(\rho_{12},\rho_{21})^\top$ were uniformly drawn from the valid parameter space (\ref{insiemevalidita}) conditional on a value of $(\phi,\rho_{11},\rho_{22})^\top$. The orange points correspond to values of $\bb{\theta}$ also fulfilling the diagonal dominance criterion. The ratio between the orange and blue points is displayed on top of each panel.}
\label{figuradd}
\end{figure}
As a consequence, we need to resort to a different approach to describe the valid parameter space (\ref{insiemevalidita}). The latter problem is equivalent to determining the values of $\bb{\theta}$ and $\bb{\tau}$ such that the minimum eigenvalue of $\bb{Q}_{n_1, n_2}(\bb{\theta},\bb{\tau})$ is strictly positive. Consequently, we will focus on the spectrum of $\bb{Q}_{n_1, n_2}(\bb{\theta}, \bb{\tau})$, which will be denoted with $\sigma(\bb{Q}_{n_1, n_2}(\bb{\theta}, \bb{\tau}))$. In addition, we will adhere to the convention of enumerating the eigenvalues in decreasing order, namely $\lambda_1(\bb{Q}_{n_1, n_2}(\bb{\theta}, \bb{\tau}))\geq \lambda_2(\bb{Q}_{n_1, n_2}(\bb{\theta}, \bb{\tau}))\geq\dots\geq \lambda_{2n}(\bb{Q}_{n_1, n_2}(\bb{\theta}, \bb{\tau}))$. From equation (\ref{matriceprecisione}), it is clear that $\bb{Q}_{n_1, n_2}(\bb{\theta},\bb{\tau})$ is positive-definite only if $\tau_1,\tau_2 > 0$. Henceforth, we will only consider the inner matrix of equation (\ref{matriceprecisione}), which depends only on $\bb{\theta}$. Moreover, we will drop the dependence on $\bb{\tau}$ in order to keep the notation as straightforward as possible.
No closed-form formulas for the eigenvalues of $\bb{Q}_{n_1, n_2}(\bb{\theta})$ are available. More specifically, only results for banded block-Toeplitz matrices are discussed (see, e.g., \cite{blocktoeplitzspectra}). In the special case $\rho_{12} = \rho_{21}$, it can be shown that there exists an orthogonal matrix $\bb{V}_{n_1,n_2}\in\mathbb{R}^{n\times n}$ not depending on $\bb{\theta}$ such that
\begin{equation}
\label{caso inutile}
\begin{pmatrix}
\bb{V}_{n_1,n_2}^\top & \\
& \bb{V}_{n_1,n_2}^\top
\end{pmatrix}\,
\bb{Q}_{n_1,n_2}(\bb{\theta})\,
\begin{pmatrix}
\bb{V}_{n_1,n_2} & \\
& \bb{V}_{n_1,n_2}
\end{pmatrix} =
\begin{pmatrix}
\bb{\Lambda}_{11}(\bb{\theta}) & \bb{\Lambda}_{12}(\bb{\theta})\\
\bb{\Lambda}_{12}(\bb{\theta}) & \bb{\Lambda}_{22}(\bb{\theta})
\end{pmatrix},
\end{equation}
where $\bb{\Lambda}_{ij}(\bb{\theta})$ is a diagonal matrix whose diagonal elements are the eigenvalues of the four $n\times n$ symmetric blocks into which the inner matrix of precision (\ref{matriceprecisione}) is partitioned. In other words, the latter are simultaneously diagonalizable through $\bb{V}_{n_1,n_2}$. This property is a consequence of the fact that
\begin{eqnarray*}
\bb{T}_{n_1,n_2}(x,y,x) & = & y\left(\bb{I}_{n_2}\otimes\bb{I}_{n_1}\right)+x\left(\bb{S}_{n_2}\otimes\,\bb{I}_{n_1}\right)+
x\left(\bb{I}_{n_2}\otimes\bb{S}_{n_1}\right),\\
\bb{S}_n & = & \operatorname{tridiag}(1, 0, 1),
\end{eqnarray*}
as also observed in \cite{paperunivariato} when dealing with univariate GMRFs with a second-order neighborhood. Up to a permutation of the rows and columns, the right-hand-side of the equation (\ref{caso inutile}) is a symmetric block-diagonal matrix whose diagonal blocks have size $2\times 2$. At this point, the determination of $\sigma(\bb{Q}_{n_1,n_2}(\bb{\theta}))$ is straightforward.
In the general case in which $\rho_{12}\neq\rho_{21}$, the four $n\times n$ blocks of the precision (\ref{matriceprecisione}) are not simultaneously diagonalizable. Therefore, the spectrum cannot be readily devised. The main idea to overcome this lack of analytical results is to introduce a perturbed precision matrix $\widetilde{\bb{Q}}_{n_1, n_2}(\bb{\theta})\in\mathbb{R}^{2n\times 2n}$, which can be partitioned into four $n\times n$ blocks which are block-circulant. This will extend the approach used in \cite{CIT-006} and is equivalent to toroidal boundary conditions on the underlying lattice.
The perturbation is carried out as follows. Bearing in mind how the function $\bb{T}_{n_1,n_2}(\cdot)$ was defined in (\ref{funzioneT}), we see that all but its sub-blocks on the main diagonal are already circulant. On the other hand, the sub-blocks on the main diagonal are by construction tridiagonal and hence can be made circulant in a natural way as follows:
\begin{equation*}
\operatorname{tridiag}(x,y,z)
\mapsto
\operatorname{circ}(x,y,z) =
\begin{pmatrix}
y & z & & & x\\
x & y & z & & \\
& x & \ddots & \ddots & \\
& & \ddots & \ddots & z \\
z & & & x & y
\end{pmatrix}.
\end{equation*}
We explicitly point out that this method of turning a tridiagonal (more generally, non-circulant Toeplitz) matrix into a circulant one is not unique. In the literature, several approaches are discussed (see, e.g., \cite{ideaminimizzarefrobenius}). However, in our case, a simulation study provided strong empirical evidence of the fact that more sophisticated approaches are advantageous only for small choices of the grid size $n$. As we are interested in an asymptotic description of $\sigma(\bb{Q}_{n_1, n_2}(\bb{\theta}))$, we decided to stick to the natural approach. We then set:
\begin{equation}
\label{bcapproximation}
\begin{aligned}
\bb{\widetilde{Q}}_{n_1,n_2}(\bb{\theta}) & =
\begin{pmatrix}
\bb{C}_{n_1,n_2}(\rho_{11}, 1, \rho_{11}) & \bb{C}_{n_1,n_2}(\rho_{21}, \phi, \rho_{12}) \\
\bb{C}_{n_1,n_2}^\top(\rho_{21}, \phi, \rho_{12}) & \bb{C}_{n_1,n_2}(\rho_{22}, 1, \rho_{22})
\end{pmatrix}\in\mathbb{R}^{2n\times 2n},\\
\bb{C}_{n_1,n_2}(x,y,z) & =
\begin{pmatrix}
\operatorname{circ}(x,y,z) & \operatorname{diag}(z) & & & \operatorname{diag}(x)\\
\operatorname{diag}(x) & \operatorname{circ}(x,y,z) & \operatorname{diag}(z) & & \\
& \operatorname{diag}(x) & \ddots & \ddots & \\
& & \ddots & \ddots & \operatorname{diag}(z)\\
\operatorname{diag}(z) & & & \operatorname{diag}(x) & \operatorname{circ}(x,y,z)
\end{pmatrix}\in\mathbb{R}^{n\times n}.
\end{aligned}
\end{equation}
It follows that the blocks $\bb{C}_{n_1,n_2}(\cdot)$ are block-circulant matrices of size $n$. Similarly to what was observed above for $\bb{Q}_{n_1,n_2}(\bb{\theta})$, the perturbed matrix (\ref{bcapproximation}) is block-circulant if and only if $\rho_{11} = \rho_{22}$.
Now, let $\Delta\bb{Q}_{n_1, n_2}(\bb{\theta}) = \bb{\widetilde{Q}}_{n_1, n_2}(\bb{\theta}) - \bb{{Q}}_{n_1,n_2}(\bb{\theta})$ be the perturbation matrix (see the right panel of Figure \ref{precisionepiccolaesempio} to inspect its structure). It has at most $8(n_1+n_2)\in\Theta(n_1+n_2)$ non-zero entries. This implies that the limit for $n_1,n_2\to +\infty$ of the ratio between the number of non-zero elements of $\bb{Q}_{n_1,n_2}(\bb{\theta})$ and of $\Delta\bb{Q}_{n_1,n_2}(\bb{\theta})$ is infinite. This property will be crucial in Lemma~\ref{lemmafacile}, when we will prove our main convergence result. Apart from this, the trace of $\Delta\bb{Q}_{n_1,n_2}(\bb{\theta})$ is zero, and its rank is greater than zero, unless ${\rho_{11} = \rho_{12} = \rho_{21} = \rho_{22} = 0}$. This implies that $\Delta\bb{Q}_{n_1, n_2}(\bb{\theta})$ is an indefinite matrix; namely that it has both strictly positive and strictly negative eigenvalues. Explicit formulas for these eigenvalues can be symbolically derived, as only their algebraic multiplicity changes when $n$ grows, but they are not reported here, as they are rather complicated and not necessary for what follows. It is also important to point out that the Weyl's inequalities \citep{paperoriginaleWeyl} imply that
\begin{equation*}
\lambda_{2n}(\bb{Q}_{n_1, n_2}(\bb{\theta})) +\lambda_{2n}(\Delta\bb{Q}_{n_1, n_2}(\bb{\theta}))
\leq
\lambda_{2n}(\widetilde{\bb{Q}}_{n_1, n_2}(\bb{\theta}))
\leq
\lambda_{2n}(\bb{Q}_{n_1, n_2}(\bb{\theta})) +\lambda_{1}(\Delta\bb{Q}_{n_1, n_2}(\bb{\theta})).
\end{equation*}
These bounds are however too loose to approximate the valid parameter space (\ref{insiemevalidita}).
We now turn to the study of $\sigma(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))$, which will allow us to characterize the set~(\ref{insiemevaliditaperturbato}). In the case of a block-circulant of size $n_1\,n_2\times n_1\,n_2$ whose sub-blocks have size $n_1\times n_1$, the eigenvalues can be efficiently computed using the bidimensional FFT, with a computational complexity of $\Theta(n_1\,n_2\,\log{(n_1\,n_2)})$, as discussed in \cite{libroFFT}. For the perturbed precision (\ref{bcapproximation}), we obtain:
\begin{align}
\begin{split}
\label{autovaloriblocco12}
\sigma(\bb{C}_{n_1,n_2}(x,y,z)) = & \left\{y + z\,\exp{\left(-2\pi\imath\left(\frac{i}{n_2} + \frac{j}{n_1}\right)\right)} + \right.\\
& \left.
+x\,
\exp{\left(-2\pi\imath\left(\frac{i(n_2-1)}{n_2} + \frac{j(n_1-1)}{n_1}\right)\right)}
\right\}_{i\in I,\ j\in J},
\end{split}
\end{align}
where $\imath$ is the imaginary unit, $I = \{0, 1, 2,\dots, n_2-1\}$ and $J = \{0,1,2,\dots,n_1-1\}$. This result readily provides the eigenvalues of the four $n\times n$ blocks of (\ref{bcapproximation}). In particular, the eigenvalues of the two diagonal blocks are real, since the latter are by construction symmetric. For instance, for the top-left block of (\ref{bcapproximation}) it holds that
\begin{equation}
\label{autovaloriblocco11}
\sigma(\bb{C}_{n_1,n_2}(\rho_{11},1,\rho_{11})) = \left\{1+2\rho_{11}\left(\cos{\left(\frac{2\pi i}{n_2}\right)} + \cos{\left(\frac{2\pi j}{n_1}\right)} \right)\right\}_{i\in I,\ j\in J}.
\end{equation}
We are now ready to explicitly derive $\sigma(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))$. Let
\begin{equation}
\label{matriceU}
\bb{U}_{n_1,n_2} = \bb{F}_{n_1}^{(n_2)}{\bb{P}_{n_1}^{(n_2)}}\bb{F}_{n_2}^{(n_1)}\bb{P}_{n_2}^{(n_1)},
\end{equation}
where, after having set $\omega = \exp{(-2\pi\imath/n)}$:
\begin{align*}
\bb{F}_n = &\, \frac{1}{\sqrt{n}}
\begin{pmatrix}
1 & 1 & 1 & \dots & 1\\
1 & \omega & \omega^2 & \dots & \omega^{n-1}\\
1 & \omega^2 & \omega^4 & \dots & \omega^{2(n-1)}\\
\vdots & \vdots & \vdots & & \vdots\\
1 & \omega^{n-1} & \omega^{2(n-1)} & \dots & \omega^{(n-1)(n-1)}
\end{pmatrix},\\
\bb{F}_{n}^{(m)} = & \,\bb{I}_{2m}\otimes\bb{F}_n,
\end{align*}
and $\bb{P}_{n}^{(m)}\in\mathbb{R}^{2mn\times 2mn}$ is a suitable permutation matrix. Of course, $\bb{F}_n$ is the normalized discrete Fourier transform (DFT) matrix, discussed in full detail in \cite{libroFFT}. Closed-form expression for $\bb{U}_{n_1,n_2}$ can be obtained, but once again they are not reported because they are not strictly necessary for what follows. It will suffice to observe that its non-zero entries are generally complex, with absolute value $1/{\sqrt{n_1n_2}}$. The non-zero entry pattern is displayed in Figure \ref{trasformazioneblocchetti}.
\begin{figure}
\centering
\vspace*{-6mm}
\includegraphics[scale = 0.38]{UniversalSmallBlock2ter}
\vspace*{-6mm}
\caption{non-zero entry pattern of $\bb{U}_{n_1,n_2}$.}
\label{trasformazioneblocchetti}
\end{figure}
In addition, $\bb{U}_{n_1,n_2}$ is by construction unitary, does not depend on $\bb{\theta}$, and is such that
\begin{equation}
\label{diagonalizzazioneblocchetti2x2}
\bb{U}_{n_1,n_2}^*\,
\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})\,
\bb{U}_{n_1,n_2} =
\operatorname{blkdiag}(
\begin{pmatrix}
\widetilde{\lambda}_{i,j}^{(n)}(\rho_{11}, 1, \rho_{11}) & \widetilde{\lambda}_{i,j}^{(n)}(\rho_{21}, \phi, \rho_{12})\\
\widetilde{\lambda}_{i,j}^{(n)}(\rho_{12}, \phi, \rho_{21}) & \widetilde{\lambda}_{i,j}^{(n)}(\rho_{22}, 1, \rho_{22})
\end{pmatrix}),
\end{equation}
where $^*$ denotes the complex conjugate operator, $i\in I, j\in J$ and $\widetilde{\lambda}_{i,j}^{(n)}(x,y,z)$ is the $(i,j)$th-~eigenvalue of the block-circulant block $\bb{C}_{n_1,n_2}(x,y,z)$ of size $n$ belonging to the matrix (\ref{bcapproximation}). From the previous result, we finally obtain closed-form expressions for $\sigma(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))$, namely
\begin{equation}
\begin{multlined}
\label{autovaloribc}
\frac{1}{2}\Bigg({\widetilde{\lambda}_{i,j}^{(n)}(\rho_{11}, 1, \rho_{11})+\widetilde{\lambda}_{i,j}^{(n)}(\rho_{22}, 1, \rho_{22})}\\
\pm{\sqrt{\left(\widetilde{\lambda}_{i,j}^{(n)}(\rho_{11}, 1, \rho_{11}) - \widetilde{\lambda}_{i,j}^{(n)}(\rho_{22}, 1, \rho_{22})\right)^2+4\big\vert\widetilde{\lambda}_{i,j}^{(n)}(\rho_{21}, \phi, \rho_{12})\big\vert^2}}\Bigg),
\end{multlined}
\end{equation}
From equations (\ref{matriceU}) and (\ref{diagonalizzazioneblocchetti2x2}), several important results follow. First, a unitary matrix $\bb{V}_{n_1,n_2}(\bb{\theta})$ of size $2n\times 2n$ exists, such that the perturbed precision matrix $\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})$ can be diagonalized by the transformation $\bb{U}_{n_1,n_2}\bb{V}_{n_1,n_2}(\bb{\theta})$. It clearly follows that the columns of $\bb{U}_{n_1,n_2}\bb{V}_{n_1,n_2}(\bb{\theta})$ are eigenvectors of $\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})$. In addition, each column of $\bb{V}_{n_1,n_2}(\bb{\theta})$ contains exactly two non-zero entries. They correspond to an eigenvector of one of the $n_1n_2$ blocks of size $2\times 2$ of equation (\ref{diagonalizzazioneblocchetti2x2}). This readily provides an estimate of the absolute value of the entries of $\bb{U}_{n_1,n_2}\bb{V}_{n_1,n_2}(\bb{\theta})$. Without loss of generality, we assume that the two non-zero entries of $(\bb{V}_{n_1,n_2}(\bb{\theta}))^m$ are its two first components. Then,
\begin{eqnarray}
\nonumber
\vert(\bb{U}_{n_1,n_2}\bb{V}_{n_1,n_2}(\bb{\theta}))_l^m\vert &
= & \vert(\bb{U}_{n_1,n_2})_l\, (\bb{V}_{n_1,n_2}(\bb{\theta}))^m\vert\\
\nonumber
& = & \vert (\bb{U}_{n_1,n_2})_l^1\, (\bb{V}_{n_1,n_2}(\bb{\theta}))_1^m + (\bb{U}_{n_1,n_2})_l^2 \,(\bb{V}_{n_1,n_2}(\bb{\theta}))_2^m\vert\\
\nonumber
& \leq & \frac{1}{\sqrt{n_1n_2}}(\vert(\bb{V}_{n_1,n_2}(\bb{\theta}))_1^m \vert + \vert (\bb{V}_{n_1,n_2}(\bb{\theta}))_2^m\vert)\\
\nonumber
& \leq & \frac{\sqrt{2}}{\sqrt{n_1n_2}}\\
& \in & \Theta\left(\frac{1}{\sqrt{n_1n_2}}\right),
\label{boundDiagonalizzazionePerturbata}
\end{eqnarray}
where $l,m=1,2,\dots,2n$. The first inequality holds due to the triangular inequality and the above stated norm of the entries of $\bb{U}_{n_1,n_2}$, while the second one holds since $\bb{V}_{n_1,n_2}(\bb{\theta})$ is unitary and hence $\vert(\bb{V}_{n_1,n_2}(\bb{\theta}))_1^m \vert^2 + \vert (\bb{V}_{n_1,n_2}(\bb{\theta}))_2^m\vert^2 = 1$.
To summarize, in this section, we solved the analogous of problem (\ref{insiemevalidita}) for the perturbed matrix~(\ref{bcapproximation}). In other words, we found closed-form expressions for the set (\ref{insiemevaliditaperturbato}). In fact, it is enough to impose that all $2n$ eigenvalues (\ref{autovaloribc}) are strictly positive. The next step is to prove asymptotically that $\bb{Q}_{n_1,n_2}(\bb{\theta})$ is positive-definite if and only if this property holds for~$\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})$.
We conclude this section by discussing the importance of the knowledge of the eigenvalues~(\ref{autovaloribc}) in the applications. The main result in this context is that we can test whether $\bb{\theta}$ belongs to the perturbed valid parameter space (\ref{insiemevaliditaperturbato}) by numerically checking that the eigenvalues (\ref{autovaloribc}) evaluated at $\bb{\theta}$ are positive. Once again, in the next section, we will show that this task is asymptotically equivalent to testing whether $\bb{\theta}$ belongs to (\ref{insiemevalidita}). Here, we highlight the usefulness of the closed-form expressions (\ref{autovaloribc}) in the applications from a general perspective. A uniform prior distribution on $\bb{\theta}$ has the following form:
\begin{equation*}
\pi(\bb{\theta})\propto\prod_{i = 1}^{2n} \bb{1}(\lambda_i(\bb{Q}_{n_1,n_2}(\bb{\theta})) >0 ),
\end{equation*}
where $\bb{1}(\cdot)$ is the indicator function. We need to evaluate all the $n$ terms of the form $\bb{1}(\lambda_i(\bb{Q}_{n_1,n_2}(\bb{\theta})) >0 )$ because the analytical expression of $\lambda_{2n}(\bb{Q}_{n_1,n_2}(\bb{\theta}))$ depends on complicated conditions on the components of $\bb{\theta}$. Through equation (\ref{autovaloribc}), this task can be accomplished with complexity $\Theta(n)$. The gain in computational efficiency is substantial. We benchmarked our approach against using the \texttt{R} package \texttt{spam} \cite{spam}, which provides numerical routines for sparse matrix algebra. We considered three grid sizes, namely $100\times 100$, $200\times 200$, and $300\times 300$. We observe that, if $\bb{\theta}$ induces a positive-definite precision matrix, it is then necessary to construct the latter (for instance, to evaluate the corresponding log-likelihood function), although our method in principle does not require this construction. On the contrary, if our approach determines that $\bb{\theta}$ does not belong to (\ref{insiemevalidita}), then it is not necessary to construct the corresponding precision, as the drawn $\bb{\theta}$ is to be discarded. In order to meaningfully assess both scenarios, for each of the aforementioned grid sizes, we drew 50 values of $\bb{\theta}$ belonging to the valid parameter (\ref{insiemevalidita}) and 50 values not belonging to this domain and recorded the correspondent computational times. For each of these 100 parameter values, we recorded the computational time in both \texttt{spam} and our approach. This task was accomplished through the \texttt{R} package \texttt{microbenchmark}~\cite{microbenchmark}. In the case of the 50 valid parameter values and for the three considered grid sizes, the median of the recorded computational times for \texttt{spam} was 2.6, 5.9, and 11.6 times the median obtained with our approach. The performance of our approach clearly improves when the grid size increases. In the case of the non-valid parameters, for which it was not necessary to construct the precision matrix, the median of \texttt{spam} was 40.6, 84.7, and 208.7 times the median of our approach. These timings can be further improved by exploiting the structure of the set (\ref{insiemevaliditaperturbato}), which is analytically provided by equation (\ref{autovaloribc}). It can be shown that, if we assume without loss of generality that $n_1<n_2$, the eigenvalues~(\ref{autovaloribc}) attain exactly $n_2$ local minima. It follows that we only need to evaluate $n_2$ elements of the form $\bb{1}(\lambda_i(\bb{Q}_{n_1,n_2}(\bb{\theta})) >0 )$ in order to test whether $\bb{\theta}$ belongs to the set~(\ref{insiemevaliditaperturbato}). From all the above-mentioned results, it is striking that our method outperforms the numerical routines of \texttt{spam}. This property is crucial, for instance, in an MCMC sampler, where the task of determining whether $\bb{\theta}$ is valid is usually repeated hundreds of thousands of times. Further aspects of our methodology are discussed in Section \ref{Discussion and Outlook}.
\section{The Main Result}
\label{The Main Result}
\todo{I would \emph{not} include Gray's weak convergence in this manuscript, for reasons to \textbf{be discussed}}
In the previous section, the perturbation $\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})$ of $\bb{Q}_{n_1,n_2}(\bb{\theta})$ was introduced. We will now formally prove our main result.
\begin{theorem}
\label{Reinhardconjecture}
Let $\bb{\theta}\in\mathbb{R}^5$. Then, for the precision matrices (\ref{matriceprecisione}) and (\ref{bcapproximation}), it holds that
\begin{equation*}
\lim_{n_1,n_2\to +\infty}\big\vert{\lambda_{2n}(\bb{Q}_{n_1,n_2}(\bb{\theta})) - \lambda_{2n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))}\big\vert = 0.
\end{equation*}
\end{theorem}
The previous result implies that, asymptotically, $\bb{Q}_{n_1,n_2}(\bb{\theta})$ is positive-definite if and only if this property holds for $\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})$.
First, we provide evidence of its novelty with respect to what is discussed in the literature in the following.
\begin{remark}
\label{remark altre convergenze}
\begin{enumerate}
\item In the literature, several convergence results are discussed for sequences of (Hermitian) matrices of the form $\{\bb{A}_n\}_{n\in\mathbb{N}}$ and $\{\bb{B}_n\}_{n\in\mathbb{N}}$, where $\bb{A}_n,\bb{B}_n\in\mathbb{C}^{n\times n}$. In general, they do not imply Theorem \ref{Reinhardconjecture}. For instance, in \cite{CIT-006}, the weak convergence is discussed, which is widely used in many other references, for instance in \cite{GMRFbook}. It is defined in terms of the scaled Frobenius norm of the difference $\bb{A}_n-\bb{B}_n$, which approaches zero as $n\to +\infty$, and the spectral norms of the two sequences, which have to be bounded by a constant not depending on $n$. However, this type of convergence only implies that, for large grid sizes, the spectra behave similarly as a whole. In other words, nothing can generally be stated for the single eigenvalues. For instance, we set $\bb{A}_n=\bb{I}_n$ and ${\bb{B}_n=\operatorname{diag}(1/2,1,1,\dots,1)\in\mathbb{R}^{n\times n}}$. It can be shown that the weak convergence holds, but clearly there is no convergence for the two sequences $\lbrace\lambda_n(\bb{A}_n)\rbrace_{n\in\mathbb{N}}$ and $\lbrace\lambda_n(\bb{B}_n)\rbrace_{n\in\mathbb{N}}$.
\item Other types of convergence were introduced in the literature, like the finite-term strong convergence discussed in \cite{convergenzafortefinita}. The goal was to overcome some of the limitations of the weak convergence in the framework of evaluating quadratic forms and log-determinants involving variance-covariance matrices with a Toeplitz structure. However, nothing can be said for the convergence of individual eigenvalues, even in this stronger framework.
\item Another classical result that, under certain technical conditions, relates the spectra of (non-circulant) Toeplitz matrices to the spectra of circulant matrices is the first Szeg\"{o} theorem, discussed in \cite{opac-b1092435}. This result was later extended in \cite{Tilli199859} to block-Toeplitz matrices. As already stated, the structure of the precision (\ref{matriceprecisione}) is more general and does not fulfill the hypotheses of the mentioned extension.
\end{enumerate}
\end{remark}
In order to prove Theorem \ref{Reinhardconjecture}, we preliminarily recall that
\begin{eqnarray}
\label{ottimizzazione1}
\lambda_{2n}(\bb{Q}_{n_1,n_2}(\bb{\theta})) & = & \min_{\bb{x}\in\mathbb{C}^{2n}\mid\Vert\bb{x}\Vert = 1}\langle \bb{Q}_{n_1,n_2}(\bb{\theta})\,\bb{x},\bb{x}\rangle,\\
\label{ottimizzazione2}
\lambda_{2n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})) & = & \min_{\bb{x}\in\mathbb{C}^{2n}\mid\Vert\bb{x}\Vert = 1}\langle \widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})\,\bb{x},\bb{x}\rangle,
\end{eqnarray}
where $\langle\cdot,\cdot\rangle$ is the Euclidean inner product on $\mathbb{C}^{2n}$ and $\Vert\cdot\Vert$ is the correspondent induced norm. These results can be straightforwardly proved by diagonalizing the matrices ${\bb{Q}}_{n_1,n_2}(\bb{\theta})$ and $\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})$, respectively. In what follows, we will denote with $\bb{v}_{n_1,n_2}(\bb{\theta})$ and $\bb{u}_{n_1,n_2}(\bb{\theta})$ a solution of (\ref{ottimizzazione1}) and (\ref{ottimizzazione2}), respectively. They are eigenvectors associated with the eigenvalues $\lambda_{2n}({\bb{Q}}_{n_1,n_2}(\bb{\theta}))$ and $\lambda_{2n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))$, respectively.
Based on equation (\ref{autovaloribc}) and (\ref{boundDiagonalizzazionePerturbata}), the following lemmas can be shown.
\begin{lemma}
\label{lemma ulteriore}
For any $\bb{\theta}\in\mathbb{R}^5$, there is a constant $C(\bb{\theta})$ such that, for any $n_1,n_2\in\mathbb{N}$, ${C(\bb{\theta})\leq\lambda_{2n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))}$ and
$$
\lim_{n_1,n_2\to +\infty} \lambda_{2n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})) = C(\bb{\theta}).
$$
\end{lemma}
\begin{proof}
From Euler’s formula applied to equation (\ref{autovaloriblocco12}), it can be derived that
\begin{align}
\begin{split}
\label{bound norma}
\Re{\left(\widetilde{\lambda}_{i,j}^{(n)}(\bb{C}(\rho_{21}, \phi, \rho_{12}))\right)} = &\ \phi + \rho_{12}\cos{\left(\frac{2\pi i}{n_2}\right)} + \rho_{21}\cos{\left(\frac{2\pi i}{n_2}\right)} \\
& + \rho_{12}\cos{\left(\frac{2\pi j}{n_1}\right)} + \rho_{21}\cos{\left(\frac{2\pi j}{n_1}\right)},\\
\Im{\left(\widetilde{\lambda}_{i,j}^{(n)}(\bb{C}(\rho_{21}, \phi, \rho_{12}))\right)} = & - \rho_{12}\sin{\left(\frac{2\pi i}{n_2}\right)} + \rho_{21}\sin{\left(\frac{2\pi i}{n_2}\right)} \\
& - \rho_{12}\sin{\left(\frac{2\pi j}{n_1}\right)} + \rho_{21}\sin{\left(\frac{2\pi j}{n_1}\right)},
\end{split}
\end{align}
where $\Re(z)$ and $\Im(z)$ denote, respectively, the real and imaginary part of a complex number $z$. Let $\mathbb{S}$ be the unitary circle in $\mathbb{C}$. From the previous equalities and equation (\ref{autovaloriblocco11}), it follows that the eigenvalues (\ref{autovaloribc}) correspond to the image of the finite subset:
\begin{equation*}
\left\{\left(\sin{\left(\frac{2\pi j}{n_1}\right)},\, \cos{\left(\frac{2\pi j}{n_1}\right)}\right)\right\}_{j\in J}
\times
\left\{\left(\sin{\left(\frac{2\pi i}{n_2}\right)},\, \cos{\left(\frac{2\pi i}{n_2}\right)}\right)\right\}_{i\in I}
\end{equation*}
of $\mathbb{S}\times\mathbb{S}$ through a continuous function defined by $\mathbb{S}\times\mathbb{S}$ to $\mathbb{R}$ by equation (\ref{autovaloribc}) itself. We shall denote this function with $\Phi(\cdot)$. Since $\mathbb{S}\times\mathbb{S}$ is a compact subset of $\mathbb{C}^2$, the existence of the constant ${C}(\bb{\theta})$ follows from the Weierstrass extreme value theorem.
Let $(\bb{x}^\ast,\bb{y}^\ast)\in\mathbb{S}\times\mathbb{S}$ such that $\Phi((\bb{x}^\ast,\bb{y}^\ast)) = C(\bb{\theta})$. From equations (\ref{autovaloriblocco11}) and (\ref{bound norma}), there is a sequence $\lbrace(\bb{x}_{n_1},\bb{y}_{n_2})\rbrace_{n_1,n_2\in\mathbb{N}}\subset\mathbb{S}\times\mathbb{S}$ of the form
\begin{align*}
\bb{x}_{n_1} = &\left\{\left(\sin{\left(\frac{2\pi j_{n_1}}{n_1}\right)},\, \cos{\left(\frac{2\pi j_{n_1}}{n_1}\right)}\right)\right\}_{n_1\in\mathbb{N}},\\
\bb{y}_{n_2} = &\left\{\left(\sin{\left(\frac{2\pi i_{n_2}}{n_2}\right)},\, \cos{\left(\frac{2\pi i_{n_2}}{n_2}\right)}\right)\right\}_{n_2\in\mathbb{N}},
\end{align*}
where $i_{n_2}\in I$ and $j_{n_1}\in J$, such that $(\bb{x}_{n_1},\bb{y}_{n_2})\to(\bb{x}^\ast,\bb{y}^\ast)$.
Then, for continuity, $${\vert \Phi((\bb{x}_{n_1},\bb{y}_{n_2}))- C(\bb{\theta})\vert\to 0}.$$ However, $C(\bb{\theta})\leq\lambda_{2n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})) \leq \Phi((\bb{x}_{n_1},\bb{y}_{n_2}))$, hence $${\vert C(\bb{\theta}) - \lambda_{2n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))\vert \leq\vert \Phi((\bb{x}_{n_1},\bb{y}_{n_2})) - C(\bb{\theta})\vert},$$ from which the second claim follows.
\end{proof}
\begin{lemma}
\label{lemmafacile}
For any $\bb{\theta}\in\mathbb{R}^5$, it holds that
\begin{equation*}
\lim_{n_1,n_2\to +\infty} \big\vert{\langle \bb{Q}_{n_1,n_2}(\bb{\theta})\,\bb{u}_{n_1,n_2}(\bb{\theta}),\bb{u}_{n_1,n_2}(\bb{\theta})\rangle -\lambda_{2n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))}\big\vert = 0.
\end{equation*}
\end{lemma}
\begin{proof}
The argument of the limit is equal to $\big\vert\langle\Delta\bb{Q}_{n_1,n_2}(\bb{\theta})\,\bb{u}_{n_1,n_2}(\bb{\theta}),\bb{u}_{n_1,n_2}(\bb{\theta})\rangle\big\vert$. Hence
\begin{eqnarray*}
\nonumber
\big\vert\langle\Delta\bb{Q}_{n_1,n_2}(\bb{\theta})\,\bb{u}_{n_1,n_2}(\bb{\theta}),\bb{u}_{n_1,n_2}(\bb{\theta})\rangle\big\vert &= &
\left\vert\sum_{l = 1}^{2n}\sum_{m = 1}^{2n}(\Delta\bb{Q}_{n_1,n_2}(\bb{\theta}))_l^m\,(\bb{u}_{n_1,n_2}(\bb{\theta}))_l\,(\bb{u}_{n_1,n_2}(\bb{\theta}))_m^*\right\vert\\
\nonumber
& \leq & \sum_{l = 1}^{2n}\sum_{m = 1}^{2n}\left\vert (\Delta\bb{Q}_{n_1,n_2}(\bb{\theta}))_l^m\right\vert\,\frac{1}{n_1n_2}\\
& \leq & 8\,K(\bb{\theta})\,\frac{n_1+n_2}{n_1 n_2},
\end{eqnarray*}
where $K(\bb{\theta})$ is a constant depending on $\bb{\theta}$ but not on $n$. This constant exists because the entries of $\Delta\bb{Q}_{n_1,n_2}(\bb{\theta})$ do not depend on the grid size $n$. In addition, we used the bound (\ref{boundDiagonalizzazionePerturbata}) and the fact that the latter matrix has at most $8(n_1+n_2)$ non-zero entries.
\end{proof}
In addition to the previous results, in order to prove Theorem \ref{Reinhardconjecture}, we will construct a $2\times 2$ block matrix $\bb{P}_{n_1,n_2}(\bb{\theta})\in\mathbb{R}^{8n\times 8n}$ with the structure that we now describe. In \cite{tesidottoratoembedding}, an approach to embed a symmetric block-Toeplitz matrix of size $
n_1n_2\times n_1n_2$ into a symmetric block-circulant matrix of size $4n_1n_2\times 4n_1n_2$ is proposed. Here, we extend this technique in a natural way to the structure of the precision matrix (\ref{matriceprecisione}). Recall that the latter is partitioned into four blocks, which are block-Toeplitz (see equation (\ref{funzioneT}) for their structure). We will start by embedding the $n_1\times n_1$ Toeplitz sub-blocks in $2n_1\times 2n_1$ circulant matrices. The sub-blocks of $\bb{T}_{n_1,n_2}(\cdot)$, which are diagonal, are embedded in a diagonal matrix of size $2n_1\times 2n_1$. In addition, we embed the $n_1\times n_1$ tridiagonal blocks in $\operatorname{circ}(x,y,z)\in\mathbb{R}^{2n_1\times 2n_1}$. For the sake of compactness of the notation, we will denote with $\operatorname{diag}_{2}(x)$ the former matrix and with $\operatorname{circ}_{2}(x,y,z)$ the latter. In what follows, it will be important to explicitly observe that
\begin{align}
\begin{split}
\label{embedding blocchetti esplicito}
\operatorname{diag}_{2}(x) = &
\begin{pmatrix}
\operatorname{diag}(x) & \bb{0}\\
\bb{0} & \operatorname{diag}(x)
\end{pmatrix},\\
\operatorname{circ}_{2}(x,y,z) = &
\begin{pmatrix}
\operatorname{tridiag}(x,y,z) & \ast\\
\ast & \operatorname{tridiag}(x,y,z)
\end{pmatrix}.
\end{split}
\end{align}
We define
\begin{equation*}
\bb{P}_{n_1,n_2}(\bb{\theta}) =
\begin{pmatrix}
\bb{E}_{n_1,n_2}(\rho_{11}, 1, \rho_{11}) & \bb{E}_{n_1,n_2}(\rho_{21}, \phi, \rho_{12})\\
\bb{E}_{n_1,n_2}(\rho_{21}, \phi, \rho_{12})^\top & \bb{E}_{n_1,n_2}(\rho_{22}, 1, \rho_{22})
\end{pmatrix}\in\mathbb{R}^{8n\times 8n},
\end{equation*}
where $\bb{E}_{n_1,n_2}(x,y,z)\in\mathbb{R}^{4n\times 4n}$ is given by
\begin{equation*}
\begin{pmatrix}
\operatorname{circ}_{2}(x,y,z) & \operatorname{diag}_{2}(z) & & & \operatorname{diag}_{2}(x)\\
\operatorname{diag}_{2}(x) & \operatorname{circ}_{2}(x,y,z) & \operatorname{diag}_{2}(z) & & \\
& \operatorname{diag}_{2}(x) & \ddots & \ddots & \\
& & \ddots & \ddots & \operatorname{diag}_2(z)\\
\operatorname{diag}_{2}(z) & & & \operatorname{diag}_2(x) & \operatorname{circ}_{2}(x,y,z)
\end{pmatrix}.
\end{equation*}
The construction of the matrix $\bb{E}_{n_1,n_2}(x,y,z)$ and the next lemma are better understood by means of the following
\begin{example}
Set $n_2 = 3$: then
\begin{equation*}
\bb{T}_{n_1, 3}(x,y,z) =
\begin{pmatrix}
\operatorname{tridiag}(x,y,z) & \operatorname{diag}(z) & \bb{0} \\
\operatorname{diag}(x) & \operatorname{tridiag}(x,y,z) & \operatorname{diag}(z)\\
\bb{0} & \operatorname{diag}(x) & \operatorname{tridiag}(x,y,z) \\
\end{pmatrix}.
\end{equation*}
The matrix $\bb{E}_{n_1,3}(x,y,z)$ is obtained by circulation from
\begin{small}
\begin{equation*}
\left(
\begin{array}{c c | c c | c c | c c | c c | c c}
\operatorname{tridiag}(x,y,z) & \ast & \operatorname{diag}(z) & \bb{0} & \bb{0} & \bb{0} & \bb{0} & \bb{0} & \bb{0} & \bb{0} & \operatorname{diag}(x) & \bb{0}\\
\ast & \operatorname{tridiag}(x,y,z) & \bb{0} & \operatorname{diag}(z) & \bb{0} & \bb{0} & \bb{0} & \bb{0} & \bb{0} & \bb{0} & \bb{0} & \operatorname{diag}(x)\\
\hline
\vdots & \vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots &\vdots
\end{array}
\right).
\end{equation*}
\end{small}
In general, in the expression above, there are $2n_2-3$ blocks of the form $\bb{0}_{2n_1\times 2n_1}$.
\end{example}
We prove the main properties of the matrix $\bb{P}_{n_1,n_2}(\bb{\theta})$ in the following
\begin{lemma}
\label{lemmaP}
Let $\bb{\theta}\in\mathbb{R}^5$. Then,
\begin{enumerate}
\item $\lambda_{8n}(\bb{P}_{n_1,n_2}(\bb{\theta})) \leq \lambda_{2n}({\bb{Q}}_{n_1,n_2}(\bb{\theta}))$
\item $\lim_{n_1,n_2\to +\infty} \vert \lambda_{2n}(\bb{\widetilde{Q}}_{n_1,n_2}(\bb{\theta}))-\lambda_{8n}(\bb{P}_{n_1,n_2}(\bb{\theta}))\vert = 0$
\end{enumerate}
\end{lemma}
\begin{proof}
\begin{enumerate}
\item Up to a permutation of the rows and columns of $\bb{P}_{n_1,n_2}(\bb{\theta})$, it holds that
\begin{equation}
\label{permutazionefurba}
\bb{P}_{n_1,n_2}(\bb{\theta}) =
\begin{pmatrix}
\bb{Q}_{n_1,n_2}(\bb{\theta}) & \ast\\
\ast & \ast
\end{pmatrix}.
\end{equation}
In order to show this claim, we introduce the set of indices given by
\begin{equation*}
\mathcal{I} = \{1,2,\dots,n_1,\,2n_1+1,\dots,3n_1,\,\dots\dots,\,2(n_2-1)n_1+1,\dots,(2n_2-1)n_1 \},
\end{equation*}
which has cardinality $n_1n_2$. Then $\bb{T}_{n_1,n_2}(\rho_{11}, 1, \rho_{11}) = (\bb{E}_{n_1,n_2}(\rho_{11}, 1, \rho_{11}))_i^i$, with $i\in\mathcal{I}$, as a consequence of (\ref{funzioneT}) and (\ref{embedding blocchetti esplicito}). This line of reasoning, when applied block-wise to $\bb{P}_{n_1,n_2}(\bb{\theta})$, implies (\ref{permutazionefurba}). The well-known Cauchy interlacing theorem, discussed in \cite{InterlacingInequalities}, yields the inequality $\lambda_{8n}(\bb{P}_{n_1,n_2}(\bb{\theta}))\leq \lambda_{2n}(\bb{Q}_{n_1,n_2}(\bb{\theta}))$.
\item The matrix $\bb{P}_{n_1,n_2}(\bb{\theta})$ is partitioned into four blocks $\bb{E}_{n_1,n_2}(\cdot)$, which are of the form $\bb{C}_{n_1,n_2}(\cdot)$, as defined in equation (\ref{bcapproximation}). Therefore, the eigenvalues of $\bb{P}_{n_1,n_2}(\bb{\theta})$ can be obtained from equation (\ref{autovaloriblocco12}) by considering a regular lattice of size $2n_1\times 2n_2$. Hence, $\lbrace\lambda_{8n}(\bb{P}_{n_1,n_2}(\bb{\theta}))\rbrace_{n\in\mathbb{N}}$ is a sub-sequence of $\lbrace\lambda_{2n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))\rbrace_{n\in\mathbb{N}}$. For Lemma \ref{lemma ulteriore}, the latter sequence is convergent, hence it is a Cauchy sequence.
\end{enumerate}
\end{proof}
Now, we have all the results that we need to show Theorem \ref{Reinhardconjecture}.
\begin{proof}[Proof of Theorem~\ref{Reinhardconjecture}]
It holds that
\begin{equation*}
\lambda_{2n}(\bb{Q}_{n_1,n_2}(\bb{\theta})) = \langle \bb{Q}_{n_1,n_2}(\bb{\theta})\,\bb{v}_{n_1,n_2}(\bb{\theta}),\bb{v}_{n_1,n_2}(\bb{\theta})\rangle
\leq
\langle \bb{Q}_{n_1,n_2}(\bb{\theta})\,\bb{u}_{n_1,n_2}(\bb{\theta}),\bb{u}_{n_1,n_2}(\bb{\theta})\rangle,
\end{equation*}
as $\bb{v}_{n_1,n_2}(\bb{\theta})$ is a solution of the variational problem (\ref{ottimizzazione1}). In addition:
\begin{equation}
\label{catenageniale}
\lambda_{2n}(\bb{\widetilde{Q}}_{n_1,n_2}(\bb{\theta}))\sim\lambda_{8n}(\bb{P}_{n_1,n_2}(\bb{\theta}))\leq
\lambda_{2n}(\bb{Q}_{n_1,n_2}(\bb{\theta}))\leq \langle\bb{Q}_{n_1,n_2}(\bb{\theta})\,\bb{u}_{n_1,n_2}(\bb{\theta}),\bb{u}_{n_1,n_2}(\bb{\theta})\rangle,
\end{equation}
where for $\{ a_n\}_{n\in\mathbb{N}}$ and $\{b_n\}_{n\in\mathbb{N}}$ we write, with slight abuse of notation, $a_n\sim b_n$ if and only if $\lim_{n\to +\infty}\vert a_n-b_n\vert = 0$. Now, $\lambda_{2n}(\bb{\widetilde{Q}}_{n_1,n_2}(\bb{\theta}))\sim\lambda_{8n}(\bb{P}_{n_1,n_2}(\bb{\theta}))$ is the second claim of Lemma~\ref{lemmaP}, while $\lambda_{8n}(\bb{P}_{n_1,n_2}(\bb{\theta}))\leq
\lambda_{2n}(\bb{Q}_{n_1,n_2}(\bb{\theta}))$ is the first claim thereof. The chain (\ref{catenageniale}) is therefore proved by applying Lemma \ref{lemmafacile}. It implies Theorem \ref{Reinhardconjecture}.
\end{proof}
\begin{remark}
\label{remarkfalsepositives}
The chain (\ref{catenageniale}) does not imply that $\lambda_{2n}(\bb{\widetilde{Q}}_{n_1,n_2}(\bb{\theta}))\leq \lambda_{2n}(\bb{{Q}}_{n_1,n_2}(\bb{\theta}))$. If it would have been possible to theoretically show this inequality, then we would not have needed to construct the matrix $\bb{P}_{n_1,n_2}(\bb{\theta})$. On the other hand, this property is beneficial in the applications. If the perturbed precision (\ref{bcapproximation}) is positive-definite, then the precision (\ref{matriceprecisione}) has the same property. This avoids the case of \lq\lq false positives,\rq\rq namely efficiently sampling from (\ref{insiemevaliditaperturbato}) values of $\bb{\theta}$ which do not belong to the domain (\ref{insiemevalidita}).
\end{remark}
From the previous proof, only a loose bound of the rate of convergence of \mbox{Theorem \ref{Reinhardconjecture}} can be found. This issue stems from the fact we had to consider the auxiliary quantities $ \langle\bb{Q}_{n_1,n_2}(\bb{\theta})\,\bb{u}_{n_1,n_2}(\bb{\theta}),\bb{u}_{n_1,n_2}(\bb{\theta})\rangle$ and $\lambda_{8n}(\bb{P}_{n_1,n_2}(\bb{\theta}))$ in order to avoid the lack of analytical results on the spectrum of ${\bb{Q}}_{n_1,n_2}(\bb{\theta})$. As a consequence, we can only get loose bounds on $\vert \lambda_{2n}(\bb{{Q}}_{n_1,n_2}(\bb{\theta})) - \lambda_{2n}(\bb{\widetilde{Q}}_{n_1,n_2}(\bb{\theta}))\vert$ in terms of the just mentioned two quantities.
In the next section, we will provide a thorough numerical study aimed at empirically obtaining a tight upper bound of the rate of convergence.
\section{Assessment of the Rate of Convergence}
\label{Assessment of the Rate of Convergence}
In this section, we numerically assess the rate of convergence in Theorem \ref{Reinhardconjecture}. As a preliminary step, we briefly inspect the asymptotic behavior in both the case of a univariate GMRF over a transect of size $n$, namely a grid constructed on a line, and over a bidimensional regular lattice of size $n_1\times n_2$. This will be helpful later to interpret the more general case of approximating the precision (\ref{matriceprecisione}) with the matrix (\ref{bcapproximation}). For a univariate GMRF, the eigenvalues of the associated precision matrix can be analytically determined. On the contrary, in the bivariate case, we will need to resort to numerical techniques, since the eigenvalues of the precision (\ref{matriceprecisione}) are not known, as already stated. We will keep the notation as close as possible to the one that we used when we introduced precision (\ref{matriceprecisione}). Since we will be concerned with matrices whose structure is similar to or the same as the one of the $n_1n_2\times n_1n_2$ diagonal blocks of (\ref{matriceprecisione}), we will therefore denote the off-diagonal non-zero elements with $\rho$.
For a univariate field over a transect of length $n$, the Toeplitz precision $\operatorname{tridiag}(\rho,1,\rho)$ is naturally approximated by $\operatorname{circ}(\rho,1,\rho)$. The eigenvalues of a Toeplitz tridiagonal matrix are well-known (see, e.g., \cite{NLA1811}). It holds that ${\lambda_n(\operatorname{tridiag}(\rho,1,\rho)) = 1 - 2\vert \rho\vert\,\cos{(\pi/(n+1))}}$. In addition, if $\rho < 0$, then ${\lambda_n(\operatorname{circ}(\rho,1,\rho)) = 1 - 2\vert \rho\vert}$, whereas, if $\rho > 0$:
\begin{equation*}
\lambda_n(\operatorname{circ}(\rho,1,\rho)) =
\begin{cases}
1 - 2\vert \rho\vert, & \text{if $n$ is even},\\
1 + 2\vert \rho\vert \,\cos{\left(\frac{2\pi}{n}\lfloor \frac{n}{2}\rfloor\right)}, & \text{if $n$ is odd}.
\end{cases}
\end{equation*}
We consider the absolute error $\epsilon_n(\rho) =\vert \lambda_n(\operatorname{tridiag}(\rho,1,\rho)) - \lambda_n(\operatorname{circ}(\rho,1,\rho)\vert$. Then, if $\rho < 0$, $1-2\vert\rho\vert=\lambda_n(\operatorname{circ}(\rho, 1, \rho))\leq\lambda_n(\operatorname{tridiag}(\rho, 1,\rho))$ and $\epsilon_n(\rho)\in\Theta(1/n^2)$. This can be proved by expanding the function $n\mapsto\cos{(\pi/(n+1))}$ in the Taylor series evaluated at $n = \infty$. Otherwise, if $\rho > 0$, the same order of convergence can be obtained if $n$ is even. If $n$ is odd, $\epsilon_n(\rho)\in\Theta(1/n^3)$, but $\lambda_n(\operatorname{tridiag}(\rho, 1,\rho))\leq\lambda_n(\operatorname{circ}(\rho, 1, \rho))$. These different rates of convergence, which depend on the parity of $n$, will allow us to better interpret the convergence pattern in the more general case of a bivariate field over a regular lattice.
We now consider the case of a univariate GMRF over a bidimensional regular lattice of size $n = n_1\times n_2$, whose precision matrix is $\bb{T}_{n_1,n_2}(\rho, 1,\rho)$. The eigenvalues of this matrix are explicitly derived in \cite{paperunivariato}. Based on our approach, $\bb{T}_{n_1,n_2}(\rho, 1,\rho)$ is approximated with $\bb{C}_{n_1,n_2}(\rho, 1,\rho)$. Similarly to what was done above, we consider the absolute error $\epsilon_n(\rho) = \vert\lambda_{n}(\bb{T}_{n_1,n_2}(\rho, 1,\rho)) - \lambda_{n}(\bb{C}_{n_1,n_2}(\rho, 1,\rho))\vert$. Using closed-form expressions (\ref{autovaloribc}), it can be shown that, if $\rho < 0$, then ${\epsilon_n \in \Theta({1}/{n})}$, regardless of the parity of $n_1$ and $n_2$. If $\rho > 0$, we need to distinguish between the following three cases:
\begin{itemize}
\item If $n_1$ and $n_2$ are even, then ${\epsilon_n \in \Theta({1}/{n})}$ and $\lambda_{n}(\bb{C}_{n_1,n_2}(\rho, 1,\rho))\leq\lambda_{n}(\bb{T}_{n_1,n_2}(\rho, 1,\rho))$. In addition, $\lambda_{n}(\bb{C}_{n_1,n_2}(\rho, 1,\rho)) = C(\rho)$, where $C(\rho)$ is a constant not depending on~$n$.
\item If exactly one of $n_1$ or $n_2$ is even, then the same rate of convergence of the previous case holds and $\lambda_{n}(\bb{C}_{n_1,n_2}(\rho, 1,\rho))\leq\lambda_{n}(\bb{T}_{n_1,n_2}(\rho, 1,\rho))$, but $\{\lambda_{n}(\bb{C}_{n_1,n_2}(\rho, 1,\rho))\}_{n\in\mathbb{N}}$ is not a constant sequence.
\item If $n_1$ and $n_2$ are odd, then ${\epsilon_n \in \Theta({1}/{n^{1.5}})}$, but, in contrast to the previous two cases, $\lambda_{n}(\bb{C}_{n_1,n_2}(\rho, 1,\rho))\geq\lambda_{n}(\bb{T}_{n_1,n_2}(\rho, 1,\rho))$.
\end{itemize}
The proof of the latter three claims is analogous to the ones of the transect case. For what follows, it is important to note that the asymptotic behavior of the error depends on the parity of $n_1$ and $n_2$ and on the sign of $\rho$. We also point out that similar results are discussed in \citep{GMRFbook} with respect to the maximum likelihood estimator of (univariate) GMRFs and the dimension of the associated lattice.
We now return to the more general case of approximating the precision (\ref{matriceprecisione}) with the matrix (\ref{bcapproximation}) in the bivariate case over a regular lattice. We will provide numerical evidence of the fact that the rate of convergence in Theorem \ref{Reinhardconjecture} is in the form $\Theta(1/n)$. Similarly to what was observed for a univariate field over a regular lattice, the pattern of convergence of the minimum eigenvalue of the perturbed precision matrix (\ref{bcapproximation}) to the minimum eigenvalue of the precision (\ref{matriceprecisione}) will depend on both the parity of $n_1$ and $n_2$ and on the value of $\bb{\theta}$. Due to the higher dimension of the set (\ref{insiemevalidita}), the fluctuation pattern as $n$ grows will require a more thorough study. Apart from this, we recall that in Remark \ref{remarkfalsepositives} we briefly discussed the importance in the applications of the absence of \lq\lq false positives.\rq\rq\ In what follows, we will also take this property into account.
We consider the following numerical experiment. First, we uniformly sample $\bb{\theta}_1,\bb{\theta}_2,\dots,\bb{\theta}_N$ from $[-1, 1]^5$, which can be shown to be a superset of (\ref{insiemevalidita}) for any $n_1,n_2\in\mathbb{N}$. Second, for each of these draws, we consider increasing grid sizes, and we numerically determine the minimum eigenvalues of the perturbed precision (\ref{bcapproximation}) and of the precision matrix for each of them (\ref{matriceprecisione}). For the former eigenvalue, we use the closed-form expressions (\ref{autovaloribc}). For the latter eigenvalue, we use the \texttt{eigs} command of the \texttt{R} package \texttt{rARPACK} \cite{bruttacopiamatlab}, which implements the Lanczos algorithm \cite{Lanczos50aniterative} to determine the extreme eigenvalues (and possibly eigenvectors) of a square complex matrix. Third, we numerically compute the constant $C(\bb{\theta})$ introduced in Lemma \ref{lemma ulteriore} and focus on the quantities
\begin{eqnarray*}
\epsilon_{n_1,n_2}(\bb{\theta}) & = & \big\vert \lambda_{2n}(\bb{\widetilde{Q}}_{n_1,n_2}(\bb{\theta})) - \lambda_{2n}(\bb{{Q}}_{n_1,n_2}(\bb{\theta}))\big\vert,\\
\delta_{n_1,n_2}(\bb{\theta}) & = & \vert \lambda_{2n}(\bb{{Q}}_{n_1,n_2}(\bb{\theta})) - C(\bb{\theta})\vert.
\end{eqnarray*}
The second quantity was introduced because we will provide numerical evidence of the fact that $\lambda_{2n}(\bb{\widetilde{Q}}_{n_1,n_2}(\bb{\theta}))\leq\lambda_{2n}(\bb{{Q}}_{n_1,n_2}(\bb{\theta}))$ definitely, hence $\epsilon_{n_1,n_2}(\bb{\theta})\leq\delta_{n_1,n_2}(\bb{\theta})$ definitely due to chain (\ref{catenageniale}). Moreover, $\delta_{n_1,n_2}(\bb{\theta})$ exhibits an asymptotic convergence pattern that does not depend on the parity of $n_1,n_2$ or on the value of $\bb{\theta}$, in contrast to $\epsilon_{n_1,n_2}(\bb{\theta})$.
In the following, we set $N = 500$ and considered the grid sizes $50\times 50$ to $300\times 300$, for a total of 251 increasing values. For each of the 500 drawn values of $\bb{\theta}$ we fitted a least-squares line of $\log_{10}{(\delta_{n_1,n_2}(\bb{\theta}))}$ against $\log_{10}{(n_1n_2)}$ (recall that the precision matrix (\ref{matriceprecisione}) has size $2n_1 n_2\times 2n_1 n_2$). As a measure of the quality of the fit, we considered the (non-adjusted) ~$R^2$. The minimum observed $R^2$ was 0.999, which provides very strong evidence for a rate of convergence of $\delta_{n_1,n_2}(\bb{\theta})$ in the form $\Theta(1/n^\alpha)$, without any fluctuations. Furthermore, since the empirical distribution of the $500$ estimated slopes had median $-0.992$ and the first and third quartiles were $-1.002$ and $-0.989$, respectively, we empirically conclude that $\delta_{n_1,n_2}(\bb{\theta})\in\Theta(1/n)$ or, equivalently, that $\alpha = 1$.
The convergence pattern of $\lambda_n(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))$ to $\lambda_n({\bb{Q}}_{n_1,n_2}(\bb{\theta}))$ is more complicated and depends on the value of $\bb{\theta}$, and on the parity of $n_1$ and $n_2$. In Figure \ref{differentialplot}, three values thereof are displayed.
\begin{figure}
\centering
\vspace*{-6pt}
\includegraphics[scale = 0.69]{PlotConvergenceStudy.pdf}
\caption{in the left column, the convergence pattern of $\lambda_{n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))$ to $\lambda_{n}({\bb{Q}}_{n_1,n_2}(\bb{\theta}))$ against increasing grid sizes is displayed for three different values of $\bb{\theta}$. In the right column, the correspondent errors $\epsilon_{n_1, n_2}(\cdot)$, and $\delta_{n_1,n_2}(\cdot)$ are plotted on a log-log scale.}
\label{differentialplot}
\end{figure}
They were chosen in order to provide a meaningful overview of all observed convergence patterns. The left column shows the obtained results on the usual scale, while the right column displays the convergence of $\delta_{n_1,n_2}(\bb{\theta})$ and $\epsilon_{n_1,n_2}(\bb{\theta})$ to zero on a log-log scale. As stated above, the convergence pattern of the former quantity is a straight line of slope $-1$, whereas the pattern of the latter depends on both the parity of $n_1,n_2$ and the value of $\bb{\theta}$. This fact is very similar to what has already been observed in the beginning of this section for univariate fields over a transect and a regular lattice. More precisely, with respect to the first considered value of $\bb{\theta}$, it holds that ${\lambda_{n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})) = C(\bb{\theta})}$ if and only if $n_1$ and $n_2$ are even, whereas $\lambda_{n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))\to C(\bb{\theta})$, but the sequence is not constant, if and only if $n_1$ and $n_2$ are odd. In addition, the absolute error $\epsilon_{n_1,n_2}(\bb{\theta})$ corresponds to a line with slope $-1$ in the log-log scale, with the intercept dependent on the parity of $n_1$ and $n_2$, and $\lambda_{n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))\leq \lambda_{n}({\bb{Q}}_{n_1,n_2}(\bb{\theta}))$ for every considered grid size. The second considered value of $\bb{\theta}$ exhibits a similar behavior for grid sizes with both $n_1$ and $n_2$ even. The spike that can be seen for odd values of $n_1$ and $n_2$ corresponds to the grid size at which the sign of ${\lambda_{n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))-\lambda_{n}({\bb{Q}}_{n_1,n_2}(\bb{\theta}))}$ changes. For larger grid-sides, it holds that ${\lambda_{n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))\leq \lambda_{n}({\bb{Q}}_{n_1,n_2}(\bb{\theta}))}$. Finally, the third considered value of $\bb{\theta}$ shows the most complicated convergence pattern. From the right panel, it is clear that the convergence patterns of grid sizes with even and odd values of $n_1$ and $n_2$ are similar. Differently from the above discussed cases, $\epsilon_{n_1,n_2}(\bb{\theta})$ does not (asymptotically) correspond to a line. In addition, it is empirically clear from the left panel that $\lambda_{n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))\leq \lambda_{n}({\bb{Q}}_{n_1,n_2}(\bb{\theta}))$, hence $\epsilon_{n_1,n_2}(\bb{\theta})\leq\delta_{n_1,n_2}(\bb{\theta})$. In other words, regardless of the complex fluctuation pattern, the absolute error is bounded from above by $\delta_{n_1,n_2}(\bb{\theta})$, which has the nice convergence pattern discussed above.
To summarize, in this section, we provided strong empirical evidence of the fact that the rate of convergence in Theorem \ref{Reinhardconjecture} is $\Theta(1/n)$. In addition, it holds that, asymptotically, ${\lambda_{n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))\leq \lambda_{n}({\bb{Q}}_{n_1,n_2}(\bb{\theta}))}$. As discussed in Remark~\ref{remarkfalsepositives}, this property implies that, for large grid sizes, our methodology does not lead to \lq\lq false positives.\rq\rq
\clearpage
\section{Discussion and Outlook}
\label{Discussion and Outlook}
In this paper, asymptotically closed-form expressions for the determination of the valid parameter space (\ref{insiemevalidita}) through a suitable approximation (\ref{insiemevaliditaperturbato}) were provided. The importance of this result in the applications is twofold. First, it provides an efficient tool in order to efficiently sample from (\ref{insiemevalidita}). This task is crucial, for instance, in the framework of a Metropolis–Hastings within the Gibbs sampler aimed at analyzing multivariate and highly correlated data with the model introduced in \cite{Sain:Furr:Cres:11} or extensions thereof. Second, current work involves the generalization of the convergence results discussed in \cite{CIT-006} by means of the eigenvalues (\ref{autovaloribc}) and Theorem~\ref{Reinhardconjecture}. The goal is to efficiently evaluate quadratic forms and log-determinants involving matrices of the form (\ref{matriceprecisione}) without the need for computing the Cholesky decomposition of large sparse matrices, for example with the routines provided by \texttt{spam} \cite{spam}. This is a very demanding task from a computational point view. The need for determining the Cholesky factor when sampling from the latent field $\bb{z}$ of the form (\ref{secondo layer}) can also be avoided by exploiting the results discussed in \cite{Aune2013}.
The rate of convergence in Theorem \ref{Reinhardconjecture} was numerically assessed to be $\Theta(1/n)$. Moreover, we provided evidence of the fact that, asymptotically, $\lambda_{2n}(\bb{\widetilde{Q}}_{n_1,n_2}(\bb{\theta}))\leq
\lambda_{2n}(\bb{Q}_{n_1,n_2}(\bb{\theta}))$. In the applications, this avoids \lq\lq false positives\rq\rq\ in the approximation of (\ref{insiemevalidita}) with (\ref{insiemevaliditaperturbato}). In fact, if $\lambda_{2n}(\widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta}))>0$, namely $ \widetilde{\bb{Q}}_{n_1,n_2}(\bb{\theta})$ is positive-definite, then $\lambda_{2n}({\bb{Q}}_{n_1,n_2}(\bb{\theta}))>0$; hence, ${\bb{Q}}_{n_1,n_2}(\bb{\theta})$ is also positive-definite.
The closed-form expressions (\ref{autovaloribc}) allow us to inspect the geometrical properties of the parameter space (\ref{insiemevalidita}) (e.g., connection and compactness), which play a central when implementing likelihood function optimizers as discussed, e.g., in \citep{bb47473, 54906}. Some preliminary insight of the shape of the set (\ref{insiemevalidita}) is already provided in Figure \ref{figuradd}. Several theoretical results on the geometrical structure of the valid parameter space (\ref{insiemevalidita}) were also obtained, but they are not included in this paper because they are beyond its scope.
From a statistical perspective, the model introduced in \cite{Sain:Furr:Cres:11} is such that the correlation structure of different locations on the lattice depends only on the parameters $\rho_{11},\rho_{22},\rho_{12}$, and $\rho_{21}$. Our knowledge of the valid parameter space (\ref{insiemevalidita}) will allow us to add a further layer to the above-mentioned model, which is in the form (\ref{modellogenerale}). For instance, $\bb{\theta}$ could be written as a suitable function of latitude and longitude. This function must ensure that the precision matrix (\ref{matriceprecisione}) is positive-definite, hence the importance of the knowledge of the eigenvalues (\ref{autovaloribc}).
The strategy used in Section \ref{The Main Result} to prove the main Theorem~\ref{Reinhardconjecture}, namely the introduction of the chain (\ref{catenageniale}), will be used to prove similar results for the more general case of $p$-variate GMRFs.
The associated precision matrix can be partitioned in $p^2$ block-Toeplitz blocks of size ${n_1n_2\times n_1n_2}$, but, depending on the parametrization of the GMRF, it may not be block-Toeplitz. Once again, we point out that no extensions of the classical convergence results described in Remark~\ref{remark altre convergenze} are (to our knowledge) available for this framework. We will therefore apply a perturbation similar to the one introduced in equation (\ref{bcapproximation}) and then obtain a closed-form expression similar to the eigenvalues (\ref{autovaloribc}). Apart from this, another interesting research question to address would be to explore the applicability of our approach for neighborhood structures of higher order in the lattice.
\section*{Acknowledgments}
This work was supported by the Swiss National Science Foundation, grant 143282. The first author would like to thank Antonio De Rosa, Giuseppe Graziani, and Salvatore Stuvard for the helpful discussions while developing the proof of the main result.
The content of this paper is part of the first author’s Ph.D. dissertation supervised by the second author.
\renewcommand{\baselinestretch}{1.}\rm
\bibliographystyle{as}
|
\section*{SUPPLEMENTARY MATERIAL}
In this supplementary material we first discuss the fabrication and characterization techniques (Section I). In Section II, we show that two synchronously bending waveguides are equivalent to two straight waveguides with some effective coupling constant. In Section III the techniques used to obtain the topological phase diagram of Fig.~\ref{fig2} (main text) are presented. We discuss wavelength tuning and the effect of disorder in Section IV before finally concluding with a study of how this disorder affects the topology in Section V.
\subsection*{I. Fabrication and characterisation}
The photonic lattice with two driving periods was fabricated inside a 70-mm-long glass (Corning Eagle$^{2000}$) substrate using the ultrafast laser inscription technique, where the refractive index profile of each waveguide was controlled using the ``slit-beam shaping" method \cite{davis1996writing, ams2005slit}. The glass substrate, mounted on $x$-$y$-$z$ translation stages, was translated at 8~mm/s through the focus of a 500~kHz train of 1030~nm femtosecond laser pulses to fabricate each waveguide. The laser inscription parameters were optimized to produce waveguides that were single-mode and well confined in the measurement wavelength range of 700-830~nm. To study the response of the edge modes in the presence of a defect at the edge, the waveguide at the (8, 4) lattice site was not fabricated. It should also be highlighted that the waveguide paths are designed such that all waveguides exhibit identical bend radii at a given $z$, although the direction of this bending is site dependent. This ensures that there is minimal site-dependent losses.
To measure the coupling constants, we fabricated five sets of each bond separately inside the same substrate. These bonds (or couplers) were then characterised in the wavelength range 705-795~nm, to obtain the variation of the mean and standard deviation of coupling strength with wavelength (see supplementary for details).
To excite the lattices with different wavelengths, a photonic crystal fiber was pumped by sub-picosecond laser pulses of 1064~nm wavelength to generate a broadband supercontinuum. A tunable monochromator placed after the supercontinuum source was used to select narrow band ($\approx 3$~nm) light, which was coupled into an optical fibre (SMF-600). The fiber was then coupled to the lattice sites. The output intensity distribution was observed using a CMOS camera. A polarizer passing only vertically polarized light is placed in front of the camera to ensure that the measurements are not affected by polarization-dependent coupling.
\subsection*{II. Coupling between bending waveguides}
In the experimental setup, the lattice constant was chosen such that, in the absence of any bending, the waveguides comprising the lattice were practically uncoupled. In order to turn on the coupling between two neighbouring sites [see Fig.~\ref{fig1}~(b) in the main text], the two waveguides were bent together and then remained straight before moving apart again [see Fig.~\ref{fig1}~(d) in the main text]. The coupling between two such synchronously bending waveguides was demonstrated in Ref.~\cite{Huang} to be equivalent to an effective coupling between two straight waveguides. This mapping is vital in order for the experiment to emulate our theoretical model, which assumes that the lattice is comprised of straight waveguides but with controllable couplings to neighbouring waveguides [see Fig.~\ref{fig1}~(b) in the main text]. In light of the importance of this result in the present work, a derivation is provided here for completeness.
The dynamics of how light propagates through an array of weakly-guiding waveguides is governed by the scalar Helmholtz equation,
\begin{equation}
\nabla^2\psi+\frac{1}{\lambdabar^2}n^2(x,y,z)\psi=0. \notag
\end{equation}
Here $\psi(x,y,z)$ is the electric field amplitude, $\lambdabar=\frac{\lambda}{2\pi}$, where $\lambda$ is the wavelength of the light in free space, and $n(x,y,z)$ is the refractive index profile in the domain of interest. If the wave propagation is primarily along the $z$-axis, then the field $E(x,y,z)$ can be represented as a slowly varying complex field amplitude and a fast oscillating wave, $\psi(x,y,z)=E(x,y,z)e^{i\frac{n_s}{\lambdabar}z}$. Substituting $\psi$ into the Helmholtz equation and making the slowly varying envelope approximation for $E$ gives
\begin{equation}
i\lambdabar\frac{\partial E}{\partial z}=-\frac{\lambdabar^2}{2n_s}(\frac{\partial^2}{\partial x^2}+\frac{\partial^2}{\partial x^2})E+\frac{1}{2n_s}(n_s^2-n^2)E.
\label{Para}
\end{equation}
The refractive index profile can be written as $n(x,y,z)=n_s+\Delta n(x,y,z)$. However, in laser-written waveguide arrays, the refractive index change caused by the laser is small and so $\frac{1}{2n_s}(n_s^2-n^2)\approx -\Delta n(x,y,z)=V[x,y,z]$.
Consider now an array of moving waveguides where $V[x,y,z]$ is given by
\begin{equation}
V[x,y,z]=\sum_k V_0^{k}
=\sum_{k} V_0[x-x_{k}^0-x_{k}(z),y-y_{k}^0].
\notag
\end{equation}
Here $V_0(x,y)$ is the refractive index profile of a single isolated waveguide, the index $k$ labels the different waveguides and $(x_k^{0}+x_k(z),y_k^{0})$ is the $z$-dependent position of the $k$th waveguide. \\
In the regime where the waveguides are far apart, compared to the width of $V_0$, the electric field can be well approximated by a superposition of the fundamental modes of the different waveguides,
\begin{equation}
E(x,y,z)=\sum_k c_k(z)u[x-x_k(z),y]e^{-i\beta z},
\notag
\end{equation}
where $u[x,y]$ is the fundamental mode of an isolated waveguide and satisfies the equation
\begin{equation}
-\beta u[x,y]=-\frac{\lambdabar^2}{2n_s}(\frac{\partial^2}{\partial x^2}+\frac{\partial^2}{\partial y^2})u[x,y]+V_0[x,y]u[x,y].
\label{Eigenmode}
\end{equation}
Inserting the electric field expansion into Eq.~\eqref{Para} gives, after some rearrangement,
\begin{align}
&i\lambdabar\sum_k(\dot{c}_k u_k+c_k\frac{\partial u_k}{\partial x}(-\dot{x}_k))\!=\!\sum_k c_k \bigg [\!-\!\frac{\lambdabar^2}{2n_s}\bigg(\frac{\partial^2}{\partial x^2}\!+\!\frac{\partial^2}{\partial y^2}\bigg)\nonumber \\
&+\beta +V_0^{k} \bigg]u_k+ \sum_{n \neq k} V_0^{n} \sum_k c_k u_k,
\end{align}
where we have introduced the notation $u_k=u[x-x_k^0-x_k(z),y-y_k^0]$. The expression in the square brackets is equal to zero as a result of Eq.~\eqref{Eigenmode}.
\\ Multiplying by $u_m$ and integrating transversely gives
\begin{equation}
i\sum_k(\dot{c}_k p_{m,k}+c_k f_{m,k})=\sum_k c_k t_{m,k},
\label{CMT}
\end{equation}
where
\begin{eqnarray}
\notag
p_{m,k}=& \int_{-\infty}^{\infty}\int_{-\infty}^{\infty} dxdy u_m u_k,\\
\notag
f_{m,k}=& -\dot{x}_k \int_{-\infty}^{\infty}\int_{-\infty}^{\infty} dxdy u_m \frac{\partial u_k}{\partial x},\\
\notag
t_{m,k}=& \frac{1}{\lambdabar}\sum_{n\neq k} \int_{-\infty}^{\infty}\int_{-\infty}^{\infty} dxdy V_0^{n} u_m u_k.
\end{eqnarray}
A number of pertinent statements can be made about these integrals:
\begin{eqnarray}
\notag
p_{m,k}=&p_{k,m},\\
\notag
f_{m,m}=&0,\\
\notag
f_{m,k}=&-\dot{x}_k I_{m,k},\\
\notag
f_{k,m}=&\dot{x}_m I_{m,k},\\
\notag
I_{m,k}=& \int_{-\infty}^{\infty}\int_{-\infty}^{\infty} dxdy u_m \frac{\partial u_k}{\partial x}.
\notag
\end{eqnarray}
These statements have the corollary that
\begin{equation}
\frac{\partial }{\partial z} p_{m,k}=\dot{p}_{m,k}=I_{m,k}(\dot{x}_{m}-\dot{x}_k).
\notag
\end{equation}
As previously mentioned, in the absence of any bending there is negligible coupling between any of the waveguides. The consequence of this is that in the presence of bending, only pairs of waveguides are coupled at any given moment. The coupled mode equations for this system can then be written as block diagonal matrices with each block comprising a $2\times 2$ matrix. This simplifies the following to a study of a two waveguide coupler as the analysis will straightforwardly generalise to the full lattice. Therefore, if there are two waveguides, labelled $1$ and $2$, that are synchronously bending, such that $\dot{x}_1=-\dot{x}_2$, then it can be readily seen that $\dot{p}_{1,2}=2f_{1,2}=2f_{2,1}$. The coupled-mode equations for this $2 \times 2$ block can therefore be written in matrix notation as
\begin{equation}
i(\textbf{P}\dot{\textbf{C}}+\frac{1}{2}\dot{\textbf{P}}\textbf{C})=\textbf{T}\textbf{C},
\label{MatForm}
\end{equation}
where the matrices $\textbf{P}$ and $\textbf{T}$ have the form
\begin{equation}
\textbf{P}=
\begin{pmatrix}
1 & X(z) \\
X(z) & 1\\
\end{pmatrix}; \quad \textbf{T}=
\begin{pmatrix}
0 & \kappa_{12}(z) \\
\kappa_{12}(z) & 0\\
\label{PMat}
\end{pmatrix}.
\end{equation}
Here $X(z)=p_{1,2}(z)$ and $\kappa_{12}(z)=t_{1,2}(z)$.
To proceed further we introduce a new set of variables by $\textbf{C}=\textbf{M}\textbf{W}$, where $\textbf{M}$ is chosen such that $\textbf{M}^{\dagger}\textbf{P}\textbf{M}=\textbf{I}$. This transformation is always possible as both $\textbf{I}$ and $\textbf{P}$ are both Hermitian, positive definite matrices and so can be connected via a similarity transformation by a non-singular matrix~\cite{Matrix}. This change of variable matrix can be written as
\begin{equation}
\textbf{M}=\frac{1}{\sqrt{2}}
\begin{pmatrix}
\frac{1}{\sqrt{1+X(z)}} & -\frac{1}{\sqrt{1-X(z)}} \\
\frac{1}{\sqrt{1+X(z)}} & \frac{1}{\sqrt{1-X(z)}}\\
\end{pmatrix}.
\label{MMat}
\end{equation}
Using this change of variables and multiplying from the left by $\textbf{M}^{\dagger}$ allows Eq.~\eqref{MatForm} to be rewritten as
\begin{equation}
i\dot{\textbf{W}}= \textbf{M}^{\dagger}\textbf{T}\textbf{M}\textbf{W},
\notag
\end{equation}
with $\textbf{M}^{\dagger}\textbf{T}\textbf{M}$ taking the diagonal form
\begin{equation}
\textbf{M}^{\dagger}\textbf{T}\textbf{M}=
\begin{pmatrix}
\frac{\kappa_{12}(z)}{1+X(z)} & 0 \\
0 & \frac{\kappa_{12}(z)}{-1+X(z)} \\
\end{pmatrix}.
\notag
\end{equation}
The equation for $\dot{\textbf{W}}$ can be straightforwardly solved to yield
\begin{equation}
\textbf{W}(z)=\textbf{T}(z)\textbf{W}_0,\\
\notag
\end{equation}
where $\textbf{T}$ has the form
\begin{align}
&\textbf{T}(z)=\begin{pmatrix}
\int_0^{z}dz \frac{\kappa_{12}(z)}{1+X(z)} & 0 \\
0 & \int_0^{z}dz \frac{\kappa_{12}(z)}{-1+X(z)}\\
\end{pmatrix}\nonumber\\
&\approx \begin{pmatrix}
\int_0^{z}dz \kappa_{12}(z) & 0 \\
0 & -\int_0^{z}dz \kappa_{12}(z)\\
\end{pmatrix}.
\notag
\end{align}
In the second equality, the cross-power term, $X(z)$, has been neglected, which is a valid approximation as long as the waveguides remain far apart throughout the bending motion. \\
A return to the original $\textbf{C}$ variables can be made by inverting the non-singular matrix $\textbf{M}$,
\begin{equation}
\textbf{C}(z)= \textbf{M}(z)\textbf{T}(z)\textbf{M}^{-1}(0)\textbf{C}_0,
\notag
\end{equation}
where $\textbf{C}_0$ are the initial conditions.
The waveguides in the experimental lattice are coming together and then moving apart.
Consequently, at the end of the bending section, $z=L$, the waveguides are at the same separation as at $z=0$ and so $\textbf{M}(L)=\textbf{M}(0)$. This allows the matrix product $\textbf{M}(L)\textbf{U}(L)\textbf{M}^{-1}(0)$ to take the form
\begin{equation}
\textbf{M}(L)\textbf{T}(L)\textbf{M}^{-1}(0)=\textbf{U}(L)=
\begin{pmatrix}
\cos(\phi) & -i\sin(\phi) \\
-i\sin(\phi) & \cos(\phi) \\
\end{pmatrix},
\label{Evo}
\end{equation}
with
\begin{equation}
\phi=\int_0^{L} dz \kappa_{12}(z).
\notag
\end{equation}
For comparison, the evolution operator for two straight waveguides of length $L$ is given by
\begin{equation}
\textbf{U}_{Str}(L)=
\begin{pmatrix}
\cos(JL) & -i\sin(JL) \\
-i\sin(JL) & \cos(JL) \\
\end{pmatrix},
\notag
\end{equation}
where $J$ is the coupling between the waveguides.
As can be readily observed, this has the same form as Eq.~\eqref{Evo} if we write $\phi$ as $\phi=LJ_\text{eff}$ where $J_\text{eff}$ is an effective coupling.
This analysis has been conducted for two waveguides ,but can be readily extended to the full lattice discussed in the main text. The experimental lattice is written such that only pairs of waveguides are coupled at any given moment. The $\textbf{P}$ and $\textbf{M}$ matrices therefore have block-diagonal matrix form with the blocks all having the form of Eq.~\eqref{PMat}.
\subsection*{III. Method to compute the topological phase diagram}
In this section the methodology used to obtain the topological phase diagram in Fig.~$\ref{fig2}$ (main text) is detailed; this method is based on Ref.~\cite{nathan2015topological}, and we refer the reader to this reference for more details on concepts related to the topological characterisation of Floquet systems. The calculation of the different phases is assisted by utilizing the property that the topology of the system can only change when there is a gap closing between the two bulk bands. The position of these gap-closing events can be found analytically by diagonalizing the evolution operator at the end of the driving period. It is found that for $\Lambda_1\!=\!\Lambda_2$ and $\Lambda_2\!=\!\frac{1}{3}(2\pi-\Lambda_1)$ the system is gapless at quasienergy zero, whilst for $\Lambda_2=\frac{1}{3}(\pi-\Lambda_1)$ and $\Lambda_2=\frac{1}{3}(3\pi-\Lambda_1)$ the system is gapless through the fundamental zone edge. The position of these gap closings thereby divides the phase space into the eight different sectors shown in Fig.~$\ref{fig2}$. The topology within these different sectors can then be defined by calculation of the winding numbers. As discussed in Eq.~\eqref{Proj}, at any time $t$ the evolution operator may be diagonalized to yield the instantaneous Bloch bands of the driven system. The eigenstates associated with these bands can be used to calculate an instantaneous Chern number for that band. In accordance with Eq.~\eqref{Winding} the winding number can then be calculated by tracking changes in the Chern number of the lowest band that occur through the zone edge throughout the driving period. This procedure is illustrated in Fig.~$\ref{Instant}$ (a) for the parameters $\Lambda_1=1$, $\Lambda_2=1.4$, which shows how the instantaneous Chern number of the lowest band changes throughout the driving period. It can be readily observed that that the Chern number changes twice within a driving period and computing the spectra at these times shows that the second of these changes occurs through the zone edge, see Fig.~$\ref{Instant}$~(b-c). This latter degeneracy causes the Chern number of the lowest band to decrease by one. This topological transition through the zone edge when combined with the Chern numbers of all the bands being zero at $t=T$ implies, in accordance with Eq.~\eqref{Winding}, that both $W_1$ and $W_2$ are equal to one.
\begin{figure}[h!]
\includegraphics[width=8.5cm]{Fig-S1.png}
\caption{{\bf Example calculation of the winding number using the evolution of the instantaneous Chern number.} (a) Evolution of the instantaneous Chern number of the lowest band, $C_1$ over one driving period; the system parameters are $\Lambda_1\!=\!1$ and $\Lambda_2\!=\!1.4$. The Chern number changes twice within a driving period (at times $t_1$ and $t_2$). The first of these topological transitions occurs when the gap closes at $\phi\!=\!0$, (b), whilst the second occurs at $\phi\!=\!\pi$, (c). This changing of the instantaneous Chern number that occurs through the zone edge is responsible for the winding numbers $W_1$ and $W_2$ being non-zero, and the Chern number of the Floquet bands to be zero. This non-trivial value for the winding number allows for the presence of edge modes in the spectrum at $t=T$, (d), even though the Chern numbers of all the Floquet bands (evaluated at $t\!=\!T$) are zero.}
\label{Instant}
\end{figure}
\subsection*{IV. Wavelength Tuning and Disorder}
In the main text it was discussed how the experimental setup provides the ability to modify the wavelength of light used as an input. This ability allows the effective coupling between waveguides to be altered without modifying the lattice. In order to understand the behaviour of the bonds $J_{2-4}$, see main text, as a function of wavelength, five sets of isolated bonds were written and the percentage of light transferred from the launch waveguide as a function of wavelength was measured, Fig.~\ref{Coupling}. This bond characterisation data illustrates how at 785~nm the bonds $J_{2-4}$ have almost equal transfer which is close to 100$\%$. This wavelength therefore closely matches the desired parameters discussed in the main text of having three bonds equal and perfectly transferring.
\begin{figure}[h!]
\includegraphics[width=8.5cm]{Fig-S2.png}
\caption{{\bf{Variation of transfer for bonds $J_{2-4}$ as a function of wavelength.}}
The effective coupling can be extracted from this characterisation data using the relation Transfer $=\sin^2(J_\text{eff}L)$, where
$L$ is the physical length of the sample. The dotted line indicates that the transfer is very close to 100$\%$ for all the three bonds at 785~nm wavelength.}
\label{Coupling}
\end{figure}
The capability of this wavelength tuning technique in our experimental setup allows for the transfer of the bonds to be changed substantially from $\approx100 \%$ right down to $\approx50\%$. This provides the ability to compare experimental and theoretical predictions for a wide range of parameters. In order to perform such a comparison, the coupling strengths extracted from the bond characterisation data were used to calculate a theoretical centre of mass drift, $\textbf{r}_{cm}$, that could be compared to experimental results. The centre of mass drift is defined as
\begin{equation}
\textbf{r}_{cm}=\sum_{m,n} I_{m,n}\textbf{r}_{m,n}-\textbf{r}_0,
\label{EqnCM}
\end{equation}
where $I_{m,n}$ is the intensity in the $(m, n)$-th site at the output, $\textbf{r}_{m,n}$ is the coordinates of the $(m, n)$-th site and $\textbf{r}_0$ are the coordinates of the launch site. Experimental data was taken at four different physical locations in the lattice. These different positions should, in the absence of disorder, produce identical results for $\textbf{r}_{cm}$, up to a sign change. However, the non-zero standard deviations measured in the bond characterisation data indicates the presence of bond strength disorder within the lattice (such a disorder will be referred to as ``off-diagonal disorder"). The effects of this disorder can be directly observed in the experimental lattice with the four different launch sites producing slightly different results for $\textbf{r}_{cm}$, see the error bars in Fig.~\ref{CM}(a). In order to theoretically model this disorder, the strengths of the bonds $J_{2-4}$ were randomly selected to lie within the range measured in the bond characterisation data. The comparison between the
experimental data and the theoretical prediction shows good agreement around $\lambda\!\approx\!785$~nm, but deviations are observed at smaller wavelengths, Fig.~\ref{CM}(a). The failure of the theoretical model in this low transfer region is further illustrated in Fig.~\ref{CM}(C$_{1-2}$), which compares the averaged experimental output image to the theoretical image averaged over 1000 disorder realisations.
The match between theory and experimental results can be improved by including a small on-site disorder term, $\Delta \beta_{m,n}$, in addition to the coupling disorder of the previous model (this form of disorder will be referred to as ``diagonal disorder"). The $\Delta \beta_{m,n}$ for the $(m,n)$-{th} site is a random number drawn from within a uniform distribution covering the interval $[-W,\!\quad\!W]$. The parameter $W$ is the disorder strength which is chosen as $WT/4=0.6$, as this particular value results in the best fit to the experimental data over the whole wavelength range investigated. The corresponding center-of-mass prediction, Fig.~\ref{CM}(b), and output facet image, Fig.~\ref{CM}(C$_{3}$), produced by this model are in closer agreement to the experimental results which points towards a disordering of this type being present in the lattice.
\begin{figure}[h!]
\includegraphics[width=8.6cm]{Fig-S3.png}
\caption{\textbf{Comparing experimental results and theoretical predictions from two different disorder models.} Model 1 has bond strength disorder only whilst model 2 has additional phase disorder. (a) and (b) compare the experimental centre of mass drifts, as defined by Eq.~\eqref{EqnCM}, to those calculated theoretically for models 1 and 2, respectively. The centre of mass drift for four different launch sites is measured experimentally and the red and blue circles indicate the average drift in the $x$ and $y$ directions whilst the black error bars indicate the range of values measured experimentally. The shaded regions in (a) and (b) indicate the theoretical disorder averaged value plus/minus one standard deviation. (C$_1$) is the experimental output image averaged over the 4 launch sites for 705~nm input whilst (C$_2$) and (C$_3$) are the disordered averaged output facet images coming from models 1 and 2 respectively.}
\label{CM}
\end{figure}
\subsection*{V. Study of Disorder and its impact on topological bands}
The analysis conducted in the previous section illustrated the presence of off-diagonal disorder as well as possibly indicating the presence of diagonal disorder terms. In the absence of any disorder, the experimental lattice should reproduce a driven-lattice model that exhibits anomalous topological edge modes (i.e.~chiral edge modes that are associated with non-zero winding numbers, while the Chern number of the Floquet bands is trivial; see main text). The inclusion of disorder in the system can allow bandgap closings to occur which can modify the topology of the system. In particular, the disorder could potentially change the Chern numbers of the Floquet bands to a non-trivial value such that the experimentally observed edge modes would no longer be ``anomalous". The elimination of this possibility thereby requires the calculation of the Chern number in a disordered system.
The presence of disorder means that quasimomentum is no longer a good quantum number which in turn precludes the use of the usual momentum space techniques in calculating the Chern number. In order to overcome this limitation we use a real-space Chern invariant, $C(\epsilon_0)$, as introduced by Bianco and Resta in Ref. \cite{Resta1}. This invariant allows for a generalisation of the ({\textit{static}}) bulk-edge correspondence to disordered and quasi-periodic lattices~\cite{Resta1,DucThanh}. In particular, when the reference energy, $\epsilon_0$, is placed in a mobility gap of the spectrum the Chern invariant will measure the number of chiral edge modes which cross this energy, as predicted by the Chern numbers of the bands below this energy. This real-space Chern invariant is based around an operator called the Chern marker, $\hat{\mathcal{C}}$, which is defined by
\begin{equation}
\hat{\mathcal{C}}=-4\pi \text{Im}[\hat{x}_Q\hat{y}_P].
\end{equation}
Here the operators $\hat{\textbf{r}}_{P}=\hat{P}\hat{\textbf{r}}\hat{Q}$ and $\hat{\textbf{r}}_{Q}=\hat{Q}\hat{\textbf{r}}\hat{P}$
are expressed in terms of the position operator $\hat{\textbf{r}}=(\hat{x},\hat{y})$ and the projection operators
\begin{equation}
\hat{P}(\epsilon_0)=\sum_{\epsilon\leq \epsilon_0} |\psi_{\epsilon} \rangle \langle \psi_{\epsilon}|=\hat{1}-\hat{Q}.
\end{equation}
In the absence of disorder a real space Chern Number can be defined by averaging the Chern marker operator over a unit cell,
\begin{equation}
C=\int_{cell} \langle \textbf{r}| \hat{\mathcal{C}}|\textbf{r}\rangle d\textbf{r}=\int_{cell} \mathcal{C}(\textbf{r}) d\textbf{r}.
\end{equation}
\begin{figure}[t!]
\includegraphics[width=8.6cm]{Fig-S4.png}
\caption{\textbf {The local Chern marker for the experimental lattice with the addition of both diagonal and off-diagonal disorder terms.} The figure shows the probability of occurrence of a particular value for $\mathcal{C}(\textbf{r})$ anywhere within the bulk of the lattice. This local Chern marker is narrowly peaked around zero which implies that the static bulk-edge correspondence would predict no robust chiral edge modes present in the system.}
\label{Chern}
\end{figure}
In the presence of disorder this real-space Chern number fluctuates depending upon the position of the unit cell. These fluctuations can be accounted for by replacing the unit-cell average by an average over an area $A$ that is located within the bulk and is large compared to the fluctuation length~\cite{DucThanh}.
These ideas thereby provide a mechanism to investigate whether the disorder terms discussed in the previous section can alter the Chern numbers of the Floquet bands to non-trivial values. In particular we calculate the local Chern marker, in the presence of these disorder terms, for a reference energy of zero, $\epsilon_0=0$. The real-space average of this local Chern marker, the Chern invariant, can be viewed as the prediction coming from the static bulk-edge correspondence for the number of chiral edge modes that are traversing the mobility gap between the two bands. The results, however, of this analysis reveal that even in the presence of disorder the local Chern marker, $\mathcal{C}(\textbf{r},0)$, is still heavily peaked around zero with this local quantity fluctuating only slightly within the bulk, Fig.~$\ref{Chern}$. The Chern invariant is therefore equal to zero independently of the size of the averaging area that is chosen. Consequently, the static bulk-edge correspondence predicts the experimental lattice should not feature any chiral edge modes.
\end{document}
|
\section{Introduction}
Throughout this paper we let $\HC$ denote a separable Hilbert space with inner product $\left\langle\cdot,\cdot\right\rangle_\HC$. Unless we explicitly state otherwise, we assume that $\HC$ is infinite-dimensional. We denote by $\LC=\LC\left(\HC\right)$ the space of bounded linear transformations on $\HC$, by $\SC^1$ the corresponding trace class, and by $\SC^2$ the Hilbert--Schmidt class. $\XC$ will be used as a generic notation for an element of the set $\left\{\HC,\LC,\SC^1,\SC^2\right\}$.
We will use $\YC$ to denote a general Banach space. By $\Hol\left(\YC\right)$ we denote the space of $\YC$-valued analytic functions on the open unit disc $\D$. For $f\in\Hol(\YC)$, we denote the $n$th Taylor coefficients at the origin by $\hat f(n)$. We denote by $\OC\left(\YC\right)$ the space of functions in $\Hol\left(\YC\right)$ that admit an analytic extension to a larger disc (centered at the origin). If $\YC=\C$, then we suppress this in our notation, i.e. $\Hol=\Hol\left(\C\right)$, and $\OC=\OC\left(\C\right)$. The same principle will apply to all function spaces discussed below.
For $p\in\left[1,\infty\right]$ and $\XC\in\left\{\HC,\SC^1\right\}$, we let $L^p\left(\T,\XC\right)$ denote the standard space of $p$-Bochner--Lebesgue integrable functions from $\T$ to $\XC$. Here $\T$ denotes the unit circle in $\C$. Similarly, we define $L^p\left(\T,\LC\right)$ as the natural WOT-analogue of $L^p(\T)$: A function $f:\T\to\LC$ belongs to $L^p\left(\T,\LC\right)$ if and only if for all $x,y\in\HC$ the function $\left\langle f\left(\cdot\right) x,y\right\rangle_\HC$ is measurable and, moreover, $\left\|f\right\|_{L^p\left(\T,\LC\right)}^p = \int_{\T}\left\|f\right\|_\LC^p\, dm<\infty$. Here $m$ denotes normalized Lebesgue measure on $\T$.
The Hardy space $H^p\left(\XC\right)$ is the space of $f\in\Hol\left(\XC\right)$ such that
\begin{equation}\label{Eq:Hp-norm}
\left\|f\right\|_{H^p\left(\XC\right)}^p=\sup_{0<r<1}\left\|f_r\right\|_{L^p\left(\T,\XC\right)}<\infty,
\end{equation}
where we have defined the function $f_r:z\mapsto f\left(rz\right)$. An important property of Hardy space functions is that they have boundary values in a natural sense, cf. Proposition \ref{Proposition:BoundaryIdentification}. We denote the boundary values of $f\in H^p\left(\XC\right)$ by $bf\in L^p(\T,\XC)$.
The space $H^2\left(\HC\right)$ is a Hilbert space, with inner product $\left\langle f,g\right\rangle = \sum_{0}^\infty \langle \hat f(n),\hat g(n)\rangle_\HC$. Of particular importance will be the set of $H^2\left(\HC\right)$-normalized functions in $\OC\left(\HC\right)$, which we denote by $\OC_1\left(\HC\right)$.
We now introduce the main topics of this paper. Initially, we consider the scalar setting, rather than the proper vectorial one.
\subsection{Hankel operators}
Given $\phi\in\Hol$ and $f\in\OC$, we define the action of the Hankel operator $\Gamma_\phi$ on $f$ by
\begin{equation}\label{Eq:HardyHankelFormula}
\Gamma_\phi f\left(z\right)=\sum_{n=0}^\infty \left(\sum_{m=0}^\infty\hat \phi\left(m+n\right)\hat f\left(m\right) \right)z^n,\quad z\in\D.
\end{equation}
A standard reference on Hankel operators is \cite{Peller2003:HankOpsBook}. We refer to $\phi$ as the symbol of $\Gamma_\phi$. We say that $\Gamma_\phi$ is bounded if it extends to a bounded operator on $H^2$.
For $\Gamma_\phi$ to be bounded it is necessary for $\phi$ to be in $H^2$. For $\phi\in H^2$, one shows by computation that $\Gamma_\phi f=P_+ \left(\phi \conjvar{f} \right)$, where $P_+$ denotes the orthogonal projection from $L^2\left(\T\right)$ onto $H^2$, and $\conjvar{ f}:z\mapsto f\left(\conj{ z}\right)$.
It is convenient to define the operation of coefficient conjugation, $f\mapsto f^\#$, $f^\# (z)=\conj{f(\conj z)}$. Note that this is an isomorphism on $H^2$. A classical result is that $H^1=H^2\cdot H^2$: If $f,g\in H^2$, then $f\cdot g\in H^1$, and $\left\|h\right\|_{H^1}\le \left\|f\right\|_{H^2}\left\|g\right\|_{H^2}$. Conversely, if $h\in H^1$, then there exists $f,g\in H^2$ such that $h=f\cdot g$ and $\left\|f\right\|_{H^2}\left\|g\right\|_{H^2}\le C\left\|h\right\|_{H^1}$, where $C>0$ is a constant independent of $f$ and $g$. Now choose $f$ so that $f^\# g = h$. By the calculation
\[
\left\langle\Gamma_\phi f,g\right\rangle = \left\langle P_+ \left(\phi\conjvar{f} \right),g\right\rangle= \left\langle \phi\conjvar{f},g\right\rangle= \left\langle\phi,f^\#g\right\rangle = \left\langle\phi,h\right\rangle,
\]
one obtains that $\Gamma_\phi$ is bounded if and only if $\phi\in\left(H^1\right)^*$.
Since $H^1$ may be identified with a subspace of $L^1(\T)$, and $\left(L^1\left(\T\right)\right)^*=L^\infty\left(\T\right)$, a straightforward application of the Hahn--Banach theorem shows that $\left(H^1\right)^*=P_+L^\infty\left(\T\right)$. The fact that $\Gamma_\phi$ is bounded if and only if $\phi\in P_+L^\infty\left(\T\right)$ is known as Nehari's theorem \cite{Nehari1957:BddBilinFrms}.
\subsection{Carleson embeddings}
Every Borel measure $\mu\ge 0$ on $\D$ corresponds to a so-called Carleson embedding $H^2\hookrightarrow L^2(\D,d\mu)$. It is a classical result \citelist{\cite{Carleson1958:InterpolProblBddAnalFcns}\cite{Carleson1962:InterpolBddAnalFcnsCoronaProbl}} in complex and harmonic analysis that boundedness of such embeddings can be characterized by a simple geometric property of $\mu$. Specifically, the Carleson embedding condition
\begin{equation}
\sup_{f\in\OC_1\left(\HC\right)} \int_\D \left|f\left(z\right) \right|^2 \, d\mu\left(z\right)<\infty
\end{equation}
holds if and only if $\mu$ satisfies the so-called Carleson intensity condition
\begin{equation}\label{Eq:CarlesonInt}
\sup_{\substack{I\subset\T\\ I\textnormal{ arc}}}\frac{\mu\left(\left\{w\in\D;\, 1-m(I)<\left|w\right|<1,\, \frac{w}{\left|w\right|}\in I\right\}\right)}{m\left(I\right)}<\infty.
\end{equation}
\subsection{Bounded mean oscillation}
A bridge connecting Hankel operators, and Carleson embeddings is given by $\textrm{BMOA}$; bounded mean oscillation of analytic functions. Suppose that $\phi\in H^1$. We then say $\phi$ belongs to the class $\textrm{BMOA}$ if and only if
\[
\left\|\phi\right\|_*=\sup_{\substack{I\subset\T\\ I\textnormal{ arc}}}\frac{1}{m\left(I\right)}\int_I \left|b\phi-\left(b\phi\right)_I \right|\, dm<\infty.
\]
Here $\left(b\phi\right)_I$ denotes the Lebesgue integral average $\frac{1}{m\left(I\right)}\int_I b\phi\, dm$. The quantity $\left\|\cdot\right\|_*$ is a semi-norm. The class $\textrm{BMOA}$ becomes a Banach space when equipped with the norm $\left\|\phi\right\|_{\textrm{BMOA}}=\left|\phi\left(0\right)\right|+\left\|\phi\right\|_*$.
A celebrated result by Fefferman \citelist{\cite{Fefferman1971:CharBMO}\cite{Fefferman-Stein1972:HpSpaces}} is that $\textrm{BMOA}$ is in fact the dual of $H^1$. Moreover, $\phi\in \textrm{BMOA}$ if and only if the measure $\mu$ given by $d\mu = \left|\phi'(z)\right|^2\left(1-\left|z\right|^2\right)\, dA(z)$ satisfies \eqref{Eq:CarlesonInt}. As a summary of this discussion we have:
\begin{proposition}
Let $\phi\in H^1$. Then the following are equivalent:
\begin{itemize}
\item[$(i)$] $\Gamma_\phi$ is $H^2$-bounded.
\item[$(ii)$] $\phi\in\left(H^1\right)^*$.
\item[$(iii)$] $\phi\in P_+L^\infty\left(\T\right)$.
\item[$(iv)$] $\phi\in \textrm{BMOA}$.
\item[$(v)$] The measure given by $d\mu = \left|\phi'(z)\right|^2\left(1-\left|z\right|^2\right)\, dA(z)$ has finite Carleson intensity.
\item[$(vi)$] The Carleson embedding $H^2\hookrightarrow L^2\left(\D,\left|\phi'(z)\right|^2\left(1-\left|z\right|^2\right)\, dA\left(z\right)\right)$ is bounded.
\end{itemize}
\end{proposition}
\subsection{The vectorial setting}
Note that \eqref{Eq:HardyHankelFormula} makes perfect sense if $\phi\in\Hol(\LC)$ and $f\in\OC(\HC)$. We take this as the definition of a vectorial Hankel operator $\Gamma_\phi$. The factorization result $H^1\left(\SC^1\right)=H^2\left(\SC^2\right)\cdot H^2\left(\SC^2\right)$, due to Sarason \cite{Sarason1967:GenInterpol}, implies that $\Gamma_\phi$ is $H^2(\HC)$-bounded if and only if $\phi\in \left(H^1\left(\SC^1\right)\right)^*$, very much like in the scalar setting.
Since $\left(L^1\left(\T,\SC^1\right)\right)^*$ is not equal to $L^\infty(\T,\LC)$ ($\LC$ does not have the so-called Radon--Nikodym property, e.g. \cite{Diestel-Uhl1977:VecMeasures}), it is not obvious that $\left(L^1\left(\T,\SC^1\right)\right)^*=P_+L^\infty\left(\T,\LC\right)$. However, this follows from a vectorial extension of Nehari's theorem, due to Page \cite{Page1970:BddCompctVecHankOps}: $\Gamma_\phi$ is $H^2\left(\HC\right)$-bounded if and only if $\phi\in P_+L^\infty\left(\T,\LC\right)$.
The space of $\LC$-valued analytic functions for which the corresponding Hankel operators are $H^2\left(\HC\right)$-bounded is commonly referred to as \emph{Nehari--Page} $\textrm{BMOA}$:
\begin{definition}\label{def:NP}
Let $\phi\in\Hol\left(\LC\right)$. We then say that $\phi\in \textrm{BMOA}_{\NC\PC}\left(\LC\right)$ if and only if
\[
\left\|\phi\right\|_{\textrm{BMOA}_{\NC\PC}}=\sup_{f\in\OC_1\left(\HC\right)}\left\|\Gamma_\phi f\right\|_{H^2\left(\HC\right)}<\infty.
\]
\end{definition}
While $\textrm{BMOA}_{\NC\PC}\left(\LC\right)$ can be identified either with $P_+L^\infty\left(\T,\LC\right)$ or with $\left(H^1\left(\SC^1\right)\right)^*$, these characterizations are of an abstract nature. Finding concrete conditions that characterize $\textrm{BMOA}_{\NC\PC}\left(\LC\right)$ has proven to be notoriously difficult. For example, define the class $\textrm{BMOA}_\OC\left(\LC\right)$ as the class of $\phi\in H^1\left(\LC\right)$ such that the oscillation condition
\[
\left\|\phi\right\|_*=\sup_{\substack{I\subset\T\\ I\textnormal{ arc}}}\frac{1}{m\left(I\right)}\int_I\left\|b\phi-\left(b\phi\right)_I\right\|_\XC\, dm<\infty
\]
holds. Then
\[
\textrm{BMOA}_\OC\left(\LC\right)\subsetneq \textrm{BMOA}_{\NC\PC}\left(\LC\right).
\]
This fact represents an area of research, where authors consider some aspect of the theory for scalar-valued $\textrm{BMOA}$ (or its harmonic or dyadic analogues), and then discuss to what extent this aspect carries over to the vector-valued case, e.g. \citelist{\cite{Blasco1988:HardySpacesVecValDuality}\cite{Blasco-Pott2008:EmbOpValDyadicBMO}\cite{Bourgain1986:VecValSingIntsHardy-BMODualityChapter}\cite{Gillespie-Pott-Treil-Volberg2004:LogGrowthHilbTransfVecHank}\cite{Mei2006:MatValParaprods}\cite{Nazarov-Pisier-Treil-Volberg2002:EstsVecCarlesonEmbThmVecParaprods}\cite{Nazarov-Treil-Volberg1997:CounterExInfDimCarlesonEmbThm}}.
Before we get to the meat of this paper, we define the differentiation operator $D:\Hol\left(\YC\right)\to\Hol\left(\YC\right)$ by $Df\left(z\right)=zf'\left(z\right)+f\left(z\right)$. With respect to the monomial basis, $D$ acts like a diagonal matrix. This presents an elementary way of taking arbitrary powers of $D$: For $\alpha\in\R$, we set
\[
D^\alpha f\left(z\right)=\sum_{n=0}^\infty \left(1+n\right)^\alpha \hat f\left(n\right) z^n,\quad z\in\D.
\]
Another convenience of working with $D$ in place of ordinary differentiation is that it does not annihilate constants. In fact we can say more: For each $\alpha\in\R$, $D^\alpha:\Hol\left(\YC\right)\to\Hol\left(\YC\right)$ is a bijection that leaves $\OC\left(\YC\right)$ invariant.
From a technical point of view, the present paper is mainly concerned with $H^2(\HC)$-boundedness of operators of the type $D^\alpha\Gamma_\phi$, with $\alpha>0$ and $\phi\in\Hol\left(\LC\right)$. The present paper was originally motivated by the natural appearance of such operators in control theory, e.g. \cite{Jacob-Rydhe-Wynn2014:WeightWeissConjRKTGenHankOps}. However, they also have implications to our understanding of $\textrm{BMOA}_{\NC\PC}\left(\LC\right)$. Our investigation motivates the definition of a class which we refer to as \emph{Carleson} $\textrm{BMOA}$:
\begin{definition}\label{def:C}
Let $\phi\in\Hol\left(\LC\right)$. We then say that $\phi\in \textrm{BMOA}_{\CC}\left(\LC\right)$ if and only if
\begin{equation}\label{Eq:AAC}
\left\|\phi\right\|_{\textrm{BMOA}_\CC}^2=\sup_{f\in\OC_1\left(\HC\right)} \int_\D \left\| \left(D\phi\right)\left(z\right) f\left(\conj{ z}\right)\right\|_\HC^2\left(1-\left|z\right|^2\right)\, dA\left(z\right)<\infty.
\end{equation}
\end{definition}
Since $D$ does not annihilate constants, $\left\|\cdot\right\|_{\textrm{BMOA}_\CC}$ is a proper norm, and not a semi-norm.
\subsection{Main result and corollaries}
\begin{theorem}\label{Theorem:HankelCarleson}
Let $\HC$ be a separable Hilbert space, $\LC$ its space of bounded linear transformations. Let $\alpha>0$ and suppose that $\phi:\D\to\LC$ is analytic. Then $D^\alpha \Gamma_\phi$ is $H^2\left(\HC\right)$-bounded if and only if $D^\alpha \phi\in \textrm{BMOA}_{\CC}\left(\LC\right)$, i.e.
\[
\left\|D^\alpha\Gamma_\phi\right\|_{H^2\left(\HC\right)\to\HC^2\left(\HC\right)}=\sup_{f\in\OC_1\left(\HC\right)}\left\|D^\alpha\Gamma_\phi f\right\|_{H^2\left(\HC\right)}<\infty
\]
if and only if
\[
\left\|D^\alpha\phi\right\|_{\textrm{BMOA}_\CC}=\sup_{f\in\OC_1\left(\HC\right)}\int_\D \left\| \left(D^{1+\alpha}\phi\right)\left(z\right) f\left(\conj{ z}\right)\right\|_\HC^2\left(1-\left|z\right|^2\right)dA\left(z\right)<\infty.
\]
Moreover,
\[
\left\|D^\alpha\Gamma_\phi\right\|_{H^2\left(\HC\right)\to\HC^2\left(\HC\right)}\approx \left\|D^\alpha\phi\right\|_{\textrm{BMOA}_\CC}.
\]
\end{theorem}
Theorem \ref{Theorem:HankelCarleson} generalizes a result by Janson and Peetre \cite{Janson-Peetre1988:Paracomms} who obtained essentially the above characterization in the case where $\HC=\C$. We point out that, in the case where $\phi$ is $\LC$-valued, we are forced to avoid the Schur multiplier techniques used in \cite{Janson-Peetre1988:Paracomms}. This is made evident by the discussion in \cite{Davidson-Paulsen1997:PolBddOps}*{Section 4}.
Operators of the type $D\Gamma_\phi$ received a lot of attention in connection to the so called Halmos problem \cite{Halmos1970:TenProbls}*{Problem 6}:
\begin{quote}
If a Hilbert space operator is similar to a Hilbert space contraction, then it is also polynomially bounded (by von Neumann's inequality). Is the converse true?
\end{quote} Following the works of many authors \citelist{\cite{Aleksandrov-Peller1996:HankOpsSimToContr}\cite{Bourgain1986:SimProblPolBddOpsHSpace}\cite{Foguel1964:CounterExSz.-NagyProbl}\cite{Paulsen1984:ComplPolBddSimContr}\cite{Peller1982:EstsFcnsPwrBddOpsHSpace}\cite{Sz.-Nagy1959:ComplContOpsUniformlyBddIterates}}, Pisier \cite{Pisier1997:PolBddNotSim} answered this question in the negative. Subsequently, different proofs of the same result have been given in several papers \citelist{\cite{Davidson-Paulsen1997:PolBddOps}\cite{Kislyakov2000:OpsDisSimContr}}. All of these proofs exploit boundedness properties of operators of the type $D\Gamma_\phi$. The following two propositions are essentially from Davidson and Paulsen \cite{Davidson-Paulsen1997:PolBddOps}:
\begin{proposition}\label{Proposition:Davidson-Paulsen1}
Let $\alpha>0$, and $\HC$ be a separable, infinite-dimensional Hilbert space, $\LC$ its space of bounded linear transformations. Then there exists an analytic function $\phi:\D\to \LC$ such that $D^\alpha\Gamma_\phi$ is bounded on $H^2\left(\HC\right)$, while $\Gamma_\phi D^\alpha$ is not. Moreover, $\phi$ may be chosen to be rank one-valued.
\end{proposition}
\begin{proposition}\label{Proposition:Davidson-Paulsen2}
Let $\alpha>0$, and $\HC$ be a separable, infinite dimensional Hilbert space, $\LC$ its space of bounded linear transformations. Then there exists a bounded analytic function $\phi:\D\to \LC$ such that $D^\alpha\Gamma_{D^{-\alpha}\phi}$ is not bounded on $H^2\left(\HC\right)$.
\end{proposition}
\begin{remark}\label{Remark:Davidson-Paulsen}
Proposition \ref{Proposition:Davidson-Paulsen1} is stated for $\alpha=1$ in \cite{Davidson-Paulsen1997:PolBddOps}*{Example 4.6}. Proposition \ref{Proposition:Davidson-Paulsen2} is essentially stated for $\alpha=1$ in \cite{Davidson-Paulsen1997:PolBddOps}*{Corollary 4.2}, but this does not explicitly mention the boundedness of $\phi$, even though it follows from the original proof. A dyadic analogue of this result has been proved by Mei \cite{Mei2006:MatValParaprods}. For the convenience of the reader, we present proofs of the above propositions in Section \ref{Sec:Davidson-Paulsen}.
\end{remark}
Combining the results by Davidson and Paulsen with Theorem \ref{Theorem:HankelCarleson}, we are able to derive several interesting results.
Given $\phi\in\Hol\left(\LC\right)$, we define the function $\phi^\#:z\mapsto \phi\left(\conj{ z}\right)^*$. This is the function obtained by taking the Hilbert space conjugate of each Taylor coefficient of $\phi$. Note that $\Gamma_\phi D=\left(D\Gamma_{\phi^\#}\right)^*$. By Proposition \ref{Proposition:Davidson-Paulsen1} and Theorem \ref{Theorem:HankelCarleson}, it follows that $\textrm{BMOA}_{\CC}\left(\LC\right)$ is not closed under coefficient conjugation (cf. \cite{Aleman-Perfekt2012:HankFrmsEmbThmsDirichletSpaces}*{Proposition 3.3}). On the other hand, $\textrm{BMOA}_{\NC\PC}\left(\LC\right)$ is obviously closed under coefficient conjugation. We obtain the following corollary:
\begin{corollary}\label{Corollary:CNP}
Let $\HC$ be a separable infinite-dimensional Hilbert space, $\LC$ its space of bounded linear transformations. Then $\textrm{BMOA}_\CC\left(\LC\right)$ is not closed under the map $\phi\mapsto\phi^\#$, where $\phi^\#\left(z\right)=\phi\left(\conj{ z}\right)^*$. In particular
\[
\textrm{BMOA}_{\CC}\left(\LC\right)\ne \textrm{BMOA}_{\NC\PC}\left(\LC\right),
\]
i.e. $H^2\left(\HC\right)$-boundedness of $\Gamma_\phi$ is not characterized by the anti-analytic Carleson embedding condition indicated by Theorem \ref{Theorem:HankelCarleson}.
\end{corollary}
Corollary \ref{Corollary:CNP} motivates the following definition:
\begin{definition}\label{def:CS}
Let $\phi\in\Hol\left(\LC\right)$. We then say that $\phi\in \textrm{BMOA}_{\CC^\#}\left(\LC\right)$ if and only if $\phi^\#\in \textrm{BMOA}_{\CC}\left(\LC\right)$.
\end{definition}
Consider now the relation
\begin{align}\label{Eq:LeibnizDecomposition}
\Gamma_{D\phi}
&=
D\Gamma_{\phi}+\left(D\Gamma_{\phi^\#}\right)^*-\Gamma_\phi,
\end{align}
which is obtained by duality, and the Leibniz rule for $D$. The operator $\Gamma_\phi$ is bounded on $H^2\left(\HC\right)$, whenever any of the other terms in \eqref{Eq:LeibnizDecomposition} is bounded, since then $D\phi$ is a Bloch function (cf. Lemma \ref{Lemma:HankelCarlesonBloch} below). In the light of Theorem \ref{Theorem:HankelCarleson}, it is then clear from \eqref{Eq:LeibnizDecomposition} that
\begin{equation}\label{Eq:CC*NP}
\textrm{BMOA}_{\CC}\left(\LC\right)\cap \textrm{BMOA}_{\CC^\#}\left(\LC\right)\subsetneq \textrm{BMOA}_{\NC\PC}\left(\LC\right).
\end{equation}
We point out that the above inclusion also follows implicitly from the proof of \cite{Nazarov-Pisier-Treil-Volberg2002:EstsVecCarlesonEmbThmVecParaprods}*{Theorem 0.8}. However, we obtain also that the inclusion is strict. To see that this is so, suppose that it is not. This would only be possible if $\textrm{BMOA}_{\NC\PC}\left(\LC\right)$ was contained in $\textrm{BMOA}_{\CC}\left(\LC\right)$. By another application of Theorem \ref{Theorem:HankelCarleson}, this would contradict Proposition \ref{Proposition:Davidson-Paulsen2}. We summarize the above discussion:
\begin{corollary}\label{Corollary:CC*NP}
Let $\HC$ be a separable Hilbert space, $\LC$ its space of bounded linear transformations. If $\phi:\D\to\LC$ is an analytic function such that
\[
\left\|\phi\right\|_{\textrm{BMOA}_\CC}=\sup_{f\in\OC_1\left(\HC\right)}\int_\D \left\| \left(D\phi\right)\left(z\right) f\left(\conj{ z}\right)\right\|_\HC^2\left(1-\left|z\right|^2\right)dA\left(z\right)<\infty,
\]
and
\[
\left\|\phi^\#\right\|_{\textrm{BMOA}_\CC}=\sup_{f\in\OC_1\left(\HC\right)}\int_\D \left\| \left(D\phi\right)\left(\conj{ z}\right)^* f\left(\conj{ z}\right)\right\|_\HC^2\left(1-\left|z\right|^2\right)dA\left(z\right)<\infty,
\]
then
\[
\left\|\Gamma_\phi\right\|_{H^2\left(\HC\right)\to\HC^2\left(\HC\right)}=\sup_{f\in\OC_1\left(\HC\right)}\left\|\Gamma_\phi f\right\|_{H^2\left(\HC\right)}<\infty.
\]
Moreover,
\[
\left\|\Gamma_\phi\right\|_{H^2\left(\HC\right)\to\HC^2\left(\HC\right)}\lesssim \left\|\phi\right\|_{\textrm{BMOA}_\CC}+\left\|\phi^\#\right\|_{\textrm{BMOA}_\CC}.
\]
If $\HC$ is infinite dimensional, then the converse statement does not hold.
\end{corollary}
Condition \eqref{Eq:AAC} states that $H^2\left(\HC\right)$ is continuously embedded into $L^2\left(\D,\HC,d\mu\right)$, where $\mu$ is a certain operator valued measure. It is natural to think of this as an embedding of anti-analytic functions, rather than analytic ones. For this reason, we call \eqref{Eq:AAC} the anti-analytic Carleson embedding, to be distinguished from the analytic one, which is given by the straightforward modification \eqref{Eq:AC} below. In the scalar case it is obvious that these two conditions are equivalent. In the general case, this is no longer obvious. In fact, whether or not the two conditions define the same class of functions was posed as an open question by Nazarov, Treil, and Volberg in \cite{Nazarov-Treil-Volberg1997:CounterExInfDimCarlesonEmbThm}. They later restated the question in a joint paper with Pisier \cite{Nazarov-Pisier-Treil-Volberg2002:EstsVecCarlesonEmbThmVecParaprods}. We answer this question in the negative:
\begin{corollary}
Let $\HC$ be a separable infinite-dimensional Hilbert space, $\LC$ its space of bounded linear transformations. Then there exists a bounded analytic function $\phi:\D\to\LC$ such that
\begin{equation}\label{Eq:AC}
\sup_{f\in\OC_1\left(\HC\right)}\int_\D \left\| \left(D\phi\right)\left(z\right) f\left(z\right)\right\|_\HC^2\left(1-\left|z\right|^2\right)dA\left(z\right)<\infty,
\end{equation}
while
\[\tag{\ref{Eq:AAC}$'$}
\sup_{f\in\OC_1\left(\HC\right)}\int_\D \left\| \left(D\phi\right)\left(z\right) f\left(\conj{ z}\right)\right\|_\HC^2\left(1-\left|z\right|^2\right)dA\left(z\right)=\infty.
\]
\end{corollary}
The proof is as follows: Since $D$ is an isomorphism from $H^2\left(\HC\right)$ to the standard weighted Bergman space $A_1^2\left(\HC\right)$, it follows from the Leibniz rule that \eqref{Eq:AC} is satisfied whenever $\phi$ is bounded. On the other hand, by Proposition \ref{Proposition:Davidson-Paulsen2} and Theorem \ref{Theorem:HankelCarleson}, there exists $\phi\in H^\infty\left(\LC\right)$ that satisfies $\left(\ref{Eq:AAC}'\right)$. \qed
Using standard arguments involving duality, $D^\alpha\Gamma_\phi$ is $H^2\left(\HC\right)$-bounded if and only if $\phi$ is in the dual of the space
\[
D^{-\alpha} \left( \left(D^\alpha H^2\left(\HC\right) \right)\hat \otimes \conj{H^2\left(\HC\right)} \right)=D^{-\alpha} \left( \left(D^\alpha H^2\left(\SC^2\right) \right) \cdot H^2\left(\SC^2\right) \right).
\]
A similar statement holds for boundedness of $D^\alpha\Gamma_{\phi^\#}$. This yields an alternative formulation of Theorem \ref{Theorem:HankelCarleson}:
\begin{corollary}\label{Corollary:Duality}
Let $\HC$ be a separable Hilbert space, $\LC$ its space of bounded linear transformations. If $\alpha>0$, then $\textrm{BMOA}_{\CC}\left(\LC\right)$ is the dual of
\[
D^{-\alpha} \left( \left(D^\alpha H^2\left(\HC\right) \right)\hat \otimes \conj{H^2\left(\HC\right)} \right)=D^{-\alpha} \left( \left(D^\alpha H^2\left(\SC^2\right) \right) \cdot H^2\left(\SC^2\right) \right),
\]
while $\textrm{BMOA}_{\CC^\#}\left(\LC\right)$ is the dual of
\[
D^{-\alpha} \left(H^2\left(\HC\right)\hat \otimes \conj{D^\alpha H^2\left(\HC\right)} \right)=D^{-\alpha} \left(H^2\left(\SC^2\right) \cdot \left(D^\alpha H^2\left(\SC^2\right) \right) \right).
\]
\end{corollary}
We return for a moment to the scalar case. By the square function characterization of $H^1$, due to Fefferman and Stein \cite{Fefferman-Stein1972:HpSpaces}, it follows that
\begin{equation}\label{Eq:Inclusion}
D^{-1} \left( \left(D^1 H^2 \right)\cdot H^2 \right)\subseteq H^1.
\end{equation}
A generalization to general $\alpha>0$, which also yields equality of function spaces in \eqref{Eq:Inclusion}, has been obtained by Cohn and Verbitsky \cite{Cohn-Verbitsky2000:FactTentSpacesHankOps}. By Corollary \ref{Corollary:Duality}, the dual inclusion becomes
\[
\textrm{BMOA}_{\NC\PC}\subseteq \textrm{BMOA}_{\CC}.
\]
Combined with Corollary \ref{Corollary:CC*NP}, this implies the well-known result that $\textrm{BMOA}_{\CC}=\left(H^1\right)^*$. For this argument to work, it suffices to use Theorem \ref{Theorem:HankelCarleson} with (say) $\alpha=1$, a special case which is substantially simpler to prove.
The paper is structured as follows: In Section \ref{Sec:Preliminaries} we fix notation, and review some preliminary material. Of particular importance are some Bergman type spaces of analytic functions. In Section \ref{Sec:MainProof} we prove Theorem \ref{Theorem:HankelCarleson}. In Section \ref{Sec:SpecialCases} we discuss and compare the special cases of $\HC$-valued, and $\HC^*$-valued symbols, and point out some significant differences between these. In Section \ref{Sec:Davidson-Paulsen} we provide proofs of the Propositions \ref{Proposition:Davidson-Paulsen1} and \ref{Proposition:Davidson-Paulsen2}.
\section{Preliminaries and further notation}\label{Sec:Preliminaries}
We use the standard notation $\Z$, $\R$, and $\C$ for the respective rings of integers, real numbers, and complex numbers. By $\N$ we denote the set of strictly positive elements of $\Z$, while $\N\cup\left\{0\right\}$ is denoted by $\N_0$.
Given two parametrized sets of nonnegative numbers $\left(A_i\right)_{i\in I}$ and $\left(B_i\right)_{i\in I}$, we use the notation $A_i\lesssim B_i$, $i\in I$ to indicate the existence of a positive constant $C$ such that $\forall i\in I$, $A_i\le CB_i$. We then say that $A_i$ is bounded by $B_i$, and refer to $C$ as a bound. Sometimes we allow ourselves to not mention the index set $I$ and instead let it be implicit from the context. If $A_i\lesssim B_i$ and $B_i\lesssim A_i$, then we write $A_i\approx B_i$. We then say that $A_i$ and $B_i$ are comparable.
The Hilbert space adjoint of $A\in\LC$ is denoted by $A^*$. We sometimes identify $x\in\HC$ with the rank one operator $\C\ni c\mapsto cx\in\HC$. Note that $x^*$ is then the linear functional $\HC\ni y\mapsto \left\langle y,x\right\rangle_\HC\in\C$.
The dual of a Banach space $\YC$ will be denoted by $\YC^*$. With Hilbert spaces in mind, we equip $\YC^*$ with an anti linear structure, rather than the standard linear ditto. Thus, the duality pairing $\left\langle y,y^*\right\rangle_\YC$, of $y\in\YC$ and $y^*\in\YC^*$, becomes anti linear in $y^*$.
We define the tensor product of two elements $x,y\in\HC$ as the rank one operator $x\otimes y:z\mapsto\left\langle z,y\right\rangle_\HC x$. The tensor product is anti linear in its second argument. The projective tensor product $\HC\hat \otimes\HC$, is the closed linear span of $\left\{x\otimes y\right\}_{x,y\in\HC}$, with respect to the norm
\[
\left\|T\right\|_{\wedge}=\inf \left\{\sum_k \left\|x_k\right\|_\HC \left\|y_k\right\|_\HC;T=\sum_k x_k\otimes y_k \right\}.
\]
The space $\HC\hat \otimes\HC$ can be isometrically identified with $\SC^1$. The dual of $\SC^1$ is isometrically identified with $\LC$ via the pairing
\[
\left\langle T,B\right\rangle_{\SC^1}=\tr \left(B^*T\right) =\sum_{n}\left\langle Te_n,Be_n\right\rangle_\HC=\sum_{k}\left\langle x_k,By_k\right\rangle_\HC,
\]
where $B\in\LC$, $\left(e_n\right)_{n=0}^\infty$ is any orthonormal basis of $\HC$, and $\sum_k x_k\otimes y_k$ is any representation of $T$, cf. Wojtaszczyk \cite{Wojtaszczyk1991:BSpacesForAnalysts}*{III.B.26}.
An important property of Hardy spaces $H^p\left(\XC\right)$ is that, given certain properties of $\XC$, $H^p\left(\XC\right)$ may be isometrically identified as a subspace of $L^p\left(\T,\XC\right)$. The precise statement is as follows:
\begin{proposition}\label{Proposition:BoundaryIdentification}
Let $p\in[1,\infty]$, and $f\in H^p\left(\XC\right)$.
\begin{itemize}
\item[$\left(i\right)$] If $\XC\in\left\{\C,\HC,\SC^1\right\}$, then there exists a function $bf\in L^p\left(\T,\XC\right)$ such that for $m$-a.e. $\zeta\in\T$, $\lim_{r\to 0}f_r\left(\zeta\right)=bf\left(\zeta\right)$ in the norm topology on $\XC$. Moreover, $f_r\to bf$ in $L^p\left(\T,\XC\right)$, and
\[
\int_{\T}\left(bf\right)\left(\zeta\right)\conj{ \zeta}^ndm\left(\zeta\right)=
\left\{
\begin{array}{cl}
\hat f\left(n\right) & \textnormal{for }n\in\N_0,
\\
0 & \textnormal{for }n\notin\N_0,
\end{array}
\right.
\]
\item[$\left(ii\right)$] If $\XC=\LC$, then there exists a function $bf\in L^p\left(\T,\XC\right)$ such that for $m$-a.e. $\zeta\in\T$, $\lim_{r\to 0}f_r\left(\zeta\right)=bf\left(\zeta\right)$ in the strong operator topology. Moreover, $\left\|bf\right\|_{L^p\left(\T,\XC\right)}=\left\|f\right\|_{H^p\left(\XC\right)}$, and all $x,y\in\HC$
\[
\int_{\T}\left\langle \left(bf\right)\left(\zeta\right)x,y\right\rangle_\HC\conj{ \zeta}^ndm\left(\zeta\right)=
\left\{
\begin{array}{cl}
\langle \hat f\left(n\right)x,y\rangle_\HC & \textnormal{for }n\in\N_0,
\\
0 & \textnormal{for }n\notin\N_0,
\end{array}
\right.
\]
\end{itemize}
In particular, we may identify the Taylor coefficients of $f$ with the Fourier coefficients of $bf$.
\end{proposition}
In the scalar case, the above result is proved in any serious introduction to Hardy spaces. We mention \cite{Garnett2007:BddAnalFcnsBook}. We refer to \cite{Nikolski2002:EasyReading} for the case $\XC=\HC$, and \cite{Rosenblum-Rovnyak1985:HardyClassesOpTheory} for the case $\XC=\LC$. The statement for $\XC=\SC^1$ holds because $\SC^1$ has the so-called analytic Radon--Nikodym property, see \citelist{\cite{Bukhvalov-Danilevich1982:BdryPropsAnalHarmFcnsValBSpace}\cite{Haagerup-Pisier1989:FactAnalFcnsNon-CommL1Spaces}}.
We define the formal duality pairing between $f\in\Hol\left(\YC\right)$ and $g\in\Hol\left(\YC^*\right)$ as
\[
\left\langle f,g\right\rangle= \sum_{n=0}^\infty \left\langle \hat f\left(n\right), \hat g\left(n\right)\right\rangle_{\YC}.
\]
The pairing is well defined if $f\in\OC\left(\YC\right)$ or $g\in\OC\left(\YC^*\right)$, and generalizes the inner product on $H^2\left(\HC\right)$. Note that $\left\langle D^\alpha f,g\right\rangle = \left\langle f, D^\alpha g\right\rangle$, and, in the case where $\YC=\HC$, $\left\langle f,\Gamma_\phi g\right\rangle=\left\langle f\otimes \conjvar{g},\phi \right\rangle$.
We will make use of two related notions of weighted Bergman spaces. For $\beta>-1$, we define two finite measures on $\D$:
\[
dA_{\beta}\left(z\right)=\frac{1+\beta}{\pi} \left(1-\left|z\right|^2 \right)^{\beta}dA\left(z\right)\quad\textnormal{and}\quad dA_{\beta,\log}\left(z\right)=\frac{1+\beta}{\pi} \left(\log \left(\frac{1}{|z|^2} \right) \right)^{\beta}dA\left(z\right).
\]
Here $dA$ denotes area measure on $\C$. For $p\in\left[1,\infty\right)$, we denote by $L_{\beta}^p\left(\YC\right)$ the space of strongly measurable functions $f:\D\to\YC$ such that
\[
\left\|f\right\|_{L_{\beta}^p\left(\YC\right)}^p=\int_\D\left\|f\left(z\right)\right\|_\YC^p\, dA_{\beta}\left(z\right)<\infty.
\]
We then define the standard weighted Bergman space $A_{\beta}^p\left(\YC\right)=L_{\beta}^p\left(\YC\right)\cap \Hol\left(\YC\right)$. We similarly define the logarithmically weighted spaces $L_{\beta,\log}^p\left(\YC\right)$ and $A_{\beta,\log}^p\left(\YC\right)$, with $dA_{\beta,\log}$ in place of $dA_{\beta}$. An enlightening reference for standard weighted Bergman spaces with $\YC=\C$ is \cite{Hedenmalm-Korenblum-Zhu2000:BergmanSpacesBook}. We remark that many of the results presented below for $\YC$-valued functions follow by the same proofs as in the scalar case.
The above two notions of Bergman spaces are to a large extent interchangeable:
\begin{proposition}\label{Proposition:InterchangeableWeights}
Let $p\in\left[1,\infty\right)$, $\beta>-1$, and $\YC$ be an arbitrary Banach space. We then have that
\[
\left\|f\right\|_{A_{\beta,\log}^p\left(\YC\right)} \approx \left\|f\right\|_{A_{\beta}^p\left(\YC\right)},\quad f\in\Hol\left(\YC\right).
\]
The corresponding bounds depend on $p$ and $\alpha$.
\end{proposition}
One of the above bounds is obtained using the pointwise estimate
\[
1-\left|z\right|^2\le \log \left(\frac{1}{\left|z\right|^2} \right),\quad z\in \D,
\]
and the other by using subharmonicity of the function $z\mapsto \left\|f\left(z\right)\right\|_\YC^p$. We refer the interested reader to the easily modified proof of \cite{Garnett2007:BddAnalFcnsBook}*{Lemma VI.3.2} for details.
A multiplier is an operator $\lambda:\Hol\left(\YC\right)\ni f\mapsto \lambda f\in\Hol\left(\YC\right)$ given by
\[
\lambda f\left(z\right)=\sum_{n=0}^\infty \lambda_n\hat f\left(n\right) z^n,\quad z\in\D,
\]
for some scalar sequence $\left(\lambda_n\right)_{n=0}^\infty$. With some abuse of the terminology in \cite{Buckley-Koskela-Vukotic1999:FracIntDiffBergmanSpaces}, we say that a multiplier is small if $\left|\lambda_n\right|\lesssim\frac{1}{1+n}$. Using ideas from the proof of \cite{Arregui-Blasco2002:MultplrsVecValBergmanSpaces}*{Theorem 3.2}, one can prove the following result, which we refer to as the small multiplier property for Bergman spaces.
\begin{proposition}\label{Proposition:SmallMultipliers}
Let $p\in\left[1,\infty\right)$, $\beta>-1$, and $\YC$ be an arbitrary Banach space. Then small multipliers act boundedly on $A_{\beta}^p\left(\YC\right)$.
\end{proposition}
The spaces $A_{\beta,\log}^2\left(\HC\right)$ and $A_{\beta}^2\left(\HC\right)$ are closed subspaces of $L_{\beta,\log}^2\left(\HC\right)$ and $L_{\beta}^2\left(\HC\right)$ respectively. The corresponding orthogonal projections are denoted by $P_{\beta,\log}$ and $P_{\beta}$. A calculation shows that if $\phi \in\Hol\left(\LC\right)$ and $f\in\Hol\left(\HC\right)$ are sufficiently regular, then
\begin{equation}\label{Eq:BergmanHankelFormulalog}
P_{\beta,\log}\left(\phi \conjvar{f} \right)\left(z\right)
=
\sum_{n=0}^\infty \left(\sum_{m=0}^\infty \left(\frac{1+n}{1+m+n} \right)^{1+\beta} \hat \phi \left(m+n\right)\hat f\left(m\right) \right)z^n,\quad z\in\D.
\end{equation}
and
\begin{equation}\label{Eq:BergmanHankelFormula}
P_{\beta}\left(\phi \conjvar{f} \right)\left(z\right)
=
\sum_{n=0}^\infty \left(\sum_{m=0}^\infty \frac{\Gamma\left(1+m+n\right)\Gamma\left(2+\beta+n\right)}{\Gamma\left(2+\beta+m+n\right)\Gamma\left(1+n\right)} \hat \phi \left(m+n\right)\hat f\left(m\right) \right)z^n,\quad z\in\D.
\end{equation}
Here $\Gamma:\C\setminus \left\{-1,-2,\ldots\right\}\to\C$ is the standard $\Gamma$-function. By \eqref{Eq:BergmanHankelFormulalog} and \eqref{Eq:BergmanHankelFormula} we are allowed to define $P_{\beta,\log}\left(\phi \conjvar{f} \right)$ and $P_{\beta}\left(\phi \conjvar{f} \right)$ as elements of $\Hol\left(\HC\right)$, whenever $\phi\in\Hol\left(\LC\right)$ and $f\in\OC\left(\HC\right)$. In this sense, they are analogues of \eqref{Eq:HardyHankelFormula}.
Using Parseval's identity we obtain
\[
\left\|f\right\|_{A_{\beta,\log}^2\left(\HC\right)}^2=\Gamma\left(2+\beta\right)\sum_{k=0}^\infty \frac{\|\hat f\left(k\right)\|_\HC^2}{\left(1+k\right)^{1+\beta}},
\]
and
\[
\left\|f\right\|_{A_{\beta}^2\left(\HC\right)}^2=\sum_{k=0}^\infty \binom{n+1+\beta}{n}^{-1}\|\hat f\left(k\right)\|_\HC^2,
\]
where $\binom{n+1+\alpha}{n}=\frac{\Gamma\left(2+\beta+n\right)}{\Gamma\left(2+\beta\right)\Gamma\left(1+n\right)}$ are generalized binomial coefficients.
The Bloch space $\BC\left(\YC\right)$ is the space of functions $f\in\Hol\left(\YC\right)$ such that
\[
\left\|f\right\|_{\BC\left(\YC\right)}=\sup_{z\in\D}\, \left(1-\left|z\right|^2\right)\left\|Df\left(z\right)\right\|_\YC<\infty.
\]
In the literature the Bloch space is typically defined by finiteness of the quantity
\[
\left\|f\left(0\right)\right\|_\YC+\sup_{z\in\D}\, \left(1-\left|z\right|^2\right)\left\|f'\left(z\right)\right\|_\YC.
\]
We leave it as an exercise to show that these definitions are equivalent.
The Bloch space has the simple property that
\begin{equation}\label{Eq:BlochWBloch}
\left\|f\right\|_{\BC\left(\YC^*\right)}=\sup_{\substack{y\in\YC\\\left\|y\right\|_\YC=1}}\left\|\left\langle y,f\right\rangle_\YC\right\|_\BC,
\end{equation}
as can be seen by interchanging the order of suprema.
The importance of the Bloch space is that $\BC\left(\YC^*\right)$ is isometric to $A_{\beta}^1\left(\YC\right)^*$ via the pairing
\[
\left\langle f,g\right\rangle_{A_{\beta}^1\left(\YC\right)} =\lim_{r\uparrow 1}\int_{\D}\left\langle f\left(rz\right),g\left(z\right)\right\rangle_{\YC}\, dA_{\beta}\left(z\right),\quad f\in A_{\beta}^1\left(\YC\right), g\in\BC\left(\YC^*\right).
\]
This follows mostly as in \cite{Hedenmalm-Korenblum-Zhu2000:BergmanSpacesBook}. The major difference is that $\BC\left(\YC^*\right)$ is the Bergman projection of a certain class of measures, rather than $L^\infty\left(\D,\YC^*\right)$, see \cite{Arregui-Blasco2003:BergmanBlochSpacesVecVal}.
\section{Proof of Theorem \ref{Theorem:HankelCarleson}}\label{Sec:MainProof}
Given $\alpha>0$, let $\beta>\max\left\{2,1+\alpha\right\}$ be an auxiliary parameter. To prove Theorem \ref{Theorem:HankelCarleson}, let $\phi\in\Hol\left(\LC\right)$ and define
\begin{align*}
\left\|\phi\right\|_{1,\alpha}&=\sup_{f\in \OC_1\left(\HC\right)} \left\|D^\alpha\Gamma_\phi f\right\|_{H^2\left(\HC\right)},
\\
\left\|\phi\right\|_{2,\alpha}&=\sup_{f\in \OC_1\left(\HC\right)} \left\|P_{2\beta-1,\log} \left(\left(D^{\beta+\alpha}\phi\right)\conjvar{ f} \right)\right\|_{A_{2\beta-1,\log}^2\left(\HC\right)},
\\
\left\|\phi\right\|_{3,\alpha}&=\sup_{f\in \OC_1\left(\HC\right)} \left\|P_{1,\log} \left(\left(D^{1+\alpha}\phi \right)\conjvar{ f} \right)\right\|_{A_{1,\log}^2\left(\HC\right)},
\\
\left\|\phi\right\|_{4,\alpha}&=\sup_{f\in \OC_1\left(\HC\right)} \left\|P_{1} \left(\left(D^{1+\alpha}\phi \right)\conjvar{ f} \right)\right\|_{A_{1,\log}^2\left(\HC\right)},
\\
\left\|\phi\right\|_{5,\alpha}&=\sup_{f\in \OC_1\left(\HC\right)} \left\|P_{1} \left(\left(D^{1+\alpha}\phi \right)\conjvar{ f} \right)\right\|_{A_{1}^2\left(\HC\right)},
\\
\left\|\phi\right\|_{6,\alpha}&=\sup_{f\in \OC_1\left(\HC\right)} \left\|\left(D^{1+\alpha} \phi\right)\conjvar{ f}\right\|_{L_{1}^2\left(\HC\right)}.
\end{align*}
Theorem \ref{Theorem:HankelCarleson} is the statement that $\left\|\phi \right\|_{1,\alpha}\approx \left\|\phi \right\|_{6,\alpha}$. We will prove that the quantities $\left\|\phi \right\|_{k,\alpha}$, $1\le k\le 6$ are pairwise comparable.
The outline of the proof is as follows. We show in detail that $\left\|\phi \right\|_{1,\alpha}\lesssim \left\|\phi \right\|_{2,\alpha}$. The reverse estimate, as well as the estimates $\left\|\phi \right\|_{2,\alpha}\approx \left\|\phi \right\|_{3,\alpha}$, and $\left\|\phi \right\|_{3,\alpha}\approx \left\|\phi \right\|_{4,\alpha}$ are similar, although the last one is substantially simpler than the preceding ones. The statement that $\left\|\phi \right\|_{4,\alpha}\approx \left\|\phi \right\|_{5,\alpha}$ is just a special case of Proposition \ref{Proposition:InterchangeableWeights}. Furthermore, it is trivial that $\left\|\phi \right\|_{5,\alpha}\le \left\|\phi \right\|_{6,\alpha}$. The reverse of this last estimate follows in a routine manner from the following remarkable result by Aleman and Perfekt \cite{Aleman-Perfekt2012:HankFrmsEmbThmsDirichletSpaces}:
\begin{lemma}\label{Lemma:AlemanPerfekt}
There exists a constant $C>0$ such that whenever $\psi\in\Hol\left(\LC\right)$ it holds that
\begin{align*}
\sup_{f\in\OC_1\left(\SC^2\right)}\int \left\| \left(D\psi\right) \conjvar{ f}\, \right\|_{\SC^2}^2\, dA_1 \le C \sup_{f,g\in\OC_1\left(\SC^2\right)} \left|\int \tr \left( \left(D\psi \right) \conjvar{ f} \left( Dg \right)^* \right)\, dA_1 \right|.
\end{align*}
\end{lemma}
To prove that $\left\|\phi \right\|_{1,\alpha}\lesssim \left\|\phi \right\|_{2,\alpha}$ we will need some lemmata. The first result gives us some preliminary control of $\phi $.
\begin{lemma}\label{Lemma:HankelCarlesonBloch}
For each $\alpha>0$ it holds that
\[
\left\|D^\alpha \phi \right\|_{\BC\left(\LC\right)}\lesssim \left\|\phi \right\|_{k,\alpha},\quad \phi \in\Hol\left(\LC\right),\ 1\le k\le 6.
\]
\end{lemma}
\begin{proof}
We consider only the case $k=1$. The other cases are similar. By \eqref{Eq:BlochWBloch} it suffices to prove that
\begin{equation}\label{Eq:EstBloch}
\left|\left\langle x, D^{1+\alpha} \phi \left(w\right)y\right\rangle_\HC\right|\lesssim \frac{\left\|\phi \right\|_{\alpha,1}\left\|x\right\|_\HC\left\|y\right\|_\HC}{1-\left|w\right|^2},\quad w\in\D,\ x,y,\in\HC.
\end{equation}
Given $w\in\D$, $x,y\in\HC$, let
\[
f\left(z\right)=\sum_{n=0}^\infty \conj{w}^n \left(1+n-n \left(\frac{n}{1+n} \right)^\alpha \right)z^ny,\quad g\left(z\right)=\sum_{n=0}^\infty w^nz^nx,\quad z\in\D.
\]
A calculation shows that $1+n-n \left(\frac{n}{1+n} \right)^\alpha$ is bounded in $n$, and so $\left\|f\right\|_{H^2}\left\|g\right\|_{H^2}\lesssim \frac{1}{1-\left|w\right|^2}$.
The definition of $\left\|\phi \right\|_{\alpha,1}$ now yields \eqref{Eq:EstBloch}.
\end{proof}
\begin{remark}
Another proof of Lemma \ref{Lemma:HankelCarlesonBloch} is to use \eqref{Eq:BlochWBloch} together with the (already known) scalar version of Theorem \ref{Theorem:HankelCarleson}. Our approach is chosen so that our results do not depend on the scalar case.
\end{remark}
The qualitative content of the next lemma is known, and due to Peller \cite{Peller1982:VecHankOps}. See also \cite{Peller2003:HankOpsBook}*{Chapter 6.9}. However, the original proof gives a much worse quantitative dependence on $l$. The proof we present is a bit lengthy, and is postponed to the next subsection.
\begin{lemma}\label{Lemma:OrderControl}
For each $\alpha>0$ it holds that
\[
\left\|D^\alpha \Gamma_{D^{-\alpha-l}\psi }D^l\right\|_{H^2\left(\HC\right)\to H^2\left(\HC\right)}\le C l \left\|\psi \right\|_{\BC\left(\LC\right)},\quad l\in\N,\ \psi \in\Hol\left(\LC\right).
\]
\end{lemma}
We are now ready for the main part of the argument. Given $f\in\OC\left(\HC\right)$, and $\phi \in\Hol\left(\LC\right)$, we use the formulas \eqref{Eq:HardyHankelFormula} and \eqref{Eq:BergmanHankelFormula} to write
\begin{align*}
\left\|D^\alpha \Gamma_\phi f\right\|_{H^2\left(\HC\right)}^2
&=
\sum_{n=0}^\infty\left(1+n\right)^{2\alpha} \left\|\sum_{k=0}^\infty \hat \phi \left(n+k\right)\hat f\left(k\right)\right\|_\HC^2
\\
&=
\sum_{n=0}^\infty\frac{1}{\left(1+n\right)^{2\beta}} \left\|\sum_{k=0}^\infty \left(\frac{1+n}{1+n+k} \right)^{\alpha+\beta} \left(D^{\alpha+\beta} \phi \right)^{\hat{}}\left(n+k\right)\hat f\left(k\right) \right\|_\HC^2
\\
&=
\left\|P_{\alpha+\beta-1,\log} \left(\left(D^\beta \psi \right)\conjvar{f} \right)\right\|_{A^2_{2\beta-1,\log}\left(\HC\right)}^2,
\end{align*}
where $\psi =D^\alpha \phi $.
A well known fact about standard weighted Bergman spaces is that there exists many bounded projections from $L_{\gamma}^p$ onto $A_{\gamma}^p$, eg. \cite{Hedenmalm-Korenblum-Zhu2000:BergmanSpacesBook}*{Theorem 1.10}. This inspires us to replace $P_{\alpha+\beta-1,\log}$ with $P_{2\beta-1,\log}$. By the triangle inequality
\begin{align*}
&\left\|P_{\alpha+\beta-1,\log} \left(\left(D^\beta \psi \right)\conjvar{f} \right)\right\|_{A^2_{2\beta-1,\log}}
\\
&\le \left\| \left(P_{\alpha+\beta-1,\log}-P_{2\beta-1,\log} \right) \left(\left(D^\beta \psi \right)\conjvar{f} \right)\right\|_{A^2_{2\beta-1,\log}}+ \left\|P_{2\beta-1,\log} \left(\left(D^\beta \psi \right)\conjvar{f} \right)\right\|_{A^2_{2\beta-1,\log}}.
\end{align*}
We carry out a few manipulations with the Taylor coefficients of $\phi $ and $f$, use the power series expansion at the origin of the function $z\mapsto \left(1-z\right)^{\beta-\alpha}$, and apply Minkowski's inequality to obtain
\begin{align*}
&\left\| \left(P_{\alpha+\beta-1,\log}-P_{2\beta-1,\log} \right) \left(\left(D^\beta \psi \right)\conjvar{f} \right)\right\|_{A^2_{2\beta-1,\log}}^2
\\
&=
\sum_{n=0}^\infty\frac{1}{\left(1+n\right)^{2\beta}} \left\|\sum_{k=0}^\infty \left[ \left(\frac{1+n}{1+n+k} \right)^{\alpha+\beta} - \left(\frac{1+n}{1+n+k} \right)^{2\beta} \right]\left(D^{\beta} \psi \right)^{\hat{}}\left(n+k\right)\hat f\left(k\right) \right\|_\HC^2
\\
&=
\sum_{n=0}^\infty \left\|\sum_{k=0}^\infty \left[ \left(\frac{1+n}{1+n+k} \right)^{\alpha} - \left(\frac{1+n}{1+n+k} \right)^{\beta} \right]\hat \psi \left(n+k\right)\hat f\left(k\right) \right\|_\HC^2
\\
&=
\sum_{n=0}^\infty \left\|\sum_{k=0}^\infty \left(\frac{1+n}{1+n+k} \right)^{\alpha} \left[1 - \left(1-\frac{k}{1+n+k} \right)^{\beta-\alpha} \right]\hat \psi \left(n+k\right)\hat f\left(k\right) \right\|_\HC^2
\\
&=
\sum_{n=0}^\infty \left\|\sum_{l=1}^\infty\binom{\beta-\alpha}{l}\left(-1\right)^l\sum_{k=0}^\infty \frac{\left(1+n\right)^\alpha \left(1+k\right)^l}{\left(1+n+k\right)^{\alpha+l}}\hat \psi \left(n+k\right)\left(\frac{k}{1+k}\right)^l\hat f\left(k\right) \right\|_\HC^2
\\
&\le
\left(\sum_{l=1}^\infty \left|\binom{\beta-\alpha}{l} \right| \left(\sum_{n=0}^\infty \left\|\sum_{k=0}^\infty \frac{\left(1+n\right)^\alpha \left(1+k\right)^l}{\left(1+n+k\right)^{\alpha+l}}\hat \psi \left(n+k\right) \left(\frac{k}{1+k} \right)^l\hat f\left(k\right) \right\|_\HC^2 \right)^{1/2} \right)^2
\\
&=
\left(\sum_{l=1}^\infty \left|\binom{\beta-\alpha}{l}\right|\left\|D^\alpha\Gamma_{D^{-\alpha-l}}D^l f_l\right\|_{H^2\left(\HC\right)} \right)^2,
\end{align*}
where $f_l$ is defined by $\hat f_l\left(k\right)= \left(\frac{k}{1+k} \right)^l\hat f\left(k\right)$. By Stirling's formula, the binomial coefficients $\binom{\beta-\alpha}{l}$ decay like $\frac{1}{l^{1+\beta-\alpha}}$, and since the map $f\mapsto f_l$ is obviously $H^2\left(\HC\right)$-contractive for each $l$, we use Lemma \ref{Lemma:OrderControl} to conclude that
\[
\left\| \left(P_{\alpha+\beta-1,\log}-P_{2\beta-1,\log} \right) \left(\left(D^\beta \psi \right)\conjvar{f} \right)\right\|_{A^2_{2\beta-1,\log}}\lesssim \left\|\psi \right\|_{\BC\left(\LC\right)}\left\|f\right\|_{H^2\left(\HC\right)},
\]
since $\beta>1+\alpha$. Lemma \ref{Lemma:HankelCarlesonBloch} then implies that
\[
\left\|P_{\alpha+\beta-1,\log} \left(\left(D^\beta \psi \right)\conjvar{f} \right)\right\|_{A^2_{2\beta-1,\log}} \lesssim \left\|\phi \right\|_{2,\alpha}\left\|f\right\|_{H^2\left(\HC\right)}.
\]
This proves that $\left\|\phi \right\|_{1,\alpha}\lesssim \left\|\phi \right\|_{2,\alpha}$. It is straightforward to use the same type of argument to show that $\left\|\phi \right\|_{2,\alpha}\lesssim \left\|\phi \right\|_{1,\alpha}$.
In order to prove that $\left\|\phi \right\|_{2,\alpha}\approx \left\|\phi \right\|_{3,\alpha}$, we note that
\[
\left\|P_{2\beta-1,\log} \left(\left(D^\beta \psi \right)\conjvar{f} \right)\right\|_{A^2_{2\beta-1,\log}}
=
\left\|P_{\beta,\log} \left(\left(D^1 \psi \right)\conjvar{f} \right)\right\|_{A^2_{1,\log}}.
\]
We repeat the above argument in order to replace $P_{\beta,\log}$ with $P_{1,\log}$. This time instead of $\beta>1+\alpha$, we use that $\beta>2$. A third application of the argument allows us to replace $P_{1,\log}$ with $P_1$, yielding $\left\|\phi \right\|_{3,\alpha}\approx \left\|\phi \right\|_{4,\alpha}$.
As was pointed out earlier, $\left\|\phi \right\|_{4,\alpha}\approx \left\|\phi \right\|_{5,\alpha}$ is just a special case of Proposition \ref{Proposition:InterchangeableWeights}, while the estimate $\left\|\phi \right\|_{5,\alpha}\le \left\|\phi \right\|_{6,\alpha}$ is trivial. For the reverse inequality, if we identify $\HC$ as a subspace of rank one operators in $\SC^2$, it is obvious that
\begin{align*}
\sup_{f\in\OC_1\left(\HC\right)}\int \left\| \left(D^{1+\alpha}\phi \right) \conjvar{f} \right\|_{\HC}^2\, dA_1 \le \sup_{f\in\OC_1\left(\SC^2\right)}\int \left\| \left(D^{1+\alpha}\phi\right) \conjvar{f} \right\|_{\SC^2}^2\, dA_1.
\end{align*}
By a simple argument
\begin{align*}
\left|\int \tr \left( \left(D^{1+\alpha}\phi \right)\conjvar{f} \left(Dg \right)^* \right)dA_1 \right|
\le
\left\|\phi \right\|_{\alpha,5}\left\|f\right\|_{H^2\left(\SC^2\right)}\left\|g\right\|_{H^2\left(\SC^2\right)}
\end{align*}
holds whenever $f,g\in\OC\left(\SC^2\right)$. By Lemma \ref{Lemma:AlemanPerfekt}, $\left\|\phi \right\|_{6,\alpha}\lesssim\left\|\phi \right\|_{5,\alpha}$. This completes the proof of Theorem \ref{Theorem:HankelCarleson}.
\subsection{Proof of Lemma \ref{Lemma:OrderControl}}
For $\alpha>0$ we define the operator $\tilde D^\alpha:\Hol\left(\YC\right)\to\Hol\left(\YC\right)$ by
\[
\tilde D^\alpha f\left(z\right)=\sum_{n=0}^\infty \frac{\Gamma\left(1+n+\alpha\right)}{\Gamma\left(1+n\right)}\hat f\left(n\right) z^n,\quad z\in\D.
\]
A calculation shows that
\begin{align*}
\left\langle \tilde D^\alpha f,\psi \right\rangle_{A_{\alpha-1}^1\left(\YC\right)}=
\Gamma\left(1+\alpha\right)\left\langle f,\psi \right\rangle,
\end{align*}
whenever $f\in\OC\left(\YC\right)$ and $\psi\in\BC\left(\YC^*\right)$.
Going to the case where $\psi\in\BC\left(\LC\right)$, $f,g\in\OC\left(\HC\right)$, we obtain that
\begin{align*}
\left\langle f,D^\alpha\Gamma_{D^{-\alpha-l}\psi }D^lg\right\rangle
&=
\left\langle D^{-\alpha-l} \left( \left(D^\alpha f \right)\otimes \conjvar{ \left(D^l g \right)} \right),\psi \right\rangle
\\
&=
\frac{1}{\Gamma\left(1+\alpha\right)} \left\langle \tilde D^\alpha D^{-\alpha}D^{-l} \left( \left(D^\alpha f \right)\otimes \conjvar{ \left(D^l g \right)} \right),\psi \right\rangle_{A_{\alpha-1}^1\left(\YC\right)}.
\end{align*}
Since $\psi \in\BC\left(\LC\right)$, we have that
\[
\left|\left\langle f,D^\alpha\Gamma_{D^{-\alpha-l}\psi }D^lg\right\rangle\right| \lesssim\left\|\psi \right\|_\BC\left\|\tilde D^\alpha D^{-\alpha}D^{-l} \left( \left(D^\alpha f \right)\otimes \conjvar{ \left(D^l g \right)} \right)\right\|_{A_{\alpha-1}^1\left(\SC^1\right)}.
\]
Following the ideas in \cite{Buckley-Koskela-Vukotic1999:FracIntDiffBergmanSpaces}, we use Stirling's formula to see that $\tilde D^\alpha D^{-\alpha}$ acts like the identity plus a small multiplier. By Propositions \ref{Proposition:InterchangeableWeights} and \ref{Proposition:SmallMultipliers}, we can now complete the proof of Lemma \ref{Lemma:OrderControl} by showing that
\[
\left\| D^{-l} \left( \left(D^\alpha f \right)\otimes \conjvar{ \left(D^l g \right)} \right)\right\|_{A_{\alpha-1,\log}^1\left(\SC^1\right)}\lesssim l \left\|f\right\|_{H^2\left(\HC\right)}\left\|g\right\|_{H^2\left(\HC\right)}.
\]
First we perform a simple decomposition of $f$ and $g$ into low and high frequencies. Assume that $f$ and $g$ are of degree at most $l$. By the triangle inequality we have
\begin{align*}
&\left\|D^{-l} \left( \left(D^\alpha f \right)\otimes \conjvar{ \left(D^l g \right)} \right)\right\|_{A_{\alpha-1,\log}^1\left(\SC^1\right)}
\\
&\le
\sum_{m,n=0}^l\left\|\hat f\left(m\right)\right\|_{\HC}\left\|\hat g\left(n\right)\right\|_{\HC} \left\| D^{-l}\left(\left(D^\alpha z^m\right)\left(D^l z^n\right)\right)\right\|_{A_{\alpha-1,\log}^1}
\\
&=
\sum_{m,n=0}^l\frac{\left(1+m\right)^\alpha\left\|\hat f\left(m\right)\right\|_{\HC}\left(1+n\right)^l\left\|\hat g\left(n\right)\right\|_{\HC}}{\left(1+m+n\right)^l} \left\| z^{m+n}\right\|_{A_{\alpha-1,\log}^1}.
\end{align*}
Using polar coordinates we compute that
\[
\left\| z^{m+n}\right\|_{A_{\alpha-1,\log}^1}=\frac{2^\alpha\Gamma\left(1+\alpha\right)}{\left(2+m+n\right)^\alpha},
\]
and so
\begin{align*}
\left\|D^{-l} \left( \left(D^\alpha f \right)\otimes \conjvar{ \left(D^l g \right)} \right)\right\|_{A_{\alpha-1,\log}^1\left(\SC^1\right)}
&\lesssim \sum_{m,n=0}^l\left\|\hat f\left(m\right)\right\|_{\HC}\left\|\hat g\left(n\right)\right\|_{\HC}
\\
&\le l\left\|f\right\|_{H^2\left(\HC\right)}\left\|g\right\|_{H^2\left(\HC\right)},
\end{align*}
by Cauchy--Schwarz's inequality. Thus the low frequencies exhibit the desired behaviour.
We now consider the high frequencies. Assume that $ \left(D^\alpha f \right)\otimes \conjvar{ \left(D^l g \right)}$ has a zero of order $l$ at the origin. We can then use Lemma \ref{Lemma:PrimitiveNorm}, followed by Cauchy-Schwarz's inequality, and Parseval's identity to obtain that
\begin{align*}
&\left\|D^{-l} \left( \left(D^\alpha f \right)\otimes \conjvar{ \left(D^l g \right)} \right)\right\|_{A_{\alpha-1,\log}^1\left(\SC^1\right)}
\\
&\le
\frac{\Gamma\left(1+\alpha\right)}{2^l\Gamma\left(1+\alpha+l\right)} \left(\frac{2+l}{1+l} \right)^l\left\| \left(D^\alpha f \right)\otimes \conjvar{ \left(D^l g \right)}\right\|_{A_{\alpha+l-1,\log}^1\left(\SC^1\right)}
\\
&\le
\frac{\Gamma\left(\alpha\right)}{2^l\Gamma\left(\alpha+l\right)2l} \left(\frac{2+l}{1+l} \right)^l\left\|D^\alpha f\right\|_{A_{2\alpha-1,\log}^2\left(\HC\right)}\left\|D^lg\right\|_{A_{2l-1,\log}^2\left(\HC\right)}
\\
&=
\frac{\Gamma\left(\alpha\right)\Gamma\left(1+\alpha\right)^{1/2}\Gamma\left(2l\right)^{1/2}}{2^l\Gamma\left(\alpha+l\right)} \left(\frac{2+l}{1+l} \right)^l\left\|f\right\|_{H^2\left(\HC\right)}\left\| g\right\|_{H^2\left(\HC\right)}
\\
&\lesssim
l^{1/4-\alpha}\left\|f\right\|_{H^2\left(\HC\right)}\left\| g\right\|_{H^2\left(\HC\right)},
\end{align*}
where in the last step, we have used Stirling's formula. Assuming Lemma \ref{Lemma:PrimitiveNorm}, this completes the proof of Lemma \ref{Lemma:OrderControl}.
\begin{lemma}\label{Lemma:PrimitiveNorm}
Let $\alpha>0$, $N\in\N_0$, and assume that $h\in\Hol\left(\YC\right)$ has a zero of order $N$ at the origin. Then
\[
\left\|D^{-l}h\right\|_{A_{\alpha-1,\log}^1\left(\YC\right)}\le \frac{\Gamma\left(1+\alpha\right)}{2^l\Gamma\left(1+\alpha+l\right)} \left(\frac{2+N}{1+N} \right)^l\left\|h\right\|_{A_{\alpha+l-1,\log}^1\left(\YC\right)},
\]
whenever $l\in\N$.
\end{lemma}
\begin{proof}
We will use an idea of Flett \cite{Flett1972:DualIneqHardyLittlewood}. Term by term integration of the power series of $h$ shows that
\[
\left(D^{-l} \right)h\left(r\zeta\right)=\frac{1}{\Gamma\left(k\right)r}\int_{s=0}^{r}h_s\left(\zeta\right) \left(\log \left(\frac{r}{s} \right) \right)^{l-1}ds,\quad r\in\left[0,1\right),\zeta\in\T.
\]
By the triangle inequality
\begin{align*}
&\left\|D^{-l}h\right\|_{A_{\alpha-1,\log}^1\left(\YC\right)}
\\
&\le \frac{2\alpha}{\Gamma\left(l\right)}\int_{r=0}^1\int_{\T}\int_{s=0}^r\left\|h_s\left(\zeta\right)\right\|_{\YC} \left(\log \left(\frac{r}{s} \right) \right)^{l-1} \left(\log \left(\frac{1}{r^2} \right) \right)^{\alpha-1}ds\, dm\left(\zeta\right) dr
\\
&=
\frac{\alpha 2^{\alpha}}{\Gamma\left(l\right)}\int_{s=0}^1\int_{\T}\left\|h_s\right\|_{\YC}\, dm \int_{r=s}^1 \left(\log \left(\frac{1}{s} \right)-\log \left(\frac{1}{r} \right) \right)^{l-1} \left(\log \left(\frac{1}{r} \right) \right)^{\alpha-1}dr\, ds.
\end{align*}
By the change of variables $\log\left(\frac{1}{r}\right)/\log\left(\frac{1}{s}\right)=x$ we obtain
\begin{multline*}
\int_{r=s}^1 \left(\log \left(\frac{1}{s} \right)-\log \left(\frac{1}{r} \right) \right)^{l-1} \left(\log \left(\frac{1}{r} \right) \right)^{\alpha-1}dr
\\
=\left(\log \left(\frac{1}{s} \right) \right)^{\alpha+l-1}\int_{x=0}^1\left(1-x\right)^{l-1}x^{\alpha-1}s^xdx.
\end{multline*}
Therefore
\begin{align*}
&\left\|D^{-l}h\right\|_{A_{\alpha-1,\log}^1}
\\
&\le
\frac{\alpha 2^{\alpha}}{\Gamma\left(l\right)}\int_{s=0}^1\int_{\T}\left\|h_s\right\|_{\YC}\, dm \left(\log \left(\frac{1}{s} \right) \right)^{\alpha+l-1} \int_{x=0}^1\left(1-x\right)^{l-1}x^{\alpha-1}s^xdx\, ds
\\
&=
\frac{\alpha 2^{\alpha}}{\Gamma\left(l\right)}\int_{x=0}^1\left(1-x\right)^{l-1}x^{\alpha-1}\int_{s=0}^1\int_{\T}\left\|h_s\right\|_{\YC}\, dm \left(\log \left(\frac{1}{s} \right) \right)^{\alpha+l-1} s^xds\, dx.
\end{align*}
We now replace the variable $s$ with $s^\delta$, where $\delta=\delta\left(x\right)$ will soon be chosen.
\begin{align*}
&\frac{\Gamma\left(l\right)}{\alpha 2^{\alpha}}\left\|D^{-l}h\right\|_{A_{\alpha-1,\log}^1}
\\
&\le
\int_{x=0}^1\left(1-x\right)^{l-1}x^{\alpha-1}\delta^{\alpha+l}\int_{s=0}^1\int_{\T}\left\|h_{s^\delta}\right\|_{\YC}\, dm \left(\log \left(\frac{1}{s} \right) \right)^{\alpha+l-1} s^{\left(1+x\right)\delta-1}ds\, dx
\\
&=
\int_{x=0}^1\left(1-x\right)^{l-1}x^{\alpha-1}\delta^{\alpha+l}\int_{s=0}^1\int_{\T}\frac{\left\|h_{s^\delta}\right\|_{\YC}}{s^{\delta N}}\, dm \left(\log \left(\frac{1}{s} \right) \right)^{\alpha+l-1} s^{\left(1+x+N\right)\delta-1}ds\, dx.
\end{align*}
Choose $\delta=\frac{2+N}{1+N+x}$. Note that $\delta\ge 1$ whenever $x\in[0,1]$. By assumption, the function $z\mapsto \frac{f\left(z\right)}{z^N}$ is analytic. It follows by subharmonicity that
\[
\int_{\T}\frac{\left\|h_{s^\delta}\right\|_{\YC}}{s^{\delta N}}\, dm\le \int_{\T}\frac{\left\|h_{s}\right\|_{\YC}}{s^{N}}\, dm,
\]
and so
\begin{align*}
& \frac{\Gamma\left(l\right)}{\alpha 2^{\alpha}}\left\|D^{-l}h\right\|_{A_{\alpha-1,\log}^1}
\\
&\le
\int_{x=0}^1\left(1-x\right)^{l-1}x^{\alpha-1}\delta^{\alpha+l}\int_{s=0}^1\int_{\T}\left\|h_s\right\|_{\YC}\, dm \left(\log \left(\frac{1}{s} \right) \right)^{\alpha+l-1} s^{\left(1+x+N\right)\delta-1-N}ds\, dx
\\
&=
\int_{x=0}^1\left(1-x\right)^{l-1}x^{\alpha-1} \left(\frac{2+N}{1+N+x} \right)^{\alpha+l}\int_{s=0}^1\int_{\T}\left\|h_s\right\|_{\YC}\, dm \left(\log \left(\frac{1}{s} \right) \right)^{\alpha+l-1} s\, ds\, dx
\\
&=
\frac{1 }{2^{l+\alpha}\left(l+\alpha\right)}\left\|h\right\|_{A_{\alpha+l-1,\log}^1}\int_{x=0}^1\left(1-x\right)^{l-1}x^{\alpha-1} \left(\frac{2+N}{1+N+x} \right)^{\alpha+l}dx.
\end{align*}
Replacing the variable $x$ with $\frac{\left(N+1\right)x}{N+2-x}$ we obtain
\begin{align*}
\int_{x=0}^1\left(1-x\right)^{l-1}x^{\alpha-1} \left(\frac{2+N}{1+N+x} \right)^{\alpha+l}dx
&= \left(\frac{2+N}{1+N} \right)^l\int_{x=0}^1\left(1-x\right)^{l-1}x^{\alpha-1}dx
\\
&=
\left(\frac{2+N}{1+N} \right)^l\frac{\Gamma\left(l\right)\Gamma\left(\alpha\right)}{\Gamma\left(l+\alpha\right)},
\end{align*}
and the proof of Lemma \ref{Lemma:PrimitiveNorm} is complete.
\end{proof}
\begin{remark}
The bound in Lemma \ref{Lemma:PrimitiveNorm} is sharp, as is seen by testing on the function $h\left(z\right)=z^N$. In particular we have that
\[
\left\|D^{-l}\right\|_{A_{\alpha-1,\log}^1\left(\YC\right)\to A_{\alpha+l-1,\log}^1\left(\YC\right)}= \frac{\Gamma\left(1+\alpha\right)}{\Gamma\left(1+\alpha+l\right)}.
\]
This shows that without the separation of $f$ and $g$ into low and high frequencies, the estimate obtained in Lemma \ref{Lemma:OrderControl} would instead be
\[
\left\|D^\alpha \Gamma_{D^{-\alpha-l}\psi }D^l\right\|_{H^2\left(\HC\right)\to H^2\left(\HC\right)}\lesssim 2^l \left\|\psi \right\|_{\BC\left(\LC\right)},\quad \psi\in\Hol\left(\LC\right),
\]
which is of course far from sufficient for proving Theorem \ref{Theorem:HankelCarleson}. Still, some of the estimates in the proof of Lemma \ref{Lemma:OrderControl} are very crude, indicating room for improvement. If Lemma \ref{Lemma:OrderControl} could be improved so that for each $l\in\N$
\[
\left\|D^\alpha \Gamma_{D^{-\alpha-l}\psi }D^l\right\|_{H^2\left(\HC\right)\to H^2\left(\HC\right)}\le C_l \left\|\psi \right\|_{\BC\left(\LC\right)},
\]
where $\sum_{l=1}^\infty \frac{C_l}{l^\gamma}<\infty$ whenever $\gamma>1$, then in the proof of Theorem \ref{Theorem:HankelCarleson} one could immediately prove that $\left\|\phi \right\|_{1,\alpha}\approx\left\|\phi \right\|_{3,\alpha}$, instead of using two iterations of the same argument.
\end{remark}
\section{$\HC$- and $\HC^*$-valued symbols}\label{Sec:SpecialCases}
A function $k_w$, where $w\in\D$, defined by
\[
k_w\left(z\right)=\frac{1}{1-\conj{w} z},\quad z\in\D,
\]
is called a reproducing kernel function for $H^2$. By Parseval's formula, $\left\langle f,k_w\right\rangle=f\left(w\right)$ whenever $f\in H^2$, and $\left\|k_w\right\|_{H^2}^2=\frac{1}{1-\left|w\right|^2}$. From \citelist{\cite{Blasco1997:VecValBMOAGeomBSpaces}\cite{Bonsall1984:BddnessHankMat}} we gather the following result:
\begin{proposition}\label{Proposition:H-Valued}
If $\phi:\D\to\HC$ is analytic, then $\phi \in H^1\left(\HC\right)^*$ if and only if either of the following conditions hold:
\begin{itemize}
\item[$\left(i\right)$]
\[
\sup_{I\subset\T}\frac{1}{m\left(I\right)}\int_I\left\|b\phi-\left(b\phi\right)_I\right\|_\HC\, dm<\infty.
\]
\item[$\left(ii\right)$]
\[
\sup_{f\in\OC_1} \int_\D \left|f\left(z\right)\right|^2\left\| \left(D\phi\right)\left(z\right)\right\|_\HC^2\left(1-\left|z\right|^2\right)dA\left(z\right)<\infty.
\]
\item[$\left(iii\right)$]
\[
\sup_{w\in\D} \left(1-\left|w\right|^2\right)\int_\D \left|k_w\left(z\right)\right|^2\left\| \left(D\phi\right)\left(z\right)\right\|_\HC^2\left(1-\left|z\right|^2\right)dA\left(z\right)<\infty.
\]
\item[$\left(iv\right)$]
\[
\sup_{f\in\OC_1} \left\|\Gamma_\phi f\right\|_{H^2\left(\HC\right)}<\infty.
\]
\item[$\left(v\right)$]
\[
\sup_{w\in\D} \left(1-\left|w\right|^2\right)\left\|\Gamma_\phi k_w\right\|_{H^2\left(\HC\right)}<\infty.
\]
\end{itemize}
Moreover, the corresponding norms are comparable.
\end{proposition}
We point out that even though the results that $\left(iii\right)\Rightarrow\left(ii\right)$ and $\left(v\right)\Rightarrow\left(iv\right)$ look similar, the relation between them is not trivial. The fact that boundedness of a Hankel operator may be determined by its action on reproducing kernels is often referred to as Bonsall's theorem, and is an example of a so called reproducing kernel thesis. It was shown in \cite{Jacob-Rydhe-Wynn2014:WeightWeissConjRKTGenHankOps} that for scalar-valued symbols, the operators $D^\alpha\Gamma_\phi :H^2\to H^2$ ($\alpha\ge 0$) have a reproducing kernel thesis, while $\left(D^\alpha \Gamma_{\phi^\#}\right)^*:H^2\to H^2$ do not. For $\HC$-valued symbols, $D^\alpha\Gamma_\phi :H^2\to H^2\left(\HC\right)$ ($\alpha\ge 0$) satisfies a reproducing kernel thesis. The proof is the same as in the scalar case. In this section, we investigate the corresponding results for Carleson embeddings.
By specializing Theorem \ref{Theorem:HankelCarleson} to the case of rank one-valued symbols, we obtain the following corollary:
\begin{corollary}\label{Corollary:HankelCarlesonrankone}
Let $\alpha>0$ and $\phi\in\Hol\left(\HC\right)$. Then the operator $D^\alpha\Gamma_\phi:H^2\to H^2\left(\HC\right)$ is bounded if and only if
\begin{equation}\label{Eq:RankOneCarlesonCondition}
\sup_{f\in\OC_1}\int_ \D \left|f\left(z\right)\right|^2\left\|D^{1+\alpha}\phi\left(z\right)\right\|_\HC^2\left(1-\left|z\right|^2\right)dA\left(z\right)<\infty.
\end{equation}
Moreover, the above supremum is comparable to $\left\|D^\alpha\Gamma_\phi\right\|_{H^2\to H^2\left(\HC\right)}^2$.
\end{corollary}
Combined with Proposition \ref{Proposition:H-Valued}, Corollary \ref{Corollary:HankelCarlesonrankone} says that $D^\alpha\Gamma_\phi:H^2\to H^2\left(\HC\right)$ is bounded if and only if $D^\alpha \phi\in \textrm{BMOA}\left(\HC\right)$, i.e. $\phi\in D^{-\alpha} \textrm{BMOA}\left(\HC\right)=\left(D^\alpha H^1\left(\HC\right)\right)^*$. This shows that Corollary \ref{Corollary:HankelCarlesonrankone} could also have been obtained from the factorization $D^\alpha H^1\left(\HC\right)=H^2\cdot D^\alpha H^2\left(\HC\right)$, see \citelist{\cite{Cohn-Verbitsky2000:FactTentSpacesHankOps}\cite{Rydhe2016:CharTriebel-LizorkinSpaces}}.
We now state the corresponding result for functional-valued symbols:
\begin{corollary}
Let $\alpha>0$ and $\phi\in\Hol\left(\HC\right)$. Then the operator $D^\alpha\Gamma_{\phi^\#}:H^2\left(\HC\right)\to H^2$ is bounded if and only if
\begin{equation}\label{Eq:CoRankOneCarlesonCondition}
\sup_{f\in\OC_1\left(\HC\right)}\int_ \D \left|\left\langle f\left(z\right),\left(D^{1+\alpha}\phi\right)\left(z\right)\right\rangle_\HC\right|^2\left(1-\left|z\right|^2\right)dA\left(z\right)< \infty.
\end{equation}
Moreover, the above supremum is comparable to $\left\|D^\alpha\Gamma_{\phi^\#}\right\|_{H^2\left(\HC\right)\to H^2}^2$.
\end{corollary}
Even though $\HC$ and $\HC^*$ are isomorphic, condition \eqref{Eq:CoRankOneCarlesonCondition} is far more subtle than \eqref{Eq:RankOneCarlesonCondition}. It is easy to show that \eqref{Eq:RankOneCarlesonCondition} implies \eqref{Eq:CoRankOneCarlesonCondition}. The reverse implication does not hold, as is seen by Theorem \ref{Theorem:HankelCarleson} together with Proposition \ref{Proposition:Davidson-Paulsen1}. This also shows that $D^\alpha H^1\left(\HC\right)\ne H^2\left(\HC\right)\cdot D^\alpha H^2$.
Motivated by Proposition \ref{Proposition:H-Valued}, it is natural to consider the condition
\begin{equation}\label{Eq:WeakBMOA}
\sup_{\substack{w\in\D,\\ x\in\HC, \left\|x\right\|_\HC=1}}\left(1-\left|w\right|^2\right)\int_ \D \left|\left\langle k_w\left(z\right)x,\left(D^{1+\alpha}\phi\right)\left(z\right)\right\rangle_\HC\right|^2\left(1-\left|z\right|^2\right)dA\left(z\right)< \infty.
\end{equation}
This weak type condition means that the functions $z\mapsto \left\langle \phi\left(z\right),x\right\rangle_\HC$ are in scalar-valued $\textrm{BMOA}$, uniformly for all $x$ in the unit ball of $\HC$. We use the conditions \eqref{Eq:RankOneCarlesonCondition}, \eqref{Eq:CoRankOneCarlesonCondition}, and \eqref{Eq:WeakBMOA} to define the respective spaces $\textrm{BMOA}_{\CC}\left(\HC\right)$, $\textrm{BMOA}_{\CC^\#}\left(\HC\right)$, and $\textrm{BMOA}_{\WC}\left(\HC\right)$. We then have the strict inclusions
\[
\textrm{BMOA}_{\CC}\left(\HC\right)\subsetneq \textrm{BMOA}_{\CC^\#}\left(\HC\right) \subsetneq \textrm{BMOA}_{\WC}\left(\HC\right).
\]
We refer to \cite{Rydhe2016:CounterExsCarlesonEmbThm} for an example showing that the last inclusion is strict.
\section{The Davidson--Paulsen results}\label{Sec:Davidson-Paulsen}
We will now present the proofs of Propositions \ref{Proposition:Davidson-Paulsen1} and \ref{Proposition:Davidson-Paulsen2}. We once again point out that these are (at most) straightforward adaptations of the arguments used in \cite{Davidson-Paulsen1997:PolBddOps}. It will be convenient to identify $H^2\left(\HC\right)$ with $l^2\left(\N_0,\HC\right)$, and let $\HC= l^2\left(\N_0\right)$. We let $\left(e_n\right)_{n= 0}^\infty$ denote the canonical basis for $l^2\left(\N_0\right)$.
\subsection{Proof of Proposition \ref{Proposition:Davidson-Paulsen1}}
Let $x\in\HC$ be a fixed vector of unit length, and consider the function $\phi:z\mapsto \sum_{n=0}^\infty \beta_n x\otimes e_n z^n$, where $\left(\beta_n\right)_{n=0}^\infty$ is some scalar sequence of moderate growth. The function $\phi$ is obviously rank one-valued, and with the right choice of $\left(\beta_n\right)_{n=0}^\infty$ it has the property that $D^\alpha\Gamma_\phi$ is bounded on $H^2\left(\HC\right)$, while $\Gamma_\phi D^\alpha$ is not.
Since the contraction $H^2\left(\HC\right)\ni f\mapsto \left\langle f,x\right\rangle_\HC\in H^2$ maps a subset of the unit sphere in $H^2\left(\HC\right)$ onto the unit sphere in $H^2$, we may instead consider boundedness of $D^\alpha\Gamma_\psi:H^2\left(\HC\right)\to H^2$, where $\psi$ is the $\HC^*$-valued function $z\mapsto \sum_{n=0}^\infty \beta_n e_n^* z^n$.
It will be simpler to consider boundedness of the operators $\left(D^\alpha\Gamma_\psi\right)^*=\Gamma_{\psi^\#}D^\alpha$ and $\left(\Gamma_\psi D^\alpha\right)^*=D^\alpha \Gamma_{\psi^\#}$. Let $X=[\beta_{m+n}e_{m+n}]_{m,n\ge 0}$ be the matrix representation of $\Gamma_{\psi^\#}$. The goal is now to show that $X\ \diag\left(\left(1+n\right)^{\alpha}\right)_{n\ge 0}$ is bounded from $l^2\left(\N_0\right)$ to $l^2\left(\N_0,\HC\right)$, while $\diag\left(\left(1+n\right)^{\alpha}\right)_{n\ge 0} X$ is not.
Obviously, the operator norm of $D^\alpha X$ is at least as big as the $l^2\left(\N_0,\HC\right)$-norm of each column of the matrix, i.e.
\[
\left\|D^\alpha X\right\|_{l^2\left(\N_0\right)\to l^2\left(\N_0,\HC\right)}^2\ge \sup_{k\in\N_0}\sum_{n=0}^\infty \left(1+n\right)^{2\alpha}\left|\beta_{n+k}\right|^2,
\]
so if for example $\sum_{n=0}^\infty \left(1+n\right)^{2\alpha}\left|\beta_{n}\right|^2=\infty$, then $D^\alpha X$ is unbounded. On the other hand,
\[
\left\langle Xe_n,X e_m\right\rangle_\HC= \left\{
\begin{array}{cl}
\gamma_n^2:=\sum_{k\ge n}\left|\beta_n\right|^2&\textnormal{for $m=n$},\\
0&\textnormal{otherwise}.
\end{array} \right.
\]
If follows that $ \left(XD^\alpha \right)^*XD^\alpha=\diag \left(\left(1+n\right)^{2\alpha}\gamma_n^2 \right)_{n\ge 0}$, and so
\[
\left\|X D^\alpha\right\|_{l^2\left(\N_0\right)\to l^2\left(\N_0,\HC\right)}^2 = \sup_{n\in\N_0}\left(1+n\right)^{2\alpha}\sum_{k\ge n}\left|\beta_n\right|^2.
\]
Now chose $\beta_n=\frac{1}{\left(1+n\right)^{\alpha+1/2}}$ to complete the proof.
\subsection{Proof of Proposition \ref{Proposition:Davidson-Paulsen2}}
Given matrices $A=[a_{mn}]_{m,n\ge 0}$ and $B=[b_{mn}]_{m,n\ge 0}$, we define the Schur product $A\star B= \left[a_{mn}b_{mn}\right]_{m,n\ge 0}$. For a fixed matrix $B$, the operator $S_B:A\mapsto A\star B$ is called a Schur multiplier. The Grothendieck-Haagerup criterion, e.g. \cite{Paulsen2002:ComplBddMapsOpAlgs}*{Corollary 8.8}, states that $S_B:\LC\left(\HC\right)\to\LC\left(\HC\right)$ is bounded if and only if there exists sequences $\left(x_n\right)_{n\ge 0}$, $\left(y_n\right)_{n\ge 0}$ in the unit ball of $\HC$ such that $b_{mn}=\left\langle x_n,y_m\right\rangle_\HC$. From this follows the so called Bennett criterion, stating that if $S_B$ is a bounded Schur multiplier, and the iterated limits $\lim_{m\to\infty}\lim_{n\to\infty}b_{mn}$ and $\lim_{n\to\infty}\lim_{m\to\infty}b_{mn}$ both exist, then the limits are equal.
Define an isometry $V:l^2\left(\N_0\right)\to H^2\left(\HC\right)$ by $Ve_n=e_n z^n$, and let $\left(E_{mn}\right)_{m,n\ge 0}$ be the scalar matrices defined by $\left\langle E_{mn}e_l,e_k\right\rangle_\HC=\delta_{mk}\delta_{nl}$. Given a scalar matrix $A=[a_{mn}]_{m,n\ge 0}$, we define the matrices $A_n=\sum_{k+l=n}a_{kl}E_{kl}$, and the function
\[
\phi\left(z\right)=\sum_{n=0}^\infty A_nz^n=\diag \left(z^k \right)_{k\ge 0}A\ \diag \left(z^l \right)_{l\ge 0}.
\]
From the above relations, $\left\|\phi\right\|_{H^\infty\left(\LC\right)}=\left\|A\right\|_\LC$. Now $\Gamma_\phi$ corresponds to the (operator-valued) Hankel matrix $X=[A_{k+l}]_{k,l\ge 0}$. A calculation shows that
\[
V^*D^\alpha\Gamma_{D^{-\alpha}\phi}V=S_{B}\left(A\right),
\]
where $B= \left[ \left(\frac{1+m}{1+m+n} \right)^\alpha\right]_{m,n\ge 0}$. It follows that
\[
\left\|D^\alpha\Gamma_{D^{-\alpha}\phi}\right\|_{\LC\left(H^2\left(\HC\right)\right)}\ge \left\|S_{B}\left(A\right)\right\|_{\LC\left(l^2\left(\N_0\right)\right)}.
\]
From Bennett's criterion, $S_B$ is not a bounded Schur multiplier, and so the right-hand side in the above inequality will be infinite for some choice of $A$. It follows that, for the same choice of $A$, $D^\alpha\Gamma_{D^{-\alpha}\phi}$ is not bounded on $H^2\left(\HC\right)$.
\section*{Acknowledgments}
The author expresses his gratitude to Sandra Pott, and Alexandru Aleman, for interesting discussions on the topics above, and also to Erik Wahlén, and the anonymous referee, for their useful comments on the presentation of this manuscript.
\bibliographystyle{amsplain}
\begin{bibdiv}
\begin{biblist}
\bib{Aleksandrov-Peller1996:HankOpsSimToContr}{article}{
author={Aleksandrov, A.~.B},
author={Peller, V.~V.},
title={Hankel operators and similarity to a contraction},
date={1996},
ISSN={1073-7928},
journal={Internat. Math. Res. Notices},
number={6},
pages={263\ndash 275},
url={http://dx.doi.org/10.1155/S1073792896000190},
review={\MR{1386078}},
}
\bib{Aleman-Perfekt2012:HankFrmsEmbThmsDirichletSpaces}{article}{
author={Aleman, A.},
author={Perfekt, K.-M.},
title={Hankel forms and embedding theorems in weighted {D}irichlet
spaces},
date={2012},
ISSN={1073-7928},
journal={Int. Math. Res. Not. IMRN},
number={19},
pages={4435\ndash 4448},
url={http://dx.doi.org/10.1093/imrn/rnr195},
review={\MR{2981715}},
}
\bib{Blasco1988:HardySpacesVecValDuality}{article}{
author={Blasco, O.},
title={Hardy spaces of vector-valued functions: duality},
date={1988},
ISSN={0002-9947},
journal={Trans. Amer. Math. Soc.},
volume={308},
number={2},
pages={495\ndash 507},
url={http://dx.doi.org/10.2307/2001088},
review={\MR{951618}},
}
\bib{Blasco1997:VecValBMOAGeomBSpaces}{article}{
author={Blasco, O.},
title={Vector-valued analytic functions of bounded mean oscillation and
geometry of {B}anach spaces},
date={1997},
ISSN={0019-2082},
journal={Illinois J. Math.},
volume={41},
number={4},
pages={532\ndash 558},
url={http://projecteuclid.org/euclid.ijm/1256068979},
review={\MR{1468865}},
}
\bib{Arregui-Blasco2002:MultplrsVecValBergmanSpaces}{article}{
author={Blasco, O.},
author={Arregui, J.-L.},
title={Multipliers on vector valued {B}ergman spaces},
date={2002},
ISSN={0008-414X},
journal={Canad. J. Math.},
volume={54},
number={6},
pages={1165\ndash 1186},
url={http://dx.doi.org/10.4153/CJM-2002-044-3},
review={\MR{1940234}},
}
\bib{Arregui-Blasco2003:BergmanBlochSpacesVecVal}{article}{
author={Blasco, O.},
author={Arregui, J.-L.},
title={Bergman and {B}loch spaces of vector-valued functions},
date={2003},
ISSN={0025-584X},
journal={Math. Nachr.},
volume={261/262},
pages={3\ndash 22},
url={http://dx.doi.org/10.1002/mana.200310109},
review={\MR{2020384}},
}
\bib{Blasco-Pott2008:EmbOpValDyadicBMO}{article}{
author={Blasco, O.},
author={Pott, S.},
title={Embeddings between operator-valued dyadic {BMO} spaces},
date={2008},
ISSN={0019-2082},
journal={Illinois J. Math.},
volume={52},
number={3},
pages={799\ndash 814},
url={http://projecteuclid.org/euclid.ijm/1254403715},
review={\MR{2546008}},
}
\bib{Bonsall1984:BddnessHankMat}{article}{
author={Bonsall, F.~F.},
title={Boundedness of {H}ankel matrices},
date={1984},
ISSN={0024-6107},
journal={J. London Math. Soc. (2)},
volume={29},
number={2},
pages={289\ndash 300},
url={http://dx.doi.org/10.1112/jlms/s2-29.2.289},
review={\MR{744100}},
}
\bib{Bourgain1986:SimProblPolBddOpsHSpace}{article}{
author={Bourgain, J.},
title={On the similarity problem for polynomially bounded operators on
{H}ilbert space},
date={1986},
ISSN={0021-2172},
journal={Israel J. Math.},
volume={54},
number={2},
pages={227\ndash 241},
url={http://dx.doi.org/10.1007/BF02764943},
review={\MR{852479}},
}
\bib{Bourgain1986:VecValSingIntsHardy-BMODualityChapter}{incollection}{
author={Bourgain, J.},
title={Vector-valued singular integrals and the {$H^1$}-{BMO} duality},
date={1986},
booktitle={Probability theory and harmonic analysis ({C}leveland, {O}hio,
1983)},
series={Monogr. Textbooks Pure Appl. Math.},
volume={98},
publisher={Dekker, New York},
pages={1\ndash 19},
review={\MR{830227}},
}
\bib{Buckley-Koskela-Vukotic1999:FracIntDiffBergmanSpaces}{article}{
author={Buckley, S.~.M},
author={Koskela, P.},
author={Vukoti{\'c}, D.},
title={Fractional integration, differentiation, and weighted {B}ergman
spaces},
date={1999},
ISSN={0305-0041},
journal={Math. Proc. Cambridge Philos. Soc.},
volume={126},
number={2},
pages={369\ndash 385},
url={http://dx.doi.org/10.1017/S030500419800334X},
review={\MR{1670257}},
}
\bib{Bukhvalov-Danilevich1982:BdryPropsAnalHarmFcnsValBSpace}{article}{
author={Bukhvalov, A.~V.},
author={Danilevich, A.~A.},
title={Boundary properties of analytic and harmonic functions with
values in a {B}anach space},
language={Russian},
date={1982},
ISSN={0025-567X},
journal={Mat. Zametki},
volume={31},
number={2},
pages={203\ndash 214, 317},
review={\MR{649004}},
}
\bib{Carleson1958:InterpolProblBddAnalFcns}{article}{
author={Carleson, L.},
title={An interpolation problem for bounded analytic functions},
date={1958},
ISSN={0002-9327},
journal={Amer. J. Math.},
volume={80},
pages={921\ndash 930},
review={\MR{0117349}},
}
\bib{Carleson1962:InterpolBddAnalFcnsCoronaProbl}{article}{
author={Carleson, L.},
title={Interpolations by bounded analytic functions and the corona
problem},
date={1962},
ISSN={0003-486X},
journal={Ann. of Math. (2)},
volume={76},
pages={547\ndash 559},
review={\MR{0141789}},
}
\bib{Cohn-Verbitsky2000:FactTentSpacesHankOps}{article}{
author={Cohn, W.~S.},
author={Verbitsky, I.~E.},
title={Factorization of tent spaces and {H}ankel operators},
date={2000},
ISSN={0022-1236},
journal={J. Funct. Anal.},
volume={175},
number={2},
pages={308\ndash 329},
url={http://dx.doi.org/10.1006/jfan.2000.3589},
review={\MR{1780479}},
}
\bib{Davidson-Paulsen1997:PolBddOps}{article}{
author={Davidson, K.~R.},
author={Paulsen, V.~I.},
title={Polynomially bounded operators},
date={1997},
ISSN={0075-4102},
journal={J. Reine Angew. Math.},
volume={487},
pages={153\ndash 170},
review={\MR{1454263}},
}
\bib{Diestel-Uhl1977:VecMeasures}{book}{
author={Diestel, J.},
author={Uhl, J.~J.~Jr.},
title={Vector measures},
publisher={American Mathematical Society, Providence, R.I.},
date={1977},
note={With a foreword by B. J. Pettis, Mathematical Surveys, No. 15},
review={\MR{0453964}},
}
\bib{Fefferman1971:CharBMO}{article}{
author={Fefferman, C.},
title={Characterizations of bounded mean oscillation},
date={1971},
ISSN={0002-9904},
journal={Bull. Amer. Math. Soc.},
volume={77},
pages={587\ndash 588},
review={\MR{0280994}},
}
\bib{Fefferman-Stein1972:HpSpaces}{article}{
author={Fefferman, C.},
author={Stein, E.~M.},
title={{$H^{p}$} spaces of several variables},
date={1972},
ISSN={0001-5962},
journal={Acta Math.},
volume={129},
number={3-4},
pages={137\ndash 193},
review={\MR{0447953}},
}
\bib{Flett1972:DualIneqHardyLittlewood}{article}{
author={Flett, T.~M.},
title={The dual of an inequality of {H}ardy and {L}ittlewood and some
related inequalities},
date={1972},
ISSN={0022-247x},
journal={J. Math. Anal. Appl.},
volume={38},
pages={746\ndash 765},
review={\MR{0304667}},
}
\bib{Foguel1964:CounterExSz.-NagyProbl}{article}{
author={Foguel, S.~R.},
title={A counterexample to a problem of {S}z.-{N}agy},
date={1964},
ISSN={0002-9939},
journal={Proc. Amer. Math. Soc.},
volume={15},
pages={788\ndash 790},
review={\MR{0165362}},
}
\bib{Garnett2007:BddAnalFcnsBook}{book}{
author={Garnett, J.~B.},
title={Bounded analytic functions},
edition={first revised},
series={Graduate Texts in Mathematics},
publisher={Springer, New York},
date={2007},
volume={236},
ISBN={978-0-387-33621-3; 0-387-33621-4},
review={\MR{2261424}},
}
\bib{Gillespie-Pott-Treil-Volberg2004:LogGrowthHilbTransfVecHank}{article}{
author={Gillespie, T.~A.},
author={Pott, S.},
author={Treil, S.},
author={Volberg, A.},
title={Logarithmic growth for weighted {H}ilbert transforms and vector
{H}ankel operators},
date={2004},
ISSN={0379-4024},
journal={J. Operator Theory},
volume={52},
number={1},
pages={103\ndash 112},
review={\MR{2091462}},
}
\bib{Haagerup-Pisier1989:FactAnalFcnsNon-CommL1Spaces}{article}{
author={Haagerup, U.},
author={Pisier, G.},
title={Factorization of analytic functions with values in noncommutative
{$L_1$}-spaces and applications},
date={1989},
ISSN={0008-414X},
journal={Canad. J. Math.},
volume={41},
number={5},
pages={882\ndash 906},
url={http://dx.doi.org/10.4153/CJM-1989-041-6},
review={\MR{1015588}},
}
\bib{Halmos1970:TenProbls}{article}{
author={Halmos, P.~R.},
title={Ten problems in {H}ilbert space},
date={1970},
ISSN={0002-9904},
journal={Bull. Amer. Math. Soc.},
volume={76},
pages={887\ndash 933},
review={\MR{0270173}},
}
\bib{Hedenmalm-Korenblum-Zhu2000:BergmanSpacesBook}{book}{
author={Hedenmalm, H.},
author={Korenblum, B.},
author={Zhu, K.},
title={Theory of {B}ergman spaces},
series={Graduate Texts in Mathematics},
publisher={Springer-Verlag, New York},
date={2000},
volume={199},
ISBN={0-387-98791-6},
url={http://dx.doi.org/10.1007/978-1-4612-0497-8},
review={\MR{1758653}},
}
\bib{Jacob-Rydhe-Wynn2014:WeightWeissConjRKTGenHankOps}{article}{
author={Jacob, B.},
author={Rydhe, E.},
author={Wynn, A.},
title={The weighted {W}eiss conjecture and reproducing kernel theses for
generalized {H}ankel operators},
date={2014},
ISSN={1424-3199},
journal={J. Evol. Equ.},
volume={14},
number={1},
pages={85\ndash 120},
url={http://dx.doi.org/10.1007/s00028-013-0209-z},
review={\MR{3169032}},
}
\bib{Janson-Peetre1988:Paracomms}{article}{
author={Janson, S.},
author={Peetre, J.},
title={Paracommutators---boundedness and {S}chatten-von {N}eumann
properties},
date={1988},
ISSN={0002-9947},
journal={Trans. Amer. Math. Soc.},
volume={305},
number={2},
pages={467\ndash 504},
url={http://dx.doi.org/10.2307/2000875},
review={\MR{924766}},
}
\bib{Kislyakov2000:OpsDisSimContr}{article}{
author={Kislyakov, S.~V.},
title={Operators that are (dis)similar to a contraction: {P}isier's
counterexample in terms of singular integrals},
language={Russian},
date={1997},
ISSN={0373-2703},
journal={Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov.
(POMI)},
volume={247},
number={Issled. po Linein. Oper. i Teor. Funkts. 25},
pages={79\ndash 95, 300},
url={http://dx.doi.org/10.1007/BF02673734},
review={\MR{1692632}},
}
\bib{Mei2006:MatValParaprods}{article}{
author={Mei, T.},
title={Notes on matrix valued paraproducts},
date={2006},
ISSN={0022-2518},
journal={Indiana Univ. Math. J.},
volume={55},
number={2},
pages={747\ndash 760},
url={http://dx.doi.org/10.1512/iumj.2006.55.2926},
review={\MR{2225451}},
}
\bib{Nazarov-Pisier-Treil-Volberg2002:EstsVecCarlesonEmbThmVecParaprods}{article}{
author={Nazarov, F.},
author={Pisier, G.},
author={Treil, S.},
author={Volberg, A.},
title={Sharp estimates in vector {C}arleson imbedding theorem and for
vector paraproducts},
date={2002},
ISSN={0075-4102},
journal={J. Reine Angew. Math.},
volume={542},
pages={147\ndash 171},
url={http://dx.doi.org/10.1515/crll.2002.004},
review={\MR{1880830}},
}
\bib{Nazarov-Treil-Volberg1997:CounterExInfDimCarlesonEmbThm}{article}{
author={Nazarov, F.},
author={Treil, S.},
author={Volberg, A.},
title={Counterexample to the infinite-dimensional {C}arleson embedding
theorem},
date={1997},
ISSN={0764-4442},
journal={C. R. Acad. Sci. Paris S\'er. I Math.},
volume={325},
number={4},
pages={383\ndash 388},
url={http://dx.doi.org/10.1016/S0764-4442(97)85621-2},
review={\MR{1467091}},
}
\bib{Nehari1957:BddBilinFrms}{article}{
author={Nehari, Z.},
title={On bounded bilinear forms},
date={1957},
ISSN={0003-486X},
journal={Ann. of Math. (2)},
volume={65},
pages={153\ndash 162},
review={\MR{0082945}},
}
\bib{Nikolski2002:EasyReading}{book}{
author={Nikolski, N.~K.},
title={Operators, functions, and systems: an easy reading. {V}ol. {I}},
series={Mathematical Surveys and Monographs},
publisher={American Mathematical Society, Providence, RI},
date={2002},
volume={92},
ISBN={0-8218-1083-9},
note={Hardy, Hankel, and Toeplitz, Translated from the French by
Andreas Hartmann},
review={\MR{1864396}},
}
\bib{Page1970:BddCompctVecHankOps}{article}{
author={Page, L.~B.},
title={Bounded and compact vectorial {H}ankel operators},
date={1970},
ISSN={0002-9947},
journal={Trans. Amer. Math. Soc.},
volume={150},
pages={529\ndash 539},
review={\MR{0273449}},
}
\bib{Paulsen1984:ComplPolBddSimContr}{article}{
author={Paulsen, V.~I.},
title={Every completely polynomially bounded operator is similar to a
contraction},
date={1984},
ISSN={0022-1236},
journal={J. Funct. Anal.},
volume={55},
number={1},
pages={1\ndash 17},
url={http://dx.doi.org/10.1016/0022-1236(84)90014-4},
review={\MR{733029}},
}
\bib{Paulsen2002:ComplBddMapsOpAlgs}{book}{
author={Paulsen, V.~I.},
title={Completely bounded maps and operator algebras},
series={Cambridge Studies in Advanced Mathematics},
publisher={Cambridge University Press, Cambridge},
date={2002},
volume={78},
ISBN={0-521-81669-6},
review={\MR{1976867}},
}
\bib{Peller1982:EstsFcnsPwrBddOpsHSpace}{article}{
author={Peller, V.~V.},
title={Estimates of functions of power bounded operators on {H}ilbert
spaces},
date={1982},
ISSN={0379-4024},
journal={J. Operator Theory},
volume={7},
number={2},
pages={341\ndash 372},
review={\MR{658618}},
}
\bib{Peller1982:VecHankOps}{article}{
author={Peller, V.~V.},
title={Vectorial {H}ankel operators, commutators and related operators
of the {S}chatten-von {N}eumann class {$\gamma _{p}$}},
date={1982},
ISSN={0378-620X},
journal={Integral Equations Operator Theory},
volume={5},
number={2},
pages={244\ndash 272},
url={http://dx.doi.org/10.1007/BF01694041},
review={\MR{647702}},
}
\bib{Peller2003:HankOpsBook}{book}{
author={Peller, V.~V.},
title={Hankel operators and their applications},
series={Springer Monographs in Mathematics},
publisher={Springer-Verlag, New York},
date={2003},
ISBN={0-387-95548-8},
url={http://dx.doi.org/10.1007/978-0-387-21681-2},
review={\MR{1949210}},
}
\bib{Pisier1997:PolBddNotSim}{article}{
author={Pisier, G.},
title={A polynomially bounded operator on {H}ilbert space which is not
similar to a contraction},
date={1997},
ISSN={0894-0347},
journal={J. Amer. Math. Soc.},
volume={10},
number={2},
pages={351\ndash 369},
url={http://dx.doi.org/10.1090/S0894-0347-97-00227-0},
review={\MR{1415321}},
}
\bib{Rosenblum-Rovnyak1985:HardyClassesOpTheory}{book}{
author={Rosenblum, M.},
author={Rovnyak, J.},
title={Hardy classes and operator theory},
series={Oxford Mathematical Monographs},
publisher={The Clarendon Press, Oxford University Press, New York},
date={1985},
ISBN={0-19-503591-7},
note={Oxford Science Publications},
review={\MR{822228}},
}
\bib{Rydhe2016:CharTriebel-LizorkinSpaces}{article}{
author={Rydhe, E.},
title={On the characterization of triebel--lizorkin type space of
analytic functions},
date={2016},
journal={arXiv:1609.09229},
status={preprint},
}
\bib{Rydhe2016:CounterExsCarlesonEmbThm}{article}{
author={Rydhe, E.},
title={Two more counterexamples to the infinite-dimensional carleson
embedding theorem},
date={2016},
journal={arXiv:1608.06728},
status={preprint},
}
\bib{Sarason1967:GenInterpol}{article}{
author={Sarason, D.},
title={Generalized interpolation in {$H^{\infty }$}},
date={1967},
ISSN={0002-9947},
journal={Trans. Amer. Math. Soc.},
volume={127},
pages={179\ndash 203},
review={\MR{0208383}},
}
\bib{Sz.-Nagy1959:ComplContOpsUniformlyBddIterates}{article}{
author={Sz.-Nagy, B.},
title={Completely continuous operators with uniformly bounded iterates},
language={Hungarian},
date={1959},
journal={Magyar Tud. Akad. Mat. Kutat\'o Int. K\"ozl.},
volume={4},
pages={89\ndash 93},
review={\MR{0108722}},
}
\bib{Wojtaszczyk1991:BSpacesForAnalysts}{book}{
author={Wojtaszczyk, P.},
title={Banach spaces for analysts},
series={Cambridge Studies in Advanced Mathematics},
publisher={Cambridge University Press, Cambridge},
date={1991},
volume={25},
ISBN={0-521-35618-0},
url={http://dx.doi.org/10.1017/CBO9780511608735},
review={\MR{1144277}},
}
\end{biblist}
\end{bibdiv}
\end{document}
|
\section{{\texttt{Magma}} code}
In this appendix we provide elementary codes that can be used with
\verb|Magma| to compute monodromy for (\ref{eqn: D-S-fibration}), i.e. the short exact sequences
$$ 1 \to F_{k(n)} \to \prod_{K^1}^n G_i \to \prod_{i=1}^n G_i \to 1 $$
for the cases $n\geq 3$. We can get the monodromy representation in this setting
for any finitely presentable finite groups by suitably changing these codes below,
which are given for Examples \ref{ex: 1} and \ref{ex: 2}.
{{\footnotesize
\begin{verbatim}
>/* ------------ FIRST EXAMPLE -------- */
>
> F<a,b,c>:=FreeGroup(3);
> J<a,b,c>:=quo<F|a^2,b^2,c^2>;
> C:=CommutatorSubgroup(J);
> G:=Rewrite(J,C);
Finitely presented group on 5 generators
Index in group J is 8 = 2^3
Generators as words in group J
G.1 = (b * a)^2
G.2 = (c * a)^2
G.3 = (c * b)^2
G.4 = a * c * b * c * b * a
G.5 = b * c * a * c * b * a
> /* How to compute the matrices in general:*/
>
> Conja:=[a* G.i * a^-1 : i in [1..5] ];
> Conjb:=[b* G.i * b^-1 : i in [1..5] ];
> Conjc:=[c* G.i * c^-1 : i in [1..5] ];
>
> /* NOW THE MATRICES FOR Conjx, x=a,b,c*/
>
> Conja1:=[G ! t : t in Conja ];
> Conjb1:=[G ! t : t in Conjb ];
> Conjc1:=[G ! t : t in Conjc ];
>
> Ra := Matrix(Integers(), [ [ Weight(r, G.j) : j in [1..5] ] : r in Conja1 ]);
> Rb := Matrix(Integers(), [ [ Weight(r, G.j) : j in [1..5] ] : r in Conjb1 ]);
> Rc := Matrix(Integers(), [ [ Weight(r, G.j) : j in [1..5] ] : r in Conjc1 ]);
>
>/* ------------ SECOND EXAMPLE -------- */
>
> F<a,b,c,d>:=FreeGroup(4);
> J1<a,b,c,d>:=quo<F|a^2,b^2,c^2,d^2,(a,b),(b,c),(c,d)>;
> C:=CommutatorSubgroup(J1);
> G:=Rewrite(J1,C);
> G;
Finitely presented group G on 5 generators
Generators as words in group J1
G.1 = (c * a)^2
G.2 = (d * a)^2
G.3 = (d * b)^2
G.4 = a * d * b * d * b * a
G.5 = c * d * a * d * c * a
> Conja:=[a * G.i * a^-1 : i in [1..5] ];
> Conjb:=[b * G.i * b^-1 : i in [1..5] ];
> Conjc:=[c * G.i * c^-1 : i in [1..5] ];
> Conjd:=[d * G.i * d^-1 : i in [1..5] ];
>
> Conja1:=[G ! t : t in Conja ];
> Conjb1:=[G ! t : t in Conjb ];
> Conjc1:=[G ! t : t in Conjc ];
> Conjd1:=[G ! t : t in Conjd ];
>
>/* NOW THE MATRICES FOR Conjx, x=a,b,c,d*/
>
> Ra := Matrix(Integers(), [ [ Weight(r, G.j) : j in [1..5] ] : r in Conja1 ]);
> Rb := Matrix(Integers(), [ [ Weight(r, G.j) : j in [1..5] ] : r in Conjb1 ]);
> Rc := Matrix(Integers(), [ [ Weight(r, G.j) : j in [1..5] ] : r in Conjc1 ]);
> Rd := Matrix(Integers(), [ [ Weight(r, G.j) : j in [1..5] ] : r in Conjd1 ]);
\end{verbatim}
}}
For related functions the handbook of \verb|Magma| functions \cite{magmahb3} is recommended.
\section{Introduction}
Studying the topology of a fibration sequence frequently involves understanding the monodromy action,
which is the action of the fundamental group of the base space on the fibre.
For example monodromy is important in topology when using the
Lyndon-Hochschild-Serre spectral sequence to compute the (co)homology with coefficients
in that representation. The homology of the fibre inherits a module structure
induced by the action of the fundamental group of the base on the fibre.
When using the spectral sequence to obtain the
homology of the fibre, then one has to consider the homology of the base with
coefficients in the homology of the fibre regarded as a module,
called the homology of the base with local coefficients.
Here the monodromy representation for a fibration $p:E\to B$ with
fibre $F$ will mean the representation
$\rho: \pi_1(B) \to {\rm Out}(H_\ast(F)).$
The goal of this paper is to study polyhedral products in connection with
monodromy representations for fibrations that arise naturally in
the field of \textit{toric topology}.
The problem of explicitly describing such representations was started in
\cite{stafa.monodromy}. The main idea is to give
geometric descriptions in terms of polyhedral products
for the spaces of the fibrations under consideration.
Numerous results on graph products of groups and polyhedral products demonstrate that the
underlying simplicial complex $K$ plays an important role in their study, and is not
only a convenient way to describe graph groups and polyhedral products.
For example Droms \cite{droms1987isomorphisms} proved that two graph groups
are isomorphic if and only if the graphs are isomorphic.
Servatius \textit{etal.} \cite{servatius1989surface}
determine the simplicial complexes for which the commutator subgroup of a
right-angled Artin groups is free.
Moreover, if $K$ is chosen carefully, one obtains classifying spaces
for various important families of discrete groups, including right-angled Artin
and Coxeter groups from geometric group theory \cite{stafa.monodromy,davis.okun}.
In another application Grbi\'c, Panov, Theriault and Wu \cite{wu.grbic.panov}
give conditions on the 1-skeleton of $K$ that determine when the face ring of $K$
is a Golod ring, or equivalently the corresponding moment-angle complex
has the homotopy type of a wedge of spheres.
Recently Panov and Veryovkin \cite{panov2016polyhedral} studied
polyhedral products that have the homotopy type of classifying spaces of right-angled
Artin groups and right-angled Coxeter groups.
\
In the present article we study further properties of the monodromy representations
associated to the fibration sequences
\begin{equation}\label{eqn: D-S-fibration INTRO1}
(\underline{EG},\underline{G})^K \to (\underline{BG},\underline{1})^K
\to \prod_{i=1}^n BG_i.
\end{equation}
Each space in \eqref{eqn: D-S-fibration INTRO1} is a polyhedral product,
depending on a simplicial complex $K$, together with a sequence of
finite groups $\underline{G}:=\{G_1,\dots,G_n\}$, their classifying spaces
$\underline{BG}:=\{BG_1,\dots,BG_n\}$, and corresponding universal
covers $\underline{EG}:=\{EG_1,\dots,EG_n\}$,
see Definition \ref{defn: polyhedral product}.
We give explicit descriptions of monodromy representations for simplicial
complexes $K$ with more than two vertices, which were described
geometrically in \cite{stafa.monodromy}.
To do this we generalize and use some results of
Panov and Veryovkin \cite{panov2016polyhedral}.
We give applications, in particular to spaces of commuting elements in
\textit{commutative transitive (CT) finite groups}, where commutativity
is a transitive relation, studied in a
celebrated paper of M. Suzuki \cite{suzuki1957nonexistence},
and to a problem related to the Feit-Thompson theorem,
which states that all groups of odd order are solvable.
Finally, we give give examples, which can be generalized easily,
using the \verb|Magma| \cite{magma} code included in the appendix.
\subsection*{Main results}
For given finite discrete groups $G_1,\dots, G_n$, we use polyhedral
products to construct monodromy representations
$ \Phi: G_1 \times \cdots \times G_n \to {\rm Out}(F_N) $
into outer automorphism groups of free groups. In particular, we obtain explicit
faithful representations of graph products of
finite groups into automorphism groups of free groups,
and faithful representations of their direct products into
linear groups ${{\rm SL}}(k,{\mathbb{Z}})$ or ${{\rm GL}}(k,{\mathbb{Z}})$.
This article presents a machinery based on polyhedral products to achieve this.
The first result is the following theorem.
\begin{thm}
Let $G_1,\dots, G_n$ be finite groups and $K$ a simplicial complex
with $n$ vertices with 1-skeleton $K^1$ a chordal graph.
Then there are faithful representations
$$\Theta_K: \prod_{K^1} G_i \to {\rm Aut}(F_{\rho_K})$$
and faithful monodromy representations
$$\Phi_K: G_1\times \cdots \times G_n \to {\rm Out}(F_{\rho_K}),$$
where $\rho_K$ is the rank of the fundamental group of the fibre in equation
(\ref{eqn: D-S-fibration INTRO1}).
\end{thm}
The case when the groups $G_1,\dots,G_n$ are abelian the representations
can be described explicitely and convenient models of polyhedral products can
then be used to show that the corresponding monodromy representations
obtained for non-abelian finite groups are also faithful.
\begin{thm}
Let $G_1,\dots, G_n$ be finite abelian groups. Then the faithful
monodromy representation $\Phi_K$ induces a faithful representation
$$\Phi_K: G_1\times\cdots \times G_n \to {{\rm SL}}({\rho_K},{\mathbb{Z}}).$$
If $G_1,\dots,G_n$ are non-abelian then $\Phi_K$ maps into ${{\rm GL}}({\rho_K},{\mathbb{Z}})$.
\end{thm}
Let $E(2,G)\subseteq EG$ and $B(2,G) \subseteq BG$ be the spaces
defined in $\S$\ref{sec: B(2,G) and E(2,G)} that classify commuting elements in
a group $G$. In particular, we use polyhedral products to study the class of
finite transitively commutative (CT) groups, a class of groups where commtutativity
is transitive. The following theorem is then an application of
polyhedral products to group theory.
\begin{thm}\label{thm: top. equiv. form CT groups INTRO}
Finite CT groups with trivial center are solvable if and only if the induced map
$H_1(E(2,G);{\mathbb{Z}}) \to H_1(B(2,G);{\mathbb{Z}})$ is not surjective.
\end{thm}
This theorem is motivated from a result of Adem, Cohen and Torres-Giese
\cite{fredb2g}, which states an equivalent topological condition to the the
Feit-Thompson theorem, namely that the theorem is true if and
only if the map $H_1(E(2,G);{\mathbb{Z}}) \to H_1(B(2,G);{\mathbb{Z}})$ is not surjective.
\subsection*{Acknowledgments} The author thanks Alina Vdovina
(University of Newcastle, UK), who was visiting the Institute
for Mathematical Research at ETH Z\"urich, during the spring
semester 2016, for our numerous conversations
on the topic, for suggesting me to use \verb|Magma|
for some computations, and for providing the first codes.
\section{Polyhedral products and related fibrations}\label{sec: polyhedral products and fibrations}
Polyhedral products are a generalized version of {\it moment angle complexes}
that appear in the work of Buchstaber and Panov \cite{buchstaber2002torus}
in the context of \textit{toric topology}. Polyhedral products were also studied
for instance by Bahri, Bendersky, Cohen and Gitler \cite{cohen.macs} and many others,
and are the main objects of study in toric topology, see the recent monograph
by Buchstaber and Panov \cite{bp2015}.
\begin{defn}\label{defn: polyhedral product}
Let $(\underline{X},\underline{A} )$ denote a sequence of
pointed $CW$-pairs $\{(X_i,A_i)\}_{i=1}^n$
and $[n]$ denote the sequence of integers $\{1,2,\dots,n\}$.
\begin{itemize}
\item A \textbf{simplex} $\sigma$ is given by an increasing
sequence of integers $\sigma = \{1 \leq i_1 < \cdots < i_q \leq n\}
\subseteq [n]$.
A {\bf simplicial complex} $K$ is a collection of simplices such that
if $\tau\subset \sigma$ and $\sigma \in K,$ then $\tau \in K.$
In particular $\emptyset \in K$.
\item The \textbf{polyhedral product} $(\underline{X},\underline{A})^K$
is the subspace of the product $X_1 \times \cdots \times X_n$
given by the colimit
$$
(\underline{X},\underline{A})^K :=
\underset{{\sigma \in K}}{{{\rm colim}}}\, \mathcal D(\sigma)
= \bigcup_{\sigma \in K} \mathcal D(\sigma) \subseteq \prod_{i=1}^n X_i,
$$
where $\mathcal D(\sigma)=\{(x_1,\dots,x_n)\in \prod_{i=1}^n X_i |
x_i\in A_i \text{ if } i \notin \sigma\}$,
the maps are the inclusions, and the topology is the
subspace topology of the product. Another standard notation for
polyhedral products is $Z_K(\underline{X},\underline{A})$.
Sometimes polyhedral products are called \textbf{$K$-powers}.
Since $\emptyset$ is in any $K$, we have
$\prod_{i=1}^n A_i \subset (\underline{X},\underline{A})^K
\subset \prod_{i=1}^n X_i.$
\item If the pairs $(X,A)$ are $(D^2,S^1)$ or $(D^1,S^0)$, then the polyhedral
products are called \textbf{moment-angle complexes} and
\textbf{real moment-angle complexes}, respectively.
\item If all the pairs in the sequence $(X_1,A_1),\dots,(X_n,A_n)$
are equal to $(X,A)$, then we
omit the underline in the notation of the polyhedral product, and write simply
$({X},{A})^K$.
\end{itemize}
\end{defn}
\begin{defn}\label{defn: many definitions}
Next we give some relevant definitions and notation:
\begin{itemize}
\item For a simplicial complex $K$, the complex $K^i$ denotes
the \textbf{$i$-skeleton} of $K$.
\item A simplicial complex $K$ is called a \textbf{flag complex}
if for any complete subgraph $\Gamma \subset K^1$,
it also contains the simplex spanned by these vertices.
The structure of moment angle complexes (or polyhedral products in general)
is better understood when $K$ is a flag complex \cite[\S 8.5]{bp2015}.
\item For a simplicial complex $K$, let ${{\rm Flag}}(K)$
denote the \textbf{clique complex} of $K^1$,
i.e. the simplicial complex whose simplices are
complete subgraphs of $K^1$.
For example a flag complex is the clique complex of its 1-skeleton.
\item A graph is called \textbf{chordal} if every cycle of length greater than three
has an edge (called a \textit{chord}) connecting two nonconsecutive vertices.
Chordal graphs are also called \textbf{triangulated graphs}, or we say they
have \textbf{perfect elimination ordering.}
\item For any group $G$ denote its abelianization by $\mathscr A(G):=G/[G,G]$, and the
abelianization map by ${\rm ab}_G: G\twoheadrightarrow \mathscr A(G)$.
\item Let $G_1,\dots,G_n$ be a sequence of groups and $\Gamma$
a simplicial graph on $[n]$. The \textbf{graph product} of
$G_1,\dots,G_n$ over $\Gamma$ is the quotient of their
free product by the normal closure of the relations
$R_{\Gamma}:=\{[g_i,g_i]: \{i,j\}\text{ is an edge in }\Gamma\}.$
The group obtained this way will be called a
\textbf{graph group}, even though in the literature
this name is sometimes used for right-angled Artin
groups. We denote it by
$$\prod_\Gamma G_i := (G_1\ast \cdots \ast G_n)/\langle R_{\Gamma} \rangle.$$
In this notation right-angled Artin groups are graph products of the group ${\mathbb{Z}}$.
\end{itemize}
\end{defn}
\begin{ex}\label{example: polyhedral products}
\
\begin{enumerate}[label=\arabic*.]
\item Let $X$ be the unit interval $[0,1]$ and $A\subset [0,1]$ be the subset
$\{0,1/2,1\}$. Let $K$ be the simplicial complex consisting of only
two vertices $\{\{v_1\},\{v_2\}\}$. Then
$\mathcal D(\{v_1\})=X\times A \subset [0,1]^2$
and $\mathcal D(\{v_2\})=A\times X \subset [0,1]^2$.
Therefore, $(X,A)^K= \mathcal D(\{v_1\}) \cup \mathcal D(\{v_2\})
= X\times A \cup A\times X$ is a graph inside the square $[0,1]^2$,
homotopy equivalent to a wedge of 4 circles $\bigvee_4 S^1$.
Similarly, we can choose $A$ to be any finite subset of the unit interval
and we obtain similar graphs.
\item Let $(X,A)=(D^2,S^1)$. If $K$ is a the boundary of the $n$-simplex
then the moment-angle complex
$(D^2,S^1)^K = \bigcup_{\sigma_i} \mathcal D(\sigma_i)$
is homeomorphic to the sphere $S^{2n+1}=\partial D^{2(n+1)}$.
It is also known \cite[Theorem 4.6]{wu.grbic.panov} that
if $K$ is a flag complex, then $(D^2,S^1)^K$ has the homotopy
of a wedge of spheres if and only if $K^1$ is chordal.
\item If $K$ is any simplicial complex on $n$ vertices, and
$\underline{\ast}=\{\ast_1,\dots,\ast_n\}$ is the sequence of basepoints
then $\underline{X}$ then $(\underline{X},\underline{\ast})^{K^0}
=\bigvee_{i=1}^n X_i.$ Therefore, in general
$\bigvee_{i=1}^n X_i \subseteq (\underline{X},\underline{\ast})^{K}
\subseteq \prod_{i=1}^n X_i$.
\end{enumerate}
\end{ex}
Let $G$ be a topological group with basepoint its identity element $\ast=1$, $BG$ be the
classifying space of $G$, and $EG$ be a contractible space with a free action of $G$
such that the quotient map $EG \to BG$ is a principal $G$-bundle.
G. Denham and A. Suciu \cite[Lemma 2.3.2]{denham} gave a natural fibration relating
the polyhedral product for the pair $(BG,1)$ to the polyhedral product for the pair $(EG,G)$.
That is, for a simplicial complex $K$ with $n$ vertices, the polyhedral product $(BG,1)^K$
fibres over the product $(BG)^n$ as follows
\begin{equation}\label{eqn: D-S-fibration 1}
(EG,G)^K \to (EG)^n \times_{G^n} (EG,G)^K \to (BG)^n ,
\end{equation}
where the total space is homotopy equivalent to $(BG,1)^K$. Note that the group $G$ acts
coordinate--wise on the fibre $(EG,G)^K \subset (EG)^n$.
This fibration is a generalization of the Davis-Januszkiewicz space
\cite{davis.januszckiewicz}, with topological group the circle $S^1$,
given by the Borel construction
$$\mathcal{DJ}(K) = (ES^1)^n \times_{(S^1)^n} (ES^1,S^1)^K,$$
which describes a cellular realization of the Stanley-Reisner ring of $K$,
in the sense that the cohomology ring of $\mathcal{DJ}(K)$ is precisely
the Stanley-Reisner ring of $K$ defined as the quotient of the polynomial ring
$R[K]=R[x_1,\dots,x_n]/I_K$ by the Stanley-Reisner ideal
$I_K=\langle x_{i_1}\cdots x_{i_t}| \{i_1,\dots,i_t\} \neq K \rangle$.
A later result of V. Buchstaber and T. Panov \cite{buch.panov} showed that
$\mathcal{DJ}(K)\simeq (BS^1,1)^K$ and
the homotopy fibre of the natural inclusion $\mathcal{DJ}(K) \to ({\mathbb{C}} P^\infty)^n$ is
equivalent to the polyhedral product $(ES^1,S^1)^K$.
If $G_1,\dots,G_n$ is a sequence of topological groups, then for any
simplicial complex $K$ with $n$ vertices, the fibration sequence
(\ref{eqn: D-S-fibration 1}) can be generalized to obtain
\begin{equation}\label{eqn: D-S-fibration}
(\underline{EG},\underline{G})^K \to (\underline{BG},\underline{1})^K \to \prod_{i=1}^n BG_i.
\end{equation}
Similarly, the fundamental group of the base space acts naturally coordinate-wise
on the homotopy fibre. The monodromy representation of this fibration is the
main object of study in this article.
Note that the homotopy type of a polyhedral product depends only on the relative
homotopy type of the pairs $(X,A)$, as observed in \cite{denham}.
We are mainly interested in the cases when $G_1,\dots,G_n$ are finite discrete groups.
If the pairs $(\underline{EG},\underline{G})$
are replaced by $(\underline{I},\underline{F})$, where $I$ is the unit interval and $F_i\subset I$
has the cardinality of $G_i$, then there is a homotopy equivalence
$(\underline{EG},\underline{G})^K\simeq (\underline{I},\underline{F})^K$.
Moreover, if $K=K^0$ is the 0-skeleton of $K$,
then it follows from Example \ref{example: polyhedral products} that
$(\underline{BG},\underline{1})^{K^0}=BG_1 \vee \cdots \vee BG_n$.
The homotopy fibre $(\underline{I},\underline{F})^{K^0}$ has the
homotopy type of a finite wedge of circles
$(\underline{I},\underline{F}) ^{K^0}\simeq \bigvee_{\rho_{K^0}} S^1, $
as shown in \cite{stafa.fund.gp}, where
\begin{equation}\label{eqn: rank of kernel for chordal graph}
\rho_{K^0}=(n-1)\prod_{i=1}^n |G_i| - \sum_{i=1}^n (\prod_{j\neq i} |G_j|)+1.
\end{equation}
This gives a topological proof of a classical theorem of J. Nielsen
\cite{nielsen} concerning the rank of the free group in the following
short exact sequence of groups
$$ 1 \to F_{\rho_{K^0}} \to G_1\ast \cdots \ast G_n \to \prod_{1\leq i \leq n} G_i \to 1.$$
Therefore, the rank $\rho_{K^0}$ depends only on the order of the
groups $G_1,\dots,G_n$, and not on their group structure.
For simplicity we simply write $\rho_{K^0}$ for
the rank, when the orders of $G_i$ are clear from the context.
More generally, it was shown in \cite{stafa.fund.gp} that if $G_1,\dots,G_n$ are countable discrete
groups then the spaces in the fibration sequence (\ref{eqn: D-S-fibration})
are Eilenberg-MacLane spaces of the type $K(\pi,1)$ if and only if $K$ is a flag complex.
Moreover, for any $K$, the fundamental group of the polyhedral product
$(\underline{BG},\underline{1})^K$ is determined
by the 1-skeleton $K^1$ and is isomorphic to
$\pi_1((\underline{BG},\underline{1})^K) \cong \prod_{K^1} G_i$,
the graph product of the groups $G_1,\dots,G_n$.
Hence there is a short exact sequence of groups
\begin{equation}\label{eqn: s.e.s.}
1 \to \pi_1((\underline{EG},\underline{G})^K) \to
\prod_{K^1} G_i \to \prod_{1\leq i \leq n} G_i \to 1.
\end{equation}
We want to study simplicial complexes $K$ for which the kernel of the
short exact sequence above is a free group. The following theorem
shows exactly which simplicial complexes $K$ have this property.
\begin{thm}
Let $G_1,\dots,G_n$ be (countable) discrete groups and $K$ be a flag
complex on $n$ vertices. Then $(\underline{EG},\underline{G})^K$
has the homotopy type of a graph if and only if $K^1$ is a chordal graph.
\end{thm}
\begin{proof}
It was shown in \cite[Theorem 1.1]{stafa.fund.gp} that
the space $(\underline{EG},\underline{G})^K$ is a $K(\pi,1)$ if and only if
$K$ is a flag complex. Therefore, we get the short exact sequence of groups
in (\ref{eqn: s.e.s.}). Panov and Veryovkin \cite[Theorem 4.3]{panov2016polyhedral}
showed that $\pi_1((\underline{EG},\underline{G})^K)$ is free if and only if
the graph $K^1$ is a chordal graph, which completes the proof. See also
\cite[Theorem 4.2]{servatius1989surface} for a relevant result.
\end{proof}
\section{Commutator subgroups of graph groups}\label{sec: commutator subgp}
The \textit{commutator subgroup} $[G,G]$ of a group $G$ is generated by
commutators $[g,h]:=ghg^{-1}h^{-1}$ with $g,h\in G$.
Let $G_1,\dots,G_n$ be finite groups and $K$ be a flag complex with $K^1$
a chordal graph. From the previous section we know that under these assumptions
the kernel of the projection map
$$p: \prod_{K^1} G_i \to \prod_{1\leq i \leq n} G_i$$
is a free group. This kernel is generated by iterated commutators of the form
$$[g_{j_1},[g_{j_2},[\dots,[g_{j_k},g_{j_{k+1}}]\dots]]],$$
where $g_{j_i}$ belong to distinct $G_{j_i}$.
The kernel ${\rm ker}(p)$ is not necessarily the commutator subgroup of
$\prod_{K^1} G_i$, if at least one of the $G_i$ is not abelian.
However, ${\rm ker}(p)$ {coincides} with the commutator subgroup of the
graph group if the groups $G_1,\dots,G_n$ are all abelian.
In this section we describe a basis for the free group
${\rm ker}(p)=F_{\rho_K}$ in terms of iterated commutators.
A basis was given in \cite[Lemma 4.7]{panov2016polyhedral},
where the groups under consideration had order 2, that is the graph
groups were right-angled Coxeter groups. Another version
of this basis of commutators was studied by Grbi\'c, Panov, Theriault, and Wu
in the context of exterior algebras in \cite[Theorem 4.3]{wu.grbic.panov}.
Before we proceed it is important to note that the commutator subgroup
of a free group can also be described by a generating set not consisting
of commutators.
One can obtain new presentations not involving commutators
using Tietze transformations
\cite{magnus2004combinatorial,vdovina1995constructing}.
Recall that the fibre in (\ref{eqn: D-S-fibration})
depends only on the order of the finite groups $G_1,\dots,G_n$. Therefore,
it suffices to describe the basis elements (i.e. iterated commutators)
of ${\rm ker}(p)=F_{\rho_K}$ only when $G_1,\dots,G_n$ are cyclic groups.
The basis for the general case of any finite groups $G_1,\dots,G_n$
can be obtained by considering the basis when $G_i$ are all cyclic
and then replacing the entries in the commutators with the nontrivial elements of $G_i$.
This observation will be used in Section \ref{sec: graph products abelian gps}.
\begin{prop}\label{prop: basis of pi_1 for K^0}
Let $G_1,\dots,G_n$ be finite groups and $K=K^0$.
Then the fundamental group of the fibre $(\underline{EG},\underline{G})^K$
in (\ref{eqn: D-S-fibration}) is the free group with basis consisting of the
following iterated commutators
\begin{equation}
[g_j,g_i],\,\, [g_{k_1},[g_j,g_i]],\,\,\dots\,,[g_{k_1},\dots,[g_{k_l},[g_j,g_i]]\dots],
\end{equation}
where $g_t\in G_t$, with $k_1<\cdots<k_{l}<j$ and $j>i\neq k_r$, for all $r$.
\end{prop}
\begin{proof}
We need to show that (1) this set of elements generates the fundamental group, and that
(2) the number of elements in the set equals the rank of the free group
in equation (\ref{eqn: rank of kernel for chordal graph}).
Since the first part of the proof is essentially the proof of
\cite[Lemma 4.7]{panov2016polyhedral}, we only give an outline here.
First recall the Hall identities for group elements $a,b,c$
\begin{equation}\label{eqn: Hall-Witt}
[a,bc]=[a,c][a,b][[a,b],c], \text{ and }
[ab,c]=[a,c][[a,c],b][b,c],
\end{equation}
and if $x$ is a commutator we can write
\begin{equation}\label{eqn: Hall-Witt derived}
[g_j,[g_i,x]]=[g_j,x] [x,[g_i,g_j]] [g_j,g_i] [x,g_i] [g_i,[g_j,x]] [x,g_j] [g_i,g_j] [g_i,x].
\end{equation}
Therefore, given the equations (\ref{eqn: Hall-Witt}) and (\ref{eqn: Hall-Witt derived})
we proceed as follows:
\begin{itemize}[leftmargin=.2in]
\item[--] we can use the identities above to switch between the commutators
$[g_j,[g_i,x]]$ and $[g_i,[g_j,x]]$ by using other commutators of lower degrees,
we can change the order of $g_{k_1},\dots,g_{k_l},g_j$ in the commutator
$[g_{k_1},\dots,[g_{k_l},[g_j,g_i]]\dots]$ to have them in increasing order, so
we can thus obtain the inequalities in the proposition;
\item[--] we can use the identities to eliminate commutators with two entries from the same
group, since we can reorder the terms to have these two entries next to each other,
then their product is in the same group, hence having a lower degree commutator;
\item[--] we can thus assume that the commutators have the prescribed order, and that $g_k,g_l$
are from different groups if $k\neq l$;
\item[--] finally we obtain a generating set for the the free group $F_{\rho(n)}$ in terms of commutators
$[g_{k_1},\dots,[g_{k_l},[g_j,g_i]]\dots]$, with $k_1<\cdots<k_{l}<j$
and $j>i\neq k_r$, for all $r$. Call this set $\mathcal S$.
\end{itemize}
Note that $\mathcal S$ does not generate the commutator subgroup, unless all $G_i$ are abelian.
Now we need to show that this generating set is minimal, that is $|\mathcal S|=\rho_K$
defined in (\ref{eqn: rank of kernel for chordal graph}).
For this we use induction on the number $n$ of vertices of $K$.
Let us denote $\rho_K=\rho(n)$ when $K=K^0$ consists of only $n$ vertices.
Assume $G_1,\dots,G_n$ have orders $m_1,\dots,m_n$, respectively.
For $n=2$, clearly $|\mathcal S|=|\{(g_j,g_i)|g_j,g_i\neq 1\}|=(m_1-1)(m_2-1)=\rho_K=\rho(2)$.
Suppose this is true for $n=k$. For $n=k+1$ we claim that
$$
\rho(k+1)=m_{k+1} \rho(k) +(m_{k+1}-1)(\prod_{1\leq i \leq k}m_i - 1).
$$
When we introduce a new group $G_{k+1}$, since it has the highest index,
according to our assumption, its elements come only second from the last
in the iterated commutator.
This yields $m_{k+1} \rho(k)$ generators, by taking a commutator and placing the elements of $G_{k+1}$
second from last in the iterated commutators, giving a higher degree commutator.
The insertion of a new non-trivial element, gives more freedom to the last element in the iterated
commutator. For each non-trivial element $1\neq g_i \in G_{k+1}$, the last entry
can take $\prod_{1\leq i \leq k}m_i - 1$ values.
Now counting for each element of the new group,
gives the second term in our claim, hence proving the claim.
Combining equation (\ref{eqn: rank of kernel for chordal graph}) and the claim,
and rearranging the terms, the minimality of $\mathcal S$ follows.
\end{proof}
\begin{ex}
Let $G_1={\mathbb{Z}}_2=\{1,x\},\, G_2={\mathbb{Z}}_2=\{1,y\},\,G_3={\mathbb{Z}}_3=\{1,z,z^2\},$
and $K=\{\{1\},\{2\},\{3\}\}.$ Then the fibre of the fibration
has fundamental group the free group $F_9$ with a minimal generating set
$\mathcal S$ given by
$$
\mathcal S = \{[z,x] , [z^2,x] , [z,y] , [z^2,y] , [y,x], [x,[z,y]] , [x,[z^2,y]] , [y,[z,x]] , [y,[z^2,x]]\}.
$$
\end{ex}
For simplicial complexes $K$ strictly larger than their 0-skeleton the
following proposition holds.
\begin{prop}\label{prop: basis of pi_1 for K=flag}
Let $G_1,\dots,G_n$ be finite groups and $K$ be a flag complex
with $n$ vertices such that $K^1$ is a chordal graph.
Then the fundamental group of the fibre in (\ref{eqn: D-S-fibration}) is a free group
with a basis the iterated commutators
$$
[g_j,g_i],\,\, [g_{k_1},[g_j,g_i]],\,\,\dots\,,[g_{k_1},\dots,[g_{k_l},[g_j,g_i]]\dots],
$$
where $g_t\in G_t$, with $k_1<\cdots<k_{l}<j$ and $j>i\neq k_r$, for all $r$, and
$i$ is the smallest vertex in a component not containing $j$ in the subcomplex
of $K$ restricted to $\{k_1,\dots,k_{l},j,i\}.$
\end{prop}
When we start introducing edges in $K^0$, then we start introducing commutator relations
$[g_i,g_j]$ whenever $\{i,j\}$ is an edge. In the iterated commutator $[g_{k_1},\dots,[g_{k_l},[g_j,g_i]]\dots]$
if $i,j$ are in the same connected component of $K$ restricted to $\{k_1,\dots,k_{l},j,i\},$
then there is a path from $i$ to $j$ with coordinates from $\{k_1,\dots,k_{l}\},$ hence
we can consider the iterated commutator induced by these vertices. Using
relations from the edges we can reduce this commutator to another commutator
of shorter length etc. Thus we can choose $i,j$ to be in different path components.
If we have two commutators where the last coordinate is in the same
component, one can show that we can write one in terms of the other.
Hence, we choose the smallest between them.
We leave it to the reader to check that the detailed arguments
in the proof of \cite[Theorem 4.5]{panov2016polyhedral}
work also for any selection of finite groups.
\begin{ex}
Let us consider an example with the symmetric group on 3 letters.
Let $G_1=\Sigma_3:=\langle s,t | s^2=t^2, (st)^3\rangle=\{1,x_1,x_2,x_3,x_4,x_5\}$,
$G_2={\mathbb{Z}}_2=\{1,y\}$, $G_3={\mathbb{Z}}_3=\{1,z,z^2\},$ and $K''=\{\{1,2\},\{3\}\}$ in
Figure \ref{fig: K and K'}. Then the fibre of the fibration
has fundamental group the free group $F_{22}$ with $\mathcal S$ given by
\begin{align*}
\mathcal S = \{ & [z,x_1] , [z,x_2], [z,x_3], [z,x_4],[z,x_5],
[z^2,x_1] , [z^2,x_2],[z^2,x_3],[z^2,x_4],[z^2,x_5], \\
& [z,y] , [z^2,y] , [x_1,[z,y]] ,[x_2,[z,y]], [x_3,[z,y]], [x_4,[z,y]], [x_5,[z,y]] \\
& [x_1,[z^2,y]] ,[x_2,[z^2,y]], [x_3,[z^2,y]], [x_4,[z^2,y]], [x_5,[z^2,y]]\}.
\end{align*}
Note that the structure of the symmetric group $\Sigma_3$
was not needed to write the generating set $\mathcal S$.
Therefore, if we replace $\Sigma_3$ with the cyclic group of
order six $C_6$, then the corresponding generating set $\mathcal S$ has the
same number and types of generators where in the commutators in $S$ we
replace the elements of the symmetric group with those of the cyclic group.
\end{ex}
\section{Examples of monodromy representations}\label{section: e.g.}
Let $G_1,\dots,G_n$ be finite discrete groups and $K$ be a flag complex with $K^1$
a chordal graph. Consider the following commutative diagram
\begin{equation}
\begin{tikzcd}
1 \arrow{r} & F_{\rho_K} \arrow{r} \arrow{d}{\Psi = \rm iso} &
\prod_{K^1} G_i \arrow{r} \arrow{d}{\Theta_K}
& \prod_{i} G_i\arrow{r} \arrow{d}{\Phi_{K}} & 1 \\
1 \arrow{r} & {\rm Inn}(F_{\rho_K}) \arrow{r} &{\rm Aut}(F_{\rho_K}) \arrow{r} &
{\rm Out}(F_{\rho_K}) \arrow{r} & 1,
\end{tikzcd}
\end{equation}
where $\Theta(g)(h)=ghg^{-1}$ and $\Psi(g)(h)=ghg^{-1}$. We are interested
in describing the maps $\Theta_{K}$ and $\Phi_{K}.$
For examples concerning only two finite groups, i.e. $n=2$, see \cite{stafa.monodromy},
where explicit answers are given. Using \verb|Magma|
we can explicitly describe faithful representations
$$\Phi_K: G_1\times \cdots \times G_n \to {{\rm SL}}({\rho_K},{\mathbb{Z}}),$$
where $G_1, \dots , G_n$ are finite abelian groups and $n\geq 3$.
In general, if $G_i$ are any finite groups (not necessarily abelian),
we obtain faithful representations of graph products of finite groups
$$ \Theta_K: \prod_{K^1} G_i \to {\rm Aut}(F_{\rho_K})$$
as well as faithful monodromy representations of direct products of finite groups
$$ \Phi_K: G_1\times \cdots \times G_n \to {{\rm GL}}({\rho_K},{\mathbb{Z}})$$
as will be shown below.
These include many interesting classes of discrete groups, such as right-angled
Coxeter groups. If one of the groups is infinite discrete, then additional
examples include hyperbolic groups as described in \cite[Theorem 5.1]{holt2012generalising},
braid groups, right-angled Artin groups and more.
Thus such representations can be realized as monodromy representations.
The rank $\rho_K$ increases {very fast} (\ref{eqn: rank of kernel for chordal graph})
if we increase the order and the number of the groups in consideration.
We concentrate on a couple of examples including right-angled Coxeter groups.
In addition we select the
simplicial complexes $K$ and $K'$ in Figure \ref{fig: K and K'},
to keep the rank of the free group small.
It is clear that by modifying the same codes in \verb|Magma| included in the appendix,
it is possible to obtain many more explicit examples, which we leave to the
interested reader. However, note that the basis generated in \verb|Magma|
is different (yet equivalent) from the basis we describe in Section
\ref{sec: commutator subgp}.
\begin{figure}[ht!]
\begin{tikzpicture}
[scale=1, vertices/.style={draw, fill=black, circle, inner sep=0.5pt}]
\node[vertices, label=below:{{\footnotesize$$1$$}}] (f) at (0,0) {};
\node[vertices, label=below:{{\footnotesize$$2$$}}] (e) at (0.5,0) {};
\node[vertices, label=below:{{\footnotesize$$3$$}}] (g) at (1,0) {};
\node[vertices, label=below:{{\footnotesize$$1$$}}] (a) at (3,0) {};
\node[vertices, label=below:{{\footnotesize$$2$$}}] (b) at (3.5,0) {};
\node[vertices, label=below:{{\footnotesize$$3$$}}] (c) at (4,0) {};
\node[vertices, label=below:{{\footnotesize$$4$$}}] (d) at (4.5,0) {};
\node[vertices, label=below:{{\footnotesize$$1$$}}] (k) at (6,0) {};
\node[vertices, label=below:{{\footnotesize$$2$$}}] (l) at (6.5,0) {};
\node[vertices, label=below:{{\footnotesize$$3$$}}] (m) at (7,0) {};
\foreach \to/\from in {a/b,b/c,c/d}
\foreach \to/\from in {k/l}
\draw [-] (\to)--(\from);
\end{tikzpicture}
\caption{$K,\,\,K',$ and $K''$}
\label{fig: K and K'}
\end{figure}
\begin{ex}\label{ex: 1}
Consider three groups of order 2 and the following short exact sequence obtained from
the fibration sequence (\ref{eqn: D-S-fibration})
$$
1 \to F_5 \to {\mathbb{Z}}_2\ast {\mathbb{Z}}_2, \ast {\mathbb{Z}}_2 \to {\mathbb{Z}}_2 \times {\mathbb{Z}}_2, \times {\mathbb{Z}}_2 \to 1,
$$
corresponding to the simplicial complex $K$ in Figure \ref{fig: K and K'},
where each of the cyclic groups is generated by $a$, $b$ and $c$, respectively.
Recall that the rank of the fibre is given by equation
(\ref{eqn: rank of kernel for chordal graph}) and in this case is 5.
Then $ F_5$ has a generating set (see \verb|Magma| in the appendix),
thus a presentation given by
$$
F_5:=\langle (ba)^2,(ca)^2,(cb)^2,acbcba,bcacba \rangle=\langle x_1,x_2,x_3,x_4,x_5\rangle.
$$
The action of ${\mathbb{Z}}_2 \times {\mathbb{Z}}_2, \times {\mathbb{Z}}_2$ on $F_5$ is determined by the following:
\begin{align*}
a\cdot x_i = \begin{cases}
x_1^{-1} \\
x_2^{-1} \\
x_4 \\
x_3 \\
x_5^{-1}
\end{cases}
b\cdot x_i = \begin{cases}
x_1^{-1} \\
x_5 x_1^{-1} \\
x_3^{-1} \\
x_1 x_4^{-1} x_1^{-1} \\
x_2 x_1^{-1}
\end{cases}
c\cdot x_i = \begin{cases}
x_3 x_5 x_4^{-1} x_2^{-1} & \text{if } i=1\\
x_2^{-1} & \text{if } i=2\\
x_3^{-1} & \text{if } i=3\\
x_2 x_4^{-1} x_2^{-1} & \text{if } i=4\\
x_3 x_1 x_4^{-1} x_2^{-1} & \text{if } i=5,
\end{cases}
\end{align*}
where $a,b,c$ act by conjugation. This gives also a faithful representation of
the right-angled Coxeter group to the automorphism group ${\rm Aut}(F_5)$.
It is straightforward to check that the induced action on
the abelianization ${\mathbb{Z}}^5$ gives a faithful representation of
the right-angled Coxeter group to the special linear group over the integers:
\begin{align*}
\Phi_K: {\mathbb{Z}}_2 \times {\mathbb{Z}}_2, \times {\mathbb{Z}}_2 & \to {{\rm SL}}(5,{\mathbb{Z}})\\
a,b,c &\mapsto X,Y,Z,
\end{align*}
where $X,Y,Z$ are the following matrices, respectively (example 1 in the appendix):
{{{{ \small
\begin{align*}
\left(\begin{array}{rrrrr}
-1 & 0 & 0 & 0 & 0 \\
0 & -1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 \\
0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 0 &-1 \\
\end{array}\right),
\left(\begin{array}{rrrrr}
-1 & 0 & 0 & 0 & 0 \\
-1 & 0 & 0 & 0 & 1 \\
0 & 0 & -1 & 0 & 0 \\
0 & 0 & 0 & -1 & 0 \\
-1 & 1 & 0 & 0 & 0 \\
\end{array}\right),
\left(\begin{array}{rrrrr}
0 & -1 & 1 & -1 & 1 \\
0 & -1 & 0 & 0 & 0 \\
0 & 0 & -1 & 0 & 0 \\
0 & 0 & 0 & -1 & 0 \\
1 & -1 & 1 & -1 & 0 \\
\end{array}\right).
\end{align*}
}}}}
Note that the generators of the fundamental group of
the polyhedral product $({E{\mathbb{Z}}_2},{{\mathbb{Z}}_2})^K$
can be described using the loops in Figure \ref{fig: the loops}
sitting in the space $({I},{F})^K$.
\begin{figure}[ht!]
\begin{tikzpicture}[scale=.5]
\draw[line width=0.3mm](0,0)--(1,1);
\draw[-,line width=0.3mm](1,1)--(3,1);
\draw[line width=0.3mm](3,1)--(2,0);
\draw[line width=0.3mm](2,0)--(0,0);
\draw[dotted](0,0)--(0,2);
\draw[dotted](0,2)--(1,3);
\draw[dotted](1,3)--(3,3);
\draw[dotted](3,3)--(2,2);
\draw[dotted](2,2)--(0,2);
\draw[-,dotted](1,1)--(1,3);
\draw[dotted](2,2)--(2,0);
\draw[dotted](3,3)--(3,1);
\end{tikzpicture}
\hspace{.2in}
\begin{tikzpicture}[scale=.5]
\draw[dotted](0,0)--(1,1);
\draw[-,dotted](1,1)--(3,1);
\draw[dotted](3,1)--(2,0);
\draw[dotted](1,1)--(1,3);
\draw[line width=0.3mm](2,0)--(0,0);
\draw[line width=0.3mm](0,0)--(0,2);
\draw[dotted](0,2)--(1,3);
\draw[dotted](1,3)--(3,3);
\draw[dotted](3,3)--(2,2);
\draw[line width=0.3mm](2,2)--(0,2);
\draw[line width=0.3mm](2,2)--(2,0);
\draw[dotted](3,3)--(3,1);
\end{tikzpicture}
\hspace{.2in}
\begin{tikzpicture}[scale=.5]
\draw[line width=0.3mm](0,0)--(1,1);
\draw[line width=0.3mm](1,1)--(1,3);
\draw[-,dotted](1,1)--(3,1);
\draw[dotted](3,1)--(2,0);
\draw[dotted](2,0)--(0,0);
\draw[line width=0.3mm](0,0)--(0,2);
\draw[line width=0.3mm](0,2)--(1,3);
\draw[dotted](1,3)--(3,3);
\draw[dotted](3,3)--(2,2);
\draw[dotted](2,2)--(0,2);
\draw[dotted](2,2)--(2,0);
\draw[dotted](3,3)--(3,1);
\end{tikzpicture}
\hspace{.2in}
\begin{tikzpicture}[scale=.5]
\draw[line width=0.3mm](0,0)--(1,1);
\draw[-,line width=0.3mm](1,1)--(3,1);
\draw[dotted](3,1)--(2,0);
\draw[line width=0.3mm](2,0)--(0,0);
\draw[dotted](0,0)--(0,2);
\draw[dotted](0,2)--(1,3);
\draw[dotted](1,3)--(3,3);
\draw[line width=0.3mm](3,3)--(2,2);
\draw[dotted](2,2)--(0,2);
\draw[-,dotted](1,1)--(1,3);
\draw[line width=0.3mm](2,2)--(2,0);
\draw[line width=0.3mm](3,3)--(3,1);
\end{tikzpicture}
\hspace{.2in}
\begin{tikzpicture}[scale=.5]
\draw[line width=0.3mm](0,0)--(1,1);
\draw[-,line width=0.3mm](1,1)--(1,3);
\draw[-,dotted](1,1)--(3,1);
\draw[line width=0.3mm](3,1)--(2,0);
\draw[line width=0.3mm](2,0)--(0,0);
\draw[dotted](0,0)--(0,2);
\draw[dotted](0,2)--(1,3);
\draw[line width=0.3mm](1,3)--(3,3);
\draw[dotted](3,3)--(2,2);
\draw[dotted](2,2)--(0,2);
\draw[dotted](2,2)--(2,0);
\draw[line width=0.3mm](3,3)--(3,1);
\end{tikzpicture}
\caption{The generators $x_1,\,x_2,\,x_3,\,x_4,$ and $x_5$, respectively}\label{fig:t3}
\label{fig: the loops}
\end{figure}
\end{ex}
\begin{ex}\label{ex: 2}
Now we consider four cyclic groups.
Construct the right-angled Coxeter group over
the simplicial complex $K'$ given in Figure \ref{fig: K and K'}.
Then equation (\ref{eqn: D-S-fibration}) gives the following short
exact sequence of groups
$$
1 \to F_{5} \to \prod_{K'}{\mathbb{Z}}_2 \to {\mathbb{Z}}_2^4 \to 1,
$$
where $F_{5} = \langle x_1,\dots,x_{5}\rangle:=\langle
(ca)^2,
(da)^2,
(db)^2,
a d b d b a,
c d a d c a,
\rangle.$
The conjugation action is then described as follows:
\begin{align*}
a\cdot x_i = \begin{cases}
x_1^{-1} \\
x_2^{-1} \\
x_4 \\
x_3 \\
x_5^{-1}
\end{cases}
b\cdot x_i = \begin{cases}
x_1 \\
x_3^{-1}x_2 x_4 \\
x_3^{-1} \\
x_4^{-1} \\
x_3^{-1} x_5 x_4
\end{cases} \,
c\cdot x_i = \begin{cases}
x_1^{-1} \\
x_5 x_1^{-1} \\
x_3 \\
x_1 x_4 x_1^{-1} \\
x_2 x_1^{-1}
\end{cases}\!\!
d\cdot x_i = \begin{cases}
x_5 x_2^{-1} \\
x_2^{-1} \\
x_3^{-1} \\
x_2 x_4^{-1} x_2^{-1} \\
x_1 x_2^{-1}
\end{cases}
\end{align*}
for all values of $i=1,2,3,4,5,$ respectively.
Then there is a representation of the right-angled Coxeter group
$\prod_{K'}{\mathbb{Z}}_2$ into the automorphism group ${\rm Aut}(F_5)$.
$$ \prod_{K'}{\mathbb{Z}}_2 \hookrightarrow {\rm Aut}(F_5) .$$
The induced action on the abelianization of the free group gives a
faithful representation
\begin{align*}
\Phi_{K'}: {\mathbb{Z}}_2^4 & \to {{\rm SL}}(5,{\mathbb{Z}})\\
a,b,c,d &\mapsto A,B,C,D
\end{align*}
where $A,B,C,D$ are the following matrices, respectively (example 2 in the appendix):
{{{{{ \small
\begin{align*}
A=&\left(\begin{array}{rrrrr}
1 & 0 & 0 & 0 & 0 \\
0 & 1 & -1 & 1 & 0 \\
0 & 0 & -1 & 0 & 0 \\
0 & 0 & 0 &-1 & 0 \\
0 & 0 & -1 & 1 & 1 \\
\end{array}\right),
B=\left(\begin{array}{rrrrr}
-1 & 0 & 0 & 0 & 0 \\
0 & -1 & 0 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 \\
0 & 0 & 1 &0 & 0 \\
0 & 0 & 0 & 0 & -1 \\
\end{array}\right), \\
C=&\left(\begin{array}{rrrrr}
-1 & 0 & 0 & 0 & 0 \\
-1 & 0 & 0 & 0 & 1 \\
0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 &1 & 0 \\
-1 & 1 & 0 & 0 & 0 \\
\end{array}\right),
D=\left(\begin{array}{rrrrr}
0 & -1 & 0 & 0 & 1 \\
0 & -1 & 0 & 0 & 0 \\
0 & 0 & -1 & 0 & 0 \\
0 & 0 & 0 &-1 & 0 \\
1 & -1 & 0 & 0 & 0 \\
\end{array}\right).
\end{align*}
}}}}}
\end{ex}
One can start with any finite groups $G_1,\dots , G_n$ with given
presentations and any simplicial complex $K$ with on $n$ vertices
such that $K^1$ is a chordal graph. If either of the
groups is not abelian, then the representations obtained
in the abelianization may not have images in ${{\rm SL}}(\rho_K,{\mathbb{Z}})$,
but rather in ${{\rm GL}}(\rho_K,{\mathbb{Z}})$ as shown in \cite[Example 2]{stafa.fund.gp}.
\begin{ex}
It is known that ${{\rm GL}}(n,{\mathbb{Z}})$ can be generated by four matrices $x,y,z,w$, see for instance
\cite[pp.44-49]{delaharpe2000topics}. For the case $n=3$ we can write
\begin{align*}
x=\left(\begin{array}{*{3}r}
0 & 0 & 1 \\
1 & 0 & 0 \\
0 & 1 & 0 \\
\end{array} \right)
y=\left(\begin{array}{*{3}r}
0 & 1 & 0 \\
1 & 0 & 0 \\
0 & 0 & 1 \\
\end{array} \right)
z=\left(\begin{array}{*{3}r}
1 & 1 & 0 \\
0 & 1 & 0 \\
0 & 0 & 1 \\
\end{array} \right)
w=\left(\begin{array}{*{3}r}
-1 & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 & 1 \\
\end{array} \right).
\end{align*}
If we define $v:=wyz$, then the groups $G=\langle x,y,w\rangle$ and $H=\langle v\rangle$
are finite of order 48 and 6, respectively, and $G$ is nonabelian. It can be shown that
${{\rm GL}}(3,{\mathbb{Z}})$ is a quotient of the free product $G\ast H$. Moreover, if $n$ is odd, then the $n\times n$ matrices
\begin{align*}
x=\left(\begin{array}{*{5}c}
0 & 0 & \cdots & 1 \\
1 & 0 & \cdots & 0 \\
\vdots & \ddots& \ddots& 0\\
0 & \cdots & 1 & 0 \\
\end{array} \right),\,
y=\left(\begin{array}{*{5}c}
1 & 1 & 0 & \cdots & 0 \\
0 & 1 & 0 & \cdots & 0 \\
0 & 0 & 1 & \cdots & 0\\
\vdots & \ddots& \ddots& &0\\
0 & \cdots & 0 & \cdots & 1 \\
\end{array} \right)
\end{align*}
generate ${{\rm SL}}(n,{\mathbb{Z}})$. A similar example can be constructed using this information.
\end{ex}
\section{Graph products of abelian groups}\label{sec: graph products abelian gps}
Every finite abelian group can be written as a
finite direct sum of finite cyclic subgroups
with order a power of a prime. Here we will
describe how to think of a graph product of
finite abelian groups over $K$ as a graph product of
cyclic groups over a new simplicial complex $K_{\underline{G}}$,
at the expense of having more vertices in the simplicial complex.
In \cite[Theorem 2.2]{stafa.monodromy} it was
shown that two cyclic subgroups yield a faithful
monodromy representation
$$C_n\times C_m \to {\rm Out}(F_{\rho_K})$$
and a faithful representation
$$\Phi_K: C_n\times C_m \to {{\rm SL}}({\rho_K},{\mathbb{Z}}),$$
where $K=\{\{1\},\{2\}\}$.
\
Consider two finite abelian groups $G,H$.
Then we can write them as direct sums if cyclic groups
$$G\cong \bigoplus_{i\in I} C_{n_i},\,\,\,\, H \cong \bigoplus_{j\in J} C_{m_j}.$$
We can then replace the simplicial complex $K=\{\{1\},\{2\}\}$
by the union of two simplices
$$K'=\Delta[|I|-1] \sqcup \Delta[|J|-1].$$
This does not change the monodromy
representation $G\times H \to {\rm Out}(F_{\rho_K})$, because the short exact sequences
of groups
$$
1 \to F_{\rho_K} \to G \ast H \to G \times H \to 1,
$$
and
$$
1 \to F_{\rho_{K'}} \to \prod_{(K')^1} C_{n_i,m_j} \to \prod_{i,j} C_{n_i,m_j} \to 1,
$$
are equivalent; note that $\prod_{i,j} C_{n_i,m_j} \cong G\times H$,
and
$$\prod_{(K')^1} C_{n_i,m_j}\cong \prod_{(\Delta[|I|-1])^1}C_{n_i}
\ast \prod_{(\Delta[|J|-1])^1}C_{m_j}\cong G\ast H.$$
More generally, if $K$ is a flag complex with $K^1$ a chordal graph,
and $G_1,\dots,G_n$ are finite abelian, write the
direct sum decomposition of each group $G_i$
$$G_i\cong \bigoplus_{j\in J_i} C_{n_j^i}.$$
Replace each vertex $i$ on the chordal graph $K^1$
by the simplex $\Delta[|J_i|-1]$. Give unique names to all vertices
in all these different simplices. Note that, if $\{i,j\}$ is an edge in $K$,
and $\Delta[|J_i|-1]=\{\{v_0^i,\cdots,v_{|J_i|-1}^i\}\}$, then we need
to add $\{v_k^i,v_t^j\}$ to the new simplicial complex for all $k,t$
since the $G_i,G_j$ commute if and only if all their subgroups commute.
We can now define the following simplicial complex.
\begin{defn}
Let $\underline{G}:=\{G_1,\dots,G_n\}$ be finite abelian groups.
Let $K$ be a simplicial complex on $n$ vertices with 1-skeleton
$K^1$ a chordal graph. Define the \textbf{simplicial complex $K_{\underline{G}}$}
to be the flag complex obtained from $K$ by the following procedure:
replace each vertex $i$ of $K$ with the full simplex
$\Delta[|J_i|-1]=\{\{v_0^i,\cdots,v_{|J_i|-1}^i\}\}$,
add an edge between the vertices in $\Delta[|J_k|-1]$ and the vertices
in $\Delta[|J_l|-1]$ if $G_k$ and $G_l$ commute, and take the
corresponding clique complex of the 1-skeleton
of this new simplicial complex.
\end{defn}
We then have the following lemma.
\begin{lemma}\label{lem: new flag complex from given K}
With the same assumptions, the graph $(K_{\underline{G}})^1$ is chordal.
\end{lemma}
\begin{proof}
This follows from the definition:
The 1-skeleton of $\Delta[|J_i|-1]$ and $K^1$ are chordal graphs.
If $c$ is a cycle of length greater than 3, then its edges are either
all in $(\Delta[|J_i|-1])^1$ for some $i$ or it moves between
various 1-skeleta $(\Delta[|J_k|-1])^1$. Suppose $c$ has length 4.
If vertices of $c$ are all in a single $(\Delta[|J_i|-1])^1$ we are done.
If vertices of $c$ lie in two distinct $(\Delta[|J_i|-1])^1$'s,
then there is one edge between $(\Delta[|J_k|-1])^1$ and
$(\Delta[|J_l|-1])^1$, thus there is an edge between all the vertices
between these two simplices, in particular between nonconsecutive vertices.
If vertices of $c$ lie in three distinct $(\Delta[|J_i|-1])^1$'s,
the same argument holds.
If vertices of $c$ lie in four distinct $(\Delta[|J_i|-1])^1$'s, then
$c$ is a replica of a cycle in $K$.
The same arguments show the triangulation of longer cycles $c$ .
\end{proof}
\begin{thm}
Let $G_1,\dots, G_n$ be finite abelian groups and $K^1$ a chordal graph.
Then the faithful monodromy representation
$G_1\times\cdots \times G_n \to {\rm Out}(F_{\rho_K})$
induces a faithful representation
$$\Phi_K: G_1\times\cdots \times G_n \to {{\rm SL}}({\rho_K},{\mathbb{Z}}).$$
\end{thm}
\begin{proof}
By Lemma \ref{lem: new flag complex from given K},
the graph $(K_{\underline{G}})^1$ is chordal and
by definition $K_{\underline{G}}$ is a flag complex. Therefore
the spaces in fibration (\ref{eqn: D-S-fibration})
are Eilenberg-MacLane spaces. Furthermore, the monodromy representation
$$\Phi_K: G_1\times\cdots \times G_n \to {\rm Out}(F_{\rho_K})$$
is equivalent to the monodromy representation
$$\Phi_{K_{\underline{G}}}: G_1\times\cdots \times G_n \to {\rm Out}(F_{\rho_{K_{\underline{G}}}}),$$
where $F_{\rho_{K_{\underline{G}}}}=F_{\rho_{K}}$
and we rewrite $$G_i\cong \bigoplus_{j\in J_i} C_{n_j^i}.$$
Each element in $G_i$ lies in
a cyclic group, which by \cite[Theorem 2.2]{stafa.monodromy} maps
faithfully into ${{\rm SL}}(\rho_K,{\mathbb{Z}}).$
The theorem follows.
\end{proof}
\begin{cor}
If $K^1$ is a chordal graph, then there is a faithful
representation of the graph product $\prod_{K^1} G_i$
into the automorphism group ${\rm Aut}(F_{\rho_K})$ of the
free groups of rank ${\rho_K}$.
In particular, this is true for any right-angled Coxeter group.
\end{cor}
Recall that there is a short exact sequence of groups
$$
1 \to {\rm IA}_{N} \to {\rm Aut}(F_{N}) \to {{\rm GL}}(N,{\mathbb{Z}}) \to 1
$$
induced by the abelianization of the automorphisms of free groups,
that is the induced map on the first homology $H_1(F_{N})$.
The group ${\rm IA}_{N}$ is the analogue of the Torelli group in
mapping class groups of surfaces.
Then we have the following immediate corollary.
\begin{cor}
If $K^1$ is a chordal graph and $G_i$ are finite discrete groups,
then the images of $\prod_{K^1} G_i$ under the faithful representations above
are not in ${\rm IA}_{\rho_K}$.
\end{cor}
\
\noindent{\bf The automorphism group ${\rm Aut}(K)$.}
Let ${\rm Aut}(K)$ denote the group of simplicial automorphisms of the
finite simplicial complex $K$, which is naturally a subgroup
of the symmetric group $\Sigma_n$ if $K$ has $n$ vertices.
For example, if $K$ is the full simplex or the trivial
simplicial complex, then ${\rm Aut}(K)=\Sigma_n$, or if $K$ is an $n$-gon
${\rm Aut}(K)$ is the dihedral group $D_{2n}$.
The group ${\rm Aut}(K)$ acts on the fibre $(\underline{EG},\underline{G})^K$.
It is natural to consider the induced action of this automorphism group
on the monodromy representation
$$
\Phi_K : G_1\times \cdots \times G_n \to {\rm Out}(F_{\rho_K}).
$$
The action of the automorphism group ${\rm Aut}(K)$ on a polyhedral product
$(\underline{X},\underline{A})^K$ was studied by A. Al-Raisi in
his thesis \cite{raisithesis}, in particular in case $K$ is an $n$-gon
with more than 3 vertices (not a chordal graph) and for $G_i \cong {\mathbb{Z}}_2.$
The automorphism group ${\rm Aut}(K)$ acts on $(\underline{X},\underline{A})^K$
by permuting coordinates
$$\sigma \cdot (x_1,\dots,x_n):=(x_{\sigma(1)},\dots, x_{\sigma(n)}).$$
By definition, if $K^1$ is the a chordal graph, then
this action permutes only coordinates in a way that preserves the
fundamental group of $(\underline{BG},\underline{1})^K$.
More precisely, let $\sigma \in {\rm Aut}(K) \leq \Sigma_n$ and $K$
be a flag complex. Note that $K={{\rm Flag}}(K^1)$, hence $K$ is determined
by its 1-skeleton. Since $\sigma$ permutes only the
coordinates which already commute with each other, then $\sigma$ permutes
the groups which already commute in the graph group
$\pi_1((\underline{BG},\underline{1})^K)=\prod_{K^1} G_i$.
Hence we obtain the following statement.
\begin{lemma}
Let $K$ be a flag complex with $n$ vertices. Then ${\rm Aut}(K)$ preserves the
homotopy type of $(\underline{BG},\underline{1})^K$,
equivalently its fundamental group and that of the fibre.
In particular it preserves the monodromy representation.
\end{lemma}
\section{Induced maps in homology}
In this section we prove the following proposition.
\begin{prop}
Let $K$ be a flag complex and $G_1,\dots,G_n$ be finite groups.
Then the induced map on first homology groups
$$H_1((\underline{EG},\underline{G})^K;{\mathbb{Z}}) \to H_1((\underline{BG},\underline{G})^K;{\mathbb{Z}}) $$
is not a surjection.
\end{prop}
\begin{proof}
The main ingredient in this proof is the fact that the
abelianizations of both the fundamental group
$\pi_1((\underline{BG},\underline{G})^K) \cong \prod_{K^1} G_i$
and the product $\prod_{1\leq i \leq n} G_i$ are the same.
For a group $G$ denote $\mathscr A(G):=G/[G,G]$ and the
abelianization map $G \to \mathscr A(G)$ by ${\rm ab}_G$.
Note that the abelianization of $\prod_{K^1} G_i$ factors
through the group $\prod_{1\leq i \leq n} G_i$:
$$
\prod_{K^1} G_i \to \prod_{1\leq i \leq n} G_i
\xrightarrow{ab_G} \mathscr A(\prod_{K^1} G_i).
$$
Let
$G:=\prod_{K^1} G_i,\, H:=\prod_{1\leq i \leq n} G_i$,
and $N=\pi_1 ((\underline{EG},\underline{G})^K).$
Note that in general, for any abelian group $A$ and a surjection
$h:G \twoheadrightarrow A$, there is a unique map
$\phi: \mathscr A(G) \twoheadrightarrow A$ such that
$\phi \circ {\rm ab}_G = h$.
Consider the following commutative diagram
\begin{center}
\begin{tikzcd}
1 \arrow{r} & N \arrow{r} {i} & G\arrow{r}{p} \arrow{d}{{\rm ab}_G} & H \arrow{r} \arrow{d}{{{\rm ab}_H}} & 1, \\
{} & {} &\mathscr{A}(G) \arrow{r}{=} & \mathscr{A}(H) &
\end{tikzcd}
\end{center}
where $p\circ i =0,$ and ${\rm ab}_G \circ i = {\rm ab}_H \circ p \circ i = 0$.
Now, since $\mathscr{A}(G)$ is an abelian group, then there is a unique map $f: \mathscr A(N) \to \mathscr A(G)$ such that ${\rm ab}_G \circ i = f \circ {\rm ab}_N$. Since ${\rm ab}_G \circ i=0 = f \circ {\rm ab}_N,$ and ${\rm ab}_N$ is clearly not trivial, then $f$ cannot be onto. Therefore $H_1((\underline{EG},\underline{G})^K;{\mathbb{Z}}) \to H_1((\underline{BG},\underline{G})^K;{\mathbb{Z}})$ is not a surjection.
\end{proof}
This proposition is in the spirit of the induced maps in homology
introduced in the next section, concerning the spaces
$B(2,G)$ and $E(2,G)$ defined below. We seek a similar result
in that case, too.
\section{CT groups and Feit-Thompson theorem}\label{sec: B(2,G) and E(2,G)}
In this section we study commutative transitive groups defined
below, and use some methods from polyhedral products to understand
the interplay between topology and group theory, and
characterize some group properties using topology.
For any group $\pi$ the descending central series is given by
a sequence of normal subgroups
$$
\pi=\Gamma^1 \triangleright \Gamma^2 \triangleright
\cdots \triangleright \Gamma^{n+1} \triangleright \cdots
$$
where inductively $\Gamma^{n+1}=[\pi,\Gamma^{n}]$ for $n\geq 2.$
If $\pi=F_n$ is the free group of rank $n$, then for any topological group
$G$ there is a filtration
$$
Hom(F_n/\Gamma^2,G) \subset Hom(F_n/\Gamma^3,G) \subset \cdots \subset G^n.
$$
The sequences of spaces given by
$$B_k(q,G):=Hom(F_k/\Gamma^q,G)\subset G^k$$
and
$$E_k(q,G):=G\times Hom(F_k/\Gamma^q,G) \subset G^{k+1}$$
have the structure of simplicial complexes (\cite{fredb2g}), respectively,
with respective geometric realizations defined as follows
\begin{align*}
B(q,G):= |B_\ast(q,G)|, \text{ and } E(q,G):=|B_\ast(q,G)|.
\end{align*}
The projections $E_k(q,G) \twoheadrightarrow B_k(q,G)$ induce a fibration
\begin{equation}\label{eqn: fibration of E(q,G) to B(q,G)}
E(q,G) \to B(q,G) \to BG
\end{equation}
and in particular, for $q=2$ we have
\begin{equation}\label{eqn: fibration of E(2,G) to B(2,G)}
E(2,G) \to B(2,G) \to BG.
\end{equation}
The total space $B(2,G)$ and the homotopy fibre $E(2,G)$
were studied by A. Adem, F. Cohen and E. Torres Giese
\cite{fredb2g}. They posed the question
whether for finite $G$ the space $B(2,G)$ is always a $K(\pi,1)$,
having showed that these spaces are occasionally $K(\pi,1)$
for the case of {\it commutative transitive groups}.
C. Okay \cite{okay,okay2015colimits} gave
classes of groups for which $B(2,G)$ is not a $K(\pi,1)$, such as
extraspecial 2-groups of order $2^{2n+1}$, for $n\geq2$, hence
answering their question. A brief survey is given in \cite[\S 9]{stafa.comm.2}.
\begin{defn}\label{defn: CT group}
A group $G$ is \textbf{commutative transitive} or \textit{CT}
if commutativity is a transitive relation in $G$. That is, if
$[a,b]=[b,c]=1$, then $[a,c]=1$ for all non-central elements
$a,\,b,\,c \in G$.
\end{defn}
The class of CT groups played an important role in the
classification of finite simple groups and were studied by M. Suzuki
\cite{suzuki1957nonexistence,suzuki1986group}, among many others,
who showed that every non-abelian simple CT-group is of even order and
isomorphic to ${{\rm PSL}}(2, 2^f)$ for some $f\geq2$.
Finite CT groups have been classified, see for example
\cite[p. 519, Theorem 9.3.12]{schmidt1994subgroup}.
In particular, if $G$ is a finite CT group with trivial center
then the following is true.
\begin{prop}[{{\cite[Cor. 8.5]{fredb2g}}}]
If there are maximal abelian subgroups $G_1,\dots, G_n$ of $G$ that cover $G$,
then there is a homotopy equivalence $B(2,G) \simeq \bigvee_i BG_i$.
\end{prop}
With the assumptions of this proposition we have the following corollary.
\begin{cor}
$B(2,G)$ has the homotopy type of the polyhedral product
$(\underline{BG},\underline{1})^{K^0}.$
\end{cor}
In what follows $G$ is assumed to be finite.
Using the five term short exact sequence from the
Lyndon-Hochschild-Serre spectral sequence Adem, Cohen and Torres Giese
\cite[Proposition 7.2]{fredb2g} showed that the non-surjectivity
of the induced map on first homology of the fibration
(\ref{eqn: fibration of E(2,G) to B(2,G)})
$$ H_1(E(2,G)) \to H_1(B(2,G))$$
is equivalent to the Feit-Thompson theorem that groups of odd
order are solvable. Hence the study of the fibration encodes
fundamental information about the group $G$.
We would like to use polyhedral products, i.e. topology,
to extract more information about this equivalent form of
the Feit-Thompson theorem \cite{feit1963chapter},
which is algebraic in nature.
\
Let $G$ be a finite CT group with trivial center
and let $G_1,\dots, G_n$ be its cover by maximal abelian subgroups as above
($n$ is called the {{\it covering number}} of $G$).
Since all spaces are $K(\pi,1)$'s, we will move frequently between fundamental groups
and their classifying spaces. Note that there are two commutative diagrams of
short exact sequences of groups:
\begin{equation}\label{eqn: first diagram}
\begin{tikzcd}
\pi_1(E(2,G)) \arrow{r}\arrow{d} & {\mathcal H} \arrow{r} \arrow{d} & {[G,G]} \arrow{d}\\
\pi_1(E(2,G)) \arrow{r}\arrow{d} & \pi_1(B(2,G)) \arrow{r} \arrow{d}& \pi_1(BG) \arrow{d}\\
\ast \arrow{r} & H_1(BG) \arrow{r} & H_1(BG),
\end{tikzcd}
\end{equation}
and
\begin{equation}\label{eqn: commutative diagram}
\begin{tikzcd}
N \arrow{r}\arrow{d} & \pi_1(\underline{EG},\underline{G})^{K_0} \arrow{r} \arrow{d}{i_2} & {[G,G]} \arrow{d}\\
\pi_1(E(2,G)) \arrow{r}{i_1} \arrow{d}{p_1} & \pi_1(B(2,G)) \arrow{r}{p_2} \arrow{d}{p_3}& \pi_1(BG) \arrow{d}\\
Q \arrow{r} & \pi_1(\prod_i BG_i) \arrow{r} & H_1(BG),
\end{tikzcd}
\end{equation}
where $Q$ is a finite abelian group and $N$ is a free group (we omit the trivial groups
on each side of the short exact sequences). The existence of the first diagram is clear,
whereas for the second diagram, even though for CT groups the map $\pi_1(B(2,G)) \to G$
does not factor through the product $\prod_i G_i$, the composition $\pi_1(B(2,G)) \to G \to H_1(G),$
being an epimorphism onto an abelian group, factors uniquely through the abelianization
of $\pi_1(B(2,G))$, which is the direct product $\prod_i G_i$.
The map $p_1$ factors uniquely through the abelianization $H_1(E(2,G))$, hence there is a
map $q_1$ such that $p_1 = q_1 \circ {\rm ab}$. Hence there is a diagram
\begin{equation}\label{eqn: small commutative diagram}
\begin{tikzcd}
\pi_1(E(2,G)) \arrow{r}{i_1} \arrow{d}{p_1} & \pi_1(B(2,G)) \arrow{d}{p_3}\\
H_1(E(2,G)) \arrow{d}{q_1} \arrow{r}{(i_1)_\ast} & \prod_i G_i \arrow{d}\\
Q \arrow{r} & \prod_i G_i ,
\end{tikzcd}
\end{equation}
where the dotted map is the one we are interested in
(we are not claiming that the lower triangle commutes).
By \cite[Proposition 8.8]{fredb2g} the group $\pi_1(E(2,G))$
is free, with rank
$$ \mathcal{N}_G=1-|G:Z(G)| + \sum_{1\leq i \leq n} \left( |G:Z(G)|-|G:G_i|\right). $$
Since $Z(G)=\{1_G\}$, by rearranging the terms of $ \mathcal{N}_G$
we obtain the following:
\begin{align*}
\mathcal{N}_G &= 1-|G:Z(G)| + \sum_{1\leq i \leq n} \left( |G:Z(G)|-|G:G_i|\right)\\
&= 1-|G| + \sum_{1\leq i \leq n} \left( |G|-|G|/|G_i|\right)\\
&= 1-|G| + n|G| - \sum_{1\leq i \leq n} |G|/|G_i|\\
&= 1+ (n-1)|G| - \sum_{1\leq i \leq n} |G|/|G_i|.
\end{align*}
Note that this is a more general version of the formula
for $\rho_K$ in equation (\ref{eqn: rank of kernel for chordal graph}),
with the special case of $|G|=\prod_i|G_i|$ giving the rank $\rho_K$
when $K$ is only a set of $n$ points; let us use the notation
$\rho_K= \rho(n)$ since $K$ becomes irrelevant. In general
$l.c.m.(|G_1|,\dots,|G_n|) \leq |G|< \prod_i |G_i| $, since the
groups $G_i$ cover $G$. Actually they divide each other from
left to right. Since $|G|$ divides $\prod_i |G_i|,$ then
$\prod_i |G_i|=C |G|$. Therefore, we have
\begin{align*}
\rho(n)-\mathcal N_G &= (n-1)|G|(C-1) - \sum_{1\leq j\leq n}(|G|(C-1))/|G_j|\\
& = |G|(C-1)\left( n-1 - \sum_{1\leq j\leq n}1/|G_j| \right).
\end{align*}
Since all $|G_i|\geq 2$, then we have
$\rho(n)-\mathcal N_G \geq |G|(C-1)(n/2-1).$
If $n=2$ then $G$ is abelian, so assume that $n\geq 3$.
Then we get $\rho(n)-\mathcal N_G >0$.
\begin{lemma}
Let $G$ be a finite CT group with trivial center. Then $\rho(n)>\mathcal N_G.$
\end{lemma}
\begin{proof}
In addition to the above argument, this is also a direct consequence
of the fact that the index of $N$ in each of the free groups is given by
the following formula \cite[p.16]{lyndon2015combinatorial}:
\begin{align*}
& [ \pi_1(E(2,G)) : N] = \frac{{\rm rank} (N) - 1}{ {\rm rank} (\pi_1(E(2,G)))-1} = |Q| = \frac{\prod_i|G_i|}{|H_1(G)|},
\text{ and,}\\
& [\pi_1((\underline{EG},\underline{G})^{K_0}):N] = \frac{{\rm rank} (N) - 1}{ {\rm rank} (\pi_1((\underline{EG},\underline{G})^{K_0})-1} = |[G,G]| = \frac{|G|}{|H_1(G)|}.
\end{align*}
Since $|G| < {\prod_i|G_i|}$ the lemma follows.
\end{proof}
Indeed the proof of this lemma tells us that
$\rho(n)/\mathcal N_G \sim |Q|/|[G,G]|$.
Before we proceed, it is clear from the diagrams
(\ref{eqn: first diagram},\ref{eqn: commutative diagram})
that if $G/[G,G]=1$, then the induced
map on homology $H_1(E(2,G)) \to H_1(B(2,G))$ is onto (without using the 5-term sequence
in homology).
\begin{prop}
{If $G$ is simple, then the following map is a surjection
$$H_1(E(2,G);{\mathbb{Z}}) \to H_1(B(2,G);{\mathbb{Z}}). $$ }
\end{prop}
Instead, using only topology we want to prove the following equivalent statements:
\textit{if the map $H_1(E(2,G)) \to H_1(B(2,G))$ is onto, then $|G|$ is even,}
or equivalently,
\textit{if $|G|$ is odd, then the map $H_1(E(2,G)) \to H_1(B(2,G))$ is not onto}.
\
Now, if $|G|$ is odd, then all $Q, \, \prod_i G_i$ and $[G,G]$ are odd. Also $N$ is a free group
of odd index in both free groups $\pi_1(E(2,G))$ and $\pi_1(EG_i,G_i)^{K_0}$.
The following results are immediate:
\begin{lemma}
Either all $\rho(n),\, \mathcal N_G,\, {\rm rank}(N)$ are even,
or, all $\rho(n),\,\mathcal N_G,\, {\rm rank}(N)$ are odd,
such that the ratios
$$
\frac{{\rm rank}(N)-1}{\mathcal N_G-1},\, \frac{{\rm rank}(N)-1}{\rho(n)-1}, \, \frac{\rho(n)-1}{\mathcal N_G-1}
$$
are odd.
\end{lemma}
\begin{proof}
Use the formulas in the proof of the previous Lemma.
\end{proof}
Next note that the map $(i_1)_\ast$ in (\ref{eqn: small commutative diagram})
can be a surjection only if $Q \lneq {\rm Im}((i_1)_\ast)$
(if not then their intersection is at most $Q$ and the image
cannot be everything). Consider the following diagram
\begin{center}
\begin{tikzcd}
& {\rm Ker}_1 \arrow{d} & \\
{\rm Ker}_2 \arrow{r} & H_1(E(2,G)) \arrow{d}{q_1}
\arrow{r}{(i_1)_\ast} & {\rm Im}((i_1)_\ast) \arrow{d}\\
& Q \arrow{r}{i} & {\rm Im}((i_1)_\ast).
\end{tikzcd}
\end{center}
The image ${\rm Im}((i_1)_\ast) $ has odd order. Since both kernels have full rank
(the quotients are both finite groups)
we have that ${\rm Ker}_2 \leq {\rm Ker}_1$. The kernels have bases as follows
$$ {\rm Ker}_1=span\{\alpha_i e_i: i \in[N_G]\}, \text{ and }
{\rm Ker}_2=span\{\beta_i e_i: \beta_i|\alpha_i,\, \forall i \in[N_G]\}.$$
Here all $\alpha_i,\beta_i$ have to be odd numbers such that $\alpha_i|\beta_i $ for all $i$.
Indeed this can be done for any (finite) sequence of subgroups
$$ Q < Q_1 < \cdots < {\rm Im}((i_1)_\ast) < \cdots <\prod_i G_i$$
as there are kernels $K_0,\, K_1,\dots,$ of full ranks corresponding to projections.
The following theorem shows that ${\rm Im}((i_1)_\ast) \lneq \prod_i G_i$ for the
case of CT groups with trivial center.
We conclude this section with the following corollary.
\begin{cor}\label{cor: top. equiv. form CT groups}
Finite CT groups with trivial center are solvable if and only if the induced map
$H_1(E(2,G);{\mathbb{Z}}) \to H_1(B(2,G);{\mathbb{Z}})$ is not a surjection.
\end{cor}
Of course this is a special case of the condition in \cite[Proposition 7.2]{fredb2g},
but for this corollary we use only the diagrams (\ref{eqn: commutative diagram},\ref{eqn: small commutative diagram}).
\begin{proof}[Proof of Corollary \ref{cor: top. equiv. form CT groups}]
Consider the commutative diagram (\ref{eqn: commutative diagram}) and
(\ref{eqn: small commutative diagram}). If the map $(i_1)_\ast$ is a
surjection, then the composition
$$H_1(E(2,G);{\mathbb{Z}}) \twoheadrightarrow H_1(B(2,G);{\mathbb{Z}}) \twoheadrightarrow G/[G,G]$$
is a surjection. On the other hand, from diagram (\ref{eqn: commutative diagram}),
the group $H_1(E(2,G);{\mathbb{Z}})$ maps trivially, hence $G/[G,G]$ is trivial.
On the other hand, if $G/[G,G]$ is trivial, then $Q=\prod_i G_i$ and $(i_1)_\ast$
is a surjection.
\end{proof}
The following question is still open for CT groups:
{\it Use Corollary \ref{cor: top. equiv. form CT groups} to show that if
$G$ is a simple finite TC group with trivial center, then $G$ has even order}.
\input{appendix}
\bibliographystyle{amsplain}
|
\section{Introduction}
Subjective well-being and life satisfaction have received a significant amount of attention in the economics \citep{blanchflower_well-being_2004, ferrer-i-carbonell_how_2004} and psychology \citep{diener_subjective_1999, argyle_causes_1999} literature over the past decades. While cardinality properties are conceptually rejected, the ordinal nature of the information conveyed in such indicators has gradually entered public policy circles. \citet{stiglitz_report_2009} argued that meaningful and reliable data on well-being can be collected and should be included in surveys and official statistics with further reports advocating its role in policy making in informing policy designs, monitoring progress and evaluating implementation of public interventions \citep{dolan_measuring_2012}.
However, despite suggestions by various authors \citep{cummins_developing_2003, kahneman_developments_2006, krueger_reliability_2008, veenhoven_four_2013}, there is little consensus in how exactly such concepts are defined
and what precisely they capture. For the past few years the Office for National Statistics in the UK monitors year-on-year improvements in reported well-being through questions on life satisfaction, feelings of self-worth, happiness and anxiety, which however provides little theoretical basis or conceptual background for such choices \citep{ons_personal_2014}. While it is clear that well-being is less of an answer to a single question and more of a composite multidimensional concept, limited empirical work has pursued such considerations. The vast majority of empirical work (for reviews see \citet{kahneman_developments_2006,dolan_we_2008}) focuses on univariate models that would not allow taking into account dependence between constituent components and their link with overall happiness, while also being unable to correctly identify direct effects of independent determinants/variables on domain and generic happiness in the presence of such outcomes' dependence. In our context, generic well-being or happiness pertains to one's satisfaction with their life overall, while domain well-being/satisfaction pertains to satisfaction individual constituent components of life, e.g. income, health, social life, family life, etc.
Conceptual work suggests bottom-up theories \citep{diener_subjective_1984} for life-satisfaction where judgements are based on assessments of satisfaction with a number of defined life domains with ensuing causal pathways running from the domains upwards \citep{schimmack_structure_2008}. Complementary positions view overall life satisfaction as the net outcome of reported satisfaction with life domains with the domains themselves seen as functions of objective outcomes/situations/covariates \citep{michalos_global_1991, easterlin_happiness_2008}. Looking at the structure of composite well-being, \citet{salvatore_appraisal_2001} conclude it is more easily interpreted in terms of generic dimensions of life (e.g. family life, social life, love life, occupational life and leisure) than attributed to the fulfilment of personal values (self-acceptance, autonomy, environmental mastery, purpose in life, and personal growth). \citet{brief_integrating_1993} argued for a mixture of processes where individual characteristics and traits are driving domain satisfaction (i.e. top-down), which in turn drives life satisfaction (i.e. bottom-up). In other words, high individual income leads to increased life satisfaction because financial satisfaction is an important component of satisfaction with life as a whole.
Looking at the economics of well-being literature, univariate models again dominate, while at the same time very few studies have focused on its composite structure. \citet{van_praag_anatomy_2003} formulate a model very close to the initial bottom-up theories, which however is estimable only under strong assumptions of no direct association between individual covariates and generic life satisfaction. However, problems of omitted variables and lack of sufficient exclusion restrictions suggest caution in drawing conclusions. Similar problems are also faced in \citet{easterlin_happiness_2008}, where again strong assumptions (both conceptual and econometric) are required in the model and raise concerns about the robustness of findings.
However, most theories developed within the well-being composition literature offer limited empirical evidence for the underlying connections posited and fail to explicitly incorporate them into their setting. In this paper we propose a multivariate framework and estimate a comprehensive relationships pattern between generic and domain satisfactions with dependence explicitly modelled through copulas. We develop a novel joint copula-based Markov model, where a set of bivariate copulas and a multivariate t (MVT) copula jointly model multivariate ordinal time-series responses with covariates. Each ordinal time-series is considered a copula-based Markov model, where a parametric bivariate copula family is used for the joint distribution of subsequent observations and which is then related to these ordinal time-series responses using an MVT to join their conditional (on the past) distributions at each time point. Note in passing that other continuous-variable models using copulas in several Markov chains exist in the literature \citep{lambert&vande02,Patton-2012,Remillard-etal-2012, Beare-etal-2015}, but in our knowledge we are the first constructing such copula-based models for ordinal time-series with covariates. The theoretical and estimation concepts in the discrete case are quite different.
Simple parametric families of copulas in more than two dimensions typically provide limited dependence \citep{Nikoloulopoulos2013a} and for discrete data it is generally hard to provide a better fit than the multivariate normal (MVN) copula, which inherits the useful properties of the MVN distribution.
However, the MVN copula is inadequate to model multivariate data with more probability in one or both joint tails.
In recent years, a popular and useful approach is the vine pair-copula construction \citep{Kurowicka-Joe-2011,nikoloulopoulos&joe&li11} which is based on $d(d-1)/2$ bivariate copulas, of which some are used to summarize conditional dependence. Vine copula constructions are suitable for modelling this kind of data by using appropriate bivariate copulas. \cite{panagiotelis&czado&joe12} and \cite{nikoloulopoulos&joe12} recently extended the idea of vine copulas to discrete data. The approach in \cite{nikoloulopoulos&joe12} involves both observed and latent variables, while the approach of \cite{panagiotelis&czado&joe12} is suitable when there are no latent variables to explain the dependence in the observed variables. In this paper to form the bivariate part of the model, a D-vine truncated at the 1-st level \citep{Brechmann-Czado-Aas-2012} has been exploited.
To develop the joint copula-based Markov model we propose the MVT copulas.
MVT copulas nest MVN copulas
and share with them the ability to accommodate any feasible pattern of association in a set of random variables. However,
the MVT copulas offer greater flexibility than MVN copulas, as they can also capture dependence in both joint tails \citep{Nikoloulopoulos&joe&li09}, which is the case in `mixtures' of population (e.g., different locations or genders). The MVT copula-based approach of this article avoids having to specify parametrically the distribution of latent heterogeneity in a non-linear setting. The MVT copulas as scale mixtures of MVN can be used to explore unobserved population heterogeneity.
Implementation of the MVT copula for discrete data is feasible, but not easy, because the MVT distribution, as a latent model for discrete response, requires rectangle probabilities based on high-dimensional integrations \citep{genest&nikoloulopoulos&rivest08}.
The probability mass function (pmf) can be
obtained by computing an MVT rectangle probability and the randomized quasi Monte Carlo methods proposed by \cite{genz&bretz02} can be used for that purpose. Computing the rectangle MVT probabilities via simulation based on the methods in \cite{genz&bretz02} is akin to using a simulated likelihood method, whose asymptotic efficiency, for the special case of MVN copula, has been studied by \citet{nikoloulopoulos13b,
nikoloulopoulos2015a} and was shown to be as good as maximum likelihood for dimension 10 or lower.
Modelling of dependence further allows revisiting a number of established relationships in the literature between covariates and well-being and examining their
association in a multivariate setting (e.g. does the importance of income on generic well-being remain when considered along income satisfaction?). In short, we establish exogenous determinants for both generic and domain satisfaction equations and separately identify structural relations among components of well-being and shed light not only into their links with generic well-being but also among themselves. This allows separate identification of the direct effect of domain characteristics and of composite domains on generic satisfaction, as well as estimation of the dependence between and among generic and domain satisfaction.
\cite{Prokhorov&Schmidt2009} showed that robust estimation can be achieved in the class of radially or reflection symmetric copulas such as the normal or t copula. \cite{nikoloulopoulos&joe&chaganty10}, \cite{Masarotto&Varin12}, \cite{Nikoloulopoulos2015d} and \cite{Nikoloulopoulos-2016-wtsc-ord} (in particular for ordinal time-series) showed robustness of the normal copula to dependence if the main interest is the univariate parameters (regression and nonregression parameters).
This type of research focussed on marginal models and on estimation of coefficients for regression models with time-series data, that are robust to the dependence structure.
Nevertheless, our manuscript is not entirely in the area of ``marginal models" (meaning specification of univariate time-series only), but also avails copula-based models for both univariate and multivariate time-series.
One of the goals of our paper is to compare dependence models in inferences involving joint probabilities.
The remainder of the paper proceeds as follows. Section \ref{themodel} introduces the joint copula-based Markov model for discrete ordinal responses with covariates and presents the conceptual framework upon which this econometric model is built. Estimation techniques and computational details are provided in Section \ref{estimation}.
Section \ref{vuong-sec} discusses the \nocite{vuong1989}Vuong's (1989) test to assess the fit of the proposed model in terms of prediction of joint probabilities. Section \ref{sec-appl} presents the application of our methodology to the British Household Panel and Section \ref{sec-discussion} concludes, followed by a technical Appendix.
\section{\label{themodel}A joint copula-based Markov model}
In this section, we construct the joint copula-based Markov model for ordinal time-series with covariates. Before that, the first subsection has some background on copula models.
\subsection{\label{overview}Overview and relevant background for copulas}
A copula is a multivariate cdf with uniform $U(0,1)$ margins \citep{joe97,joe2014,nelsen06}.
If $F$ is a $d$-variate cdf with univariate margins $F_1,\ldots,F_d$,
then Sklar's (1959) theorem\nocite{sklar1959} implies that there is a copula $C$ such that
$$F(y_1,\ldots,y_d)= C\Bigl(F_1(y_1),\ldots,F_d(y_d)\Bigr).$$
The copula is unique if $F_1,\ldots,F_d$ are continuous, but not
if some of the $F_j$ have discrete components.
If $F$ is continuous and $(Y_1,\ldots,Y_d)\sim F$, then the unique copula
is the distribution of $(U_1,\ldots,U_d)=\left(F_1(Y_1),\ldots,F_d(Y_d)\right)$ leading to
$$C(u_1,\ldots,u_d)=F\Bigl(F_1^{-1}(u_1),\ldots,F_d^{-1}(u_d)\Bigr),
\hspace{2ex} 0\le u_j\le 1, j=1,\ldots,d,$$
where $F_j^{-1}$ are inverse cdfs. In particular,
if $\mathcal{T}_d(\cdot;\mathbf{R})$
is the MVT cdf with correlation matrix $\mathbf{R}=(\rho_{jk}: 1\le j<k\le d)$ and $\nu$ degrees of freedom, and $\mathcal{T}$ is the univariate Student t cdf with $\nu$ degrees of freedom,
then the MVT copula is
\begin{equation}\label{MVNcdf}
C(u_1,\ldots,u_d)=\mathcal{T}_d\Bigl(\mathcal{T}^{-1}(u_1),\ldots,\mathcal{T}^{-1}(u_d);\mathbf{R}\Bigr).
\end{equation}
\subsection{\label{markov}Copula-based Markov models for ordinal time-series }
For a latent variable $Z\sim \mathcal{F}$ such that $Y=y$ if
$\alpha_{y-1}+\mathbf{x}^T\boldsymbol{\beta}\leq Z\leq \alpha_{y}+\mathbf{x}^T\boldsymbol{\beta},\,y=1,\ldots,K,$
with $K$ being the number of categories of $Y$, $\boldsymbol{\beta}$ the $p$-dimensional regression vector, $t=1 \dots T$ the ``panel'' dimension, $i=1 \dots n$ the number of clusters (note that varying cluster sizes can be accommodated by the theory) and $p$ the number of covariates (i.e. the dimension of a covariate vector $\mathbf{x}$), the response $Y$ is assumed to have density
$$f(y;\mu,\boldsymbol{\gamma})=\mathcal{F}(\alpha_{y}+\mu)-\mathcal{F}(\alpha_{y-1}+\mu),$$
where $\mu=\mathbf{x}^T\boldsymbol{\beta}$ is a function of $\mathbf{x}$, $\boldsymbol{\beta}$ is a $p$-dimensional regression vector and $\boldsymbol{\gamma}=(\alpha_1,\ldots,\alpha_{K-1})$ is the $q$-dimensional vector of the univariate cutpoints ($q=K-1$) with $\alpha_0=-\infty$ and $\alpha_K=\infty$. Choosing normal or logistic for $\mathcal{F}$ leads to the ordered probit and cumulative logit models, respectively.
For data $(y_{itj}, \mathbf{x}_{itj})$, where $j$ is an index for the ordinal responses, the univariate marginal model for
$Y_{itj} $ is $f_j(y_{itj}; \mu_{itj},\boldsymbol{\gamma}_j)$ where $\mu_{itj}=\mathbf{x}_{itj}^\top\boldsymbol{\beta}_j$ and $\boldsymbol{\gamma}_j$ of dimension $q_j$. If for each $t$, $Y_{i1j},\ldots,Y_{iTj}$ are independent, then the
log-likelihood for each univariate ordinal response is
\begin{eqnarray}\label{indlik}\ell_{j}= \sum_{i=1}^n\sum_{t=1}^T\, \log f_j(y_{itj};\mu_{itj},\boldsymbol{\gamma}_j).
\end{eqnarray}
If the ordinal data are observed in a time-series sequence, then the ordinal regression model can be adapted in two ways:
\begin{enumerate}
\item[1.] add previous observations as covariates;
\item[2.] make use of some models for stationary ordinal time-series.
\end{enumerate}
Here we adapt the methodology for case 2.
For dependent $Y_{i1j},\ldots,Y_{iTj}$, estimation of $\boldsymbol{\beta}_j$ and $\boldsymbol{\gamma}_j$ involves copula-based Markov models \citep[page 244]{joe97} for ordinal time-series with covariates. The joint distribution of subsequent observations is modelled through a parametric copula family with the corresponding transition probabilities subsequently elicited.
The advantages of a time-series regression model are explicitly mentioned in \cite{Joe-2015-proceedings} and also reproduced below:
\begin{itemize}
\itemsep=0pt
\item The class of autocorrelation functions is much wider than those based on an ordered probit with lagged dependent variables appearing as explanatory variables.
\item Prediction in regressions with time dependent observations is simpler as they can be formulated with or without the preceding observations.
\item Serial dependence (positive or negative) can be modelled through suitable copula families.
\item The non-linearity of the conditional expectations allows for various patterns to be replicated, while the conditional expectation and variance for large values is determined by the choice of copula family and corresponding tail behaviour.
\item Incorporating covariates in time-series models is more straightforward in univariate regression models.
\item Extending the framework to Markov orders higher than one is straightforward.
\item Copula families with an easy (e.g. closed) form allow for easier likelihood inference.
\end{itemize}
Note in passing that copula-based Markov models for continuous response data have been studied before in \citet{chen-fan-06}.
Assuming a copula based Markov model, the transition cdf of $Y_{tj}$ given $Y_{t-1,j}$ is
\begin{eqnarray}\label{transition-cdf}F_{j|t}(y_{tj}|y_{t-1,j})&=&P(Y_{tj}\leq
y_{tj}|Y_{t-1,j}=y_{t-1,j})\\&=&\Bigl[C_{j|t}\bigl(F(y_{t-1,j}),F(y_{tj})\bigr)-C_{j|t}\bigl(F(y_{t-1,j}-1),F(y_{tj})\bigr)\Bigr]/f_j(y_{t-1,j}),\nonumber
\end{eqnarray}
and the transition pmf is
$$f_{j|t}(y_{tj}|y_{t-1,j})=P(Y_{tj}=y_{tj}|Y_{t-1,j}=y_{t-1,j})=\frac{f(y_{tj},y_{t-1,j})}{f_j(y_{t-1,j})},$$
where
$f(y_t,y_{t-1})=C_{j|t}\bigl(F(y_t),F(y_{t-1})\bigr)-C_{j|t}\bigl(F(y_t-1),F(y_{t-1})\bigr)-C_{j|t}\bigl(F(y_t),F(y_{t-1}-1)\bigr)+C_{j|t}\bigl(F(y_t-1),F(y_{t-1}-1)\bigr)$.
Then the log-likelihood for each ordinal time-series is
\begin{equation}\label{serlik}\ell_{j|t}= \sum_{i=1}^n\left(\log f_j(y_{i1j};\mu_{i1j},\boldsymbol{\gamma}_j)+\sum_{t=2}^T\, \log f_{j|t}(y_{itj}|y_{i,t-1,j};\mu_{itj},\mu_{i,t-1,j},\boldsymbol{\gamma}_j)\right).
\end{equation}
Such framework incorporates the BVN copula as a special case, i.e. ``autoregressive-to-anything'' in \cite{biller&nelson2005} as also acknowledged by \cite{joe2014}. Stronger clustering of consecutive large/small values than expected in the BVN would require alternative copulas to obtain more appropriate transition probabilities.
\subsubsection{\label{sec-families}Choices of parametric families of copulas} In our candidate set, families that are in line with the conditions under which a copula function generates a stationary Markov chain that satisfies mixing conditions at a geometric rate \citep{chen-etal-2009,Beare2010} are used. These families
have different strengths of tail behaviour (see e.g., \cite{nikoloulopoulos&joe&li11,
nikoloulopoulos&joe12}):
\footnote{A bivariate copula $C$ is {\it reflection symmetric}
if its density $c(u_1,u_2)=\partial^2 C(u_1,u_2)/\partial u_1\partial u_2$
satisfies $c(u_1,u_2)=c(1-u_1,1-u_2)$ for all $0\leq u_1,u_2\leq 1$
and {\it reflection asymmetric} otherwise often with more probability in the
joint upper tail or joint lower tail. {\it Upper tail dependence} implies
$c(1-u,1-u)=O(u^{-1})$ as $u\to 0$ and {\it lower tail dependence}
that $c(u,u)=O(u^{-1})$ as $u\to 0$.
If $(U_1,U_2)\sim C$ for a bivariate copula $C$, then $(1-U_1,1-U_2)\sim
C_{180^0}$, with $C_{180^0}(u_1,u_2)=u_1+u_2-1+C(1-u_1,1-u_2)$ being the survival (or rotated by 180 degrees) copula of $C$. The ``reflection"
of each uniform $U(0,1)$ random variable by about $1/2$ changes the direction
of tail asymmetry.}
\begin{itemize}
\itemsep=0pt
\item Frank copula is reflection symmetric satisfying tail independence
$C(u,u)=O(u^2)$ and ${\overline C}(1-u,1-u)=O(u^2)$ as $u\to 0$,
with cdf
$$C(u_1,u_2;\theta)=-\theta^{-1}\log \left\{1+\frac{(e^{-\theta u_1}-1)(e^{-\theta
u_2}-1)}{e^{-\theta}-1} \right\},\hspace{2ex} \theta \in (-\infty,\infty)\setminus\{0\}.$$
\item Gumbel extreme value copula is reflection asymmetric with upper tail dependence and cdf
$$C(u_1,u_2;\theta)=\exp\Bigl[-\Bigl\{(-\log u_1)^{\theta}
+(-\log u_2)^{\theta}\Bigr\}^{1/\theta}\Bigr],\hspace{2ex} \theta\geq 1.$$
Compared to the BVN copula, the resulting model has more probability in the joint upper tail accommodating more dependence of large ordinal values than expected in the BVN.
\item Survival Gumbel (s.Gumbel) copula is reflection asymmetric with lower tail dependence and cdf
$$C(u_1,u_2;\theta)=u_1+u_2-1 + \exp\Bigl[-\Bigl\{\bigl(-\log (1-u_1)\bigr)^{\theta}
+\bigl(-\log (1-u_2)\bigr)^{\theta}\Bigr\}^{1/\theta}\Bigr],\hspace{2ex} \theta\geq 1.$$
Compared to the BVN copula, this resulting model has more probability in the joint lower tail accommodating more dependence of small ordinal values than expected in the BVN.
\item Bivariate Student t (BVT) copula with reflection symmetric upper and lower tail dependence and cdf
$$C(u_1,u_2;\theta)=\mathcal{T}_2\Bigl(\mathcal{T}^{-1}(u_1;\nu),\mathcal{T}^{-1}(u_2;\nu);\theta,\nu\Bigr),\hspace{2ex}-1\leq\theta\leq 1,$$
where $\mathcal{T}(;\nu)$ is the univariate Student t cdf with (non-integer) $\nu$ degrees of freedom, and $\mathcal{T}_2$ is the
cdf of a bivariate Student t distribution with $\nu$ degrees of freedom and correlation parameter $\theta$.
Small values of $\nu$ (i.e. $1\le \nu\le 5$) lead to models with more probabilities in the joint upper and joint lower tails accommodating more dependence of large and small ordinal values that would be expected with BVN.
\end{itemize}
\citet{Nikoloulopoulos&karlis08CSDA} have shown that, when using real data, copulas with similar (tail) dependence properties provide similar fit making selection among them cumbersome.
With tail dependence properties being copula family specific, upper/lower tail dependence becomes one way to differentiate among families. Contour plots of copula densities with standard normal margins and dependence parameters corresponding to Kendall's $\tau=0.6$ are given in Figure \ref{contours} to depict concepts of reflection (a)symmetric tail (in)dependence.
\begin{figure}[!t]
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=0.3\textwidth]{bvn.pdf}
&
\includegraphics[width=0.3\textwidth]{bvt.pdf}\vspace{-0.5cm}\\
\multicolumn{2}{c}{\includegraphics[width=0.3\textwidth]{frank.pdf}}\vspace{-0.5cm}
\\
\includegraphics[width=0.3\textwidth]{gumbel.pdf}
&
\includegraphics[width=0.3\textwidth]{sgumbel.pdf}
\\
\end{tabular}
\caption{\label{contours}Contour plots of BVN, BVT with 4 degrees of freedom (df), Frank, Gumbel and s.Gumbel copulas with standard normal margins and dependence parameters corresponding to Kendall's $\tau$ value of $0.6$. }
\end{center}
\end{figure}
\subsection{Joint copula-based Markov models for ordinal time-series}
So far we treat the $d$ ordinal time-series responses separately as if they were independent. In this Section, we propose relating these responses using an MVT copula to join their conditional (on the past) distributions at each time point.
Consider a multivariate discrete regression setup in which the $d \geq 2$ dependent ordinal time-series $Y_{t1}, \ldots, Y_{td}$ are observed together with a vector $\mathbf{x} \in \mathbb{R}^p$ of explanatory variables. If $C(\cdot;\mathbf{R})$ is the MVT copula (or any other parametric family of copulas) and $F_{j|t}(y_{tj}|y_{t-1,j})$, as defined in (\ref{transition-cdf}),
is the parametric model for the $j$th univariate ordinal time-series then $$C\Bigl(F_{1|t}(y_{t1}|y_{t-1,1}),\ldots,F_{d|t}(y_{td}|y_{t-1,d});\mathbf{R}\Bigr)$$ is a multivariate parametric model with univariate margins $F_{1|t},\ldots,F_{d|t}$. For copula models, the response vector $\mathbf{Y}=(Y_1,\ldots,Y_d)$ can be discrete \citep{Nikoloulopoulos2013a,nikoloulopoulos&joe12}.
Then it follows that the joint pmf is
\begin{equation}\label{jointserpmf}
f_{1\ldots d|t}(\mathbf{y};\boldsymbol{\beta},\boldsymbol{\gamma},\mathbf{R})
=\int_{\mathcal{T}^{-1}(F_{1|t}^{-})}^{\mathcal{T}^{-1}(F_{1|t}^{+})}\cdots
\int_{\mathcal{T}^{-1}(F_{d|t}^{-})}^{\mathcal{T}^{-1}(F_{d|t}^{+})} t_d(z_1,\ldots,z_d;\mathbf{R}) dz_1\ldots dz_d,\nonumber
\end{equation}
where $F_{j|t}^{-}:=F_{j|t}(y_{tj}-1|y_{t-1,j})$, $F_{1|t}^{+}:=F_{j|t}(y_{tj}|y_{t-1,j})$ and $t_d(\cdot;\mathbf{R})$ denotes the MVT density with latent correlation matrix $\mathbf{R}$ and $\nu$ degrees of freedom. The MVN case can be treated as a special case of the MVT with a large value of $\nu$.
For the joint copula-based Markov model, we let $C_{j|t},\,j=1,\ldots,d$ and be parametric bivariate copulas, say with parameters $\theta_j,\,j=1,\ldots,d$ and $C$ be an MVT copula. For the set of all parameters, let $\boldsymbol{\theta}=\{\boldsymbol{\beta}_j,\boldsymbol{\gamma}_j,\theta_j,\mathbf{R}: j=1,\ldots,d\}$. We model the joint distribution in terms of $d$ bivariate copulas and an MVT copula. Note that the copula $C_{j|t}$ models the time-series for the $j$th response and the copula $C$ links that $j$ ordinal time-series responses. Our general statistical model allows for selection of $C_{j|t}$ independently among a variety of parametric copula families, i.e., there are no constraints in the choices of parametric copulas $\{C_{j|t}: j=1,\ldots,d\}$.
\subsection{\label{concept}Conceptual framework}
Let $(Y_1,\ldots,Y_m)$ and $(Y_{m+1},\ldots,Y_d)$ denote the generic and domain satisfactions, respectively. We propose an expansive set of relationships: objective individual characteristics and covariates
directly relate to both generic and domain satisfactions, with $(Y_{m+1},\ldots,Y_d)$ further influencing $(Y_1,\ldots,Y_m)$, while inter-dependencies among them are also allowed. Such latent correlations capture the residual dependence (i.e. over and above the effect of covariates) among equations/outcomes. Figure \ref{fig:schematic} gives a schematic of our structural model. Such specification allows for direct, indirect and ripple (spill-over) effects on well-being, e.g. capturing at least three possible ways high individual income could affect generic satisfaction: a) a direct effect of income, b) an indirect effect through income satisfaction and c) an indirect effect through increased income satisfaction that itself is the result of an improved, for example, leisure satisfaction that was caused by the initial increase in income.
\begin{figure}[!h]
\begin{center}
\begin{tabular}{cc}
\includegraphics[trim = 20mm 170mm 20mm 20mm, scale=.7]{schematic.pdf}
\end{tabular}
\vspace{-1.5cm}
\caption{Life and domain satisfactions conceptual framework where solid and dotted lines indicate regression coefficients and latent correlations, respectively.}
\label{fig:schematic}
\end{center}
\end{figure}
\section{\label{estimation}Estimation techniques and computational details}
For estimation purposes we propose a maximum simulated likelihood method, which is based on evaluating the multidimensional integrals of the likelihood with randomized quasi Monte Carlo methods; an analysis of asymptotic properties of the estimators is shown in the Appendix.
\subsection{Simulated likelihood}
The log-likelihood of the joint copula-based Markov model is
\begin{equation}\label{jointserlik}\ell_{1\ldots d|t}(\boldsymbol{\theta})= \sum_{i=1}^n\left(\log f_{1\ldots d}(y_{i11},\ldots,y_{i1d};\boldsymbol{\theta}) + \sum_{t=2}^T\, \log f_{1\ldots d|t}(y_{it1},\ldots,y_{itd};\boldsymbol{\theta})\right).
\end{equation}
where $f_{1\ldots d|t}(\cdot)$ is given in (\ref{jointserpmf}) and
\begin{equation}\label{jointserpmf1}
f_{1\ldots d}(\mathbf{y}_1;\boldsymbol{\theta})=
\int_{\mathcal{T}^{-1}[F_{1}(y_{1}-1;\boldsymbol{\beta}_1,\boldsymbol{\gamma})]}^{\mathcal{T}^{-1}[F_{1}(y_{1};\boldsymbol{\beta}_1,\boldsymbol{\gamma})]}\cdots
\int_{\mathcal{T}^{-1}[F_{d}(y_{d}-1;\boldsymbol{\beta}_d,\boldsymbol{\gamma})]}^{\mathcal{T}^{-1}[F_{d}(y_{d};\boldsymbol{\beta}_d,\boldsymbol{\gamma})]} t_d(z_1,\ldots,z_d;\mathbf{R}) dz_1\ldots dz_d.
\end{equation}
We develop and implement a maximum simulated likelihood estimator (MSLE). There exist general results on asymptotics of simulated likelihood based
estimators (see, e.g., \citealp{Gourieroux&Monfort-1991}). They usually involve a rate assumption on the number of simulations versus the sample size.
Nevertheless,
we propose a simulated likelihood method, where the rectangle MVT probabilities in (\ref{jointserpmf}) and (\ref{jointserpmf1}) are computed using a quasi Monte Carlo method proposed by \cite{genz&bretz02}.
\cite{genz&bretz02} achieve error reduction of Monte Carlo methods through variance reduction techniques such as (a) transforming to a bounded integrand, (b) using antithetic variates, and (c) using a randomized quasi Monte Carlo method. The test results in \cite{genz&bretz02,genz&bretz2009} show that their method is very efficient, compared to other methods in the literature.
The method in \cite{genz&bretz02} is ``optimized" in the {\tt mtvnorm} R package \citep{genz-etal-2012}. Hence, on the calculation of the MSLE, one doesn't need to worry about the selection, for example, of the number of simulated quasi points.
The estimated parameters can be obtained by maximizing the simulated log-likelihood in (\ref{jointserlik}) over the model parameters $\boldsymbol{\theta}$. The method was initially proposed for the analysis of discrete (binary and count) longitudinal data by \cite{nikoloulopoulos13b} and extended to a high-dimensional context in \cite{nikoloulopoulos2015a}. We refer the interested reader to these papers for more details including studies of small-sample and asymptotic efficiency for Bernoulli, Poisson, and negative binomial regression models. In addition to that we study here the asymptotic properties of the maximum simulated likelihood estimators for ordinal regression models in an Appendix.
\subsection{\label{sec-compdet}Computational details}
The MSLEs can be derived using a three-step procedure:
\begin{enumerate}
\itemsep=0pt
\item For each $j$:
\begin{enumerate}
\itemsep=0pt
\item Assuming time independence, $\ell_j(\boldsymbol{\beta}_j,\boldsymbol{\gamma}_j)$ in (\ref{indlik}) is maximized over the univariate marginal parameters $\boldsymbol{\beta}_j,\boldsymbol{\gamma}_j$.
\item Keeping the univariate parameters $\boldsymbol{\beta}_j,\boldsymbol{\gamma}_j$ fixed at the values estimated in (a), the $\ell_{j|t}(\boldsymbol{\beta}_j,\boldsymbol{\gamma}_j,\theta_j)$ in (\ref{serlik}) is maximized over the copula parameter $\theta_j$.
\item Finally, using starting values from the estimates above the $\ell_{j|t}(\boldsymbol{\beta}_j,\boldsymbol{\gamma}_j,\theta_j)$ in (\ref{serlik}) is maximized over both the univariate $\boldsymbol{\beta}_j,\boldsymbol{\gamma}_j$ and copula $\theta_j$ parameters.
\end{enumerate}
\item Setting all parameters to their estimated values from the first step, the $\ell_{1\ldots d|t}(\boldsymbol{\theta})$ in (\ref{jointserlik}) is maximized over $\mathbf{R}$.
\item At the third and final step the $\ell_{1\ldots d|t}(\boldsymbol{\theta})$ in (\ref{jointserlik}) is maximized over $\boldsymbol{\theta}$ with initial parameters the estimates for the preceding steps.
\end{enumerate}
Given the typical large number of estimable (univariate and copula) parameters in multivariate models one can restrict themselves to only the first two steps of the method to make inference computationally feasible.
This two-step approach is known in the copula literature as the Inference Function of Margins (IFM) method \citep{joe&xu1996,joe97} and its asymptotic
efficiency has been established \citep{joe05}. Hence using only the two first steps,
\begin{itemize}
\item the model parameters can be efficiently (in the sense of computing time and asymptotic variance) estimated;
\item cross-model comparisons with respect to dependence structure and subsequently predictions and inferences can be performed.
\end{itemize}
Note also in passing that compared to the (simulated) maximum likelihood, the IFM method is not as punishing for misspecification of the dependence structure \citep{joe&xu1996,xu96}.
Each of the estimated parameters can be obtained by using a quasi-Newton \citep{nash90} method applied to the log-likelihood. This numerical method requires only the objective function, i.e., the joint log-likelihood, while the gradients
are computed numerically and the Hessian matrix of the second order derivatives is updated in each iteration.
Since the estimation of parameters in MVT copula-based models is obtained using a quasi-Newton routine \citep{nash90}
applied to the log-likelihood in (\ref{jointserlik}), the use of randomized quasi Monte Carlo simulation to four decimal place
accuracy for evaluations of integrals works poorly, because numerical derivatives of the log-likelihood with respect to
the parameters are not smooth. In order to achieve smoothness, the same set of uniform random variables should be used for every rectangle probability that comes up in the optimization of the simulated likelihood \citep{nikoloulopoulos13b,
nikoloulopoulos2015a}.
\section{\label{vuong-sec}Vuong's test for model comparison}
A methodology for the comparison of non-nested models using the Vuong's test \citep{vuong1989} is formulated below to test if:
\begin{enumerate}
\item the copula-based Markov models with different choices of bivariate copulas outperform the copula-based Markov model with BVN (i.e. ``autoregressive-to-anything'' model);
\item the joint copula-based Markov models with MVT copulas provide better fit than their special case, namely the MVN copula.
\end{enumerate}
The Vuong's test is appropriate for parametric non-nested models comparisons and has often been used in the copula literature (e.g., \citealp{belgorodski10,Brechmann-Czado-Aas-2012,joe2014,Nikoloulopoulos2015b}).
Assume models 1 and 2 with parametric densities $f^{(1)}$ and $f^{(2)}$, respectively. Comparison of
$$\Delta_{1f^\maltese}=N^{-1}\Bigl[\sum_i\{E_{f^\maltese}[\log f^\maltese(\cdot)]-E_{f^\maltese}[\log f^{(1)}(\cdot;\boldsymbol{\theta}^{(1)})]\}\Bigr],$$
and
$$\Delta_{2f^\maltese}=N^{-1}\Bigl[\sum_i\{E_{f^\maltese}[\log f^\maltese(\cdot)]-E_{f^\maltese}[\log f^{(2)}(\cdot;\boldsymbol{\theta}^{(2)})]\}\Bigr],$$
where $\boldsymbol{\theta}^{(1)},\boldsymbol{\theta}^{(2)}$ are the parameters in models 1 and 2 respectively that lead to the closest Kullback-Leibler divergence to the true $f^\maltese$. Model 1 is closer to the true $f^\maltese$, i.e., fits better if $\Delta=\Delta_{1f^\maltese}-\Delta_{2f^\maltese}<0$, while model 2 fits better if $\Delta>0$. The sample version of $\Delta$ with estimates $\hat\boldsymbol{\theta}^{(1)},\hat\boldsymbol{\theta}^{(2)}$ is
$$\bar D=\sum_{i=1}^N D_i/N,$$
where $D_i=\log\left[\frac{f^{(2)}\left(\cdot;\hat\boldsymbol{\theta}^{(2)}\right)}{f^{(1)}\left(\cdot;\hat\boldsymbol{\theta}^{(1)}\right)}\right]$.
\cite{vuong1989}
has shown that asymptotically under the null hypothesis $H_0:\Delta=0$, i.e., models 1 and 2 have the same parametric densities $f^{(1)}$ and $f^{(2)}$,
$$z_0=\sqrt{N}\bar D/s\widesim{H_0}\mathcal{N}(0,1),$$
where $s^2=\frac{1}{N-1}\sum_{i=1}^N(D_i-\bar D)^2$.
Rejection on the null hypothesis follows if $\abs{z_0}$ is greater than the critical value from the standard normal distribution, denoted $\mathcal{N}(0,1)$.
\section{\label{sec-appl} British Household Panel Survey}
For the estimation of the model in Figure \ref{fig:schematic} we use data from the British Household Panel Survey (BHPS). The BHPS was an annual longitudinal survey (now superseded by Understanding Society) carried out by the Institute for Social and Economic Research sampling about 10,000 individuals aged 16 years or over from 1991 to 2008. However, the survey modules required for our model are collected on a bi-annual basis starting in wave 6 (i.e. 1996) and finishing in wave 18 (i.e. 2008), resulting in a maximum of 7 measurements per individual.
To ease coding, we use individuals observed at all seven time points resulting in a sample size of 4186 individuals, though our methodology does not depend on a constant ``cluster" size $T$.
\subsection{Equations and covariates}
The actual model estimated is based on one generic $Y_1$ and six domain satisfaction $(Y_2,Y_3,\ldots,Y_7)$ questions each answered on a 1 (not satisfied at all) to 6 (completely satisfied) likert scale. The seven equations (outcomes) are:
\begin{description}
\itemsep=0pt
\item[$Y_1$:] Satisfaction with Life overall
\item[$Y_2$:] Satisfaction with Health
\item[$Y_3$:] Satisfaction with Income
\item[$Y_4$:] Satisfaction with House/flat
\item[$Y_5$:] Satisfaction with Spouse/partner
\item[$Y_6$:] Satisfaction with Job
\item[$Y_7$:] Satisfaction with Leisure
\end{description}
Each of $(Y_2,Y_3,\ldots,Y_7)$ is conditioned upon individual characteristics (i.e. age, age square, gender, household size, number of kids, education and geographical region within the UK) that are common controls in the well-being literature and domain specific factors that appear only on the respective domain equation and would enhance the ability to capture domain specific variation.
For example, number of health problems are used for the satisfaction with health equation, disaggregated sources of income in the satisfaction with income equation and so on for the rest of the equations. Table \ref{tab:desstat} offers definitions and breakdowns for domain specific variables, as well as the reference levels for all categorical variables. Note that the original {\emph{region}} variable is a geographical identifier that splits UK into 19 areas and which is aggregated into 5 ``super''-regions for the regressions. Further, income rank captures relative income and is determined by the rank of the individual in their original region (i.e. one of 19 areas) according to their total annual income.
For equation $Y_1$ we use as covariates all common and domain specific variables capturing the direct effects that each characteristic has on overall life satisfaction. This formulation allows the separate identification of the direct effect of
domain characteristics and the effect of composite domains on generic satisfaction, while, as the econometric model poses, residual dependence between and among $Y_1$ and $(Y_2,Y_3,\ldots,Y_7)$ is captured by estimable latent correlation parameters.
Given the multivariate nature of our model we restrict the estimation sample to those for whom information on each equation is available, i.e. married, employed individuals (up to 70 years old) who indicate to have a partner. Further, as discussed earlier, albeit not a requirement of the model, we only keep those individuals that appear in all seven waves of the data. Descriptive statistics for all equation outcomes, and covariates are available in Table \ref{tab:desstat} where informal sample selection comparisons can also be drawn between the estimation and full sample.
Respondents tend to be most satisfied with their partners, followed by satisfaction with their houses and then with life overall. Income and leisure are the least satisfactory dimensions. Overall, our sample is 46\% males and on average 45 years old, with 12\% having a higher degree and 7\% residing in London, 20\% in South England, 16\% in the Midlands and 19\% in North England. Comparing restricted and full samples the former is slightly less educated and with a different income sources structure but on average individual characteristics and domain covariates are largely comparable across the two samples.
\begin{table}[!b]
\centering
\caption{\label{tab:desstat}Descriptive statistics for generic and domain satisfaction outcomes and all common and domain specific covariates for estimation and full samples.}
\begin{tabular}{lccc}
\hline & Estimation sample & & Full sample \\ \cline{2-2} \cline{4-4} \\
Equations & & & \\ \cline{1-1}
$Y_1$: Satisfaction with Life overall & 4.25 & & 4.29 \\
$Y_2$: Satisfaction with Health & 4.00 & & 4.23 \\
$Y_3$: Satisfaction with Income & 3.64 & & 3.78 \\
$Y_4$: Satisfaction with House/flat & 4.45 & & 4.44 \\
$Y_5$: Satisfaction with Spouse/partner & 5.24 & & 5.27 \\
$Y_6$: Satisfaction with Job & 4.05 & & 4.01 \\
$Y_7$: Satisfaction with Leisure & 3.93 & & 3.77 \\
& & & \\
Covariates & & & \\ \cline{1-1}
Age & 4.53 & & 4.36 \\
Age$^2$ & 23.98 & & 19.72 \\
Sex (1 if male) & 0.46 & & 0.55 \\
Household size & 0.29 & & 0.33 \\
\# of kids & 0.06 & & 0.09 \\
Education (ref category: Uni degree) & & & \\
\hspace{2ex}hnd,hnc, a/o levels, cse & 0.55 & & 0.69 \\
\hspace{2ex}No education & 0.33 & & 0.11 \\
Region (ref category: London) & & & \\
\hspace{2ex}South & 0.20 & & 0.30 \\
\hspace{2ex}Midlands & 0.16 & & 0.22 \\
\hspace{2ex}North & 0.19 & & 0.32 \\
\hspace{2ex}RUK & 0.38 & & 0.10 \\
Health & & & \\
\hspace{2ex}\# Health problems & 0.12 & & 0.08 \\
Income & & & \\
\hspace{2ex}ln(Labour Income) & 7.53 & & 10.47 \\
\hspace{2ex}ln(Pension Income) & 1.98 & & 0.68 \\
\hspace{2ex}ln(Benefit Income) & 5.89 & & 4.80 \\
\hspace{2ex}ln(Transfer Income) & 0.70 & & 0.41 \\
\hspace{2ex}ln(Investment Income) & 3.46 & & 4.19 \\
\hspace{2ex}Regional incomne rank (stadardized) & 0.00 & & -0.07 \\
House type (ref category: Detached) & & & \\
\hspace{2ex}Semi-detached & 0.32 & & 0.43 \\
\hspace{2ex}Terraced & 0.27 & & 0.21 \\
\hspace{2ex}Other & 0.15 & & 0.02 \\
House value (ref category: 0-50K) & & & \\
\hspace{2ex}50K-100K & 0.35 & & 0.28 \\
\hspace{2ex}100K-175K & 0.23 & & 0.25 \\
\hspace{2ex}175K-250K & 0.14 & & 0.19 \\
\hspace{2ex}$>$250K & 0.11 & & 0.15 \\
\hline \multicolumn{4}{r}{{Continued on next page}}
\end{tabular}
\end{table}
\setcounter{table}{0}
\begin{table}[!t]
\centering
\caption{Continued.}
\begin{tabular}{lccc}
\hline & Estimation sample & & Full sample \\ \cline{2-2} \cline{4-4} \\
Spouse Charateristics & & & \\
\hspace{2ex} Age of spouse & 4.70 & & 4.37 \\
\hspace{2ex} Age$^2$ of spouse & 24.46 & & 19.92 \\
\hspace{2ex} Sex of spouse & 0.50 & & 0.45 \\
\hspace{2ex} Education (ref category: Uni degree) & & & \\
\hspace{4ex} hnd,hnc, a/o levels, cse & 0.55 & & 0.67 \\
\hspace{4ex} No education & 0.32 & & 0.16 \\
\hspace{2ex}ln(Labour Income) & 6.35 & & 8.46 \\
\hspace{2ex}\# Health problems & 0.11 & & 0.09 \\
\hspace{2ex}Hours of wk housework & 1.19 & & 1.18 \\
Satisfaction with job pay (ref category: Low) & & & \\
\hspace{2ex}Medium & 0.33 & & 0.33 \\
\hspace{2ex}Very & 0.45 & & 0.49 \\
Satisfaction with job security (ref category: Low) & & & \\
\hspace{2ex}Medium & 0.26 & & 0.31 \\
\hspace{2ex}Very & 0.61 & & 0.57 \\
Satisfaction with work itself (ref category: Low) & & & \\
\hspace{2ex}Medium & 0.28 & & 0.30 \\
\hspace{2ex}Very & 0.62 & & 0.61 \\
Satisfaction with hours worked (ref category: Low) & & & \\
\hspace{2ex}Medium & 0.31 & & 0.33 \\
\hspace{2ex}Very & 0.54 & & 0.51 \\
Leisure activities (1 if several times a year) & & & \\
\hspace{2ex}walk/swim/play sport & 0.67 & & 0.75 \\
\hspace{2ex}watch live sport & 0.25 & & 0.33 \\
\hspace{2ex}cinema & 0.47 & & 0.53 \\
\hspace{2ex}theatre/concert & 0.36 & & 0.37 \\
\hspace{2ex}out for a drink & 0.69 & & 0.84 \\
\hspace{2ex}work in garden & 0.64 & & 0.86 \\
\hspace{2ex}diy, car maintenance & 0.53 & & 0.72 \\
\hspace{2ex}attend evening classes & 0.27 & & 0.30 \\
\hspace{2ex}attend local groups & 0.20 & & 0.21 \\ \hline
\end{tabular}
\end{table}
\subsection{Fitted copula-based Markov models for each ordinal time-series}
We fit the
copula-based Markov model with BVN, Gumbel, s.Gumbel, and BVT bivariate linking copulas. For BVT, choices of $\nu$ were $1,2,\ldots,10$. To make it easier to compare the dependence parameters, we convert the estimated parameters to Kendall's $\tau$'s in $(0,1)$ via the relations
$\tau=\frac{2}{\pi}\arcsin{\theta}$,
$\tau=1+4\theta^{-1}\left[\frac{1}{\theta}\int_0^{\theta}\frac{t}{e^t-1}dt-1\right],$
and
$\tau=1-\theta^{-1}$
for elliptical, Frank and Gumbel copulas in \cite{HultLindskog02}, \cite{genest87}, and \cite{genest&mackay86}, respectively.
Note that Kendall's tau only accounts for the dependence dominated by the middle of the data, and it is expected to be similar amongst different families of copulas. However, the tail dependence varies, as explained in Section \ref{sec-families}, and is a property to consider when choosing amongst different families of copulas. For the model with BVT we summarize the choice of integer $\nu$ with the largest log-likelihood.
Given the equality in number of parameters between models, the log-likelihood at estimates can be used as a measure for goodness of fit across all models. We further compute the \nocite{vuong1989}Vuong's (1989) test to formally assess if more probability is accumulated in the joint tails than one would expect via a BVN copula.
For these data, if a respondent thinks about the maximum or minimum satisfaction at year $t$ it seems natural to think about the maximum or minimum satisfaction at year $t+1$ and year $t-1$. That is, based on data descriptions, we could expect a priori that a model with $C_{j|t}$ being the BVT copulas might be plausible, as data have more probability in the joint tails.
\begin{table}[!b]
\centering \caption{\label{tab:lfsato} Estimated parameters and joint log-likelihoods $\ell_{j|t}$ for the copula-based Markov models for ordinal time-series with covariates for satisfaction with life overall $Y_1$, where a parametric copula family $C_{j|t}$ is used for the joint distribution of subsequent observations. For the best fit, the standard errors (SE) of the estimates, Wald tests ($Z$) and $p$-values are also presented. }
\begin{scriptsize}
\begin{tabular}{lcccccccc} \hline
& BVN & Frank & Gumbel & s.Gumbel & \multicolumn{4}{c}{BVT, $\nu=4$} \\ \cline{2-9}
& Est. & Est. & Est. & Est. & Est. & SE & $Z$ & $p$-value \\ \cline{2-9}
$\alpha_{1}$ & -2.93 & -3.24 & -2.88 & -2.63 & -2.63 & 0.83 & -3.18 & 0.01 \\
$\alpha_{2}$ & -2.07 & -2.38 & -1.98 & -1.90 & -1.81 & 0.82 & -2.20 & 0.03 \\
$\alpha_{3}$ & -1.33 & -1.64 & -1.22 & -1.23 & -1.09 & 0.82 & -1.32 & 0.19 \\
$\alpha_{4}$ & -0.14 & -0.47 & -0.03 & -0.09 & 0.09 & 0.83 & 0.11 & 0.92 \\
$\alpha_{5}$ & 1.44 & 1.10 & 1.48 & 1.50 & 1.66 & 0.83 & 2.01 & 0.04 \\
Age & -0.74 & -0.84 & -0.58 & -0.70 & -0.55 & 0.31 & -1.79 & 0.07 \\
Age$^2$ & 0.10 & 0.11 & 0.08 & 0.10 & 0.08 & 0.03 & 2.23 & 0.03 \\
Sex & 0.13 & 0.45 & 0.32 & -0.09 & 0.08 & 0.37 & 0.20 & 0.84 \\
Household size & -0.72 & -0.70 & -0.78 & -0.64 & -0.70 & 0.32 & -2.16 & 0.03 \\
\# of kids & 0.10 & 0.01 & -0.03 & 0.08 & -0.03 & 0.34 & -0.10 & 0.93 \\
Education & & & & & & & & \\
\hspace{2ex}No education & 0.08 & 0.09 & 0.09 & 0.11 & 0.11 & 0.11 & 1.06 & 0.29 \\
\hspace{2ex} hnd,hnc, a/o levels, cse & 0.03 & 0.03 & 0.05 & 0.03 & 0.05 & 0.07 & 0.73 & 0.46 \\
Region & & & & & & & & \\
\hspace{2ex}South & 0.22 & 0.25 & 0.28 & 0.18 & 0.24 & 0.11 & 2.08 & 0.04 \\
\hspace{2ex}Midlands & 0.12 & 0.18 & 0.17 & 0.09 & 0.12 & 0.12 & 1.03 & 0.30 \\
\hspace{2ex}North & 0.16 & 0.15 & 0.22 & 0.16 & 0.21 & 0.12 & 1.82 & 0.07 \\
\hspace{2ex}RUK & 0.30 & 0.32 & 0.29 & 0.31 & 0.29 & 0.16 & 1.85 & 0.07 \\
Health & & & & & & & & \\
\hspace{2ex}\# Health problems & -1.43 & -1.53 & -1.46 & -1.24 & -1.34 & 0.19 & -7.05 & 0.00 \\
Income & & & & & & & & \\
\hspace{2ex}ln(Labour Income) & 0.05 & 0.01 & 0.03 & 0.07 & 0.05 & 0.06 & 0.95 & 0.34 \\
\hspace{2ex}ln(Pension Income) & 0.01 & 0.00 & 0.00 & 0.00 & 0.00 & 0.01 & 0.20 & 0.84 \\
\hspace{2ex}ln(Benefit Income) & -0.01 & -0.00 & -0.01 & -0.01 & -0.01 & 0.01 & -1.01 & 0.31 \\
\hspace{2ex}ln(Transfer Income) & -0.01 & -0.01 & -0.01 & -0.00 & -0.00 & 0.01 & -0.29 & 0.77 \\
\hspace{2ex}ln(Investment Income) & 0.01 & 0.01 & 0.01 & 0.01 & 0.01 & 0.01 & 1.29 & 0.20 \\
\hspace{2ex}Regional income rank (standardized) & -0.01 & 0.02 & 0.01 & -0.03 & -0.02 & 0.05 & -0.33 & 0.74 \\
House type & & & & & & & & \\
\hspace{2ex}Semi-detached & -0.03 & -0.049 & -0.04 & -0.04 & -0.04 & 0.05 & -0.80 & 0.42 \\
\hspace{2ex}Terraced & -0.09 & -0.12 & -0.09 & -0.11 & -0.10 & 0.07 & -1.56 & 0.12 \\
\hspace{2ex}Other & -0.45 & -0.44 & -0.41 & -0.44 & -0.43 & 0.15 & -2.80 & 0.01 \\
House value & & & & & & & & \\
\hspace{2ex}50K-100K & -0.06 & -0.03 & -0.06 & -0.07 & -0.07 & 0.06 & -1.21 & 0.23 \\
\hspace{2ex}100K-175K & -0.12 & -0.07 & -0.10 & -0.15 & -0.13 & 0.07 & -1.79 & 0.07 \\
\hspace{2ex}175K-250K & -0.15 & -0.12 & -0.12 & -0.17 & -0.15 & 0.08 & -1.82 & 0.07 \\
\hspace{2ex}$>$250K & -0.12 & -0.09 & -0.09 & -0.14 & -0.12 & 0.10 & -1.20 & 0.23 \\
Spouse Charateristics & & & & & & & & \\
\hspace{2ex} Age & 0.07 & -0.03 & 0.03 & 0.07 & 0.04 & 0.28 & 0.13 & 0.90 \\
\hspace{2ex} Age$^2$ & -0.03 & -0.01 & -0.02 & -0.03 & -0.03 & 0.03 & -0.81 & 0.42 \\
\hspace{2ex} Sex & 0.13 & 0.46 & 0.31 & -0.07 & 0.08 & 0.37 & 0.22 & 0.83 \\
\hspace{2ex}Education & & & & & & & & \\
\hspace{4ex}No education & 0.10 & 0.06 & 0.10 & 0.12 & 0.12 & 0.10 & 1.16 & 0.25 \\
\hspace{4ex}hnd,hnc, a/o levels, cse & -0.05 & -0.03 & -0.09 & -0.04 & -0.09 & 0.07 & -1.15 & 0.25 \\
\hspace{2ex}ln(Labour Income) of spouse & 0.01 & 0.01 & 0.01 & 0.01 & 0.01 & 0.01 & 1.17 & 0.24 \\
\hspace{2ex}\# Health problems of spouse & -0.27 & -0.25 & -0.19 & -0.26 & -0.19 & 0.17 & -1.11 & 0.27 \\
\hspace{2ex}Hours of wk housework & -0.02 & -0.02 & -0.02 & -0.01 & -0.01 & 0.02 & -0.54 & 0.59 \\
Satisfaction with job pay & & & & & & & & \\
\hspace{2ex}Medium & 0.14 & 0.18 & 0.15 & 0.13 & 0.14 & 0.04 & 3.15 & 0.00 \\
\hspace{2ex}Very & 0.23 & 0.28 & 0.22 & 0.20 & 0.20 & 0.05 & 4.44 & 0.00 \\
Satisfaction with job security & & & & & & & & \\
\hspace{2ex}Medium & 0.10 & 0.10 & 0.09 & 0.11 & 0.10 & 0.05 & 2.14 & 0.03 \\
\hspace{2ex}Very & 0.18 & 0.18 & 0.16 & 0.17 & 0.16 & 0.05 & 3.42 & 0.00 \\
Satisfaction with work itself & & & & & & & & \\
\hspace{2ex}Medium & 0.14 & 0.16 & 0.12 & 0.14 & 0.14 & 0.05 & 2.62 & 0.01 \\
\hspace{2ex}Very & 0.41 & 0.41 & 0.37 & 0.39 & 0.38 & 0.05 & 7.03 & 0.00 \\\hline
\multicolumn{9}{r}{{Continued on next page}}
\end{tabular}
\end{scriptsize}
\end{table}
\setcounter{table}{1}
\begin{table}[!t]
\centering
\caption{Continued.}
\begin{scriptsize}
\begin{tabular}{lcccccccc} \hline
& BVN & Frank & Gumbel & s.Gumbel & \multicolumn{4}{c}{BVT, $\nu=4$} \\ \cline{2-9}
& Est. & Est. & Est. & Est. & Est. & SE & $Z$ & $p$-value \\ \cline{2-9}
Satisfaction with hours worked & & & & & & & & \\
\hspace{2ex}Medium & 0.21 & 0.19 & 0.19 & 0.20 & 0.19 & 0.04 & 4.32 & 0.00 \\
\hspace{2ex}Very & 0.32 & 0.29 & 0.30 & 0.29 & 0.30 & 0.05 & 6.21 & 0.00 \\
Frequency of leisure activities & & & & & & & & \\
\hspace{2ex}walk/swim/play sport & -0.02 & 0.00 & -0.02 & -0.03 & -0.04 & 0.04& -1.07 & 0.29 \\
\hspace{2ex}watch live sport & 0.06& 0.06 & 0.05 & 0.06 & 0.05 & 0.04 & 1.18 & 0.24 \\
\hspace{2ex}cinema & -0.03 & -0.03 & -0.02 & -0.04 & -0.03 & 0.04 & -0.99 & 0.32 \\
\hspace{2ex}theatre/concert & 0.01 & -0.01 & -0.00 & 0.02 & 0.01& 0.03 & 0.25 & 0.80\\
\hspace{2ex}out for a drink & 0.06 & 0.06 & 0.07 & 0.05 & 0.06 & 0.05 & 1.25 & 0.21 \\
\hspace{2ex}work in garden & 0.06 & 0.05 & 0.04 & 0.04 & 0.03 & 0.05 & 0.56 & 0.58 \\
\hspace{2ex}diy, car maintenance & 0.01 & 0.01 & 0.02 & -0.01 & 0.00 & 0.04 & -0.01 & 0.10 \\
\hspace{2ex}attend evening classes & 0.04 & 0.05 & 0.05 & 0.04 & 0.05 & 0.03 & 1.50 & 0.13 \\
\hspace{2ex}attend local groups & 0.01 & 0.00 & 0.01 & 0.01 & 0.01 & 0.04 & 0.22 & 0.82 \\ \hline
$\tau$ & 0.36 & 0.40 & 0.39 & 0.37 & 0.38 & 0.01 & 28.5 & 0.00 \\
$-\ell_{j|t}$ & 4734.1 & 4734.2 & 4719.1 & 4740.0 & \multicolumn{4}{c}{4689.2} \\ \hline
$z_0$ ($p$-value)&{-}& -0.010 (0.992) &1.159 (0.246) &-0.576 (0.565) &\multicolumn{4}{c}{3.419 (0.001)}\\\hline
\end{tabular}
\end{scriptsize}
\end{table}
\begin{table}[!h]
\centering \caption{\label{tab:lfsat1} Estimated parameters and joint log-likelihoods $\ell_{j|t}$ for the copula-based Markov models for ordinal time-series with covariates for satisfaction with health $Y_2$, where a parametric copula family $C_{j|t}$ is used for the joint distribution of subsequent observations. For the best fit, the standard errors (SE) of the estimates, Wald tests ($Z$) and $p$-values are also presented.}
\begin{scriptsize}
\begin{tabular}{lcccccccc} \hline
& BVN & Frank & Gumbel & s.Gumbel & \multicolumn{4}{c}{BVT, $\nu=6$} \\ \cline{2-9}
& Est. & Est. & Est. & Est. & Est. & SE & $Z$ & $p$-value \\ \cline{2-9}
$\alpha_1$ & -3.85 & -3.85 & -4.01 & -3.38 & -3.75 & 0.46 & -8.20 & 0.00 \\
$\alpha_2$ & -3.21 & -3.20 & -3.35 & -2.82 & -3.12 & 0.46 & -6.87 & 0.00 \\
$\alpha_3$ & -2.67 & -2.65 & -2.79 & -2.32 & -2.59 & 0.45 & -5.69 & 0.00 \\
$\alpha_4$ & -1.86 & -1.85 & -1.98 & -1.54 & -1.78 & 0.45 & -3.92 & 0.00 \\
$\alpha_5$ & -0.64 & -0.65 & -0.81 & -0.31 & -0.56 & 0.45 & -1.24 & 0.22 \\
Age & -0.75 & -0.72 & -0.82 & -0.63 & -0.74 & 0.21 & -3.49 & 0.00 \\
Age$^2$ & 0.08 & 0.08 & 0.09 & 0.07 & 0.08 & 0.02 & 3.44 & 0.00 \\
Sex & 0.02 & 0.01 & 0.01 & 0.01 & 0.02 & 0.05 & 0.32 & 0.75 \\
Household size & -0.49 & -0.44 & -0.31 & -0.51 & -0.42 & 0.29 & -1.47 & 0.14 \\
\# of kids & 0.36 & 0.26 & 0.22 & 0.43 & 0.34 & 0.31 & 1.10 & 0.27 \\
Education & & & & & & & & \\
\hspace{2ex}No education & 0.00 & -0.04 & 0.01 & 0.03 & 0.03 & 0.09 & 0.33 & 0.74 \\
\hspace{2ex} hnd,hnc, a/o levels, cse & 0.02 & 0.02 & 0.02 & 0.02 & 0.02 & 0.06 & 0.38 & 0.70 \\
Region & & & & & & & & \\
\hspace{2ex}South & 0.19 & 0.12 & 0.17 & 0.20 & 0.20 & 0.10 & 1.97 & 0.05 \\
\hspace{2ex}Midlands & 0.10 & 0.05 & 0.09 & 0.11 & 0.12 & 0.11 & 1.08 & 0.28 \\
\hspace{2ex}North & 0.07 & 0.01 & 0.04 & 0.11 & 0.09 & 0.10 & 0.87 & 0.38 \\
\hspace{2ex}RUK & 0.24 & 0.24 & 0.20 & 0.25 & 0.23 & 0.12 & 1.93 & 0.05 \\
Health & & & & & & & & \\
\hspace{2ex}\# Health problems & -3.59 & -3.45 & -3.34 & -3.45 & -3.42 & 0.20 & -16.8 & 0.00 \\ \hline
$\tau$ & 0.34 & 0.37 & 0.36 & 0.34 & 0.35 & 0.01 & 28.3 & 0.00 \\
$-\ell_{j|t}$ & 5752.2 & 5741.8 & 5739.9 & 5773.2 & \multicolumn{4}{c}{5725.9} \\ \hline
$z_0$ ($p$-value)&{-}&1.112 (0.266)& 1.147 (0.251)& -1.916 (0.055)& \multicolumn{4}{c}{3.009 (0.003)}\\\hline
\end{tabular}
\end{scriptsize}
\end{table}
\begin{table}[!h]
\centering \caption{\label{tab:lfsat2} Estimated parameters and joint log-likelihoods $\ell_{j|t}$ for the copula-based Markov models for ordinal time-series with covariates for satisfaction with income $Y_3$, where a parametric copula family $C_{j|t}$ is used for the joint distribution of subsequent observations. For the best fit, the standard errors (SE) of the estimates, Wald tests ($Z$) and $p$-values are also presented.}
\begin{scriptsize}
\begin{tabular}{lcccccccc} \hline
& BVN & Frank & Gumbel & s.Gumbel & \multicolumn{4}{c}{BVT, $\nu=4$} \\ \cline{2-9}
& Est. & Est. & Est. & Est. & Est. & SE & $Z$ & $p$-value \\ \cline{2-9}
$\alpha_1$ & 1.79 & 1.26 & 1.84 & 1.57 & 1.83 & 0.67 & 2.76 & 0.01 \\
$\alpha_2$ & 2.42 & 1.89 & 2.49 & 2.13 & 2.45 & 0.67 & 3.67 & 0.00 \\
$\alpha_3$ & 3.06 & 2.51 & 3.15 & 2.73 & 3.09 & 0.67 & 4.62 & 0.00 \\
$\alpha_4$ & 4.04 & 3.46 & 4.11 & 3.71 & 4.07 & 0.67 & 6.07 & 0.00 \\
$\alpha_5$ & 5.18 & 4.60 & 5.14 & 4.90 & 5.20 & 0.67 & 7.71 & 0.00 \\
Age & -0.39 & -0.52 & -0.35 & -0.24 & -0.23 & 0.22 & -1.06 & 0.29 \\
Age$^2$ & 0.04 & 0.06 & 0.03 & 0.03 & 0.02 & 0.02 & 0.86 & 0.39 \\
Sex & -0.01 & -0.02 & -0.01 & -0.03 & -0.02 & 0.05 & -0.45 & 0.66 \\
Household size & -1.44 & -1.30 & -1.36 & -1.23 & -1.27 & 0.30 & -4.19 & 0.00 \\
\# of kids & 0.85 & 0.89 & 0.73 & 0.70 & 0.68 & 0.32 & 2.13 & 0.03 \\
Education & & & & & & & & \\
\hspace{2ex}No education & -0.19 & -0.17 & -0.12 & -0.24 & -0.20 & 0.10 & -1.95 & 0.05 \\
\hspace{2ex} hnd,hnc, a/o levels, cse & -0.16 & -0.15 & -0.11 & -0.18 & -0.15 & 0.07 & -2.22 & 0.03 \\
Region & & & & & & & & \\
\hspace{2ex}South & 0.08 & 0.06 & 0.18 & -0.04 & 0.07 & 0.12 & 0.56 & 0.57 \\
\hspace{2ex}Midlands & 0.04 & -0.04 & 0.17 & -0.09 & 0.05 & 0.12 & 0.46 & 0.65 \\
\hspace{2ex}North & 0.00 & -0.07 & 0.14 & -0.15 & 0.01 & 0.12 & 0.05 & 0.96 \\
\hspace{2ex}RUK & -0.02 & -0.13 & 0.12 & -0.14 & 0.01 & 0.15 & 0.08 & 0.94 \\
Income & & & & & & & & \\
\hspace{2ex}ln(Labour Income) & 0.46 & 0.43 & 0.45 & 0.40 & 0.43 & 0.05 & 8.73 & 0.00 \\
\hspace{2ex}ln(Pension Income) & 0.02 & 0.02 & 0.02 & 0.02 & 0.02 & 0.01 & 2.31 & 0.02 \\
\hspace{2ex}ln(Benefit Income) & -0.01 & -0.01 & -0.01 & -0.01 & -0.01 & 0.01 & -2.11 & 0.03 \\
\hspace{2ex}ln(Transfer Income) & -0.01 & -0.01 & -0.01 & -0.01 & -0.01 & 0.01 & -1.07 & 0.29 \\
\hspace{2ex}ln(Investment Income) & 0.04 & 0.03 & 0.03 & 0.03 & 0.03 & 0.01 & 5.84 & 0.00 \\
\hspace{2ex}Regional income rank (standardized) & 0.02 & 0.02 & 0.00 & 0.02 & 0.01 & 0.04 & 0.13 & 0.89 \\ \hline
$\tau$ & 0.40 & 0.43 & 0.42 & 0.41 & 0.41 & 0.01 & 33.2 & 0.00 \\
$-\ell_{j|t}$ & 5750.1 & 5756.6 & 5756.3 & 5726.6 & \multicolumn{4}{c}{5698.1} \\ \hline
$z_0$ ($p$-value)&{-}&-0.549 (0.583)& -0.526 (0.599) &2.008 (0.045)& \multicolumn{4}{c}{4.272 (0.000)}\\\hline
\end{tabular}
\end{scriptsize}
\end{table}
\begin{table}[!h]
\centering \caption{\label{tab:lfsat3} Estimated parameters and joint log-likelihoods $\ell_{j|t}$ for the copula-based Markov models for ordinal time-series with covariates for satisfaction with house/flat $Y_4$, where a parametric copula family $C_{j|t}$ is used for the joint distribution of subsequent observations. For the best fit, the standard errors (SE) of the estimates, Wald tests ($Z$) and $p$-values are also presented.}
\begin{scriptsize}
\begin{tabular}{lcccccccc} \hline
& BVN & Frank & Gumbel & s.Gumbel & \multicolumn{4}{c}{BVT, $\nu=5$} \\ \cline{2-9}
& Est. & Est. & Est. & Est. & Est. & SE & Z & $p$-value \\ \cline{2-9}
$\alpha_1$ & -3.19 & -3.36 & -3.22 & -2.71 & -2.97 & 0.47 & -6.36 & 0.00 \\
$\alpha_2$ & -2.65 & -2.79 & -2.65 & -2.27 & -2.46 & 0.47 & -5.27 & 0.00 \\
$\alpha_3$ & -2.11 & -2.25 & -2.09 & -1.80 & -1.94 & 0.47 & -4.15 & 0.00 \\
$\alpha_4$ & -1.21 & -1.36 & -1.17 & -0.94 & -1.04 & 0.47 & -2.22 & 0.03 \\
$\alpha_5$ & 0.06 & -0.10 & 0.05 & 0.35 & 0.24 & 0.47 & 0.51 & 0.61 \\
Age & -0.48 & -0.54 & -0.41 & -0.39 & -0.39 & 0.22 & -1.78 & 0.07 \\
Age$^2$ & 0.06 & 0.07 & 0.05 & 0.06 & 0.05 & 0.02 & 2.11 & 0.04 \\
Sex & 0.02 & 0.01 & 0.00 & 0.03 & 0.01 & 0.05 & 0.24 & 0.81 \\
Household size & -0.70 & -0.72 & -0.75 & -0.47 & -0.57 & 0.29 & -1.95 & 0.05 \\
\# of kids & 0.27 & 0.39 & 0.32 & 0.10 & 0.17 & 0.32 & 0.55 & 0.58 \\
Education & & & & & & & & \\
\hspace{2ex}No education & 0.36 & 0.33 & 0.37 & 0.26 & 0.32 & 0.10 & 3.25 & 0.00 \\
\hspace{2ex} hnd,hnc, a/o levels, cse & 0.03 & 0.04 & 0.01 & 0.03 & 0.01 & 0.06 & 0.16 & 0.87 \\
Region & & & & & & & & \\
\hspace{2ex}South & -0.03 & 0.00 & -0.01 & -0.07 & -0.06 & 0.11 & -0.53 & 0.60 \\
\hspace{2ex}Midlands & -0.05 & 0.01 & 0.00 & -0.11 & -0.09 & 0.12 & -0.73 & 0.46 \\
\hspace{2ex}North & -0.03 & 0.03 & -0.01 & -0.12 & -0.09 & 0.11 & -0.82 & 0.41 \\
\hspace{2ex}RUK & 0.07 & 0.15 & 0.09 & 0.00 & 0.01 & 0.13 & 0.10 & 0.92 \\
House Type & & & & & & & & \\
\hspace{2ex}Semi-detached & -0.27 & -0.29 & -0.29 & -0.21 & -0.25 & 0.05 & -4.75 & 0.00 \\
\hspace{2ex}Terraced & -0.49 & -0.50 & -0.50 & -0.42 & -0.47 & 0.07 & -7.21 & 0.00 \\
\hspace{2ex}Other & -1.15 & -1.26 & -1.07 & -1.05 & -1.06 & 0.15 & -7.10 & 0.00 \\
House value & & & & & & & & \\
\hspace{2ex}50K-100K & 0.16 & 0.12 & 0.13 & 0.15 & 0.14 & 0.06 & 2.34 & 0.02 \\
\hspace{2ex}100K-175K & 0.25 & 0.18 & 0.21 & 0.20 & 0.19 & 0.07 & 2.87 & 0.00 \\
\hspace{2ex}175K-250K & 0.31 & 0.24 & 0.26 & 0.25 & 0.25 & 0.08 & 3.24 & 0.00 \\
\hspace{2ex}$>$250K & 0.42 & 0.40 & 0.37 & 0.32 & 0.32 & 0.09 & 3.65 & 0.00 \\ \hline
$\tau$ & 0.35 & 0.38 & 0.37 & 0.36 & 0.37 & 0.01 & 29.28 & 0.00 \\
$-\ell_{j|t}$ & 5407.6 & 5394.8 & 5396.3 & 5411.4 & \multicolumn{4}{c}{5373.0} \\ \hline
$z_0$ ($p$-value) &-& 1.328 (0.184) &1.049 (0.294)& -0.390 (0.696) & \multicolumn{4}{c}{3.139 (0.002)}\\\hline
\end{tabular}
\end{scriptsize}
\end{table}
\begin{table}[!h]
\centering \caption{\label{tab:lfsat4} Estimated parameters and joint log-likelihoods $\ell_{j|t}$ for the copula-based Markov models for ordinal time-series with covariates for satisfaction with spouse/partner $Y_5$, where a parametric copula family $C_{j|t}$ is used for the joint distribution of subsequent observations. For the best fit, the standard errors (SE) of the estimates, Wald tests ($Z$) and $p$-values are also presented.}
\begin{scriptsize}
\begin{tabular}{lcccccccc} \hline
& BVN & Frank & \multicolumn{4}{c}{Gumbel} & s.Gumbel & BVT, $\nu=3$ \\ \cline{2-9}
& Est. & Est. & Est. & SE & Z & $p$-value & Est. & Est. \\ \cline{2-9}
$\alpha_1$ & -3.75 & -4.16 & -4.09 & 0.72 & -5.65 & 0.00 & -3.19 & -3.63 \\
$\alpha_2$ & -3.29 & -3.62 & -3.55 & 0.72 & -4.92 & 0.00 & -2.83 & -3.20 \\
$\alpha_3$ & -2.90 & -3.18 & -3.12 & 0.72 & -4.33 & 0.00 & -2.53 & -2.85 \\
$\alpha_4$ & -2.34 & -2.59 & -2.52 & 0.72 & -3.50 & 0.00 & -2.05 & -2.33 \\
$\alpha_5$ & -1.47 & -1.73 & -1.64 & 0.72 & -2.28 & 0.02 & -1.20 & -1.46 \\
Age & -0.39 & -0.28 & -0.32 & 0.36 & -0.89 & 0.37 & -0.53 & -0.50 \\
Age$^2$ & 0.05 & 0.03 & 0.04 & 0.04 & 1.00 & 0.32 & 0.07 & 0.06 \\
Sex & -0.02 & 0.13 & 0.09 & 0.49 & 0.19 & 0.85 & -0.18 & -0.11 \\
Household size & -0.98 & -1.01 & -1.08 & 0.31 & -3.51 & 0.00 & -0.73 & -0.88 \\
\# of kids & -0.39 & -0.43 & -0.41 & 0.33 & -1.24 & 0.22 & -0.44 & -0.52 \\
Education & & & & & & & & \\
\hspace{2ex}No education & 0.48 & 0.52 & 0.56 & 0.13 & 4.28 & 0.00 & 0.30 & 0.45 \\
\hspace{2ex} hnd,hnc, a/o levels, cse & 0.06 & 0.09 & 0.13 & 0.09 & 1.54 & 0.12 & 0.01 & 0.08 \\
Region & & & & & & & & \\
\hspace{2ex}South & 0.18 & 0.28 & 0.18 & 0.12 & 1.49 & 0.14 & 0.06 & 0.06 \\
\hspace{2ex}Midlands & 0.13 & 0.21 & 0.11 & 0.13 & 0.85 & 0.39 & 0.04 & 0.02 \\
\hspace{2ex}North & 0.24 & 0.27 & 0.15 & 0.12 & 1.23 & 0.22 & 0.18 & 0.12 \\
\hspace{2ex}RUK & 0.31 & 0.33 & 0.27 & 0.14 & 1.90 & 0.06 & 0.21 & 0.20 \\
Spouse Characteristics & & & & & & & & \\
\hspace{2ex} Age & -0.06 & -0.34 & -0.21 & 0.33 & -0.62 & 0.53 & 0.27 & 0.14 \\
\hspace{2ex} Age$^2$ & -0.01 & 0.02 & 0.00 & 0.03 & 0.12 & 0.91 & -0.05 & -0.03 \\
\hspace{2ex} Sex & -0.25 & -0.10 & -0.13 & 0.48 & -0.27 & 0.78 & -0.41 & -0.35 \\
\hspace{2ex}Education& & & & & & & & \\
\hspace{4ex}No education & 0.16 & 0.09 & 0.07 & 0.12 & 0.57 & 0.57 & 0.15 & 0.10 \\
\hspace{4ex}hnd,hnc, a/o levels, cse & -0.09 & -0.12 & -0.13 & 0.09 & -1.47 & 0.14 & -0.08 & -0.11 \\
\hspace{2ex}ln(Labour Income) & 0.01 & 0.01 & 0.01 & 0.01 & 1.49 & 0.14 & 0.01 & 0.01 \\
\hspace{2ex}\# Health problems & -0.14 & -0.23 & -0.18 & 0.17 & -1.02 & 0.31 & -0.08 & -0.10 \\
\hspace{2ex}Hours of wk housework & 0.02 & 0.02 & 0.02 & 0.02 & 1.00 & 0.32 & 0.01 & 0.02 \\ \hline
$\tau$ & 0.51 & 0.53 & 0.57 & 0.01 & 48.6 & 0.00 & 0.49 & 0.53 \\
$-\ell_{j|t}$ & 4007.0 & 3984.7 & \multicolumn{4}{c}{3967.3} & 4011.8 & 3968.2 \\ \hline
$z_0$ ($p$-value) &-& 1.558 (0.119) & \multicolumn{4}{c}{4.789 ($<0.001$)}& -0.514 (0.607)& 2.992 (0.003)\\\hline
\end{tabular}
\end{scriptsize}
\end{table}
\begin{table}[!h]
\centering \caption{\label{tab:lfsat5} Estimated parameters and joint log-likelihoods $\ell_{j|t}$ for the copula-based Markov models for ordinal time-series with covariates for satisfaction with job $Y_6$, where a parametric copula family $C_{j|t}$ is used for the joint distribution of subsequent observations. For the best fit, the standard errors (SE) of the estimates, Wald tests ($Z$) and $p$-values are also presented.}
\begin{scriptsize}
\begin{tabular}{lcccccccc} \hline
& BVN & Frank & Gumbel & s.Gumbel & \multicolumn{4}{c}{BVT, $\nu=8$} \\ \cline{2-9}
& Est. & Est. & Est. & Est. & Est. & SE & Z & $p$-value \\ \cline{2-9}
$\alpha_1$ & -0.29 & -0.39 & -0.43 & -0.35 & -0.47 & 0.43 & -1.10 & 0.27 \\
$\alpha_2$ & 0.40 & 0.31 & 0.27 & 0.33 & 0.22 & 0.43 & 0.52 & 0.60 \\
$\alpha_3$ & 1.26 & 1.17 & 1.12 & 1.17 & 1.07 & 0.43 & 2.52 & 0.01 \\
$\alpha_4$ & 2.46 & 2.37 & 2.30 & 2.35 & 2.26 & 0.43 & 5.28 & 0.00 \\
$\alpha_5$ & 3.88 & 3.78 & 3.72 & 3.77 & 3.69 & 0.43 & 8.61 & 0.00 \\
Age & -0.10 & -0.13 & -0.12 & -0.17 & -0.17 & 0.20 & -0.87 & 0.39 \\
Age$^2$ & 0.02 & 0.02 & 0.02 & 0.02 & 0.02 & 0.02 & 1.09 & 0.28 \\
Sex & -0.03 & -0.05 & -0.03 & -0.03 & -0.03 & 0.04 & -0.73 & 0.46 \\
Household size & -0.58 & -0.55 & -0.67 & -0.46 & -0.55 & 0.30 & -1.85 & 0.06 \\
\# of kids & 0.41 & 0.36 & 0.49 & 0.33 & 0.40 & 0.32 & 1.26 & 0.21 \\
Education & & & & & & & & \\
\hspace{2ex}No education & -0.08 & -0.09 & -0.14 & -0.02 & -0.08 & 0.08 & -1.01 & 0.31 \\
\hspace{2ex} hnd,hnc, a/o levels, cse & -0.01 & -0.02 & -0.02 & 0.02 & 0.00 & 0.05 & 0.05 & 0.96 \\
Region & & & & & & & & \\
\hspace{2ex}South & -0.19 & -0.19 & -0.17 & -0.22 & -0.20 & 0.09 & -2.12 & 0.03 \\
\hspace{2ex}Midlands & -0.16 & -0.16 & -0.12 & -0.17 & -0.15 & 0.10 & -1.54 & 0.12 \\
\hspace{2ex}North & -0.26 & -0.26 & -0.24 & -0.27 & -0.27 & 0.09 & -2.87 & 0.00 \\
\hspace{2ex}RUK & -0.29 & -0.29 & -0.25 & -0.31 & -0.29 & 0.11 & -2.67 & 0.01 \\
Satisfaction with job pay & & & & & & & & \\
\hspace{2ex} Medium & 0.38 & 0.38 & 0.39 & 0.39 & 0.39 & 0.05 & 7.97 & 0.00 \\
\hspace{2ex} Very & 0.74 & 0.73 & 0.74 & 0.74 & 0.74 & 0.05 & 14.3 & 0.00 \\
Satisfaction with job security & & & & & & & & \\
\hspace{2ex} Medium & 0.41 & 0.43 & 0.40 & 0.40 & 0.41 & 0.05 & 7.43 & 0.00 \\
\hspace{2ex} Very & 0.64 & 0.65 & 0.61 & 0.63 & 0.62 & 0.05 & 11.3 & 0.00 \\
Satisfaction with work itself & & & & & & & & \\
\hspace{2ex} Medium & 0.88 & 0.87 & 0.84 & 0.89 & 0.86 & 0.06 & 13.3 & 0.00 \\
\hspace{2ex} Very & 1.62 & 1.61 & 1.56 & 1.62 & 1.57 & 0.07 & 22.8 & 0.00 \\
Satisfaction with hours worked & & & & & & & & \\
\hspace{2ex} Medium & 0.18 & 0.17 & 0.16 & 0.18 & 0.17 & 0.05 & 3.41 & 0.00 \\
\hspace{2ex} Very & 0.46 & 0.44 & 0.42 & 0.46 & 0.43 & 0.05 & 8.14 & 0.00 \\ \hline
$\tau$ & 0.20 & 0.22 & 0.20 & 0.18 & 0.21 & 0.01 & 15.2 & 0.00 \\
$-\ell_{j|t}$ & 5245.5 & 5241.5 & 5234.8 & 5261.7 & \multicolumn{4}{c}{5228.9} \\ \hline
$z_0$ ($p$-value) &-&0.705 (0.481)& 1.336 (0.181) &-2.261 (0.024) &\multicolumn{4}{c}{ 2.454 (0.014)}\\\hline
\end{tabular}
\end{scriptsize}
\end{table}
\begin{table}[!h]
\centering \caption{\label{tab:lfsat8} Estimated parameters and joint log-likelihoods $\ell_{j|t}$ for the copula-based Markov models for ordinal time-series with covariates for satisfaction with use of leisure time $Y_7$, where a parametric copula family $C_{j|t}$ is used for the joint distribution of subsequent observations. For the best fit, the standard errors (SE) of the estimates, Wald tests ($Z$) and $p$-values are also presented.}
\begin{scriptsize}
\begin{tabular}{lcccccccc} \hline
& BVN & Frank & Gumbel & s.Gumbel & \multicolumn{4}{c}{BVT, $\nu=4$} \\ \cline{2-9}
& Est. & Est. & Est. & Est. & Est. & SE & Z & $p$-value \\ \cline{2-9}
$\alpha_1$ & -2.67 & -2.77 & -2.68 & -2.33 & -2.43 & 0.46 & -5.33 & 0.00 \\
$\alpha_2$ & -2.05 & -2.12 & -2.03 & -1.78 & -1.82 & 0.46 & -3.99 & 0.00 \\
$\alpha_3$ & -1.37 & -1.46 & -1.35 & -1.14 & -1.15 & 0.46 & -2.52 & 0.01 \\
$\alpha_4$ & -0.51 & -0.63 & -0.51 & -0.28 & -0.29 & 0.46 & -0.63 & 0.53 \\
$\alpha_5$ & 0.47 & 0.36 & 0.39 & 0.75 & 0.69 & 0.46 & 1.52 & 0.13 \\
Age & -0.58 & -0.62 & -0.49 & -0.54 & -0.47 & 0.21 & -2.23 & 0.03 \\
Age$^2$ & 0.07 & 0.08 & 0.06 & 0.07 & 0.06 & 0.02 & 2.49 & 0.01 \\
Sex & 0.10 & 0.07 & 0.05 & 0.13 & 0.10 & 0.05 & 1.85 & 0.06 \\
Household size & -0.85 & -0.88 & -0.92 & -0.71 & -0.80 & 0.28 & -2.84 & 0.00 \\
\# of kids & -0.43 & -0.45 & -0.33 & -0.50 & -0.47 & 0.30 & -1.55 & 0.12 \\
Education & & & & & & & & \\
\hspace{2ex}No education & -0.02 & 0.00 & -0.07 & 0.03 & -0.03 & 0.09 & -0.31 & 0.75 \\
\hspace{2ex} hnd,hnc, a/o levels, cse & 0.04 & 0.05 & 0.00 & 0.11 & 0.06 & 0.06 & 0.96 & 0.34 \\
Region & & & & & & & & \\
\hspace{2ex}South & 0.08 & 0.04 & 0.07 & 0.11 & 0.12 & 0.11 & 1.11 & 0.27 \\
\hspace{2ex}Midlands & 0.06 & 0.04 & 0.05 & 0.07 & 0.08 & 0.11 & 0.69 & 0.49 \\
\hspace{2ex}North & -0.01 & -0.07 & 0.02 & -0.01 & 0.04 & 0.11 & 0.35 & 0.72 \\
\hspace{2ex}RUK & 0.09 & 0.02 & 0.10 & 0.09 & 0.14 & 0.12 & 1.14 & 0.25 \\
Frequency of leisure activities & & & & & & & & \\
\hspace{2ex}walk/swim/play sport & 0.12 & 0.11 & 0.09 & 0.11 & 0.09 & 0.03 & 2.67 & 0.01 \\
\hspace{2ex}watch live sport & 0.14 & 0.16 & 0.12 & 0.14 & 0.12 & 0.04 & 3.30 & 0.00 \\
\hspace{2ex}cinema & -0.04 & -0.03 & -0.03 & -0.03 & -0.03 & 0.03 & -0.93 & 0.35 \\
\hspace{2ex}theatre/concert & 0.04 & 0.03 & 0.03 & 0.04 & 0.03 & 0.03 & 1.07 & 0.28 \\
\hspace{2ex}out for a drink & 0.14 & 0.14 & 0.10 & 0.13 & 0.11 & 0.04 & 2.43 & 0.02 \\
\hspace{2ex}work in garden & -0.04 & -0.04 & -0.03 & -0.03 & -0.02 & 0.05 & -0.46 & 0.64 \\
\hspace{2ex}diy, car maintenance & -0.03 & -0.01 & -0.02 & -0.02 & -0.02 & 0.04 & -0.40 & 0.69 \\
\hspace{2ex}attend evening classes & 0.16 & 0.15 & 0.14 & 0.14 & 0.14 & 0.03 & 4.47 & 0.00 \\
\hspace{2ex}attend local groups & 0.06 & 0.06 & 0.06 & 0.04 & 0.05 & 0.04 & 1.40 & 0.16 \\ \hline
$\tau$ & 0.36 & 0.39 & 0.39 & 0.38 & 0.38 & 0.01 & 31.52 & 0.00 \\
$-\ell_{j|t}$ & 6141.7 & 6124.7 & 6137.3 & 6113.1 & \multicolumn{4}{c}{6077.5} \\ \hline
$z_0$ ($p$-value) &1.462 (0.144)& 0.384 (0.701)& 2.189 (0.029)& \multicolumn{4}{c}{3.999 ($<0.001$)}\\ \hline
\end{tabular}
\end{scriptsize}
\end{table}
Tables \ref{tab:lfsato} - \ref{tab:lfsat8} give the estimated parameters and joint log-likelihoods $\ell_{j|t}$ for the copula-based Markov models for ordinal time-series with covariates for the seven satisfaction equations, where a parametric copula family is used for the joint distribution of subsequent observations. For the best fit, according to the likelihood principle, we also calculate standard errors (SE) and corresponding Wald tests and $p$-values.
SEs of estimates have been obtained via the gradients and the Hessian computed numerically during the maximization process. Assuming that the usual regularity conditions \citep{serfling80} for asymptotic maximum likelihood theory hold for the bivariate model as well as for its margins we have that the estimates are asymptotically normal. Therefore we also build Wald tests to statistically judge the effect of any covariate.
\cite{liang&zeger86} noted even though parameter estimates from univariate analysis ignoring the association remain consistent, they are inefficient. When using copula terms in the likelihood improves asymptotic efficiency over the independence estimating equations.
\cite{Prokhorov&Schmidt2009} acknowledged that the efficiency gains will come at the expense of an asymptotic bias if the joint distribution is misspecified.
However, our results show that the effect of misspecifying the copula choice can be seen as minimal for both the univariate parameters and Kendall’s tau, since (a) the univariate parameters are a univariate inference, and hence, it is the univariate marginal distribution that matters and not the type of the copula, and (b) Kendall’s tau only accounts for the dependence dominated by the middle of the data, and it is expected to be similar amongst different families of copulas. In essence, given that tail dependence varies, the effect of different tail behaviours is reflected in predictive inferences that depend on the joint distribution, e.g., the Vuong's statistic.
The best fit for the ordinal time-series, as expected, is based on BVT copulas with a small $\nu$ (according to the likelihood principle), where there is a big and statistical significant improvement over the autoregressive-to-anything (BVN copula-based Markov) model according to Vuong's statistics. This result suggests skewness to both upper and lower tail for subsequent (in time) observations, i.e. more probability in both joint tails of the various univariate time-series. In particular, the BVT copula, which is a radially symmetric, provides the best fit for the 6 out 7 univariate time-series. Hence, according to
\cite{Prokhorov&Schmidt2009} our models are robust to the estimation of the regression parameters. Some interpretation of the estimated regression coefficients for each ordinal times-series is provided below.
In Table \ref{tab:lfsato}, generic satisfaction results suggest a drop (at an increasing rate) with age while gender and education do not seem to matter. Smaller households are happier and with the exception of Midlands everywhere is happier than London. Improved health (i.e. fewer health problems) and higher satisfaction with job characteristics significantly improve life satisfaction. However, absolute or relative income, house type, choice of partner or frequency of leisure activities have little direct effect on life satisfaction over and above what they might have through their respective domain satisfactions.
As expected, in Table \ref{tab:lfsat1} satisfaction with health goes down (at an increasing rate) with age and number of health problems, which has a large and highly significant effect. No gender or education differences are observed, while South of England and non-English regions report higher values than London.
In Table \ref{tab:lfsat2}, satisfaction with income is significantly lower for the less educated and those in larger households but exhibits no age or gender variation. Looking at domain specific variables, with the exception of transfer and benefit income, all absolute income variables improve income well-being with the labour component exhibiting the strongest effect. Transfer income plays little role, whereas benefit income has a negative sign suggesting that those with higher benefit income are less satisfied with their financial well-being. Finally, relative position of the individual within their region (based on total annual income) bears no influence in determining domain income satisfaction.
In Table \ref{tab:lfsat3}, house satisfaction decreases (at an increasing rate) with age and size of the household, while it increases with education and house value and does not seem to change with region. Compared to a detached house all other house types result in lower satisfaction level. Similarly, compared to a very low house value, all value increments imply improvement in house domain satisfaction.
Satisfaction with spouse is the only exception in our data, with the Gumbel copula marginally providing the best fit and suggesting more probability in the upper joint tail. Overall age, gender and number of kids do not influence satisfaction with partner, whereas those less or not educated tend to be happier with the partners, as do those living in the Midlands and rest of UK compared to London. Looking at partners, little variation in domain satisfaction is explained by partners' individual characteristics (Table \ref{tab:lfsat4}).
In Table \ref{tab:lfsat5}, none of the demographics characteristics nor education are important in explaining domain satisfaction with one's job. However, all regions report lower satisfaction compared to London, while improved satisfaction with job pay, security, work hours and work itself all significantly and strongly improve overall job satisfaction.
In Table \ref{tab:lfsat8}, satisfaction with use of leisure time goes down with age (at an increasing rate) and being lower for females and those in larger households, whereas education, more kids or geographical location have no effect. Increasing frequency of various leisure activities, on the other hand, indeed improve domain satisfaction with sports (playing or watching), out for drinks and evening classes begin the most significant.
\subsection{Fitted joint copula-based Markov models}
A joint copula-based Markov model joins the various satisfaction ordinal time-series, where the best fit bivariate copula family from the preceding subsection is used for the joint distribution of subsequent observations for each ordinal time-series. Since the sample is a mixture of populations (e.g., unobserved individual traits/characteristics) the MVT copula would be in theory a potential model to join the univariate time-series.
Once again, a number of different copulas are tried to form the joint distribution, i.e. the MVN and MVT with $\nu=\{5,10,15\}$.
Given that all together we have $197$ $\bigl($Step 1(c)$\bigr)$ $+21$ (Step 2) parameters to be estimated, makes the third step of the estimation approach in Section \ref{estimation} infeasible. Hence the estimated latent correlations, their SEs, and joint log-likelihoods $\ell_{1\ldots d|t}$ from these joint models in Table \ref{tab:corrs} are derived using the second step of the estimation procedure. The SEs of the estimated latent correlations are obtained by the inversion of the Hessian matrix. These SEs are adequate to assess the flatness of the log-likelihood. Proper SEs that account for the estimation of all parameters can be obtained by jackknifing the two-stage estimation procedure.
\begin{table}[!h]
\centering
\caption{\label{tab:corrs}Estimated latent correlations, their SEs, and joint log-likelihoods $\ell_{1\ldots d|t}$ for the joint copula-based Markov models for ordinal time-series with covariates, where the best fit bivariate copula family is used for the joint distribution of subsequent observations for each ordinal time-series, and, an MVT parametric copula family is used for the joint distribution of joint observations.}
\begin{tabular}{ccccccccc} \hline
& \multicolumn{2}{c}{MVN} & \multicolumn{2}{c}{MVT, $\nu=5$} & \multicolumn{2}{c}{MVT, $\nu=10$} & \multicolumn{2}{c}{MVT, $\nu=15$} \\ \cline{2-9}
&Est. & SE&Est. & SE&Est. & SE&Est. & SE \\ \cline{2-9}
$\rho_{12}$ & 0.381 & 0.015 & 0.365 & 0.017 & 0.379 & 0.016 & 0.383 & 0.016 \\
$\rho_{13}$ & 0.300 & 0.016 & 0.297 & 0.017 & 0.308 & 0.017 & 0.310 & 0.017 \\
$\rho_{14}$ & 0.328 & 0.016 & 0.323 & 0.018 & 0.335 & 0.017 & 0.337 & 0.017 \\
$\rho_{15}$ & 0.461 & 0.016 & 0.440 & 0.018 & 0.456 & 0.017 & 0.461 & 0.017 \\
$\rho_{16}$ & 0.357 & 0.015 & 0.361 & 0.017 & 0.369 & 0.016 & 0.369 & 0.017 \\
$\rho_{17}$ & 0.442 & 0.014 & 0.441 & 0.015 & 0.451 & 0.015 & 0.452 & 0.014 \\
$\rho_{23}$ & 0.295 & 0.016 & 0.290 & 0.017 & 0.300 & 0.017 & 0.301 & 0.016 \\
$\rho_{24}$ & 0.222 & 0.016 & 0.217 & 0.018 & 0.229 & 0.017 & 0.230 & 0.017 \\
$\rho_{25}$ & 0.191 & 0.018 & 0.173 & 0.019 & 0.188 & 0.019 & 0.192 & 0.019 \\
$\rho_{26}$ & 0.215 & 0.016 & 0.212 & 0.018 & 0.221 & 0.018 & 0.223 & 0.017 \\
$\rho_{27}$ & 0.284 & 0.015 & 0.274 & 0.017 & 0.286 & 0.017 & 0.289 & 0.016 \\
$\rho_{34}$ & 0.338 & 0.015 & 0.344 & 0.016 & 0.351 & 0.016 & 0.350 & 0.016 \\
$\rho_{35}$ & 0.184 & 0.018 & 0.173 & 0.020 & 0.186 & 0.019 & 0.189 & 0.020 \\
$\rho_{36}$ & 0.262 & 0.016 & 0.256 & 0.017 & 0.268 & 0.017 & 0.270 & 0.017 \\
$\rho_{37}$ & 0.220 & 0.016 & 0.218 & 0.018 & 0.228 & 0.017 & 0.229 & 0.017 \\
$\rho_{45}$ & 0.271 & 0.018 & 0.262 & 0.019 & 0.274 & 0.019 & 0.277 & 0.019 \\
$\rho_{46}$ & 0.172 & 0.017 & 0.178 & 0.019 & 0.185 & 0.017 & 0.185 & 0.019 \\
$\rho_{47}$ & 0.262 & 0.016 & 0.253 & 0.017 & 0.265 & 0.017 & 0.268 & 0.017 \\
$\rho_{56}$ & 0.231 & 0.019 & 0.223 & 0.019 & 0.233 & 0.020 & 0.235 & 0.019 \\
$\rho_{57}$ & 0.275 & 0.018 & 0.263 & 0.018 & 0.276 & 0.019 & 0.279 & 0.019 \\
$\rho_{67}$ & 0.188 & 0.016 & 0.190 & 0.018 & 0.198 & 0.017 & 0.199 & 0.019 \\ \hline
$-\ell_{1\ldots d|t}$ & \multicolumn{2}{c}{34950.99} & \multicolumn{2}{c}{34841.91} & \multicolumn{2}{c}{34747.89} & \multicolumn{2}{c}{34766.48} \\ \hline
Vuong's & $z_0$ & $p$-value & $z_0$ & $p$-value & $z_0$ & $p$-value & $z_0$ & $p$-value\\
test & \multicolumn{2}{c}{-}&2.606 & 0.009&7.634 & $<0.001$ & 9.125 & $<0.001$\\
\hline
\end{tabular}
\begin{flushleft}
\begin{scriptsize}
Subscript 1 denotes satisfaction with life overall; 2 satisfaction with health; 3 satisfaction with income; 4 satisfaction with house/flat; 5 satisfaction with spouse/partner; 6 satisfaction with job; 7 satisfaction with use of leisure time.
\end{scriptsize}
\end{flushleft}
\end{table}
According to the likelihood principle the best fit is based on an MVT copula with 10 degrees of freedom.
In this example, it is highlighted that a joint model with an MVT copula is plausible for a population that is a mixture of subpopulations, while a MVN model might be adequate for smaller homogeneous subgroups. This is confirmed by the Vuong's statistic of 7.634 (p-value $<$ 0.001) reported in the final row of Table \ref{tab:corrs}, which establishes clear superiority of the MVT over the MVN.
The fact that the best-fitting copula for the joint model is the MVT with $\nu=10$ (instead of BVN) suggests positive tail dependence in the data, i.e. individuals reporting high satisfaction tend to do so across multiple domains, while correspondingly for those reporting low satisfaction.
Furthermore, a joint copula-based Markov model leads to better inferences than a copula-based Markov model with independence among the different satisfactions since the likelihood has been improved by
$2012.0=-34747.9-(-4689.2-5725.9-5698.1-5373.0-3967.3-5228.9-6077.5)$.
Overall, all latent correlations are positive and highly statistically significant providing evidence for the multivariate nature of well-being, while suggesting that increases in life satisfaction in one domain result in further increase in satisfaction in another and the presence of potential multiplying ripple effects. The strongest latent correlations are realised between generic satisfaction and various domain satisfaction responses, i.e. Spouse, Leisure, Health and Job domain satisfaction appearing to have the strongest links to overall well-being, followed by House and lastly Income domain satisfaction. Moving on, cross-domain associations are slightly mitigated but still various pairs emerge as important. Between domain satisfactions of income-house, health-income and health-leisure exhibit the strongest links, followed by spouse-leisure, income-job and house-leisure.
\section{\label{sec-discussion}Discussion}
In this paper, we develop a comprehensive conceptual model of life satisfaction and its constituents where a number of direct and indirect links between objective covariates and domain and generic components of well-being are captured. Modelling dependence allows revisiting previously estimated relationships in univariate frameworks and testing their association in a structural setting. In order to apply such structural framework, a joint copula-based Markov econometric model for ordinal time-series with covariates is developed, where each ordinal time-series is considered a copula-based Markov model with a parametric bivariate copula for the joint distribution of subsequent observations and whose conditionals are subsequently joined through an MVT copula. We have implemented a simulated likelihood method, where the rectangles are converted to bounded integrands via the error reduction methods in \cite{genz&bretz02}, and hence the statistical efficiency of simulated likelihood is as good as maximum likelihood.
Comparing with past literature we replicate the U-shaped effect for age \citep{blanchflower_is_2008} and strong health effects \citep{dolan_we_2008}. Yet, our results fail to confirm many of the findings of past studies. For income, which has long been one of the main variables of interest driving the well-being literature, no indicator appears statistically significant for overall well-being \citep{clark_satisfaction_1996, clark_relative_2008}. Similarly for individual characteristics such as gender \citep{Alesina2004} or education \citep{blanchflower_well-being_2004}. With the exception of health and job related characteristics that appear significant in both generic and domain equations, most covariates that have been argued to directly influence overall satisfaction (i.e. housing, spouse, exercise, social life and leisure activities etc) fail to achieve significance \citep{powdthavee_putting_2008, powdthavee_i_2009, mentzakis_allowing_2011}.
This lack of direct relationships between covariates and overall well-being is suggestive of alternative underlying mechanisms that influence life satisfaction. In relation to income, past literature has put forward the importance of one's relative position and individual perception in evaluating their well-being \citep{stutzer_role_2004, clark_relative_2008, mentzakis_poor_2009}. Given the nature of domain satisfaction questions (i.e. themselves a relative measure of the well-being an individual perceives themselves as possessing in this aspect of their lives), it is mostly through them that any influence is exerted on generic satisfaction. The strong significant latent correlations confirm the presence of such links across the spectrum of domain satisfactions. The fact that job characteristics indicators (themselves satisfaction questions capturing relative position and perception) are significant in the overall well-being marginal model provides further support for this proposition.
At the extreme, our results would suggest, that covariates are less likely to directly impact generic well-being but can do so through changing individuals' satisfaction with domain well-being, which would be akin to changing individuals' relative position or perception of their status. This observation follows the rationale behind Easterlin's paradox, where once a basic level of need has been met, aspirations increase along with income with the relative position being the main aspect that continues to affect well-being \citep{easterlin_will_1995}. Such mechanism would also explain why peer and network effects have been shown to have strong effects in the literature \citep{kahneman_would_2006, clark_happy_2011}.
Looking at the literature on the structure of well-being \citep{argyle_causes_1999, van_praag_anatomy_2003}, our results empirically validate past findings with aspects of family life, social life, love life, occupational life and leisure coming up as important \citep{salvatore_appraisal_2001}, while the significance of both direct and indirect effects in the case of health lends credence to the conceptual framework and the mixture of patterns in the well-being structure \citep{brief_integrating_1993}.
Specifically, positive tail dependence suggests positive latent correlations of reported satisfaction levels over time with past high satisfaction more likely to spill-over to future periods. In other words, individuals on high satisfaction trends are more likely to continue reporting high well-being (correspondingly for those reporting low satisfaction). Positive temporal association could point to resilience or adaptation in individual happiness parallel to the``set point'' theory of happiness, in which individuals are believed to have a set happiness level that they rerun to over time after positive and negative events \citep{graham_happiness_2008, bradford_getting_2010}. Parallelly, positive temporal association could indicate that individuals of consistently high or low happiness are less likely to experience life events that will move them to the opposite happiness spectrum potentially pointing to the role of habit formation \citep{easterlin_will_1995}.
However, in the presence of shocks in domain satisfaction, overall happiness is prone to follow, something also posited within set-point theory \citep{graham_happiness_2008}. Positive tail dependence across well-being dimensions implies a drop in domain satisfaction is accompanied by corresponding changes in overall and other-domain satisfaction although the latter effects dissipate for certain domain pairs. Looking at this from a slightly different angle, overall happiness requires happiness in all aspects of life or, alternatively, full happiness cannot be achieved without meeting a basic satisfaction level in important domains of one's life \citep{graham_happiness_2008}. The signifiant effect of health in overall well-being would suggest a complementary story to that of the latent correlations, where the former provide some type of necessary conditions for happiness that can only be realized when the latter hold. In other words, having objectively good health improves the chances of satisfaction with life overall or maybe set the foundations for a happy life \citep{frijters_money_2004}, which however can only be enjoyed through certain personal and environmental conditions. These latter conditions that emerge from the latent correlation patterns could be usefully stylized into common clich{\'e}s, such as happiness in personal life (i.e. spousal relations) is the key to true happiness \citep{powdthavee_i_2009}, or that life is all about having fun (i.e. enjoy leisure time), or finally that income does not buy happiness, i.e., limited effect of financial satisfaction \citep{mentzakis_poor_2009}. Nevertheless, despite similar messages with past literature our results reveal substantially different underlying mechanisms that rely on the multivariate dependence component of our model that has been mostly ignored to date.
Finally, an interpretation that brings close the underlying mechanisms discussed so far in relation to the positive temporal and cross-domain dependencies would be the effect individual behavioural traits have on the evaluation of well-being. The classification of individuals according to optimistic (i.e. positively correlated high satisfaction) and pessimistic (i.e. positively correlated low satisfaction) mental predispositions and its subsequent effect on reporting patterns in subjective satisfaction questions could be the driver behind positive dependencies and offer some economic/psychological intuition behind the best fit of the MVT copula and its mixture of populations interpretation.
In conclusion, theories of generic and domain satisfaction suggest new insights can be obtained through dependence modelling with copulas offering a powerful and flexible tool to accommodate all necessary relationships and dependencies.
\section*{\label{sec-asym}Appendix}
Assume a multivariate ordinal regression in which $d \geq 2$ dependent ordinal random variables $Y_{1}, \ldots, Y_{d}$ are observed together with a vector $\mathbf{x} \in \mathbb{R}^p$ of explanatory variables.
If $C(\cdot;\mathbf{R})$ is the MVT copula (or any other parametric
family of copulas) and $F_j(\cdot\; ;\; \mu,\boldsymbol{\gamma})$,
where $\mu=\mathbf{x}^T\boldsymbol{\beta}$ is a function of $\mathbf{x}$
and the $p$-dimensional regression vector $\boldsymbol{\beta}$, and $\boldsymbol{\gamma}=(\alpha_1,\ldots,\alpha_{K-1})$ is the $q$-dimensional vector of the univariate cutpoints ($q=K-1$),
is a parametric model for the $j$th univariate margin
then
$$C\Bigl(F_1(y_1;\mu_1,\boldsymbol{\gamma}),\ldots,F_d(y_d;\mu_d,\boldsymbol{\gamma});\mathbf{R}\Bigr)$$
is a multivariate parametric model with univariate margins $F_1,\ldots,F_d$.
For data $\mathbf{y}_1,\ldots,\mathbf{y}_n$ and sample size $n$,
the MVT copula model joint log-likelihood is
\begin{equation}\label{MLlik}
\ell(\boldsymbol{\beta},\boldsymbol{\gamma},\mathbf{R})\\
= \sum_{i=1}^{n}
\log{h(y_{i1},\ldots,y_{id};\boldsymbol{\beta},\boldsymbol{\gamma},\mathbf{R})},
\end{equation}
where $h(\cdot;\boldsymbol{\beta},\boldsymbol{\gamma},\mathbf{R})$ is the
joint pmf of the multivariate ordinal response vector $\mathbf{Y}=(Y_{1}, \ldots, Y_{d})$, which
can be computed through the rectangle probability:
\begin{equation}
\label{MVNpmf}
h(\mathbf{y};\boldsymbol{\beta},\boldsymbol{\gamma},\mathbf{R})=\int_{\mathcal{T}^{-1}[F_{1}(y_1-1;\mu_1,\boldsymbol{\gamma})]}^{\mathcal{T}^{-1}[F_{1}(y_1;\mu_1,\boldsymbol{\gamma})]}\cdots
\int_{\mathcal{T}^{-1}[F_{d}(y_d-1;\mu_d,\boldsymbol{\gamma})]}^{\mathcal{T}^{-1}[F_{d}(y_d;\mu_d,\boldsymbol{\gamma})]} t_d(z_1,\ldots,z_d;\mathbf{R}) dz_1\ldots dz_d.
\end{equation}
In the following, we are studying the asymptotic properties of the proposed simulated likelihood for the limit (as the number of clusters increases to infinity) of the maximum simulated likelihood estimate (MSLE). We restrict ourselves to a MVN copula model with a positive exchangeable structure, that is we took $\mathbf{R}$ as $(1-\rho)\mathbf{I}_d+ \rho \mathbf{J}_d$,
where $\mathbf{I}_d$ is the identity matrix of order $d$ and $\mathbf{J}_d$ is the $d\times d$ matrix of 1s. In this special case, $d$-dimensional integrals collapse to 1-dimensional integrals \citep[p. 48]{Johnson&Kotz72} resulting in fast and accurate MVN rectangle probabilities. Using the 1-dimensional integral method \citep{Johnson&Kotz72} to calculate rectangle MVN probabilities (\ref{MLlik}) results in a numerically accurate likelihood method that is valid for any dimension \citep{nikoloulopoulos13b}.
By varying factors such as dimension $d$, the amount of discreteness (number of ordinal categories), and latent correlation for exchangeable structures, we demonstrate patterns in the asymptotic bias of the MSLE, and assess the performance of the simulated likelihood.
When computing the probability limits we take a constant dimension $d$ that increases and use discrete covariates where finite number of distinct values are assumed. Finally, without any loss of generality, we consider the case where the marginal parameters are common to different univariate margins.
Let the $T$ distinct cases for the ordinal response and the covariates be denoted as $$(\mathbf{y}^{(1)},\mathbf{x}^{(1)}),\ldots, (\mathbf{y}^{(T)},\mathbf{x}^{(T)}),$$ where $\mathbf{y}^{(t)}=(y_1^{(t)},\ldots,y_d^{(t)}),\,\mathbf{x}^{(t)}=(\mathbf{x}_1^{(t)},\ldots,\mathbf{x}_d^{(t)}),\, t=1,\ldots, T.$ In a random sample of size $n$, let the corresponding frequencies be denoted as $n^{(1)},\ldots, n^{(T)}$. Assuming a probability distribution on the covariates, for $t = 1,\ldots, T$, let $p^{(t)}$ be the limit in probability of $n^{(t)}/n$ as $n\to \infty$.
For the simulated likelihood in (\ref{MLlik}), we have the limit,
\begin{equation}\label{limitslik}
n^{-1}\ell(\boldsymbol{\beta},\boldsymbol{\gamma},\rho)\to \sum_{t=1}^{T} p^{(t)}
\log{h(y_{1}^{(t)},\ldots,y_{d}^{(t)};\boldsymbol{\beta},\boldsymbol{\gamma},\mathbf{R})}),
\end{equation}
where $h(\mathbf{y}^{(t)};\boldsymbol{\beta},\boldsymbol{\gamma},\mathbf{R})$ is computed through the method proposed in \cite{genz&bretz02}. As $n\to \infty$, the limit of the MSLE, $(\boldsymbol{\beta}^{SL},\boldsymbol{\gamma}^{SL},\rho^{SL})$, is the maximum of (\ref{limitslik}), while the limit of the standard maximum likelihood estimator (MLE) is the maximum of (\ref{limitslik}) where $h(\mathbf{y}^{(t)};\boldsymbol{\beta},\boldsymbol{\gamma},\mathbf{R})$ is computed through the 1-dimensional integral method in \cite{Johnson&Kotz72}.
We compute these limiting MSLE in a variety of situations to show clearly if the SL method is good. By using these limits, we do not need Monte Carlo simulations for comparisons, and we can quickly vary parameter values and see the effects. The $p^{(t)}$ in (\ref{limitslik}) are the model based probabilities $h(\mathbf{y}^{(t)};\boldsymbol{\beta},\boldsymbol{\gamma},\mathbf{R})$, and computed with the 1-dimensional integral method in \cite{Johnson&Kotz72}.
Our results are in line with the ones in \cite{nikoloulopoulos13b} for binary and count regression with dependent data; that is the proposed MSLE are identical with MLE up to four decimal places.
\baselineskip=12pt
|
\section{Introduction}
In earlier contributions, we have highlighted the importance of identifying fossils groups or dynamically relaxed groups and clusters \citep{Khosroshahi2006,Khosroshahi2007,Khosroshahi2014,Raouf2014,Miraghaei2014,Gozaliasl2014,Khosroshahi2016}. Adapting these classifications and continued studies enable us (I) to explore if these systems truly follow a different evolutionary path in their galaxy or halo properties, (II) to employ these galaxy systems and their statistical properties to identify the best possible model of galaxy formation and evolution, generally implemented in cosmological simulations and (III) to better understand the galaxy-halo connection.
Dark matter simulations have shown that galaxies in compact groups should merge into a single massive galaxy within a Gyr \citep{Barnes1989, Bode1993}. Consequently, an elliptical galaxy is formed, developing a large luminosity gap while the X-ray emitting halo remains unaffected by merging \citep{Ponman1994}. Such groups are known as fossil groups in which the essential observational tracers have been identified including the luminosity gap between the first and second brightest galaxy group members and the presence of an extended, i.e. group scale, X-ray emission with a luminosity of at least $L_{X,bol}\approx 10^{42} $~h$_{50}^{-2}$ ~erg~s$^{-1}$ \citep{Jones2003}. There are several studies in the literature focusing on the detailed characterization and properties of fossil groups base on X-ray and optical observations \citep{Khosroshahi2004,Sun2004,Ulmer2005,Khosroshahi2006,Khosroshahi2007,Miraghaei2014}, cosmological simulations \citep{Yoshioka2004,Milosavljevic2006,VandenBosch2007,VonBenda-Beckmann2008,Deason2013}, semi-analytical models \citep{Sales2007,Dariush2007,Diaz2008,Dariush2010,Raouf2014} and hydrodynamical simulations \citep{DOnghia2005,Cui2011}. Recent study of \cite{Khosroshahi2014} reveals that a diffuse extended X-ray emission beyond the optical size of the brightest group galaxy, exists specially when a large magnitude gap is present.
\citet{Khosroshahi2006} presented evidences that the majority of Brightest Group Galaxy (BGG) dominating fossil galaxy groups have non-boxy isophotes which could point to wet, or gas rich, nature of galaxy merger in their evolutionary history. \cite{Smith2010} employed a large sample of BGGs observed with the Hubble Space Telescope and found the trend in the luminosity gap (as an indication for the dynamical age of the system) and the isophotal shape of the BGGs, to be consistent with earlier study of \citet{Khosroshahi2006}. Furthermore, in comparison with the general population of galaxy groups, \citet{Khosroshahi2007} show that for a given optical luminosity, fossil groups not only contain hotter Intra Galactic Medium(IGM) for a given halo mass, but also their dark matter halo is more concentrated, all pointing at their relatively earlier formation epoch. In addition, the study of scaling laws in fossil groups indicate that they mostly follow the trend of galaxy clusters which is likely to be driven by dynamically relaxed state of cluster core. It worths highlighting an apparent conflict, as \citet{Voevodkin2010} show that there is no noticeable difference between the X-ray luminosity of the fossils and non-fossils for a given optical luminosity. While more recent studies support the latter \citep{Aguerri2011,Proctor2011,Harrison2012,Girardi2014}, however, the apparent contradiction could primarily be originated from the sample selection, and is due to the fundamental differences between galaxy groups, which forms the basis for \citet{Khosroshahi2007}, and galaxy cluster sample by \cite{Santos2007} which forms the basis for contrasting studies. In a Lambda-CDM model galaxy clusters are generally young assembly of galaxies while galaxy groups can be old and young depending on whether they survive major mergers during the hierarchical cosmic evolution.
\begin{figure}
\includegraphics[width=0.47\textwidth]{redshift_mag_SDSS_FC.pdf}
\caption{Absolute r-band magnitude distribution as a function of redshift based on SDSS DR10 group catalogue of \citet{Tempel2014}. The red triangles show the BGGs belonging to fossil groups ($\Delta M_{12}\gtrsim2$) while blue diamonds show non-fossil ($\Delta M_{12}<0.5$) galaxy groups. In addition, The red and blue dots represent the members of fossil and non-fossil galaxy groups, respectively.}
\label{SDSS:fig}
\end{figure}
In recent years, cosmological simulations have offered the necessary tools to address open questions, regarding the formation and evolution of galaxies. This is generally achieved through Semi-Analytical Models (SAMs) and hydrodynamical models. In semi-analytic approach the baryonic matter properties are calculated on the basis of analytical prescription in a post-processing procedure built on the merger tree \citep{Croton2006,Bower2006,DeLucia2007,Guo11}. In hydrodynamical approach, baryons directly interact and co-evolve with the dark matter particles within the cosmological volume. Although, the hydrodynamic approach has the upper hand in dealing with baryonic matter that can be directly linked to the gas properties ( such as cooling , heating and feedback process in and around galaxy halos \citep[and their references]{Springel2005a,Vogelsberger2014a}, but the semi-analytic approach is computationally inexpensive compared to the hydrodynamic and facilitate to construct sample of galaxies which are an order of magnitude larger than the same allowed by hydrodynamical simulations. Furthermore, the SAMs are more suitable for adding in new physics and assessing the impact.
A number of authors have suggested ways in which, radio-AGN heating is powerful enough to expel a fraction of baryons from the galaxy groups or clusters \citep{Croton2006,Bower2008}. Observationally, some studies show that radio-AGN heating could account for the missing baryons in galaxy groups \citep{Oklopcic2010,Giodini2010}. A useful approach for understanding the role of AGN feedback in galaxy evolution is to connect the astrophysical parameters related to the AGN feedback to observable quantities and make predictions which can be verified by the existing or future observations.
In a recent study \citep{Raouf2014}, we established a set of four observationally measurable parameters using the semi-analytic models of \cite{Guo11}, based on the Millennium Simulation, which can be used in combination, to identify a subset of galaxy groups which are dynamically old, with a very high statistical probability. We argued that a sample of fossil groups selected based on luminosity gap will result in a contaminated sample of old galaxy groups. However, by adding constraints on the offset between the group luminosity centroid and the BGG position, we considerably improved the age-dating method for galaxy groups and clusters, in comparison to the method based on the luminosity gap only.
\begin{figure}
\includegraphics[width=0.5\textwidth]{Gap12_compare.pdf}
\caption{Distribution of luminosity gap ($\Delta m_{12}$) for 3 categories; SDSS (brown dotted line), MS-Guo+11 (blue dashed line) and IS-1 (red dotted-dashed line). Colour filled regions represent the poison errors for each bin in the three categories. The IS-1 distribution shows that a large fraction of galaxy groups have large luminosity gap in contrast of the other two. The fraction of fossil groups ($\Delta m_{12} > 1.7 \ mag$) in IS-1, MS-Guo+11 and SDSS are $\approx$47, 22 and 27 per cent, respectively. The vertical gray dashed line marks the luminosity gap for fossil galaxy groups and the black dashed line marks the completeness limit of the luminosity gap for the SDSS sample. }
\label{gap12}
\end{figure}
\begin{figure}
\includegraphics[width=0.47\textwidth]{BF_diagram_all.pdf}
\caption{Absolute r-band magnitude of the first (triangle) and the second ranked (diamond) galaxies as function of luminosity gap, ($\Delta m_{12}$), for 3 categories of SDSS (brown dotted line), MS-Guo+11 (blue dashed-line) and IS-1 (red dotted-dashed-line). The data points illustrate the average bin of data and the error bars present the standard deviation of bind data points in each categories.}
\label{BF:fig}
\end{figure}
In this study, the main focus is to explore the Illustris in the context of luminosity gap formation and the advantage that this simulation may offer in providing a hydro based measure of accretion rate and super massive black hole mass. Thus we prepare an observed sample of galaxy systems using the Sloan Digital Sky Survey (SDSS) data release 10 \citep{Ahn2014} and similarly in the Millennium Simulation \citep{Springel2005} joined with \cite{Guo11} Semi-Analytic model and high resolution gravitational and hydrodynamical simulation of Illustris-1 \citep{Vogelsberger2014a}. Section 2 describes the data that we are using for simulation and observation. In Section 3 we describe our analysis and compare the luminosity gap of two brightest group galaxies and the map of luminosity gap -- centroid off-set. Finally in Section 4, we present summery of our results and conclusion. In this paper a $\Lambda CDM$ cosmology with $\Omega_m = 0.3$, $\Omega_{\Lambda} = 0.7$ and $H_0 = 100 h$ km $s^{-1}$ $Mpc^{-1}$ where $h=0.7$ is assumed.
\section{Data \& Simulations}
\subsection{Observations: SDSS group catalogues}
We use the legacy archive of the Sloan Digital Sky Survey-III, Data Release 10 \citep[SDSSIII-DR10] {Ahn2014} which covers 14,555 $deg^2$ in imaging data containing 469053874 unique objects for which 3276914 spectra were measured.
In this study, we use the FOF group/cluster catalogue of \citet{Tempel2014} for redshift range between 0.015 and 0.05 (Figure \ref{SDSS:fig}). Morphologically, all BGGs in our sample are elliptical with absolute r-band magnitude of $M_r(BGG)< -22$ mag and reside in halos with masses (referred in the catalogue as "massNFW") $\geq 10^{13} M_{\odot}$. Groups contain at least 4 spectroscopic members. With these constraints the observational sample contain 300 galaxy groups/clusters.
We estimate the offset ($D_{off}$) between the BGG location and the luminosity centroid, using the r-band magnitude of group's spectroscopic members and their coordinates. The luminosity gap in the r-band is obtained within 500 kpc/h radius from the BGG. Figure \ref{SDSS:fig} shows the distribution of absolute magnitude in the r-band vs. the redshift for all galaxies (gray points) as well as the selected sample in the SDSS (blue and red points).
\subsection{Simulations: Illustris-1 and Millennium Simulations}
We use the public release of the Illustris-1 Simulation \citep[Hereafter: IS-1]{Vogelsberger2014a}
a series of gravity as well as hydrodynamics realizations of a $(106.5$ $Mpc)^3$ cosmological volume that contains $1820^{3}$ gas cell and $1820^{3}$ dark matter particles, run with the AREPO code \citep{Springel2010}.
The highest-resolution run of the Illustris-1 handles the dark matter (DM) component with a mass resolution of $m_{DM} = 6.3 \times 10^6M_{\odot}$ and a baryonic component with $m_{baryon} = 1.6 \times 10^6M_{\odot}$ in 136 snapshots from z = 127 to z = 0 by adopting cosmological parameters consistent with the latest Wilkinson Microwave Anisotropy Probe 9 observations \citep[WMAP-9]{Hinshaw2013}. Halos, subhalos, and their basic properties have been identified with the FOF and SUBFIND algorithms \citep{Davis1985,Springel2001,Dolag2009} at every stored snapshots. Based on halo mass limit of $M_{halo}\geq 10^{13}-10^{14} M_{\odot}$ on galaxy systems with BGG absolute r-band magnitude of $M_r(BGG)< -22$ mag and multiplicity of at least 4 members for groups, the number of galaxy group in the present epoch is reduced to $\sim$ 190 systems containing $\sim$ 15000 galaxies. See \cite{Nelson2015} for more detailed description of the galaxy group catalogue properties.
In addition, we are using the Millennium Simulation \citep[MS]{Springel2005} joined with the \cite{Guo11} Semi-Analytical Model (Hereafter: MS-Guo+11) to extract galaxy properties. The cosmological model adopted in the Millennium Simulation is consistent with the first Wilkinson Microwave Anisotropy Probe 1 data \citep[WMAP-1]{Spergel2003} (note that the value of ${\sigma}_8$ is assumed to be greater than its present value of 0.82 given by WMAP-9 that is not strongly affects in this study). The simulation box $(500 h^{-1} Mpc)^{3}$ contains $2160^{3}$ particles and presents the mass resolution of $8.6 \times 10^8 h^{-1}M_{\odot}$. The dark matter merger trees within each simulation snapshot (64 snapshots in total) are spanned approximately logarithmically in time between $z=127$ and $z=0$ and extracted from the simulation using combination of FoF \citep{Davis1985} and SUBFIND \citep{Springel2001} halo finders algorithms. The gas and stellar components of galaxies in dark matter halos are constructed semi-analytically, based on laying a series of couples differential equations on top of the halo merger trees. In this study, we use \citet{Guo11} semi-analytical model at the present epoch which contains $\sim 23000$ galaxy groups/clusters with at least 4 members and halo mass above 10$^{13}$~M$_{\odot}$ to $\sim 10^{14} M_{\odot}$ with BGG absolute r-band magnitude of $M_r(BGG)< -22$ mag and $\sim2$ million galaxies.
Luminosity centroid for the simulations is defined base on $X_L= \sum X_i L_i/ \sum L_i$, where $L_i$ is the luminosity of a galaxy within a group in the r-band and $X_i$ is the projected coordinate of each galaxy within the radius of $r_{200}$.
Finally, we use the r-band magnitude of the group members and their coordinates to obtain the luminosity gaps within 500 kpc/h radius in each simulation.
\section{Result}
\subsection{The Luminosity Gap} \label{gap:sec}
In the previous studies of fossil groups, the luminosity gap between the two most luminous galaxies, located within a given physical radius of the group center (e.g. 0.5 R$_{200}$), has been used as a statistical tool to probe the accuracy of a number of semi-analytic galaxy formation models in cosmological simulations \citep{Dariush2007,Dariush2010,Smith2010,Raouf2014,Gozaliasl2014}. \cite{Dariush2007} used \cite{Croton2006} SAM in MS studies, to predict that fossil systems could be found in significant numbers (3--4 per cent of the population) even in quite rich clusters. Other probes have also been proposed, for instance \cite{Dariush2010} introduced $\Delta$m$_{14} \geq 2.5$, i.e. the luminosity gap between the first and fourth brightest galaxies within $0.5R_{200}$, as opposed to the conventional $\Delta$m$_{12} \geq 2.0$.
\cite{Smith2010} combined a series of observational data to study the luminosity gap statistics within a radius of $\sim$640 kpc a sample of 59, intermediate mass, galaxy clusters. They show that base on the luminosity gap parameters, 8$\pm3$ per cent of the sample are fossil systems. Recently, \cite{Gozaliasl2014} studied luminosity gap distribution using a large sample of X-ray galaxy groups (129 groups) spanned over redshift $z\le 1$ in the XMM-LSS X-ray observations and the CFHT follow-up optical observations. They found that 22$\pm6$ per cent of groups at z $\leq$ 0.6 are fossils.
\begin{figure}
\includegraphics[width=0.5\textwidth]{lum_gap_separation.pdf}
\caption{Distribution of the luminosity gap parameter as function of inverse centroid offset ($1/D_{off}$) for the SDSS, IS-1 and MS-Guo+11 samples. The colour coding is base on the halo age. For the SDSS sample this is based on the method of age assignment in \citep{Raouf2014}. In the simulations, this is based on the fraction of halo mass in z=1 to the final mass (z=0). Thus, the red colour represent the oldest galaxy groups and the blue marks the youngest galaxy groups. The gray dotted, dot-dashed and dashed lines show the mean of the binned data points with the standard deviation error bar ($\sigma$) for SDSS, IS-1 and MS-Guo+11, respectively. }
\label{gap_off}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.8\textwidth]{toy-model-age1.pdf}
\caption{Median evolutionary track of galaxy groups in the plane of the luminosity gap and the BGG off-set ($\Delta M_{12}- D_{off}$). The evolutionary tracks are colour coded based on the age of the halo as in fig. \ref{gap_off}. Two filled symbols within the circles mark is are used to highlight the luminosity gap (size of the symbols) and also to indicate whether the BGG is located centrally or not. The dotted-line, dotted-dashed-line and dashed-line are the mean data points of SDSS, IS-1 and MS-Guo+11, respectively.}
\label{gap_off_model}
\end{figure*}
All of these studies rule out the possibility that a large luminosity gap has a statistical origin.
In this study, we estimate the luminosity gap parameter within a radius of 500 kpc/h for 3 galaxy group samples; SDSS (observational), MS-Guo+11 (simulations; semi-analytic) and IS-1 (simulations; hydrodynamical). In both observation and simulations, the luminosity gap is measured within a projected distance from the most luminous galaxy in the group.
Figure \ref{gap12} shows the distribution of the luminosity gap for the SDSS groups and the two aforementioned model predictions. The distributions show that the fossil groups, i.e. galaxy groups with a large luminosity gap, are overproduced by the IS-1 in comparison to the observations and semi-analytic model used here. For instance, the fraction of fossil galaxy groups base on the definition of luminosity gap ($\Delta m_{12}\geq$ 1.7 mag; vertical gray dashed-line) in IS-1, MS-Guo+11 and SDSS are $\approx$ 47, 22 and 27 per cent, respectively. Moreover, galaxy groups with small magnitude gaps are under produced in contrast to observation and previous studies \citep{Smith2010}. Note that, $\Delta m_{12} = 4$ mag upper limit has been adopted base on the redshift completeness in the SDSS galaxies.
As in \citet{Smith2010} and \citet{Tavasoli2011} , we present in Figure \ref{BF:fig} the distribution of the luminosity gap as a function of the SDSS absolute $r$-band magnitudes \citep[rest frame;][]{Stoughton2002} for the first and second ranked group galaxies in the SDSS, MS-Guo+11 and IS-1. The figure suggests that in contrast to IS-1, MS-Guo+11 and SDSS follow the same trend.
We note here that there are several adjustments that one could make in order to make SAMs consistent with a set of observational properties of galaxies. For example although all SAMs based on the Millennium simulation \citep[e.g.][]{Croton2006,Bower2006,DeLucia2007,Guo11} well reproduce general observed characteristics of galaxies such as luminosity function and colour bi-modality, which indeed are important factors to properly recover majority of observed galaxy properties, they do not fully agree on the prediction of observable parameters such as the luminosity gap \citep[e.g. $\Delta m_{12}$ or $\Delta m_{13}$; ][]{Dariush2010}. The source of such a discrepancy, to some extent, depends on the ways different treatments have been implemented into SAMs to account for the dynamical friction (as well as other physical processes involved) in order to predict the faith of infall galaxies in groups/clusters. Indeed the main reason that the \cite{Guo11} model has been adopted for the purpose of current study is its robustness against the luminosity gap measurements \citep[e.g.][]{Gozaliasl2014b}
\begin{table}
\centering
\caption{Median luminosity gap ($\Delta M_{12}$) and centroid offset ($D_{off}$) in different redshift bins for 2 categories of galaxy groups shown in fig. \ref{gap_off_model} (1) sky blue and (2) red. Column 1: redshift; Column 2,4: magnitude gap between the two most luminous galaxies in the group; Column 3,5: physical separation between the BGG and the luminosity centroid of the group, centroid offset.}
\begin{tabular}{ccccc}
\hline
redshit & $\Delta M_{12}(1)$ & $D_{off}(1)$ & $\Delta M_{12}(2)$ & $D_off(2)$\\
-- & [mag] & [Mpc] & [Mag] & [Mpc] \\
\hline\hline
0.98871& 0.88478& 0.13739& 0.86139 & 0.06097\\
0.90546& 0.81897& 0.12588& 0.95807 & 0.05898\\
0.8277 & 0.87772& 0.13164& 0.9606 & 0.06625\\
0.75504& 0.83571& 0.13995& 0.91877 & 0.06247\\
0.68711& 0.82057& 0.12339& 0.81165 & 0.05092\\
0.62359& 0.87549& 0.12641& 0.94561 & 0.04388\\
0.56418& 0.89932& 0.12322& 0.9324 & 0.04949\\
0.50859& 0.90665& 0.12058& 1.03455 & 0.04188\\
0.45658& 0.98908& 0.11449& 1.0715 & 0.03431\\
0.4079 & 1.00568& 0.12119& 1.15871 & 0.02996\\
0.36234& 1.07564& 0.11808& 1.24518 & 0.0296\\
0.3197 & 1.09738& 0.11 & 1.25695 & 0.02564\\
0.2798 & 1.11613& 0.12398& 1.38156 & 0.02494\\
0.24247& 1.17151& 0.12331& 1.58907 & 0.02589\\
0.20755& 1.19065& 0.14448& 1.74282 & 0.02078\\
0.1749 & 1.24234& 0.15149& 1.97987 & 0.01951\\
0.14438& 1.27901& 0.21507& 2.12837 & 0.0189\\
0.11588& 1.30518& 0.40083& 2.26004 & 0.01662\\
0.08929& 1.35794& 0.82458& 2.30102 & 0.01515\\
0.06449& 1.3548 & 0.9089 & 2.34741 & 0.0124\\
0.0414 & 1.39455& 0.93671& 2.41839 & 0.01003\\
0.01993& 1.4333 & 0.95849& 2.49909 & 0.00796\\
0.0 & 1.42636& 1.00297& 2.57957 & 0.00624\\
\hline
\end{tabular}
\label{gap-off:tab}
\end{table}
\subsection{Magnitude gap vs. BGG centroid offset}
In \cite{Raouf2014}, we show that the luminosity gap and the offset between the location of the BGG and the luminosity centroid are useful indicators for the dynamical age or virialization state of galaxy groups.
Thus in Figure \ref{gap_off}, we show the correlation of the luminosity gap with the inverse of centroid offset ($1/D_{off}$) for three group catalogues of SDSS, IS-1 and MS-Guo+11. Data points are colour coded according to the halo age indicators i.e.: mass assembly history ($M_{200, z=1}/M_{200, z=0}$) in case of the simulation and the age probability map in the 3D parameter plane of $\Delta m_{12}$, centroid off-set and $M_r$(BGG) (associated to different range of group galaxy luminosities) as an indicator of the halo dynamical age in case of the SDSS data, as explained in \citet{Raouf2014}.
The gray lines in each panel shows the mean of binned data points with the standard deviation errors. Broadly, the three panels in figure \ref{gap_off} show similar trends in the $\Delta M_{12}- D_{off}$ relation. In comparison to the observations, the IS-1 appears to perform more successfully in predicting the observed distribution, however, as mentioned before it significantly over-predicts the fraction of large luminosity gap galaxy systems. A possible explanation is the highly enhanced dark matter particle mass resolution of IS-1 (by a factor of $\sim100\times$) compared to the Millennium dark matter simulation as this makes the former more robust in handling the dynamics of baryon particles.
\subsubsection{An evolutionary track for galaxy groups}
As stated above the luminosity gap and the luminosity offset complement each other to target the most evolved galaxy groups. To better understand it we superimpose all panels in Figure \ref{gap_off} on top of each other. This is shown in Figure \ref{gap_off_model} where data points represent the mean and standard deviation error-bars associated to SDSS (gray dotted-line), IS-1(gray dotted-dashed) and MS-Guo+11(gray dashed-line) catalogues. We trace back the evolution of galaxy groups located at the top-right of the Figure (red model i.e. high magnitude gap $\Delta M_{12}>2$ and small centroid offset $D_{off} < 50 kpc$) from the present epoch at $z=0$ to the snapshot corresponding to $z=1$ in MS-Guo+11. Median of the evolutionary tracks, colour coded according to the halo dynamical age (similar to bottom panel of Figure \ref{gap_off})), are also shown in Figure \ref{gap_off_model} . Hence the red colour indicates a high probability for a group to be old.
Similarly, we trace back the evolution of galaxy groups with large luminosity gaps ($1<\Delta M_{12}<2$) and a large centroid offset ($D_{off} > 300 kpc$) (sky blue-model) between z=0 and z=1. These type of galaxy groups also end up being originated from a population of young galaxy groups with a small luminosity gap.
Likewise, the median of the evolutionary track of sample groups is colour coded (sky-blue to blue) based on the age of halos. As indicated by colour of the track, these galaxy groups are not entirely populated by old groups. The evolution of the magnitude gap and centroid offset against the redshift in red and sky-blue galaxy models (indices 1 and 2 respectively) are summarised in Table \ref{gap-off:tab}.
\subsection{Black hole feedback} \label{BH:sec}
In the IS-1, black holes are implemented as sink particles \citep{Bellovary2010} and thus grow in mass by accreting surrounding gas or through black hole mergers and accretion. The black hole accretion is described by a Bondi-Hoyle-Lyttleton by eq. \ref{eq:BHaccr}
\begin{equation}
\dot M_\mathrm{BH} = \frac{4 \pi \alpha G^2 M_\mathrm{BH}^2 \rho}{(c_s^2 + v_\mathrm{BH}^2)^{3/2}}
\label{eq:BHaccr}
\end{equation}
where $\rho$ and $c_s$ are density and sound speed of the surrounding gas, respectively, and $v_\mathrm{BH}$ is the black hole velocity relative to the gas. Also, $\alpha$ and G are the stagnation point and the gravitational constants, respectively. In the IS-1, they use a repositioning scheme for black hole sink particles that connects them to the minimum of gravitational potential, in which case, they disregard the relative gas velocity term, $v_\mathrm{BH}$, in the accretion rate \citep[See also][]{Vogelsberger2013}.
In the IS-1, the AGN feedback regulates the star formation in galaxy formation process through thermal quasar-mode (cold-mode), thermal-mechanical radio-mode(hot-mode), and radiative mode of black hole accretion. At a high accretion rate with respect to the Eddington rate, cold-mode accretion, the black hole mass grows substantially. In contrast, the low accretion rate or radio mode, the AGN jets expand hot bubbles in the surrounding halo. The radiative AGN also known as electro-magnetic feedback impact on photo-ionisation and photo-heating rates which represent the net cooling rates for a short interval of cosmic time. Moreover, this feedback only present for each black hole which locate at the state of highest accretion around the Eddington limit \citep{Sijacki2007}.
In a recent observational study of dynamically relaxed (old) and unrelaxed (young) galaxy groups, we show that relaxed systems are less luminous in radio emissivity compared to unrelaxed galaxy groups \citep{Miraghaei2014}. In addition \cite{Suresh2015} study the central galaxies of FOF groups, based on hydrodynamical simulation, and show that the environment of central galaxies (e.g. BGGs) is influenced by the AGN feedback. They show that the radio mode feedback which inflates large hot bubbles, heats the environment of the BGG and reduces the fraction of cold gas for star formation. Another study by \cite{Genel2014} shows that AGN radio mode feedback operates as a powerful ejecting gas in most massive halos below $z = 1$ such that halos are almost devoid of gas, in disagreement with observations. Moreover, \cite{Vogelsberger2013} show that the radio mode feedback requires more power to suppress efficient cooling in massive halos compared to previous studies.
In Figure \ref{BH:fig}, we present the distribution of the instantaneous accretion ($dM_{BH}/dt$) of all black holes (top-panel), black hole mass (middle-panel) and gas fraction $f_{gas}$ in the sub-halo of the BGG (bottom-panel) as function of stellar mass for the BGG of old and young halos in the IS-1. As seen in Figure \ref{BH:fig}, the brightest group galaxy in the dynamically relaxed galaxy groups display a lower accretion rate compared to the brightest groups galaxies in dynamically young groups. At the same time the black hole mass in the BGGs dominating the dynamically relaxed groups is larger than in the BGGs of dynamically unrelaxed or young groups. This means that the mass assembly history of the group halos has a significant impact on the supermassive black hole of the brightest group galaxy. The BGGs dominating the dynamically old galaxy groups, seem to be very efficient in black hole growth by consuming the gas which could have been generally found with a higher density in early stages of the halo formation (as shown in bottom-panel of Figure \ref{BH:fig}). While this was argued in earlier studies of fossil groups, the Illustris provides the first direct numerical evidence.
\begin{figure}
\includegraphics[width=0.5\textwidth]{BGG-accration-mass_age.pdf}
\includegraphics[width=0.5\textwidth]{BGG-mass-BHmass_age.pdf}
\includegraphics[width=0.5\textwidth]{BGG-gasfrac-mass.pdf}
\caption{From top to bottom: mean values for the black hole accretion ($\dot{M}_{BH}$), black hole mass ($M_{BH}$) and gas fraction ($f_{gas}$) as a function of the BGG stellar mass for old (red) and young (blue) galaxy groups. The red line and blue dashed-line show the linear regressions to the old and young systems, respectively. The error bars in all panels are base on the standard deviation over mean, $\sigma / \sqrt{N}$. Accretion to the central black hole in the brightest group galaxies (top) is generally higher in young groups while their BGG black hole mass appears to be lower (middle). The fraction of gas in the BGG of young galaxy groups is higher than the BGG of old galaxy groups (bottom). }
\label{BH:fig}
\end{figure}
Note that, we define the relaxed (old) and unrelaxed (young) halos base on the accumulation of $> 50$ per cent and $< 30$ per cent of their final mass at z $\sim$ 1, respectively. Base on this definition, we find that 33 and 29 per cent of the halos out of about 190 halos with mass over $\sim 10^{13} M_{\odot}$ within IS-1, fall in the categories of old and young groups, respectively. The remaining halos form an intermediate population.
This is a new finding as the Illustris simulations is the first simulation, in cosmological scale, allowing us to study the growth of the supermassive black hole in the fossil dominant galaxies and also relative to giant elliptical galaxies with similar masses but in groups with a small luminosity gap. This finding makes direct connection between the dynamical state and thus the dynamical age of the halo and the growth of black hole mass. Observationally also \cite{Khosroshahi2016} show that there is a relation between the dynamical age of groups and radio luminosity of BGGs. This is following a study by \cite{Miraghaei2014} based on smaller sample. A popular argument to support the findings is based on the lack of recent on going galaxy mergers in dynamically old and fossil groups compared to their rivals, the dynamically young groups where there brightest group galaxy is expected to be surrounded by other massive galaxies \citep{Smith2010,Khosroshahi2007}.
\subsubsection{IGM temperature}
In the IS-1, the gas temperature in each cell is obtained from the internal energy $u$ and the electron abundance $x_e$. At the first, we are estimating the mean molecular weight using eq. \ref{molec:eq}
\begin{equation}
\mu = \frac{4}{1+3 X_H + 4 X_H x_e},
\label{molec:eq}
\end{equation}
where $X_H$ is equal to 0.76 and present the hydrogen mass fraction. Therefore, the temperature of cells in kelvin is estimated by eq. \ref{temp:eq}
\begin{equation}
T = (\gamma - 1) \frac{u}{K_B}(\mu m_p),
\label{temp:eq}
\end{equation}
where $ \gamma = 5/3 $ is the adiabatic index, and $m_p$ and $K_B$ are the proton mass and the Boltzmann constant, respectively.
The top-panel in Figure \ref{Temp:fig}, shows the Intra Galactic Medium (IGM) gas temperature as function of radial distance from the center in units of $r_{200}$ for dynamically old and young galaxy groups with halo masses over $10^{13} ~M_{\odot}$ at present epoch, z = 0. A comparison of the median temperature profile of old (red solid-line) and young (blue dashed-line) galaxy groups in IS-1, suggests that the IGM temperature in halos with earlier formation epoch, is systematically higher than the same in halos formed recently. In order to see if such an observed difference is due to a systematic bias in halo mass selection, we present in the lower-panel of Figure \ref{Temp:fig} the mean value of the IGM temperature, estimated within $r_{200}$, as a function of the halo mass, i.e. the $M-T$ scaling relation.
From X-ray and optical observation of a sample of galaxy groups, \citet{Khosroshahi2007} show that for the same halo mass, the IGM in fossil groups is hotter, compared to non-fossil groups.
Their sample was constructed from a small sample of fossil groups, e.g. $\Delta m_{12} \geq 2$, in which the brightest group galaxy was located at the X-ray emission pick suggesting a high degree of dynamical relaxation. Although the scale in which the IGM temperature and mass were measured in the observations, we compare to, differ from the scales shown in this study, however, we find that the results are robust and consistent with the observations. The scaling properties of relaxed and unrelaxed galaxy groups in Illustris Simulation and comparison with the observations will be presented in a separate study.
In most SAMs, analyses of the black hole feedback are estimated in a crude way by suppression of the gas cooling provided by the radio mode of AGN \citep{Croton2006}, then act to make stellar mass function close match to the observations. The gas temperature also has to be constant and equal to the virial temperature in all processes and there are simple assumptions in it which do not let us to use the temperature and feedback analysis for an instantaneous estimation. We try to address this issue in a separate study \citep[In prep.]{Raouf2016} by using the Semi Analytic Galaxy Evolution code \citep[SAGE;][]{Croton2016} which helps to use the temperature analysis and feedback process in a more physical way which closely match the previous X-ray observations.
\begin{figure}
\includegraphics[width=0.5\textwidth]{profile_temp_OL_mean.pdf}
\includegraphics[width=0.5\textwidth]{Group-M-T-age.pdf}
\caption{Top panel: IGM radial temperature profile for old (solid-line) and young (dashed-line) galaxy groups, for all the IS-1 halos with a halo mass above $10^{13} M_{\odot} h^{-1}$ at $z = 0$. The radius is given in units of the $r_{200}$. The red-line and blue dashed-line refer to the IGM median temperature profile of old and young halos, respectively. Early formed galaxy groups have a hotted IGM in comparison to late formed galaxy groups. Bottom panel: Mean value of $M-T$ relation for old (red-triangle) and young (blue-diamond) galaxy groups. The error bars present the standard deviation over mean of bind data points in each categories. The red-line and blue-dashed line show the linear regressions to the old and young systems, respectively.}
\label{Temp:fig}
\end{figure}
\section{Summary and discussion}
In this study we probe the distribution of the luminosity gap in Illustris, a new cosmological simulation, in which the properties of galaxies is dealt with hydrodynamically as opposed to a more economic, semi-analytic computation \citep{Vogelsberger2014a}. We find that galaxy groups with relatively large luminosity gaps are overproduced in Illustris with $\approx$ 47 per cent of groups having large magnitude gap ($\Delta m_{12}$) whereas the same is $\approx$ 22 per cent and $\approx 27$ per cent in the semi-analytic model of \cite{Guo11} and the SDSS based \citet{Tempel2014} group catalogue, respectively. However, we find that the Illustris recovers the observed trend in the plane of the luminosity gap and the offset between location of brightest group galaxy and halo center of mass, as two independent indicators for the halo dynamical state.
We show for the first time the evolutionary track of galaxy groups in the plane of luminosity gap ($\Delta M_{12}$) vs. BGG off-set ($D_{off}$) indicating that galaxy groups with large luminosity gaps, regardless of the position of the BGG within the group, are originated from small luminosity gap groups. However, majority of groups with BGGs at the centre of their halos are early-formed systems.
One could argue that the higher production of large luminosity gap systems in Illustris is due to an inefficient AGN feedback. Theoretical studies suggest that AGN activity should supply enough energy to prevent gas from being accumulated in the central regions of galaxy clusters and therefore quenching the formation of stars \citep{Tabor1993,Ciotti1997,Silk1998} . According to semi-analytical models, where the AGN feedback is paired with N-body simulations, most of the current stellar mass in the brightest cluster galaxies are assembled through dry minor mergers, following a phase of quiescent star formation influenced by feedback processes \citep{Croton2006,DeLucia2007,Bower2006,Guo11}.
Our study of the central black hole mass and the amount of mass it accreted, in old and young galaxy groups and for a given stellar mass, shows that the central black hole in fossil dominant galaxy (BGG) is noticeably more massive than a similar mass BGG in a non-fossil or a young galaxy group (e.g. Fig. \ref{BH:fig}). Furthermore, the black hole accretion in fossil dominant galaxies, on average, occurs at slower rates compared to systems with smaller luminosity gaps (i.e. non-fossil systems). This is consistent with our earlier observational findings in which fossil groups are less luminous in radio emission due to relatively less cold mode accretion \citep{Miraghaei2014}.
If galaxies in fossil groups are produced in major, possibly multiple, mergers at high redshifts, their supermassive black holes could well be more massive for their total stellar mass and as a result of no recent major merger or interaction with massive galaxies, the accretion to their central black hole could occur in a slower rate compared to galaxies which are subject to stronger interactions with more massive counterparts in groups with smaller luminosity gap \citep{Capelo2015,Lackner2014,Ellison2011,Silverman2011}. This argument is supported by a recent study in which over a dozen fossil groups were studies with a focus on their IGM properties \citep{Bharadwaj15} .
The Illustris Simulation does let us for the first time to study the IGM temperature profile, suggesting that the IGM in dynamically old galaxy systems is hotter in comparison to the IGM in dynamically young/evolving halos. This is in agreement with the previous observational study of \cite{Khosroshahi2007} in which a hotter IGM was found to be associated to fossil galaxy groups, compared to non-fossil groups, with the same halo mass.
\section*{Acknowledgments}
We thank the anonymous reviewer for constructive comments which helped us to improve the manuscript. We benefited from discussions with Joseph Silk and Gary Mamon in this study.
Authors acknowledge ISEF support for this study. The Illustris project acknowledges support from many sources.
VS acknowledges support by the DFG Research Centre SFB-881 "The Milky Way System" through project A1, and by the European Research Council under ERC-StG EXAGAL-308037. GS acknowledges support from the HST grants program, number HST-AR- 12856.01-A. Support for program 12856 was provided by NASA through a grant from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS 5-26555. LH acknowledges support from NASA grant NNX12AC67G and NSF grant AST-1312095. DX acknowledges support from the Alexander von Humboldt Foundation. SB was supported by NSF grant AST-0907969. DN acknowledges support from XSEDE grant AST-130032, which is supported by National Science Foundation grant number OCI-1053575. Thankfully acknowledge Dylan Nelson for facilitating the access to the data.
The Millennium database used in this paper and the web application providing online access to them were constructed as part of the activities of the German Astrophysical Virtual Observatory.
And www.sdss3.org/. SDSS-III is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS-III. Authors thank the anonymous referee for useful comments which helped the improvement of the text.
|
\section{Introduction}
Throughout this paper, we work over the field $\mathbb{C}$ of complex numbers. It has been well known by people working in higher dimensional geometry that there is an analogue between the local object Kawamata log terminal (klt) singularities $(X,o)$ and log Fano varieties as the global counterpart (cf. e.g. \cite{Sho00, Xu14} etc.). From this comparison, since the stability theory of Fano varieties has been a central object in people's study in the last three decades, it is natural to expect there is a local stability theory on singularities. The primary goal of this preprint is to develope such a theory. In another word, we want to investigate singularities using the tools from the theory of K-stability, a notion which was first defined in \cite{Tia97} and later algebraically formulated in \cite{Don02}. We note that this interaction between birational geometry and K-stability theory has been proved to significantly fertilize both subjects (cf. \cite{Oda12, Oda13, LX14, WX14, LWX14, Fuj15} etc.).
For the stability theory of log Fano varieties, a crucial ingredient is the CM weight. Philosophically, the stability of log Fano varieties is equivalent to minimizing the CM weight. In the stability theory of singularities, we fix the singularity $(X,o)$ and look for `the most stable' valuations $v$. Thus the first step of establishing a local stability theory for $(X,o)$ would be to find the right counterpart of CM weight in the local setting.
In the work \cite{Li15a}, the first named author defined the normalized volume function ${\widehat{\rm vol}}_{X,o}$ on the space of valuations centered at $o$, which we believe should be the right invariant. In fact, its derivative at the canonical divisorial valuation over the cone singularity along certain tangent direction gives the the CM weight. So in the local picture, the normalized volume function indeed carries more information than the CM weight!
By the above discussion, inspired by the global theory, we focus on studying the valuation minimizing the normalized volume function, which is conjectured to uniquely exist and ought to be thought as the `(semi-)stable' object. This picture is understood well in the case of Sasakian geometry where one only considers the valuations coming from the Reeb vector fields on the torus (e.g. \cite{MSY08, CoSz16}). Here we can naturally define the stability of the singularity using the one for the base. However, this requires the extra cone structure. By investigating the minimizer of the normalized volume function on all valuations, our plan, as we mentioned, built on the previous work (\cite{Li15a, Li15b, LL16}), is to establish an intrinsic stability which only depends on the isomorphic class of the singularity.
To revisit the cone singularity case, in fact it was shown in \cite{Li15b, LL16} that a Fano manifold $X$ is K-semistable, if and only if that among all valuations over the vertex $o$ of the cone $C(X)$ given by a multiple of $-K_X$, the canonical valuation defined to be the one obtained by blowing up the vertex $o\in C(X)$ minimizes the normalized volume function. This gives evidence to justify that at least for such singularities, our study is on the right direction.
For an {\it arbitrary} klt singularity, there is no direct way to associate a global object. Nevertheless, in differential geometry, when there is a `canonical' metric, the metric tangent cone around the singularity is the stable object in the category of metric space. With a similar philosophy, we expect the minimizer of the normalized volume function always gives {\it a degeneration} to a K-semistable Fano cone singularity in the Sasakian setting, and such degeneration should be provided by the minimizer of ${{\widehat{\rm vol}}}_{X,o}$. In the current paper, we work out this picture in the case that the minimizer is divisorial, by implementing the machinery of the minimal model program (based on the foundational results in \cite{BCHM10}). So our treatment will be purely algebraic though it is strongly inspired by profound analytic results.
One ingredient we introduce is to define the volume associated to a birational model and then connect it to the normalized volume of a valuation. For studying the divisorial valuations, the class of models which play a central role here are the ones obtained by the construction of {\it Koll\'ar} component (cf. \cite{Xu14}): for an arbitrary $n$-dimensional klt singularity $(X,o)$, we can use minimal model program to construct a birational model whose exceptional locus is an $(n-1)$-dimensional $\mathbb{Q}$-Fano variety. We will systematically develop the tools of using Koll\'ar components to understand the normalized local volume and its minimizer. In fact, Koll\'ar components can be considered as the local analogue of special degenerations studied in \cite{LX14}.
In the case of Sasakian geometry, one Reeb vector gives a Koll\'ar component if and only if it is rational, i.e., it is quasi-regular.
Therefore, to summarize, the aim of this paper is of twofolds: on one hand, we aim at using the construction of Koll\'ar components to get information of the valuations space, especially the minimizer of the normalized volume functions; On the other hand, in the reverse direction, we want to use the viewpoint of stability to study the birational geometry construction of Koll\'ar components, and search out a more canonical object under suitable assumptions.
We also expect for any klt singularity $(X,o)$, even when the minimizer is not necessarily divisorial, we can still use suitable birational models to degenerate $(X,o)$ to a K-semistable (possibly irregular) Sasakian singularity. However, it seems to involve a significant amount of new technical issues.
In the below, we will give more details.
\subsection{Koll\'ar components}
\begin{defn}[Koll\'ar component]\label{d-kollar}
Let $o\in (X,D)$ be a klt singularity. We call a proper birational morphism $\mu:Y\to X$ provides a {\bf Koll\'ar component $E$}, if $\mu$ is isomorphic over $X\setminus \{o\}$, and $\mu^{-1}(o)$ is an irreducible divisor $S$, such that $(Y,S+\mu^{-1}_*D)$ is pure log terminal and $-S$ is $\mathbb{Q}$-Cartier and ample over $X$.
\end{defn}
We easily see the birational model $Y$ is uniquely determined once the divisorial valuation $S$ is fixed, and if we denote
$$(K_Y+S+\mu_*^{-1}D)|_S=K_S+\Delta_S,$$
then $(S,\Delta_S)$ is a klt log Fano pair.
Given any klt singularity $(X,o)$, after the necessary minimal model program type result is established (see \cite{BCHM10}), we know that there always exists a Koll\'ar component (see \cite{Pro00} or \cite[Lemma 1]{Xu14}), but it is often not unique (nevertheless, see the discussion in \ref{e-example}.4 for some known special cases for the uniqueness). From what we have discussed, instead of an arbitrary Koll\'ar component, we want to study those which are `more stable', and show if it exists, it provides a more canonical object. Indeed, we prove that if there is a K-semistable Koll\'ar component, then it gives the unique minimizer of ${\widehat{\rm vol}}(X,o)$ among all Koll\'ar components.
\vskip 1mm
\vskip 1mm
Compared to the global theory of degeneration of Fano varieties, this fits into the philosophy that K-stability provides a more canonical degeneration(cf. \cite{LWX14, SSY14}) and it should minimize the CM weight among all degenerations. But one surprising thing to us is that K-semistability is enough, instead of K-polystability which was required in the global setting.
\bigskip
The following theorem is our main theorem.
\begin{thm}\label{t-main1}Let $o\in (X,D)$ be klt singularity.
A divisorial valuation
${\rm ord}_S$ is a minimizer of ${\widehat{\rm vol}}_{X,D, o}$ if and only if the following conditions are satisfied
\begin{enumerate}
\item
$S$ is a Koll\'{a}r component;
\item
$(S, \Delta_S)$ is K-semistable.
\end{enumerate}
Moreover, such a minimizing divisorial valuation is unique.
\end{thm}
More precisely, we will prove Theorem \ref{t-main1} by proving the following four theorems. For each of them, we need somewhat different techniques.
First we prove
\begin{thmx}\label{t-main}
If $o\in (X,D)$ is an algebraic klt singularity. Let $S$ be a Koll\'ar component over $X$.
If $(S,\Delta_S)$ is (log-)$K$-semistable. Then ${\widehat{\rm vol}}_{X,o}$ is minimized at the valuation ${\rm ord}_S$.
\end{thmx}
This extends the main theorem in \cite{LL16} from cone singularities to the more general setting. For the proof, we indeed degenerate a general singularity to a cone singularity induced by its Koll\'ar components. However, instead of degenerating the valuation, we degenerate the associated ideals and then use the result in \cite{Liu16}. An extra subtlety is to treat the case of the cone singularity, we can not directly use \cite{LL16} as there they only proved the result for the cone singularity over an analytic K-semistable $\mathbb{Q}$-Fano variety, which we still do not know to be equivalent to the algebraic definition. As a redemption, we first show that it suffices to concentrate on the equivariant data and then use \cite{Li15b} to finish the argument.
In Section \ref{s-exam}, we use this criterion to find minimizers for various examples of singularities including: quotient singularities, $A_k$ and $E_k$ singularities etc.
\bigskip
Next, we turn to the result on the uniqueness.
\begin{thmx}\label{t-main2}
If $o\in (X,D)$ is an algebraic klt singularity. Assume $S$ is a Koll\'ar component over $X$ such that $(S,\Delta_S)$ is $K$-semistable. Then
$${\widehat{\rm vol}}_{X,o}({\rm ord}_S)<{\widehat{\rm vol}}_{X,o}({\rm ord}_T)$$ for any other divisorial valuation $T$.
\end{thmx}
This is done by a detailed study of the geometry when the equality holds. In the cone singularity case, we investigate the equality condition in the calculation in \cite{Li15b}. It posts a strong assumption which enables us to compute the corresponding invariants including nef thresholds and pseudo-effective thresholds. The argument is partially inspired by the work in \cite{Liu16}. Once this is clear, the rest follows from a simple application of Kawamata's base point free theorem. And the general case can be again reduced to the case of cone singularity using a degeneration process.
\bigskip
Now we consider the converse direction.
For any klt singularity, a minimizer of the normalized volume function always exists by \cite{Blu16b}. The following theorem says that if a minimizer is divisorial, it always yields a Koll\'ar component. We indeed prove slightly more for a general rational rank 1 minimizer.
\begin{thmx}\label{t-divisor}
Given an arbitrary algebraic klt singularity $o\in (X,D)$ where $X=\mathrm{Spec}(R)$. Let $v$ be a valuation that minimizes ${\widehat{\rm vol}}_{X,o}$. Assume the valuation group of $v$ is isomorphic to $\mathbb{Z}$, i.e., $v$ has rational rank one, and one of the following two assumptions holds
\begin{enumerate}
\item $v$ is a multiple of a divisorial valuation; or
\item the graded family of valuative ideals $$\mathfrak{a}_{\bullet}=\{\mathfrak{a}_k\} \mbox{\ where \ } \mathfrak{a}_k=\{f\in R\ |\ v(f)\ge k\} $$ is finitely generated, i.e., there exists $m\in \mathbb{N}$ such that $\mathfrak{a}_{mk}=(\mathfrak{a}_m)^k$ for any $k\in \mathbb{N}$.
\end{enumerate}
Then up to a rescaling, $v$ is given by the divisorial valuation induced by a Koll\'ar component $S$.
\end{thmx}
The fact that the assumption 1 above implies $v$ is given by a Koll\'ar component is also independently proved in \cite{Blu16b}.
We note that a minimizer is conjectured to be quasi-monomial ((cf. \cite[Conjecture 6.1.3]{Li15a})) and the graded family of valuative ideals for a minimizer of the normalized volume function is conjectured to be always finitely generated (cf. \cite[Conjecture 6.1.5]{Li15a}). So granted any one of these two conjectures, this result should presumably characterize all the cases with minimizers of rational rank 1.
The proof uses the definition of the volumes of a model, and run the decreasing process of the volumes given by the minimal model program as in \cite{LX14}.
\bigskip
Next we turn to the stability of the minimizer. By using the techniques from the toric degeneration (see \cite{Cal02, AB04, And13}) and the relation between CM weight and normalized volumes, we
will prove
\begin{thmx}\label{t-mintok}
We use the same notation as in Theorem \ref{t-divisor}. Let $\mu\colon Y\to X$ be the morphism which extracts $S$, and write $(K_Y+S+\mu_*^{-1}D)|_S=K_S+\Delta_S$, then $(S,\Delta_S)$ is a K-semistable log Fano pair.
\end{thmx}
\subsection{Approximation}
In a different direction, we also obtain results which describe the minimizer of the normalized volume function from the viewpoint of Koll\'ar components. We show
for a general klt singularity, although the minimizer of its associated normalized volume function might not be given by a Koll\'ar component, but we can always approximate it by a sequence of Koll\'ar components.
\begin{thm}\label{t-approx}
Given an arbitrary algebraic klt singularity $o\in (X,D)$, and a minimizer $v^{\rm m}$ of ${\widehat{\rm vol}}_{X,o}$, there always exists a sequence of Koll\'ar components $\{ S_i\}$ and positive numbers $c_j$ such that
$$\lim_{j\to \infty}c_j\cdot {\rm ord}_{S_j}\to v\mbox{\ in }{\rm Val}_{X,o} \mbox{\ \ \ and \ \ \ } \lim_{i\to \infty} {\widehat{\rm vol}}({\rm ord}_{S_i})={\widehat{\rm vol}}(v^{\rm m}).$$
\end{thm}
Here ${\rm Val}_{X,o}$ consists of all valuations centered at $o$, and is endowed with the weakest topology as in \cite[Section 4.1]{JM12}.
\subsection{Equivariant K-semistability}
By relating a Fano variety and the cone over it, we can compare the calculation in \cite{Li15b} for a cone and \cite{Fuj16} for its base. Then an interesting by product of our method is the following theorem.
\begin{thm}\label{t-equiK}
Let $T\cong (\mathbb{C}^*)^r$ be a torus. Let $(V,\Delta) $ be a log Fano variety with a $T$-action. Then $(V,\Delta)$ is K-semistable if and only if any $T$-equivariant special test configuration $\mathcal{S}\to \mathbb{A}^1$ of $(V,\Delta)$ has its generalized Futaki invariant ${\rm Fut}(\mathcal{V})\ge 0$.
\end{thm}
When $S$ is smooth and $\Delta=0$, this follows from the work of \cite{DS15} with an analytic argument. Our proof is completely algebraic. It again uses the techniques of degenerating any ideal to an equivariant one and showing it has a smaller invariant.
\bigskip
The paper is organized in the following way: In Section \ref{s-pre}, we give some necessary backgrounds. In Section \ref{s-vmodel}, we introduce one key new tool: the volume of a model. By combining the normalized volume function on valuations with the local volume defined in \cite{Ful13}, and applying minimal model program, we prove Theorem \ref{t-approx} and Theorem \ref{t-divisor}. In Section \ref{s-min}, we prove Theorem \ref{t-main}, by connecting it to the infimum of ${\rm lct}(X,D;\mathfrak{a})^n \cdot {\rm mult}(\mathfrak{a})$ for all $\mathfrak{m}$-primary ideals $\mathfrak{a}$ centered on $o$. We note that this latter invariant indeed has also been studied in other context (cf. \cite{dFEM04}). In Section \ref{s-uni}, we prove Theorem \ref{t-main2}. We first prove it for the cone singularity case, using heavily the ideas and calculations in \cite{LL16}. Then we use a degeneration argument
to reduce the general case to the case of cone singularities. In Section \ref{s-Ksta}, we prove Theorem \ref{t-mintok}, which verifies the K-semistability of a minimizing Koll\'ar component. In Section \ref{s-exam}, we give some examples on how to apply our techniques to calculate the minimizer for various classes of klt singularities.
\vspace{5mm}
\noindent {\it History:} Since \cite{Li15a}, the study of the minimizer of the normalized volume function moves forward rapidly. Our preprint is inspired by the earlier work of \cite{Li15a, Li15b} and \cite{LL16}. It also uses ideas from \cite{Liu16}, which is written in the same period as the first version of our preprint. After we posted our preprint, the existence of the minimizer is completely settled in \cite{Blu16b}. Inspired by this work, in the revision, we improve our work by showing that for a $\mathbb{C}^*$-equivariant singularity, we only need to consider the equivariant valuations for minimizing the normalized local volume. We also include his result on the existence of the minimizer in the exposition. In particular, this allows us to stay with the algebraic definition of K-semistability. Another major improvement we achieve in the revision is that we can indeed show that any Koll\'ar component which minimizes the normalized local volume is always K-semistable.
\vspace{5mm}
\noindent
{\bf Acknowledgement}: We thank Yuchen Liu, Dhruv Ranganathan and Xiaowei Wang for helpful discussions and many useful suggestions. We especially want to thank Harold Blum and Mircea Musta\c{t}\v{a} for pointing out a gap in an earlier draft. CL is partially supported by NSF DMS-1405936. CX is partially supported by `The National Science Fund for Distinguished Young Scholars (11425101)'. Part of the work was done when CX visited Imperial College London and Massachusetts Institute of Technology. He wants to thank Paolo Cascini and Davesh Maulik for the invitation and providing a wonderful environment.
\section{Preliminary}\label{s-pre}
\noindent{\bf Notation and Conventiones:} We follow the standard notation in \cite{Laz, KM98, Kol13}. A {\it log Fano} pair $(X,D)$ is a projective klt pair such that $-K_X-D$ is ample.
For a local ring $(R,\mathfrak{m})$ and $\mathfrak{a}$ an $\mathfrak{m}$-primary ideal, we denote by $l_R(R/\mathfrak{a})$ the length of $R/\mathfrak{a}$.
For the K-semistability of a log Fano pair, see \cite{Tia97, Don02} (also see \cite{Oda13,LX14}).
\subsection{Normalized volume}
Let $(X,o)$ be a normal algebraic singularity and $v\in {\rm Val}_{X,o}$, which is the space of all valuations centered on $o$. Let $D\ge 0$ be a $\mathbb{Q}$-divisor such that $K_X+D$ is $\mathbb{Q}$-Cartier. We can define the volume ${\rm vol}_{X,o}(v)$ and the log discrepancy $A_{X,D}(v)$ (if the context is clear, we will abbreviate it as ${\rm vol}(v)$ and $A(v)$) following \cite{ELS03} and \cite{JM12} (see e.g. \cite[Section 1.1]{Li15a}). In particular, if $S$ is a divisor with center on $X$ to be $o$, we have
$$A_{X,D}(S):=A_{X,D}({\rm ord}_S)=a(S; X,D)+1$$
the same as the standard log discrepancy.
\begin{defn}Notation as above. We denote the {\bf normalized volume} by ${\widehat{\rm vol}}_{X,D,o}(v)$ (or ${\widehat{\rm vol}}_{X,o}(v)$ if $D$ is clear or ${\widehat{\rm vol}}_{X,D}(v)$ if $o$ is clear or simply ${\widehat{\rm vol}}(v)$ if there is no confusion) to be
$${\rm vol}_{X,o}(v)\cdot A^n(v)$$
if $A^n(v)<+\infty$; and $+\infty$ if $A^n(v)=+\infty$.
\end{defn}
Recall that it was conjectured that ${\widehat{\rm vol}}(v)$ achieves the minimum at a valuation $v\in {\rm Val}_{X,o}$. In \cite{Li15a}, it is showed that the space
$$\{v\in {\rm Val}_{X,o}|\ v(\mathfrak{m})=1, {\widehat{\rm vol}}(v) \le C \}$$
for any constant $C>0$ forms a compact set. However, in general the volume function ${\rm vol}$ is only upper continuous on ${\rm Val}_{X,o}$ although we expect it is continuous at a minimizer of ${\widehat{\rm vol}}$.
\begin{prop}\label{p-upperconti}
If $\{v_i\}$ is a sequence of valuations, such that $v_i\to v$ in the weakest topology, then
$${\rm vol}(v)\ge \limsup_i {\rm vol}(v_i). $$
\end{prop}
\begin{proof}The valuation $v$ determines a sequence of graded ideas
$$\mathfrak{a}_k=\mathfrak{a}_k(v)=\{f\in R\ |\ v(f)\ge k\}.$$
By \cite{Mus02}, we know that for any $\epsilon>0$, there exists a sufficiently large $k$ such that
$$\frac{1}{k^n}{\rm mult}(\mathfrak{a}_k)< {\rm vol}(v)+\epsilon.$$
Since $R$ is Noetherian, we know that there exist finitely many generators $f_p$ ($1\le p\le j$) of $\mathfrak{a}_k=(f_1,...,f_j)$. As $v(f_p)\ge k$, we know that for any $\delta$, there exists sufficiently large $i_0$ such that for any $i\ge i_0$, $v_i(f_p)\ge k-\delta$. Thus
$$\mathfrak{a}^{(i)}_{k-\delta}=\{f\in R\ |\ v_i(f)\ge k-\epsilon \}\supset \mathfrak{a}_k .$$
Therefore,
$${\rm vol}(v_i)\le \frac{1}{(k-\delta)^n}{\rm mult}(\mathfrak{a}^{(i)}_{k-\epsilon})\le \frac{1}{(k-\delta)^n}{\rm mult}(\mathfrak{a}_{k})\le \frac{k^n}{(k-\delta)^n}({\rm vol}(v)+\epsilon).$$
\end{proof}
\begin{prop}\label{p-valuation}
Let $(X,o)=(\mathrm{Spec}(R),\mathfrak{m})$ be a singularity. Let $v$ and $v'$ be two real valuations in ${\rm Val}_{X,o}$. Assume
$${\rm vol}(v)={\rm vol}(v')>0\qquad \mbox{and} \qquad v(f)\ge v'(f)$$ for any $f\in R^*$, then $v=v'$.
\end{prop}
\begin{proof} We prove it by contradiction. Assume this is not true, we fix $g\in R$ such that
$$v(g)=l>v'(g)=s.$$
Denote by $r= l-s>0.$ Fix $k\in \mathbb{R}_{>0}$. Consider
$$\mathfrak{a}_k:=\{f\in R | \ v(f)\ge k\}\qquad\mbox{ and} \qquad{\mathfrak{b}}_k:=\{f\in R |\ v'(f)\ge k\}. $$
So ${\mathfrak{b}}_k\subset \mathfrak{a}_k$, and we want to estimate the dimension of
$$\dim(R/{{\mathfrak{b}}_k})-\dim(R/{\mathfrak{a}_k})=\dim (\mathfrak{a}_k/{\mathfrak{b}}_k). $$
Fix a positive integer $m<\frac{k}{l}$ and a set
$$g^{(1)}_{m},...,g^{(k_m)}_{m} \in {\mathfrak{b}}_{k-ml}$$
whose images in ${\mathfrak{b}}_{k-ml}/{\mathfrak{b}}_{k-ml+r}$ form a $\mathbb{C}$-linear basis.
We claim that
$$\{f^m\cdot g^{(j)}_{m}\}\ \ (1\le m \le \frac{k}{l}, 1\le j\le k_m)$$
are $\mathbb{C}$-linear independent in $\mathfrak{a}_k/{\mathfrak{b}}_k$. Granted this for now,
we know since ${\rm vol}(v)>0$, then
$$\limsup_{\lim k\to \infty} \frac{1}{k^n}\sum_{1\le m \le \frac{k}{l}}k_m=\limsup_{\lim k\to \infty} \sum_{1\le m\le \frac{k}{l}} \frac{1}{k^n} \dim ({\mathfrak{b}}_{k-ml}/{\mathfrak{b}}_{k-ml+r})>0, $$
which then implies ${\rm vol}(v)>{\rm vol}(v')$.
\bigskip
Now we prove the claim.
\noindent{\bf Step 1:} For any $1\le m \le \frac{k}{l}, 1\le j\le k_m$,
\begin{eqnarray*}
v(f^m\cdot g^{(j)}_{m})&=&v(f^m)+v(g^{(j)}_{m})\\
&\ge& ml+v'(g^{(j)}_{m})\\
&\ge & ml+k-ml\\
&\ge &k.
\end{eqnarray*}
Thus $f^m\cdot g^{(j)}_{m}\in \mathfrak{a}_k$.
\vspace{3mm}
\noindent{\bf Step 2:} If $$\{f^m\cdot g^{(j)}_{m}\}\ \ (1\le m \le \frac{k}{l}, 1\le j\le k_m)$$ are not $\mathbb{C}$-linear independent in $\mathfrak{a}_k/{\mathfrak{b}}_k$,
then there is an equality
$$\sum_{m}h_m= b \in {\mathfrak{b}}_k,$$
where there exists $c_j\in \mathbb{C}$, such that
$$h_m=f^m\cdot \sum_{1\le j \le k_m} c_jg^{(j)}_{m}$$
and some $h_m\neq 0$. Consider the maximal $m$, such that $h_m\neq 0$.
Since
\begin{eqnarray*}
v'(h_m)&=&v'(f^m\cdot \sum_{1\le j \le k_m} c_jg^{(j)}_{m}) \\
&=& v'(f^m)+v'(\sum_{1\le j \le k_m} c_j g^{(j)}_{m})\\
&< & ms+k-ml+r\\
&= & k-(m-1)l+(m-1)s,
\end{eqnarray*}
where the third inequality follows from that
$$ \sum_{1\le j \le k_m} c_jg^{(j)}_{m} \notin {\mathfrak{b}}_{k-ml+r} .$$
However, we have
\begin{eqnarray*}
v'(h_m)&=&v'(b-\sum_{j<m}h_j) \\
&\ge & \min \big\{ v'(b), v'(h_1),....v'(h_{m-1}) \big\}\\
&= & \min_{1\le j \le m-1}\{ k , js+k-jl\} \\
&=& k-(m-1)l+(m-1)s,
\end{eqnarray*}
which is a contradiction.
\end{proof}
Several results in our work depend on a relation between the normalized volume and some normalized multiplicity. The latter quantity was first considered in the smooth case in \cite{dFEM04}, and since then it has been studied in many other work, including positive characteristic versions (see e.g. \cite{TW04}). Its relevance to the normalized volume appeared in \cite[Example 3.7]{Li15a}. We have the following more precise observation in \cite{Liu16}.
\begin{prop}[{\cite[Section 4.1]{Liu16}, cf. \cite[Example 3.7]{Li15a}}]\label{p-inf}
Let $(X,o)=({\rm Spec}R, \mathfrak{m})$ and $D\ge 0$ a $\mathbb{Q}$-divisor, such that $o\in (X,D)$ is a klt singularity. Then we have
\begin{equation}\label{eq-vol2mul}
\inf_{v} {\widehat{\rm vol}}_{X,o}(v)=\inf_{\mathfrak{a}}{\rm lct}^n(X,D; \mathfrak{a})\cdot{\rm mult}(\mathfrak{a}),
\end{equation}
where on the left hand side the infimum runs over all the valuations centered at $o$, and on the right hand side it runs over all the $\mathfrak{m}$-primary ideals. Moreover,
the left hand side can be replaced by $\inf_{v\in {\rm Div}_{X,o}} {\widehat{\rm vol}}_{X,o}(v)$ where ${\rm Div}_{X,o}$ denotes the space of all divisorial valuations with center at $o$.
\end{prop}
For the reader's convenience we provide a sketch of the proof.
\begin{proof
We first use the same argument as in \cite[Example 2.7]{Li15a}) to prove that the left hand side is greater or equal to the right hand side. For any real valuation $v$, consider
the graded family of valuative ideals
$$\mathfrak{a}_k=\mathfrak{a}_k(v)=\{f\in R\ |\ v(f)\ge k\}.$$
Then $v(\mathfrak{a}_k)\ge k$ and we can estimate:
\begin{eqnarray*}
A_{X,D}(v)^n \cdot \frac{{\rm mult}(\mathfrak{a}_k)}{k^n}&\ge & \left(\frac{A_{X,D}(v)}{v(\mathfrak{a}_k)}\right)^n \cdot {\rm mult}(\mathfrak{a}_k)\ge {\rm lct}^n(X,D; \mathfrak{a}_k)\cdot {\rm mult}(\mathfrak{a}_k).
\end{eqnarray*}
Since for $\mathfrak{a}_\bullet=\{\mathfrak{a}_k\}$ is a graded family of $\mathfrak{m}$-primary ideals on $X$,
\[
{\rm mult}(\mathfrak{a}_\bullet)=\lim_{k\rightarrow+\infty} \frac{l_R(R/\mathfrak{a}_k)}{k^n}=\lim_{k\rightarrow+\infty}\frac{{\rm mult}(\mathfrak{a}_k)}{k^n}.
\]
(see e.g. \cite{ELS03, Mus02, LM09, Cut12}), as $k\rightarrow +\infty$, the left hand side converges to ${\widehat{\rm vol}}(v)$ and we get one direction.
For the other direction of the inequality, we follow the argument in \cite{Liu16}. For any $\mathfrak{m}$-primary ideal $\mathfrak{a}$, we can choose a divisorial valuation $v$ calculating
${\rm lct}(\mathfrak{a})$. Then $v$ is centered at $o$. Assume $v(\mathfrak{a})=k$, or equivalently $\mathfrak{a}\subseteq \mathfrak{a}_k(v)$. Then we have $\mathfrak{a}^l \subseteq \mathfrak{a}_k(v)^l\subseteq \mathfrak{a}_{kl}(v)$ for any $l\in \mathbb{Z}_{>0}$. So we can estimate:
\begin{eqnarray*}
{\rm lct}^n(X,D; \mathfrak{a})\cdot {\rm mult}(\mathfrak{a})&=&\frac{A_{X,D}(v)^n}{k^n}\cdot {\rm mult}(\mathfrak{a})=A_{X,D}(v)^n\cdot\frac{{\rm mult}(\mathfrak{a}) l^n}{(kl)^n}\\
&=&A_{X,D}(v)^n \cdot \frac{{\rm mult}(\mathfrak{a}^l)}{(kl)^n}\ge A_{X,D}(v)^n\cdot \frac{{\rm mult}(\mathfrak{a}_{kl})}{(kl)^n}.
\end{eqnarray*}
As $l\rightarrow +\infty$, then again the right hand side converges to
$$A_{X,D}(v)^n \cdot {\rm mult}(\mathfrak{a}_\bullet(v))={\widehat{\rm vol}}(v).$$ So we have proved the right-hand-side is bigger than the left-hand-side in \eqref{eq-vol2mul}.
The last statement follows easily from the above proof.
\end{proof}
In a very recent preprint \cite{Blu16b}, it is proved that a minimizer always exists.
\begin{thm}[Blum] For any klt singularity $o\in (X,D)$, ${\widehat{\rm vol}}_{X,D}(v)$ always has a minimizer $v^{\rm m}$ in ${\rm Val}_{X,o}$.
\end{thm}
\subsection{Properties of Koll\'ar component}
The concept of {\it Koll\'ar component} is defined in Definition \ref{d-kollar}. It always exists (cf. see \cite{Pro00} or \cite[Lemma 1]{Xu14}). In this section, we establish some of their properties using the machinery of the minimal model program.
The following statement is the local analogue of \cite[Theorem 1.6]{LX14}, which can be also easily obtained by following the proof of the existence of Koll\'ar component. (See e.g. the proof of \cite{Xu14}.)
\begin{prop}\label{p-special}Let $o\in (X,D)$ be a klt singularlty. Let $\mu\colon Y\to X$ be a model, such that $\mu$ is an isomorphism over $X\setminus\{o\}$ and $(Y,E+\mu_*^{-1}D)$ is dlt where $E$ is the divisorial part of $\mu^{-1}(o)$.
Then we can choose a model $W\to Y$ and run MMP to obtain $W\dasharrow Y'$, such that $Y'\to X$ gives a Koll\'ar component $S$ with $a(Y,E; S)=-1$.
\end{prop}
We also have the following straightforward lemma.
\begin{lem}\label{l-inter}
If $S$ is a Koll\'{a}r component, then ${\rm vol}({\rm ord}_S)=(-S|_S)^{n-1}$ and ${\widehat{\rm vol}}({\rm ord}_S)=(-(K_Y+S+\mu_*^{-1}D)|_S)^{n-1}\cdot A_{X,D}(S)^n$.
\end{lem}
\begin{proof}
For any $k\ge 0$, we have an exact sequence,
$$0\to \mathcal{O}_{Y}(-(k+1)S)\to \mathcal{O}_{Y}(-kS)\to \mathcal{O}_{S}(-kS)\to 0.$$
Because $-S$ is ample over $X$, we have the vanishing
$$R^1f_*(\mathcal{O}_{Y}(-(k+1)S))=0,$$
from which we get
\[
H^0(S, -kS|_S)\cong \frac{H^0(Y, -kS)}{H^0(Y, -(k+1)S)}=\frac{\mathfrak{a}_k({\rm ord}_S)}{\mathfrak{a}_{k+1}({\rm ord}_S)}.
\]
So we get the identity:
\begin{eqnarray*}
\dim_{\mathbb{C}}\left(\mathcal{O}_{X,o}/\mathfrak{a}_{m}({\rm ord}_S)\right)&=&\sum_{k=0}^{m-1} \dim H^0(S, -kS|_S).
\end{eqnarray*}
Then the result follows easily from the Hirzebruch-Riemann-Roch formula.
As $K_Y+S+\mu_*D\sim_{\mathbb{Q},X} A_{X,D}(S)\cdot S$, the second identity is implied by the first statement.
\end{proof}
\begin{rem}Inspired by the above simple calculation, we indeed can extend the definition of normalized volumes to any model $f:Y\to (X,o)$, such that $f$ is isomorphic over $X\setminus \{o\}$. See Section \ref{s-vmodel}.
\end{rem}
\begin{lem}\label{l-finite}
Let $f\colon (X',o')\to (X,o)$ be a finite morphism, such that $f^*(K_X+D)=K_{X'}+D'$ for some effective $\mathbb{Q}$-divisors. We assume $(X,D)$ and $(X',D')$ are klt. If $S$ is a Koll\'ar component given by $Y\to X$ over $o$, then $Y\times_XX'\to X'$ induces a Koll\'ar component $S'$ over $o'\in (X',D')$.
Conversely, if $X'\to X$ is Galois with Galois group $G$, then any $G$-invariant Koll\'ar component $S'$ over $o\in (X',D')$ is the pull back from a Koll\'ar component over $o\in (X,D)$.
\end{lem}
\begin{proof} The first part is standard. In fact, denote by $S'=f^{-1}(S)$, then $(Y',\mu'^{-1}_*D'+S')$ is log canonical, such that if we restrict to $T$ a component of $S'$,
$$(K_{Y'}+\mu'^{-1}_*D'+S')|_{T}=K_{T}+\Delta_{T},$$
then $(T,\Delta_{T})$ is klt, which by Koll\'ar-Shokurov connectedness theorem implies that $T=S'$.
For the converse, let $$L\sim_{X'}-m(K_{Y'}+\mu'^{-1}_*D'+E')$$ be a divisor of general position for sufficiently divisible $m$ and $H:=\frac{1}{m}L$, then $(Y', E'+\mu'^{-1}_*D'+H)$ is plt. Replacing $H$ by $H_G:=\frac{1}{|G|}(\sum_{g\in G} g^*H)$, we know that $(X', D'+ \mu_*H_{G})$ is $G$-invariant, and there exists a $\mathbb{Q}$-divisor $H_X\ge 0$, such that
$$f^*(K_X+D +H_X)=K_{X'}+D'+\mu_*H_G.$$
Therefore, $(X,D+H_X)$ is plt, and its unique log canonical place is a divisor $S$ which is a Koll\'ar component over $o\in (X,D)$ whose pull back gives the Koll\'ar component $S'$ over $o'\in (X',D')$.
\end{proof}
We prove a change of volume formula for Koll\'ar components under a finite map.
\begin{lem}\label{l-finitevolume}With the same notation as in Lemma \ref{l-finite}, then
$$d\cdot {\widehat{\rm vol}}_{X,D}({\rm ord}_S)={\widehat{\rm vol}}_{X',D'}({\rm ord}_{S'}),$$
where $d$ is the degree of $X'\to X$.
\end{lem}
\begin{proof}Since the pull back of $S$ is $S'$ which is irreducible by Lemma \ref{l-finite}, let the degree of $S'\to S$ be $a$ and the ramified degree be $r$, we have
$$ar=d \qquad \mbox{and}\qquad rA_{X,D}({\rm ord}_S)=A_{X',D'}({\rm ord}_{S'})$$
(see \cite[5.20]{KM98}).
By Lemma \ref{l-inter}, we know that
\begin{eqnarray*}
d\cdot {\widehat{\rm vol}}_{X,D}({\rm ord}_S)& = & ar\cdot A_{X,D}({\rm ord}_S)\cdot ((K_Y+S+\mu_*^{-1}D)|_{S})^{n-1}\\
&=&(rA_{X,D}({\rm ord}_S))\cdot \big( (a(K_Y+S+\mu_*^{-1}D)|_{S})^{n-1}\big)\\
&= &A_{X',D'}({\rm ord}_{S'})\cdot \big( ((K_{Y'}+S'+\mu_*'^{-1}D')|_{S'})^{n-1}\big),\\
&=&{\widehat{\rm vol}}_{X',D'}({\rm ord}_{S'}),
\end{eqnarray*}
where for the third equality we use the projection formula of intersection numbers.
\end{proof}
\subsection{Deformation to cone}\label{ss-deformation}
Let $(X, o)=({\rm Spec}(R), \mathfrak{m})$ be an algebraic singularity such that $(X,D)$ is klt for a $\mathbb{Q}$-divisor $D\ge 0$.
Let $S$ be a Koll\'{a}r component and $\Delta=\Delta_S$ be the different divisor defined by adjunction $(K_Y+S+\mu_*^{-1}D)|_S=K_S+\Delta_S$ where $Y\rightarrow X$ is the extraction of $S$.
Denote $v_0:={\rm ord}_S$ and
$$T=\bigoplus_{k=0}^{+\infty} \mathfrak{a}_k(v_0)/\mathfrak{a}_{k+1}(v_0)=\bigoplus_{k=0}^{+\infty} T_k$$ and the $d$-th truncation
$$T^{(d)}=\bigoplus_{k=0}^{+\infty} \mathfrak{a}_{dk}(v_0)/\mathfrak{a}_{dk+1}(v_0)=\bigoplus_{k=0}^{+\infty} T_{dk}\qquad \mbox{for $d\in \mathbb{N}$}.$$
Now we give a more geometric description of $\mathrm{Spec}(T)$ and $\mathrm{Spec}(T^{(d)})$ using the idea of degenerating $o\in (X,D)$ to an (orbifold) cone over the Koll\'ar component $S$. Assume $\mu\colon Y\rightarrow X$ is the extraction of the Koll\'{a}r component $S$ of $(X, o)$. Then
$\mu_{\mathbb{A}^1}\colon Y\times{\mathbb{A}}^1\rightarrow X\times {\mathbb{A}}^1$ has the exceptional divisor $S\times {\mathbb{A}}^1$. The divisor $S$ is not necessarily Cartier, but only $\mathbb{Q}$-Cartier. Thus we can take the index 1 covering Deligne-Mumford stack $\pi: \mathcal{Y}\to Y$ for $S$. So $\pi$ is isomorphic over $Y\setminus S$ and $\pi^*(S)=\mathcal{S}$ is Cartier on ${\mathcal{Y}}$.
We consider the deformation to the normal cone construction for ${\mathcal{S}}\subset {\mathcal{Y}}$. More precisely, we consider the blow up $\pi_1: {\mathcal{Z}} \rightarrow {\mathcal{Y}}\times {\mathbb{A}}^1$ along ${\mathcal{S}}\times\{0\}$. Denote by $E$ the exceptional divisor of $\pi_1$ and by $\mathcal{T}$ the strict
transform of ${\mathcal{S}}\times{\mathbb{A}}^1$. We note that $E$ has a stacky structure along the 0 and $\infty$ section, but a scheme structure at other places.
Then $\mathcal{T}\subset {\mathcal{Z}}$ is a Cartier divisor which is proper over ${\mathbb{A}}^1$ and can be contracted to a normal Deligne-Mumford stack $\Psi_1\colon {\mathcal{Z}}\to {\mathcal{W}}$ and in this way we get a flat family ${\mathcal{W}}\rightarrow {\mathbb{A}}^1$ such that ${\mathcal{W}}_t\cong X$ and ${\mathcal{W}}_0\cong \bar{{\mathcal{C}}}\cup \mathcal{Y}_0$, where $\mathcal{Y}_0$ is the birational transform of $Y\times\{0\}$. To understand ${\mathcal{C}}$ if we denote by ${\mathcal{Z}}^0:={\mathcal{Z}}\setminus {\mathcal{Y}}_0$, then the fiber ${\mathcal{Z}}^0$ over $0$ is isomorphic to ${\mathcal{C}}$ which is an affine orbifold cone over ${\mathcal{S}}$ with the polarization given by $\mathcal{O}_{{\mathcal{Y}}}(-{\mathcal{S}})|_{{\mathcal{S}}}$.
Moreover, $\bar{{\mathcal{C}}}$ is the projective orbifold cone completing ${\mathcal{C}}$.
Let $d$ be a positive integer such that $d\cdot S$ is Cartier in $Y$, then ${\mathcal{C}}^{(d)}$ given by the cone over $\mathcal{O}_{{\mathcal{Y}}}(-d\cdot{\mathcal{S}})|_{{\mathcal{S}}}$ is a degree $d$ cyclic quotient of ${\mathcal{C}}$, which is a usual ($\mathbb{A}^1$-)cone over ${\mathcal{S}}$.
We denote by $C$ and $C^{(d)}$ the underlying coarse moduli space of ${\mathcal{C}}$ and ${\mathcal{C}}^{(d)}$. We also denote $S$ to be the coarse moduli space of ${\mathcal{S}}$.
Applying the exact sequence,
$$0\to \mathcal{O}_Y(-(k+1)S)\to \mathcal{O}_Y(-kS)\to \mathcal{O}_{S}(-kS)\to 0,$$
since $h^1(\mathcal{O}_Y(-(k+1) S))=0$ by the Grauert-Riemenschneider vanishing theorem, we get:
\[
H^0(S, \mathcal{O}(-k S|_S))\cong H^0(\mathcal{O}_Y(-kS))/H^0(\mathcal{O}_Y(-(k+1)S)).
\]
Notice that the right hand side is equal to:
\[
\frac{\mu_*\mathcal{O}_Y(-kS)}{\mu_*\mathcal{O}_Y(-(k+1)S)}=\frac{\mathfrak{a}_{k}(v_0)}{\mathfrak{a}_{k+1}(v_0)}.
\]
In particular, $C={\rm Spec} (T)$. Similarly, $C^{(d)}={\rm Spec}(T^{(d)})$.
There is also a degree $d$ cyclic quotient morphism $h\colon C\to C^{(d)}$, and we know that
$$h^*(K_{\bar{C}^{(d)}}+C^{(d)}_\Delta+C^{(d)}_{D})=K_{\bar{C}}+C_{D},$$
where $C_D$ is the intersection of $\bar{C}$ with the birational transform of $D\times \mathbb{A}^1$ and $C^{(d)}_{\Delta}$ (resp. $C^{(d)}_{D}$) on $\bar{C}^{(d)}$ is the induced cone over the branched $\mathbb{Q}$-divisor on $S$ of ${\mathcal{S}}\to S$ (resp. $\mu^{-1}_*D|_S$). In particular, $C^{(d)}_{\Delta}+C^{(d)}_{D}$ is the cone over $\Delta_S$.
\subsection{Filtrations and valuations}\label{sec-filtration}
Assume ${\mathcal{F}}$ is a $\mathbb{Z}$-graded filtration on $R$. We have the Rees algebra and extended Rees algebra:
\begin{equation}
{\mathcal{R}}:={\mathcal{R}}({\mathcal{F}})=\bigoplus_{k=0}^{+\infty} ({\mathcal{F}}^k R) t^{-k}, \quad {\mathcal{R}}':={\mathcal{R}}'({\mathcal{F}})=\bigoplus_{k=-\infty}^{+\infty} ({\mathcal{F}}^k R) t^{-k},
\end{equation}
and the associated graded ring:
\begin{equation}
{\rm gr}_{\mathcal{F}}(R)={\mathcal{R}}'/t {\mathcal{R}}'=\bigoplus_{k=0}^{+\infty} ({\mathcal{F}}^k R/{\mathcal{F}}^{k+1} R) t^{-k}=: \bigoplus_{k=0}^{+\infty} A_k.
\end{equation}
Assuming ${\mathcal{R}}'$ is finitely generated, ${\mathcal{X}}:={\rm Spec}_{\mathbb{C}[t]} ({\mathcal{R}}')$ can be seen as a $\mathbb{C}^*$-equivariant flat degeneration of $X$ into
${\mathcal{X}}_0={\rm Spec}_{\mathbb{C}} ({\mathcal{R}}'/t{\mathcal{R}}')={\rm Spec}_{\mathbb{C}}({\rm gr}_{\mathcal{F}} R)$. Denote $E={\rm Proj} ({\rm gr}_{\mathcal{F}}(R))$, $\tilde{X}={\rm Proj}_{R} {\mathcal{R}}$. Then the natural map $\tilde{X}\rightarrow X$ is the filtered blow up associated with the ${\mathcal{F}}$ such that $E$ is the exceptional divisor. Moreover $\tilde{X}$ can be seen as a flat deformation of a natural filtered blow up on ${\mathcal{X}}_0$. Indeed following \cite[5.15]{TW89}, we have a filtration ${\mathcal{F}}$ on ${\mathcal{R}}'$:
\[
{\mathcal{F}}^m {\mathcal{R}}'=\left\{\sum_{k=-\infty}^{+\infty} \left({\mathcal{F}}^{\max(k,m)} R\right) t^{-k} \right\}.
\]
The objects associated to the corresponding Rees algebra and graded algebra over ${\mathcal{R}}'$ are:
\[
\tilde{{\mathcal{X}}}={\rm Proj}_{{\mathcal{R}}'} \bigoplus_{r=0}^{+\infty}({\mathcal{F}}^r {\mathcal{R}}' ) T^{-r}, \quad \mathcal{E}={\rm Proj}_{\mathbb{C}} \bigoplus_{r=0}^{+\infty} ({\mathcal{F}}^r {\mathcal{R}}' /{\mathcal{F}}^{r+1} {\mathcal{R}}') T^{-r}.
\]
Then we have the following commutative diagram (see \cite[Proposition 5.17]{TW89}):
\[
\begin{CD}
&& E @>>> {\mathcal{E}} @<<< E \\
&& @VVV @VVV @VVV \\
\tilde{\mathbb{C}}^p@<<< \tilde{X} @>>> \tilde{{\mathcal{X}}} @<<< \tilde{{\mathcal{X}}}_0\\
@VVV @VVV @VVV @VVV\\
\mathbb{C}^p @<<< X@>>> {\mathcal{X}} @<<< {\mathcal{X}}_0
\end{CD}
\]
There is a natural $\mathbb{C}^*$-action on ${\mathcal{X}}_0$ associated to the natural $\mathbb{N}$-grading such that the quotient is isomorphic to $E$. Let $\mathcal{J}=\bigoplus_{k\ge 0} {\mathcal{F}}^{k+1}t^{-k}=t {\mathcal{R}}'\cap {\mathcal{R}}$ so that ${\mathcal{R}}/\mathcal{J}\cong {\rm gr}_{{\mathcal{F}}}(R)\cong {\mathcal{R}}'/t {\mathcal{R}}'$.
Now we assume furthermore that
$E$ is a normal projective variety. This implies both ${\mathcal{R}}$ and ${\mathcal{R}}'$ are normal (see \cite{TW89}).
Let $\mathfrak{P}$ be the unique minimal prime ideal of ${\mathcal{R}}$ over $\mathcal{J}$ that corresponds to $E$, and $w$ the valuation of $K(t)$ attached to $\mathfrak{P}$. Then the restriction of $w$ to $R$ is equal to $b \cdot {\rm ord}_E$. Assume $a=w(t)$. Thus the filtration ${\mathcal{F}}$ is equivalent to the filtration that is given by:
\[
(t^m {\mathcal{R}}') \cap R=\{f\in R; {\rm ord}_E(f)\ge m a/b\}.
\]
\begin{rem}
There is a general Valuation Theorem about the relation between finitely generated filtrations and valuations proved by Rees for which we refer the reader to \cite{Ree88}. See also \cite{BHJ15}.
\end{rem}
\section{Volume of models}\label{s-vmodel}
One very useful tool for us to study the minimzer of the normalized local volume is the concept of a local volume of a model. It is this concept which enables us to apply the machinery of the minimal model program to construct different models, especially those yielding Koll\'ar components.
\subsection{Local volume of models}\label{s-lvmodel}
In this section, we extend the definition of volume to models in the `normalized' sense. We use the concept of local volumes as in \cite{ELS03, Ful13}. Let us first recall the definition, which is from \cite{Ful13}.
\begin{defn}[Local volume](cf. See \cite{Ful13})
Let $X$ be a normal algebraic variety of dimension at least two and let $o$ be a point on $X$. Fixing a proper birational map $\mu\colon Y \to X$, for an arbitrary Cartier divisor $D$ on $Y$, we define the local volume of $D$ at $x$ to be
$${\rm vol}^F_o(D) = \limsup_{m\to \infty} \frac{h^1_o(mD)} {m^n /n!},$$
where $h^1_o(mD):=\dim H^1_{\{o\}}(X, f_*\mathcal{O}_Y(mD)).$
\end{defn}
We can define the volume of a $\mathbb{Q}$-Cartier divisor $D$ to be
$${\rm vol}^F_{o}(D):=\frac{{\rm vol}^F_o(mD)}{m^n},$$
for sufficiently divisible $m$.
\begin{lem}\label{l-pushforward}
Let $\mu\colon Y\to X$ be a birational morphism. If $D\ge 0$ is an exceptional $\mathbb{Q}$-divisor, such that ${\rm Supp}(D)\subset \mu^{-1}(o)$.
Then
$${\rm vol}^F_{o}(-D)=\limsup_{k\to \infty} \frac{l_R(\mathcal{O}_X/\mathfrak{a}_k)}{k^n/n!},$$
where $k$ is sufficiently divisible and $\mathfrak{a}_k=\mu_*(\mathcal{O}_Y(-kD))$.
\end{lem}
\begin{proof}This follows from \cite[Remark 1.1(ii)]{Ful13} (see also \cite[Remark 1.31 and 1.32]{Ful13}).
\end{proof}
The right hand side of the above display is also the volume ${\rm vol}(\mathfrak{a}_{\bullet})$ defined in \cite[Definition 3.1, Proposition 3.11]{ELS03}. In particular, given a prime divisor $D$ over $o$ with log discrepancy $a$, then we see that
$${\rm vol}^{F}_o(-aD)={\widehat{\rm vol}}_{X,o}({\rm ord}_D).$$
\begin{defn}\label{d-modelv}Let $o\in (X,D)$ be a klt singularity.
Let $\mu\colon Y\to (X,o)$ be a birational morphism, such that $\mu$ is an isomorphism over $X\setminus \{o\}$. Let $E=\sum_i G_i$ be the divisorial part of $\mu^{-1}(o)$. Then we define the volume ${\rm vol}_{(X,D),o}(Y)$ (abbreviated as ${\rm vol}_{X,o}(Y)$ or ${\rm vol}(Y)$ if $(X,D; o)$ is clear) of $Y$ to be
$${\rm vol}_{(X,D),o}(Y):={\rm vol}^F_{o}(-K_Y-E-\mu_*^{-1}(D))={\rm vol}^F_{o}(\sum_i -a_iG_i),$$
where $a_i=A_{X,D}(G_i)$ is the log discrepancy.
\end{defn}
We mainly combine the above definition with the following construction.
\begin{defn}\label{d-dlt}
For a klt pair $(X,D)$ with an idea $\mathfrak{a}$, we denote by $c$ its log canonical threshold ${\rm lct}(X,D;\mathfrak{a})$. We say $\mu\colon Y\to X$ is a {\it dlt modification} of $(X,D+c\cdot \mathfrak{a})$, if
\begin{enumerate}
\item denote the divisorial part of $\mu^*(\mathfrak{a})$ by $\mathcal{O}(-\sum m_iG_i)$ and denote by $\mu^*(K_X+D)=K_Y+D_Y$, then
$$D_Y+c\cdot \sum m_i G_i=\mu^{-1}_*(D)+E$$ where $E$ is the reduced divisor on ${\rm Ex}(\mu)$;
\item $(Y,D_Y+c\cdot \sum m_i G_i)$ is dlt; and
\item for any divisor $F$,
$$a(F; Y,D_Y+c\cdot \sum m_i G_i)=-1\mbox{ if and only if }a(F; X,D+c\cdot\mathfrak{a})=-1.$$
\end{enumerate}
By the argument in \cite{OX12}, we know that it follows the MMP result in \cite{BCHM10} that the dlt modification of $(X,D+c\cdot \mathfrak{a})$ always exists. More precisely, we can choose general elements $f_j\in\mathfrak{a}$ $(1\le j \le l)$ which generate $\mathfrak{a}$ such that $l> \frac{1}{c}$. Let $D_j=(f_j)$, then $Y$ is the dlt modification of $(X,D+c\cdot\frac{1}{l}\sum^l_{j=1}D_j)$.
\end{defn}
We do not need property 3 in our argument, but it seems to us it is natural to require this by comparing with \cite[Theorem 1.34]{Kol13}.
\begin{lem}\label{l-gooddlt}
We can indeed assume that $-K_Y-{\mu}_*^{-1}D-E$ is nef over $X$.
\end{lem}
\begin{proof}Since $(X,D)$ is klt, we know that
$$K_Y+\mu_*^{-1}D+E\sim_{\mathbb{Q},X} \sum a_iG_i$$
with $a_i=A_{X,D}(G_i)>0$. Running an MMP with scaling by an ample divisor, we obtain a relative minimal model $Y\dasharrow Y'$ of
$$
K_Y+\mu^{-1}_*(D+c\cdot \frac{1}{l}\sum_{j=1}^l D_j)+E \sim_{{\mathbb{Q}}, X} - c\cdot \sum m_i G_i+\sum_i a_i G_i=0.
$$
So we have
$$
K_Y+\mu^{-1}_*(D+c\cdot \frac{1}{l}\sum_{j=1}^l D_j)+E-\epsilon \sum a_i G_i=-\epsilon (K_Y+\mu_*^{-1}D+E),
$$
and hence $-K_{Y'}-{\mu'}_*^{-1}D-E'$ is nef over $X$. Furthermore, since
$$K_Y+\mu_*^{-1}D+E\sim_{\mathbb{Q},X}-c \cdot\frac{1}{l} \mu_*^{-1} \sum^l_{j=1}D_j,$$
$Y'$ also gives a minimal model of the dlt pair $$\big(Y, \mu_*^{-1}(D+c(1+\epsilon) \cdot\frac{1}{l}\sum^l_{j=1}D_j)+E\big),$$
which implies $(Y', {\mu'}_*^{-1}D+E')$ is a dlt modification of $(X,D+c\cdot\frac{1}{l}\sum^l_{j=1}D_j)$. Therefore, we can replace $ Y$ by $Y'$.
\end{proof}
When $E$ is irreducible, then ${\rm vol}_{X,o}(Y)={\widehat{\rm vol}}_{X,o}({\rm ord}_E)$. We can also generalize Lemma \ref{l-inter} to the dlt case.
\begin{lem}\label{l-inter2}In the setting of Definition \ref{d-modelv}, if we assume that $-K_Y-\mu_*^{-1}D-E$ is nef over $X$.
Then
$${\rm vol}_{X,o}(Y)=\sum_i a_i \big((-K_Y-\mu_*^{-1}D-E)|_{E_i}\big)^{n-1}.$$
\end{lem}
\begin{proof}Let $m$ be sufficiently divisible such that $L:=m(K_Y+\mu_*^{-1}D+E)$ is Cartier. Denote by $F$ the effective Cartier divisor $F:=\sum_i ma_iG_i$.
Then
$$0\to \mathcal{O}_Y(L^{\otimes -(k+1)})\to \mathcal{O}_Y(L^{\otimes -k})\to \mathcal{O}_F(L^{\otimes -k})\to 0.$$
Since $L^{-1}$ is nef, we know that $R^1\mu_*(L^{\otimes -(k+1)})=0$.
Thus
$${\rm vol}_o^F(L)={\rm vol}(L|_{F}),$$
then we conclude by dividing $m^n$ in both sides.
\end{proof}
\begin{lem}\label{l-model1}
Let $\mathfrak{a}$ be an $\mathfrak{m}$-primary ideal, we define $c={\rm lct}(X,D;\mathfrak{a})$ and $(Y,E)\to X$ be the dlt modification of $ (X,D+c\cdot \mathfrak{a})$. Then
$${\rm vol}_{X,o }(Y)\le {\rm lct}^n(\mathfrak{a} )\cdot{\rm mult}(\mathfrak{a} ).$$
\end{lem}
\begin{proof}Write $K_Y+\mu_*^{-1}D+E=\mu^*(K_X+D)+\sum_i a_iG_i,$ where $E$ is the reduced divisor on ${\rm Ex}(\mu)$. If we denote the vanishing order of $\mu^* \mathfrak{a}$ along $G_i$ by $m_i$, then since $c$ is the log canonical threshold and for every $i$, $G_i$ computes the log canonical threshold, we know that $c\cdot m_i=a_i$. Thus
$$\mathfrak{a}^k\subset \mu_*\mathcal{O}_Y(-\sum_i km_iG_i)=_{\rm def} {\mathfrak{b}}_k.$$
It suffices to show that
$${\rm mult}({\mathfrak{b}}_{\bullet})={\rm vol}_o^F (-\sum m_iG_i),$$
but this follows Lemma \ref{l-pushforward}.
\end{proof}
\begin{lem}\label{l-kollar}
Notations as above. Then there exists a Koll\'ar component $S$, such that
$${\widehat{\rm vol}}({\rm ord}_S)\le {\rm vol}(Y).$$
\end{lem}
\begin{proof}It follows from Proposition \ref{p-special} that we can choose a model $W\to Y$ and run MMP to obtain $W\dasharrow Y'$, such that $Y'\to X$ gives a Koll\'ar component $S$ with $a(Y,E; S)=-1$.
If we fix a common resolution $p\colon W'\to Y $ and $q\colon W'\to Y'$, then since $-(K_Y+E)$ is nef and $A_{Y, E}(S)=0$,
we know $-p^*(K_Y+E)+q^*(K_{Y'}+S)$ is $q$-nef and $q$-exceptional. By the negativity lemma, we get
$$p^*(K_Y+E)\ge q^*(K_{Y'}+S).$$
Thus
$${\widehat{\rm vol}}({\rm ord}_S)={\rm vol}(-K_{Y'}-S)\le {\rm vol}(-K_Y-E)={\rm vol}(Y).$$
\end{proof}
\subsection{Approximating by Koll\'ar components}
With the above discussions, we can start to prove our theorems.
\begin{proof}[Proof of Theorem \ref{t-approx}]
By Proposition \ref{p-inf}, we know
$$\inf_{v} {\widehat{\rm vol}}_{X,o}(v)=\inf_{\mathfrak{a}}{\rm lct}^n(\mathfrak{a})\cdot{\rm mult}(\mathfrak{a}).$$
By the above construction in Lemma \ref{l-model1} and \ref{l-kollar}, for any $\mathfrak{m}$-primary ideal $\mathfrak{a}$, we know that there exists a Koll\'ar component $S$,
such that
$${\widehat{\rm vol}}({\rm ord}_S)\le {\rm lct}^n(\mathfrak{a})\cdot{\rm mult}(\mathfrak{a}). $$ This finishes the proof of the first part of Theorem \ref{t-approx}.
\bigskip
We continue to prove the second part of the theorem. Let $\{\mathfrak{a}_k\}_{k\in \Phi}$ be the associated graded family of valuation ideals induced by $v$ where $\Phi\in \mathbb{R}$ is the value semigroup. For each $\mathfrak{a}_k$ ($k\in \Phi$), we denote by
$$c_k:={\rm lct}(X,D; \mathfrak{a}_k).$$
Let $\mu_k\colon Y_k\to X$ be a dlt modification of $(X,D; c_k\cdot \mathfrak{a}_k)$ and $E_k$ the exceptional divisor of $Y_k$ over $X$. Assume the model we obtain from Lemma \ref{l-kollar} is $Y'_k$ with the Koll\'ar component $S_k$.
We consider the valuation
$$v_k=\frac{c_k\cdot k}{A_{X,D}(S)}{\rm ord}_{S_k}.$$
By \cite{JM12}, we know
$$c_k\cdot k={\rm lct} (X,D; \frac{1}{k}\mathfrak{a}_k)=\inf_{v'}\frac{A_{X,D}(v')}{\frac{1}{k}v'(\mathfrak{a}^k)}\le A_{X,D}(v) <\infty.$$
Then by the Izumi type estimate in \cite[Proposition 1.2]{Li15a}, we know that
$$v_k(\mathfrak{m}){\rm ord}_{o}\le v_k \le c \cdot {\rm ord}_{o},$$
for some positive constant $c$ and all $k$. And by \cite[Theorem 2.3]{Li15a} and the fact that ${\widehat{\rm vol}}(v_k)$ is bounded from above, we know that $v_k(\mathfrak{m})$ is bounded from below. In particular, by the compactness result \cite[Proposition 5.9]{JM12} and Proposition \ref{p-sequcom}, we know that there is an infinite sequence $\{v_{k_j}\}_{k_j\in \Phi}$ with $k_j\to +\infty$ which has a limit in ${\rm Val}_{X,o}$.
Denote by
$$v'=\lim_{i\to \infty} v_{k_i},$$ then we know that
$$A_{X,D}(v')\le \liminf_{i\to \infty} A_{X,D}(v_{k_i})=\lim_{i\to \infty} c_{k_i}\cdot {k_i} \le A_{X,D}(v)$$ as $A_{X,D}$ is lower semicontinuous (see \cite[Lemma 5.7]{JM12}). We claim for any $f$, we have
$$v'(f)\ge v(f),$$
and this clearly implies that ${\rm vol}(v')\le {\rm vol}(v)$, which then implies ${\widehat{\rm vol}}(v')\le {\widehat{\rm vol}}(v)$.
Denote by $v(f)= p$. For a fixed $k_j$, choose $l$ such that
$$(l-1)p< k_j\le lp .$$ Let $k=k_j$ in the previous construction, then
\begin{eqnarray*}
v(f)= p&\Longrightarrow & v(f^l)= pl,\\
&\Longrightarrow & f^l\in \mathfrak{a}_{pl},\\
&\Longrightarrow & f^l\in \mathfrak{a}_{k_j},\\
&\Longrightarrow &l\cdot {\rm ord}_{E_i}(f)\ge m_{k_j,i} \mbox{\ \ for any $i$},\\
&\Longrightarrow& l\cdot {\rm ord}_{S_{k_j}}(f)\ge A_{X,D}(S_{k_j})\cdot \frac{1}{c_{k_j}},\\
&\Longrightarrow&v_{k_j}(f)\ge \frac{k_j}{l}>p-\frac{p}{l}.
\end{eqnarray*}
The fourth arrow is because if $f^l\in \mathfrak{a}_{k_j}$, then $f^l$ vanishes along $m_{k_j,i}E_i$; and
the fifth arrow is because that
$$K_{Y_{k_j}}+Y_{k_j}+\mu_{k_j*}^{-1}D+E_{k_j}\sim_{\mathbb{Q},X} c_{k_j}\cdot \sum m_{k_j,i}E_i,$$
and the pull back of $K_{Y_{k_j}}+Y_{k_j}+\mu_{k_j*}^{-1}D+E_{k_j}$ is larger than the one from
$$K_{Y'_{k_j}}+\mu_{k_j*}^{'-1}D+S_{k_j}\sim_{\mathbb{Q},X} A_{X,D}(S_{k_j})S_{k_j}.$$
Thus $v'(f)=\lim v_{k_j}(f)\ge p= v(f)$.
\end{proof}
\begin{prop}\label{p-sequcom} Let $o\in (X,D)$ be a klt singularity. Let $a$ and $b$ be two positive numbers. Then the subset $K_{a,b}$ of ${\rm Val}_{x,X}$ which consists of all valuations with
$$a\le v(\mathfrak{m}) \qquad \mbox{and} \qquad A_{X,D}(v)\le b$$
is sequential compact.
\end{prop}
\begin{proof} Let $\{v_i\}$ be a sequence contained in $K$. Let $\{\mathfrak{a}_{i,k}\}$ be its associated graded valuative ideals for $k\in \Phi_i$. We can find a countably generated field $F\subset \mathbb{C}$, such that $R={\rm Spec}(R_F)\times_F \mathbb{C}$ for some finitely generated $F$-algebra $R_F$ and $D$, $x$ are defined over $F$. Furthermore, we can assume for each pair $(i,k)$, ${\mathfrak{a}_{i,k}}=({\mathfrak{a}_{i,k}})_F\times_F {\mathbb{C}}$, for some ideal $({\mathfrak{a}_{i,k}})_F\subset R_F$. Denote by $X_F:={\rm Spec}(R_F)$ and $D_F$ the divisor of $D$ descending on $X_F$.
Now let $(v_i)_F$ be the restriction of $v_i$ on $R_F$. By our definition, we know that
$$\mathfrak{a}_{i,k}=\{f\in R_F \ | (v_i)_F(f)\ge k\},$$
and $(v_i)_F\in (K_{a,b})_F$ where $(K_{a,b})_F$ is defined for all $v\in {\rm Val}_{x,X_F}$ with $a\le v(\mathfrak{m}_F)$ and $A_{X_F,D_F}(v)\le b$. By \cite[Theorem 1.1]{HLP12}, ${\rm Val}_{x,X_F}$ has the same topology as a set of some Euclidean space, thus $(K_{a,b})_F$ is sequential compact as it is compact by \cite[Proposition 5.9]{JM12}. Therefore after passing through a subsequence, $(v_i)_F$ has a limit $(v_{\infty})_F$, which can be extended to a valuation $v_{\infty}:=(v_{\infty})_F\otimes \mathbb{C}$. In fact, $v_{\infty}$ is defined as follows: for any $f\in R$ it can be written $f=\sum^m_{j=1}f_j\otimes_F h_j$ such that $0\neq f_j\in R$ and $h_1,...,h_m\in \mathbb{C}$ are linearly independent over $F$, then
$$v_{\infty}(f)=\min^m_{j=1} (v_{\infty})_F(f_j).$$ We claim $v_i=(v_i|_{R_F})\otimes_F\mathbb{C}$. In fact, for any $f$, if $v_i(f)=k$, then $f\in \mathfrak{a}_{i,k}=(\mathfrak{a}_{i,k})_F\otimes_F \mathbb{C}$, thus $(v_i|_{R_F})\otimes_F\mathbb{C}(f)=k$.
To see that for any $f$, $v_{\infty}(f)=\lim v_i(f)$, we know for some $j$,
$$v_{\infty}(f)=(v_{\infty})_F(f_j)=\lim_{i}(v_i|_{R_F})(f_j)\ge \limsup_i v_i(f). $$
For another direction, if we have a subsequence of $i$, such that $\lim_i v_i(f)<v_{\infty}(f)$, after passing to a subsequence again, we can find a $j$, such that $$\lim_iv_i(f)=\lim_iv_i(f_j)=\lim_i (v_{\infty})_F(f_j)\ge v_{\infty}(f), $$
a contradiction.
\end{proof}
As Example \ref{ex-irreg} shows that for a general klt singularity $(X,o)$, the minimum is not always achieved by a Koll\'ar component. Thus we have to take a limit process. However, if the minimizer $v$ has rational rank 1, then it should always yield a Koll\'ar component. First we have the following result which is inspired by the work in \cite{Blu16a}.
We note that it is also independently obtained in \cite{Blu16b}.
\begin{lem}\label{l-lcplace}
If $E\in {\rm Val}_{X,o}$ such that ${\rm ord}_E$ minimizes ${\widehat{\rm vol}}(X,D)$. Then the Rees algebra associated to ${\rm ord}_E$ is finitely generated.
\end{lem}
\begin{proof}If we let $\{\mathfrak{a}_{\bullet}\}$ be the graded valuative ideas associated to ${\rm ord}_E$, then we know that
\begin{eqnarray*}
{\widehat{\rm vol}}({\rm ord}_{E})&= &\lim_{k\to \infty}A_{X,D}({\rm ord}_E)^n \cdot \frac{{\rm mult}(\mathfrak{a}_k)}{k^n}\\
&\ge& \lim_{k\to \infty}{\rm lct} (X,D;\mathfrak{a}_k)^n \cdot {\rm mult}(\mathfrak{a}_k)\\
& \ge&{\widehat{\rm vol}}({\rm ord}_{E})
\end{eqnarray*}
by Proposition \ref{p-inf} and our assumption that ${\rm ord}_E$ is a minimizer of ${\widehat{\rm vol}}(X,D)$. So we conclude that (see \cite{Mus02})
$${\rm lct}(X,D;\mathfrak{a}_{\bullet}):=\lim_{k\to \infty} k\cdot {\rm lct} (X,D;\mathfrak{a}_k)=A_{X,D}({\rm ord}_E),$$
which we denote by $c$. Therefore, we can choose $\epsilon$ sufficiently small, such that the discrepancy $a(E;X,D+(1-\epsilon)c\cdot \mathfrak{a}_{\bullet})\in (-1,0),$
On the other hand, we know
$${\rm lct}(X,D;\mathfrak{a}_{\bullet})=\lim_{m\to \infty} m\cdot {\rm lct}(X,D;\mathfrak{a}_{m}).$$ So for sufficiently large $m$, we know that for all $G$, the discrepancy
$$a(G;X, D+\frac{1}{m}(1-\epsilon)c\cdot\mathfrak{a}_m)>-1.$$
We also have $a(E;X, D+\frac{1}{m}(1-\epsilon)c\cdot\mathfrak{a}_m)<0$.
Then similar to the discussion in \ref{d-dlt}, we can find general $\mathbb{Q}$-divisor $\Delta$, such that
$(X, D+\frac{1}{m}(1-\epsilon)c \cdot \Delta)$ is klt and $a(E;X, D+\frac{1}{m}(1-\epsilon)c\cdot \Delta)<0$. In particular, we can apply \cite{BCHM10} to obtain a model $\mu\colon Y\to X$ such that ${\rm Ex}(\mu)=E$ and $-E$ is $\mu$-ample, which implies the finite generation.
\end{proof}
\begin{proof}[Proof of Theorem \ref{t-divisor}] Applying Lemma \ref{l-lcplace}, the assumption in Case 1 which says $v$ is a divisorial valuation implies the assumption in Case 2, thus we only need to treat the Case 2.
\bigskip
By the proof of Proposition \ref{p-inf}, we know that if we let $\mathfrak{a}_k$ be the ideal of elements with values at least $k$,
then
\begin{eqnarray*}
A_{X,D}(v)^n \cdot \frac{{\rm mult}(\mathfrak{a}_k)}{k^n}&\ge & \left(\frac{A_{X,D}(v)}{v(\mathfrak{a}_k)}\right)^n \cdot {\rm mult}(\mathfrak{a}_k)\ge {\rm lct}^n(X,D; \mathfrak{a}_k)\cdot {\rm mult}(\mathfrak{a}_k).
\end{eqnarray*}
By the finite generation assumption, we know that $\mathfrak{a}_{kl}=\mathfrak{a}^l_{k}$ for sufficiently divisible $k$ and any $l$.
So replace $k$ by $kl$ in the above display and let $l\to +\infty$, we know that
$${\widehat{\rm vol}}_{X,o}(v)\ge {\rm lct}^n(X,D; \mathfrak{a}_k)\cdot {\rm mult}(\mathfrak{a}_k)\ge {\widehat{\rm vol}}_{X,o}(v).$$
Take $\mu\colon Y\to X$ to be the dlt modification of $(X,D+{\rm lct}(X,D, \mathfrak{a}_k)\cdot \mathfrak{a}_k)$ as given in Lemma \ref{l-gooddlt}. The above discussion then implies that
$$ {\rm lct}^n(X,D; \mathfrak{a}_k)\cdot {\rm mult}(\mathfrak{a}_k)\ge {\widehat{\rm vol}}_{X,o}(v)={\rm vol}_{X,o}(Y).$$
It follows from Proposition \ref{p-special}, that we can choose a model $W\to Y$ and running MMP to obtain $W\dasharrow Y'$, such that $\mu'\colon Y'\to X$ gives a Koll\'ar component $S$ with $a(S; Y,\mu^{-1}D_*+E)=-1$. We only need to show that if $Y'$ and $Y$ are not isomorphic in codimension 1, then
$${\rm vol}_{X,o}(Y')<{\rm vol}_{X,o}(Y).$$
This is the the local analog of the argument in \cite[Proposition 5]{LX14}. We give the details for the reader's convenience.
Let $\pi\colon Y\to Y^{\rm c}$ be the canonical model of $-K_Y-{\mu}_*^{-1}D-E$ over $X$, which exists because
$$-\epsilon (K_Y+{\mu}_*^{-1}D+E)\sim_{\mathbb{Q},X}K_Y+{\mu}_*^{-1}(D+c\cdot\frac{1}{l}\sum D_j)+E-\epsilon \sum_i A_{X,D}(G_i)G_i$$
is a klt pair for $\epsilon$ sufficiently small. The assumption that $Y'$ and $Y$ are not isomorphic in codimension 1 implies $Y^{\rm c}\neq Y$.
Take $p\colon \hat{Y}\to Y$ and $q\colon \hat{Y}\to Y'$ a common log resolution, and write
$$p^*(K_Y+{\mu}_*^{-1}D+E)=q^*(K_{Y'}+{\mu'}_*^{-1}D+S)+G.$$
By negativity lemma (cf. \cite[3.39]{KM98}), we conclude that $G\ge 0$. Since
$$K_Y+{\mu}_*^{-1}D+E\sim_{\mathbb{Q},X}\sum_i A_{X,D}(G_i)G_i $$
and
$$K_{Y'}+{\mu'}_*^{-1}D+S\sim_{\mathbb{Q},X} A_{X,D}(S),$$
we know that
$$p^*(\sum_iA_{X,D}(G_i)G_i)=q^*(A_{X,D}(S))+G.$$
For $0\le \lambda \le 1$, let
$$L_{\lambda}=q^*(A_{X,D}(S))+\lambda G=\sum_{i} b_i(\lambda) F_i,$$
where $F_i$ runs over all divisor supports on $\hat{Y}_{o}:=\hat{Y}\times_X\{o\}$, and $-L_{\lambda}|_{\hat{Y}_o}$ is nef.
Define
$$f(\lambda) = \sum_i b_i(\lambda)(-L_{\lambda}|_{F_i})^{n-1},$$
thus $f(\lambda)$ is non-decreasing as $G\ge 0$. By Lemma \ref{l-inter} and \ref{l-inter2}, we know that
$$f(1)={\rm vol}_{X,o}(Y)\qquad\mbox{and}\qquad f(0)={\rm vol}_{X,o}(Y').$$
Since $Y\dasharrow Y'$ are not isomorphic incodimension 1, it must contract some component $G_1$ of $E$, and the coefficient of $G_1$ in $G$ is
$$a:=A_{Y',{\mu'}_*^{-1}D+S}(G_1)>0.
$$
Then
\begin{eqnarray*}
\frac{df (\lambda)}{d \lambda}|_{\lambda=1}&=& n\cdot G\cdot \big(-p^*(K_Y+{\mu}_*^{-1}D+E)\big)^{n-1}\\
&\ge&n \cdot aG_1 \cdot \big(-\pi_*(K_Y+{\mu}_*^{-1}D+E)\big)^{n-1}\\
&>&0.
\end{eqnarray*}
Thus ${\rm vol}_{X,o}(Y')=f(0)<f(1)={\rm vol}_{X,o}(Y)$.
\end{proof}
With all this discussion, we also obtain the following result, which characterizes the equality condition in Proposition \ref{p-inf} and is a corresponding generalization of \cite[Theorem 1.4]{dFEM04} (see Remark \ref{r-sharp}) for smooth point. See \cite[9.6]{LazII} for more background.
\begin{thm}\label{p-equality}
Let $(X,o)=(\mathrm{Spec} (R),\mathfrak{m})$. Assume $(X,D)$ is a klt singularity for a $\mathbb{Q}$-divisor $D\ge 0$.
If there exists an $\mathfrak{m}$-primary ideal ${\mathfrak{b}}$, such that ${\rm lct}^n(X,D;{\mathfrak{b}})\cdot {\rm mult} ({\mathfrak{b}})$ is equal to $\inf_{v\in {\rm Val}_{X,o}}{\widehat{\rm vol}}(v)$, then $(X,D)$ is minimized at a Koll\'ar component.
Moreover,
for an ideal $\mathfrak{a}$,
$${\rm lct}^n(X,D;\mathfrak{a})\cdot {\rm mult} (\mathfrak{a})=\inf_{v\in {\rm Val}_{X,o}}{\widehat{\rm vol}}(v),$$
if and only if there exists a Koll\'ar component $S$ and a positive integer $k$ such that ${\rm ord}_S$ computes the log canonical threshold of $\mathfrak{a}$ on $o\in (X,D)$ with
$${\widehat{\rm vol}}({\rm ord}_S)=\inf_{v\in {\rm Val}_{X,o}}{\widehat{\rm vol}}(v), $$
and the only associated Rees valuation of $\mathfrak{a}^k$ is ${\rm ord}_S$.
\end{thm}
\begin{proof}By the argument in Theorem \ref{t-divisor}, we see that
$${\rm lct}^n(X,D;{\mathfrak{b}})\cdot {\rm mult} ({\mathfrak{b}})$$
reaches the minimum of ${\widehat{\rm vol}}_{X,o}$ if and only if the dlt modification of
$$(X,D; c\cdot {\mathfrak{b}}) \qquad \mbox {where $c={\rm lct}(X,D; {\mathfrak{b}})$} $$
is a model $\mu:Y\to X$ which only extracts a Koll\'ar component $S$ of $(X,D)$ such that ${\rm ord}_S$ is a minimizer of ${\widehat{\rm vol}}_{X,o}$.
Now we fix an ideal $\mathfrak{a}$ such that $${\rm lct}^n(X,D;\mathfrak{a})\cdot {\rm mult} (\mathfrak{a})=\inf_{v\in {\rm Val}_{X,o}}{\widehat{\rm vol}}(v),$$ and $S$ is the Koll\'ar component we obtain as above.
Let $\mu^*\mathfrak{a}$ have the vanishing order $m$ along $S$. Since $S$ is $\mathbb{Q}$-Cartier, then let $k$ be a positive integer such that $mk S$ is Cartier. We claim that
$$ \mu^*(\mathfrak{a}^k)=\mathcal{O}_Y(-mkS).$$
Granted this for now, then we know that $Y$ coincides with the normalized blow up $X^{+}\to X$ of $\mathfrak{a}^k$, i.e., $S$ is the only associated Rees valuation for $\mathfrak{a}^k$.
To see the claim, since $-mkS$ is Cartier, we know that
$$ \mu^*(\mathfrak{a}^k)={\mathfrak{c}}\cdot \mathcal{O}_Y(-mkS)\qquad \mbox{for some ideal ${\mathfrak{c}}\subset \mathcal{O}_Y$},$$
and we aim to show that ${\mathfrak{c}}$ is indeed trivial. If not, we take a normalized blow up $\phi\colon Y^+\to Y$ of ${\mathfrak{c}}$, so $\phi^*{{\mathfrak{c}}}=\mathcal{O}_{Y^+}(-E)$ for some effective Cartier divisor $E$. Since $-S$ is ample over $X$, we can choose $l$ sufficiently big, such that
$$-D:=-\phi^*(mklS)-E$$ on $Y^+$ is ample over $X$.
Since
$$(\mu\circ \phi)^*\mathfrak{a}^{kl}=\mathcal{O}_{Y^+}(-\phi^*(mklS)-lE)\subset \mathcal{O}_{Y^+}(-\phi^*(mklS)-E),$$
we know that
\begin{eqnarray*}
{\rm mult}(\mathfrak{a}^{kl})&\ge & {\rm vol}^{F}_{o}(-\phi^*(mklS)-E)\\
&= & {\rm vol}^{F}_o(-D)\\
&=& mkl (-D|_{\phi^*S})^{n-1}+ (-D|_{E})^{n-1}\\
&>& mkl (mkl(-S)|_{S})^{n-1}\\
&=&(mkl)^n{\rm vol}({\rm ord}_S).
\end{eqnarray*}
Since $ {\rm lct}^n(X,D;\mathfrak{a})=\frac{1}{l}\cdot A_{X,D}({\rm ord}_S)$, we can easily see the above inequality is contradictory to the assumption that
$${\rm lct}^n(X,D;\mathfrak{a})\cdot {\rm mult} (\mathfrak{a})={\widehat{\rm vol}}({\rm ord}_S).$$
Here the inequality in the fourth row comes from a similar but easier calculation as in the proof of of Theorem \ref{t-divisor}.
\bigskip
For the converse direction, since $\mathfrak{a}$ reaches the minimum if and only if $\mathfrak{a}^k$ reaches the minimum, then we can replace $\mathfrak{a}$ by $\mathfrak{a}^k$ and assume its only associated Rees valuation is ${\rm ord}_{S}$, i.e., we know that the normalized blow up $\mu\colon X^{+}\to X$, has the property that $\mu^*(\mathfrak{a})=\mathcal{O}_{X^+}(-mS)$. Then the valuative ideal
$$\mathfrak{a}_{mk}= \overline{\mathfrak{a}^k},$$
where $\overline{\mathfrak{a}^k}$ means the integral closure of $\mathfrak{a}^k$. And ${\rm lct}(X,D;\mathfrak{a})=\frac{A_{X,D}(S)}{m}.$
We claim that
$${\rm mult}(\mathfrak{a})=\lim_{k\to +\infty}\frac{n!\cdot l_R(R/\overline{\mathfrak{a}^{k}})}{k^n} ,$$
and this immediately implies that $${\widehat{\rm vol}}({\rm ord}_S)={\rm lct}(X,D;\mathfrak{a})^n\cdot {\rm mult}(\mathfrak{a}).$$
To see the claim, if we denote by $\mathcal{J}(\mathfrak{a}^k)=\mathcal{J}(X,D;\mathfrak{a}^k)$ the multiplier ideal, then we know that
$${\rm mult}(\mathfrak{a})=\lim_{k\to +\infty}\frac{n!\cdot l_R(R/\mathcal{J}(\mathfrak{a}^k) )}{k^n} ,$$
by the local Skoda Theorem \cite[9.6.39]{LazII}. On the other hand, since $(X,D)$ is klt, we have
$$\overline{\mathfrak{a}^k} \subset \mathcal{J}(\mathfrak{a}^k), $$
thus we are done.
\end{proof}
\begin{rem}\label{r-sharp}
In the proof, we indeed show that for any $k$ such that $mkS$ is Cartier on $Y$, the integral closure $\overline{\mathfrak{a}^k}$ coincides with the valuative ideal $\mathfrak{a}_{mk}$ of ${\rm ord}_{S}$.
\end{rem}
\section{K-semistability implies minimum}\label{s-min}
\subsection{Degeneration to initial ideals}\label{s-degin}
Let $(X, o)=({\rm Spec}(R), \mathfrak{m})$ be an algebraic singularity such that $(X,D)$ is klt for a $\mathbb{Q}$-divisor $D\ge 0$. Denote by $\mathfrak{m}$ the maximal ideal of $o\in X$. Suppose ${\mathfrak{b}}$ is an $\mathfrak{m}$-primary ideal on $X$. We consider its flat degeneration $W/{\mathbb{A}}^1$ where $W$ is the underlying coarse moduli space of ${\mathcal{W}}$ defined in Section \ref{ss-deformation}. We keep the notation in Section \ref{ss-deformation} and denote by $v_0$ the valuation ${\rm ord}_S$.
Next we will describe explicitly a way of obtaining an ideal $\mathfrak{B}$ on ${\mathcal{X}}$ such that $\mathfrak{B}\otimes \mathcal{O}_{X\times\mathbb{C}^*}={\mathfrak{b}}$ and $\mathfrak{B}\otimes \mathcal{O}_C\cong {\bf in}({\mathfrak{b}})$ by considering the closure of ${\mathfrak{b}}\times \mathbb{C}^*$ on the pull back ${\mathcal{Z}}$.
For this we consider the extended Rees algebra (see \cite[6.5]{Eis94}):
\[
\mathcal{R}'=\bigoplus_{k\in \mathbb{Z}}{\mathcal{R}}'_k:=\bigoplus_{k\in \mathbb{Z}} \mathfrak{a}_k t^{-k}\subset R[t, t^{-1}].
\]
Notice that if $k\le 0$, then $\mathfrak{a}_k=R$. It is well known that:
\[
\mathcal{R}'\otimes_{\mathbb{C}[t]}\mathbb{C}[t,t^{-1}]\cong R[t, t^{-1}], \quad \mathcal{R}'\otimes_{\mathbb{C}[t]}\mathbb{C}[t]/(t)\cong \bigoplus_{k=0}^{+\infty}(\mathfrak{a}_k/\mathfrak{a}_{k+1})t^{-k}\cong T.
\]
Geometrically this exactly means $W=\mathrm{Spec} (\mathcal{R}')$ and
\[
W\times_{{\mathbb{A}}^1}({\mathbb{A}}^1\setminus\{0\})=X\times ({\mathbb{A}}^1\setminus\{0\}), \quad W\times_{\mathbb{A}^1}\{0\}=C.
\]
Notice that there is a natural ${\mathbb{G}}_m$-action on ${\mathcal{R}}'$ given by the $\mathbb{Z}$-grading.
For any $f\in R$, supposing $v_0(f)=k$ then we define
$$\tilde{f}=t^{-k}f\in \mathfrak{a}_k t^{-k}\subset {\mathcal{R}}',$$ and denote
$${\bf in}(f)=[f]=[f]_{\mathfrak{a}_{k+1}}\in \mathfrak{a}_k/\mathfrak{a}_{k+1}=T_k,$$
where we use $[f]_{\mathfrak{a}}$ to denote the image of $f$ in $R/\mathfrak{a}$.
Then we define the ideal ${\mathfrak{B}}$ to be the ideal in ${\mathcal{R}}'$ generated by $\{\tilde{f}; f\in {\mathfrak{b}}\}$, and ${\bf in}({\mathfrak{b}})$ the ideal of $T$ generated by
$\{ {\bf in} (f); f\in {\mathfrak{b}} \}$.
The first two items of the following lemma is similar to (but not the same as) \cite[Theorem 15.17]{Eis94} and should be well known to experts. Notice that here we degenerate both the ambient space and the ideal. A version of the equality \eqref{eqdim} was proved in \cite[Proposition 4.3]{Li15b}.
\begin{lem}[]
\begin{enumerate}
\item There are the identities:
\begin{equation*}\label{eqflat}
\left({\mathcal{R}}'/{\mathfrak{B}}\right) \otimes_{\mathbb{C}[t]}\mathbb{C}[t,t^{-1}]\cong (R/{\mathfrak{b}})[t, t^{-1}], \quad \left({\mathcal{R}}'/{\mathfrak{B}}\right)\otimes_{k[t]}k[t]/(t)\cong T/{\bf in}({\mathfrak{b}}).
\end{equation*}
\item
The $\mathbb{C}[t]$-algebra
${\mathcal{R}}'/{\mathfrak{B}}$ is free and thus flat as a $\mathbb{C}[t]$-module. In particular, we have the identity of dimensions:
\begin{equation}\label{eqdim}
\dim_{\mathbb{C}} \left(R/{\mathfrak{b}}\right)=\dim_{\mathbb{C}} \left(T/{\bf in}({\mathfrak{b}})\right).
\end{equation}
\item
If ${\mathfrak{b}}$ is $\mathfrak{m}_R$-primary, then ${\bf in}({\mathfrak{b}})$ is an $\mathfrak{m}_T$-primary homogeneous ideal.
\end{enumerate}
\end{lem}
\begin{proof}
The statement (1) follows easily from the definition.
\bigskip
Next we prove (2). Denote by ${\mathfrak{c}}_k=T_k\cap {\bf in}({\mathfrak{b}})$ the $k$-th homogeneous piece of ${\bf in}({\mathfrak{b}})$.
We fix a basis $\left\{{\bf in}(f^{(k)}_i); 1\le i\le d_k \right\}$ of $T_k/{\mathfrak{c}}_k$. We want to show that
\[
{\mathcal{A}}':=\left\{\left.\left[\widetilde{f^{(k)}_i}\right]=\left[f^{(k)}_i\right]_{{\mathfrak{B}}}\; \right|\; 1\le i\le d_k \right\}\subset {\mathcal{R}}'/{\mathfrak{B}}
\]
is a $\mathbb{C}[t]$-basis of ${\mathcal{R}}'/{\mathfrak{B}}$.
We first verify that ${\mathcal{A}}'$ is a linearly independent set. To prove this, we just need to show that ${\mathcal{A}}'$ is a $\mathbb{C}[t, t^{-1}]$-linearly independent subset of $(R/{\mathfrak{b}})[t,t^{-1}]$.
It is then enough to show that
\begin{equation}\label{eqcA}
{\mathcal{A}}:=\left\{[f^{(k)}_i]=[f^{(k)}_i]_{\mathfrak{b}} \; |\; 1\le i\le d_k\right\}\subset R/{\mathfrak{b}}.
\end{equation}
is $\mathbb{C}$-linearly independent, which can be verified directly as in \cite[Proposition 4.3]{Li15b}. See also \cite[Proposition 15.3]{Eis94}.
So we just need to show that ${\mathcal{A}}'$ spans ${\mathcal{R}}'/{\mathfrak{B}}$. Equivalently, we need to show that for any $f\in R$, $[\tilde{f}]=[\tilde{f}]_{{\mathfrak{B}}} \in {\mathcal{R}}'/{\mathfrak{B}}$ is in the $\mathbb{C}[t]$-span of ${\mathcal{A}}'$.
This can be shown again with the help of ${\mathcal{A}}$ in \eqref{eqcA}, that is, it is enough to prove that
${\mathcal{A}}$ $\mathbb{C}$-spans $R/{\mathfrak{b}}$. Indeed, assuming the latter, for any $f\in R$, there exists a linear combination $g=\sum_{i,k}c_{ik} f^{(k)}_i$ such that
$f-g=:h\in {\mathfrak{b}}$. If $m=v_0(f)$, then
\[
\tilde{f}=t^{-m}f=\sum_{i,k}c_{ik}t^{-m}f^{(k)}_i+t^{-m}h
\]
Because $t^{-m}h\in {\mathfrak{B}}$, the above indeed implies $[\tilde{f}]$ is in the $\mathbb{C}[t]$-span of ${\mathcal{A}}'$.
To prove that ${\mathcal{A}}$ indeed $\mathbb{C}$-spans $R/{\mathfrak{b}}$, we first claim that the following set is finite:
\[
\{v_0(g)\; |\; g\in R-{\mathfrak{b}}\}.
\]
Indeed because ${\mathfrak{b}}$ is $\mathfrak{m}$-primary, there exists $N>0$ such that $\mathfrak{m}^N\subseteq {\mathfrak{b}}\subseteq \mathfrak{m}$. So $R-{\mathfrak{b}}\subseteq R-\mathfrak{m}^N$. Now the claim follows from the fact that for any element $f\in \mathfrak{m}^N$,
$$v_0(f)\le c \cdot A(v_0)\cdot N$$
by Izumi's theorem, where $c$ is a uniform constant not depending on $f$.
If there is $[f]\neq 0\in R/{\mathfrak{b}}$ that is not in the span of ${\mathcal{A}}$, then we can choose a maximal $k=v_0(f)$ such that this happens. There are two cases:
\begin{enumerate}
\item If ${\bf in}(f)\in T_k\setminus {\mathfrak{c}}_k$, then because ${\bf in}(f^{(k)}_i)$ is a basis of $T_k/{\mathfrak{c}}_k$, there exists $t_j\in \mathbb{R}$ such that ${\bf in}(f)-\sum_{j=1}^{d_k} t_j {\bf in}(f^{(k)}_j)={\bf in}(g) \in {\mathfrak{c}}_k$ for some $g\in{\mathfrak{b}}$. So we get:
\[
v_0\left(f-\sum_{j=1}^{d_k} t_j f^{(k)}_j-g\right)>k.
\]
By maximality
of $k$, $[f-\sum_{j=1}^{d_k} t_j f^{(k)}_j-g]=[f]-\sum_{j=1}^{d_k} t_j [f^{(k)}_j]$ and hence $[f]$ is in the span of ${\mathcal{A}}$. Contradiction.
\item If ${\bf in}(f)\in {\mathfrak{c}}_k={\bf in}({\mathfrak{b}})\cap T_k$. Then ${\bf in}(f)={\bf in}(g)$ for some $g\in {\mathfrak{b}}$. So $v_0(f-g)>k$ and hence $[f-g]$ is in the span
of ${\mathcal{A}}$ by the maximal property of $k$. But then $[f]=[f-g]+[g]=[f-g]$ is in the span of ${\mathcal{A}}$. Contradiction.
\end{enumerate}
\bigskip
To prove part 3 of the Lemma, we need to show that there exists $N\in \mathbb{Z}_{>0}$ such that $\mathfrak{m}_T^{N}\subseteq {\bf in}({\mathfrak{b}})\subseteq \mathfrak{m}_T$. Because
${\mathfrak{b}}$ is $\mathfrak{m}_R$ primary, there exists $N_1\in \mathbb{Z}_{>0}$ such that $\mathfrak{m}_R^{N_1}\subseteq {\mathfrak{b}} \subseteq \mathfrak{m}_R$. By Izumi's theorem,
there exists $l\in\mathbb{Z}_{>0}$ such that $\mathfrak{a}_{l m}\subseteq \mathfrak{m}_R^m$ for any $m\in \mathbb{Z}_{>0}$. By letting $N=l N_1$, it's easy to see that $\mathfrak{m}_{T}^{N}\subseteq {\bf in}({\mathfrak{b}})\subseteq\mathfrak{m}_T$.
\end{proof}
\begin{lem}
If ${\mathfrak{b}}_\bullet=\{{\mathfrak{b}}_k\}$ is a graded family of ideals of $R$, then ${\bf in}({\mathfrak{b}}_\bullet):=\{{\bf in}({\mathfrak{b}}_k)\}$ is also a graded family
of ideals of $T$.
\end{lem}
\begin{proof}
We need to show that:
\[
{\bf in}({\mathfrak{b}}_k)\cdot {\bf in}({\mathfrak{b}}_l)\subseteq {\bf in}({\mathfrak{b}}_{k+l}).
\]
If $v_0(f)=k$ and $v_0(g)=l$, then $v_0(fg)=k+l$.
\[
{\bf in}(f)\cdot {\bf in}(g)=[f]_{\mathfrak{a}_{k+1}}\cdot [g]_{\mathfrak{a}_{l+1}}=[f g]_{\mathfrak{a}_{k+l+1}}={\bf in}(f\cdot g).
\]
\end{proof}
\begin{lem}\label{l-deg}
If ${\mathfrak{b}}_\bullet$ is a graded family of ideals, then
\begin{equation}\label{eqdeg}
{\rm lct}^n({\mathfrak{b}}_\bullet)\cdot {\rm mult}({\mathfrak{b}}_\bullet)\ge {\rm lct}^n({\bf in}({\mathfrak{b}}_\bullet))\cdot {\rm mult}({\bf in}({\mathfrak{b}}_\bullet)).
\end{equation}
\end{lem}
\begin{proof}
By the flatness of ${\mathfrak{B}}$ and the lower semicontinuity of log canonical thresholds, we have ${\rm lct}({\mathfrak{b}}_k)\ge {\rm lct}({\bf in}({\mathfrak{b}}_k))$. Therefore, by \eqref{eqdim}
\begin{eqnarray*}
{\rm lct}^n({\mathfrak{b}}_k)\cdot l_R(R/{\mathfrak{b}}_k)&\ge& {\rm lct}^n({\bf in}({\mathfrak{b}}_k))\cdot l_T(T/{\bf in}({\mathfrak{b}}_k)).
\end{eqnarray*}
Taking limits as $k\rightarrow+\infty$, we then get the inequality \eqref{eqdeg}.
\end{proof}
\subsection{Equivariant K-semistability and minimizing}\label{s-equiv}
In this section, we will take a detour to show the discussion in Section \ref{s-degin} can be used to study the equivariant K-semistability. Here for a $\mathbb{Q}$-Fano variety $(X,D)$ with an action by an algebraic group $G$, we call it {\it $G$-equivariantly K-semistable (resp. Ding semistable)} if for any $G$-equivariant test configuration, its generalized Futaki (resp. Ding) invariant is non-negative. Let $T=(\mathbb{C}^*)^k$ be a torus. First we improve the two approximating results to the equivariant case.
\begin{lem}\label{l-Tmini}
Let $(X,o)=({\rm Spec}R, \mathfrak{m})$ and $D\ge 0$ a $\mathbb{Q}$-divisor, such that $o\in (X,D)$ is a klt singularity. Assume $x\in (X,D)$ admits a $T$-action. Then we have
\begin{equation}\label{eq-Tvol2mul}
\min_{v} {\widehat{\rm vol}}_{X,o}(v)=\inf_{\mathfrak{a}}{\rm lct}^n(X,D; \mathfrak{a})\cdot{\rm mult}(\mathfrak{a})=\inf_{S}{\widehat{\rm vol}}_{X,o}({\rm ord}_{S}),
\end{equation}
where on the left hand side the infimum runs over all the valuations centered at $o$, and on the middle it runs over all the $T$-equivariant $\mathfrak{m}$-primary ideals; and at the end, it runs over all $T$-equivariant Koll\'ar components.
\end{lem}
\begin{proof} Let $\{\mathfrak{a}_{\bullet}\}=\{\mathfrak{a}^k\}$ be a graded sequence for an $\mathfrak{m}$-primary ideal $\mathfrak{a}$.
The discussion in Section \ref{s-degin} implies that for $\{{\mathfrak{b}}_{\bullet}\}=_{\rm defn} {\bf in}(\mathfrak{a}_{\bullet})$
$${\rm lct}^n(X,D; {\mathfrak{b}}_{\bullet})\cdot{\rm mult}({\mathfrak{b}}_{\bullet})\le {\rm lct}^n(X,D; \mathfrak{a}_{\bullet})\cdot{\rm mult}(\mathfrak{a}_{\bullet}).$$
Since
$${\rm lct}^n(X,D; {\mathfrak{b}}_{\bullet})\cdot{\rm mult}({\mathfrak{b}}_{\bullet})=\lim_{m} {\rm lct}^n(X,D; {\mathfrak{b}}_{m})\cdot{\rm mult}({\mathfrak{b}}_{m}),$$
we conclude the first inequality as a corollary of Proposition \ref{p-inf}.
\medskip
For the second equality, we just need to show that the construction in Section \ref{s-lvmodel} can be established $T$-equivariantly. This is standard, which relies on two facts: first, we can always take an equivariant log resoltuion of $(X,D, \mathfrak{a})$ (see \cite{Kol07}); second, as $T$ is a connected group, for any curve $C$ in a $T$-variety and any $t\in T$, $t\cdot C$ will always be numerically equivalent to $C$; as the minimal model program only depends on the numerical class $[C]$, we know that any MMP sequence is automatically $T$-equivariant.
Therefore, for any $T$-equivariant $\mathfrak{m}$-primary ideal $\mathfrak{a}$, we can find a $T$-equivariant dlt modication $Y\to X$ and then a $T$-equivariant Koll\'ar component $S$, such that
$${\rm lct}^n(X,D; \mathfrak{a})\cdot{\rm mult}(\mathfrak{a})\ge {\rm vol}(Y)\ge {\widehat{\rm vol}}( {\rm ord}_{S}).$$
\end{proof}
In the below, we also need use the main idea from Fujita's work (see \cite{Fuj15, Fuj16}).
\begin{prop}[\cite{Fuj16}]\label{p-fujita}
Let $(V, \Delta)$ be an $(n-1)$-dimensional log-Fano pair which is $T$-equivariantly log-Ding-semistable. Let $\delta$ be a positive rational number such that $-\delta^{-1}(K_V+\Delta)$ is Cartier.
Let $I_M\subset \cdots\subset I_1\subset \mathcal{O}_V$ be a sequence of $T$-invariant coherent ideal sheaves and assume
$${\mathcal{I}}:=I_M+I_{M-1} t^1+\cdots+I_1 t^{M-1}+(t^M)\subset \mathcal{O}_{V\times \mathbb{C}}$$
form a flag ideal. Let
$\Pi: {\mathcal{V}}\rightarrow V\times\mathbb{C}$ be the blowup along ${\mathcal{I}}$. Denote ${\mathcal{D}}=\Pi^*(\Delta\times \mathbb{C})$. Let $E\subset {\mathcal{X}}$ be the Cartier divisor defined by $\mathcal{O}_{{\mathcal{V}}}(-E)={\mathcal{I}}\cdot \mathcal{O}_{{\mathcal{V}}}$, and let
\[
{\mathcal{L}}:=\Pi^*\mathcal{O}_{V\times \mathbb{C}}\left(-\delta^{-1}(K_{V\times\mathbb{C}/\mathbb{C}}+\Delta\times \mathbb{C})\right)\otimes \mathcal{O}_{{\mathcal{X}}}(-E).
\]
Assume that ${\mathcal{L}}$ is semi ample over $\mathbb{C}$. Then $({\mathcal{V}}, {\mathcal{D}}; {\mathcal{L}})$ is naturally seen as a (possibly non-normal) semi test configuration of $(S, \Delta; -\delta^{-1}(K_V+D))$. Under these conditions,
$((V\times\mathbb{C}^1, \Delta\times \mathbb{C}); {\mathcal{I}}^\delta\cdot (t)^d)$ must be sub log canonical, where
\[
d:=1+\frac{\delta^{n}\bar{{\mathcal{L}}}^n}{n (-K_V-\Delta)^{n-1}}.
\]
\end{prop}
The following argument is essentially in \cite{Li15b}.
\begin{prop}\label{thm-Ksemi}
If $(V,\Delta)$ is a projective log Fano variety and $o\in (X,D)$ is the cone of $(V,\Delta)$ induced by some ample Cartier divisor $L=-r^{-1}(K_V+\Delta)$. Then the canonical valuation obtained by blowing up the vertex minimizes ${\widehat{\rm vol}}_{X,D}$ if and only if $(V,\Delta)$ is K-semistable.
\end{prop}
\begin{proof}
First we assume $(S, \Delta)$ is $(n-1)$-dimensional log-K-semistable and prove the volume minimizing. By Lemma \ref{l-Tmini}, we only need to prove that for any $\mathbb{C}^*$-invariant divisorial valuation $v$ over $(X, o)$,
$${\widehat{\rm vol}}({\rm ord}_V)\le {\widehat{\rm vol}}(v).$$ Let $c_1=v(V)>0$ and $R=\bigoplus_{k=0}^{+\infty}R_k=\bigoplus_{k=0}^{+\infty}H^0(V, k L)$ thus $X={\rm Spec}(R)$.
On $R$, we define a filtration
$${\mathcal{F}} R^{(t)} = \bigoplus_k {\mathcal{F}}^{kt}R_k, $$ where
$${\mathcal{F}}^xR_k:=H^0(V, L^{\otimes k}\otimes \mathfrak{a}_x), \qquad \mbox{and \ \ } \mathfrak{a}_x=\{f\in \mathcal{O}_V\ |\ v(f)\ge x\}. $$
Then the volume is defined to be
$${\rm vol}({\mathcal{F}} R^{(t)}) := \limsup_{m\to \infty} \frac{\dim_{\mathbb{C}} {\mathcal{F}}^{mt}R_m}{m^n/n!}.$$
By \cite[(18)]{Li15b}, we get a formula for
${\rm vol}(v)$:
\begin{eqnarray*}
{\rm vol}(v)&=&\lim_{m\rightarrow+\infty} \frac{n!}{m^n}\dim_{\mathbb{C}} R/\mathfrak{a}_m(v)\\
&=&\frac{L^{n-1}}{c_1^n}-\int^{+\infty}_{c_1}{\rm vol}\left({\mathcal{F}} R^{(t)}\right)\frac{dt}{t^{n+1}}\\
&=&-\int^{+\infty}_{c_1}\frac{d {\rm vol}\left({\mathcal{F}} R^{(t)}\right)}{t^n}.
\end{eqnarray*}
Then we consider the following function
\begin{eqnarray*}
\Phi(\lambda, s)&=&\frac{L^{n-1}}{(\lambda c_1 s+(1-s))^n}-n \int^{+\infty}_{c_1}{\rm vol}\left({\mathcal{F}} R^{(t)}\right)\frac{\lambda s dt}{(1-s+\lambda st)^{n+1}}\\
&=&\int^{+\infty}_{c_1}\frac{-d \; {\rm vol}({\mathcal{F}} R^{(t)})}{((1-s)+\lambda st)^n}.
\end{eqnarray*}
$\Phi(\lambda, s)$ satisfies the following properties:
\begin{enumerate}
\item For any $\lambda \in (0, +\infty)$, we have:
\[
\Phi(\lambda, 1)={\rm vol}(\lambda v)=\lambda^{-n} {\rm vol}(v), \quad \Phi(\lambda, 0)={\rm vol}(v_0)=L^{n-1}.
\]
\item For any $\lambda \in (0, +\infty)$, $\Phi(\lambda, s)$ is continuous and convex with respect to $s\in [0, 1]$.
\item The directional derivative of $\Phi(\lambda, s)$ at $s=0$ is equal to:
\[
\Phi_s(\lambda, 0)=n \lambda L^{n-1} \left(\lambda^{-1}-c_1-\frac{1}{L^{n-1}}\int^{+\infty}_{c_1} {\rm vol}\left({\mathcal{F}} R^{(t)}\right)dt\right).
\]
\end{enumerate}
Let $\lambda_*=\frac{r}{A_{(X, C(D))}(v)}$. By Item 1 and 2, we just need to prove $\Phi_s(\lambda_*, 0)\ge 0$.
Let $\bar{v}=v|_{\mathbb{C}(S)}$ be the restriction of $v$ under the inclusion $\mathbb{C}(S)\hookrightarrow \mathbb{C}(X)$.
It is known that $\bar{v}=b\cdot {\rm ord}_E$ where $b>0$ and ${\rm ord}_E$ is a divisorial valuation on $\mathbb{C}(S)$.
Using Adjunction formula, it is easy to show that:
\[
\lambda^{-1}-c_1=\frac{A_{(X, D)}(v)}{r}-c_1= A_{V,\Delta}(\bar{v})=b\cdot A_{V,\Delta}(E).
\]
By change of variables we get:
\[
\int^{+\infty}_{c_1}{\rm vol}\left({\mathcal{F}} R^{(t)}\right)dt=\int^{+\infty}_0 {\rm vol}\left({\mathcal{F}}_{\bar{v}} R^{(t)}\right) dt.
\]
where
\[
{\mathcal{F}}_{\bar{v}} R^{(t)} = \bigoplus_k H^0(V, L^{\otimes k}\otimes \mathfrak{a}_{kt}), \qquad \mbox{and \ \ } \mathfrak{a}_{kt}=\{f\in \mathcal{O}_V\ |\ \bar{v}(f)\ge kt\}.
\]
So we get:
\begin{eqnarray*}
\Phi_s(\lambda_*, 0)&=&n \lambda_* L^{n-1}\left(A_{(V,\Delta)}(\bar{v})-\frac{1}{L^{n-1}}\int^{+\infty}_{0} {\rm vol}\left({\mathcal{F}}_{\bar{v}}R^{(t)}\right)dt\right)\\
&=&n \lambda_* L^{n-1} b\left(A_{(V,\Delta)}(E)-\frac{1}{L^{n-1}}\int_0^{+\infty} {\rm vol}\left({\mathcal{F}}_{{\rm ord}_E}R^{(t)}\right)dt \right).
\end{eqnarray*}
By the valuative criterion of (log-)K-semistability derived in \cite{Li15b, Fuj16, LL16}, we get $\Phi_s(\lambda_*, 0)\ge 0$ (see e.g. Proposition \ref{p-fujita}).
Conversely the statement follows immediately by the valuative criterion for log-K-semistability in \cite{Li15b, Fuj16, LL16} by choosing a family of valuation $v_s$ emanating from $v_0={\rm ord}_S$ in the direction of ${\rm ord}_E$ as in \cite[Section 7]{Li15b}.
\end{proof}
\begin{rem}
The fact that K-semistability implies the canonical valuation is a minimizer was proved in \cite{LL16}, where $(V,\Delta)$ is assumed to be analytic K-semistable, i.e., if there exists a special test configuration $({\mathcal{X}}, \mathcal{D})$ that degenerates $(X,D)$ to a log Fano $(X_0,D_0)$ with a conic K\"ahler-Einstein metric. If $(V,\Delta)$ is analytic K-semistable, then we know $(V,\Delta)$ is K-semistable. The converse implication conjecturally true, and known for smooth $X$ when $D=0$ (see \cite{CDS15, Tia15}).
\end{rem}
With all the techniques we have, we can prove Theorem \ref{t-equiK}.
\begin{proof}[Proof of Theorem \ref{t-equiK}]
Let $(X,D)$ be a cone of $-r(K_V+\Delta)$ over $(V,\Delta)$ for some sufficiently divisible positive integer $r$. And we consider the minimizing problem of the normalized local volume at the $T$-singularity $o$ which is the vertex. We aim to show that if $(V, \Delta)$ is $T$-equivariantly K-semistable then ${\rm ord}_V$
minimizes ${\widehat{\rm vol}}_{X,D}$.
Following the proof of Lemma \ref{l-Tmini}, by degenerating the ideal $r$ times, we can find a sequence of $T$-equivariant ideals $\{\mathfrak{a}\}_{i\in I}$ such that
$$\inf_{i\in I}{\rm lct}^n(X,D; \mathfrak{a})\cdot{\rm mult}(\mathfrak{a})=\min_{v\in {\rm Val}_{X,o}} {\widehat{\rm vol}}(v).$$ Then using the equivariant resolution and running an MMP process, we can find a sequence of $T$-equivariant Koll\'ar components $S_i$ such that
$$\inf_{i} {\widehat{\rm vol}}({\rm ord}_{S_i})= \min_{v\in {\rm Val}_{X,o}} {\widehat{\rm vol}}(v). $$
Since we assume that for any $T$-equivariant special degeneration of $ (S,\Delta)$, the generalized Futaki invariant is non-negative, and for special test configuration, the generalized Futaki invariant is the same as Ding invariant, we know that the Ding invariant for any $T$-equivariant special test configuration is non-negative. Using the fact that MMP decreases Ding invariant (see \cite[Corollary 3.4]{Fuj16} and its proof), we know this implies that the Ding invariant for any $T$-equvariant test configuration is nonnegative.
Then for any $T$-equivariant Koll\'ar component $S_i$, we consider $v={\rm ord}_{S_i}\in {\rm Val}_{X,o}$. Denote its induced divisorial valuation $b\cdot {\rm ord}_E$ on $S$.
Restricting the calculation in \cite{Fuj16} to the equivariant setting, we conclude that
$$A_{(V,\Delta)}(E)-\frac{1}{L^{n-1}}\int_0^{+\infty} {\rm vol}\left({\mathcal{F}}_{{\rm ord}_E}R^{(t)}\right)dt\ge 0.$$
Running the construction as in the proof of Proposition \ref{thm-Ksemi}, then we know that $\Phi_s(\lambda_*, 0)\ge 0$
which gives that $${\widehat{\rm vol}}(v_*)\le {\widehat{\rm vol}}({\rm ord}_{S_i}).$$
This implies that the canonical component is a minimizer which then implies that $(V,\Delta)$ is K-semistable by Proposition \ref{thm-Ksemi} .
\end{proof}
\subsection{Proof of Theorem \ref{t-main}}
Let $(X, o)=({\rm Spec}(R), \mathfrak{m})$ be an algebraic singularity such that $(X,D)$ is klt for a $\mathbb{Q}$-divisor $D\ge 0$.
Let $S$ be a Koll\'{a}r component and $\Delta=\Delta_S$ be the different divisor defined by adjunction $(K_Y+S+\mu_*^{-1}D)|_S=K_S+\Delta_S$ where $Y\rightarrow X$ is the extraction of $S$. We follow the notation in Section \ref{ss-deformation}.
\begin{lem}\label{l-ksemi}
Let ${\mathfrak{b}}_\bullet$ be a graded sequence of $\mathfrak{m}_0$-primary ideal.
If $(S,\Delta_S)$ is K-semistable, then we have
$$ {\rm lct}^n({\mathfrak{b}}_\bullet)\cdot {\rm mult}({\mathfrak{b}}_\bullet)\ge {\widehat{\rm vol}}_{C,C_{\Delta},o}({\rm ord}_S).$$
\end{lem}
\begin{proof}
Using the result in \cite{JM12}, we have
\begin{eqnarray*}
{\rm lct}^n({\mathfrak{b}}_\bullet)\cdot {\rm mult}({\mathfrak{b}}_\bullet)&=&\lim_{k \to +\infty} \big(k\cdot {\rm lct}({\mathfrak{b}}_k) \big)^n\cdot \frac {{\rm mult}({\mathfrak{b}}_k)}{k^n}\\
&=&\lim_{k\to +\infty} {\rm lct}^n({\mathfrak{b}}_k)\cdot {\rm mult}({\mathfrak{b}}_k).
\end{eqnarray*}
By Proposition \ref{p-inf}, it suffices to show that ${\widehat{\rm vol}}_{C,C_{\Delta},o}({\rm ord}_S)$ is equal to
$$\min_{v} {\widehat{\rm vol}}_{C,C_{\Delta},o}(v)$$
for $v$ runs over valuations centered on $o$.
It follows from Proposition \ref{thm-Ksemi} that if we choose $d$ sufficiently divisible, such that $C^{(d)}=C(S,H)$ is constructed as the cone over $S$ with some ample Cartier divisor $H$ proportional to $-(K_{S}+\Delta_{S})$, then the canonical valuation ${\rm ord}_S$ is a minimizer of ${\widehat{\rm vol}}_{C^{(d)}, C^{(d)}_{\Delta}+C^{(d)}_D}$. By Proposition \ref{p-cover}, this implies the same for $C$.
\end{proof}
\begin{prop}\label{p-cover}
Under the above notation, ${\widehat{\rm vol}}_{({C}, C_D)}$ minimizes at ${\rm ord}_S$ if and only if ${\widehat{\rm vol}}_{({C}^{(d)}, C^{(d)}_{\Delta}+C^{(d)}_D)}$ minimizes at ${\rm ord}_{S^{(d)}}$.
\end{prop}
\begin{proof}The degree $d$ cover $h\colon C\to C^{(d)}$ is a fiberwise map and the Galois group $G=_{\rm defn}\mathbb{Z}/d$ is a natural subgroup $\mathbb{C}^*$.
Let $E$ be a Koll\'ar component over $C^{(d)}$, by Lemma \ref{l-finite} we know $h^*(E)$ is a Koll\'ar component over $ C$, and it follows from Lemma \ref{l-finitevolume} (or \cite[Lemma 6.9]{Li15b}) that
$$d\cdot {\widehat{\rm vol}}({\rm ord}_E)={\widehat{\rm vol}}(h^*E). $$
So if ${\rm ord}_S$ minimizes ${\widehat{\rm vol}}_{C,C_{\Delta}}$, then the corresponding canonical valuation also minimizes ${\widehat{\rm vol}}_{{C}^{(d)}, C^{(d)}_{\Delta}+C^{(d)}_D}$.
For the converse, let $E$ be a $T$-invariant Koll\'ar component over $C$, since it is $G$-invariant, by Lemma \ref{l-finite} we know it is a pull back of a Koll\'ar component $F$ over $C^{(d)}$. If the canonical valuation minimizes ${\widehat{\rm vol}}_{{C}^{(d)}, C^{(d)}_{\Delta}+C^{(d)}_D}$, then we see over $C$, ${\widehat{\rm vol}}({\rm ord}_S)$ is less or equal to ${\widehat{\rm vol}}({\rm ord}_E)$ for any $T$-equivariant Koll\'ar component $E$, therefore ${\rm ord}_S$ is a minimizer of ${\widehat{\rm vol}}_{C,C_{\Delta}}$ by Lemma \ref{l-Tmini}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{t-main}]
With the notation as in Section \ref{ss-deformation}, we denote by $ \mathcal{D}$ the birational transform of $ D\times \mathbb{A}^1$ on $\mathcal{Z}^0 $ and write $a\mathcal{T}\sim_{\mathbb{Q},{\mathcal{X}}}K_{{\mathcal{Z}}^0}+\mathcal{T}$. Restricting over a general fiber and taking the coarse moduli spaces, we obtain
$$aS\sim_{X,{\mathbb{Q}}} K_Y+S+\mu^{-1}_*(D),$$
then $a=A_{X,D}(S)$. Similarly, over the central fiber over $0$, we get
$$aS_0\sim_{X,{\mathbb{Q}}} K_{Y_0}+S_0+\mu^{-1}_{0,*}C_D,$$
thus $a=A_{C,C_D}(S_0)$, where $\mu_0\colon Y_0\to C$ is the blow up of the vertex with the exceptional divisor $S_0$.
We also know that
$${\rm vol}_{X,o}({\rm ord}_S)=(-S|_S)^{n-1}=(-S_0|_{S_0})^{n-1}={\rm vol}_{C,p}({\rm ord}_{S_0}).$$
Combining all the above, we know that for any ideal ${\mathfrak{b}}$ on $X$, if we let ${\mathfrak{b}}_{\bullet}=\{{\mathfrak{b}}^k\}$, then
\begin{eqnarray*}
{\widehat{\rm vol}}_{X,o}({\rm ord}_S) &= &{\rm vol}_{X,o}({\rm ord}_S)\cdot A^n_{X,D}(S) \\
& = &{\rm vol}_{C,p}({\rm ord}_{S_0})\cdot A^n_{C,{C_{D}}}(S_0)\\
&= & \inf_v {\rm vol}_{C,p}({\rm ord}_{v})\cdot A^n_{C,{C_{D}}}(v)\\
&\le & {\rm lct}^n({\bf in}({\mathfrak{b}}_\bullet))\cdot {\rm mult}({\bf in}({\mathfrak{b}}_\bullet))\\
&\le & {\rm lct}^n({\mathfrak{b}}) \cdot {\rm mult} ({\mathfrak{b}}),
\end{eqnarray*}
where the last two inequalities follow from Lemma \ref{l-deg} and \ref{l-ksemi}.
Thus we conclude that
$${\widehat{\rm vol}}_{X,o}({\rm ord}_S)\le \inf_{{\mathfrak{b}}} {\rm lct}^n({\mathfrak{b}}) \cdot {\rm mult} ({\mathfrak{b}})=\inf_v {\widehat{\rm vol}}_{X,o}(v),$$
where the second equality follows from Proposition \ref{p-inf}.
\end{proof}
\section{Uniqueness}\label{s-uni}
In this section, we will prove Theorem \ref{t-main2} about the uniqueness of the minimizers among all Koll\'ar components. There are two steps: first we prove this for cone singularities; then for a general singularity, we combine the deformation construction with some results from the minimal model program to essentially reduce it to the case of cone singularities.
\subsection{Case of cone singularity}\label{ss-ucone}
We first settle the case of cone singularities. It can be proved using Proposition \ref{p-Tequiv} and \cite[Theorem 3.4]{Li15b}. Here we give a different proof, which analyzes the geometry in more details. A similar argument in the global case appears in the proof of \cite[Theorem 3]{Liu16}, where a characterization of quotients of $\mathbb{P}^n$ is given by achieving the maximal possible volumes among all K-semistable $\mathbb{Q}$-Fano varieties with only quotient singularities.
\bigskip
Let $(V,\Delta)$ be an $(n-1)$-dimensional log Fano variety and $-(K_V+\Delta)=r H$ for an ample Cartier divisor $H$ and $r\le n$. Assume $X^0:=C(V, H)$ is the affine cone over the base $V$ with the vertex $o$ and let $X$ be the projective cone and $D$ be the cone divisor over $\Delta$ on $X$. In the below, for a variety $\bullet$, we denote the product $\bullet \times \mathbb{A}^1$ by $\bullet_{\mathbb{A}^1}$.
Consider a Koll\'ar component $S$ over $o\in (X,D)$ with the extraction morphism $\mu\colon Y\to X$. Let
$$\mu_{\mathbb{A}^1}(=\mu \times {\rm id})\colon Y_{\mathbb{A}^1}\rightarrow X_{\mathbb{A}^1}$$
be the extraction of $S_{\mathbb{A}^1}$. We carry out the deformation process in Section \ref{ss-deformation} with respect to $S$. Here $X$ is a projective variety instead of a local singularity, but the construction is exactly the same. We denote by $Z$ (resp. $W$) the coarse moduli space of ${\mathcal{Z}}$ (resp. ${\mathcal{W}}$), so there are morphisms, $\psi_1:Z\to W$, $\phi_1\colon Z\to Y_{\mathbb{A}^1}$ and $\pi\colon W\to X_{\mathbb{A}^1}$. We denote by $\phi=\mu_{\mathbb{A}^1}\circ\phi_1.$
\begin{center}
\begin{figure}[h]
\centering
\includegraphics[width=10cm, height=6cm]{bir.jpg}
\end{figure}
\end{center}
Then we have the following equalities:
\begin{enumerate}
\item
$K_{Y_{\mathbb{A}^1}}+(\mu_{\mathbb{A}^1})_*^{-1}(D_{\mathbb{A}^1})=\mu_{\mathbb{A}^1}^* (K_{X_{\mathbb{A}^1}}+D_{\mathbb{A}^1})+a S_{\mathbb{A}^1} $ with $a=A_{X,D}(S)-1$;
\item
$K_{Z}+\phi_*^{-1}(D_{\mathbb{A}^1})=\phi_1^*\big(K_{Y_{\mathbb{A}^1}}+(\mu_{\mathbb{A}^1})_*^{-1}D_{\mathbb{A}^1}\big)+{\mathcal{F}}$;
\item
$K_{Z}+\phi_*^{-1}D_{\mathbb{A}^1}=\psi_1^* \big(K_{W}+(D_{\mathbb{A}^1})_W\big)+a S_{\mathbb{A}^1}$, where $(D_{\mathbb{A}^1})_W:=\psi_{1*}\phi_*^{-1}(D_{\mathbb{A}^1})$.
\end{enumerate}
The first two equalities imply:
\begin{eqnarray*}
K_{Z}+\phi_*^{-1}(D_{\mathbb{A}^1})&=&\phi_1^* \big(K_{Y_{\mathbb{A}^1}}+(\mu_{\mathbb{A}^1})_*^{-1}D_{\mathbb{A}^1}\big)+ {\mathcal{F}}\\
&=&\phi_1^*\mu_{\mathbb{A}^1}^* (K_{X_{\mathbb{A}^1}}+D_{\mathbb{A}^1})+a \phi_1^*S_{\mathbb{A}^1}+{\mathcal{F}}\\
&=&\phi^*(K_{X_{\mathbb{A}^1}}+D_{\mathbb{A}^1})+a S_{\mathbb{A}^1}+(a+1){\mathcal{F}}.
\end{eqnarray*}
So $A_{X_{\mathbb{A}^1},D_{\mathbb{A}^1}}({\mathcal{F}})=a+2=A_{X,D}(S)+1$. This implies:
\[
K_{W}+(D_{\mathbb{A}^1})_W=\pi^*K_{X_{\mathbb{A}^1},D_{\mathbb{A}^1}}+A_{X,D}(S) \bar{C}.
\]
Recall that $-K_X-D=(1+r) L$, where $L=\mathcal{O}(1)$ for the cone construction, so we get:
\[
K_{W}+(D_{\mathbb{A}^1})_W=-(1+r)\rho^*L+A_{X,D}(S) \bar{C},
\]
where $\rho\colon W\to X$ the composite of $\pi\colon W\to X_{\mathbb{A}^1}$ with the first projection $X_{\mathbb{A}^1}\to X$.
\begin{rem}
\begin{enumerate}
\item As in \cite{LL16}, we define the cone angle parameter $\beta=\frac{r}{n}$. Denote by $V_{\infty}$ the section at infinite place of $X$ and ${\mathbb{V}}_{\infty}$ the birational transform of $(V_{\infty})_{\mathbb{A}^1}$. Then
$$-(K_X+D+(1-\beta)V_\infty)=\delta L$$ with $\delta=r\frac{n+1}{n}$, we get:
\begin{eqnarray}\label{lclY2}
& &K_{W}+(D_{\mathbb{A}^1})_W+(1-\beta){\mathbb{V}}_\infty\\
&=&\pi^*\big(K_{X_{\mathbb{A}^1}}+D_{\mathbb{A}^1}+(1-\beta) {\mathbb{V}}_\infty\big)+A_{X,D}(S)\bar{C}\nonumber\\
&=&-\delta\rho^*L+A_{X,D}(S)\bar{C}.
\end{eqnarray}
\item
Because $K_{Y}+\mu_*^{-1}D=\mu^*(K_X+D)+(A_{X,D}(S)-1)S$, we have:
\begin{eqnarray*}
K_{Y}+\mu^{-1}_*D+(1-\beta)V_\infty&=&\mu^*\big(K_X+D+(1-\beta)V_\infty\big)+(A_{X,D}(S)-1)S\\
& =&-\delta \mu^*L+(A_{X,D}(S)-1)S.
\end{eqnarray*}
\end{enumerate}
\end{rem}
The above construction works for any Koll\'{a}r component. From now on we assume that $(V,\Delta)$ is K-semistable and $S$ minimizes the normalized volume, i.e. it satisfies
\begin{equation}
{\widehat{\rm vol}}({\rm ord}_{S})={\widehat{\rm vol}}({\rm ord}_{V_0})=r^n (H^{{n-1}}),
\end{equation}
where $V_0$ denotes the canonical divisor obtained by blowing up the vertex of the cone and we aim to show $S=V$ is the canonical component. We note that by Proposition \ref{thm-Ksemi} , we know that ${\widehat{\rm vol}}({\rm ord}_{V_0})$ is the minimal normalized volume.
Then we have:
\[
{\rm vol}({\rm ord}_{S})=\frac{{\widehat{\rm vol}}({\rm ord}_{S})}{A_{X,D}(S)^n}=\frac{r^n (H^{n-1})}{A_{X,D}(S)^n}.
\]
In Section \ref{s-equiv}, we have used the filtration induced by a valuation (see also \cite{BHJ15, Fuj15}). Here we use the same construction but for sections on the projective cone instead of the base.
\begin{defn}[Filtration by valuation]
Fixed a valuation $v$. Let $S_m=H^0(X, L^{\otimes m})$. Define ${\mathcal{F}}^xS_m\subset S_m$ to be a decreasing filtration (with respect to $x$) as follows:
$${\mathcal{F}}^xS_m:=H^0(X, L^{\otimes m}\otimes \mathfrak{a}_x), \qquad \mbox{where \ } \mathfrak{a}_x=\{f\in \mathcal{O}_X\ |\ v(f)\ge x\}. $$
On $\bigoplus_{m=0} S_m$, we define ${\mathcal{F}} S^{(t)} = \bigoplus {\mathcal{F}}^{kt}S_k$. Then the volume is defined to be
$${\rm vol}({\mathcal{F}} S^{(t)}) := \limsup_{m\to \infty} \frac{\dim_{\mathbb{C}} ({\mathcal{F}}^{mt}S_m)}{m^n/n!}.$$
\end{defn}
\begin{prop}\label{p-fanocone}
With the above notation, if the case $(V,\Delta)$ is log K-semistable, then $(X,D+(1-\beta)V_{\infty})$ is log K-semistable. As a consequence,
\[
A_{X, D}(S)-\frac{\delta }{(L^{n})}\int_0^{+\infty}{\rm vol}({\mathcal{F}} S^{(x)})dx\ge 0.
\]
\end{prop}
\begin{proof}By Theorem \ref{t-equiK}, we only need to check this is true for $\mathbb{C}^*$-equvariant special test configuration $(\mathcal{X},\mathcal{D}+(1-\beta){\mathcal{V}})/\mathbb{P}^1$ of $(X,D+(1-\beta)V_{\infty})$, where we consider the test configuration over $\mathbb{P}^1$ by adding a trivial fiber over $\{\infty \}$.
Consider the closure of $\mathcal{V}\supset V_{\infty}\times( \mathbb{P}^1\setminus \{0 \})$. As $(1+r)V\sim_{\mathbb{Q}} -(K_X+D) $, we know that if we denote by ${\mathcal{L}}$ the polarization on the test configuration extending $L$, then
$$(1+r)\mathcal{V}\sim_{\mathbb{Q},\mathbb{P}^1}-K_{\mathcal{X}}-\mathcal{D}=(1+r)\mathcal{L}$$
as the fiber over $0$ is irreducible.
Let $\Delta_{\infty}(=\Delta)=V_{\infty}\cap D$ and $\mathcal{E}$ be the closure of $\Delta_{\infty}\times ( \mathbb{P}^1\setminus \{0\})$, then
$$(K_{\mathcal{X}}+\mathcal{D}+{\mathcal{V}})|_{{\mathcal{V}}}=K_{{\mathcal{V}}}+\mathcal{E}$$ as $\mathcal{X}$ is smooth along the codimension 2 points over $0$ and so there is no different divisor.
Then the generalized Futaki invariant of $(\mathcal{X},\mathcal{D}+(1-\beta){\mathcal{V}})/\mathbb{P}^1$ is
$${\rm Fut}(\mathcal{X})= -\frac{1}{(n+1)(\delta L)^n}(-K_{\mathcal{X}/\mathbb{P}^1}-\mathcal{D}-(1-\beta)\mathcal{V})^{n+1}. $$
Since $V_{\infty}\sim L$ and $\delta=r\frac{n+1}{n}$, it's easy to verify that the generalized Futaki invariant of the induced test configuration of $(V, \Delta)$ is
$${\rm Fut}(\mathcal{V})= -\frac{1}{nr^{n-1} H^{n-1}}((-K_{\mathcal{V}/\mathbb{P}^1}-\mathcal{E})|_{\mathcal{V}})^{n}={\rm Fut}(\mathcal{X}). $$
Recall the log-K-semistability is equivalent to the log-Ding-semistablity (see e.g. \cite{Fuj16}). Then the second part is a standard generalization of \cite[Proposition 4.5]{LL16} to the log setting, and we skip it.
\end{proof}
The key calculations are given by the following results proved in in \cite{LL16}.
\begin{prop}\label{p-equal}
Suppose $(V, \Delta)$ is log-K-semistable. If $S$ is a Koll\'{a}r component obtaining the minimum of ${\widehat{\rm vol}}$ over $(X, o)$, then the graded filtration induced by $S$ satisfies the following two conditions:
\begin{enumerate}
\item
\[
A_{X, D}(S)-\frac{\delta }{(L^{n})}\int_0^{+\infty}{\rm vol}({\mathcal{F}} S^{(x)})dx=0.
\]
\item Denote $\tau:=\sqrt[n]{\frac{(L^{ n})}{{\rm vol}({\rm ord}_{S})}}$. We have:
\[
{\rm vol}\left({\mathcal{F}} S^{(x)}\right)={\rm vol}_Y(\mu^*L-xS)=(L^{n})-{\rm vol}({\rm ord}_{S})x^n
\text{ for any } x\in [0, \tau].
\]
\end{enumerate}
\end{prop}
\begin{proof}By Proposition \ref{p-fanocone}, we know $(X,D+(1-\beta)V_{\infty})$ is log K-semistable. Then it follows from precisely the equality case of the Formula (25) and (29) in \cite{LL16}.
\end{proof}
\begin{lem}We have $\tau=\frac{A_{X,D}(S)}{r}$.
\end{lem}
\begin{proof}
Combining 1 and 2 in Proposition \ref{p-equal}, we know that
\[
A_{X,D}(S)-\frac{r(1+n)}{n\cdot L^n}\int^{\tau}_{0}\big(L^n-{\rm vol}({\rm ord}_S)x^n\big)dx= A_{X,D}(S)-r\cdot \tau= 0.
\]
\end{proof}
By arguing as in \cite{Fuj15} (see also \cite{Liu16}), we know that:
\begin{lem}\label{lemamp}
We know $\tau$ is the nef threshold, i.e.
$$\tau=\sup\left\{x| \ \mu_{\mathbb{A}^1}^*L-xS \text{ is ample } \right\}.$$
\end{lem}
\begin{proof} When the point is smooth, this follows from \cite[Theorem 2.3(2)]{Fuj15}. Exactly the same argument can be used to treat the current case.
\end{proof}
\begin{thm}\label{t-cone}
If $S$ is a Koll\'{a}r component obtaining the minimum of the normalized volume, then $S$ is the canonical component $V_0$.
\end{thm}
We first show the following statements.
\begin{lem}
\begin{enumerate}
\item
$\rho^*L-\tau \bar{C}$ is base-point-free, and contracts $Y$ to $S_{\infty}(\cong S)\subset \bar{C}$ as the section at the infinite place.
\item
$A_{X,D}(S)=r$ and there is a special test configuration degenerating $(V, \Delta)$ into $(E, \Delta_E)$.
Moreover, there is a special test configuration of $(X, D+ (1-\beta)V_\infty; L)$
to $(\bar{C}, D_{\bar{C}}+(1-\beta) S_{\infty}; {\mathcal{L}}_0)$ which is indeed an isomorphism, where $D_{\bar{C}}$ is the intersection of $ \bar{C}$ with $(D\times {\mathbb{A}}^1)_W$.
\end{enumerate}
\end{lem}
\begin{proof}
The proof of this part is along the similar line in \cite[Proof of Lemma 33]{Liu16}.
First we observe the following restrictions of $\rho^*L-x\bar{C}$:
\begin{enumerate}
\item
$\left.\rho^*L-x\bar{C}\right|_{X_t}=L$, $t\neq 0$;
\item
$\left.\rho^*L-x \bar{C}\right|_{Y_0}=\mu^*L-x S$;
\item
$\left.\rho^*L-x\bar{C}\right|_{\bar{C}}=-x\bar{C}|_{\bar{C}}=x Y_0|_{\bar{C}}=x S_{\infty}=x\mathcal{O}_{\bar{C}}(1)$.
\end{enumerate}
So by Lemma \ref{lemamp}, it is easy to see that $\rho^*L-x\bar{C}$ is ample when $x\in (0, \tau)$.
To show that $\rho^*L-\tau\bar{C}$ is base-point-free, we calculate by using \eqref{lclY2}:
\begin{eqnarray*}
m(\rho^*L-x \bar{C})-K_{W}-(D_{{\mathbb{A}}^1})_W&=&m(\rho^*L-x\bar{C})+(1+r) \rho^*L-A_{X,D}(S)\bar{C}\\
&=&(m+1+r)\left(\rho^*L-\frac{mx+A_{X,D}(S)}{m+1+r}\bar{C}\right).
\end{eqnarray*}
Notice that:
\[
\frac{mx+A_{X,D}(S)}{m+1+r}<\tau= \frac{A_{X,D}(S)}{r}
\]
if and only if
\[
x< \left(1+\frac{1}{m}\right)\frac{A_{X,D}(S)}{r}.
\]
Because this is satisfied for
$$x=\tau=\frac{A_{X,D}(S)}{r}\qquad \mbox{for any\ } m>0,$$ the first statement holds by base-point-free theorem \cite[Theorem 3.13]{KM98}.
Next we claim that
\begin{equation}\label{eqcontr1}
H^0(Y, m(\mu^*L-\tau S))\cong H^0(S, -m\tau S)
\end{equation}
for any $m$ sufficiently divisible. To see this, we consider the exact sequence:
\begin{equation}
0\rightarrow \mathcal{O}_{Y}(m(\mu^*L-\tau S)-S)\rightarrow \mathcal{O}_{Y}(m(\mu^*L-\tau S))\rightarrow \mathcal{O}_{Y}(m(\mu^*L-\tau S))\otimes\mathcal{O}_S\rightarrow 0,
\end{equation}
and its associated long exact sequence of cohomology groups.
By the above discussion, and
\[
m(\mu^*L-\tau S)-S-K_Y=m(\mu^*L-\frac{A_{X,D}(S)}{r}S)+(1+r)\mu^*L-A_{X,D}(S)S
\]
is ample, it follows from the Kawamata-Viehweg vanishing theorem that
\[
H^1\big(Y, m(\mu^*L-\tau S)\otimes \mathcal{O}(-S)\big)=0 \text{ for any } m\ge 0.
\]
We also have
\[
H^0\big(Y, m(\mu^*L-\tau S)\otimes \mathcal{O}(-S)\big)=0 \text{ for any } m\ge 0.
\]
as $\tau$ is also the pseudo-effective threshold.
Thus we know $|m(\rho^*L-\tau \bar{C})|$ contracts the fiber $W\times_{\mathbb{A}^1} \{0 \}$ to $\bar{C}$ for sufficiently divisible $m$. This finishes the proof of (1). We denote by $ \theta\colon W\to \mathcal{X}$ the induced morphism and there is an ample line bundle ${\mathcal{L}}$ on ${\mathcal{X}}$ such that $\theta^*{\mathcal{L}}=\rho^*L-\tau \bar{C}$.
\bigskip
Next we prove (2). Let $(D_{\mathbb{A}^1})_{{\mathcal{X}}}$ be the push forward of $(D_{\mathbb{A}^1})_{W}$ on ${\mathcal{X}}$. Then $-K_{{\mathcal{X}}}-(D_{\mathbb{A}^1})_{{\mathcal{X}}}$ and $(1+r) {\mathcal{L}}$ coincide outside ${\mathcal{X}}_0$, they must be linearly equivalent on the whole
${\mathcal{X}}$ because ${\mathcal{X}}_0$ is irreducible. In particular, they are linearly equivalent when restricted to ${\mathcal{X}}_0$.
Since
$$ (K_{Y}+\mu^{-1}_*D+S)|_S=K_S+\Delta_S\sim_{\mathbb{Q}}A_{X,D}(S) \cdot S|_S,$$
we know that
$$-K_{{\mathcal{X}}}-(D_{\mathbb{A}^1})_{{\mathcal{X}}}|_{{\mathcal{X}}_0}=-K_{\bar{C}}-D_{\bar{C}}\sim_{\mathbb{Q}}(1+A_{X,D}(S))S_{\infty}.$$
Similarly, we have
${\mathcal{L}}|_{{\mathcal{X}}_0}\sim_{{\mathbb{Q}}}\tau S$ with $\tau=\frac{A_{X,D}(S)}{r}$.
Therefore,
\[
1+A_{X,D}(S)=(1+r)\frac{A_{X,D}(S)}{r},
\]
which implies $A_{X,D}(S)=r$ and $\tau=1$.
The degree of $V_{\infty}$ under ${\mathcal{L}}$ is
\begin{eqnarray*}
{\mathcal{L}}|_{{\mathcal{X}}_0}^{n-1}\cdot V_\infty&=&L^{n-1}\cdot V_\infty\\
&=&L^{ n},
\end{eqnarray*}
while the degree of $S$ is
\begin{eqnarray*}
{\mathcal{L}}|_{{\mathcal{X}}_0}^{ n-1}\cdot S &=&\tau^{-1}{\mathcal{L}}_0^{ n}=L^n_t=L_0^n.
\end{eqnarray*}
The restriction $\theta|_{V_{\infty}} \colon V_{\infty}\to S $ is finite since
$$(\rho^*L-\tau \bar{C})|_{V_{\infty}}=L|_{V_{\infty}}$$
is ample. And the degree is 1 by the above calculation on degrees, which implies this is an isomorphism. We claim $Y$ is indeed the $\mathbb{P}^1$-bundle over $V_{\infty}$ induced by blowing up the vertex of $X$, $S$ is a section and the morphism $\theta$ is just contracting the $\mathbb{P}^1$-bundle.
Granted this for now,
we then indeed have an isomorphism from
$(X, D+(1-\beta)V_\infty; L)$
to $(\bar{C}, D_{\bar{C}}+(1-\beta) S_{\infty}; {\mathcal{L}}_0)$.
To see the claim, let $l$ be a curve contracted by $\theta$, we want to show that it is the birational transform of a ruling line of $X$. To see this, since $(\rho^*(L)-\bar{C})\cdot l=0$, we know that $\rho^*(L)\cdot l=1$. So the image $\rho_*l$ of $l$ in $X$ is a line, and it passes through the vertex. Therefore, it is a ruling of the cone.
\end{proof}
By the above proof, let ${\mathcal{V}}$ be the birational transform of $(V_{\infty})_{\mathbb{A}^1}$ on ${\mathcal{X}}$, and ${\mathcal{H}}$ the extension of $H_{\mathbb{A}^1}\setminus\{0\}$ on ${\mathcal{X}}$, we know that:
\[
{\mathcal{X}}={\rm Proj}_{{\mathcal{V}}}\left( \bigoplus_{k=0}^{+\infty} {\mathcal{S}}_k\right),
\]
where
\[
{\mathcal{S}}_k=\bigoplus_{i=0}^{k} (H^0({\mathcal{V}}, i {\mathcal{H}})\cdot u^{k-i}).
\]
From this we easily see that
$S$ and $V$ give the same component over the vertex.
\subsection{The general case}
In this section, we prove Theorem \ref{t-main2} in the general case.
We first show that the cone case we prove in Section \ref{ss-ucone} can be generalized to orbifold cone. Let $T=\mathbb{C}^*$.
\begin{prop}\label{p-Tequiv}
Let $x\in (X,D)$ be a $T$-singularity. Assume a minimizer $v$ of ${\widehat{\rm vol}}_{X,x}$ is given by a rescaling of ${\rm ord}_{F}$ for a Koll\'ar component $F$, then $v$ is $T$-equivariant.
\end{prop}
\begin{proof}
Let $\mathfrak{a}$ be an ideal whose normalized blow up gives the model of extracting the Koll\'ar component $F$ (see the proof of Proposition \ref{p-equality}). Denote the degeneration of $\{\mathfrak{a}_{\bullet}\}:=\{\mathfrak{a}^p\}$ by ${\mathfrak{b}}:=\{{\bf in} (\mathfrak{a}^p)\}$ (which in general is not necessarily equal to but only contains $({\bf in}(\mathfrak{a}))^p$ ).
Since
$${\rm mult}(\mathfrak{a})\cdot {\rm lct}^n(X,D; \mathfrak{a})={\widehat{\rm vol}}({\rm ord}_F)\ge {\rm mult}({\mathfrak{b}}_{\bullet})\cdot {\rm lct}^n(X,D; {\mathfrak{b}}_{\bullet}),$$
But ${\rm mult}(\mathfrak{a})={\rm mult}({\mathfrak{b}}_{\bullet})$ and $ {\rm lct}(X,D,\mathfrak{a})\ge {\rm lct}(X,D,{\mathfrak{b}}_{\bullet})$, we know that
$$ {\rm lct}(X,D,\mathfrak{a})= {\rm lct}(X,D,{\mathfrak{b}}_{\bullet})=\lim_{k\to \infty}{\rm lct}(X,D, \frac{1}{k}{\mathfrak{b}}_k),$$
which we denote by $c$.
In particular, for we can choose $\epsilon$ sufficiently small, and $k$ sufficiently large, such that the log discrepancy
$$a_l(F, X,D+(c-\epsilon)\mathfrak{a})<1 $$
and $({X},D+(c-\epsilon)\frac{1}{k}{\mathfrak{b}}_k)$ is klt.
Thus by \cite{BCHM10} we can construct a model $Z\to W=_{\rm defn} X\times \mathbb{A}^1$ extracting only the irreducible divisor $ F_Z$ which gives $F$ over the generic fiber. Furthermore, we can assume $-F_W$ is ample over $W$ and we denote by $W_0=X\times \{0 \}$.
Therefore, $W$ and $Y\times \mathbb{A}^1$ where $Y=W\times_{\mathbb{A}^1}\{t\}$ are isomorphic incodimension 1, with the exceptional divisors antiample over $X\times \mathbb{A}^1$. Thus we conclude that they are isomorphic.
\end{proof}
\begin{prop}\label{p-cover2}
Under the notation in Section \ref{ss-deformation}, $S$ is the unique minimizer among all Koll\'ar components for ${\widehat{\rm vol}}_{({C},C_{\Delta})}$ if and only if the same holds for ${C}^{(d)}$ on
${\widehat{\rm vol}}_{({C}^{(d)},C^{(d)}_{\Delta}+C^{(d)}_D)}$.
\end{prop}
\begin{proof}By Proposition \ref{p-Tequiv}, any minimizing Koll\'ar component $E$ of ${\widehat{\rm vol}}_{X,D}$ is $T$-equivalent, therefore it is $G=\mathbb{Z}/d$ equivalent. So $E$ is the pull back of a Koll\'ar component on $C^{(d)}$ by Lemma \ref{l-finite}, which can be only the canonical component obtained by blowing up the vertex by our assumption and Lemma \ref{l-finitevolume}.
\end{proof}
\begin{proof}[Proof of Theorem \ref{t-main2}] We first notice that by Theorem \ref{t-cone} and Proposition \ref{p-cover2}, we know that for the coarse moduli space of an orbifold cone over a K-semistable log Fano pair, the only Koll\'ar component which minimizes the normalized volume function is given by the canonical component.
Now we consider the case of a general klt singularity $o\in (X,D)$. Let us assume that there is another component $F$ such that
$${\widehat{\rm vol}}_{X,o}({\rm ord}_S)={\widehat{\rm vol}}_{X,o}({\rm ord}_F),$$
in particular, they take the minimal value of ${\widehat{\rm vol}}_{X,o}$ by Theorem \ref{t-main}.
Let $W\to B$ be the family which degenerates $X$ to $X_0=Y_0\cup \bar{C}$, where $Y_0\cong Y$ extracting $S$ over $X$ and $\bar{C}$ is the the coarse moduli space of the orbifold cone over $S=\bar{C}\cap Y_0$. Then we argue as in the proof Proposition \ref{p-Tequiv}: Let $\mathfrak{a}$ be an ideal whose normalized blow up gives the model of extracting the Koll\'ar component $F$ (see the proof of Proposition \ref{p-equality}). Denote the degeneration of $\{\mathfrak{a}_{\bullet}\}:=\{\mathfrak{a}^p\}$ by ${\mathfrak{b}}_{\bullet}:=\{{\bf in} (\mathfrak{a}^p)\}$.
Since
$${\rm mult}(\mathfrak{a})\cdot {\rm lct}^n(X,\mathfrak{a})={\widehat{\rm vol}}_{X_0}({\rm ord}_F)= {\widehat{\rm vol}}_{Y_0}({\rm ord}_{S_0})\le {\rm mult}({\mathfrak{b}}_{\bullet})\cdot {\rm lct}^n(\bar{C},{\mathfrak{b}}_{\bullet}),$$
where the last inequality is from the assumption that $S_0$ is K-semistable and Theorem \ref{t-main}.
But ${\rm mult}(\mathfrak{a})={\rm mult}({\mathfrak{b}}_{\bullet})$ and $ {\rm lct}(X,\mathfrak{a})\ge {\rm lct}(\bar{C},{\mathfrak{b}}_{\bullet})$, we know that
$$ {\rm lct}(X,\mathfrak{a})= {\rm lct}(\bar{C},{\mathfrak{b}}_{\bullet})=\lim_{k\to \infty}{\rm lct}(\bar{C}, \frac{1}{k}{\mathfrak{b}}_k),$$
which we denote by $c$.
In particular, for we can choose $\epsilon$ sufficiently small, and $k$ sufficiently large, such that the log discrepancy
$$a_l(F, X,D+(c-\epsilon)\mathfrak{a})<1 $$
and $({C},C_D+(c-\epsilon)\frac{1}{k}{\mathfrak{b}}_k)$ is klt.
Thus by \cite{BCHM10} we can construct a model $Z\to W$ extracting only the irreducible divisor $ F_Z$ which gives $F$ over the generic fiber. Furthermore, we can assume $-F_W$ is ample over $W$ and after a base change, we can assume $F_W|_{X_0}$ is reduced.
We claim that $Z_0\to W_0$ also only extracts a Koll\'ar component. In fact, locally over the vertex $v$ of $\bar{C}$, since
$$(-F_W|_{X})^n=(-F_W|_{X_0})^n,$$
then by Lemma \ref{l-inter2},
$${\widehat{\rm vol}}_X({\rm ord}_F)={\widehat{\rm vol}}(Z_0/W_0)= {\widehat{\rm vol}}_{Y_0}({\rm ord}_{S_0}).$$
Thus we conclude that the volume of the model $Z_0$ is equal to ${\widehat{\rm vol}}_{X,o}(F)$, which is equal to the minimum of the normalized volume ${\widehat{\rm vol}}_{\bar{C},v}$. It follows from the argument in the proof of Theorem \ref{t-divisor} that over $v$, $Z_0\to W_0$ yields a Koll\'ar component $F_0$. By the proof in the cone case Theorem \ref{t-cone}, $F_0$ has to be the same as the canonical component $S_0$. In particular, this implies that the birational transform ${\mathbb{P}}$ of $\bar{C}$ in $Z$ is the extract of the canonical component.
Thus there is a morphism $\mathbb{P}\to S$. Let $l$ be the fiber class of $\mathbb{P}\to S$. Consider $K_{Z}+\mu_*^{-1}(D_{\mathbb{A}^1})+F_Z$, which satisfies that
$$\big(K_{Z}+\mu_*^{-1}(D_{\mathbb{A}^1})+F_Z\big)|_{F_0}=K_{F_0}+\Delta_{F_0}$$
is anti-ample,
$$ \big(K_{Z}+\mu_*^{-1}(D_{\mathbb{A}^1})+F_Z\big)|_S=K_{S}+\Delta_S$$
is anti-ample, and
$$\big(K_{Z}+\mu_*^{-1}(D_{\mathbb{A}^1})+F_Z\big)\cdot l=0.$$
Thus $l$ is an extremal ray in $N_1(Z/X_{\mathbb{A}^1})$.
Hence we know that there is a morphism $Z\to W'$ which contracts $\mathbb{P}$ and $W'$ admits a morphism $\chi\colon W'\to X_{\mathbb{A}^1}$. Restricting over $0$, the central fiber is the birational model $\mu\colon Y\to X$ which extracts $S$. On the other hand, let $\mu_F\colon Y_F\to X$ be the birational model which extracts the Koll\'ar component $F$.
As $Y_F\times \mathbb{A}^1$ and $W'$ is isomorphic in codimension 1, if we denote by $F_{W'}$ the push forward of $F_Z$ on $W'$, we have
\begin{eqnarray*}
W' &=&{\rm Proj}\bigoplus_m{\mu_1}_* \mathcal{O}_{W'}\big(-m F_{W'}\big)\\
& =& {\rm Proj}\bigoplus_m{\mu_2}_* \mathcal{O}_{Y_F\times {\mathbb{A}}^1}\big(-m(F\times {\mathbb{A}}^1)\big)\\
&=&Y_F\times {\mathbb{A}}^1.
\end{eqnarray*}
Consider the central fiber over $0$, this implies that $Y_F=Y$.
\end{proof}
\section{Minimizing Koll\'{a}r component is K-semistable}\label{s-Ksta}
In this section, we aim to prove the a Koll\'ar component is minimizing only if it is K-semistable. The method used in the proof of this result is motivated by a method used in the study of toric degenerations (see \cite[Section 3.2]{Cal02}, \cite[Proposition 2.2]{AB04} and \cite[Proposition 3]{And13}). In particular this method allows us to reduce two-step degenerations to a one-step degeneration.
\begin{proof}[Proof of Theorem \ref{t-mintok}]
By Proposition \ref{thm-Ksemi}
we know that the canonical valuation of $({C}^{(d)}, C^{(d)}_{\Delta}+C^{(d)}_D)$ minimizes the normalized local volume if and only if $(S,\Delta_S)$ is K-semistable. Thus by Proposition \ref{p-cover}, to show that $(S,\Delta_S)$ is K-semistable, it suffices to show that the canonical component is a minimizer of ${\widehat{\rm vol}}_{X_0,D_0}$ for $ (X_0,D_0)=_{\rm defn} (C,C_D)$, which is the special degeneration associated to $S$.
Let $(X, D)$ be a $\mathbb{Q}$-Gorenstein klt singularity with $X={\rm Spec}_{\mathbb{C}} (R)$. Assume that $S$ is a Koll\'{a}r component that minimizes ${\widehat{\rm vol}}_{(X, D; o)}$ and appears as the exceptional divisor in a plt blow-up $X'\rightarrow X$. Let $\Delta_S$ be the divisor on $S$ satisfying $K_{X'}+S|_S=K_S+\Delta_S$.
By Theorem \ref{t-approx} (and Lemma \ref{l-Tmini}), to show that $(S, \Delta_S)$ is K-semistable, it suffices to show that
$${\widehat{\rm vol}}_{X_0,D_0}({\rm ord}_S) \le {\widehat{\rm vol}}_{X_0,D_0}({\rm ord}_{F})$$
for any $\mathbb{C}^*$-invariant Koll\'ar component $F$ over $o'\in (X_0,D_0)$.
Let $({\mathcal{Y}}, {\mathcal{E}})$ be the associated special degeneration which degenerates $(X_0, D_0)$ to a pair $(Y_0, E_0)$ where $Y_0$ is an orbifold cone over $(F, \Delta_F)$.
Then we have a $ \mathbb{Z}_{\ge 0}\times \mathbb{Z}_{\ge 0}$-valued function on $R$:
\begin{eqnarray}\label{eq-Z2val}
w: R& \longrightarrow & \mathbb{Z}_{\ge 0} \times \mathbb{Z}_{\ge 0}\\
f & \mapsto &\left( {\rm ord}_S(f), {\rm ord}_F({\bf in}(f)) \right).\nonumber
\end{eqnarray}
We give $ \mathbb{Z}_{\ge 0}\times \mathbb{Z}_{\ge 0}$ the following lexicographic order: $(m_1, u_1)<(m_2, u_2)$ if and only if $m_1<m_2$, or $m_1=m_2$ and $u_1<u_2$. If we denote by $\Gamma$ the valuative monoid and denote the associated graded ring by
\[
{\rm gr}_w R=\bigoplus_{(m, u)\in \Gamma} R_{\ge (m, u)}/ R_{> (m, u)},
\]
then it's easy to see that ${\mathcal{Y}}_0={\rm Spec}_{\mathbb{C}} \left({\rm gr}_w R\right)$.
We will also denote:
\[
A=\bigoplus_{m\in \mathbb{N}} R_{\ge m}/R_{>m}=\bigoplus_{m\in \mathbb{N}} A_m.
\]
Then ${\rm Spec}_{\mathbb{C}} (A)={\mathcal{X}}_0=Y$. Moreover if we define the extended Rees ring of $A$ with respect to the filtration associated to ${\rm ord}_{F}$:
\[
\mathcal{A}':=\bigoplus_{k\in \mathbb{Z}} {\mathcal{A}}_k:=\bigoplus_{k\in \mathbb{Z}} \mathfrak{b}_k t^{-k} \subset A[t, t^{-1}],
\]
where $\mathfrak{b}_k=\{f\in A; {\rm ord}_{F}(f)\ge k\}$. Then the flat family ${\mathcal{Y}}\rightarrow \mathbb{C}$ is given by the ${\rm Spec}_{\mathbb{C}[t]}\left(\mathcal{A}'\right)$. In particular, we have
\[
{\mathcal{A}}' \otimes_{\mathbb{C}[t]}\mathbb{C}[t, t^{-1}]\cong A[t, t^{-1}], \quad {\mathcal{A}}' \otimes_{\mathbb{C}[t]}\mathbb{C}[t]/(t)\cong {\rm gr}_w R.
\]
\bigskip
Pick up a set of homogeneous generators $\bar{f}_1, \dots, \bar{f}_p$ for ${\rm gr}_w R$ with $\deg(\bar{f}_i)=(m_i, u_i)$. Lift them to generators $f_1, \dots, f_p$ for $A$ such that $f_i\in A_{m_i}$.
Set $P=\mathbb{C}[x_1, \dots, x_p]$ and give $P$ the grading by $\deg(x_i)=(m_i, u_i)$ so that the surjective map
$$P\rightarrow {\rm gr}_w R \qquad\mbox{ given by }\qquad x_i\mapsto f_i$$ is a map of graded rings.
Let $\bar{g}_1, \dots, \bar{g}_q\in P$ be a set of homogeneous generators of the kernel and assume $\deg(\bar{g}_j)=(n_j, v_j)$.
Since $\bar{g}_j(\bar{f}_1, \dots, \bar{f}_p)=0 \in {\rm gr}_wR$, it follows
$$\bar{g}_j(f_1, \dots, f_p) \in (A_{n_j})_{>v_j} \qquad \mbox{ for each } j.$$
By
the flatness of ${\mathcal{A}}$ over $\mathbb{C}[t]$, there exist liftings $g_j\in \bar{g}_j+(P_{n_j})_{> v_j}$ of the relation $\bar{g}_j$ such that
$$g_j(f_1, \dots, f_p)=0 \mbox{ for } 1\le j\le q .$$
So $g_j{}'s$ form a Gr\"{o}bner basis of $J$ with respect to the order function ${\rm ord}_V$, where $J$ is the kernel of the
surjection $P\rightarrow A$. In other words, if we let $K=(\bar{g}_1, \dots, \bar{g}_q)$ denote the kernel $P\rightarrow {\mathcal{A}}_0$, then $K$ is the initial ideal of $J$ with respect to the order determined by
${\rm ord}_F$. As a consequence, we have:
\[
{\mathcal{A}}=P[\tau]/(\tilde{g}_1, \dots, \tilde{g}_q),
\]
where $\tilde{g}_j=\tau^{v_j} g_j(\tau^{-u_1} x_1, \dots, \tau^{-u_p} x_p)$.
\bigskip
Now we lift $f_1, \dots, f_p$ further to generators $F_1, \dots, F_p$ of $R$. Then we have:
$$g_j(F_1, \dots, F_p)\in R_{> m_j}.$$ By the flatness of ${\mathcal{R}}$ over $\mathbb{C}[t]$, there exist $G_j\in g_j+P_{> n_j}$ such that
$$G_j(F_1, \dots, F_p)=0.$$ Let $I$ be the kernel of $P\rightarrow R$. Then $G_j{}'s$ form a Gr\"{o}bner basis with respect to the order function $w$ in \eqref{eq-Z2val} and the associated initial ideal is $I$. As a consequence, we have:
\[
{\mathcal{R}}=P[\zeta]/(\tilde{G}_1, \dots, \tilde{G}_q)
\]
where $\tilde{G}_j=\zeta^{n_j} G_j(\zeta^{-m_1} x_1, \dots, \zeta^{-m_p} x_p)$.
In summary we have $(\mathbb{C}^*)^2$ action on $\mathbb{C}^p$ generated by two 1-parameter subgroups $\lambda_0(t)=t^{\bold m}$ and $\lambda'(t)=t^{\bold u}$. $\lambda_0$ degenerates $(X, D)$ to $(X_0, D_0)$ and $\lambda'$ degenerates $(X_0, D_0)$ further to $(Y_0, E_0)$.
\bigskip
We now claim that, for $0<\epsilon\ll 1$, there is a family of one parameter subgroups $\lambda_\epsilon(t)\colon \mathbb{C}^*\to (\mathbb{C}^*)^{2}$ such that the following conditions are satisfied:
\begin{enumerate}
\item $\lambda_\epsilon(t)$ degenerates $X$ to $Y_0$. For this to happen, we need to make sure that the initial term of $G_j$ with respect to the weight function $\pi_\epsilon$ defined by $\lambda_\epsilon (t)$ is exactly $\bar{g}_j$. For the latter condition to hold it suffices:
$$\pi_\epsilon (n'_j, v'_j)> \pi_\epsilon (n_j, v_j) \mbox{\ where \ } (n'_j, v'_j)=\deg(G_j-\bar{g}_j).$$
\item
As $\epsilon\rightarrow 0$, $\lambda_\epsilon\rightarrow \lambda_0$ in the sense that $\lambda_\epsilon(t)\rightarrow \lambda_0(t)$ for any $t\in \mathbb{C}^*$.
\item
For $0<\epsilon\ll 1\in {\mathbb{Q}}$, $S_\epsilon=Y_0/\lambda_\epsilon(t)$ is a Koll\'{a}r component over $(X,o)$. Moreover as $\epsilon\rightarrow0$, ${\rm ord}_{T_\epsilon}\rightarrow {\rm ord}_{S}$.
\end{enumerate}
Denote by $B\subset \mathbb{Z}\times \mathbb{Z}$ be the finite set consisting of the differences $(n'_j, v'_j)-(n_j, v_j)$, together with $0$ and the two generators of $\mathbb{N}\times\mathbb{N}$. Let $M$ be a positive integer that is larger than all coordinates of $(m, u)-(n, v)$ for all pairs of elements $(m, u), (n, v)\in B$ and let $\epsilon$ be sufficiently small such that $1> M \epsilon$.
Define
$$\pi_\epsilon=e^*_0-\epsilon e^*_{1}.$$
Then for $\epsilon$ sufficiently small, $\pi_\epsilon$ is a linear projection $\pi: \mathbb{Z}\times \mathbb{Z}\rightarrow \mathbb{Z}$ such that $0<\pi(n_j, v_j)<\pi(n'_j, v'_j)$.
So by choosing $M\gg 1$ so that $\epsilon\ll 1$ we see that indeed there is a family of linear projections $\pi_\epsilon$ approaches $\pi_0$ which corresponds to ${\rm ord}_S$. We can define $\lambda_\epsilon(t)$ to be one parameter subgroup corresponding to $\pi_\epsilon$.
Let $\lambda_\epsilon=\lambda_\epsilon(t)$ be the $\mathbb{C}^*$-action induced by a choice of rational $\epsilon$. Now we claim that $Y_0/(\lambda_\epsilon(t))$ yields a Koll\'{a}r component $S_\epsilon$ over both $(X, o)$ and $(X_0, D_0)$. Moreover the associated special degeneration has the central fibre equal to $(Z, \lambda_\epsilon)$.
Because ${\rm ord}_{S}$ is a minimizer on ${\widehat{\rm vol}}_{X,D}$, we have:
\[
{\widehat{\rm vol}}_{X,D}({\rm ord}_{S_\epsilon})\ge {\widehat{\rm vol}}_{X,D}({\rm ord}_{S}).
\]
On the other hand, we have
\[
{\widehat{\rm vol}}_{X,D}({\rm ord}_{S_\epsilon})={\widehat{\rm vol}}_{X_0,D_0}({\rm ord}_{S_\epsilon})={\widehat{\rm vol}}_{Y_0,E_0}({\rm ord}_{S_\epsilon}).
\]
It's known that ${\widehat{\rm vol}}_{Y_0, E_0}$ is a convex function with respect to $\epsilon$. So we conclude
${\widehat{\rm vol}}_{Y_0, E_0}({\rm ord}_F)\ge {\widehat{\rm vol}}_{Y_0, E_0}({\rm ord}_{T_0})$. As a consequence we have:
\[
{\widehat{\rm vol}}_{X_0, D_0}({\rm ord}_F)={\widehat{\rm vol}}_{Y_0, E_0}({\rm ord}_F)\ge {\widehat{\rm vol}}_{Y_0, E_0}({\rm ord}_{T_0})={\widehat{\rm vol}}_{X_0, D_0}({\rm ord}_S).
\]
To see the construction of $S_\epsilon$,
we define a filtration:
\begin{eqnarray*}
{\mathcal{F}}^N R&=&{\rm Span}_{\mathbb{C}}\left\{f_1^{a_1}\dots f_p^{a_p}; \pi_\epsilon \left( \sum_{i=1}^p a_i (m_i, u_i) \right) \ge N\right\}\\
&=& \{ f; \text{ there exists } F\in P \text{ such that } F|_X=f \text{ and } \lambda_\epsilon (F)\ge N \}.
\end{eqnarray*}
Then $\{{\mathcal{F}}^N R\}$ is the weighted filtration induced by the weighted blow up $\widehat{\mathbb{C}}^p\rightarrow \mathbb{C}^p$ and the associated graded ring of the above filtration is isomorphic to ${\rm gr}_w R$ with the grading given by the weight function $\lambda_\epsilon$. Denote the strict transform of $X$ by $\hat{X}$. Then the exceptional divisor $\hat{X}\rightarrow X$ is isomorphic to $S_\epsilon=Y_0/\lambda_\epsilon$ by the discussion in Section \ref{sec-filtration}. By Proposition \ref{prop-Kolseq}, $(S_\epsilon, \Delta_{\epsilon})=(Y_0, E_0) / \lambda_\epsilon$ is a klt log-Fano-variety and a Koll\'{a}r component over $(X, D; o)$.
\end{proof}
\begin{prop}\label{prop-Kolseq}
For any $0<\epsilon\ll 1$ with $\epsilon\in \mathbb{Q}_{+}$, let $(S_\epsilon, \Delta_\epsilon)=Y_0/\lambda_\epsilon$. Then $S_\epsilon$ is a Koll\'{a}r component over $o\in (X,D)$.
\end{prop}
\begin{proof}
For $0<\epsilon\ll 1$ with $\epsilon\in \mathbb{Q}_{+}$, $\xi_\epsilon$ generates a $\mathbb{C}^*$-action. We have a log orbifold $\mathbb{C}^*$-bundle
$\pi: (Y_0^{\circ}, E_0^{\circ}):=(Y_0\setminus \{v\}, E_0\setminus \{v\}) \rightarrow (S_\epsilon, \Delta_\epsilon)$ where $v$ is the vertex of $Y_0$. The Chern class of this orbifold $\mathbb{C}^*$-bundle, denoted by $c_1(Y_0^{\circ} /S_\epsilon)$, is contained in ${\rm Pic}(S_\epsilon)$ and is ample. $Y_0^\circ$ can be compactified by adding the zero section $S_\epsilon$ so that we get a birational morphism $\mu: Y_\epsilon\rightarrow Y_0$ with the exceptional divisor isomorphic to be $S_\epsilon$.
Because $Y_0$ has klt singularity, by \cite[40-42]{Kol04} (see also \cite{BG08}) we know that $c_1(Y_0^{\circ}/S_\epsilon)=-r^{-1} (K_{S_\epsilon}+\Delta_\epsilon)$ and $(S_\epsilon, \Delta_\epsilon)$ has klt singularities.
So $(S_\epsilon, \Delta_\epsilon)$ is a Koll\'{a}r component over $v\in (Y_0, E_0)$.
To transfer this to $(X, o)$, we notice that the graded ring of $w_\epsilon$ is isomorphic to ${\rm gr}_{{\rm wt}_\epsilon} \mathbb{C}[Y_0]$. The exceptional divisor of the filtered blow-up associated to $w_\epsilon$ is isomorphic to ${\rm Proj}({\rm gr}_{{\rm wt}_\epsilon} \mathbb{C}[Y_0])$ which is isomorphic to $(S_\epsilon, \Delta_\epsilon)$. Since $(S_\epsilon, \Delta_\epsilon)$ is klt, by inversion of adjunction we know that the filtered blow up is indeed a plt blow up and hence $S_\epsilon$ is a Koll\'{a}r component over $(X, o)$.
\end{proof}
\section{Examples}\label{s-exam}
In this section, we find out the minimizer for some examples of klt singularities $(X,o)=(\mathrm{Spec} R,\mathfrak{m})$. We note that by Proposition \ref{p-inf} and \ref{p-equality}, this also explicit calculates
$$\inf_{\mathfrak{a}} {\rm lct}(X,\mathfrak{a})^n\cdot {\rm mult}(\mathfrak{a})$$ for all $\mathfrak{m}$-primary ideals $\mathfrak{a}$ and gives the equality condition, which generalizes the results in \cite{dFEM04} on a smooth point.
\begin{exmp}\label{e-example}
We explicitly compute the minimizer for quotient, $A_k$, $E_k$ and weakly exceptional singularities in the below.
\end{exmp}
\begin{enumerate}
\item (cf. \cite[Example 4.9]{LL16}) Let $(X,o)=(\mathbb{C}^n,0)/G$ be an $n$-dimensional quotient singularity. Let $E\cong \mathbb{P}^{n-1}$ be the exceptional divisor over $\mathbb{C}^n$ obtained by blowing up $0$. Then denote by $S$ the valuation over $(X,o)$ which is the quotient of $E$ by $G$. Applying Lemma \ref{l-finite} to the pull back of Koll\'ar components on $X$, we know that
$${\widehat{\rm vol}}_{X,o}({\rm ord}_S)\le {\widehat{\rm vol}}_{{\rm ord}_E}$$ for any Koll\'ar component $E$ over $(X,o)$. So ${\widehat{\rm vol}}_{X,o}$ minimizes at ${\rm ord}_S$ with
$${\widehat{\rm vol}}_{X,o}({\rm ord}_S)=\frac{n^n}{|G|}.$$
\item
Consider the $n$-dimensional $A_{k-1}$ singularity:
$$
X=A^{n}_{k-1}:=\{z_1^2+\cdots+z_n^2+z_{n+1}^k=0\}.
$$
We consider cases when $k>\frac{2(n-1)}{n-2}$ (for other cases, see \cite[Example 4.7]{LL16}). We want to show that the valuation corresponding to the weight
$w_*=(n-1, \cdots, n-1, n-2)$ is a minimizer among all valuations in ${\widehat{\rm vol}}_{X,o}$. In \cite[Example 2.8]{Li15a}, these are computed out as the minimizer among all valuations obtained by weighted blow ups on the ambient space $\mathbb{C}^{n+1}$.
We notice that under the weighted blow up corresponding to $w_*$, we have a
birational morphism $Y\rightarrow X$ with exceptional divisor $S$ isomorphic to the weighted hypersurface
$$
S:=\{Z_1^2+\cdots+Z_n^2=0\}\subset {\mathbb{P}}(n-1, \cdots, n-1, n-2)=:{\mathbb{P}}_{w_*}.
$$
Because ${\mathbb{P}}_{w_*}\cong {\mathbb{P}}(1,\cdots, 1, n-2)$, it is easy to see that $S$ is isomorphic to $\bar{C}(Q, -K_Q)$ where
$Q=Q^{n-2}=\{Z_1^2+\cdots+Z_n^2=0\}\subset{\mathbb{P}}^{n-1}$ (notice that $K_{Q}^{-1}=(n-2)H$). On the other hand, because ${\mathbb{P}}_{w_*}$ is not well-formed, we have
codimensional 1 orbifold locus along the infinity divisor $Q_\infty\subset S$ with the isotropy group $\mathbb{Z}/(n-1)\mathbb{Z}$. So the corresponding Koll\'{a}r component
is the log Fano pair $\left(S, (1-\frac{1}{n-1})Q_\infty\right)$. Because $Q_\infty$ has KE, by \cite{LL16} there is a conical KE on the pair $\left(S, (1-\frac{1}{n-1})Q_\infty\right)$.
So by Theorem \ref{t-main} and Theorem \ref{t-main2}, ${\rm ord}_S$ is indeed a global minimizer of ${\widehat{\rm vol}}$ that is the unique minimizer among all Koll\'{a}r components.
Notice that for any higher order klt perturbation of these singularities, $w_*$ is also a minimizer.
\item
We can also use Theorem \ref{t-main} to verify that the valuations in \cite[Example 2.8]{Li15a} for $E_k$ (k=6,7,8) are indeed minimizers in ${\widehat{\rm vol}}_{X,o}$, which are unique among Koll\'{a}r components. To avoid repetition, we will only do this for $E_7$
singularities. The argument for other two cases are similar. So consider the $(n+1)$-dimensional $E_7$ singularity:
\[
X^{n+1}=\{z_1^2+z_2^2+\cdots+z_n^2+z_{n+1}^3z_{n+2}+z^3_{n+2}=0\}\subset\mathbb{C}^{n+2}.
\]
\begin{enumerate}
\item If $n+1=2$, then $X^2$ is a quotient singularity $\mathbb{C}^2/E_7$ and so we get the unique polystable component by \cite[Example 4.9]{LL16} and example 1 above.
\item If $n+1=3$, then $X^3=\{z_1^2+z_2^2+z_3^3z_4+z_4^3=0\}\subset \mathbb{C}^4\cong \{w_1 w_2+w_3^3w_4+w_4^3=0\}\subset \mathbb{C}^4$ by the change of variables. This singularity has a
$(\mathbb{C}^*)^2$-action and is an example of $T$-variety of complexity one. By the recent work in \cite[Theorem 7.1 (II)]{CoSz16}, $X^3$ indeed has a Ricci flat cone K\"{a}hler metric
associated to the canonical $\mathbb{C}^*$-action associated to $w_*$. So by \cite[Theorem 1.7]{LL16}, the unique K-polystable Koll\'{a}r component is given by the orbifold $X^3/\langle w_*\rangle$.
\item $n+1=4$, then under the weighted blow up corresponding to $w_*=(9,9,9,5,6)$, we have a birational morphism
$\hat{X}\rightarrow X$ with exceptional divisor $E$ isomorphic to the weighted hypersurface
$$
E=\{z_1^2+z_2^2+z_3^2+z_5^3=0\}\subset {\mathbb{P}}(9, 9, 9, 5, 6)={\mathbb{P}}(w_*).
$$
Since ${\mathbb{P}}(w_*)$ is not well-formed, we have:
$$
E\cong \{z_1^2+z_2^2+z_3^2+z_5^3=0\}\subset {\mathbb{P}}(3, 3, 3, 5, 2)={\mathbb{P}}'.
$$
with orbifold locus of isotropy group $\mathbb{Z}/3\mathbb{Z}$ along $$
V=\{z_1^2+z_2^2+z_3^2+z_5^3=0\}\subset {\mathbb{P}}(3,3,3,2).
$$
Alternatively, $E$ is a weighted projective cone over the weighted hypersurface. It is easy to see that as an orbifold $(V, \Delta) \cong \left({\mathbb{P}}^2, (1-\frac{1}{3})Q\right)$ where $Q=\{z_1^2+z_2^2+z_3^2=0\}\subset {\mathbb{P}}^2$.
By \cite{LS14}, there exists an orbifold K\"{a}hler-Einstein metric on $(V, \Delta)$.
Notice that $-(K_V+\Delta)=3L-\frac{4}{3}L=\frac{5}{3}L$ where $L$ is the hyperplane bundle of ${\mathbb{P}}^2$. Denoting by $H$ the hyperplane bundle of ${\mathbb{P}}'$, then
$H|_V=L/3$. If $V$ is considered as a divisor of $E$, then
$$V|_V=\big(\{z_4=0\}\cap E\big)=5H|_V=\frac{5}{3}L.$$ So $-(K_V+\Delta)= V|_V$.
Then by \cite[Theorem 1.7]{LL16}, there exists an orbifold K\"{a}hler-Einstein metric on $E$ because the cone angle at infinity is $\beta=1/3$. Thus the unique log-K-semistable (actually log-K-polystable) Koll\'{a}r component is given by the pair $\left(E, \left(1-\frac{1}{3}\right)V\right)$.
\item
$n+1=5$, under the weighted blow up corresponding to $w_*=(3, 3, 3, 3, 2, 2)$, we have a birational morphism $\hat{X}\rightarrow X$ with exceptional divisor $E$ isomorphic to the
weighted hypersurface:
\[
E=\{z_1^2+z_2^2+z_3^2+z_4^2+z_6^3=0\}\subset {\mathbb{P}}(3, 3, 3, 3, 2, 2)=:{\mathbb{P}}(w_*).
\]
This is a weighted projective cone over the weighted hypersurface:
\[
V=\{z_1^2+z_2^2+z_3^2+z_4^2+z_6^3=0\}\subset {\mathbb{P}}(3, 3, 3, 3, 2).
\]
As orbifold, we have $(V, \Delta)=\left({\mathbb{P}}^3, (1-\frac{1}{3})Q\right)$. By \cite{LS14, Li13}, $(V, \Delta)$ is log-K-semistable and degenerates to a conical K\"{a}hler-Einstein pair. So by \cite{LL16}, we know
that $(E, (1-\beta)V_\infty)$ is log-K-semistable. To determine $\beta$, we notice that
$$-(K_V+\Delta)=4L-\frac{4}{3}L=\frac{8}{3}L=4\cdot \frac{2}{3}L=4\cdot V_\infty|_V.$$
So $\beta=1$ and we conclude that the unique (strictly) K-semistable Koll\'{a}r component is indeed the ${\mathbb{Q}}$-Fano variety $E$.
\item $n+1\ge 6$. Under the weighted blow up corresponding to $w_*=(n-1, \dots, n-1, n-2, n-2)$, we have a birational morphism $\hat{X}\rightarrow X$ with exceptional divisor $E$ isomorphic to
the weighted hypersurface:
\[
E=\{z_1^2+\cdots+z_n^2=0\}\subset {\mathbb{P}}(n-1, \cdots, n-1, n-2, n-2)=:{\mathbb{P}}(w_*).
\]
This is the weighted projective cone over
\[
V=\{z_1^2+\cdots+z_n^2=0\}\subset {\mathbb{P}}(n-1, \cdots, n-1, n-2).
\]
By the discussion in the above $A^n_{k-1}$ singularity case, we know that as an orbifold,
$(V, \Delta)=\left(\bar{C}(Q, -K_Q), (1-\frac{1}{n-1})Q_\infty\right)$, which has an orbifold K\"{a}hler-Einstein metric. Notice that
\[
-(K_V+\Delta)=(n(n-1)+n-2)H|_V-2(n-1)H|_V=n (n-2)H|_V.
\]
By \cite[Theorem 1.7]{LL16}, the ${\mathbb{Q}}$-Fano variety $E$ indeed has an orbifold K\"{a}hler-Einstein metric ($\beta=n/n=1$ at infinity) and hence by Theorem \ref{t-main} is the unique K-semistable (actually K-polystable) Koll\'{a}r component.
\end{enumerate}
We remark that, however, in the case of $D_{k+1}$ singularities, since the valuations computed out in \cite[Example 2.8]{Li15a} could be irrational, our method can not directly tell whether it is a minimizer in ${\rm Val}_{X,o}$.
\item A notion called weakly-exceptional singularity is introduced in \cite{Pro00}. As the name suggested, this is a weaker notion than the exceptional singularity introduced in \cite{Sho00}, which forms a special class of singularities in the theory of local complements. In our language, a singularity $(X,o)$ is {\it weakly-exceptional} if and only if it has a unique Koll\'ar component $S$. We know that if a singularity is weakly-exceptional, then the log $\alpha$-invariant for the log Fano $(S,\Delta_S)$ is at least 1 (see \cite[Theorem 4.3]{Pro00} and \cite{CS14}). In particular, we know that $(S,\Delta_S)$ is K-semistable (see \cite[Theorem 1.4]{OS14} or \cite[Theorem 3.12]{Ber13}). And by Theorem \ref{t-main} and \ref{t-main2}, we know ${\rm ord}_S$ is the unique minimizer of ${\widehat{\rm vol}}(S)$ among all Koll\'ar components. See \cite{CS14} for examples of weakly exceptional singularities.
\end{enumerate}
\bigskip
The following example is indeed a prototype of our study.
\begin{exmp}\label{ex-irreg}
If a log terminal singularity $o\in X$ has a quasi-regular Sasakian-Einstein metric, then the $\mathbb{C}^*$-quotient provides a Koll\'ar component which minimizes ${\widehat{\rm vol}}_{x,X}$.
\end{exmp}
Assume $X={\rm Spec} R$ is an affine variety with effective torus action and ${\mathbb{Q}}$-Gorenstein klt isolated singularity at $o\in X$. Denote the torus by ${\mathbb{T}}:=({\mathbb{G}}_m)^d$.
Then we have a weight decomposition:
\[
R=\bigoplus_{\alpha\in \Gamma} R_\alpha,
\]
where
\begin{itemize}
\item
For any $\alpha\in \mathbb{Z}^d$, $R_\alpha=\{f\in R\; ; \; t\cdot f=t^{\alpha}f \text{ for any } t\in ({\mathbb{G}}_m)^d\}$.
\item
$\Gamma=\{\alpha\in \mathbb{Z}^d; R_\alpha\neq 0\}$.
\end{itemize}
The Reeb cone of $X$ is the following conic subset of the Lie algebra of ${\mathbb{T}}=({\mathbb{G}}_m)^d$:
$$
\mathcal{RC}:=\{
\xi\in \mathbb{R}^d; \langle \alpha, \xi\rangle>0 \text{ for any } \alpha\in \Gamma
\}.
$$
Notice that the elements in the Reeb cone $\mathcal{RC}$ can be considered as holomorphic vector fields on $X$ via the map
$\xi\mapsto \sigma_\xi:=\sum_{i=1}^d \xi_i \sigma_i$
where $\sigma_i$ is the infinitesimal generator of the action of the $i$-th factor of $({\mathbb{G}}_m)^d$.
If we denote by $\{e_i\}$ the standard basis of $\mathbb{R}^d$, then
\[
\sigma_i(p)=\left.\frac{d}{ds}\right|_{s=0}\exp (s e_i)\circ p - \sqrt{-1} \left.\frac{d}{ds}\right|_{s=0}\exp(\sqrt{-1} s e_i) \circ p.
\]
Suppose that there is a K\"{a}hler cone metric $g_\xi=dr^2+r^2 g_{L}$ on $X$ where
\begin{enumerate}
\item
$r: X-\{o\}\rightarrow \mathbb{R}_{>0}$ is a smooth radius function, $L=\{r=1\}\subset X$ is the link of the isolated singularity $(X,o)$.
\item $r\frac{\partial}{\partial r}- \sqrt{-1} J(r\frac{\partial}{\partial r})=\sigma_\xi$.
\end{enumerate}
Now assume that $g_\xi$ is a Ricci-flat K\"{a}hler metric, equivalently the link $L$ has a Sasaki-Einstein metric (see \cite{BG08}). There are two different cases:
\begin{enumerate}
\item (quasi-regular case)
If the generic orbit of $\partial_\theta:=J(r\frac{\partial}{\partial r})$ is closed, $g_L$ is called a quasi-regular Sasaki-Einstein metric. $\sigma_\xi$ generates a ${\mathbb{G}}_m$-subgroup of ${\mathbb{T}}$. the quotient $X/{\mathbb{G}}_m$ is an Fano orbifold $(V, \Delta)$ with orbifold K\"{a}hler-Einstein metric. By \cite{LL16}
${\rm ord}_V$ is a minimizer of ${\widehat{\rm vol}}_{X,o}$. By Theorem \ref{t-main2}, this is the unique minimizer among all Koll\'{a}r components.
\item (irregular case)
If the generic orbit of $\partial_\theta$ is not closed, $g_L$ is called an irregular Sasaki-Einstein metric. Explicit examples can be found in \cite{GMSW04a, GMSW04b, FOW09}.
Since $\xi\in \mathbb{R}^d$ is an irrational vector, we can approximate it by a sequence of rational vectors $\{\xi_k\}\subset \mathcal{RC}\cap {\mathbb{Q}}^d$. Each $\xi_k$ generates a ${\mathbb{G}}_m$-subgroup $G_k$ of ${\mathbb{T}}$, and we get a sequence of quotients $X/G_k=(V_k, \Delta_k)$. It is easy to see that $V_k$ are all Koll\'{a}r components. It is possible to show that a suitable rescaling $c_k\cdot {\rm ord}_{V_k}$ converges to a minimizer of ${\widehat{\rm vol}}_{(X,o)}$, which corresponds to the vector field $\sigma_\xi$. The details will be discussed in a future work.
\end{enumerate}
\bigskip
As we mentioned, there is a related differential geometry study on the metric tangent cone.
\begin{exmp}\label{ex-metric}
In the notation of \cite{DS15}, let $p\in Z$ be a singularity appearing on the Gromov-Hausdorff limit of a sequence of KE Fano manifolds. Assume the Reeb vector of the weight tangent cone $W$ is quasi-regular, then the $\mathbb{C}^*$-quotient of $W$ indeed gives a minimizer of $p\in Z$.
\end{exmp}
Assume $(X, o)$ is an algebraic germ on a normal affine variety that is embedded into $\mathbb{C}^N$. Assume that we have a weight vector $w=(a_1, \dots, a_N)\in \mathbb{Z}_{>0}^N$ satisfying that there is no common factor of $a_i$. Then it defines a filtration on $R=\mathcal{O}_{X,o}$:
\begin{equation}\label{eq-wf}
{\mathcal{F}}^k=\{f\in R; \text{ there exists } g\in \mathcal{O}_{\mathbb{C}^N,o} \text{ with } g|_X=f \text{ and } w(g)\ge k\}.
\end{equation}
From this we can form the Rees algebra and extended Rees algebra:
\[
{\mathcal{R}}=\bigoplus_{k=0}^{+\infty} {\mathcal{F}}^k t^{-k}, \quad {\mathcal{R}}'=\bigoplus_{k\in \mathbb{Z}} {\mathcal{F}}^k t^{-k},
\]
where ${\mathcal{F}}^k=R$ if $k\le 0$. We also have the associated graded ring of $(R, {\mathcal{F}})$:
\[
G=\bigoplus_{k=0}^{+\infty} \left({\mathcal{F}}^k/{\mathcal{F}}^{k+1}\right) \cdot t^{-k}\cong {\mathcal{R}}'/t{\mathcal{R}}'.
\]
In our discussion, we always assume that $G$ is a finitely generated $\mathbb{C}$-algebra that is also a normal domain. Then we can get a normal affine variety $W={\rm Spec}(G)$. As in \cite{DS15}, we call $W$ to be a weighted tangent cone of $(X, o)$. Notice that there is a natural $\mathbb{C}^*$-action on $G$ and hence it is an orbifold cone over the base $E={\rm Proj}(G)$ (by Pinkham-Demazure's construction).
On the other hand, the weight $w$ defines a weighted blow up $\pi_w: \widehat{\mathbb{C}^N} \rightarrow \mathbb{C}^N$ given by:
\[
\widehat{\mathbb{C}^N}={\rm Proj} \left(\bigoplus_{k=0}^{+\infty} \mathfrak{A}_k\cdot t^{-k}\right),
\]
where
\[
\mathfrak{A}_k=\{g\in \mathbb{C}[z_1, \cdots, z_n]; w(g)\ge k\}.
\]
Notice that the exceptional divisor of $\pi_w$ is the weighted projective space with weight $w$. We have the following observation, which implies that the discussion can be put in the setting of
filtered blowing-ups as studied in \cite{TW89}.
\begin{lem}
\begin{enumerate}
\item
If $Y$ denotes the strict transform of $X$ under $\pi_w$, then
$Y={\rm Proj}({\mathcal{R}})$. In particular the Rees algebra ${\mathcal{R}}$ is finitely generated. Moreover, the exceptional divisor of the natural birational morphism
$Y\stackrel{\psi}{\longrightarrow} X$ is isomorphic to $E$.
\item
The filtration ${\mathcal{F}}=\{{\mathcal{F}}^k\}_{k\ge 0}$ is the same as the filtration induced by the divisorial valuation ${\rm ord}_E$.
\end{enumerate}
\end{lem}
\begin{proof}
Denoting by $\iota_X: X\rightarrow \mathbb{C}^N$ the embedding morphism, then ${\mathcal{F}}^k$ in \eqref{eq-wf} coincides with the inverse image ideal $\iota_X^{-1}\mathfrak{A}_k\cdot \mathcal{O}_{X,o}$ (cf. \cite[7.15]{Har77}). So $Y$ is indeed
the strict transform of $X$ under the weighted blow up $\pi_w$. The second statement is standard (see \cite[1.2.1]{TW89}). Item 2 follows from the discussion in Section \ref{sec-filtration} (see \cite[(2.2)]{TW89}).
\end{proof}
If $(X,o)$ is a klt singularity appearing in a Gromov-Hausdorff limit of a sequence of K\"{a}hler-Einstein Fano manifolds, Donaldson-Sun (\cite{DS15})
constructed an affine variety $W$ using the metric structure on the Gromov-Hausdorff limit and showed that it specially degenerates to the metric tangent cone $C(Y)$. The degeneration can be realized under a common embedding of $W$ and $C(Y)$ into some ambient $\mathbb{C}^N$. More precisely $W$ is a weighted tangent cone associated to some weight under the common embedding. The weight is determined by the Reeb vector field of the singular Ricci flat metric on $C(Y)$. If the Reeb vector field is quasi-regular, that is, if it generates a $\mathbb{C}^*$-action, then the weight can be normalized to have integer components so that we are in the situation discussed above. Moreover it is shown in this case $E=W\setminus \{ 0 \}/\mathbb{C}^*$ specially degenerates to $C(Y)\setminus \{ 0\}/\mathbb{C}^*$ and $C(Y)\setminus \{ 0\} /\mathbb{C}^*$ admits a weak K\"{a}hler-Einstein metric (see \cite{DS15}). This implies $E$ is analytically K-semistable, which then implies it is (algebraically) K-semistable. So we can apply our result Theorem \ref{t-main} to see that ${\rm ord}_E$ is indeed a global minimizer of ${\widehat{\rm vol}}$ over $(X, o)$, which is the unique minimizer among all Koll\'{a}r components.
|
\section{\label{sec:level1}INTRODUCTION}
\section{\label{sec:level1}INTRODUCTION}
Two-dimensional layered materials are becoming the focus of experimental and theoretical investigations aiming to realize the potential applications of these atomically thin structures\cite{2dmaterial_rev1,2dmaterial_rev2,2dmaterial_rev3}. The library of these layered materials is still expanding with properties that range from metals and semi-metals to semiconductors and insulators. Graphene, a semi-metallic atomically-thin sheet of carbon atoms in the honeycomb lattice\cite{graphene_2004,graphene_rev}, is an important member of the layered materials family. Flakes of graphene can be obtained by the exfoliation method from a graphite crystal\cite{graphene_exfoliation} or synthesized by methods such as chemical vapor deposition\cite{graphene_CVD}. In addition to its outstanding electronic and mechanical properties, graphene is also an interesting platform to investigate the quasi-relativistic strongly-interacting many-body physics near the charge-neutrality point (CNP) when screening between charges is reduced\cite{graphen_dirac_fluid}. This electron-hole plasma, known as the Dirac fluid, behaves differently from the conventional Fermi liquid, when the Fermi level is far from the CNP. The inclusion of graphene in the van der Waals (vdW) heterostructures, that is, stacks of various layered materials, can serve several purposes, for instance as an active layer, a spacer or an electrode. The different layered materials exhibit a wide variety of physical properties such as topological phases\cite{TMDC_TI}, superconductivity\cite{TMDC_sc}, magnetism\cite{TMDC_magnetism} and charge density waves\cite{TMDC_cdw}. Different ways of stacking, manipulating these materials and intercalating with foreign atoms in the vdW heterostructure open even wider possibilities for interesting physics phenomena and for novel nano-device applications\cite{2dmaterial_rev2,layer_pseudospin}. Other control knobs also include electrical gating, an external magnetic field, various contacts and twist angles which affect strongly the Brillouin zone (BZ) alignment and the coupling between layers\cite{twist_graphene_review}.
To investigate the electronic properties of these fascinating systems, large-scale density functional theory (DFT)\cite{graphene_sc_dft1,graphene_sc_dft2}, tight-binding models\cite{graphene_sc_tbh1,graphene_sc_tbh2,graphene_sc_tbh3,graphene_sc_tbh4}, and low-energy k $\cdot$ p expansions\cite{graphene_sc_kp1,graphene_sc_kp2,graphene_sc_kp3,graphene_sc_kp4} have been employed. The parameter-free DFT approach is computationally demanding, while the computationally efficient tight-binding methods and low-energy k $\cdot$ p expansions are hampered by the absence of a universal form of the interlayer couplings. The interlayer couplings employed by empirical methods are often parametrized as functions of only the interatomic distances or more elaborate forms that depend on the local bonding environment\cite{graphene_sc_tbh4,carbon_bond} with the values of parameters obtained by fitting the band structure of selected crystal configurations. A set of such interlayer hopping terms for graphene has been determined from {\it ab initio} calculations, but they are extracted only from a restricted {\em subset} of all possible bilayer orientations\cite{macdonald_graphene_bilayer,macdonald_moire}. The dependence of interlayer hopping on both the distance between pairs of atoms and the relative orientation of bonds, as exemplified by the $\gamma_3$ and $\gamma_4$ terms in the Slonczewski-Weiss-McClure model\cite{graphene_rev}, have not yet been addressed properly. An accurate and transferable theory of interlayer coupling would not only provide an efficient way of evaluating electronic properties, but would also shed light on transport properties across layers\cite{moire_transport} and on the derivation of effective low-energy theories for arbitrary graphene stacking sequences.
We provide here a comprehensive and quantitive understanding of interlayer coupling in two prototypical bilayers, graphene-graphene (G-G) and graphene-hexagonal born nitride (G-hBN). For the first time, this type of {\it ab initio} modeling based on the Wannier transformation is applied to derive a transferable potential applicable to bilayer configurations with arbitrary translations and twists between the two layers. In contrast to similar analysis applied to transition metal dichalcogenides, in which interlayer coupling was shown to have a simple, orientation-independent scaling form that depends only on the distance between pairs of atoms\cite{tmdc_tbh}, the G-G and G-hBN interlayer couplings include a dependence on both pair distances and relative orientations. The ensuing angular dependence is related to the crystal field distortions of the atomic $p_z$ Wannier orbitals. We derive the form of effective coupling terms that is suitable in the general vdW heterostructure with arbitrary twists and translations when some layers are integrated out. This scheme is relevant for obtaining the proximity effects on graphene due to other layered materials in the vdW heterostructure. These interlayer coupling models provide efficient ways for obtaining electronic properties and effective low-energy theories, as well as for estimating proximity effects in vdW heterostructures, especially when graphene layers are included in the stacks.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{FIG1}
\caption{(a) The generic graphene bilayer configuration with arbitrary translations and twists. The constituent monolayer crystal is described by the primitive vectors $\vec{a}_1$ and $\vec{a}_2$. For an interlayer pair, the coupling dependence is characterized by the projected distance $r$ and the angles $\theta_{12}$ and $\theta_{21}$ between $\vec{r}$ and the nearest neighbor bonds. (b) Decomposition of the Wannier function for monolayer graphene into the constant ($m=0$), $\cos(3\theta)$ ($m=\pm 3 $) and $\cos(6\theta)$ ($m=\pm 6 $) components.}
\label{fig:Wannier_orbital}
\end{figure}
The paper is structured as follows: In Sec. II, we introduce the numerical methods employed for the DFT calculations and the Wannier transformation. In Sec. III, we investigate the hamiltonians for weakly interacting bilayers, using G-G and G-hBN interfaces as the prototypical examples. In Sec. IV we discuss the physics of twisted bilayer graphene and in Sec. V we derive the effective theory for proximity effects by integrating out the neighboring layers. Our concluding Sec. VI gives a summary of the main points, makes comparisons with similar approaches in the literature, and contains some remarks on possible extensions and future applications.
\section{\label{sec:level1} NUMERICAL METHODS}
The approach we adopt here is to derive the {\it ab initio} tight binding hamiltonian based on the Wannier transformation of DFT calculations. Within DFT, we obtain the Bloch wavefunctions and energies using VASP\cite{vasp1,vasp2} with pseudo-potentials of the Projector Augmented-Wave (PAW) type, the exchange-correlation functional of Perdew, Burke and Ernzerhof (PBE)\cite{pbe}, a plane-wave energy cutoff 500 eV and a 17 $\times$ 17 $\times$ 1 reciprocal space grid. A $20$ \AA $ $ distance is used to eliminate the coupling between periodic images of the layers in the direction perpendicular to the atomic planes. The diagonal Kohn-Sham hamiltonian in Bloch basis from the DFT calculations is then transformed into a basis of maximally-localized Wannier functions (MLWF)\cite{mlwf} implemented in the Wannier90 code. In our modeling, only $p_z$-like orbitals at each atomic site are projected out and retained in the Wannier basis. The short-ranged {\it ab initio} tight-binding hamiltonian we construct is an accurate and reliable way to obtain model parameters by preserving the phase and the orbital information from the DFT calculations.
\section{\label{sec:level1} HAMILTONIAN FOR WEAKLY INTERACTING BILAYERS}
Before modeling the interlayer coupling, we first reconstruct the {\it ab-initio} tight-binding hamiltonian $\mathcal{H}_0$ for a graphene monolayer\cite{macdonald_graphene_mono}. The unit cell for monolayer graphene is spanned by $\vec{a}_1=(\sqrt{3}\hat{x}-\hat{y})a/2$ and $\vec{a}_2=(\sqrt{3}\hat{x}+\hat{y})a/2$ with the lattice constant $a$=2.46 \AA. Two basis atoms are situated at $\vec{\delta}_{\rm A}=0$ and $\vec{\delta}_{\rm B}=(\vec{a}_1+\vec{a}_2)/3$. We extract the intralayer couplings up to the eighth nearest neighbors, which shows good agreements with DFT results. The numerical parameters for $t_i$, the intralayer hopping parameter to the $i$-th nearest neighbor, are listed in the left block $\mathcal{H}_0$ of Table \ref{table:tbh_parm}.
\begin{table}[]
\centering
\caption{Graphene intralayer ($\mathcal{H}_0$) and interlayer ($\mathcal{H}'$ ) TBH parameters with the on-site energy $\epsilon_C=0.3504$ eV. $t_i$ (for the locations of these neighbors, see Ref. \cite{macdonald_graphene_mono}) and $\lambda_i$ are in eV with $a=2.46$ \AA $ $ for $\bar{r}$; $\xi_i$, $x_i$, $\kappa_i$ are dimensionless parameters.}
\label{table:tbh_parm}
\begin{tabular}{cccccc|c|c|c|}
\cline{7-9}
\multicolumn{4}{c}{$\mathcal{H}_0$} & & $\mathcal{H}'$ & \textbf{$V_0(r)$} & \textbf{$V_3(r)$} & \textbf{$V_6(r)$} \\ \cline{1-4} \cline{6-9}
\multicolumn{1}{|c|}{\textbf{$t_1$}} & \multicolumn{1}{c|}{$-2.8922$} & \multicolumn{1}{c|}{\textbf{$t_5$}} & \multicolumn{1}{c|}{0.0524} & \multicolumn{1}{c|}{$\; \; \; $} & \textbf{$\lambda_i$} & $0.3155$ & $-0.0688$ & $-0.0083$ \\ \cline{1-4} \cline{6-9}
\multicolumn{1}{|c|}{\textbf{$t_2$}} & \multicolumn{1}{c|}{0.2425} & \multicolumn{1}{c|}{\textbf{$t_6$}} & \multicolumn{1}{c|}{$-0.0209$} & \multicolumn{1}{c|}{$\; \; \; $} & \textbf{$\xi_i$} & $1.7543$ & $3.4692$ & $2.8764$ \\ \cline{1-4} \cline{6-9}
\multicolumn{1}{|c|}{\textbf{$t_3$}} & \multicolumn{1}{c|}{$-0.2656$} & \multicolumn{1}{c|}{\textbf{$t_7$}} & \multicolumn{1}{c|}{$-0.0148$} & \multicolumn{1}{c|}{$\; \; \; $} & \textbf{$x_i$} & $-$ & $0.5212$ & $1.5206$ \\ \cline{1-4} \cline{6-9}
\multicolumn{1}{|c|}{\textbf{$t_4$}} & \multicolumn{1}{c|}{0.0235} & \multicolumn{1}{c|}{\textbf{$t_8$}} & \multicolumn{1}{c|}{$-0.0211$} & \multicolumn{1}{c|}{$\; \; \; $} & \textbf{$\kappa_i$} & $2.0010$ & $-$ & $1.5731$ \\ \cline{1-4} \cline{6-9}
\end{tabular}
\end{table}
The shape of the localized basis, also known as the Wannier orbital provides intuition for the chemical bonding, hybridization and the symmetry of the crystal. For monolayer graphene, the constructed Wannier orbital has a dominant $p_z$ character but the azimuthal symmetry is broken by the crystal field distortion from the neighboring atoms. Locally, at the position of the carbon atom, the three-fold rotation symmetry is restored. Thus, the angular momentum is defined up to modulo 3, which means that there is hybridization within each sector of angular momentum states. In Fig. \ref{fig:Wannier_orbital}(b), we decompose the Wannier function for graphene into $m=0$ (dominant $p_z$), $m=\pm3$ and $m=\pm6$ angular momentum components. This decomposition shows the range and the strength of each component, and the characteristic radius gets larger for larger angular momentum components.
When two or more monolayers are brought into contact, the shape of the Wannier function has implications for the interlayer coupling. These couplings are described by the matrix elements: $\langle \psi_2|\mathcal{H}|\psi_1 \rangle$ with $\psi_1$ ($\psi_2$) the Wannier orbital of the first (second) layer and $\mathcal{H}$ the total hamiltonian. The angular momentum mixing as shown in Fig. \ref{fig:Wannier_orbital}(b) for the Wannier orbital in graphene translates into the angular dependence of such interlayer couplings, in addition to the usual dependence on the distance of the pair. Without loss of generality, we assume the projected vector $\vec{r}$ from $\psi_1$ to $\psi_2$ on the plane is along the positive $x$ axis, and $\theta_1$ ($\theta_2$) is the angle relative to $\vec{r}$ needed to determine the orientation of the crystal of the layer to which the Wannier orbital $\psi_1$ ($\psi_2$) belongs. The interlayer coupling can then be written as the function $t(r,\theta_1,\theta_2)$. If the underlying crystal and the embedded Wannier orbital has N-fold rotation symmetry, then $\theta$ is only defined up to modulo $2\pi/N$. The above interlayer coupling can be simplified to:
\begin{equation}
t(r,\theta_1,\theta_2) = \sum_{m_1,m_2=-\infty} ^{\infty}f_{m_1,m_2}(r) e^{im_1 N_1 \theta_1+im_2 N_2 \theta_2}
\label{eqn:interlayer_general}
\end{equation} with integers $m_i$. For real $t$, $f_{\bar{m}_1,\bar{m}_2}(r)=f^*_{m_1,m_2}(r)$. This decomposition can be viewed as the multi-channel interlayer hopping process.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{FIG2}
\caption{Hoppings for the shifted graphene bilayer of (a) $AA$-type and (b) $AB$-type pairs as functions of the pair distance $r$; the spread of the hoppings at fixed $r$ indicates angular dependence. Decomposition into the constant $m=0$ (solid line), $\cos(3\theta)$ (circle), and $\cos(6\theta)$ (cross) components for the (c) $AA$-type pairs and (d) $AB$-type pairs. These curves are modeled by the $V_i(r)$ in Eq. (\ref{eqn:TBH_interlayer_fit}).}
\label{fig:GG_AA_AB}
\end{figure}
We next apply this general analysis to bilayers of graphene (G-G) and of graphene/hexagonal boron nitride (G-hBN). We consider two specific stackings, $AA$ and $AB$ defined by the relative position of the basis $A$ or $B$ atom of the top layer to that of the basis $A$ atom of the bottom layer. The bilayers are assumed to be flat with the same constant separation $c=3.35$ \AA $ $ in the $z$ direction\cite{graphene_rev}. Since in these two specific cases the two layers are not rotated with respect to each other, each primitive unit cell contains four atoms. After carrying out the DFT and Wannier transformation,, the {\it ab initio} tight binding hamiltonian $\mathcal{H}=\mathcal{H}_0^t+\mathcal{H}_0^b+\mathcal{H}'$ can be decomposed into the intralayer $\mathcal{H}_0^{t}$, $\mathcal{H}_0^{b}$ and the interlayer $\mathcal{H}'$ parts. The interlayer coupling of any atomic pair can be obtained from elements of $\mathcal{H}'$ in the Wannier basis. We then apply a lateral translation $\vec{\Delta}=r\cos(\theta) \hat{x}+r\sin(\theta)\hat{y}$ to the top layer with the vertical separation $c$ fixed. This translation will affect both the distance and the relative orientation of the interlayer bonds while keeping the underlying crystal orientation untouched. The interlayer hoppings are extracted between the basis $A$ atom of the bottom layer at the origin and the shifted $A$ ($B$) basis atom of the top layer in the translated $AA$ ($AB$) structure.
The extracted interlayer hoppings as functions of the projected interlayer bond distance $r$ are plotted in Fig. \ref{fig:GG_AA_AB}(a), (b) for the graphene $AA$ and $AB$ bilayers, respectively. At a given distance $r$, the spread of the interlayer hopping indicates strong angular dependence. Notably, hopping in the $AA$ type bilayer is different from that in the $AB$ type bilayer. Due to the three-fold rotation symmetry of the underlying crystal, these interlayer hoppings are invariant under $\theta \rightarrow \theta \pm 2\pi/3$. We further decompose the angular dependence at fixed $r$ into its Fourier components of the constant term, $\cos(3\theta)$ and $\cos(6\theta)$ terms. Higher order $\cos(3N\theta)$ terms do exist but are vanishingly small.
The hopping is given by the superposition of the interlayer terms that involve the symmetric combination of the following parameters
\begin{equation}
\begin{split}
t(\vec{r})= & V_0(r) +V_3(r) [\cos(3\theta_{\rm 12})+\cos(3\theta_{\rm 21})] \\
& + V_6(r) [\cos(6\theta_{\rm 12})+\cos(6\theta_{\rm 21})]
\end{split}
\label{eqn:TBH_interlayer}
\end{equation} where $\vec{r}$ the two-dimensional (projected) vector connecting the two atoms, $r=|\vec{r}|$, and $\theta_{12}$ and $\theta_{21}$ the angles between the projected interlayer bond and the in-plane nearest neighbor bond as defined in Fig. \ref{fig:Wannier_orbital}(a). The result does not depend on which nearest neighbor bond is used. Compared with Eq. (\ref{eqn:interlayer_general}), these correspond to non-zero values for the terms $f_{0,0}$, $f_{\pm1,0}$, $f_{0,\pm1}$, $f_{\pm2,0}$ and $f_{0,\pm2}$, with $N_1=N_2=3$. We use the following fitting functions for $V_i(r)$ with $\bar{r}=r/a$
\begin{equation}
\begin{split}
V_0(r) & = \lambda_0 e^{-\xi_0 (\bar{r})^2} \cos(\kappa_0 \bar{r}) \\
V_3(r) & = \lambda_3 \bar{r}^2 e^{-\xi_3(\bar{r}-x_3)^2} \\
V_6(r) & = \lambda_6 e^{-\xi_6(\bar{r}-x_6)^2} \sin(\kappa_6 \bar{r})
\end{split}
\label{eqn:TBH_interlayer_fit}
\end{equation}
These Fourier projected components are plotted in Fig. \ref{fig:GG_AA_AB}(c) and (d) for the $AA$ and $AB$ stackings respectively. For the constant term, the $AA$/$AB$ hoppings are very similar to each other, and we define $V_0(r)$ to be the average of the two. Projection into $\cos(3\theta)$ is significant for the $AB$ type but vanishes identically for the $AA$ type, and the curve in the $AB$ case is defined as $2V_3(r)$. The two stackings have similar behavior for the much smaller $\cos(6\theta)$ term, and $2V_6(r)$ is the average of the two. The values of the fitting parameters are given in Table \ref{table:tbh_parm} from the analysis of the translated $AA$/$AB$ structures.
A proper analysis of the inherent symmetry of the two-layer system provides the justification for the form of the interlayer hoppings and enables us to generalize the model to arbitrary configuration. Specifically, the three-fold symmetry of the crystal field allows mixing between $p_z$ and $m=\pm 3N$ orbitals with $N$ an integer. Due to the crystal symmetry, these components acquire a phase $(-1)^N$ when $A$ and $B$ basis atoms are interchanged, which are related by a $yz$ mirror operation. Wannier orbital viewed as composite objects of mixed angular momentum, the hopping between two such objects is determined by the superposition of the individual hopping channels between each component, within the two center approximation\cite{slater}. The dominant channel is between the two $m=0$ components which gives the constant part in the interlayer hopping. There is a $\cos(3\theta)$ term from the coupling between $m=0$ and $m=\pm 3$ channels. Due to the symmetry, the two terms add up constructively (destructively) for the $AB$ ($AA$) type. This can also be seen from the additional minus sign in exchanging $A$ and $B$ basis atoms. There are two types of contribution to the $\cos(6\theta)$ term, and they can be generated from the coupling between $m=0$ and $m=\pm 6$, or between $m=\pm 3$ components of the two atoms. By symmetry, the first (second) part of the contribution is even (odd) in $AA$/$AB$. We can model the contribution of the first type from the average $\cos(6\theta)$ term of AA/AB in Fig. \ref{fig:GG_AA_AB}(c), (d). The channel between two $m=\pm 3$ can in general produce complicated angular dependence, but it is only a small correction (a few meV) and hence we ignore it in our model.
Following similar steps, we derive the tight-binding hamiltonian for interlayer coupling in the case of a bilayer G-hBN. In the G-hBN interlayer coupling compared to the G-G coupling the symmetry is lower since the atoms are not identical, and this affects the amplitude for each angular momentum channel. From similar analysis for the shifted $AA$/$AB$ couplings in a G-hBN bilayer, we can model the coupling as:
\begin{equation}
\begin{split}
t^{\rm CX}(\vec{r})= &V^{\rm CX}_0(r) + V^{\rm CX}_{3, \rm XC}(r) \cos(3\theta_{\rm XC}) \\
& +V^{\rm CX}_{3,\rm CX}(r) \cos(3\theta_{\rm CX})
\end{split}
\label{eqn:TBH_interlayer_GhBN}
\end{equation} where X=B,N atoms. $V^{\rm CX}_0(r) $, $V^{\rm CX}_{3, \rm CX}(r)$ and $V^{\rm CX}_{3, \rm XC}(r)$ share the same functional form as the ones in the G-G case, and the corresponding values of the parameters are tabulated in Table \ref{table:tbh_parm_hBN}.
\begin{widetext}
\begin{table}[]
\centering
\caption{hBN intralayer and G-hBN interlayer tight-binding parameters. $\epsilon_B=2.2021$ eV, $\epsilon_N=-1.9124$ eV with the convention $a=2.46$ \AA $ $ for $\bar{r}$ in the fitting expression Eq. (\ref{eqn:TBH_interlayer_fit}).}
\label{table:tbh_parm_hBN}
\begin{tabular}{cccccc|c|c|c|c|c|c|}
\cline{7-12}
\multicolumn{4}{c}{$\mathcal{H}_0^B$ ($\mathcal{H}_0^N$)} & & $\mathcal{H}'$ & $V_0^{\rm CB}(r)$ & $V_{3,\rm BC}^{\rm CB}(r)$ & $V_{3,\rm CB}^{\rm CB}(r)$ & $V_0^{\rm CN}(r)$ & $V_{3, \rm NC}^{\rm CN}(r)$ & $V_{3,\rm CN}^{\rm CN}(r)$ \\ \cline{1-4} \cline{6-12}
\multicolumn{1}{|c|}{$t_1$} & \multicolumn{1}{c|}{$-2.6490$} & \multicolumn{1}{c|}{$t_5$} & \multicolumn{1}{c|}{0.0344 (0.0301)} & \multicolumn{1}{c|}{$\; \; \; $} & $\lambda_i$ & 0.3905 & $-0.0588$ & $-0.0651$ & 0.2517 & $-0.0606$ & $-0.0465$ \\ \cline{1-4} \cline{6-12}
\multicolumn{1}{|c|}{$t_2$} & \multicolumn{1}{c|}{0.0594 (0.2276)} & \multicolumn{1}{c|}{$t_6$} & \multicolumn{1}{c|}{$-0.0374 (-0.0240)$} & \multicolumn{1}{c|}{$\; \; \; $} & $\xi_i$ & 1.5426 & 3.0827 & 3.7998 & 1.6061 & 3.3502 & 3.0464 \\ \cline{1-4} \cline{6-12}
\multicolumn{1}{|c|}{$t_3$} & \multicolumn{1}{c|}{$-0.2163$} & \multicolumn{1}{c|}{$t_7$} & \multicolumn{1}{c|}{$-0.0053$} & \multicolumn{1}{c|}{$\; \; \; $} & $x_i$ & $-$ & 0.6085 & 0.6341 & $-$ & 0.5142 & 0.5264 \\ \cline{1-4} \cline{6-12}
\multicolumn{1}{|c|}{$t_4$} & \multicolumn{1}{c|}{0.0502} & \multicolumn{1}{c|}{$t_8$} & \multicolumn{1}{c|}{$-0.0133$} & \multicolumn{1}{c|}{$\; \; \; $} & $\kappa_i$ & 1.8229 & $-$ & $-$ & 2.1909 & $-$ & $-$ \\ \cline{1-4} \cline{6-12}
\end{tabular}
\end{table}
\end{widetext}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{FIG3}
\caption{Comparison between the tight binding hamiltonian (red lines) and {\it ab initio} DFT (blue circles) band structure calculations along $\Gamma$-M-K-$\Gamma$ in a $AB$ stacking (a) bilayer, (b) bulk; (c) $(M,N)$=(6,5) twisted super-structure ($\theta^{(6,5)} \approx 6.01^{\circ}$) and comparison to the folded monolayer band structure (green crosses).}
\label{fig:TBH_DFT_bands}
\end{figure}
To validate our model with the intra- and inter-layer couplings, in Fig. \ref{fig:TBH_DFT_bands}(a), (b) we compare the band structure obtained from DFT and from our tight-binding hamiltonian in the conventional $AB$-stacking bilayer and bulk graphite. The two band structures show good agreement over a large energy region around the Fermi level. The discrepancies away from the Fermi level are due to the hybridization of $p_z$ orbitals and other orbitals such as $sp_2$ which are not included in our Wannier model. When one monolayer is twisted relativ to the other, a supercell structure can be constructed in the commensurate case\cite{graphene_sc_dft1,SC_def}, labeled by ($M$,$N$) with twist angle $\theta^{(M,N)}$. The two layers are separated by a constant height $c=3.35$ \AA. In Fig. \ref{fig:TBH_DFT_bands}(c), we compare the result from DFT and tight-binding calculations for the ($M$,$N$)=$(6,5)$ twisted super-cell ($\theta\approx 6.01^{\circ}$); the model hamiltonian reproduces the DFT band structure well. We also include in Fig. \ref{fig:TBH_DFT_bands}(c) a comparison with the band structure of the folded BZ for a single monolayer, which is quite different, showing the importance of having an accurate description of interlayer coupling.
\section{\label{sec:level1} TWISTED BILAYER GRAPHENE PHYSICS}
The coupling between layers is weak for $\theta \approx 30^\circ$, and gets stronger when the twist approaches angles near $0^\circ$ or $60^\circ$.\cite{graphene_sc_tbh1} When two layers are twisted, the bands are formed from the hybridization of monolayer bands\cite{graphene_sc_tbh3} as in the schematic diagram of Fig. \ref{fig:twist_physics}(a) inset. The characteristic kinetic energy scale is defined by $\hbar v_F \Delta K$, with $v_F=8.22 \times 10^5$ m/s the Fermi velocity of the monolayer graphene, and $\Delta {\rm K} =|{\rm K}-{\rm K}^\theta| = \frac{8\pi}{3a} \sin(\theta/2)$ the distance between displaced K points of two layers (${\rm K}=\frac{4\pi}{3a} \hat{y}$, ${\rm K}^\theta=\frac{4\pi}{3a} (\cos(\theta)\hat{y}-\sin(\theta)\hat{x})$). Under the hybridization at large twist angles, the bilayer bands retain the linear Dirac dispersion, but with a different slope around the Dirac point compared to the folded bands of the monolayer graphene as in Fig. \ref{fig:TBH_DFT_bands}(c) and \ref{fig:twist_physics}(b). The Fermi velocity is renormalized by the interlayer coupling as can be seen from the effective low-energy theory around the Dirac points. A pair of states around the Dirac point with near zero energy are coupled to three pairs of states of energy $\pm \hbar v_F \Delta K$. The effective theory from a Schrieffer-Wolff transformation for the near zero energy doublet states has corrections to linear order in $k$ which renormalize the velocity\cite{graphene_sc_kp1}. We constructed a series of twisted supercell structures with decreasing angles from $30^\circ$, and the renormalized Fermi velocity calculated from the bands along $\Gamma$-K indeed follows the theoretical prediction $\tilde{v}_F /v_F = 1-C /\sin^{2} (\theta/2)$ with $C = 1.953 \times 10^{-4}$.\cite{graphene_sc_kp1}
Another feature for the band structure of the twisted superstructure is the van Hove singularities (VHS) in the density of states (DOS) near the Fermi level\cite{graphene_sc_tbh3}, which often leads to electronic instabilities such as superconductivity\cite{vanhove_sc} and magnetism\cite{vanhove_mag} in the many-body system. In Fig. \ref{fig:twist_physics}(c), the DOS of a $(M,N)$=(6,5) supercell with $\theta \approx 6.01^\circ$ is compared to a monolayer graphene with the singular points corresponding to the energy extrema in Fig. \ref{fig:twist_physics}(b). This VHS is due to gap opening from hybridization between states in the overlap between the Dirac cones of the two layers \cite{graphene_sc_tbh3}. The advantage of twisted bilayers is that the location of VHS can be controlled by varying the twist angle\cite{van_hove}, and are roughly centered at $E=\pm \frac{1}{2}\hbar v_F \Delta K$.
When the twist angle is even smaller, such as with a $(M,N)$=(31,30) supercell structure ($\theta \approx 1.08^\circ$), the Fermi velocity is close to zero, and nearly flat bands are observed at the Fermi level\cite{graphene_sc_tbh2} in Fig. \ref{fig:twist_physics}(d). The electronic states in these nearly dispersionless bands show highly localized charge density at the $AA$-sites\cite{graphene_sc_tbh1} as in Fig. \ref{fig:twist_physics}(e), referred to as Moir\'{e} pattern. In experiments, the VHS and the localization of electrons in twisted graphene layers have been measured by scanning tunneling spectroscopy\cite{van_hove,moire_exp,moire_exp2}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{FIG4}
\caption{(a) The renormalized Fermi velocity, $\tilde{v}_F/v_F$, as a function of the twist angle $\theta$ and VHS in the DOS (inset). (b) The band structure along K-M-K' in a $(M,N)$=(6,5) supercell ($\theta \approx 6.01^\circ$) compared to the folded monolayer bands (black dots). (c) The DOS for a $(M,N)$=(6,5) supercell (green) compared with the DOS of the monolayer (black). The energy extrema in the bands in (b) correspond to the singular points of DOS. (d) Similar calculations for the $(M,N)$=(31,30) supercell ($\theta\approx1.08^\circ$). (e) The states around the Fermi level are localized at $AA$-sites forming a Moir\'{e} pattern.}
\label{fig:twist_physics}
\end{figure}
In the discussion so far we have assumed the layers to be flat and focused only on their electronic properties. Structural relaxations such as rippling or more drastic commensurate-incommensurate transitions with domain-line formation, as in the G-hBN bilayer\cite{ghbn_com_incom}, could be relevant for small twist angles. They are driven by the different local mechanical energy for $AA$ and $AB$ stackings\cite{graphene_com_incom}. Though the prediction of mechanical deformations is beyond the scope of the current work, we comment that more general forms of interlayer couplings can be modeled by incorporating variable height or strain by modifying the initial $AA$/$AB$ sliding bilayers, which will allow proper description of the effects of structural deformations.
\section{\label{sec:level1} EFFECTIVE THEORY FROM PROXIMITY EFFECTS}
As a final comment on how our model can be applied, we discuss how to construct effective theories with limited degrees of freedom instead of having to solve the hamiltonian in the full Hilbert space of the combined layers. In graphene bilayers, the low-energy hamiltonian has the form of non-abelian gauge theory\cite{non_abelian_gauge}. To define and formulate the problem, we consider a vdW bilayer heterostructure with layer 1 as the main component where the low-energy degrees of freedom at the Fermi level $E_{F}$ reside. Layer 2 is brought close to layer 1 to introduce the desired proximity effects. In general, the presence of layer 2 will affect the electronic properties of layer 1 in two ways: (1) by introducing a direct additional potential generated by the neighboring atoms; (2) by introducing virtual interlayer hopping processes through hybridization to the states of the neighboring layer. The general vdW heterostructure hamiltonian takes the form
\begin{equation}
\mathcal{H}_{\rm vdW}=\begin{bmatrix}
\hat{H}_{1} + \Delta \hat{H}_{12} & \hat{T} \\
\hat{T}^\dagger & \hat{H}_{2} + \Delta \hat{H}_{21}
\end{bmatrix}
\end{equation} $\Delta \hat{H}_{12}$ and $\Delta \hat{H}_{21}$ are the direct corrections of the first type, and $\hat{T}$ is the interlayer coupling in the heterostructure. Integrating out the second layer gives a perturbation term for the first layer
\begin{equation}
\Delta V_{12}(E)=\hat{T} \frac{1}{E-(\hat{H}_2+\Delta \hat{H}_{21})} \hat{T}^\dagger
\label{eqn:eff_layer1}
\end{equation} where, since $\hat{T}$ is already small, and $\Delta \hat{H}_{21}$ is a higher order correction that disrupts the lattice translation symmetry in $\hat{H}_2$, this term can be ignored. The effective potential takes the following form in the spatial representation, evaluated at $E=E_F$
\begin{equation}
\begin{split}
& \Delta V_{12}(r_2,r_1,E_F) = \sum_{s_2,s_1} t_{r_2,s_2} \langle s_2| \frac{1}{E_F-\hat{H}_2} |s_1 \rangle t^\dagger_{s_1,r_1} \\
&=\sum_{s_2,s_1} \frac{t_{r_2,s_2} t^\dagger_{s_1,r_1} }{\Omega_k} \int_{\rm BZ} d^2 \vec{k} \langle \psi_{\vec{k},\delta_2}| \frac{ e^{i \vec{k}\cdot (\vec{s}_{2}-\vec{s}_1)} }{E_F-\hat{H}_2(\vec{k})}|\psi_{ \vec{k},\delta_1} \rangle
\end{split}
\end{equation} with $\Omega_k$ the BZ area, $r_i$ ($s_i$) the localized orbitals of layer 1 (2), which include both the position vector $\vec{r}_{i}$ ($\vec{s}_i$) and the orbital index $\delta_i$, and $t_{i,j}$ are the interlayer coupling from $j$ to $i$ orbitals between the layers. In the usual perturbation framework, this expression describes hopping across the layers from $r_1$ to $s_1$, allowing for all paths $s_1$ to $s_2$ within layer 2, and hopping back to layer 1, from $s_2$ to $r_2$. When applied to the G-hBN bilayer, the sub-lattice symmetry breaking mass terms of $\Delta V_{12}$ from hBN to carbon sites have opposite sign from the direct term $\Delta H_{12}$: the carbon site above a BN layer experiences the same sub-lattice potential as the BN layer itself in the direct contribution $\Delta \hat{H}_{12}$ while $\Delta V_{12}$ is opposite due to level repulsion in the framework of perturbation theory. The use of this effective potential will enable application to very large systems without loss of accuracy.
\section{\label{sec:level1} CONCLUSION}
In summary, we derived the {\it ab initio} G-G and G-hBN interlayer couplings based on the maximally localized Wannier function transformation of DFT calculations. We show that these interlayer couplings have both pair-distance and angular-orientation dependence. In contrast, the conventional way of modeling such couplings by fitting band structure calculations leads to ambiguities in the functional form and its dependence on important structural variables\cite{graphene_sc_tbh1,graphene_sc_tbh2,graphene_sc_tbh3,graphene_sc_tbh4, carbon_bond,fit_tb}. The success of the latter, simpler approach is due to the small number of parameters needed in effective low-energy theories near the Dirac energy\cite{graphene_sc_kp1,graphene_sc_kp2,graphene_sc_kp3,graphene_sc_kp4} implying that only one set of dominant Fourier components for interlayer coupling is relevant; this set of components, however, is not enough to constrain its functional form and its dependence on key variables. In the work by Jeil Jung {\it et al.} \cite{macdonald_graphene_bilayer,macdonald_moire}, such interlayer couplings were extracted with the use of the Wannier transformation but the crystal configuration of the bilayer was held at fixed orientation which means that it can only be applied to layered stacks that involve only translations and small relative twist angles.
In our work, we elucidate the physics from the extracted couplings by analyzing the multi-angular-momentum channel contributions. This enabled us to generalize the interlayer coupling model to arbitrary stacking orientations with that involve any possible relative translation or rotation of the layers. Our model can also be generalized to incorporate local variations of in-plane strain and interlayer distance by varying the reference configurations in a systematic way. We expect our model to be relevant in investigating the derivation of low-energy theories appropriate for layer stackings\cite{graphene_sc_kp1,graphene_sc_kp2,graphene_sc_kp3,graphene_sc_kp4,non_abelian_gauge}, the band gap introduced by the presence of hBN\cite{mac_donald_gap_ghbn}, optical absorption\cite{optical_absorp}, vertical transport across layers\cite{moire_transport}, phenomena like the quantum Hall effects and Hofstadter's butterfly from the competition between magnetic field and supercell length scales\cite{QHE_TwBLG,moire_butterfly,kim_butterfly}. The systematic Wannier approach also allows for further generalization to other two-dimensional layered materials.
\begin{acknowledgements}
We thank Dennis Huang, Daniel Massatt, Stephen Carr, Min-Feng Tu, Sharmila N. Shirodkar, Georgios A. Tritsaris, Paul Cazeaux, Eric Cances, Mitchell Luskin, Bertrand Halperin and Philip Kim for useful discussions. This work was supported by the STC Center for Integrated Quantum Materials, NSF Grant No. DMR-1231319 and by ARO MURI Award W911NF-14-0247. We used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by NSF Grant No. ACI-1053575.
\end{acknowledgements}
|
\section{Introduction}
\label{sec:intro}
Galactic winds have been known to be common features of star-forming
galaxies for many years. While particularly spectacular winds, such as
those of M82, NGC~1482 and NGC~253 \citep[eg][]{1999ApJ...513..156M,
2004ApJ...606..829S}, are exceptional, galaxies with modest winds
are widespread \citep{2005ARAA..43..769V}, as revealed by extraplanar
diffuse x-ray
\citep{1990ApJ...355..442F,1995ApJ...445..666A,2004ApJS..151..193S}
H$\alpha$ \citep{1990ApJS...74..833H,2003ApJS..148..383M}, and dust
emission \citep{1999AJ....117.2077H} in several galaxies.
Interest in the physical nature of the winds and their prevalence has
increased since it has been recognised they appear to play key roles
in galaxy formation and the distribution of metals in the
Intergalactic Medium (IGM). Both photoionization and mechanical
feedback from star-forming regions have long been expected to limit
star-formation on small scales \citep{2007ARAA..45..565M}. Wind
feedback has also been invoked to impede gas accretion and so limit
the efficiency of star formation in galaxy formation models to account
for disagreement between model predictions and observations
\citep[eg][]{1986ApJ...303...39D, 2012MNRAS.425.2027K,
2015MNRAS.453.3499K, 2008MNRAS.387..577O,
1998MNRAS.300..773W}. Winds extending over hundreds of kiloparsecs
may account for intergalactic metal absorption systems
\citep[see the review by][]{RevModPhys.81.1405}.
It is widely believed galactic winds result from the collective impact
of supernovae in compact star-forming regions on their surroundings
\citep[see the review by][]{2005ARAA..43..769V}. The details of the
physical mechanism driving the winds, however, are still unknown. Most
models are based on the injection of energy and matter by supernovae
and stellar winds in a distributed region \citep{1968MNRAS.140..241B,
1971ApJ...165..381J, 1971ApJ...170..241M, 1985Natur.317...44C}. The
most straightforward model of pressure-driven expanding gas expelled
by the supernovae fails because it greatly over-predicts the gas
temperature as inferred from x-ray spectra of the winds, and
underpredicts the x-ray luminosities. Mass loading, in which
interstellar gas is incorporated into the flow, is recognized as the
most plausible explanation for the moderate temperatures and high
x-ray luminosities detected \citep{1993PASJ...45..513T,
1996ApJ...463..528S}. This possibility receives observational
support from measurements of velocity-broadened metal absorption line
systems \citep{2015ApJ...809..147H}. Sources for the mass include
pre-existing interstellar gas in the vicinity of the supernovae from
stellar winds, hydrodynamic ablation of gas clouds entrained in the
outflow and thermal evaporation from gas clouds
\citep{1996ApJ...463..528S, 2000MNRAS.314..511S, 2001AA...367.1000P,
2005MNRAS.362..626M}. Alternative wind models have also been
considered, such as momentum-driven winds \citep{2005ApJ...618..569M}
and winds driven by cosmic ray streaming \citep{1975ApJ...196..107I,
2012MNRAS.423.2374U}.
Based on the stellar wind model of \citet{1975ApJ...200L.107C} and
\citet{1977ApJ...218..377W} for isolated stars, the galactic
superbubble model of \citet{1987ApJ...317..190M} and
\citet{1988ApJ...324..776M} naturally incorporates mass loading
sourced by thermal evaporation off the wall of the wind cavity
produced by supernovae, resulting in a self-similarly expanding
supershell. Analytic modelling and numerical computations suggest the
initially spherically expanding superbubble soon develops into a
biconal outflow within the stratified interstellar medium of a disc
galaxy \citep{1985ApJ...299...24S, 1988ApJ...324..776M,
1988ApJ...330..695T}. Asymptotically, at large distances from the
galactic plane, the outflow may develop into a more spherical
superwind.
A more sophisticated approach to modelling winds is to use numerical
hydrodynamical computations to evolve a galactic wind from first
principles, but uncertainties in the detailed structure of the
interstellar medium limit the generality of the computations. The
computational demands imposed by the spatial resolution necessary to
capture all of the essential physics moreover precludes a fully
self-consistent treatment using cosmological simulations with current
resources. An element of \lq sub-grid physics' is ultimately required
\citep[eg][]{2003MNRAS.339..289S, 2012MNRAS.426..140D}. Critically,
all the numerical models suffer from one key deficiency:\ the unknown
physical mechanism driving the wind. In particular, the roles of both
clouds and thermal heat conduction, particularly as they affect the
amount of mass loading, are unknown. While the interaction between
winds and clouds has been investigated using numerical simulations
\citep[eg][]{2008ApJ...674..157C, 2015ApJ...805..158S}, most
simulations neglect thermal heat conduction \citep[but
see][]{1999MNRAS.309..941D, 2005MNRAS.362..626M,
2016arXiv160201843B}. The superbubble model has recently been
incorporated into cosmological simulations
\citep{2014MNRAS.442.3013K}, producing galaxy properties in good
agreement with observations \citep{2015MNRAS.453.3499K}.
A principal goal of this paper is to distinguish between models with
and without thermal heat conduction. The implicit justification for
neglecting heat conduction is that magnetic fields tangled by
turbulence will suppress heat conductivity. On the other hand, winds
will tend to comb out magnetic field lines, allowing some thermal heat
conduction along the wind direction. Which of these effects dominates
is not known.
The analytic modelling here is performed in the context of a
homogeneous wind. The homogeneity of wind gas is currently not well
constrained, although there is clear evidence for the presence of
entrained gas clouds. Even if initially the gas is clumped into
clouds, large mass-loading factors resulting from hydrodynamical
ablation or thermal evaporation may render the hot gas interior to the
wind bubble sufficiently smooth for the homogeneous models still to
provide a good approximation to the wind structure. Observations
suggest radiative losses, not included in the models, are generally
too small to affect the energetics of the winds, although they may
affect the metal column densities of embedded clouds
\citep{2001ApJ...554.1021H, 2002ApJ...577..691H}. Gravity will play a
secondary role in slowing the flow, an effect not readily incorporated
into an analytic treatment \citep[but see][]{2016ApJ...819...29B},
while it may limit the outflow velocity of ram-pressure driven clouds
\citep{2015ApJ...809..147H}. The principal role of gravity is in
stratifying the galactic disc gas, as will be discussed below.
The {\it Chandra X-ray Observatory} has enabled the development of
observing campaigns to systematically investigate the x-ray properties
of star-forming galaxies \citep{2004ApJS..151..193S,
2005ApJ...628..187G, 2012MNRAS.419.2095M, 2012MNRAS.426.1870M,
2014MNRAS.437.1698M}.
The primary quantity focussed on in this paper is the radiative x-ray
efficiency of the wind as quantified by the specific {\it diffuse}
x-ray energy generated per solar mass of stars formed. The x-ray
emission profiles extend to typical scales of several kiloparsecs
\citep{2004ApJS..151..193S, 2012MNRAS.426.1870M}, well beyond the
active star-forming regions driving the outflow. Winds on these scales
form bipolar cones through the stratified disc gas and become
increasingly inhomogeneous due to the onset of Kelvin-Helmholtz and
Rayleigh-Taylor instabilities, details not amenable to an analytic
treatment. Since the diffuse x-ray profiles are strongly centrally
peaked, however, the models considered here should still capture a
fair fraction of the total x-ray luminosity, particularly for dwarf
starbursts \citep{2005ApJ...628..187G}. This paper seeks to quantify
the radiative x-ray efficiency from the central region within the disc
of the galaxies, allowing for a range in star-formation rates and wind
properties. Although analytic models are approximate, they provide
insight into the origin of observational trends in terms of the
physical properties of the winds. They also provide invaluable
guidance into the design and interpretation of numerical simulations.
In the next section, approximate analytic scaling relations are
derived for the structure and some observational signatures of the
winds. In Sec.~\ref{sec:numeval_xray}, x-ray luminosity predictions
are presented using full numerical integrations of the models. This is
followed by predictions for radio luminosities in
Sec.~\ref{sec:numeval_rad} and for metal absorption column densities
in Sec.~\ref{sec:metals}. The results are discussed in
Sec.~\ref{sec:Discussion}, followed by a summary of the key results in
a conclusions section.
\section{Analytic estimates of specific x-ray emission}
\label{sec:estimates}
The x-ray emission from a wind arises from both thermal free-free and
line emission. To understand the dependence of the emission on the
properties of the sources and the surrounding gas, it is helpful first
to estimate the thermal free-free component analytically. In the
following section, more accurate results from numerical computations
are provided including line emission. Estimates use the model of
\citet{1985Natur.317...44C} for a steady-state wind and of
\citet{1987ApJ...317..190M} and \citet{1988ApJ...324..776M} for
superbubbles. Both assume gravitational acceleration is negligible, a
good approximation in the central regions of a galaxy where the gas is
much hotter than the galactic virial temperature. Gravity, however,
produces stratification of the galactic disc gas, which results in
biconal outflow. This is a limitation of both models:\ the
steady-state wind assumes a homogeneous source, limited therefore to a
region small compared with the scale-height of the disc. Comparison
with numerical simulations, however, suggest the wind produced by a
starburst in a cylindrical region in a disc may be rescaled to the
spherically symmetric solution to high accuracy
\citep{2009ApJ...697.2030S}. The superbubble model assumes a
homogeneous surrounding medium. Hydrodynamical simulations show this
tends to limit the growth of the superbubble to within the disc as
pressure-driven lobes emerge vertically
\citep{1989ApJ...337..141M}. The analysis here concentrates on the
structure of the superbubble when it reaches a size comparable to the
scale-height of the disc.
\subsection{Steady-state wind}
\label{subsec:ssw}
The characteristic ejecta energy and mass of a core-collapse supernova
are taken by \citet{1985Natur.317...44C} to be $E=10^{51}\,{\rm erg}$
and $M=3\,{\rm M_\odot}$. Assuming a Salpeter stellar initial mass function
(IMF), a lower progenitor mass limit of $8\,{\rm M_\odot}$, $\nu_{\rm SN}$ for
a core-collapse supernova gives a rate of about 1 core-collapse
supernova per 100 solar masses of stars formed. (The value would rise
by about 40 percent for a Kroupa IMF.) Characterizing the supernova
rate as $0.01\nu_{100}$ supernova per solar mass of stars formed, the
energy and mass injection rates for a star-formation rate $\dot M_*$
are $\dot E=(10^{49}\epsilon\,{\rm erg}\, {\rm M_\odot}^{-1})\nu_{100}\dot
M_*=\epsilon\dot E_1$ and $\dot M=(0.03\beta)\nu_{100}\dot
M_*=\beta\dot M_1$, where $\epsilon$ and $\beta$ allow for
uncertainties in the mechanical energy and mass-loading,
respectively. For $\dot M_*$ expressed in ${\rm M_\odot\, yr^{-1}}$,
$\dot E\simeq3.17\times10^{41}\,{\rm
erg\,s^{-1}}\,\epsilon\nu_{100}{\dot M_*}$.
Most of the bolometric thermal free-free emission originates in the
central source region at $r<R$. The analytic estimate is based on
emission from this region. Using the results in the Appendix for a
$\gamma=5/3$ gas, the central hydrogen density is
\begin{equation}
n_{{\rm H}0}\simeq0.00658\,{\rm cm}^{-3}\,\frac{\beta^{3/2}}{\epsilon^{1/2}}\nu_{100}\dot M_* R_{100}^{-2},
\label{eq:nH0}
\end{equation}
\noindent where $\dot M_*$ is the star formation rate in units of
${\rm M_\odot}\,{\rm yr}^{-1}$ and $R_{100}$ is the radius of the star
forming region in units of 100~pc. The central temperature is
\begin{equation}
T_0 = \frac{2}{5}\frac{\bar m}{k_{\rm B}}\frac{\epsilon}{\beta}{\dot
E_1}{\dot M_1}^{-1}
\simeq4.76\times10^8\,{\rm K}\,\frac{\epsilon}{\beta}
\simeq14.2v_\infty^2,
\label{eq:T0}
\end{equation}
\noindent where $\bar m$ is the mean mass per particle, and in the
last expression the temperature is characterized by the asymptotic
wind velocity at $r\gg R$,
\begin{equation}
v_\infty = 2^{1/2}\frac{\epsilon^{1/2}}{\beta^{1/2}}{\dot
E_1^{1/2}}{\dot M_1}^{-1/2}
\simeq5790\,{\rm km\,s^{-1}}\,\frac{\epsilon^{1/2}}{\beta^{1/2}}.
\label{eq:vinf}
\end{equation}
In terms of $v_\infty$, the central hydrogen density is
$n_{{\rm H}0}\simeq1.28\,{\rm cm^{-3}}\,(v_\infty/1000\,{\rm
km\,s^{-1}})^{-3}\epsilon\nu_{100}{\dot M}_* R_{100}^{-2}$.
Radiative cooling places a lower limit on the asymptotic wind
velocity. The energy injection rate must exceed the cooling rate
within the central region of the wind. The cooling rate is
$n_en_{\rm H}\Lambda_R(T)$ for electron and hydrogen number densities
$n_e$ and $n_{\rm H}$, respectively, where
$\Lambda_R(T)\simeq1.0\times10^{-22}\,{\rm erg\,cm^3\,s^{-1}}
T_6^{-0.7}\zeta_m$
with $T_6=T/10^6$~K and $\zeta_m$ the metallicty relative to solar
\citep{1988ApJ...324..776M}. Requiring
${\dot E}> n_en_{\rm H}\Lambda_R(T)(4\pi R^3/3)$ imposes the robust
restriction
\begin{equation}
v_\infty > 550\,{\rm km\,s^{-1}}\left(\epsilon\nu_{100}\zeta_m{\dot M_*}
R_{100}^{-1}\right)^{0.14},
\label{eq:vinfmin}
\end{equation}
corresponding to the limit on the mass-loading factor
$\beta<110\epsilon^{0.73}(\nu_{100}\zeta_m{\dot M_*}/
R_{100})^{-0.27}.$
In the literature, a more commonly defined mass-loading factor is the
ratio of mass outflow rate to star-formation rate. In terms of this
ratio, designated here by $\beta_*(=0.03\beta\nu_{100})$, the cooling
restriction imposes
\begin{equation}
\beta_*=\left(\frac{1000\,{\rm {km\, s^{-1}}}}{v_\infty}\right)^2\epsilon\nu_{100}<3.3\left(\epsilon\nu_{100}\right)^{0.73}\left({\dot
M_*}\zeta_m R_{100}^{-1}\right)^{-0.27}.
\label{eq:betas}
\end{equation}
A similar restriction is derived by \citet{2014ApJ...784...93Z}. The
mass injection rate is then limited to ${\dot M} <
3.3(\epsilon\nu_{100}{\dot M_*})^{0.73}(R_{100}/
\zeta_m)^{0.27}\,{\rm M_\odot}\,{\rm yr}^{-1}$.
As shown in the Appendix, the core density and temperature are nearly
uniform within $r< 0.98R$. Approximating the density and temperature
as constant within $r<R$, the total thermal free-free luminosity is
\begin{equation}
L^{\rm
ff}_\nu\simeq3\times10^{16}\,{\rm erg\,s^{-1}\,Hz^{-1}}\, e^{-0.0244e_{\rm keV}(\beta/\epsilon)}
\frac{\beta^{7/2}}{\epsilon^{3/2}}(\nu_{100}{\dot M}_*)^2R_{100}^{-1},
\label{eq:Lnuffssw}
\end{equation}
for x-rays of energy $e_{\rm keV}$ in keV. Integrating
Eq.~(\ref{eq:Lnuffssw}) over the energy band $(e_1-e_2)$~keV, the
x-ray energy produced per solar mass of stars formed is then
\begin{eqnarray}
\frac{E_{[e_1-e_2]\,{\rm keV}}}{M_*}&\simeq&8\times10^{42}\,{\rm
erg}\,{M_\odot}^{-1}\,\nu_{100}^2{\dot M}_*R_{100}^{-1}\nonumber\\
&&\times\frac{\beta^{5/2}}{\epsilon^{1/2}}\Biggl[\exp\left(-0.0244\frac{\beta}{\epsilon}e_{1,{\rm keV}}\right)\nonumber\\
&&-\exp\left(-0.0244\frac{\beta}{\epsilon}e_{2,{\rm keV}}\right)\Biggr].
\label{eq:ExMsssw}
\end{eqnarray}
The model predicts a linear increase with the star formation
rate. Taking $\epsilon=1$ and a typical mass-loading factor of
$\beta\simeq100$, corresponding to the asymptotic wind velocity
$v_\infty\simeq600\,{\rm km\,s^{-1}}$, gives
\begin{equation}
\frac{E_{[0.5-2]\,{\rm keV}}}{M_*}\simeq2\times10^{47}\,{\rm
erg}\,{M_\odot}^{-1}\,\nu_{100}^2{\dot M}_*R_{100}^{-1}.
\label{eq:ExMssswex}
\end{equation}
\subsection{Superbubble with thermal heat conduction}
\label{subsec:ssbtc}
Allowing for an ambient interstellar medium and equilibration of the
temperature interior to the bubble cavity by thermal heat conduction,
\citet{1987ApJ...317..190M} and \citet{1988ApJ...324..776M} model the
superbubble as a self-similar expanding stellar wind.
Normalized by the typical mechanical luminosity of an OB association,
$L=10^{38}L_{38}\,{\rm erg\,s^{-1}}$, and for an ambient hydrogen
density outside the wind bubble $n_{{\rm H}, 0}$, the bubble radius
increases, assuming no radiative losses, like
\begin{equation}
R_{\rm B}\simeq 66\,{\rm pc}\left(\frac{L_{38}t_6^3}{n_{{\rm H},0}}\right)^{1/5},
\label{eq:SB_rad}
\end{equation}
where $t_6$ is the age of the bubble in units of $10^6$~yr. Adopting
the thermal conductivity coefficient
$\kappa(T)=6\times10^{-7}f_T\,{\rm erg\,s^{-1}\,cm^{-1}\,K^{-7/2}}$,
including a possible conductivity suppression factor $f_T$, the
interior bubble temperature and ionized hydrogen number density are
given in terms of the similarity variable $x=r/R_{\rm B}$ for radius
$r$ by
\begin{equation}
T \simeq (5.2\times10^6\,{\rm K})f_T^{-2/7}L_{38}^{8/35}n_{{\rm H},0}^{2/35}t_6^{-6/35}(1-x)^{2/5},
\label{eq:SB_T}
\end{equation}
and
\begin{equation}
n_{\rm H} \simeq (0.016\,{\rm cm^{-3}})f_T^{2/7}L_{38}^{6/35}n_{{\rm H}, 0}^{19/35}t_6^{-22/35}(1-x)^{-2/5}.
\label{eq:SB_nH}
\end{equation}
The bubble will cool primarily by line radiation at its surface. The
characteristic radiative cooling time is
\begin{equation}
t_{\rm R}\simeq(15\times10^6\,{\rm yr})L_{38}^{3/11}n_{{\rm
H},0}^{-8/11}f_T^{-25/22}(\zeta_m+0.15)^{-35/22}.
\label{eq:tcool}
\end{equation}
The factor 0.15 has been added to $\zeta_m$ to account for hydrogen
and helium cooling, where care is taken near the surface to ensure
cooling is cut off below the recombination temperatures for helium and
hydrogen for collisionally ionized gas, and
$\Lambda_R(T)\sim T^{-1/2}$ was adopted for the surface layer,
following \citet{1988ApJ...324..776M}. For high ambient hydrogen
densities, cooling will limit the radius of the bubble to be smaller
than the characteristic scale height of the stratified interstellar
medium perpendicular to the disc. At lower densities, the bubble
radius may reach the disc scale height. The wind will then evolve into
a bipolar outflow perpendicular to the disc, and the expansion into
the plane of the disc ceases, or may even reverse
\citep{1988ApJ...324..776M, 1989ApJ...337..141M}.
Expressing the temperature and density of the gas interior to the
bubble as $T=T_cu^{2/5}$ and $n_{\rm H}=n_{{\rm H}c}u^{-2/5}$, where $u=1-x$,
the thermal free-free emission emitted by a wind bubble of radius
$R_{\rm B}=100R_{\rm B, 100}$~pc is
\begin{align}
L^{\rm ff}_\nu &\simeq& 5\times10^{25}\,{\rm
erg\,s^{-1}\,Hz^{-1}}n_{{\rm H}c}^2T_c^{-1/2}R_{\rm B, 100}^3\nonumber\\
&\times&\int_0^1\,\frac{du}{u}(1-u)^2\exp\left[-\left(\frac{h_{\rm P}\nu}{k_{\rm
B}T_c}\right)u^{-2/5}\right] \label{eq:Lnuffssbtc}\\
&\simeq&\begin{cases}
10^{26}\,{\rm erg\,s^{-1}\,Hz^{-1}}n_{{\rm
H}c}^2T_c^{-1/2}R_{\rm B, 100}^3\nonumber\\
\times\left[\log\left(\frac{k_{\rm
B}T_c}{h_{\rm P}\nu}\right)-\frac{3+5\gamma}{5}\right]
& (h_{\rm P}\nu\ll k_{\rm B}T_c);\nonumber\\
10^{27}\,{\rm erg\,s^{-1}\,Hz^{-1}}n_{{\rm
H}c}^2T_c^{-1/2}R_{\rm B, 100}^3\nonumber\\
\times\left(\frac{k_{\rm
B}T_c}{h_{\rm P}\nu}\right)^3e^{-h_{\rm P}\nu/k_{\rm
B}T_c} &(h_{\rm P}\nu\gg k_{\rm B}T_c),
\end{cases}
\end{align}
where $h_{\rm P}$ is the Planck constant, $k_{\rm B}$ is the Boltzmann
constant, here $\gamma\simeq0.5772$ is Euler's constant, and the
integral has been evaluated with its asymptotic leading order
behaviour retained for the two limiting cases shown. A characteristic
central temperature of $\sim4\times10^7$~K gives a transition energy
between the two cases of about 3~keV.
At high densities, the growth of the wind bubble will be limited by
cooling once the energy radiated matches the total mechanical energy
deposited by the wind. This may be quantified as follows. At the
cooling time $t=t_{\rm R}$, the central hydrogen density and gas
temperature take on the values
\begin{equation}
n_{{\rm H}c}\simeq0.003\,{\rm cm^{-3}}\,n_{{\rm H},0}f_T(\zeta_m+0.15),
\label{eq:nHcc}
\end{equation}
independent of $L_{38}$, and
\begin{equation}
T_c\simeq1.4\times10^7\,{\rm K}\,(\epsilon\nu_{100} n_{{\rm H},0}\dot
M_*)^{2/11}f_T^{-1/11}(\zeta_m+0.15)^{3/11},
\label{eq:Tcc}
\end{equation}
respectively, where $L_{38}$ has been converted to the star formation
rate $\dot M_*$ (${\rm M_\odot\,yr^{-1}}$) using
$L_{38}\simeq3170\epsilon\nu_{100}{\dot M_*}$
(Sec.~\ref{subsec:ssw}). The cooling radius may be expressed as
\begin{equation}
R_{\rm B, cool}\simeq6.3\,{\rm kpc}\,{n_{{\rm H},0}}^{-7/11}{\epsilon\nu_{100}\dot
M_*}^{4/11}f_T^{-15/22}(\zeta_m+0.15)^{-21/22}.
\label{eq:Rbc}
\end{equation}
The corresponding bubble expansion velocity
$\dot R_{\rm B, cool}=(3/5)R_{\rm B, cool}/t_{\rm R}$ at this time is
\begin{equation}
\dot R_{\rm B, cool}\simeq27\,{\rm km\,s^{-1}}\,(\epsilon\nu_{100}{\dot
M_* n_{{\rm H},0}})^{1/11}f_T^{5/11}(\zeta_m+0.15)^{7/11}.
\label{eq:vbc}
\end{equation}
The mass interior to the bubble is dominated by the evaporation off
the bubble wall into the hot cavity at the rate
$\dot M_{\rm ev}=(16\pi/25)[{\bar m}\kappa(T_c)/ k_{\rm B}]R_{\rm B}$
\citep{1975ApJ...200L.107C}, where $\kappa(T_C)$ is the thermal
conductivity coefficient. The mass loading factor in the wind core
referenced to the star formation rate becomes
\begin{eqnarray}
\beta_*&=&\frac{\dot M_{\rm ev}}{\dot M_*}\simeq2.1\left({\dot M_*}n_{{\rm
H},0}\right)^{-2/11}\nonumber\\
&&\times (\epsilon\nu_{100})f_T^{1/11}(\zeta_m+0.15)^{-3/11}.
\label{eq:betastcc}
\end{eqnarray}
From Eq.~(\ref{eq:Lnuffssbtc}), the x-ray energy in the 0.5--2~keV
band per solar mass of stars formed is then
\begin{equation}
\frac{E_{[0.5-2]\,{\rm keV}}}{M_*}\simeq2\times10^{47}\,{\rm
erg}\,{M_\odot}^{-1}\,(\epsilon\nu_{100})(\zeta_m+0.15)^{-1},
\label{eq:ExMssswtcc}
\end{equation}
independent of the star formation rate, the ambient hydrogen density
and the rate of thermal heat conduction. It corresponds to 2 percent
of the mechanical energy radiated as x-rays in this band. The wind
will not immediately cease as the momentum of the outflow will
continue to sweep up material, but at a reduced rate
\citep{1992ApJ...388...93K}.
At lower densities, the bubble radius will be limited by the scale
height of the gas perpendicular to the plane. The central hydrogen
density and gas temperature when the bubble reaches a radius $R_{\rm B}$ are
\begin{equation}
n_{{\rm H}c}\simeq0.23\,{\rm cm}^{-3}\,(\epsilon\nu_{100})^{8/21}f_T^{2/7}n_{{\rm H},0}^{1/3}{\dot
M_*}^{8/21}R_{\rm B, 100}^{-22/21}
\label{eq:nHw}
\end{equation}
and
\begin{equation}
T_c\simeq4.6\times10^7\,{\rm K}\,\left(\epsilon\nu_{100}{\dot M_*}/
f_TR_{\rm B, 100}\right)^{2/7},
\label{eq:Tcw}
\end{equation}
respectively. The corresponding wind velocity is
\begin{equation}
{\dot R}_{\rm B}\simeq430\,{\rm km\,s^{-1}}\,\left(\epsilon\nu_{100}{\dot
M_*}/n_{{\rm H},0}R_{\rm B, 100}^2\right)^{1/3}.
\label{eq:vww}
\end{equation}
The mass loading factor in the bubble is
\begin{equation}
\beta_*\simeq0.64(\epsilon\nu_{100})^{5/7}\left(R_{\rm B, 100}f_T/{\dot M_*}\right)^{2/7},
\label{eq:betastcsc}
\end{equation}
independent of the ambient gas density. The x-ray energy in the
0.5--2~keV band per solar mass of stars formed is then
\begin{eqnarray}
\frac{E_{[0.5-2]\,{\rm keV}}}{M_*}&\simeq&5\times10^{45}\,{\rm
erg}\,{M_\odot}^{-1}\,(\epsilon\nu_{100})^{13/21}f_T^{5/7}\nonumber\\
&&\times n_{{\rm H},0}^{2/3}{\dot M_*}^{-8/21}R_{\rm B, 100}^{22/21},
\label{eq:ExMssswtc}
\end{eqnarray}
decreasing weakly with increasing star formation rate.
The x-ray energy produced per solar mass of stars formed may then take
on a wide range of values, depending on $n_{{\rm H},0}$. For
$n_{{\rm H,0}}>n_{\rm H,0,R}\simeq670\,{\rm
cm^{-3}}\,(\epsilon\nu_{100}\dot
M_*)^{4/7}f_T^{-15/14}(\zeta_m+0.15)^{-11/7}R_{\rm B, 100}$,
the rate will be near $10^{47}\,{\rm erg\, M_\odot^{-1}}$, where it
reaches a peak value independent of the ambient hydrogen density and
the star formation rate once radiative cooling restricts the bubble
growth.
It is instructive to compute the thermal heat conduction saturation
parameter for these two limiting cases. Following
\citet{1977ApJ...211..135C}, a consideration of the ratio of the mean
free path of the electrons to the temperature scale height for the
wind, expressed as an \lq inverted cloud,' shows that the surrounding
density and temperature in the cloud case should be replaced by the
central temperature and density of the wind. For a wind limited by
radiative cooling, the saturation parameter becomes
$\sigma_0=(T_c/1.54\times10^7\,{\rm K})^2f_T/(n_{\rm
Hc}R_{\rm B}\phi_s)\simeq(0.05/\phi_s)f_T^{1/2}(\zeta_m+0.15)^{1/2}$,
where $\phi_s\sim1$ characterizes the uncertainty in the saturated
heat flux. This is nearly identical to the value
\citet{1977ApJ...215..213M} derive for interstellar clouds, below
which clouds will cool and condense rather than evaporate. If the wind
bubble is limited instead by the scale height of the disc to a radius
$100\,{\rm pc}\,R_{\rm B, 100}$, the saturation parameter becomes
$\sigma_0\simeq(0.39/\phi_s)f_T^{1/7}(\epsilon\nu_{100}\dot
M_*)^{4/21}n_{\rm H,0}^{-1/3}R_{\rm B, 100}^{-11/21}$.
Thermal heat conduction is thus close to being saturated
($\sigma_0>1$) for typical values of the parameters. Only models with
$\sigma_0<1$ are considered here so that the classical heat conduction
description applies.
In the following section, more precise numerical predictions are made
for the models, including the contribution from metal emission
lines. Comparisons with observations are also drawn.
\section{Numerical evaluation of specific x-ray emission}
\label{sec:numeval_xray}
\subsection{Data and modelling}
\label{subsec:model_xray}
The high angular and spectral resolution of the {\it Chandra X-ray
Observatory} have enabled quantification of the correlation between
the soft x-ray diffuse emission associated with star forming regions
within galaxies and the star formation rate. From measurements of 6
disc galaxies, \citet{2009MNRAS.394.1741O} find
$L_{[0.3-1]\,{\rm keV}}\simeq (2-10)\times10^{38}\,{\rm
erg\,s^{-1}}\dot M_*\,({\rm M_\odot\, yr^{-1}})$.
Only the most luminous point sources were removed, so that their value
may be conservatively viewed as an upper limit to the x-ray luminosity
of a gas component. \citet{2013MNRAS.435.3071L} find a similar
correlation between diffuse galactic coronal emission, corrected for
observed or estimated stellar contributions, and the star formation
rate of
$L_{[0.5-2]\,{\rm keV}}\simeq1.4^{+1.1}_{-0.8}\times10^{39}\,{\rm
erg\,s^{-1}}\dot M_*\,({\rm M_\odot\, yr^{-1}})$
for 53 nearby disc galaxies. Based on star formation rate estimates
from infra-red and UV measurements restricted to the same projected
region as the diffuse x-ray emission, in a sample of galaxies cleaned
of those showing evidence of an active nucleus and with detected or
the estimated contribution of unresolved high mass x-ray binaries
removed, \citet{2012MNRAS.426.1870M} obtained
$L_{[0.5-2]\,{\rm keV}}\simeq(8.3\pm0.1)\times10^{38}\,{\rm
erg\,s^{-1}}\dot M_*\,({\rm M_\odot\, yr^{-1}})$
for a sample of 21 late-type galaxies. (A Salpeter stellar initial
mass function was assumed.) On fitting a two-component thermal model
to the spectra, they find a correlation between the gaseous
contribution to the diffuse x-ray luminosity and the star formation
rate of
\begin{equation}
L_{[0.5-2]\,{\rm keV}}\simeq(5.2\pm0.2)\times10^{38}\,{\rm erg\,s^{-1}}\dot
M_*\,({\rm M_\odot\, yr^{-1}}),
\label{eq:LxSFR}
\end{equation}
or
$E_{[0.5-2]\,{\rm keV}}/M_*\simeq(1.6\pm0.1)\times10^{46}\,{\rm erg\,
M_*^{-1}}$.
For 9 galaxies, they find spectral evidence for substantial absorption
internal to the galaxies. Using these systems, they estimate the {\it
intrinsic} diffuse gaseous emission to be
\begin{equation}
L_{[0.3-10]\,{\rm keV}}\simeq(7.3\pm1.3)\times10^{39}\,{\rm erg\,s^{-1}}\dot
M_*\,({\rm M_\odot\, yr^{-1}}),
\label{eq:LxSFRcorr}
\end{equation}
or
$E_{[0.3-10]\,{\rm keV}}/M_*\simeq(2.3\pm0.4)\times10^{47}\,{\rm erg\,
M_*^{-1}}$.
Since the x-ray emission in the wind models peaks within the energy
bands used to measure the emission, a more precise comparison between
the models and the measurements requires numerical integration of the
models. In addition to thermal free-free, x-ray line emission also
contributes substantially to the overall x-ray budget. The
\texttt{CHIANTI} rates \citep{2012ApJ...744...99L} for collisionally
ionized gas are adopted from Cloudy (13.03)
\citep{2013RMxAA..49..137F}, and emission tables for solar and
half-solar metallicity computed. Numerical integrations of the models
interpolate on the tables. Comparisons with measurements are made
separately below for the steady-state wind model and the superbubble
model.
\subsection{Steady-state wind}
\label{subsec:numssw_xray}
\begin{figure*}
\scalebox{0.45}{\includegraphics{fig01_a.eps}\includegraphics{fig01_b.eps}}
\caption{X-ray emission per solar mass of stars formed for a
steady-state wind. Left panel:\ Shown for the x-ray band $0.5-2$~keV
as a function of the star formation rate, for asymptotic wind
velocities $v_\infty=500$, 1000 and 1500~${\rm km\,s^{-1}}$. The
data points are from \citet{2012MNRAS.426.1870M}. The error bars
represent uncertainties in the distances to the galaxies. Right
panel:\ As in the left panel, but for the x-ray band
$0.3-10$~keV. The data points are from \citet{2012MNRAS.426.1870M},
including correction for internal absorption. The triangles indicate
the upper bound imposed by radiative cooling (see text). A source
region 100~pc in radius and solar metallicity are assumed for both
panels.
}
\label{fig:EmissSSW}
\end{figure*}
The specific x-ray emission in the bands $0.5-2$~keV and $0.3-10$~keV
for the steady-state wind model is shown in Fig.~\ref{fig:EmissSSW} as
a function of star formation rate. A source region of radius
$R=100$~pc is adopted, with solar metallicity. Emission from outside
the source region is included, although it diminishes rapidly with
distance outward. The mass-loading factor is expressed in terms of the
asymptotic wind velocity $v_\infty$.
The specific emissivity is a decreasing function of
$v_\infty$. Expanding the source region to $R=200$~pc is found to
decrease the emission in the $0.5-2$~keV band by about 30 percent, the
same trend, but with a somewhat weaker dependence, as predicted by
Eq.~(\ref{eq:ExMsssw}). The x-ray emission is diminished by 30--50
percent on going from solar to half-solar metallicity for
$v_\infty>500\,{\rm km\,s^{-1}}$.
In the broader energy band $0.3-10$~keV, the specific emissivity,
shown in the right panel of Fig.~\ref{fig:EmissSSW}, decreases with
increasing volume of the source region, varying nearly as rapidly as
$1/R$, as in Eq.~(\ref{eq:ExMsssw}). The specific emissivity
varies nearly linearly with metallicity, except for
$v_\infty\gsim1000\,{\rm km\,s^{-1}}$, for which the specific
emissivity depends only weakly on the metallicity.
The predicted linearly increasing trend with star formation rate is
not consistent with the observations. The data from
\citet{2012MNRAS.426.1870M} suggests a constant amount of x-ray energy
emitted per unit mass of star formed. In the energy band $0.5-2$~keV,
this is matched by allowing a tight correlation between the asymptotic
wind velocity and the star formation rate according to
$v_\infty\simeq1000 {\dot M_*}^{1/6}\,\,{\rm {km\, s^{-1}}}$, corresponding to a
central hydrogen density
$n_{\rm H,0}\simeq1.3 {\dot M_*}^{1/2}\,{\rm cm^{-3}}$. This results
in an increasing amount of mass loading for a decreasing star
formation rate, a general requirement recognized by
\citet{2014ApJ...784...93Z}.
Results for galaxies corrected for internal absorption are shown in
the right hand panel of Fig.~\ref{fig:EmissSSW} for the band
$0.3-10$~keV. The near constancy of the specific emissivity with star
formation rate persists in the data. Agreement with the data may again
be achieved if the wind velocity were tightly correlated with the star
formation rate according to
$v_\infty\simeq700{\rm km\,s^{-1}}{\dot M_*}^{1/6}$, corresponding to
a central hydrogen density
$n_{\rm H,0}\simeq4 {\dot M_*}^{1/2}\,{\rm cm^{-3}}$. The scaling with
star formation rate is bracketed by that expected from
Eqs.~(\ref{eq:Lnuffssw}) and (\ref{eq:ExMsssw}), which give for a
constant luminosity per rate of star formation the approximate
analytic scaling
$v_\infty\sim(\epsilon^2\nu_{100}^2{\dot M_*}/R_{100})^\alpha$ with
$\alpha=1/7-1/5$. The velocity correlations are close to the cooling
restriction Eq.~(\ref{eq:vinfmin}), suggesting a narrow range is
allowed for viable winds \citep[cf][]{2014ApJ...784...93Z}.
\subsection{Superbubble with thermal heat conduction}
\label{subsec:numsbtc_xray}
\begin{figure*}
\scalebox{0.45}{\includegraphics{fig02_a.eps}\includegraphics{fig02_b.eps}}
\caption{X-ray emission per solar mass of stars formed for a
superbubble including thermal heat conduction. Left panel:\ Shown
for the x-ray band $0.5-2$~keV as a function of the star formation
rate, for external hydrogen densities $n_{\rm H,0}=1$, 10, 100 and
1000~${\rm cm\,s^{-3}}$. The data points are from
\citet{2012MNRAS.426.1870M}. Right panel:\ As in the left panel, but
for the x-ray band $0.3-10$~keV. The data points are from
\citet{2012MNRAS.426.1870M}, including correction for internal
absorption. The error bars represent uncertainties in the distances
to the galaxies. A maximum wind radius of 100~pc and solar
metallicity are assumed for both panels.
}
\label{fig:EmissSSTC}
\end{figure*}
Allowing for thermal evaporation from a surrounding medium results in
a much narrower range in specific emissivity compared with the steady
state model. Results assuming a maximum radius of 100~pc for the
expanding bubble and for solar metallicity are shown in
Fig.~\ref{fig:EmissSSTC}. The results shown are time-averaged over the
duration of the spherical expansion of the wind, assumed to cease at
the cooling time, Eq.~(\ref{eq:tcool}), or when it reaches the maximum
radius. X-ray emission only from within the maximum radius is
computed. The emission will fall off rapidly away from the plane if
the superbubble expands out of the disc, but emission from an extended
region may be comparable to that from within the disc. A
multi-dimensional model is required to estimate the full emission more
accurately.
The specific x-ray emissivity increases with the ambient hydrogen
density approximately as $n_{\rm H,0}^{0.7}$, in agreement with
Eq.~(\ref{eq:ExMssswtc}), except at the highest density and low star
formation rate where the wind expansion is cooling limited. At high
densities, the specific emissivity becomes nearly independent of the
star formation rate and gas density, in agreement with
Eq.~(\ref{eq:ExMssswtcc}).
For low star formation rates and low ambient hydrogen densities, the
specific emissivity nearly halves on going from solar to half-solar
metallicity. The x-ray emission is dominated by line emission. The
difference is much more moderate at high star formation rates and high
ambient densities, for which line emission no longer dominates. At low
ambient hydrogen densities, the specific emissivity increases linearly
with the maximum bubble radius, but less rapidly at higher densities
as cooling becomes important, especially for low star formation rates,
in accordance with Eqs.~(\ref{eq:ExMssswtcc}) and
(\ref{eq:ExMssswtc}).
As shown in the left hand panel of Fig.~\ref{fig:EmissSSTC},
comparison with the measured specific emissivities using the data from
\citet{2012MNRAS.426.1870M}, assuming no internal absorption from the
galaxies, shows good agreement for ambient gas densities of
$n_{\rm H,0}\simeq1-10\,{\rm cm^{-3}}$. For a fixed star formation
rate, the required $n_{\rm H,0}$ will scale like $R_{\rm B}^{-11/7}$
according to Eq.~(\ref{eq:ExMssswtc}). Since the measured values
likely exceed the emission from the inner region within the disc by a
factor of a few, the implied hydrogen densities are likely somewhat
smaller. Allowing for internal absorption, agreement with the data in
the right hand panel shows values of
$n_{\rm H,0}\simeq10-100\,{\rm cm^{-3}}$ are preferred. No other
parameters need be adjusted:\ the model predicts the specific x-ray
emissivity is only weakly dependent on the star formation rate, in
agreement with the data.
\section{Specific radio emission}
\label{sec:numeval_rad}
\subsection{Data and modelling}
\label{subsec:model_rad}
The radio continuum radiation emitted by star-forming galaxies scales
with the star formation rate, at least for large radio luminosities
\citep{1992ARAA..30..575C, 2003ApJ...586..794B}. The physical origin
of the emission is unknown, but it is suspected to arise both from
shocks driven by stellar winds and supernovae and from cosmic rays in
a large-scale magnetic field. Measurements suggest that 90 percent of
the continuum emission at 1.4~GHz is synchrotron and 10 percent
thermal free-free in nature, suggesting a component from \hbox{H~$\scriptstyle\rm II\ $}\ regions
as well \citep{1992ARAA..30..575C}. Modelling all these effects is
well beyond the scope of this paper. Here only the synchrotron and
free-free radio emission from the wind regions are estimated. In
comparing with radio data, it is unclear from which scale to take the
emission. The correlation between the radio continuum and the star
formation rate is based on extended regions that likely include
emission from large-scale interstellar cosmic rays. A representative
value is
\begin{equation}
L_{1.4\,{\rm GHz}} ({\rm erg\,s^{-1}\,Hz^{-1}}) \simeq 8.4\times10^{27}\dot
M_*\,({\rm M_\odot\, yr^{-1}})
\label{eq:LrSFR}
\end{equation}
\citep{2002AJ....124..675C}. By contrast, the dominant emission from
shocks within the wind region would be much more centrally
concentrated.
The {\it FIRST} radio survey \citep{1995ApJ...450..559B} includes data
that matches the scale of the x-ray and star-forming regions,
typically up to a few arcminutes, measured by
\citet{2012MNRAS.426.1870M}. We compare the models with two 1.4~MHz
continuum measurements from the {\it FIRST} survey, the large-scale
value centred on each galaxy and the brightest unresolved peak value
in the nucleus of the galaxy, corresponding typically to a region
within the central 100--500~pc of the galaxy for a source at a
distance of 10--20~Mpc. As shown below, the large scale values agree
well with Eq.~(\ref{eq:LrSFR}), corresponding to
$L_{1.4\,{\rm GHz}}/{\dot M_*}\simeq2.6\times10^{35}\,{\rm
erg\,Hz^{-1}\,M_\odot^{-1}}$.
The synchrotron and thermal free-free emission are computed for the
models as in \citet{2013MNRAS.430.2854M}. In brief, a power-law energy
distribution $dn/d\epsilon\sim\epsilon^{-p_e}$ is assumed for the
relativistic electrons, with an energy density a fraction $f_e$ of the
local thermal energy density. The magnetic field energy density is
also taken to be $f_e$ for simplicity, corresponding to approximate
equipartition. The relativistic electrons are allowed to cool by
synchrotron and thermal free-free radiation and by excitation of
plasmon waves following the passage of the wind-driven shock front
into the interstellar gas. Thermal free-free and synchrotron
self-absorption are included, although for the frequencies of interest
these are generally negligible in the models considered. Observations
of supernova remnant spectra suggest typical values for the
relativistic electron energy index of $2<p_e<3$
\citep{1998ApJ...499..810C, 1986ApJ...301..790W}, while representative
model values for the relativistic electron energy density fraction
range over $0.001 < f_e < 0.2$ \citep{2006ApJ...641.1029C}. The
predictions of the wind models for radio emission are estimated to
check they do not exceed the observed limits for plausible
parameters. Virtually all the emission predicted by the models is
synchrotron radiation; the thermal free-free component is two to three
orders of magnitude smaller.
\subsection{Steady-state wind}
\label{subsec:numssw_rad}
\begin{figure*}
\scalebox{0.45}{\includegraphics{fig03_a.eps}\includegraphics{fig03_b.eps}}
\caption{Radio continuum emission at 1.4~GHz per solar mass of stars
formed for a steady-state wind. Shown for a relativistic electron
energy index $p_e=2$ and energy density fraction $f_e=0.03$ (left
panel) and $p_e=3$, $f_e=0.1$ (right panel), as a function of the
star formation rate, for asymptotic wind velocities $v_\infty=500$,
1000 and 1500~${\rm km\,s^{-1}}$. The emission adopting
$v_\infty=700{\dot M_*}^{1/6}\,\,{\rm {km\, s^{-1}}}$ is shown by the solid (black)
lines. The data points are from the {\it FIRST} radio survey. The
error bars represent distance uncertainties. Open points represent
the large-scale emission; filled points represent the peak
unresolved emission (see text). A wind source region 100~pc in
radius is assumed for both panels.
}
\label{fig:RadioSSW}
\end{figure*}
The x-ray measurements require models with mass-loading from a gas
reservoir surrounding the supernovae. The wind will then drive a shock
into the surroundings. Since the wind is in a steady state, however,
the time since the shock passed a given radius is undetermined in the
model. To allow an estimate of the synchrotron emission, the wind is
arbitrarily assumed to have reached a distance of 10~kpc from the
source region, corresponding to an age of $t_{\rm age}=10\,{\rm
kpc}/v_\infty$.
For a characteristic wind age of $10^7t_{w,7}\,{\rm yr}$, for low star
formation rates the synchrotron emission is dominated by the source
region $r<R$, giving, for $p_e=2$,
\begin{eqnarray}
\frac{S_{\rm 1.4GHz}}{\dot M_*}&\simeq&5\times10^{37}\,{\rm
erg\,Hz^{-1}\,M_\odot^{-1}}\,\nu_{\rm GHz}^{-1/2}f_e^{7/4}R_{100}^{-1/2}\nonumber \\
&&\times \left(\frac{1000\,{\rm
km\,s^{-1}}}{v_\infty}\right)^{7/4}\left(\epsilon\nu_{100}\right)^{7/4}
{\dot M_*}^{3/4}.
\label{eq:SnuMss}
\end{eqnarray}
For a typical value $f_e=0.01$, $S_{\rm 1.4GHz}/{\dot
M_*}\simeq1.2\times10^{34}\,{\rm erg\,Hz^{-1}\,M_\odot^{-1}}$.
At higher star formation rates, the density becomes sufficiently high
that two additional effects become important:\ synchrotron and plasmon
generation losses deplete the central region of relativistic
electrons; for sufficiently strong synchrotron losses the critical
frequency falls to several gigahertz or less. Emission then arises
only from outside the source region. The criterion for plasmon and
synchrotron losses not to deplete the population of relativistic
electrons is $f_en_{\rm H}Pt_{w,7}^2<8\times10^{-11}\,{\rm
dyne\,cm^{-5}}$, where $P$ is the gas pressure in the wind. Using
the asymptotic limits for density and pressure, this requires emission
to arise only from radii
\begin{equation}
\frac{r}{R}>1.1f_e^{3/16}R_{100}^{-3/4}t_{w,7}^{3/8}\left(\epsilon\nu_{100}{\dot
M_*}\right)^{3/8}\left(\frac{1000\,{\rm
km\,s^{-1}}}{v_\infty}\right)^{3/4},
\label{eq:xspl}
\end{equation}
resulting in a rate of radio energy generation at frequency $10^9\nu_{\rm
GHz}$~Hz per solar mass of stars formed of
\begin{eqnarray}
\frac{S_\nu}{\dot M_*}&\simeq&4\times10^{37}\,{\rm
erg\,Hz^{-1}\,M_\odot^{-1}}\, \nu_{\rm
GHz}^{-1/2}f_e^{39/32}R_{100}^{13/8}\nonumber\\
&&\times \left(\frac{v_\infty}{1000\,{\rm
km\,s^{-1}}}\right)^{3/8}\left(\epsilon\nu_{100}\right)^{11/16}{\dot
M_*}^{-5/16}t_{w,7}^{-17/16}.
\label{eq:Snuspl}
\end{eqnarray}
At even higher densities, once synchrotron losses lower the synchrotron
critical frequency into the gigahertz range, emission above
frequency $\nu_{\rm GHz}$ occurs only at radii
\begin{equation}
\frac{r}{R}>3.2\nu_{\rm GHz}^{1/5}f_e^{3/10}R_{100}^{1/5}t_{w,7}^{2/5}\left(\epsilon\nu_{100}{\dot
M_*}\right)^{3/10}\left(\frac{1000\,{\rm
km\,s^{-1}}}{v_\infty}\right)^{3/10},
\label{eq:xs}
\end{equation}
resulting in a rate of radio energy generation per solar mass of stars formed
\begin{eqnarray}
\frac{S_\nu}{\dot M_*}&\simeq&3\times10^{35}\,{\rm
erg\,Hz^{-1}\,M_\odot^{-1}}\, \nu_{\rm
GHz}^{-16/15}f_e^{9/10}R_{100}^{-16/15}\nonumber\\
&&\times \left(\frac{1000\,{\rm
km\,s^{-1}}}{v_\infty}\right)^{9/10}\left(\epsilon\nu_{100}\right)^{9/10}{\dot
M_*}^{-1/10}t_{w,7}^{-17/15}.
\label{eq:Snus}
\end{eqnarray}
Results from numerically integrating the model are shown in
Fig.~\ref{fig:RadioSSW}. The 1.4~GHz emission is displayed as a
function of the star-formation rate and $v_\infty$ for $p_e=2$,
$f_e=0.03$ and $p_e=3$, $f_e=0.1$. All three trends with the star
formation rate are apparent:\ a rise ($\dot M_*^{3/4}$), followed by a
decline ($\dot M_*^{-5/16}$) once synchrotron and plasmon losses pinch
off the relativistic electron distribution within the wind source
region, and finally a near constant level ($\dot M_*^{-1/10}$) once
synchrotron losses restrict the generation of 1.4~GHz power to regions
well outside the core. (The power at large star formation rates for
the $v_\infty=1500\,{\rm {km\, s^{-1}}}$ case differs from the trend with $v_\infty$ in
Eq.~(\ref{eq:Snus}) because the radius at which the critical frequency
exceeds 1.4~GHz lies just outside the source region, where the
asymptotic decrease of pressure with radius is no longer a good
approximation.) For a high star formation rate, the radio spectrum for
$p_e=2$ is found to steepen to $\nu^{-\alpha_S}$ with
$\alpha_S\simeq1.5$, rather than the expected
$\alpha_S=(p_e-1)/2=1/2$, as is found for low star formation rates:\
the truncation of the emitting volume at high densities steepens the
spectrum.
The predicted trend adopting the correlation
$v_\infty\simeq700\,{\rm {km\, s^{-1}}}\, {\dot M_*}^{1/6}$ is shown by the solid
(black) curves. For $p_e=2$, the fluxes for the unresolved radio
sources are largely recovered or exceeded. The high flux values for
the highest star-formation rates are matched only for
$f_e\simeq0.1-0.2$, the limiting values inferred for supernova
remnants. These high fluxes, however, are derived from distant
galaxies within regions unresolved on the scales of 1--2~kpc, so the
radio emission may be contaminated by emission from cosmic rays
interacting with large-scale galactic magnetic fields. The predicted
excess emission for unresolved sources with low flux values may
indicate a reduced volume filling factor of emitting electrons,
resulting in small volume-averaged values for the relativistic energy
fraction. At high star formation rates, the radio flux also decreases
inversely with the bubble size, according to
Eq.~(\ref{eq:Snus}). Alternatively, the relativistic electron energy
distribution may be steeper. For $p_e=3$, $f_e\approx0.1$ recovers the
lower flux values of the unresolved sources.
\subsection{Superbubble with thermal heat conduction}
\label{subsec:numsbtc_rad}
\begin{figure*}
\scalebox{0.45}{\includegraphics{fig04_a.eps}\includegraphics{fig04_b.eps}}
\caption{Radio continuum emission at 1.4~GHz per solar mass of stars
formed for the superbubble model. Shown for relativistic electron
distribution index $p_e=2$ and energy density fraction $f_e=0.01$
(left panel) and $p_e=3$, $f_e=0.1$ (right panel), as a function of
the star formation rate, for external hydrogen densities
$n_{\rm H,0}=1$, 10, 100 and 1000~${\rm cm\,s^{-3}}$. The data
points are from the {\it FIRST} radio survey. Open points represent
the large-scale emission; filled points represent the peak
unresolved emission (see text). A maximum bubble radius of 100~pc is
adopted and solar metallicity.
}
\label{fig:RadioSSTC}
\end{figure*}
For the superbubble model, the characteristic radio emission for a
scale-height limited superbubble from within $R_{\rm B, 100}$ is given
approximately by
\begin{eqnarray}
\frac{S_\nu}{\dot M_*}&\simeq&1.0\times10^{38}\,{\rm
erg\,Hz^{-1}\,M_\odot^{-1}}\, \nu_{\rm
GHz}^{-1/2}f_e^{7/4}\nonumber\\
&&\times(\epsilon\nu_{100})^{7/6}n_{\rm H,0}^{7/12}{\dot
M_*}^{1/6}R_{\rm B, 100}^{2/3},
\label{eq:Snustc}
\end{eqnarray}
nearly independent of the star formation rate. The radio emission is
independent of any suppression of thermal conductivity since the radio
power is determined by the thermal energy density, not the gas density
or temperature separately in the absence of significant
attenuation. Results from numerical integration of the wind equations
are shown in Fig.~\ref{fig:RadioSSTC}. The bubble region recovers the
full range of measured large-scale radio power for $p_e=2$ and
$f_e=0.01$. The larger flux values, at high star formation rates,
require a high ambient hydrogen density of
$n_{{\rm H},0}\sim100\,{\rm cm^{-3}}$, the upper value required by the
x-ray data. The lower power from unresolved regions is over-predicted
for the higher star formation rates, possibly indicating a reduced
volume filling factor of emitting electrons. The steeper electron
distribution case with $p_e=3$ requires increasing the relativistic
electron energy fraction to $f_e=0.1$, approaching the limit from
supernova remnant modelling. A near constant specific radio power is
found for a given ambient hydrogen density, only weakly dependent on
the star formation rate, in agreement with Eq.~(\ref{eq:Snustc}).
\section{Metal column densities}
\label{sec:metals}
\subsection{Data and modelling}
\label{subsec:model_metals}
An estimate of the column densities of metal ions within the winds may
be made as follows. For solar metallicity, abundances by number of
commonly detected metal atoms compared with hydrogen include:\
$\log_{10}\xi_{\rm He} = -1.07$, $\log_{10}\xi_{\rm C} = -3.57$,
$\log_{10}\xi_{\rm N} = -4.17$, $\log_{10}\xi_{\rm O} = -3.31$,
$\log_{10}\xi_{\rm Si} = -4.49$ and $\log_{10}\xi_{\rm S} = -4.88$
\citep{2009ARAA..47..481A}. Any given ionization state will dominate
at a particular temperature where it contributes most to the column
density of that ion, although for some species neighbouring ionization
states share substantially in the ionization. Temperatures at which
commonly measured ions peak include: \hbox{C~$\scriptstyle\rm III\ $} at $10^{4.8}$~K
($10^{-3.6}$), \hbox{C~$\scriptstyle\rm IV\ $} at $10^{5.0}$~K ($10^{-4.1}$), \hbox{N~$\scriptstyle\rm II\ $} at
$10^{4.4}$~K ($10^{-4.2}$), \hbox{N~$\scriptstyle\rm III\ $} at $10^{4.9}$~K ($10^{-4.2}$), \hbox{O~$\scriptstyle\rm VI\ $}
at $10^{5.5}$~K ($10^{-4.2}$), \hbox{Si~$\scriptstyle\rm III\ $} at $10^{4.5}$~K ($10^{-4.6}$),
\hbox{Si~$\scriptstyle\rm IV\ $} at $10^{4.8}$~K ($10^{-5.0}$), SIII at $10^{4.7}$~K
($10^{-5.0}$) and \hbox{S~$\scriptstyle\rm IV\ $} at $10^{5.0}$~K ($10^{-5.1}$), where the peak
abundance fractions by number relative to hydrogen,
$\xi_{i,{\rm max}}$, are indicated in parentheses. It is noted these
values will be modified if the gas is not in collisional ionization
equilibrium, as appears to be the case for some high velocity clouds
in the Galactic halo \citep{2011ApJ...739..105S}.
\begin{figure*}
\scalebox{0.45}{\includegraphics{fig05_a.eps}\includegraphics{fig05_b.eps}}
\caption{Column densities for selected metal ions, shown for the
steady state model for $v_\infty=1000\,{\rm {km\, s^{-1}}}$ (left panel) and the
superbubble model for $n_{\rm H,0}=10\,{\rm cm}^{-3}$ (right
panel). Also shown is the \hbox{H~$\scriptstyle\rm I\ $}\ column density. A source region of
100~pc radius is adopted for the steady state model, and a maximum
bubble radius of 100~pc for the superbubble model. Solar metallicity
is assumed for both.
}
\label{fig:MetalsSSTC}
\end{figure*}
For the steady-state wind model with $v_\infty>500\,{\rm {km\, s^{-1}}}$ and wind core
radius $R=100$~pc, the temperatures at which these ion abundances peak
are achieved only outside the core, where the gas density is rapidly
declining.
The column density of a typical ion like
$\hbox{C~$\scriptstyle\rm III\ $}$ is negligibly small, $N_{\rm CIII}\simeq10^7\,{\rm
cm^{-2}}$. Of the ions listed above, only \hbox{O~$\scriptstyle\rm VI\ $}\ would achieve a
measurable column density within the core, $N_{\rm
OVI}\simeq10^{14}\,{\rm cm^{-2}}$. It is noted, however, that at the
interface of the wind shock and the interstellar medium in the
galactic disc, detectable levels of absorption may arise
\citep{1996ApJS..102..161D}. Such systems could possibly be
distinguished from those produced by superbubbles, discussed below,
through their kinematics.
For $r\gg R$, the wind temperature is
$T\simeq 6.7v_\infty^2(R/r)^{4/3}$. For a given ion $i$, the radius
and density at which $T=T_{\rm max}=10^5\,{\rm K}\,T_{\rm max,5}$ are
$r_{\rm max}\simeq23 R T_{\rm max, 5}^{-3/4}(v_\infty/1000\,{\rm {km\, s^{-1}}})^{3/2}$
and
$n_{\rm H, max}\simeq4.4\times10^{-4}\,{\rm cm^{-3}}T^{3/2}_{\rm max,
5}(v_\infty/1000\,{\rm {km\, s^{-1}}})^{-6}\epsilon\nu_{100}{\dot M_*}R_{100}^{-2}.$
The corresponding column density will be
\begin{eqnarray}
N_i&\simeq& n_{\rm H,max}r_{\rm max}\xi_{i, {\rm max}}\nonumber\\
&\simeq&5.7\times10^{14}\,{\rm
cm^{-2}}\epsilon\nu_{100}{\dot
M_*}R_{100}^{-1}\left(\frac{1000\,{\rm {km\, s^{-1}}}}{v_\infty}\right)^{9/2}\nonumber\\
&&\times\left(\xi_{i,{\rm max}}T_{\rm max}^{3/4}\right).
\label{eq:coldenSS}
\end{eqnarray}
The column densities decline very rapidly with
$v_\infty$. Representative values for some common ions, computed
numerically by integrating along the full wind solution at $r>R$, are
shown in Fig.~\ref{fig:MetalsSSTC} (left panel). These represent the
maximum column densities that would arise from the homogeneous wind
for lines of site passing through the region $r\simeq r_{\rm max}$. At
larger distances, the column densities will rapidly decline.
In the superbubble model, the column densities are dominated by
absorption from a very thin layer at the bubble interface with the
interstellar medium. For the scale-height limited case with a bubble
radius $R_{\rm B}$, the characteristic temperature scale height at the
position where $T=T_{\rm max}$ for a given ion is $L_T=\vert dr/d\log
T\vert_{\rm max}=(5/2)R_{\rm B}(T_{\rm
max}/T_c)^{5/2}\simeq5.5\times10^{-7}R_{\rm B}(f_T R_{\rm B,
100}/\epsilon\nu_{100}\dot M_*)^{5/7}T_{\rm max,5}^{5/2}$,
corresponding to a hydrogen density $n_{\rm H, max}\simeq110\,{\rm
cm^{-3}}\,n_{\rm H,0}^{1/3}{\epsilon\nu_{100}\dot M_*}^{2/3}R_{\rm
B, 100}^{-4/3}T_{\rm max,5}^{-1}$. The column density of ion $i$ is
then
\begin{eqnarray}
N_i&\simeq& n_{\rm H, max}\xi_{i, {\rm max}} L_T \label{eq:coldenTC}\nonumber\\
&\simeq& 5.7\times10^8\,{\rm cm^{-2}}\,n_{\rm H,0}^{1/3}(\epsilon\nu_{100}{\dot
M_*})^{-1/21}R_{\rm B, 100}^{8/21}f_T^{5/7}\nonumber\\
&&\times\left(\xi_{i, {\rm max}}T_{\rm max}^{3/2}\right),
\end{eqnarray}
nearly independent of the star formation rate and only weakly
sensitive to the ambient gas density and size of the
bubble. Representative values, computed numerically by integrating
through the centre of the region $r<R_{\rm B}$, are shown in
Fig.~\ref{fig:MetalsSSTC} (right panel). This is a minimal value that
must arise in the thermal conductive interface with the interstellar
medium. Substantially lower values would be evidence against the
superbubble model. The column density for \hbox{H~$\scriptstyle\rm I\ $}\ is also a minimal
value, and it will generally be small compared with the \hbox{H~$\scriptstyle\rm I\ $}\ column
density through the surrounding disc.
\section{Discussion}
\label{sec:Discussion}
\subsection{X-ray constraints}
\begin{figure*}
\scalebox{0.45}{\includegraphics{fig06_a.eps}\includegraphics{fig06_b.eps}}
\caption{X-ray emission in the 2--10~keV band per solar mass of stars
formed for a steady-state wind (left panel) and for a superbubble
(right panel), both as a function of the star formation rate. A
source region of 100~pc radius is adopted for the steady state
model, and a maximum wind radius of 100~pc for the
superbubble. Solar metallicity is assumed for both models.
}
\label{fig:EmissSSWTC}
\end{figure*}
Both the freely expanding steady-state wind model and the
self-similarly expanding superbubble model with thermal heat
conduction may account for the measured amount of diffuse soft x-ray
energy generated per unit mass in stars formed. An additional
assumption of a tight correlation between the asymptotic wind velocity
and the star formation rate, however, is required for the steady state
model. By contrast, thermal evaporation from the cavity walls in the
superbubble model naturally accounts for the measured amount of difuse
x-ray energy per unit mass in stars formed for characteristic
interstellar gas densities $1<n_{\rm H,0}<100\,{\rm cm^{-3}}$, with
the higher values favoured if much of the soft x-ray emission is
absorbed internally to the galaxies.
In the steady state model, the required correlation between the
asymptotic wind velocity and the star formation rate, particularly for
the high x-ray luminosities when internal galactic absorption is
allowed for, is close to the minimum wind velocity
(Eq.~[\ref{eq:vinfmin}]) for which a steady-state wind may be
maintained against radiative cooling within the star-forming region,
consistent with the narrow range in observed radiative
efficiencies. But it does not provide a reason for the narrow
range. One possibility is that the winds are driven by
superbubbles. Once a superbubble expands to the scale-height of the
galactic disc, its thermal pressure drives a vertical conical outflow
rather than further expansion into the disc
\citep{1985ApJ...299...24S, 1988ApJ...324..776M}. Simulations suggest
the outflow is nearly adiabatic \citep{2015MNRAS.453.3499K}, so that
it may be approximated by the steady state model with a superbubble as
the source. The rate of mechanical energy injected by supernovae and
the rate of evaporative mass loss from the disc may be used to define
an effective asymptotic wind velocity for the superbubble:
\begin{equation}
v_\infty^{\rm eff}=\left(\frac{2\dot E}{\dot M_{\rm
ev}}\right)^{1/2}=850\,\,{\rm {km\, s^{-1}}} \left(\epsilon\nu_{100}{\dot M_*}/f_TR_{\rm
B, 100}\right)^{1/7},
\label{eq:vinfeff}
\end{equation}
close to the required relation found for the steady-state wind
solution. This may reconcile the two models:\ thermal heat conduction
sets the source terms that initiate the wind, which then \lq blows
out' vertically into a steady-state outflow
\citep{1989ApJ...337..141M, 1999ApJ...513..142M, 2003ApJ...599...50F,
2004ApJ...613..159F, 2014MNRAS.442.3013K}.
For a bubble to blow out, two criteria must be satisfied:\ the cooling
radius must exceed the disc scale height and the bubble velocity must
exceed the sound speed in the surrounding medium. Both of these may be
expressed as a restriction on the average star formation rate per
superbubble cross-sectional area, $\dot\Sigma_{\rm B}={\dot M_*}/\pi
R_{\rm B}^2$. The cooling criterion gives
\begin{eqnarray}
\dot\Sigma_{\rm B}>\dot\Sigma_{\rm B, R} &\simeq&0.0004\,{\rm M_\odot}\,{\rm
yr^{-1}\, kpc^{-2}}\,n_{\rm
H,0}^{7/4}R_{\rm B, 100}^{3/4}\nonumber\\
&&\times(\epsilon\nu_{100})^{-11/4}f_T^{15/8}(\zeta_m+0.15)^{21/8}.
\label{eq:SBR}
\end{eqnarray}
The hydrodynamical computations of \citet{1988ApJ...324..776M} suggest
for the dynamical criterion that, in terms of their dynamical variable
$D$, the bubble velocity must exceed the disc sound speed by a factor
$D^{1/3}\simeq4.6$, for $D>100$. This gives
\begin{equation}
\dot\Sigma_{\rm B}>\dot\Sigma_{\rm B, R} \simeq0.003\,{\rm M_\odot}\,{\rm
yr^{-1}\, kpc^{-2}}\,T_{\rm d,4}^{3/2}\frac{n_{\rm H,
0}}{\epsilon\nu_{100}},
\label{eq:SBD}
\end{equation}
where $T_{\rm d,4}$ is the temperature of the ambient disc gas in
units of $10^4$~K. The dynamical criterion is similar to the estimate
of \citet{2004ApJ...606..829S}. For
$n_{\rm H, 0}\sim1-10\,{\rm cm^{-3}}$, these criteria give
$\dot\Sigma_{\rm B}>0.003-0.03\,{\rm M_\odot}\,{\rm yr^{-1}\, kpc^{-2}}$,
comparable to the minimum observed star formation surface density in
galaxies with winds \citep{2005ARAA..43..769V}. This raises the
question:\ do the proxies for star formation actually probe the larger
region of a superbubble, so that the minimum observed star formation
surface density in galaxies with winds may be identified with the
minimum $\dot\Sigma_{\rm B}$ required for blowout?
Differences in the hard x-ray luminosities of the wind cores are
expected between the two models. In the steady-state model, the
exponential sensitivity to the star formation rate results in a rapid
decrease with decreasing star formation rate of the 2--10~keV
luminosity when the correlation $v_\infty\simeq700{\dot M_*}^{1/6}$ is
imposed, as shown in Fig.~\ref{fig:EmissSSWTC}. At
$E_{[2-10]{\rm keV}}/M_*\simeq2\times10^{45}\,{\rm erg\,M_\odot^{-1}}$
for $\dot M_*=1\,{\rm M_{\odot}\,yr^{-1}}$, the specific emissivity is
up to an order of magnitude smaller than the predicted value for a
superbubble with $10<n_{\rm H,0}<100\,{\rm cm^{-3}}$. For
$M_*=0.1\,{\rm M_{\odot}\,yr^{-1}}$, at
$E_{[2-10]{\rm keV}}/M_*\simeq2\times10^{43}\,{\rm erg\,M_\odot^{-1}}$
the predicted specific emissivity in the steady state model is nearly
two orders of magnitude smaller than for a superbubble with
$n_{\rm H,0}\simeq1\,{\rm cm^{-3}}$. The hard x-ray luminosity of the
wind cores may thus serve to discriminate between the models.
Both models have been scaled to a supernova ejecta mechanical energy
of $10^{51}\epsilon\,{\rm erg}$. Observations suggest a range of
$0.6<\epsilon<1.5$ \citep{1989ARA&A..27..629A}. The effects on the
values of $v_\infty$ in the steady state model or $n_{\rm H,0}$ in the
superbubble model required to match observations may be inferred from
the scaling relations in Eqs~(\ref{eq:ExMsssw}) and
(\ref{eq:ExMssswtc}). This is a systematic uncertainty of the models,
and applies as well to the predictions for radio power and metal
column densities.
\subsection{Radio constraints}
The radio spectrum in the steady state model is expected to steepen
due to the depletion of relativistic electrons by synchrotron and
plasmon losses over a timescale of $\sim10$~Myr. The superbubble model
is computed only until the bubble emerges from the galactic disc,
which is typically a much shorter time. An assessment of the long term
energy losses of the relativistic electron population may be
approximated by imposing energy losses over a characteristic age of
10~Myr on the bubble. Doing so shows that, because of the low gas
density inside the bubble cavity, the relativistic electron population
is not depleted by synchrotron or plasmon losses, so that the emission
persists without steepening. Although the intrinsic electron
distribution may provide a steep spectrum, this is not required. A
hard spectrum near 1.4~GHz may then provide a means of discriminating
between the superbubble model and the conduction-free steady state
model.
While both the steady state and the superbubble models are consistent
with the measured radio powers for plausible values of the synchrotron
model parameters, the data do not require these values. Other
combinations will match the data. It is also possible other
mechanisms, such as large-scale cosmic ray emission, dominate the
measured power if $f_e\ll1$. Radio measurements that resolve the
star-forming regions in starbursts may help to clarify the
contribution of shock-accelerated electrons within the wind-generating
region to the total synchrotron radio emission.
\subsection{Metal column density constraints}
\begin{figure}
\scalebox{0.45}{\includegraphics{fig07.eps}}
\caption{Column densities for selected metal ions, shown for the
steady state model with $v_\infty=700 {\dot M_*}^{0.17}\,{\rm {km\, s^{-1}}}$. Also
shown is the \hbox{H~$\scriptstyle\rm I\ $}\ column density. A source region of 100~pc radius
and solar metallicity are adopted.
}
\label{fig:MetalsSSC}
\end{figure}
\begin{figure}
\scalebox{0.45}{\includegraphics{fig08.eps}}
\caption{Mass loading factor $\beta_*(\zeta_m/R_{100})^{2/7}$ as a
function of star formation rate. To allow for variations in the size $R=100R_{100}$~pc of the star-forming
regions and the gas metallicity $\zeta_m$, the mass loading factor has
been scaled by the factor $(\zeta_m/R_{100})^{2/7}$ (see text). The
upper limit imposed by radiative cooling in the freely-expanding
steady state model is shown for $\epsilon=1$ (filled triangles) and
$\epsilon=1.5$ (open triangles). The predicted value in the steady
state model for $v_\infty=700\,{\rm {km\, s^{-1}}}(\epsilon^2\nu^2_{100}\dot
M_*/R_{100})^{1/7}$ is shown for $\zeta_m=1$ (thick blue solid line)
and $\zeta_m=0.25$ (thin blue solid line). The prediction for the
superbubble model when the bubble reaches size $R$ is shown for
$\zeta_m=1$ (thick magenta broken line) and $\zeta_m=0.25$ (thin
magental broken line). The data are from
\citet{2015ApJ...809..147H}:\ the dark (blue) circles represent
strong cloud outflows and light (green) circles represent weak or no
cloud outflows.
}
\label{fig:betas}
\end{figure}
Imposing the approximate analytic velocity correlation
$v_\infty\simeq700\,{\rm {km\, s^{-1}}}(\epsilon^2\nu_{100}^2\dot M_*/R_{100})^{1/7}$
(Sec.~\ref{subsec:numssw_xray}) for the steady state model results in
ion column densities depending only weakly on the star formation rate
and size of the wind-generating region,
\begin{eqnarray}
N_i&\simeq&2.8\times10^{15}\,{\rm
cm^{-2}}(\epsilon\nu_{100})^{-2/7}\nu_{100}({\dot
M_*}/R_{100})^{5/14}\nonumber\\
&&\times\left(\xi_{i,{\rm max}}T_{\rm max}^{3/4}\right).
\label{eq:coldenSSv700}
\end{eqnarray}
The metal column densities computed after imposing the similar
numerical correlation $v_\infty=700\,{\rm {km\, s^{-1}}}{\dot M_*}^{1/6}$ on the
numerically integrated model with $R=100$~pc are shown in
Fig.~\ref{fig:MetalsSSC}. Commonly detected ions will have column
densities between $10^{14}-10^{16}\,{\rm cm^{-2}}$, comparable to
those measured \citep{2001ApJ...554.1021H}. The ion abundances peak,
however, at characteristic temperatures near the peak in the cooling
curve, so that the gas will be thermally unstable. The computed column
densities would then represent minimal values:\ as the gas condenses,
the column densities will increase, although the covering fraction of
the absorption systems will decline as the gas fragments. Measurements
suggest the column densities of measured systems are determined by the
velocity at which they flow through coronal temperatures
\citep{1986ApJ...310L..27E, 2002ApJ...577..691H}, for which the column
density will be of the order
$(1/\sigma_i){\mathcal M}_{\rm
cool}$\footnote{\citet{1986ApJ...310L..27E}
and \citet{2002ApJ...577..691H} argue the cooling column density is
$3kT v_{\rm cool}/\Lambda(T)$, where $v_{\rm cool}$ is the flow
velocity of the cooling gas and
$\Lambda(T)\sim E_i\sigma_i(T) v_e(T)$ is the cooling coefficient
due to collisional excitation of a transition with energy $E_i$, and
$v_e$ is the velocity of the electrons. Coronal temperatures
correspond to $k_{\rm B}T\sim E_i$.}, where $\sigma_i(T)$ is the
characteristic collisional cross section of the ion and
${\mathcal M}_{\rm cool}(T)$ is the Mach number of the flow.
The neutral hydrogen column density arises from a recombination layer
at $T\sim10^4$~K that will occur at radial distances
$r\simeq76R(\epsilon\nu_{100})^{3/7}(\dot M_*/R_{100})^{3/14}$, typically
5--15~kpc. The layer will be optically thick to any penetrating
external photoionizing radiation that may modify the predicted metal
column densities, although they would still be susceptible to
photoionizing radiation from the starburst. Thermal instabilities,
however, may cause the layer to fragment so that its structure may
become porous.
The superbubble model predicts metal column densities much smaller
than measured. This was recognised by \citet{2001ApJ...554.1021H}, who
suggested dynamical instabilities as the superbubble emerges from the
disc will develop turbulent mixing layers that, in sufficient number,
could produce the measured column densities.
Measurements of absorption lines produced by clouds in wind outflows
have been used by \citet{2015ApJ...809..147H} to assess the amount of
mass loading in the winds.\footnote{The velocities of the clouds
determined from the widths of their absorption features are not
necessarily to be identified with the outflow velocity of the wind
gas impinging on them; the clouds will generally be accelerated by
ram pressure to lower velocities than the flow because of their
higher densities \citep[eg][]{2015ApJ...805..158S}.} Their results
are shown in Fig.~\ref{fig:betas} along with the predictions of the
models. The measured values have been rescaled to account for the
variation in the star-forming region sizes and the
metallicities. Following the terminology of these authors, the points
have also been designated as either corresponding to a strong outflow
or to a weak or no outflow. The upper values lie close to the maximum
allowed by cooling in the steady state model, and include many of the
weak or no outflows. Adopting the correlation
$v_\infty=700\,{\rm {km\, s^{-1}}} (\epsilon^2\nu_{100}^2\dot M_*/R_{100})^{1/7}$
results in values matching the median of those measured, corresponding
to strong outflows.
The lowest mass loading values lie near the value for the superbubble
model, and correspond to many of the weak or no outflows. (Here, the
measured star formation region is identified with the superbubble
radius.) This supports the possibility that the steady-state wind is
initiated by a superbubble. The high metal column densities may then
arise within an outflow region at coronal temperatures, as for the
steady-state wind. Further mass loading from the hydrodynamical
ablation or thermal evaporation of clouds above the plane would
account for the larger mass loading factors. Observational evidence
suggests thermal conduction is active within halo clouds
\citep{2005MNRAS.362..626M}. The clouds may have pre-existed or been
accreted from larger scales, as simulations suggest entrained clouds
have too short survival times to have originated in the disc
\citep{1994ApJ...420..213K, 2015ApJ...805..158S} unless possibly
supported by magnetic field pressure \citep{1994ApJ...433..757M,
2015MNRAS.449....2M}.
\section{Conclusions}
\label{sec:conclusions}
Two models for the generation of galactic winds are compared. While
both model the winds as powered by supernovae explosions in active
areas of star formation, the models differ fundamentally in their
microphysical assumptions. One neglects thermal heat conduction and
treats the wind in a steady state of free expansion, as may be
expected for an age large compared with the flow time. The second
includes thermal heat conduction, resulting in time-dependent
self-similar expansion into a surrounding medium. Observational
predictions of the models are compared with x-ray, radio and UV metal
line absorption measurements. The principal conclusions are summarised
below.
\subsection{Steady-state wind}
\label{subsec:ssm}
The steady state model is characterised by two principal parameters,
chosen here to be the star formation rate $\dot M_*$ and the
asymptotic wind velocity $v_\infty$. The wind is assumed to be
generated by uniformly distributed supernovae within a core region of
radius $R=100R_{100}$~pc. The requirement that the energy generation
rate exceed the cooling rate imposes a lower limit on the asymptotic
velocity of
$v_\infty>550\,{\rm km\,s^{-1}}(\epsilon\nu_{100}{\dot M_*}\zeta_m
R_{100}^{-1})^{0.14}$,
where $\zeta_m$ is the metallicity of the wind gas relative to solar
and $\nu_{100}$ and $\epsilon$ are scaling parameters describing the
expected supernovae rate per star formed and their amount of energy
injection. This restriction constrains the mass-loading factor to
$\beta_*<3.3(\epsilon\nu_{100})^{0.73}(\dot
M_*\zeta_mR_{100}^{-1})^{-0.27}$.
The amount of soft x-ray energy generated per unit mass of stars
formed scales approximately as $E_x/M_*\sim {\dot M_*}/
(v_\infty^{1/\alpha} R)$ with $\alpha=1/7-1/5$. This is contrary to
the measurements of \citet{2012MNRAS.426.1870M}, who find a constant
amount of soft x-ray energy generation per unit mass of stars formed,
independent of the star formation rate. The amount measured in the
0.5--2~keV band of $2\times10^{46}\,{\rm erg\, M_\odot^{-1}}$ is
matched if the asymptotic wind velocity correlates with the star
formation rate according to $v_\infty\simeq 1000\,{\rm {km\, s^{-1}}}
(\epsilon^2\nu_{100}^2\dot M_*/R_{100})^\alpha$ with
$\alpha\sim1/7-1/5$. A lower asymptotic speed is required if the
measured x-ray luminosities are reduced by extinction internal to the
galaxies. Matching to the extinction-corrected 0.3--10~keV band value
of $2\times10^{47}\,{\rm erg\, M_\odot^{-1}}$ requires $v_\infty\simeq
700\,{\rm {km\, s^{-1}}} (\epsilon^2\nu_{100}^2\dot M_*/R_{100})^\alpha$ with
$\alpha\sim1/7-1/5$, close to the correlation required to sustain a
steady-state wind against radiative cooling and suggesting only a
narrow range of mass loading is allowed for a given star formation
rate. The predicted amount of hard x-ray energy generated in the
2--10~keV band declines steeply with decreasing star formation rate,
falling from $\sim2\times10^{45}\,{\rm erg\, M_\odot^{-1}}$ at $\dot
M_*=1\,{\rm M_\odot\,yr^{-1}}$ to $\sim2\times10^{43}\,{\rm erg\,
M_\odot^{-1}}$ at $\dot M_*=0.1\,{\rm M_\odot\,yr^{-1}}$.
The model gives a near constant amount of radio synchrotron energy at
1.4~GHz generated per unit mass of stars formed. The value is
sensitive to the assumed parameters of the energy distribution
function of relativistic electrons, but agreement with the data may be
achieved for typical values inferred from supernova remnants. The
spectrum, however, is expected to be steeper than would be given by
the electron energy index for a typical wind age of 10~Myr because of
the depletion of relativistic electrons within the wind-generating
core by synchrotron and plasmon wave energy losses.
Except for high ionization species like \hbox{O~$\scriptstyle\rm VI\ $}, commonly measured metal
absortion lines such as \hbox{C~$\scriptstyle\rm III\ $}, \hbox{Si~$\scriptstyle\rm III\ $}, \hbox{S~$\scriptstyle\rm III\ $}\ and \hbox{S~$\scriptstyle\rm IV\ $} are expected to
be undetectable within the hot core of the wind. (Since the wind core
resides within the galactic disc, these ions may form in detectable
amounts in the disc within the interface between the wind shock front
and the interstellar medium.) Detectable levels of these ions are
expected from the outflow region with typical column densities of
$10^{14}-10^{16}\,{\rm cm^{-2}}$, weakly increasing with the star
formation rate after allowing for the correlation between the
asymptotic wind velocity and the star formation rate required to match
the soft x-ray data. Since the gas is likely thermally and dynamically
unstable in the temperature regime at which the ions form in greatest
abundance, these values may set lower limits on the column densities
of an increasingly fragmenting outflow.
Estimates of the amount of mass-loading from absorption line data show
a maximum mass-loading factor comparable to the maximum imposed by the
cooling restriction. For the approximate analytic velocity correlation
$v_\infty\simeq 700\,{\rm {km\, s^{-1}}} (\epsilon^2\nu_{100}^2\dot
M_*/R_{100})^{1/7}$, the mass loading factor is somewhat smaller than
the cooling limit,
$\beta_*\simeq2.0(\epsilon\nu_{100})^{3/7}(R_{100}/\dot M_*)^{2/7}$,
comparable to measured values and corresponding to a mass injection
rate into the wind of $\dot M\simeq2.0{\rm M_\odot}\,{\rm
yr}^{-1}\,(\epsilon\nu_{100})^{3/7}{\dot M_*}^{5/7}R_{100}^{2/7}$.
\subsection{Superbubble with thermal heat conduction}
\label{subsec:sbmtc}
The superbubble model may be parametrized by the star formation rate
and the ambient hydrogen density into which the bubble expands within
a disc galaxy. The model ceases to be directly applicable once the
bubble breaks out of the disc:\ hydrodynamical simulations have shown
instead the bubble vents its thermal energy into conical winds
vertical to the disc. This will occur for expected hydrogen densities
within the discs before the bubble growth is restricted by the onset
of radiative cooling. The subsequent evolution of the wind is not
directly computed here.
Mass loading by thermal evaporation naturally results in an amount of
soft x-ray generation per solar mass of stars formed only weakly
dependent on the star formation rate. Matching to the measured value
in the 0.5--2~keV band requires ambient hydrogen densities of
$1-10\,{\rm cm^{-3}}$, while higher densities of
$10-100\,{\rm cm^{-3}}$ are required to match the measured value in
the 0.3--10~keV band, after allowing for extinction corrections for
internal galactic absorption. (Since extended emission out of the disc
will contribute a comparable amount to the measured x-ray luminosity,
the required densities may be somewhat lower.) For these densities,
cooling and dynamical criteria require the star formation rate per
superbubble cross-sectional area to exceed
$0.003-0.03\,{\rm M_\odot}\,{\rm yr^{-1}\,kpc^{-2}}$. A higher level of hard
x-ray emission in the 2--10~keV band is expected for a superbubble
compared with the wind core in the steady-state wind model for low
star formation rates, with a value in excess of
$\sim10^{45}\,{\rm erg\, M_\odot^{-1}}$, and possibly as much as an
order of magnitude larger or more, sustained by the superbubble for
star formation rates $\dot M_*\sim0.1\,{\rm M_\odot\,yr^{-1}}$. Such
high radiative efficiencies measured in the wind cores would favour
the superbubble model.
If the bubble blows out into the halo, it will do so with a thermal
pressure corresponding to an effective asymptotic wind velocity, in
terms of the steady-state wind parametrization, of $v_\infty^{\rm
eff}\simeq 850\,\,{\rm {km\, s^{-1}}} (\epsilon\nu_{100}{\dot M_*}/f_TR_{\rm B,
100})^{1/7}$. The effective wind velocity is suggestively close to
the value required for the steady state model to match the soft x-ray
data. This may indicate that large-scale winds are indeed driven by
superbubbles. The amount of mass loading through thermal evaporation,
however, is somewhat
low:\ $\beta_*\simeq0.64(\epsilon\nu_{100})^{5/7}(R_{\rm B,
100}f_T/{\dot M_*})^{2/7}$, corresponding to a mass injection rate
of $\dot M\simeq0.64{\rm M_\odot}\,{\rm yr}^{-1}(\epsilon\nu_{100}\dot
M_*)^{5/7}(R_{\rm B, 100}f_T)^{2/7}$, just somewhat smaller than the
amount above for the steady state model to match the amount of x-ray
energy generated per unit mass of stars formed. The value for mass
loading does correspond to, and so may set, the minimal values
measured from metal absorption line data. The larger measured values,
however, would require additional mass loading as the wind left the
disc, such as from hydrodynamical ablation or the thermal evaporation
of clouds, for which there is observational evidence.
As for the steady state model, the measured 1.4~GHz radio luminosities
of galaxies with winds may be recovered for plausible parameters of
the relativistic electron energy distribution. Unlike for the steady
state model, the relativistic electron distribution is not depleted
within the wind core so that steep radio spectra are not
required. This provides a possible means of distinguishing between the
models.
Detectable levels of \hbox{Si~$\scriptstyle\rm III\ $}, \hbox{C~$\scriptstyle\rm III\ $}, \hbox{S~$\scriptstyle\rm III\ $}\, \hbox{S~$\scriptstyle\rm IV\ $} and \hbox{O~$\scriptstyle\rm VI\ $} in
absorption are expected through the superbubble, with typical column
densities of $10^{11}-10^{14}\,{\rm cm^{-2}}$, nearly independent of
the star formation rate and only weakly dependent on the ambient gas
density and size of the superbubble. Although an ionization layer
produced by a wind shock impacting on the interstellar medium of a
galaxy in the steady state model may produce detectable levels of
metal absorption as well, differences in the column density ratios and
kinematics between the models would be expected. It is emphasized that
the thermal evaporation layer in the superbubble model sets a lower
limit to the metal absorption, which should have a covering fraction
of order unity. The absence of the minimal predicted levels of
absorption would be a telling indicator against the superbubble
model. Larger values may also arise outside the disc in gas at coronal
temperatures in a superbubble-driven outflow.
|
\section{Introduction and statements of results}
\label{approx1}
Consider any simply connected domain $O$ in the complex plane. Fix $p\in O$ and consider the operator
of complex integration defined on $H(O)$, the set of functions analytic in~$O$:
$$
J_p f\,(z)= \int_p^z f(\zeta)\, d \zeta, \quad f\in H(O), z\in O.
$$
This operator is related to a generalized Volterra operator acting
on $H(\mathbb D)$, where $\mathbb D$ is the unit disk. Let $g\in H(\mathbb D)$ and
define the operator $T_g:H(\mathbb D)\to H(\mathbb D)$ by
$$
T_gf(w)=\int_0^w f(t)g'(t)\, dt, \quad f\in H(\mathbb D),\,\,\zeta\in \mathbb D.
$$
In the case that $g$ is univalent, the change of variable $\zeta=g(t)$ transforms the operator
$T_g$ on $H(\mathbb D)$ to the operator $J_{g(0)}$ on $H(g(\mathbb D))$.
The operator $T_g$ has been studied on many Banach spaces $X\subset H(\mathbb D)$. For such $X$, define
$$
T[X]=\{g\in H(\mathbb D)\,:\, T_g \text{ is bounded on } X\}.
$$
C. Pommerenke's short proof of the analytic John-Nirenberg inequality
in \cite{Pom1}, based on his observation that $T[H^2]=$ BMOA, attracted considerable interest.
Subsequently, $T[X]$ has been identified for a variety of spaces $X$, including the Hardy spaces
($0<p<\infty$), Bergman spaces, and BMOA; see \cite{AC}, \cite{AS}, \cite{AS2} and \cite{SZ}.
The study of $T[H^{\infty}]$, where $H^\infty=H^\infty(\mathbb D)$ is the usual space of bounded
analytic functions on $\mathbb D$, was begun in \cite{AJS}.
We note also the paper \cite{NZ}, which gives sufficient conditions for the boundedness of the operators under investigation.
It is clear that
$T[H^{\infty}]\supseteqq$ BRV, the space of functions analytic on $\mathbb D$ with bounded radial variation
$$
\text{BRV}=\{g\in \text{H}(\mathbb D)\,:\, \sup_\theta\int_0^1|g'(re^{i\theta})|\,dr<\infty\},
$$
and in the article \cite{AJS} it was conjectured that $T[H^{\infty}]=$ BRV.
In the case that $g$ in univalent, this becomes a conjecture about when the operator
$J_{g(0)}$ is bounded on $H^\infty(g(\mathbb D))$. Our main result, Theorem \ref{oper} below, confirms
this conjecture when the symbol $g$ is univalent.
Recently, a discussion between
Fedor Nazarov, Paata Ivanisvili, Alexander Logunov, and one of the authors (D. Stolyarov), resulted in a
counterexample to the general conjecture in \cite{AJS} about when $T_g$ is bounded. Thus it is now known
that BRV $\subsetneqq T[H^{\infty}]$. We thank F. Nazarov, P. Ivanisvili, and A. Logunov for permission
to include the counterexample at the end of this paper.
Recently a paper \cite{CPPR} addressing the action
of the operator $T_g$ from a Banach space $X$ into $H^\infty$ has appeared in arXiv.
It has many interesting results, including
a characterization of when $T_g$ is bounded on $H^\infty$:
\cite[Theorem 1.2]{CPPR}
$g\in T[H^\infty]$ if and only if $\sup_{z\in\mathbb D}\|G_{g,z}\|_{\mathcal K}<\infty$, where
$$
\overline{G_{g,z}(w)}= \int_0^z g'(\zeta)\overline{K_\zeta(w)}\,d\zeta.
$$
Here $K_\zeta$ denotes the reproducing kernel for $H^2$ and $\mathcal K$ denotes the space of Cauchy transforms.
By the interior diameter of $O$ we understand the following quantity:
$$
\text{diam}_I \,O := \sup_{z_0,z_1\in O}\,\, \inf_{\gamma\in\Gamma(z_0,z_1)} \int_0^1 |\gamma'(t)| dt,
$$
where $\gamma\in\Gamma(z_0,z_1)$ means that $\gamma\colon [0,1]\to O$ is a smooth path with
$\gamma(0)=z_0$ and $\gamma(1)=z_1$.
\begin{theorem}
\label{oper}
Let $O$ be a simply connected domain in the plane, and let $p\in O$.
The operator $J_p$ is bounded on $H^\infty(O)$ if and only if ${\rm diam}_I \,O<\infty$.
\end{theorem}
Our proof of Theorem \ref{oper} is based on a new result on uniform approximation of Bloch functions,
the connection being that when $g$ is univalent, $\log g'$ is a Bloch function.
Recall that a function $f\in H(\mathbb D)$ is said to be a Bloch function if $(1-|z|)|f'(z)|$ is
bounded on $\mathbb D$.
For the statements, we introduce notation for the usual partial differentiation operators
$$
\partial=\frac{\partial}{\partial z} =
\frac12\left(\frac{\partial}{\partial x}-i\frac{\partial}{\partial y}\right), \quad
\bar{\partial}=\frac{\partial}{\partial \bar z} =
\frac12\left(\frac{\partial}{\partial x}+i\frac{\partial}{\partial y}\right).
$$
In particular, when $f$ is analytic $\partial f=f'$ and $\bar{\partial}f =0$.
Let $\Omega_{\alpha}^r$ denote the domain $\{z\colon |z|<r, \, \arg\, z\in (-\alpha/2, \alpha/2)\}$. We abbreviate $\Omega_\alpha =\Omega_\alpha^1$.
By $\mathcal B(\Omega_\alpha^r)$ we denote the class of functions analytic in $\Omega_\alpha^r$ and such that
\begin{equation}
\label{Bl}
|\partial F(z)| \le \frac{C_F}{|z|},\, z\in \Omega_\alpha^r.
\end{equation}
For a harmonic function $u$ on $\Omega_\beta$, denote by~$\tilde{u}$ the harmonic conjugate of~$u$
with $\tilde u (1/2) =0$.
\begin{theorem}
\label{realpart}
Let $0<\alpha<\beta<\pi$, $\varepsilon>0$, and let $F\in \mathcal B(\Omega_\alpha^{1/2})$. Then there exists a harmonic function $u$ in $\Omega_\beta$ such that
\textup{1)} $|u(x)- \mbox{\rm Re\,} F(x)|\le \varepsilon,\, x\in (0, \delta(\varepsilon)]$.
\textup{2)} $|\tilde u(z)| \le C(\varepsilon,\alpha,\beta,C_F)<\infty$ for all $z\in \Omega_{\beta}$.
\end{theorem}
In Section 2, we will assume Theorem \ref{realpart} and use it
to prove Theorem \ref{oper}. The proof of Theorem \ref{realpart} will be given in Section 3.
The example showing that BRV $\subsetneqq T[H^{\infty}]$
is in Section 4.
\medskip
{\it Notation for constants.} The letter $C$ will be used throughout the paper to denote various positive constants
which may vary at each occurrence but do not depend on the essential parameters.
The dependence of $C$ on important variables will be often indicated by placing the
variables in parentheses.
For $X$ and $Y$ nonnegative quantities, the notation $X\lesssim Y$ or $Y\gtrsim X$
means $X\le CY$ for some inessential constant $C$. Similarly, $X\approx Y$ means that
both $X\lesssim Y$ and $Y\lesssim X$ hold.
\noindent{\bf Acknowledgments.} We already mentioned that we are grateful to P. Ivanisvili, A. Logunov and F. Nazarov for the discussion (mentioned above) that led to the first example in Section \ref{examples}.
We are thankful to the referee for making valuable remarks that significantly improved the exposition, in particular, we are grateful to the referee for the remark at the end of Section 3.
\section{The proof of Theorem \ref{oper}, assuming Theorem
\ref{realpart}}
In this section we assume that Theorem \ref{realpart} holds and show that Theorem \ref{oper} is a consequence.
The proof of one implication in Theorem \ref{oper} does not require Theorem \ref{realpart}.
Suppose that
diam$_I\, O<\infty$, and let $z\in O$ and $f\in H^\infty(O)$ be arbitrary. Then, for
any smooth path $\gamma:[0,1]\to O$ connecting $p$ to $z$, we have
$$
\left|J_p f\,(z)\right|=\left|\int_\gamma f(\zeta)\,d\zeta\right|\le \|f\|_\infty\int_0^1|\gamma'(t)|\,dt.
$$
Thus, taking the infimum over all such paths $\gamma$
shows that $J_p: H^\infty(O)\to H^\infty(O)$ is a bounded operator with
$\|J_p\|\le \text{diam}_I \,O$.
For the other implication in Theorem \ref{oper},
assume that diam$_I\, O=\infty$. In the case that $O=\mathbb{C}$, by considering its action on
the constant function 1, it is clear that $J_p$ is unbounded on $H^\infty(O)$. On the other
hand, if $O$ is a
proper subset of $\mathbb C$, let
$\varphi:\mathbb D\to O$ be a Riemann
map, which we may assume is normalized so that $\varphi'(0)=1$.
Since the interior diameter is infinite,
given an integer $N$ there exists a radius, which we may assume is $[0,1)$, such that
\begin{equation}
\label{geN}
\int_0^1|\varphi'(x)| \,dx\ge N\,.
\end{equation}
Consider the function
\begin{equation}
\label{f}
f(z):= e^{-i(u(z)+i\tilde u(z))}\in \mathcal B(\Omega_\beta),
\end{equation}
where $u, \tilde u$ satisfy Theorem \ref{realpart} with $F$ chosen as follows.
Fix some $\alpha <\beta < \pi$.
We first denote by $\psi_\beta: \Omega_\beta\to \mathbb D$ the conformal map with $\psi_\beta(1/2) =0$ and $\psi_\beta(0) =1$. Then $|z\psi_\beta'(z)|\approx|1-\psi_\beta(z)|$, for $z\in \Omega_\beta$ and $|z|\le 1/2$, where the
constants suppressed depend only on $\beta$. Since also $|z\psi_\beta'(z)|\lesssim 1$, it follows that $|z\psi_\beta'(z)|\lesssim|1-\psi_\beta(z)|$, for $z\in \Omega_\beta$.
Hence restricting to $\Omega_\alpha^{1/2}$ gives
$$
|z\psi_\beta'(z)|\lesssim1-|\psi_\beta(z)|^2, \quad z\in \Omega_\alpha^{1/2},
$$
with constants depending only on $\alpha$ and $\beta$.
Now consider the composition
$$
F(z):= i \log \varphi' \circ \psi_\beta (z)\,.
$$
Using the well known inequality (see, for example, \cite[p. 9]{Pom}) that
$$
(1-|z|^2)\frac{|\varphi''(z)|}{|\varphi'(z)|}\le 6, \quad z\in \mathbb D,
$$
for the univalent function $\varphi$,
it follows that the restriction to $\Omega_\alpha^{1/2}$ of the function $F$ satisfies \eqref{Bl} (with $r=1/2$),
with constant $C_F=C(\alpha,\beta)$ depending only on $\alpha$ and $\beta$.
Thus Theorem~\ref{realpart} is applicable, and we get an approximate $\Phi= u+i\tilde u$ defined in $\Omega_\beta$.
By this theorem (with $\varepsilon =\pi/4$)
$$
|\mbox{\rm Re\,} F(x) - u(x)| \le \pi/4, \quad x\in (0,\delta(\pi/4)],
$$
and therefore we have that
$$
|\arg \varphi'(t) - u(\psi_\beta^{-1}(t))|\le \pi/4, \quad r_\beta<t<1,
$$
where $r_\beta\in(0,1)$ and depends only on $\beta$.
From the Koebe distortion theorem and our assumption that $\varphi'(0)=1$, there is a constant $C_1(\beta)$
such that
\begin{equation}
\label{lerbeta}
\int_0^{r_\beta}|\varphi'(t)|\,dt\le C_1(\beta).
\end{equation}
We also have from Theorem \ref{realpart} that
$$
|\tilde u|\le C_2(\alpha,\beta),\quad z\in \Omega_\beta.
$$
Therefore, if $f$ is the function from \eqref{f}, then function
$$
g:= f\circ \psi_\beta^{-1}\in H^\infty(\mathbb D),
$$
with $\|g\|_\infty< \exp(C_2(\alpha,\beta))$. We now estimate
\begin{align*}
\mbox{\rm Re\,} \int_{r_\beta}^1 g(x) \varphi'(x) dx &= \mbox{\rm Re\,} \int_{r_\beta}^1 e^{i (\arg \varphi'(x) -u(\psi_\beta^{-1}(x))} |\varphi'(x)| e^{\tilde u(\psi_\beta^{-1}(x))} dx\\
&\ge \cos (\pi/4) e^{-C_2(\alpha,\beta)} \int_{r_\beta}^1 |\varphi'(x)| dx \\
&\ge
\cos (\pi/4) e^{-C_2(\alpha,\beta)} (N-C_1(\beta)),
\end{align*}
from \eqref{geN} and \eqref{lerbeta}. Since also
$$
\left| \mbox{\rm Re\,} \int_0^{r_\beta} g(x) \varphi'(x) dx\right| \le C_1(\beta)\|g\|_\infty\le C_1(\beta)\exp(C_2(\alpha,\beta)),
$$
it follows that
$$
\mbox{\rm Re\,} \int_0^1 g(x) \varphi'(x) dx\ge \cos (\pi/4) e^{-C_2(\alpha,\beta)} (N-C_1(\beta))-C_1(\beta)\exp(C_2(\alpha,\beta)).
$$
Since the integer $N$ was arbitrary, and $\|g\|_\infty< \exp(C_2(\alpha,\beta))$,
this means that the operator $J_p$ is unbounded on $H^\infty(O)$.
Theorem \ref{oper} is proved.
\section{The proof of Theorem \ref{realpart}}
\label{prooffixed}
We separate out the main part of the proof of
Theorem \ref{realpart} into the following lemma.
\begin{lemma}
\label{fixed}
Let $0<\alpha<\beta<\pi$, and let $\varepsilon>0$. Given a function $F\in\mathcal B( \Omega_\alpha^{1/2})$, one can find analytic
$\Phi$ such that
\textup{1)} $|F(x)-\Phi(x)|\le \varepsilon, \quad x\in (0, \delta(\varepsilon)];$
\textup{2)} $\Phi\in\mathcal B( \Omega_\beta)$ and $C_\Phi =C(\varepsilon, \alpha,\beta, C_F)$.
\end{lemma}
\subsection{The proof of Theorem \ref{realpart}, assuming Lemma \ref{fixed}}
\label{prooffixed2}
Given $F=U+iV\in \mathcal B(\Omega_\alpha^{1/2})$, consider its symmetrization $F^*(z)= (F(z) + \overline{F(\bar{z})})/2
= U^*(z)+iV^*(z)$. It obviously belongs to $\mathcal B(\Omega_\alpha^{1/2})$ as well,
and we apply Lemma \ref{fixed}
to it to obtain a
function (let us call it) $\Phi^*$ which satisfies the derivative estimate in a larger domain $\Omega_\beta$.
Moreover, we may assume $\Phi^*$ is symmetric in the sense
$\Phi^*(z)= (\Phi(z) + \overline{\Phi(\bar{z})})/2$ for some~$\Phi$ satisfying the same bounds.
Let $\Phi^*(z):= u(z) + i\tilde u(z)$.
Then $V^*(x)=0$, $x\in \mathbb R\cap \Omega_\alpha^{1/2}$, $\tilde u(x)=0$, $x\in \mathbb R\cap \Omega_\beta$ , so we have
$$
|U(x) - u(x) |=|U^*(x)-u(x)|= |F^*(x) - \Phi^*(x) |\le \varepsilon,\quad x\in (0, \delta(\varepsilon)].
$$
We now use that
$$
|\nabla \tilde u(z)| \approx |\partial \Phi^*(z)| \le \frac{C_1}{|z|},\quad z\in \Omega_\beta,
$$
in conjunction with
$$
\tilde u (x)=0,\, x\in \mathbb R\cap \Omega_\beta,
$$
to conclude that
\begin{equation}
\label{bdd}
|\tilde u(z)|=|\tilde u(z)-\tilde u (x)|\le |y|\frac{C_2}{|z|} \le C_2, \quad z=x+iy\in \Omega_{\beta}.
\end{equation}
We have deduced Theorem \ref{realpart} from Lemma \ref{fixed}.
\medskip
Next, we present two lemmas that will be used in our proof of
Lemma \ref{fixed}. While these two lemmas are certainly well known to experts, we include the proofs
since we do not know good references.
\begin{lemma}
\label{JB}
Let $\varphi$ on $I_0:=[-1, 1]$ have Lipschitz norm $N$. Then $\varphi$ can be approximated by polynomials of degree $N^{3/2}$ with the error at
most $cN^{-1/2}$.
\end{lemma}
\begin{proof}
Given $\theta \in [-\pi, \pi]$ we introduce the new function $\Phi(e^{i\theta})= \varphi (\cos \theta)$.
Since $|\cos\theta_1 -\cos\theta_2|\le | e^{i\theta_1}- e^{i\theta_2}|$,
the modulus of continuity of
$\Phi$ is not greater than the modulus of continuity $\omega_\varphi$ of $\varphi$.
Therefore, using the Jackson--Bernstein theorem we can find a trigonometric polynomial $S(e^{i\theta})$
of degree $K$ such that
$$
|\Phi(e^{i\theta})- S(e^{i\theta})|\le A \omega_\varphi(1/K).
$$
Notice that $\Phi(e^{i\theta})$ is even by construction, so $S$ can be just a linear combination of $\cos k\theta$, $k=0,\dots, K$. Now we substitute $\theta= \arccos x$, and get the combination $P_K$ of Chebyshev polynomials $T_k(x)= \cos (k\arccos x)$, $k= 0, \dots, K$, such that
$$
|\varphi(x)- P_K(x) |\le A \omega_\varphi(1/K).
$$
Applying this inequality to a Lipschitz function $\varphi$ with Lipschitz constant at most $N$ and with $K= N^{3/2}$
completes the proof.
\end{proof}
\begin{lemma}
\label{pol}
Let a polynomial $P$ of degree $d$ satisfy $|P(x)|\le 1$, for $x\in I_0:=[-1, 1]$.
Then \textup{1)}~$P$ satisfies the uniform estimate~$|P(z)| \leq |16z|^d$ when~$|z| \geq 1$\textup{;} \textup{2)} $P'$ satisfies the same estimate with a slightly bigger constant in place of~$16$.
\end{lemma}
\begin{proof}
We use the Lagrange interpolation formula
\begin{equation}\label{Lagrange}
P(z) = \sum\limits_{j=0}^d P\Big(\frac{j}{d}\Big)\frac{\prod_{i\ne j}(z-\frac{i}{d})}{\prod_{i\ne j}(\frac{j-i}{d})}
\end{equation}
to establish the inequality (for~$|z| > 1$)
\begin{equation*}
|P(z)| \leq |2z|^d\Big(\sum\limits_{j=0}^d \frac{d^d}{j!(d-j)!}\Big) = |2x|^d \frac{(2d)^d}{d!} \leq |16 z|^d
\end{equation*}
since~$d^d < 4^d d!$. The estimate for the derivative can be proved in the same manner after one differentiates~\eqref{Lagrange}.
\end{proof}
\subsection{The proof of Lemma \ref{fixed}}
\label{prooffixed}
Let us change the variable: use $w$ for the variable in $\Omega_\beta$ and put $w=e^{-z}$, where
$$
z\in \Pi_\beta:=\{ z=x+iy: x>0, |y|<\beta/2\}.
$$
The same change of variable relates $\Pi_\alpha^{\log 2} $ to $\Omega_\alpha^{1/2}$:
$$
z\in \Pi_\alpha^{\log 2} :=\{ z=x+iy: x> \log 2, |y|<\alpha/2\}.
$$
Condition \eqref{Bl} for $F$ becomes the following condition for $f(z):= F(e^{-z})$ in $\Pi_\alpha^{\log 2}$:
\begin{equation}
\label{BlP}
|\partial f(z)| \le C_1,\, z\in \Pi_\alpha^{\log 2},
\end{equation}
which is precisely the Lipschitz assumption on a function $f$ analytic in $\Pi_\alpha^{\log 2}$.
We need to find {\it analytic} $h\in \Pi_\beta$ such that for some large number $\Delta(\varepsilon)$
\begin{equation}
\label{h}
|f(x)-h(x)| \le \varepsilon, \,\,x\ge \Delta(\varepsilon)\quad\text{and}\quad |\partial h(z)| \le C_2,\, z\in \Pi_\beta.
\end{equation}
We begin with a Lipschitz extension of $f$ from $\Pi_\alpha^{\log 2}$ into $\Pi_\alpha$.
For example we can extend $f$ by symmetry with respect to the vertical line $x=\log 2$.
Namely, $f$ extends by \eqref{BlP} to be continuous on the closure of ${\Pi_\alpha^{\log 2}}$ and then,
given $z=x+iy, 0<x<\log 2,$ we define $z^*= (2\log 2 -x)+iy$ and put
$$
f^*(z)=\begin{cases} f(z^*), \, z=x+iy, 0<x<\log 2, \, z\in \Pi_\alpha;\\
f(z), \, \,z=x+iy, x\ge \log 2, \,z\in \overline{\Pi_\alpha}\,.
\end{cases}
$$
It is easy to see that the new function $f^*$ is a Lipschitz function in the whole strip $\Pi_\alpha$, and it extends the analytic function $f$ defined on $\Pi_\alpha^{\log 2}$.
Then $f^*$ is not differentiable at the points
$z=\log 2+iy\in \Pi_\alpha$, but a standard smoothing of $f^*$ will have bounded gradient on $\Pi_\alpha$
and be an extension of the restriction of $f$ to $\Pi_\alpha^{1} :=\{ z\in \Pi_\alpha: \mbox{\rm Re\,} z> 1\}$.
Below the symbol $f$ denotes this extension.
Consider now
$$
H(x, y):= f\big(x+ i \frac{\alpha}{\beta} y\big), \, x+iy \in \Pi_\beta.
$$
It satisfies
$$
|\nabla H(z)|\le C_3,\, \, z \in \Pi_\beta \quad\text{and}\quad |f(x)- H(x)|=0,\,\, x\ge 0,
$$
but it is not analytic.
We claim that there is a function $g$ such that
\begin{equation}
\label{gH}
\bar{\partial} g =\bar{\partial} H;\text{ and } |g(x)| \le \varepsilon, x\ge \Delta(\varepsilon);
\text{ and } |\partial g(z)| \le C_4,\, z\in \Pi_\beta.
\end{equation}
The existence of such a function $g$ will complete the proof of Lemma \ref{fixed}. Indeed, setting
$$
h:= H-g,
$$
we have for $x\ge \Delta(\varepsilon)$ that
$$
|f(x)- h(x)|= |H(x)- (H(x)- g(x))|=|g(x)|\le \varepsilon.
$$
Also $h$ is analytic: $ \bar{\partial} h = \bar{\partial} H -\bar{\partial} g =0$. Moreover
$$
|\partial h |= |\partial H- \partial g| \le C_3+ C_4\,.
$$
This will establish that (\ref{h}) holds, and thus complete the proof of Lemma \ref{fixed}.
Hence it suffices to construct a function $g$ that satisfies (\ref{gH}).
\subsection{Analytic partition of unity}
There exists a number $b>0$ such that in $\Pi_\beta$ the function
$$
w(z)=\sum_{k=0}^\infty e^{-(z/b -k)^2}
$$
is uniformly bounded away from zero in absolute value. In fact, $\sum_{k=0}^\infty e^{-(z -k)^2}$ is
bounded away from zero on $\mathbb R_+$ and its derivative obviously is uniformly bounded in a fixed thin strip around $\mathbb R_+$.
Then in a smaller but fixed strip it is uniformly bounded away from zero in absolute value. Thus, with
$b$ chosen to be sufficiently large, $w$ will be uniformly bounded
away from zero in absolute value on $\Pi_\beta$.
We now introduce the notation
$$
e_k(z)= e^{-(z/b -k)^2}/ w(z)\, ,
$$
and note that
\begin{equation}
\label{modul}
\sum_{k=0}^\infty e_k(z) =1 \quad\text{and}\quad\sum_{k=0}^\infty |e_k(z)| \le C_5,\quad z\in\Pi_\beta\,.
\end{equation}
\subsection{The first modification of $H$}
As a step toward \eqref{gH}, let us first modify $H$ to $H_0$ in $\Pi_\beta$ in such a way that $\bar{\partial} H_0 =\bar{\partial} H$, but that also
\begin{equation}
\label{bddH}
|\nabla H_0(z)|\le C_6,\,\,|H_0(z) |\le C_7\,\quad z\in\Pi_\beta\,.
\end{equation}
Here is the formula for $H_0(z)$:
$$
H_0(z) = \sum_{k=0}^\infty e_k(z) (H(z)- H(k))\,.
$$
The Lipschitz property of $H$ (remember that $|\nabla H|\le C_3 $ in $\Pi_\beta$) and
\eqref{modul} prove that $H_0$ is bounded, and of course $\bar{\partial} H=\bar{\partial} H_0$.
We also have that
\begin{equation}
\label{modul2}
\sum_{k=0}^\infty |\partial e_k(z)| \le C_7,\quad z\in\Pi_\beta\,,
\end{equation}
which can be used to estimate $|\nabla H_0|$ in the same way \eqref{modul} was used to estimate $| H_0|$.
\subsection{The second step of the modification of $H$, from $H_0$ to $g$}
Re-writing (\ref{gH}) in terms of $H_0$, we need to find $g$ such that
\begin{equation}
\label{gH0}
\bar{\partial} g =\bar{\partial} H_0; \text{ and } |g(x)| \le \varepsilon, x\ge \Delta(\epsilon); \text{ and } |\partial g(z)| \le C_4,\, z\in \Pi_\beta.
\end{equation}
Let $m$ be a large integer to be fixed later. Consider functions
$$
h_{k, m} (t):= H_0(btm+ bk), \,\, t\in[-1,1],
$$
where $k$ and $m$ are integers and $b$ is the parameter from the partition of unity introduced in
section 3.3.
These are functions on the interval $I_0:= [-1,1]$ with Lipschitz norm bounded by $Cbm$, where $C=C_6$ from \eqref{bddH}.
We now apply Lemma \ref{JB} to the functions $h_{k, m}$ defined above to get polynomials $\mathcal P_k$ of degree
bounded by $ (Cbm)^{3/2}:=\lambda m^{3/2}$ such that
\begin{equation}
\label{a1}
|\mathcal P_k(t) - h_{k, m}(t) |\lesssim m^{-1/2}, \quad t\in I_0,
\end{equation}
which translates to
\begin{equation}
\label{a2}
\Big|\mathcal P_k\big(\frac{x-bk}{bm}\big) - H_0(x)\Big|\lesssim m^{-1/2},
\end{equation}
whenever $|x-bk|\le bm$.
We can now give the formula for the function $g$ that will satisfy (\ref{gH0}):
$$
g(z):= \sum_{j=0}^\infty e_j(z) \Big(H_0(z) - \mathcal P_j\big(\frac{z-bj}{bm}\big)\Big).
$$
From (\ref{modul}) it is clear that $\bar{\partial} g =\bar{\partial} H_0$, so it remains to estimate
1) $|g(x)|$ when $x\in \mathbb R$ is large, and 2) $|\partial g(z)|$ when $z\in \Pi_\beta$.
\medskip
Fix $x_0>0$ and let $k_0$ be the integer such that $|x_0- bk_0|\le b$.
We split the sum in the definition of $g(x_0)$ into three parts:
$$
\Sigma_1:= \sum_{j: |j-k_0| \le m-10} e_j(x_0) \Big(H_0(x_0) - \mathcal P_j\big(\frac{x_0-bj}{bm}\big)\Big);
$$
$$
\Sigma_2:= \sum_{j: j-k_0 \ge m-9} e_j(x_0) \Big(H_0(x_0) - \mathcal P_j\big(\frac{x_0-bj}{bm}\big)\Big);
$$
$$
\Sigma_3:= \sum_{j\ge 0: k_0-j \ge m-9} e_j(x_0) \Big(H_0(x_0) - \mathcal P_j\big(\frac{x_0-bj}{bm}\big)\Big).
$$
For the indices $j$ occurring in $\Sigma_1$, we have
$$
\frac{|x_0 - bj|}{bm} \le \frac{|x_0-bk_0 |}{bm}+ \frac{|bk_0 - bj|}{bm}
\le\frac1m+ \left(1- \frac{10}{m}\right) \le 1.
$$
Hence \eqref{a2} applies to each term in $\Sigma_1$. As the sum $\sum_{j\ge 0} e_j(z)$ converges absolutely in our strip, we get that
\begin{equation}
\label{s1}
\left| \Sigma_1 \right|\le Cm^{-1/2}.
\end{equation}
To estimate $\Sigma_2$ and $ \Sigma_3$ we need the following estimate of $\mathcal P_r$ and $\mathcal P'_r$, $r\ge 0$:
\begin{equation}
\label{Pr}
|\mathcal P_r(z)|+|\mathcal P'_r(z)|\le (C|z|)^{\lambda m^{3/2}}, \quad |z|\ge 1.
\end{equation}
Here the constant $C$ is independent of $r\ge 0$.
This follows from Lemma \ref{pol} and \eqref{a1}, since $H_0$ is bounded.
Next, notice that the part of $\Sigma_2$, $\Sigma_2':= \sum_{j: j-k_0 \ge m-9} e_j(x_0) H_0(x_0)$ is obviously small if $m$ is large. In fact, $|H_0|\le C_7$ from (\ref{bddH}), and so
\begin{align*}
|\Sigma_2'|&\le C_7\sum_{j: j-k_0 \ge m-9} |e_j(x_0)|
\lesssim C_7 \sum_{j: j-k_0 \ge m-9} e^{-\frac{|x_0- bj|^2}{b^2}}\\
&\le C\sum_{j: j-k_0 \ge m-9} e^{-\frac{|bk_0- bj|^2}{b^2}}= C\sum_{j: j-k_0 \ge m-9} e^{-|k_0- j|^2}\,
\end{align*}
which converges to zero as $m\to\infty$ by convergence of the series.
To estimate the other part of $\Sigma_2$, namely $\Sigma_2'':= \sum_{j: j-k_0 \ge m-9} e_j(x_0) \mathcal P_j(\frac{x_0-bj}{bm})$,
we notice that for the indices involved, we have (as $|x_0-bk_0|\le b$)
$$
\frac{|x_0-bj|}{bm}\le m^{-1}(1+ |j-k_0|)\le 1+ |j-k_0|.
$$
Therefore,
from \eqref{Pr} we get
\begin{equation}
\label{cPj}
\Big|\mathcal P_j\big(\frac{x_0-bj}{bm}\big)\Big| \le (C+C|j-k_0|)^{\lambda m^{3/2}}.
\end{equation}
For later use, we note that the derivative can be estimated in the same way, using \eqref{Pr}:
\begin{equation}
\label{diffcPj}
\Big|\mathcal P'_j\big(\frac{z-bj}{bm}\big)\Big| \le C(\beta)b m (C+C|j-k_0|)^{\lambda m^{3/2}},
\quad |z-x_0| \le \beta.
\end{equation}
Hence,
\begin{align*}
|\Sigma_2''|&\le \sum_{j: j-k_0 \ge m-9} |e_j(x_0)| |\mathcal P_j(\frac{x_0-bj}{bm})| \le C \sum_{j: j-k_0 \ge m-9} e^{-\frac{|x_0- bj|^2}{b^2}}|\mathcal P_j(\frac{x_0-bj}{bm})| \\
& \lesssim
\sum_{j: j-k_0 \ge m-9}\!\!\!\!\!\!(C+ C|j-k_0|)^{\lambda m^{3/2}} e^{-\frac{|bk_0- bj|^2}{b^2}}\\
&= \sum_{j: j-k_0 \ge m-9}\!\!\!\!\!\! (C+ C|j-k_0|)^{\lambda m^{3/2}} e^{-|k_0- j|^2}.
\end{align*}
This is small if $m$ is chosen to be large, and combined with the previous estimate for $|\Sigma_2'|$
we get that $|\Sigma_2|\to0$ as $m\to\infty$.
Notice that the same argument shows that $\Sigma_3$ is small when $m$ is large. Combining these estimates with \eqref{s1}, we obtain
\begin{equation}
\label{small_g}
|g(x)| =|\Sigma_1+\Sigma_2+\Sigma_3|\le\varepsilon, \quad\text{whenever}\quad x\ge bm\,,
\end{equation}
where $\varepsilon=\varepsilon(m)\to 0$ as $m\to \infty$. This is the required estimate for $|g(x)|$.
\subsection{The estimate of $\partial g$}
\label{deriv}
It remains to estimate $\partial g$.
The terms in $\partial g$ with $\frac{w'(z)}{w^2(z)}$ are estimated precisely as before,
since $\frac{w'(z)}{w^2(z)}$ is bounded on $\Pi_\beta$.
The terms with $(z/b-k) e^{-(z/b-k)^2}$ can be estimated along verbatim the same lines as before.
What is left, is to estimate
$$
\sum_{k=0}^\infty e_k(z) (\partial H_0(z) - \frac{1}{bm}\mathcal P'_k(\frac{z-bk}{bm})).
$$
From the estimate $|\nabla H_0|\le C_6$ and the absolute convergence of $\sum e_k$, we need only to prove
$$
\sum_{k=0}^\infty |e_k(z)| |\mathcal P'_k(\frac{z-bk}{bm})|\le C,\quad z\in \Pi_\beta.
$$
Let $z= x_0 +iy, |y|\le \beta/2$, and $k_0$ an integer with $|x_0-bk_0|\le b$.
Using \eqref{diffcPj}, we can estimate
$$
\sum_{k=0}^{\infty} |e_k(z)| |\mathcal P'_k(\frac{z-bk}{bm})|
\lesssim
C(\beta) bm\sum_{k=0}^{\infty} |e_k(bk_0)| (C+ C|k-k_0|)^{\lambda m^{3/2}}.
$$
Since $|e_k(bk_0)|\le C \exp(-|k-k_0|^2)$,
this sum is clearly bounded by some $C(m, \beta, b)<\infty$.
This completes the proof of the estimate $|\partial g(z)|\le C(m,\beta,b)$, $z\in \Pi_\beta$.
Hence (\ref{gH0}) has been established, the proof of the lemma is complete.
\begin{remark} In fact, one can modify the $\bar\partial $ proofs of this section to get a better claim. Namely, one can obtain the following statement by modifying the proofs above.
Let $0<\alpha' <\alpha<\beta<\pi/2$, $\varepsilon>0$, and $0<\delta<1/2$. Given a function
$F \in \mathcal B(\Omega_\alpha^{1/2})$ one can find analytic $\Phi$ such that
1) $|F(z)-\Phi(z)|\le\varepsilon$, $z\in \Omega_{\alpha'} \cap \{w\in \mathbb C : |w|\le\delta\}$; 2) $\Phi \in \mathcal B(\Omega_\beta)$ and $C_\Phi = C(\varepsilon,\delta,\alpha',\alpha,\beta,C_F)$.
\end{remark}
We are grateful to the referee for telling us that this better result was available.
\section{Examples: the proof that BRV $\subsetneqq T[H^\infty]$}
\label{examples}
We now present the example showing that BRV $\subsetneqq T[H^\infty]$. Again, we thank
F. Nazarov, P. Ivanisvili, and A. Logunov for permission to include it here.
\begin{prop}\label{MainEx}
There exists an analytic function~$g\colon \mathbb{D} \to \mathbb{C}$ such that the operator~$T_g$ is bounded on~$H^{\infty}(\mathbb{D})$ and
\begin{equation*}
\int\limits_{-1}^0 |g'(z)|\,dz = +\infty,
\end{equation*}
where integration is along the radius~$(-1,0]$.
\end{prop}
It will be more convenient to work with another domain. Let $\mathbb{D}_1$ be the disk with center $1$ and
radius 1, and denote by $\log$ the branch of the logarithm on
$\mathbb{D}_1$ that preserves the real numbers. Then
\begin{equation*}
z\in\mathbb{D}_1 \quad \Leftrightarrow \quad \zeta \in \Omega = \Big\{(x,y) \in\mathbb{R}^2\,:\; x < \log(2\cos y), y \in \Big(\frac{\pi}{2},\frac{\pi}{2}\Big)\Big\}
\end{equation*}
if~$\zeta = \log z$.
\begin{prop}\label{StOmega}
There exists an analytic function~$f\in H^{\infty}(\Omega)$ such that
\begin{equation}\label{CondOnf}
\int\limits_{-\infty}^{0}|f(\xi)|\,d\xi = \infty, \quad \hbox{but}\quad \bigg|\int\limits_{\zeta}^{0}f(\xi) h(\xi)\,d\xi\bigg| \lesssim \|h\|_{H^{\infty}(\Omega)}
\end{equation}
for all $\zeta $ real and negative, and for all $h \in H^{\infty}(\Omega)$.
\end{prop}
\begin{proof}[Proof of Proposition~\textup{\ref{MainEx}} assuming Proposition~\textup{\ref{StOmega}}:]
We claim that the function
$$
g(z) = \int_1^z \frac{f(\log w)}{w}\,dw, \quad w\in \mathbb{D}_1,
$$
is almost the one described in Proposition~\ref{MainEx}
(the only difference is that its domain is~$\mathbb{D}_1$). First, its variation along the radial segment~$(0,1]$ is infinite:
\begin{equation*}
\int\limits_{0}^1 |g'(r)|dr = \int\limits_{0}^1 \Big|\frac{f(\log r)}{r}\Big|dr = \int_{-\infty}^0|f(\xi)|\,d\xi = +\infty.
\end{equation*}
Second, for any function~$\tilde{h} \in H^{\infty}(\mathbb{D}_1)$,
\begin{equation*}
-T_g[\tilde{h}](z) = \int\limits_{z}^1 \tilde{h}(s)g'(s)\,ds = \int\limits_{\zeta}^0 \tilde{h}(e^{\xi})f(\xi)\,d\xi,\quad \zeta = \log z.
\end{equation*}
We did not specify the curve of integration in the line above because the integral is path independent.
Since $|\zeta|<\log2 + \pi/2$ when $\mbox{\rm Re\,} \zeta \geq 0$,
it is clear that the integral above is bounded by~$C\|f\|_{H^{\infty}}\|\tilde{h}\|_{H^{\infty}}$
in this case.
For the case that $\mbox{\rm Re\,} \zeta<0$, we pick a specific contour: starting at~$\zeta \in \Omega$, we
integrate first along the vertical segment~$v_{\zeta} = [\zeta,\mbox{\rm Re\,} \zeta]$, and then integrate along
the horizontal segment~$[\mbox{\rm Re\,} \zeta, 0]$. This leads to the bound (let~$\tilde{h}(e^{\xi})$ be simply~$h(\xi)$)
\begin{align*}
\big|T_g[\tilde{h}](z)\big| &\leq \bigg|\int\limits_{v_{\zeta}} h(\xi)f(\xi)\,d\xi\bigg| + \bigg|\int\limits_{\mbox{\rm Re\,} \zeta}^0 h(\xi)f(\xi)\,d\xi\bigg| \\
&\lesssim \|f\|_{H^{\infty}}\|h\|_{H^{\infty}} + \|h\|_{H^{\infty}} \lesssim \|\tilde{h}\|_{H^{\infty}}.
\end{align*}
To bound the first summand we used that the
length of $v_{\zeta}$ is at most ${\pi}/{2}$; the bound for the second summand came from
Proposition~\textup{\ref{StOmega}}.
Now we only have to shift~$\mathbb{D}_1$ to transfer~$g$ to~$\mathbb{D}$.
\end{proof}
\begin{proof}[Proof of Proposition~\textup{\ref{StOmega}}]
The function~$f$ will be given by the formula
\begin{equation}\label{FormulaForf}
f(\xi) = \sum\limits_{k=0}^{\infty} a_k e^{i\lambda_k(\xi - \zeta_k)}e^{-(\xi - \zeta_k)^2},
\end{equation}
where the sequence~$\{\lambda_k\}_k$ is real-valued and tends to~$+\infty$, and the sequence~$\{\zeta_k\}_k$ is real-valued and tends rapidly to~$-\infty$. We require that the following conditions are satisfied:
\begin{equation}\label{SumAk}
\sum\limits_{k=0}^{\infty}|a_k| = +\infty;
\end{equation}
\begin{equation}\label{FiniteSum}
\sum\limits_{k=0}^{\infty}\frac{|a_k|}{\lambda_k} < +\infty;
\end{equation}
\begin{equation}\label{Product}
|a_k|e^{2\lambda_k} \lesssim\ 1.
\end{equation}
For example, we may take~$a_k = e^{-2\lambda_k}$,~$\lambda_k = \frac12(\log k + \log\log k)$. With this choice of~$\lambda_k$ and~$a_k$, we may take~$\zeta_k = -2^k$. We will work with this particular choice of the sequences, however, the main properties we will use are~\eqref{SumAk}, \eqref{FiniteSum}, \eqref{Product}, and the fact that~$\{\zeta_k\}_k$ decreases sufficiently fast.
First, the function~$f$ is uniformly bounded on $\Omega$:
\begin{equation*}
|f(\xi)|\leq\Big|\sum\limits_{k=0}^{\infty} a_k e^{i\lambda_k(\xi - \zeta_k)}e^{-(\xi - \zeta_k)^2}\Big|
{\lesssim} \sum\limits_{k=0}^{\infty}|a_k|e^{\frac{\pi}{2}\lambda_k}\left|e^{-(\xi -\zeta_k)^2}\right|\lesssim 1,
\end{equation*}
where \eqref{Product} was used to get the last estimate.
Next, we prove the first part of~\eqref{CondOnf}.
By~\eqref{SumAk}, it suffices to show that
\begin{equation*}
|f(\xi)| \gtrsim |a_k|,\quad \xi \in [\zeta_k-1,\zeta_k + 1],
\end{equation*}
provided~$k$ is sufficiently large. This is easy: for $\xi \in [\zeta_k-1,\zeta_k + 1]$,
\begin{equation*}
\begin{aligned}
|f(\xi)| &\geq \frac{|a_k|}{e^2} - \sum\limits_{l < k}|a_l|e^{-(\zeta_k - \zeta_{l} + 1)^2} - \sum\limits_{l > k}|a_l|e^{-(\zeta_l - \zeta_k + 1)^2}\\& \geq
\frac{|a_k|}{e^2} - e^{-2^k}\sum\limits_{l < k}\frac{1}{l\log l} - \sum\limits_{l > k}\frac{1}{l\log l}e^{-(2^l - 2^k+1)^2} \gtrsim \frac{1}{k\log k} = |a_k|.
\end{aligned}
\end{equation*}
Last, we prove the inequality in~\eqref{CondOnf}.
Let $\zeta\in(-\infty,0)$.
Due to~\eqref{FiniteSum}, it suffices to prove that
\begin{equation*}
\bigg|\int\limits_{\zeta}^{0}e^{i\lambda_k(\xi - \zeta_k)}e^{-(\xi - \zeta_k)^2} h(\xi)\,d\xi\bigg| \lesssim \frac{\|h\|_{H^{\infty}(\Omega)}}{\lambda_k}.
\end{equation*}
Integration by parts gives
\begin{align*}
\int\limits_{\zeta}^{0}e^{i\lambda_k(\xi - \zeta_k)}e^{-(\xi - \zeta_k)^2} h(\xi)\,d\xi = &\frac{1}{i\lambda_k}e^{i\lambda_k(\xi - \zeta_k)}e^{-(\xi - \zeta_k)^2} h(\xi)\bigg|_{\xi = \zeta}^0 \\
&- \frac{1}{i\lambda_k}\int\limits_{\zeta}^0 e^{i\lambda_k(\xi - \zeta_k)}\Big(e^{-(\xi - \zeta_k)^2} h(\xi)\Big)'\,d\xi.
\end{align*}
The first term is clearly bounded by $|\lambda_k|^{-1}\|h\|_{H^{\infty}(\Omega)}$,
and the required bound for the integration term follows
from the estimate
\begin{equation*}
\bigg|\Big(e^{-(\xi - \zeta_k)^2} h(\xi)\Big)'\bigg| \lesssim (1 + |\xi - \zeta_k|)e^{-(\xi - \zeta_k)^2}\|h\|_{H^{\infty}(\Omega)}.
\end{equation*}
This estimate is a consequence of the inequality $|h'(\xi)|\lesssim \|h\|_{H^\infty(\Omega)}$,
$\xi\in (-\infty,0]$,
which follows
from the Cauchy integral formula for the derivative.
\end{proof}
\medskip
We end the paper with an example related to the proof of Theorem \ref{oper}. That proof
would have been easier if diam$_I\, \varphi(\mathbb D)=\infty$ implied that there is a radius $[0,e^{i\theta})$ of $\mathbb D$
such that $\varphi( [0,e^{i\theta}))$ is not rectifiable.
Here is an example that shows this may not be the case.
\begin{ex}\label{examp}
Form the domain $$O= U\setminus\bigcup_{k=2}^\infty \ell_k,$$ from
the domain $U= \{x+iy\,:\, 0<x<\infty,\,0<y<e^{-x}\}$ with the rays
$\ell_k=\{x+ie^{-k}/2\,:\, 2\le x<\infty\}$ removed, and let $\varphi:\mathbb D\to O$ be a
Riemann map with $\varphi(0)=z_0=1+ie^{-1}/2$.
Then the image under $\varphi$ of every radius of $\mathbb D$
is rectifiable, but diam$_I \,O=\infty$. Furthermore, $O$ may be modified to obtain a bounded
domain with the same properties.
\end{ex}
\begin{proof}
That diam$_I\, O=\infty$ is clear. Now consider any of the degenerate prime ends of $O$,
i.e. any prime end that corresponds to a single point $Q$ on the boundary of $O$. Clearly there is a rectifiable
curve in $\gamma_Q\subset O$ connecting $Q$ to the point $z_0$ and with length $\Lambda_1(\gamma_Q)< 2+|Q|$.
Let $e^{i\theta}$ be the point on the unit circle that corresponds to $Q$ under the map $\varphi$.
By a theorem of Gehring and Hayman (see \cite{GH}, or \cite[Theorem 4.20]{Pom}), there is an absolute constant
$K$ such that
$\Lambda_1(\varphi[0,e^{i\theta}))\le K \Lambda_1(\gamma_Q)< K(2+|Q|)<\infty$. There is only one prime
end left to consider, the one with impression $[2,\infty)$ and principal point $P=2$. Then, for the point $e^{i\eta}$
corresponding to this prime end under the map $\varphi$, we have
$$
\lim_{r\to 1}\varphi(re^{i\eta})=P,
$$
see \cite[Theorem 2.16]{Pom}, and once again the theorem of Gehring and Hayman tells us that
$\Lambda_1(\varphi[0,e^{i\eta}))<\infty$.
We now modify $O$ to obtain a bounded domain with the same properties. First form the domain
$$
\widetilde O = \{x+i(y+e^{-x})\,:\, x+iy\in O\},
$$
and let $\widetilde\varphi$ be a Riemann map with $\widetilde\varphi(0)=z_0=1+i3e^{-1}/2$. Next, notice that the restriction
of the function $e^{iz}$ to $\widetilde O$ is univalent. Indeed, if
$e^{iz_1}=e^{iz_2}$, then $z_2= z_1+2n\pi$ for some integer $n$. We need to show that
at most one of these points can
be in $\widetilde O$. So assume that $n>0$ and $z_1= x+iy\in \widetilde O$. Then
$
e^{-x}<y<2e^{-x},
$
and hence
$
y>2e^{-(x+2n\pi)}.
$
This means that $z_2\notin \widetilde O$, which completes that demonstration that $e^{iz}$ is univalent on
$\widetilde O$.
Applying to $\widetilde\varphi$ the
analysis that was just applied to $\varphi$ on $O$ now
shows that the image under $\widetilde\varphi$ of every radius of $\mathbb D$
is rectifiable. This is preserved under composition with the map $e^{iz}$, and hence
the Riemann map $e^{i\widetilde\varphi}$ from $\mathbb D$ to the bounded
domain $\exp(i\widetilde O)$ that spirals out to the unit circle is the example we are looking for.
\end{proof}
|
\section{Time Upper-bound for an Optimal Pebbling of Complete Binary Trees}
In this section, we improve time upper bounds for optimally
pebbling complete binary trees. It is known that the optimal pebbling
number of complete binary trees is $\log(h) + \theta(\log^{*}(h))$,
where $h$ is the height of the tree and $\log^{*}$ is the iterated
logarithmic function(\cite{Kra01}). We give an optimal pebbling of
complete binary trees that takes at most $n^{O(\log\log(n))}$ steps,
where $n$ is the number of nodes in the tree. Our pebbling is
essentially the same as in \cite{Kra01}. Our main contribution is to
show that the pebbling given in \cite{Kra01} is optimal. This proof ,
like the proof of Theorem~\ref{thm:main}, uses the equivalence between
the reversible pebble game and the Dymond-Tompa pebble game.
\begin{proposition}
The following statements hold.
\begin{enumerate}
\item $\rev{Bt_h} \geq \rev{Bt_{h-1}} + 1$
\item $\rev{Bt_{h}} \geq h + 2$ for $h \geq 3$
\item (\cite{Ben89}) $\rev{Ch_{n}} \leq \lceil\log_{2}(n)\rceil + 1$ for all $n$
\end{enumerate}
\label{prop:bt-ch-rbp}
\end{proposition}
\begin{proof}
(1) In any persistent pebbling of $Bt_{h}$, consider the earliest
time after pebbling the root at which one of the subtrees of the
root node has $\vrev{Bt_{h-1}}$ pebbles. At this time, there is a
pebble on the root and there is at least one pebble on the other
subtree of the root node. So, in total, there are at least
$\vrev{Bt_{h-1}} + 2 \geq \rev{Bt_{h-1}} + 1$ pebbles on the tree.
(2) Item (1) and the fact that $\rev{Bt_{3}} = 5$.\rule{7pt}{7pt}
\end{proof}
\begin{theorem}
There exists an optimal pebbling of $Bt_{h}$ that takes at most
$n^{O(\log\log(n))}$ steps.
\label{thm:bt-time-upperbound}
\end{theorem}
\begin{proof}
We will describe an optimal upstream pebbler in a pebbler-challenger
game who pebbles $root(Bt_{h})$, $left(root(Bt_{h}))$,
$left(right(root(Bt_{h})))$ and so on. In general, the pebbler
pebbles $left(right^{i-1}(root(Bt_{h})))$ in the $i^{\text{th}}$
step for $1 \leq i < h - \log(h)$. An upper bound on the number of
steps taken by the reversible pebbling obtained from this game
(which is, recursively pebble $left(right^{i-1}(root(Bt_{h})))$ for
$0 \leq i < h - \log(h)$ and optimally pebble the remaining tree
$Ch_{h - \log(h)} + Bt_{\log(h)}$ using any algorithm) is given
below. Here the term ${(2h - \log(h) + 1)}^{3\log(h)}$ is an upper
bound on the number of different pebbling configurations with
$3\log(h)$ pebbles, and therefore an upper bound for time taken for
optimally pebbling the tree $Ch_{h - \log(h)} + Bt_{\log(h)}$.
\begin{align*}
t(h) &\leq 2\left[t(h-1) + t(h-2) + \ldots + t(\log(h) + 1)\right] + {(2h - \log(h) + 1)}^{3\log(h)}\\
&\leq 2ht(h-1) + {(2h - \log(h) + 1)}^{3\log(h)}\\
&= O\left({(2h)}^{h}{(2h)}^{3\log(h)}\right)\\
&= (\log(n))^{O(\log(n))} = n^{O(\log\log(n))}
\end{align*}
In the first step, the pebbler will place a pebble on
$left(root(Bt_{h}))$ and the challenger will re-challenge the root
node. These moves are optimal. Before the $i^{\text{th}}$ step, the
tree has pebbles on the root and $left(right^{j}(root(Bt_{h})))$ for
$0 \leq j < i - 1$. We argue that if $i < h - \log(h)$, placing a
pebble on $left(right^{i-1}(root(Bt_{h})))$ is an optimal move. If
the pebbler makes this move, then the cost of the game is
$\max(\rev{Bt_{h_{1} - 1}}, \rev{Ch_{i} + Bt_{h_{1} - 1}}) =
\rev{Ch_{i} + Bt_{h_{1} - 1}} \leq \rev{Bt_{h_{1} - 1}} + 1 = p$,
where $h_{1} = h - i + 1$. Note that the inequality here is true
when $i < h - \log(h)$ by Prop~\ref{prop:bt-ch-rbp}. We
consider all other possible pebble placements on $i^{\text{th}}$
step and prove that all of them are inferior.
\begin{itemize}
\item \emph{A pebble is placed on the path from the root to
$right^{i-1}(root(Bt_{h}))$ (inclusive)}: The challenger will
challenge the node on which this pebble is placed. The cost of this
game is then at least $\rev{Bt_{h_{1}}} \geq p$.
\item \emph{A pebble is placed on a node with height less than $h_{1} -
1$}: The challenger will re-challenge the root node and the cost of
the game is at least $\rev{Ch_{i} + Bt_{h_{1} - 1}}$.
\end{itemize}
The theorem follows. For completeness, the following figure represents the optimal pebbler strategy used in the proof of Theorem~\ref{thm:bt-time-upperbound} for proving time upper bounds for complete binary tree.
\begin{figure}[ht]
\centering
\input{bt-time-fig}
\caption{An Optimal Pebbling for Complete Binary Trees\label{fig:bt}}
\end{figure}
\end{proof}
\section{Almost Optimal Pebblings of Complete Binary Trees}
In light of Theorem~\ref{thm:bt-time-upperbound}, the natural question
to ask is whether there are polynomial time optimal pebblings for
complete binary trees. In this section, we show that we can get
arbitrarily close to optimal pebblings for complete binary trees using
a polynomial number of steps (Theorem~\ref{thm:complete-trees-poly}).
\begin{theorem}
For any constant $\epsilon > 0$, we can pebble $Bt_{h}$ using at
most $(1 + \epsilon)h$ pebbles and $n^{O(\log(1/\epsilon))}$ steps
for sufficiently large $h$.
\label{thm:complete-trees-poly}
\end{theorem}
\begin{proof}
Let $k \geq 1$ be an integer. Then consider the following pebbling
strategy parameterized by $k$.
\begin{enumerate}
\item Recursively pebble the subtrees rooted at
$left(right^{i}(root(Bt_{h})))$ for $0 \leq i \leq k-1$ and
$right^{k}(root(Bt_{h}))$.
\item Leaving the $(k+1)$ pebbles on the tree (from the previous
step), pebble the root node using an additional $k$ pebbles in $2k-1$
steps.
\item Retaining the pebble on the root, reverse step (1) to remove
every other pebble from the tree.
\end{enumerate}
The number of pebbles and the number of steps used by the above
strategy on $Bt_{h}$ for sufficiently large $h$ is given by the
following recurrences.
\[ S(h) \leq S(h-k) + (k + 1) \leq \frac{(k+1)}{k}h \]
\[ T(h) \leq 2\left[\sum_{i = 1}^{k} T(h-i)\right] + (2k + 2)
\leq {(2k)}^{h}(2k+2) \leq n^{\log(k) + 1}(2k+2) \]
where $n$ is the number of nodes in $Bt_{h}$.
If we choose $k > 1/\epsilon$, then the theorem follows
\end{proof}
\section{Discussion \& Open Problems}
We studied reversible pebbling on trees. Although there are polynomial
time algorithms for computing black and black-white pebbling numbers
for trees, it was unclear, prior to our work, whether the reversible
pebbling number for trees could be computed in polynomial time. We also
established that almost optimal pebbling can be done in polynomial
time.
We conclude with the following open problems.
\begin{itemize}
\item Prove or disprove that there is an optimal pebbling for
complete binary trees that takes at most $O\left(n^k\right)$ steps for a fixed
$k$.
\item Prove or disprove that the there is a constant $k$ such that
optimal pebbling for any tree takes at most $O\left(n^k\right)$ (for black and
black-white pebble games, this statement is true with $k = 1$).
\item Give a polynomial time algorithm for computing optimal pebblings
of trees that take the smallest number of steps.
\end{itemize}
\section{Pebbling meets Coloring}
In this section, we prove our main theorem which states that the
reversible pebbling number of any tree is exactly one more than the
edge rank coloring number of its underlying undirected tree. It is
helpful to think about how to solve $\lang{TREE\hyphen PEBBLE}$ in polynomial time or even
${\sf NP}$. The first attempt would be to try and use the pebbling sequence
as a certificate that the input tree has low pebbling number. But,
this approach fails because trees are not guaranteed to have optimal
pebbling sequences of polynomial number of steps. We propose the
\emph{strategy tree} (Definition~\ref{def:strat-tree}) as a succinct
encoding of pebbling sequences. A strategy tree describes a pebbling
sequence. The key property is that for any tree, there is an optimal
pebbling sequence that can be described using a strategy tree
(Lemma~\ref{lem:rev=strat}).
\begin{definition}{(Strategy Tree)}
Let $T$ be a rooted directed tree. If $T$ only has a single node $v$,
then any strategy tree for $T$ only has a single node labeled
$v$. Otherwise, we define a strategy tree for $T$ as any tree
satisfying
\begin{enumerate}
\item The root node is labelled with some edge $e = (u, v)$ in $T$.
\item The left subtree of root is a strategy tree for $T_{u}$ and the
right subtree is a strategy tree for $T\setminus T_{u}$.
\end{enumerate}
\label{def:strat-tree}
\end{definition}
The following properties are satisfied by any strategy tree $S$ of $T
= (V, E)$.
\begin{enumerate}
\item Each node has 0 or 2 children.
\item \label{stratprop:bij} There are bijections from $E$ to internal nodes of $S$ and from
$V$ to leaves of $S$.
\item \label{stratprop:subtree} Let $v$ be any node in $S$. Then the
subtree $S_{v}$ corresponds to the subtree of $T$ spanned by the
nodes labeling the leaves of $S_{v}$. If $u$ and $v$ are two nodes
in $S$ such that one is not an ancestor of the other, then the
subtrees in $T$ corresponding to $u$ and $v$ are vertex-disjoint.
\end{enumerate}
\begin{lemma}
Let $T$ be a rooted directed tree. Then $\rev{T} \leq k$ if and only if there
exists a strategy tree for $T$ of depth at most $k$.
\label{lem:rev=strat}
\end{lemma}
\begin{proof}
We prove both directions by induction on $|T|$. If $T$ is a single
node tree, then the statement is trivial.
(if) Assume that the root of a strategy tree for $T$ of depth $k$ is
labelled by an edge $(u, v)$ in $T$. The pebbler then pebbles the
node $u$. If the challenger challenges $u$, the pebbler follows the
strategy for $T_{u}$ given by the left subtree of root. If the
challenger re-challenges, the pebbler follows the strategy for
$T\setminus T_{u}$ given by the right subtree of the root. The
remaining game takes at most $k-1$ pebbles by the inductive
hypothesis. Therefore, the total number of pebbles used is at most
$k$.
(only if) Consider an upstream pebbler that uses at most $k$
pebbles. We are going to construct a strategy tree of depth at most
$k$. Assume that the pebbler pebbles $u$ in the first move where
$e = (u, v)$ is an edge in $T$. Then the root node of $S$ is
labelled $e$. Now we have
$\rev{T_{u}}, \rev{T \setminus T_{u}} \leq k-1$. Let the left
(right) subtree be the strategy tree obtained inductively for
$T_{u}$ ($T \setminus T_{u}$). Since the pebbler is upstream, the
pebbler never places a pebble outside $T_{u}$ ($T \setminus T_{u}$)
once the challenger has challenged $u$ (the root).
\end{proof}
We now introduce a new game called the matching game played on
undirected trees (Definition~\ref{def:contr}). This game acts as a
link between the reversible pebble game and edge rank coloring.
\begin{definition}{(Matching Game)}
Let $U$ be an undirected tree. Let $T_{1} = U$. At each step of the
matching game, we pick a matching $M_{i}$ from $T_{i}$ and contract
all the edges in $M_{i}$ to obtain the tree $T_{i+1}$. The game ends
when $T_{i}$ is a single node tree. We define the \emph{contraction
number} of $U$, denoted \contr{U}, as the minimum number of
matchings in the matching sequence required to contract $U$ to the
single node tree.
\label{def:contr}
\end{definition}
\begin{lemma}
Let $T$ be a rooted directed tree and let $U$ be the underlying
undirected tree for $T$. Then $\rev{T} = k+1$ if and only if $\contr{U} = k$.
\label{lem:strat=contr}
\end{lemma}
\begin{proof}
First, we describe how to construct a matching sequence of length
$k$ from a strategy tree $S$ of depth $k + 1$. Let the leaves of $S$
be the level 0 nodes. For $i \geq 1$, we define the level $i$ nodes
to be the set of all nodes $v$ in $S$ such that one child of $v$ has
level $i-1$ and the other child of $v$ has level at most
$i-1$. Define $M_{i}$ to be the set of all edges in $U$
corresponding to level $i$ nodes in $S$. We claim that $M_{1},
\ldots ,M_{k}$ is a matching sequence for $U$. Define $S_{i}$ as the
set of all nodes $v$ in $S$ such that the parent of $v$ has level at
least $i+1$. Let $Q(i)$ be the statement ``$T_{i+1}$ is obtained from
$T_{1}$ by contracting all subtrees corresponding to nodes (See
Property~\ref{stratprop:subtree}) in $S_{i}$''. Let $P(i)$ be the
statement ``$M_{i+1}$ is a matching in $T_{i+1}$''. We will prove
$Q(0)$ and $Q(i) \implies P(i)$ and $(Q(i) \wedge P(i)) \implies
Q(i+1)$. Indeed for $i = 0$, we have $Q(0)$ because $T_{1} = U$ and
$S_{0}$ is the set of all leaves in $S$ or nodes in $T$
(Property~\ref{stratprop:bij}). To prove $Q(i) \implies P(i)$,
observe that the edges of $M_{i+1}$ correspond to nodes in $S$ where
both children are in $S_{i}$. So these edges correspond to edges in
$T_{i+1}$ (by $Q(i)$) and the fact that these edges are pairwise disjoint since no
two nodes in $S$ have a common child).
To prove that $(Q(i) \wedge P(i)) \implies Q(i+1)$, consider the
tree $T_{i+2}$ obtained by contracting $M_{i+1}$ from
$T_{i+1}$. Since $Q(i)$ is true, this is equivalent to contracting
all subtrees corresponding to $S_{i}$ and then contracting the edges
in $M_{i+1}$ from $T_{1}$. The set $S_{i+1}$ can be obtained from
$S_{i}$ by adding all nodes in $S$ corresponding to edges in
$M_{i+1}$ and then removing both children (of these newly added
nodes) from $S_i$. This is equivalent to combining the subtrees
removed from $S_{i}$ using the edge joining them. This is because
$M_{i+1}$ is a matching by $P(i)$ and hence one subtree in $S_{i}$
will never be combined with two other subtrees in $S_{i}$. But then
contracting subtrees in $S_{i+1}$ from $T_{1}$ is equivalent to
contracting $S_{i}$ followed by contracting $M_{i+1}$.
We now show that a matching sequence of length at most $k$ can be
converted to a strategy tree of depth at most $k+1$. We use proof by
induction. If the tree $T$ is a single node tree, then the statement
is trivial. Otherwise, let $e$ be the edge in the last matching
$M_{k}$ in the sequence and let $(u, v)$ be the corresponding edge
in $T$. Label the root of $S$ by $e$ and let the left (right)
subtree of root of $S$ be obtained from the matching sequence
$M_{1}, \ldots ,M_{k-1}$ restricted to $T_{u}$ ($T\setminus T_{u}$).
By the inductive hypothesis, these subtrees have height at most
$k-1$.
\end{proof}
\begin{lemma}
For any undirected tree $U$, we have \contr{U} = \erank{U}.
\label{lem:contr=erank}
\end{lemma}
\begin{proof}
Consider an optimal matching sequence for $U$. If the edge $e$ is
contracted in $M_{i}$, then label $e$ with the color $i$. This is an
edge rank coloring. Suppose for contradiction that there exists two
edges $e_{1}$ and $e_{2}$ with label $i$ such that there is no edge
labelled some $j \geq i$ between them. We can assume without loss of generality that there
is no edge labelled $i$ between $e_{1}$ and $e_{2}$ since if there
is one such edge, we can let $e_{2}$ to be that edge. Then $e_{1}$
and $e_{2}$ are adjacent in $T_{i}$ and hence cannot belong to the
same matching.
Consider an optimal edge rank coloring for $U$. Then in the
$i^{\text{th}}$ step all edges labelled $i$ are contracted. This
forms a matching since in between any two edges labelled $i$, there
is an edge labelled $j > i$ and hence they are not adjacent in
$T_{i}$.
\end{proof}
The theorems in this section are summarized in Fig.~\ref{fig:equiv}
\begin{figure}
\centering
\begin{subfigure}{0.45\textwidth}
\centering
\input{bt3}
\caption{The complete binary tree of height 3}
\label{fig:bt3}
\end{subfigure}\hfill
\begin{subfigure}{0.45\textwidth}
\centering
\input{erank3}
\caption{Optimal edge rank coloring}
\label{fig:erank3}
\end{subfigure}
\begin{subfigure}{0.9\textwidth}
\centering
\input{strattree3}
\caption{Optimal strategy tree}
\label{fig:st3}
\end{subfigure}
\begin{subfigure}{0.9\textwidth}
\centering
\input{matching3}
\caption{Optimal matching sequence}
\label{fig:matching3}
\end{subfigure}
\caption{This figure illustrates the equivalence between persistent
reversible pebbling, matching game and edge rank coloring on trees
by showing an optimal strategy tree and the corresponding matching
sequence and edge rank coloring for height 3 complete binary
tree.\label{fig:equiv}}
\end{figure}
\begin{theorem}
Let $T$ be a rooted directed tree and let $U$ be the underlying
undirected tree for $T$. Then we have $\rev{T} = \erank{U} + 1$.
\label{thm:main}
\end{theorem}
\begin{corollary}
\vrev{T} and \rev{T} along with strategy trees achieving the optimal
pebbling value can be computed in polynomial time for trees.
\label{cor:vis=pers}
\end{corollary}
\begin{proof}
We show that \lang{TREE\hyphen PEBBLE}\ and \lang{TREE\hyphen VISITING\hyphen PEBBLE}\ are polynomial time equivalent. Let
$T$ be an instance of \lang{TREE\hyphen PEBBLE}. Pick an arbitrary leaf $v$ of $T$ and
root the tree at $v$. By Theorem~\ref{thm:main}, the reversible
pebbling number of this tree is the same as that of $T$. Let $T'$ be
the subtree rooted at the child of $v$. Then we have
$\rev{T} \leq k \iff \vrev{T'} \leq k-1$.
Let $T$ be an instance of \lang{TREE\hyphen VISITING\hyphen PEBBLE}. Let $T'$ be the tree obtained by
adding the edge $(r, r')$ to $T$ where $r$ is the root of $T$. Then
we have $\vrev{T} \leq k \iff \rev{T'} \leq k + 1$.
The statement of the theorem follows from Theorem~\ref{thm:main} and
the linear-time algorithm for finding an optimal edge rank coloring
of trees\cite{Lam98optimaledge}.
\end{proof}
The following corollary is immediate from the equivalence of pebble
games (Theorem \ref{lem:chan:dt=rp}).
\begin{corollary}
For any rooted directed tree $T$, we can compute \DT{T} and \RM{T}
in polynomial time.
\end{corollary}
An interesting consequence of Theorem~\ref{thm:main} is that the
persistent reversible pebbling number of a tree depends only on its
underlying undirected graph. A natural question would be to ask
whether this fact generalizes to DAGs. The following proposition shows
that this is not the case.
\begin{proposition}
There exists two DAGs with the same underlying undirected graph and
different pebbling numbers.
\end{proposition}
\begin{proof}
Consider the following two DAGs
\begin{figure}[ht]
\begin{subfigure}{0.45\textwidth}
\centering
\input{dag1}
\caption{$\rev{G_1} = 5$}
\end{subfigure}\hfill
\begin{subfigure}{0.45\textwidth}
\centering
\input{dag2}
\caption{$\rev{G_2} = 6$}
\end{subfigure}
\end{figure}
DAGs $G_1$ and $G_2$ have the same underlying undirected graph and different persistent pebbling numbers.
\end{proof}
\section{Introduction}
Pebbling games of various forms on graphs abstracts out resources in
different combinatorial models of computation (See
\cite{Chan13-thesis}). A rooted DAG can be used to model computation
as follows -- Each node in the DAG represents a value obtained during
computation, the source nodes represent input values, the internal
nodes represent intermediate values, and the root node represents the
output value. A pebble placed on a vertex in a graph corresponds to
storing the value at that node, and an edge $(a,b)$ in the graph would
represent a data-dependency - namely, the value at $b$ can be computed
only if the value at $a$ is known (or stored). Devising the rules of
the pebble game to capture the rules of the computation, and
establishing bounds for the total number of pebbles used at any point
in time, gives rise to a combinatorial approach to proving bounds on
the \emph{space} used by the computation. The Dymond-Tompa and
Raz-Mckenzie pebble games depict some of the combinatorial barriers in
improving upper bounds for depth (or parallel time) of Boolean
circuits (or parallel algorithms).
Motivated by applications in the context of reversible computation
(for example, quantum computation), Bennett\cite{Ben89} introduced the
reversible pebble game. Given any DAG $G$ with a unique sink node $r$,
the reversible pebble game starts with no pebbles on $G$ and ends with
a pebble (only) on $r$. Pebbles can be placed or removed from any node
according to the following two rules.
\begin{enumerate}
\item To pebble $v$, all in-neighbors of $v$ must be pebbled.
\item To unpebble $v$, all in-neighbors of $v$ must be pebbled.
\end{enumerate}
The goal of the game is to pebble the sink node $r$ using the minimum
number of pebbles (also using the minimum number of steps).
Recently, Chan\cite{Chan13} showed that for any DAG $G$ the number of
pebbles required for the reversible pebble game is exactly the same as
the number of pebbles required for the Dymond-Tompa pebble game and
the Raz-Mckenzie pebble game. However, connections between the
reversible pebble game and graph parameters not arising from
computational considerations were not known. For irreversible pebble
games, we know that the black white pebbling number of trees is
closely related to min-cut linear arrangements of trees\cite{Yan85}.
On the computational complexity front, Chan\cite{Chan13} also studied
the complexity of the following problem -- Given a DAG $G = (V, E)$
with a unique sink $r$ and an integer $1 \leq k \leq |V|$, check if
$G$ can be pebbled using at most $k$ pebbles. He showed that this
problem is \PSPACE-complete. Determining the irreversible black and
black-white pebbling number are known to be \PSPACE-complete on DAGs
(See \cite{Gilbert}, \cite{Hertel}). If we restrict the irreversible
black pebble game to be read-once (each node is pebbled only once),
then the problem becomes {\sf NP}-complete (See \cite{Sethi75}). However,
if we restrict our attention to trees, the irreversible black pebble
game\cite{loui} and black-white pebble game\cite{Yan85} are solvable
in polynomial time. The key insight is that the optimal
\emph{irreversible} (black or black-white) pebbling number of trees
can be achieved by read-once pebblings. Deciding whether the pebbling
number is at most $k$ for a given tree is in {\sf NP}\ since the optimal
pebbling serves as the certificate. We cannot show that determining
the reversible pebbling number is in {\sf NP}\ using the same argument as
we do not know whether the optimal value can always be achieved using
pebblings taking only polynomially many steps.
\paragraph{Our Results:} In this paper, we study the reversible pebble
game on trees. For an undirected tree $T$, the edge rank coloring
number of the tree is the minimum number of colors required to color
the edges of $T$ using integers such that for any two edges in $T$
having the same color $i$, there is at least one edge on the path
between those edges that has a higher color. We show that the
reversible pebbling number of any tree is exactly one more than the
edge rank coloring number of the underlying undirected tree. Besides,
the reversible pebbling number, another interesting parameter related
to reversible pebble game is the number of steps required to optimally
pebble the given DAG. For example, it is known that paths can be
optimally pebbled in $O(n\log n)$ steps. We show that the connection
with Dymond-Tompa pebble game can be exploited to show that complete
binary trees have optimal pebblings that take at most
$n^{{O(\log\log(n))}}$ steps. This is a significant improvement over
the previous upper bound of $n^{{O(\log(n))}}$ steps. It remains open
whether complete binary trees can be pebbled in polynomial number of
steps. Towards this end, we show that ``almost'' (within a factor of
$(1+\epsilon)$ for any constant $\epsilon > 0$) optimal pebblings of
complete binary trees can be done in polynomial number of steps. We
also generalize a time-space trade-off result given for paths by
Kr\'alovic to families of bounded degree trees showing that for any
constant $\epsilon > 0$, such families can be pebbled using
$O(n^{\epsilon})$ pebbles in $O(n)$ steps.
\paragraph{Complexity of Reversible Pebbling Number on
Trees:} We show that the reversible pebbling number of trees along
with strategies achieving the optimal value can be computed in
polynomial time. This is obtained by combining our main result with
the linear-time algorithm given by Lam and Yue \cite{Lam98optimaledge}
for finding an optimal edge rank coloring of the underlying undirected
tree. Our proof of the main result also shows how to convert an
optimal edge rank coloring into an optimal reversible pebbling.
\section{Preliminaries}
We assume familiarity with basic definitions in graph theory, such as
those found in \cite{West}. A directed tree $T = (V, E)$ is called a
\emph{rooted directed tree} if there is an $r \in V$ such that $r$ is
reachable from every node in $T$. The node $r$ is called the root of
the tree.
An \emph{edge rank coloring} of an undirected tree $T$ with $k$
colors $\{ 1, \ldots ,k\}$ labels each edge of $T$ with a color such
that if two edges have the same color $i$, then the path between
these two edges consists of an edge with some color $j > i$. The
minimum number of colors required for an edge rank coloring of $T$
is denoted by \erank{T}.
\begin{definition}{(Reversible Pebbling\cite{Ben89})}
Let $G$ be a rooted DAG with root $r$. A \emph{reversible pebbling
configuration} of $G$ is a set $P \subseteq V$ (the set of pebbled
vertices). A \emph{reversible pebbling} of $G$ is a sequence of
reversible pebbling configurations $P = (P_{1}, \ldots ,P_{m})$ such
that $P_{1} = \phi$ and $P_{m} = \{ r \}$ and for every
$i, 2 \leq i \leq m$, we have
\begin{enumerate}
\item $P_{i} = P_{i-1} \cup \{ v \}$ or
$P_{i-1} = P_{i} \cup \{ v \}$ and $P_{i} \neq P_{i-1}$ (Exactly
one vertex is pebbled/unpebbled at each step).
\item All in-neighbors of $v$ are in $P_{i-1}$.
\end{enumerate}
The number $m$ is called the time taken by the pebbling $P$. The
number of pebbles or space used in a reversible pebbling of $G$ is
the maximum number of pebbles on $G$ at any time during the
pebbling. The \emph{persistent reversible pebbling number} of $G$,
denoted by \rev{G}, is the minimum number of pebbles required to
persistently pebble $G$.
A closely related notion is that of \emph{visiting} reversible pebbling, where the pebbling $P$ satisfies (1) $P_{1} = P_{m} = \phi$ and (2) there exists a $j$ such that $r \in P_{j}$. The minimum number
of pebbles required for a visiting pebbling of $G$ is denoted by
\vrev{T}.
\end{definition}
It is easy to see that $\vrev{G} \leq \rev{G} \leq \vrev{G} + 1$ for
any DAG $G$.
\begin{definition}{(Dymond-Tompa Pebble Game \cite{DT85})}
Let $G$ be a DAG with root $r$. A Dymond-Tompa pebble game is a
two-player game on $G$ where the two players, the pebbler and the
challenger takes turns. In the first round, the pebbler pebbles the
root node and the challenger challenges the root node. In each
subsequent round, the pebbler pebbles a (unpebbled) node in $G$ and
the challenger either challenges the node just pebbled or
re-challenges the node challenged in the previous round. The pebbler
wins when the challenger challenges a node $v$ and all in-neighbors
of $v$ are pebbled.
The Dymond-Tompa pebble number of $G$, denoted \DT{G}, is the
minimum number of pebbles required by the pebbler to win against
an optimal challenger play.
\end{definition}
The Raz-Mckenzie pebble game is also a two-player pebble game played
on DAGs. The optimal value is denoted by \RM{G}. A definition for the
Raz-Mckenzie pebble game can be found in \cite{RM99}. Although the
Dymond-Tompa game and the reversible pebble game look quite
different. The following theorem reveals a surprising connection
between them.
\begin{theorem}{(Theorems 6 and 7, \cite{Chan13})}
For any rooted DAG $G$, we have $\DT{G} = \rev{G} = \RM{G}$.
\label{lem:chan:dt=rp}
\end{theorem}
\begin{definition}{(Effective Predecessor \cite{Chan13})}
Given a pebbling configuration $P$ of a DAG $G$ with root $r$, a
node $v$ in $G$ is called an \emph{effective predecessor} of $r$ iff
there exists a path from $v$ to $r$ with no pebbles on the vertices
in the path (except at $r$).
\end{definition}
\begin{lemma}{(Claim 3.11, \cite{Chan13})}
Let $G$ be any rooted DAG. There exists an optimal pebbler strategy
for the Dymond-Tompa pebble game on $G$ such that the pebbler always
pebbles an effective predecessor of the currently challenged node.
\label{lem:chan:ep}
\end{lemma}
The height or depth of a tree is defined as the maximum number of
nodes in any root to leaf path. We denote by $Ch_{n}$ the rooted
directed path on $n$ nodes with a leaf as the root. We denote by
$Bt_{h}$ the complete binary tree of height $h$. We use
$root(Bt_{h})$ to refer to the root of $Bt_{h}$. If $v$ is any node in
$Bt_{h}$, we use $left(v)$ ($right(v)$) to refer to the left (right)
child of $v$. We use $right^{i}$ and $left^i$ to refer to iterated
application of these functions. We use the notation $Ch_{i} + Bt_{h}$
to refer to a tree that is a chain of $i$ nodes where the source node
is the root of a $Bt_{h}$.
\begin{definition}
We define the language \lang{TREE\hyphen PEBBLE}\ as the set of all tuples $(T, k)$,
where $T$ is a rooted directed tree and k is a integer satisfying
$1 \leq k \leq n$, such that $\rev{T} \leq k$. The language \lang{TREE\hyphen VISITING\hyphen PEBBLE}\
is the same as \lang{TREE\hyphen PEBBLE}\ except that the goal is to check whether
$\vrev{T} \leq k$.
\end{definition}
In the rest of the paper, we use the term pebbling to refer to
\textit{persistent reversible pebbling} unless explicitly stated
otherwise.
\section*{Appendix}
\end{document}
\section{{\sc Tree-Pebble} $\in$ co-{\sf NP}}
\subsection{Equivalence between matching and reversible pebbling}
\section{{\sc Tree-Pebble} $\in {\sf NP}$}
\subsection{Strategy trees}
\subsection{{\sf NP} Algorithm}
\subsection{Constant-factor approximation algorithm}
\subsection{Poly-time equivalence between Visiting and Persistent pebbling}
\section{Time-space Trade-offs for Bounded-degree Trees}
In \cite{Kra01}, it is shown that there are linear time pebbling
sequences for paths that use only $n^{\epsilon}$ pebbles for any
constant $\epsilon > 0$. In this section, we generalize this result to
bounded degree trees (Theorem~\ref{thm:tree-lin-time}).
\begin{theorem}
For any constant positive integer $k$, a bounded-degree tree $T$
consisting of $n$ vertices can be pebbled using at most
$O\left(n^{1/k}\right)$ pebbles and $O(n)$ pebbling moves.
\label{thm:tree-lin-time}
\end{theorem}
\begin{proof}
Let us prove this by induction on the value of $k$. In the base case
($k=1$), we are allowed to use $O(n)$ pebbles. So, the best strategy
would to place a pebble on every vertex of $T$ in bottom-up fashion,
starting from the leaf nodes. After the root is pebbled, we unpebble
each node in exactly the reverse order, while leaving the root
pebbled.
In this strategy, clearly, each node is pebbled and unpebbled at
most once. Hence the number of pebbling moves must be bounded by
$2n$. Hence, a tree can be pebbled using $O(n)$ pebbles in $O(n)$ moves.
Now consider that for $k \leq k_0 - 1$, where $k_0$ is an integer $\geq 2$, any bounded-degree
tree $T$ with $n$ vertices can be pebbled using $O\left(n^{1/k}\right)$ pebbles in $O(n)$ moves.
Assume that we are allowed $O\left(n^{1/k_0}\right)$ pebbles. To apply induction, we will be decomposing the tree into smaller components. We prove the following claim first.
\begin{claim}
Let $T'$ be any bounded-degree tree with $n' > n^{(k_0-1)/k_0}$ vertices and maximum degree $\Delta$. There
exists a subtree $T''$ of $T'$ such that the number of vertices in
$T''$ is at least $\lfloor n^{(k_0-1)/k_0}/2 \rfloor$ and at most
$\lceil n^{(k_0-1)/k_0} \rceil$.
\label{lem:partition}
\end{claim}
\begin{proof}
From the classical tree-separator theorem, we know that $T'$ can be divided into two subtrees, where the larger subtree has between $\lfloor n'/2 \rfloor$ and $\lceil n' \cdot \dfrac{\Delta}{\Delta + 1} \rceil$ vertices. The key is to recursively subdivide the tree in this way and continually choose the larger subtree. However, we need to show that in doing this we will definitely strike upon a subtree with the number of vertices within the required range. Let $T_1',T_2', \ldots$ be the sequence of subtrees we obtain in these iterations. Also let $v_i$ be the number of vertices in $T_i'$ for every $i$. Note that $\forall i, \lfloor v_i/2 \rfloor \leq v_{i+1} \leq \lceil v_i \cdot \dfrac{\Delta}{\Delta+1} \rceil$. Assume that $j$ is the last iteration where $v_j > \lceil n^{(k_0-1)/k_0} \rceil$. Clearly $v_{j+1} \geq \lfloor n^{(k_0-1)/k_0}/2 \rfloor$. Also, by the definition of $j$, $v_{j+1} \leq \lceil n^{(k_0-1)/k_0} \rceil$. Hence the proof
\end{proof}
The final strategy will be as follows:
\begin{enumerate}
\item Separate the tree into $\theta(n^{1/k_0})$ connected subtrees, each containing $\theta(n^{(k_0-1)/k_0})$ vertices. Claim ~\ref{lem:partition} indicates that this can always be done.
\item Let us number these subtrees in the following inductive fashion: denote by $T_1$, the `lowermost' subtree, i.e. every path to the root of $T_1$ must originate from a leaf of $T$. Denote by $T_i$, the subtree for which every path to the root originates from either a leaf of $T$ or the root of some $T_j$ for $j<i$. Also, let $n_i$ denote the number of vertices in $T_i$.
\item Pebble $T_1$ using $O\left(n_1^{1/(k_0-1)}\right) = O\left(n^{1/k_0}\right)$ pebbles. From the induction hypothesis, we know that this can be done using $O(n_1)$ pebbling moves.
\item Retaining the pebble on the root node of $T_1$, proceed to pebble $T_2$ in the same way as above. Continue this procedure till the root node of $T$ is pebbled. Then proceed to unpebble every other vertex by executing every pebble move up to this instant in reverse order.
\end{enumerate}
Now we argue the bounds on the number of pebbles and pebbling moves of the algorithm. Recall that the number of these subtrees is $O\left(n^{1/k_0}\right)$. Therefore, the number of intermediate pebbles at the root nodes of these subtrees is $O\left(n^{1/k_0}\right)$. Additionally, while pebbling the last subtree, $O\left(n^{1/k_0}\right)$ pebbles are used. Therefore, the total number of pebbles at any time remains $O\left(n^{1/k_0}\right)$. Each of the subtrees are pebbled and unpebbled once (effectively pebbled twice). Therefore the total number of pebbling moves is at most $\sum_i 2O(n_i)=O(n)$.
\end{proof}
|
\section{Asymptotically Optimal Algorithm}
\label{sec:bss}
In addition to sampling by online leverage scores, there is also a variant of the ``BSS'' method \cite{batson2012twice} that applies in our setting. Like the original \cite{batson2012twice}, this approach removes the $\log d$ factor from the row count of the output spectral approximation, matching the lower bound for online sampling given in Theorem \ref{thm:lower_bound}.
Unlike \cite{batson2012twice} itself, our algorithm is randomized -- it is similar to, and inspired by, the randomized version of BSS from \cite{lee2015linear}, especially the simpler ``Algorithm 1'' from that paper (the main difference from that is considering each row separately). In fact, this algorithm is of the same form as the basic sampling algorithm, in that when each row comes in, a probability $p_i$ is assigned to it, and it is kept (and rescaled) with probability $p_i$ and rejected otherwise. The key difference is the definition of the $p_i$.
There are also some differences in the nature of the algorithm and its guarantees. Notably, the $p_i$ cannot be computed solely based on the row sample output so far--it is necessary to ``remember'' the entire matrix given so far. This means that the BSS method is not memory efficient, using $O(d^2)$ space. Additionally, online leverage score sampling gives bounds on both the size of the output spectral approximation and its accuracy with high probability. In contrast, this method gives an \emph{expected} bound on the output size, while it \emph{never} fails to output a correct spectral approximation. Note that these guarantees are essentially the same as those in the appendix of \cite{lee2015linear}.
One may, however, improve the memory dependence in some cases simply by running it on the output stream of the online leverage score sampling method. This reduces the storage cost to the size of that spectral approximation. The BSS method still does not produce an actual space \emph{savings} (in particular, there is a still a $\log d$ factor in space), but it does reduce the number of rows in the output stream while only blowing up the space usage by $O(1/\epsilon^2)$ (due to requiring the storage of an $\epsilon$-quality approximation rather than only $O(1)$).
The BSS method maintains two matrices, $\bv{B}^U_i$ and $\bv{B}^L_i$, acting as upper and lower ``barriers''. The current spectral approximation will always fall between them:
\begin{equation*}
\bv{B}^L_i \prec \bv{\tilde{A}}_i^T \bv{\tilde{A}}_i^T \prec \bv{B}^U_i.
\end{equation*}
This guarantee, at the end of the algorithm, will ensure that $\bv{\tilde{A}}$ is a valid approximation.
Below, we give the actual BSS algorithm and its performance guarantees.
\begin{theorem}
\leavevmode
\label{thm:bss}
\begin{enumerate}
\item
The online BSS algorithm always outputs $\tilde{A}$ such that
\begin{equation*}
(1-\epsilon) \bv{A}^T \bv{A} - \delta \bv{I} \prec \bv{\tilde{A}}^T \bv{\tilde{A}}^T \prec (1+\epsilon) \bv{A}^T \bv{A} + \delta \bv{I}
\end{equation*}
\item
The probability that a row $\bv{a}_i$ is included in $\bv{\tilde{A}}$ is at most $\frac{8}{\epsilon^2} l_i$, where $l_i$ is the online $\frac{2 \delta}{\epsilon}$-ridge leverage score of $\bv{a}_i$. That is $l_i = \min(\bv{a}_i^T \left ( \bv{A}_i^T \bv{A}_i + \frac{2 \delta}{\epsilon} I \right )^{-1} \bv{a}_i, 1)$. The expected number of rows in $\bv{\tilde{A}}$ is thus at most $\frac{8}{\epsilon^2} \sum_{i=1}^n l_i =\ensuremath{\mathcal{O}}(d \log (\epsilon\norm{\bv A}_2^2/\delta) / \epsilon^2)$.
\end{enumerate}
\end{theorem}
\begin{figure}[ht]
\noindent
\centering
\fbox{
\begin{minipage}{6in}
\noindent $\bv{\tilde{A}} = \textsc{Online-BSS} (\bv{A}, \epsilon, \delta)$,
where $\bv{A}$ is an $n \times d$ matrix with rows $\bv{a}_1, \ldots, \bv{a}_n$, $\epsilon \in (0, 1)$, $\delta > 0$.
\begin{enumerate}
\item Set $c_U = \frac{2}{\epsilon}+1$ and $c_L = \frac{2}{\epsilon}-1$.
\item Let $\bv{\tilde{A}}_0$ be a $0 \times d$ matrix, $\bv{B}^U_0 = \delta \bv{I}$, $\bv{B}^L_0 = -\delta \bv{I}$.
\item For $i = 1, \ldots, n$:
\begin{enumerate}
\item Let $\bv{X}^U_{i-1} = (\bv{B}^U_{i-1}-\bv{\tilde{A}}_{i-1}^T \bv{\tilde{A}}_{i-1})$, $\bv{X}^L_{i-1} = (\bv{\tilde{A}}_{i-1}^T \bv{\tilde{A}}_{i-1} - \bv{B}^L_{i-1})$.
\item Let $p_i := \min(c_U \bv{a}_i^T (\bv{X}^U_{i-1})^{-1} \bv{a}_i + c_L \bv{a}_i^T (\bv{X}^L_{i-1})^{-1} \bv{a}_i, 1)$.
\item Set
$
\bv{\tilde{A}}_i :=
\begin{cases}
\begin{bmatrix} \bv{\tilde{A}}_{i-1}\\\bv{a}_i / \sqrt{p_i}\end{bmatrix} &\mbox{ with probability $p_i$,} \vspace{0.2cm}\\
\bv{\tilde{A}}_{i-1} &\mbox{ otherwise.}\\
\end{cases}
$
\item Set $\bv{B}^U_i = \bv{B}^U_{i-1} + (1+\epsilon) \bv{a}_i \bv{a}_i^T$, $\bv{B}^L_i = \bv{B}^L_{i-1} + (1-\epsilon) \bv{a}_i \bv{a}_i^T$.
\end{enumerate}
\item Return $\bv{\tilde{A}} := \bv{\tilde{A}}_n$.
\end{enumerate}
\end{minipage}
}
\caption{The Online BSS Algorithm}
\label{fig:online-bss}
\end{figure}
\Proofof{Theorem \ref{thm:bss} part 1}
We first note the basic invariant that $\bv{X}^U_i$ and $\bv{X}^L_i$ always remain positive definite--or equivalently,
\begin{equation*}
\bv{B}^L_i \prec \bv{\tilde{A}}_i^T \bv{\tilde{A}}_i^T \prec \bv{B}^U_i.
\end{equation*}
We may prove this by induction on $i$. The base case follows from the initialization of $\bv{\tilde{A}}_0$, $\bv{B}^U_0$ and $\bv{B}^L_0$. For each successive step, we consider two possibilities.
The first is that $p_i = 1$. In that case, $\bv{\tilde{A}}^T \bv{\tilde{A}}$ always increases by exactly $\bv{a}_i \bv{a}_i^T$, $\bv{B}^U$ by $(1+\epsilon) \bv{a}_i \bv{a}_i^T$ and $\bv{B}^L$ by $(1-\epsilon) \bv{a}_i \bv{a}_i^T$. Thus $\bv{X}^U$ and $\bv{X}^L$ increase by exactly $\epsilon \bv{a}_i \bv{a}_i^T$, which is positive semidefinite, and so remain positive definite.
In the other case, $p_i < 1$. Now, $\bv{X}^U$ decreases by at most the increase in $\bv{\tilde{A}}_i^T \bv{\tilde{A}}_i^T$, or
\begin{equation*}
\bv{M}_i = \frac{\bv{a}_i \bv{a}_i^T}{p}.
\end{equation*}
Since $c_U > 1$, $p > \bv{a}_i^T (\bv{X}^U_{i-1})^{-1} \bv{a}_i$, so $\bv{a}_i \bv{a}_i^T \prec p \bv{X}^U_{i-1}$ and $\bv{M}_i \prec \bv{X}^U_{i-1}$. Subtracting this then must leave $\bv{X}^U$ positive definite.
Similarly, $\bv{X}^L$ decreases by at most the increase in $\bv{B}^L$, which is $(1-\epsilon) \bv{a}_i \bv{a}_i^T \prec \bv{a}_i \bv{a}_i^T$. Since $c_L > 1$ and $p < 1$, $\bv{a}_i^T (\bv{X}^L_{i-1})^{-1} \bv{a}_i < 1$, and $\bv{a}_i \bv{a}_i^T \prec \bv{X}^L_{i-1}$. Subtracting this similarly leaves $\bv{X}^L$ positive definite.
Finally, we note that
\begin{align*}
\bv{B}^U_n &= (1+\epsilon) \bv{A}^T \bv{A} + \delta \bv{I} \\
\bv{B}^L_n &= (1-\epsilon) \bv{A}^T \bv{A} - \delta \bv{I}.
\end{align*}
This gives the desired result.
\QED
To prove part 2, we will use quantities of the form $\bv{v}^T \bv{X}^{-1} \bv{v}$. We will need a lemma describing how this behaves under a random rank-1 update:
\begin{lemma}
\label{lem:rankoneupdate}
Given a positive definite matrix $\bv{X}$, two vectors $\bv{u}$ and $\bv{v}$, two multipliers $a$ and $b$ and a probability $p$, define the random variable $\bv{X}'$ to be $X - a \bv{u} \bv{u}^T$ with probability $p$ and $X - b \bv{u} \bv{u}^T$ otherwise. Then if $\bv{u}^T \bv{X}^{-1} \bv{u} = 1$,
\begin{equation*}
\expct{}{}{\bv{v}^T \bv{X}'^{-1} \bv{v} - \bv{v}^T \bv{X}^{-1} \bv{v}] = (\bv{v^T} \bv{X}^{-1} \bv{u})^2 \frac{pa + (1-p)b - ab}{(1-a)(1-b)}}.
\end{equation*}
\end{lemma}
\Proof
We apply the Sherman-Morrison formula to each of the two possibilities (subtracting $a \bv{u} \bv{u}^T$ and $b \bv{u} \bv{u}^T$ respectively). These give $\bv{X}'$ values of respectively
\begin{equation*}
\bv{X}^{-1} + a \frac{\bv{X}^{-1} \bv{u} \bv{u}^T \bv{X}^{-1}}{1 - a \bv{u}^T \bv{X}^{-1} \bv{u}} = \bv{X}^{-1} + \frac{a}{1-a} \bv{X}^{-1} \bv{u} \bv{u}^T \bv{X}^{-1}
\end{equation*}
and
\begin{equation*}
\bv{X}^{-1} + b \frac{\bv{X}^{-1} \bv{u} \bv{u}^T \bv{X}^{-1}}{1 - b \bv{u}^T \bv{X}^{-1} \bv{u}} = \bv{X}^{-1} + \frac{b}{1-b} \bv{X}^{-1} \bv{u} \bv{u}^T \bv{X}^{-1}.
\end{equation*}
The values of $\bv{v}^T \bv{X}'^{-1} \bv{v} - \bv{v}^T \bv{X}^{-1} \bv{v}$ are then respectively
\begin{equation*}
\frac{a}{1-a} \bv{v}^T \bv{X}^{-1} \bv{u} \bv{u}^T \bv{X}^{-1} \bv{v} = (\bv{v^T} \bv{X}^{-1} \bv{u})^2 \frac{a}{1-a}
\end{equation*}
and
\begin{equation*}
\frac{b}{1-b} \bv{v}^T \bv{X}^{-1} \bv{u} \bv{u}^T \bv{X}^{-1} \bv{v} = (\bv{v^T} \bv{X}^{-1} \bv{u})^2 \frac{b}{1-b}.
\end{equation*}
Combining these gives the stated result.
\QED
\Proofof{Theorem \ref{thm:bss} part 2}
First, we introduce some new matrices to help in the analysis:
\begin{align*}
\bv{C}^U_{i,j} &= \delta \bv{I} + \frac{\epsilon}{2} \bv{A}_i^T \bv{A}_i + \left ( 1 + \frac{\epsilon}{2} \right ) \bv{A}_j^T \bv{A}_j \\
\bv{C}^L_{i,j} &= -\delta \bv{I} - \frac{\epsilon}{2} \bv{A}_i^T \bv{A}_i + \left ( 1 - \frac{\epsilon}{2} \right ) \bv{A}_j^T \bv{A}_j.
\end{align*}
Note that $\bv{C}^U_{i,i} = \bv{B}^U_i$, $\bv{C}^L_{i,i} = \bv{B}^L_i$, and for $j \leq i$, $\bv{C}^U_{i,j} \succeq \bv{B}^U_j$ and $\bv{C}^L_{i,j} \preceq \bv{B}^L_j$.
We can then define:
\begin{align*}
\bv{Y}^U_{i,j} &= \bv{C}^U_{i,j} - \bv{\tilde{A}}_j^T \bv{\tilde{A}}_j \\
\bv{Y}^L_{i,j} &= \bv{\tilde{A}}_j^T \bv{\tilde{A}}_j - \bv{C}^L_{i,j}.
\end{align*}
We then have, similarly, $\bv{Y}^U_{i,i} = \bv{X}^U_i$, $\bv{Y}^L_{i,i} = \bv{X}^L_i$, and for $j \leq i$, $\bv{Y}^U_{i,j} \succeq \bv{X}^U_j$ and $\bv{Y}^L_{i,j} \succeq \bv{X}^L_j$.
We will assume that $l_i < 1$, since otherwise the claim is immediate (as probabilities cannot exceed 1). Now, note that
\begin{align*}
\bv{a}_i^T (\bv{Y}^U_{i,0})^{-1} \bv{a}_i &= \bv{a}_i^T (\bv{Y}^L_{i,0})^{-1} \bv{a}_i \\
&= \bv{a}_i^T \left ( \frac{\epsilon}{2} \bv{A}_i^T \bv{A}_i + \delta I \right )^{-1} \bv{a}_i \\
&= \frac{2}{\epsilon} \left ( \bv{A}_i^T \bv{A}_i + \frac{2 \delta}{\epsilon} I \right )^{-1} \bv{a}_i \\
&= \frac{2}{\epsilon} l_i.
\end{align*}
Next, we will aim to show that for $j < i-1$,
\begin{align*}
\expct{}{}{\bv{a}_i^T \bv{Y}^U_{i-1,j+1} \bv{a}_i} &\leq \expct{}{}{\bv{a}_i^T \bv{Y}^U_{i-1,j} \bv{a}_i} \\
\expct{}{}{\bv{a}_i^T \bv{Y}^L_{i-1,j+1} \bv{a}_i} &\leq \expct{}{}{\bv{a}_i^T \bv{Y}^L_{i-1,j} \bv{a}_i}
\end{align*}
In particular, we will simply show that conditioned on any choices for the first $j$ rows, the expected value of $\bv{a}_i^T \bv{Y}^U_{i-1,j+1} \bv{a}_i$ is no larger than $\bv{a}_i^T \bv{Y}^U_{i-1,j} \bv{a}_i$, and analogously for $\bv{Y}^L$.
Similar to the proof of part 1, we separately consider the case where $p_{j+1} = 1$. In that case, the positive semidefinite matrix $\frac{\epsilon}{2} \bv{a}_j \bv{a}_j^T$ is simply added to $\bv{Y}^U$ and $\bv{Y}^L$. Adding this can only decrease the values of $\bv{a}_i^T \bv{Y}^U \bv{a}_i$ and $\bv{a}_i^T \bv{Y}^L \bv{a}_i$.
The $p_{j+1} < 1$ case is more tricky. Here, we define the vector $\bv{w}_{j+1} = \frac{\bv{a}_{j+1}}{\sqrt{p_{j+1}}}$. Importantly
\begin{gather*}
p_{j+1} \geq c_U \bv{a}_{j+1}^T (\bv{X}^U_j)^{-1} \bv{a}_{j+1} \geq c_U \bv{a}_{j+1}^T (\bv{Y}^U_{i-1,j})^{-1} \bv{a}_{j+1} \\
p_{j+1} \geq c_L \bv{a}_{j+1}^T (\bv{X}^L_j)^{-1} \bv{a}_{j+1} \geq c_L \bv{a}_{j+1}^T (\bv{Y}^L_{i-1,j})^{-1} \bv{a}_{j+1}.
\end{gather*}
This means that
\begin{align*}
\bv{w}_{j+1}^T (\bv{Y}^U_{i-1,j})^{-1} \bv{w}_{j+1}^T &\leq \frac{1}{c_U} \\
\bv{w}_{j+1}^T (\bv{Y}^L_{i-1,j})^{-1} \bv{w}_{j+1}^T &\leq \frac{1}{c_L}.
\end{align*}
Now, we additionally define
\begin{align*}
s^U_{j+1} &= \bv{w}_{j+1}^T (\bv{Y}^U_{i-1,j})^{-1} \bv{w}_{j+1}^T \\
s^L_{j+1} &= \bv{w}_{j+1}^T (\bv{Y}^L_{i-1,j})^{-1} \bv{w}_{j+1}^T \\
\bv{u}^U_{j+1} &= \frac{\bv{w}_{j+1}}{\sqrt{s^U_{j+1}}} \\
\bv{u}^L_{j+1} &= \frac{\bv{w}_{j+1}}{\sqrt{s^L_{j+1}}}.
\end{align*}
We then deploy Lemma \ref{lem:rankoneupdate} to compute the expectations. For the contribution from the upper barrier, we use $\bv{X} = \bv{Y}^U_{i-1,j}$, $\bv{u} = \bv{u}^U_{j+1}$, $\bv{v} = \bv{a}_i^T$, $a = -s^U_{j+1} (1 - p_{j+1} (1 + \epsilon / 2))$, $b = s^U_{j+1} p_{j+1} (1 + \epsilon / 2)$, $p = p_{j+1}$. For the lower barrier, we use $\bv{X} = \bv{Y}^L_{i-1,j}$, $\bv{u} = \bv{u}^L_{j+1}$, $\bv{v} = \bv{a}_i^T$, $a = s^L_{j+1} (1 - p_{j+1} (1 - \epsilon / 2))$, $b = -s^L_{j+1} p_{j+1} (1 - \epsilon / 2)$, $p = p_{j+1}$. In both cases we can see that the numerator of the expected change is nonpositive.
Finally, this implies that the probability that row $i$ is sampled is
\begin{align*}
\expct{}{}{p_i} &= c_U \expct{}{}{\bv{a}_i^T (\bv{X}^U_{i-1})^{-1} \bv{a}_i} + c_L \expct{}{}{\bv{a}_i^T (\bv{X}^L_{i-1})^{-1} \bv{a}_i} \\
&= c_U \expct{}{}{\bv{a}_i^T (\bv{Y}^U_{i-1,i-1})^{-1} \bv{a}_i} + c_L \expct{}{}{\bv{a}_i^T (\bv{Y}^L_{i-1,i-1})^{-1} \bv{a}_i} \\
&\leq c_U \expct{}{}{\bv{a}_i^T (\bv{Y}^U_{i-1,0})^{-1} \bv{a}_i} + c_L \expct{}{}{\bv{a}_i^T (\bv{Y}^L_{i-1,0})^{-1} \bv{a}_i} \\
&= \frac{2}{\epsilon} (c_U + c_L) l_i \\
&= \frac{8}{\epsilon^2} l_i
\end{align*}
as desired.
\QED
\section{Future Work}
An obvious open question arising from our work is if one can prove that the algorithm of \cite{kelner2013spectral} works despite dependencies arising due to the row pruning step. By operating in the online setting, our algorithm avoids row pruning, and hence is able to skirt these dependencies, as the probability that a row is sampled only depends on earlier rows in the stream. However, because the streaming setting offers the potential for sampling fewer rows than in the online case, obtaining a rigorous proof of \cite{kelner2013spectral} would be very interesting.
While our work focuses on spectral approximation, variants on (ridge) leverage score sampling and the BSS algorithm are also used to solve low-rank approximation problems, including column subset selection \cite{boutsidis2014optimal,cohen2015ridge} and projection-cost-preserving sketching \cite{cohen2014dimensionality,cohen2015ridge}. Compared with spectral approximation, there is less work on streaming sampling for low-rank approximation, and understanding how online algorithms may be used in this setting would an interesting extension of our work.
\section{Introduction}
\label{sec:intro}
\subsection{Background}
A spectral approximation to a tall $n \times d$ matrix $\bv A$ is a smaller, typically $\tilde \ensuremath{\mathcal{O}}(d) \times d$ matrix $\bv{\tilde A}$ such that $\norm{\bv{\tilde A} \bv{x}}_2 \approx \norm{\bv{ A} \bv{x}}_2$ for all $\bv{x}$. Typically one asks for a multiplicative approximation, which guarantees that $(1-\epsilon)\norm{\bv{ A} \bv{x}}^2_2 \le \norm{\bv{\tilde A} \bv{x}}^2_2 \le (1+\epsilon)\norm{\bv{ A} \bv{x}}^2_2$. In other notation, $(1-\epsilon)\bv{A} \preceq \bv{\tilde A} \preceq (1+\epsilon) \bv{A}$.
Such approximations have many applications, most notably for solving least squares regression over $\bv A$ \cite{Clarkson:2013,cohen2015uniform}. If $\bv{A}$ is the vertex edge incidence matrix of a graph, $\bv{\tilde A}$ is a \emph{spectral sparsifier} \cite{spielman2004nearly}. It can be used to approximate effective resistances, spectral clustering, mixing time and random walk properties, and many other computations.
A number of recent papers focus on fast algorithms for spectral approximation. Using sparse random subspace embeddings \cite{Clarkson:2013,DBLP:conf/focs/NelsonN13,meng2013low}, it is possible to find $\bv{\tilde A}$ in input sparsity time, essentially by randomly recombining the rows of $\bv{A}$ into a smaller number of rows. In some cases these embeddings are not enough, as it is desirable for the rows of $\bv{\tilde A}$ to be a subset of rows sampled from $\bv{A}$. If $\bv{A}$ is sparse, this ensures that $\bv{\tilde A}$ is also sparse. If $\bv{A}$ represents a graph, it ensures that $\bv{\tilde A}$ is also a graph, specifically a weighted subgraph of the original.
It is well known that sampling $\ensuremath{\mathcal{O}}(d\log d/\epsilon^2)$ rows of $\bv{A}$ with probabilities proportional to their \emph{leverage scores} yields a $(1+\epsilon)$ multiplicative factor spectral approximation to $\bv{A}$. Further, this sampling can be done in input sparsity time, either using subspace embeddings to approximate leverage scores, or using iterative sampling techniques \cite{li2013iterative}, some that only work with subsampled versions of the original matrix \cite{cohen2015uniform}.
\subsection{Streaming and Online Row Sampling}
When $\bv{A}$ is very large, input sparsity runtimes are not enough -- memory restrictions also become important. Hence, recent work has tackled row sampling in a streaming model of computation. \cite{kelner2013spectral} presents a simple algorithm for sampling rows from an insertion only stream, using space approximately proportional to the size of the final approximation. \cite{kapralov2014single} gives a sparse-recovery based algorithm that works in dynamic streams with row insertions and deletions, also using nearly optimal space. Unfortunately, to handle dynamic streams, the algorithm in \cite{kapralov2014single} is complex, requires additional restrictions on the input matrix, and uses significantly suboptimal runtime to recover a spectral approximation from its low memory representation of the input stream.
While the algorithm in \cite{kelner2013spectral} is simple and efficient, we believe that its proof is incomplete, and do not see an obvious way to fix it. The main idea behind the algorithm is to sample rows by their leverage scores with respect to the stream seen so far. These leverage scores may be coarse overestimates of the true scores. However as more rows are streamed in, better estimates can be obtained and the sampled rows pruned to a smaller set. Unfortunately, the probability of sampling a row becomes dependent on which other rows are sampled. This seems to break the argument in that paper, which essentially claims that their process has the same distribution as would a single round of leverage score sampling.
In this paper we initiate the study of row sampling in an \emph{online setting}. As in an insertion stream, we read rows of $\bv A$ one by one. However, upon seeing a row, we immediately decide whether it should be kept in the spectral approximation or discarded, without ever retracting these decisions.
We present a similar algorithm to \cite{kelner2013spectral}, however, since we never prune previously sampled rows, the probability of sampling a row only depends on whether previous rows in the stream were sampled. This limited dependency structure allows us to rigorously argue that a spectral approximation is obtained.
In addition to addressing gaps in the literature on streaming spectral approximation, our restricted model extends work on online algorithms for a variety of other machine learning and data analysis problems, including principal component analysis \cite{boutsidis2015online}, clustering \cite{liberty2014algorithm}, classification \cite{bordes2005huller,crammer2006online}, and regression \cite{crammer2006online}. In practice, online algorithms are beneficial since they can be highly computationally and memory efficient. Further, they can be applied in scenarios in which data is produced in a continuous stream and intermediate results must be output as the stream is processed.
Spectral approximation is a widely applicable primitive for approximate learning and computation, so studying its implementation in an online setting is a natural direction.
\subsection{Our Results}
Our primary contribution is a very simple algorithm for leverage score sampling in an online manner. The main difficultly with row sampling using leverage scores is that leverage scores themselves are not easy to compute. They are given by $l_i = \bv{a}_i^T (\bv{A}^T\bv{A})^{-1} \bv{a}_i$, and so require solving systems in $\bv{A}^T\bv{A}$ if computed naively. This is not only expensive, but also impossible in an online setting, where we do not have access to all of $\bv{A}$.
A critical observation is that
it always suffices to sample rows by overestimates of their true leverage scores. The number of rows that must be sampled is proportional to the sum of these overestimates.
Since the leverage score of a row can only go up when we remove rows from the matrix, a
simple way to obtain an overestimate is to compute leverage score using just a subset of the other rows of $\bv{A}$. That is, letting $\bv{A}_j$ contain just $j$ of $\bv{A}$'s $n$ rows, we can overestimate $l_i$ by $\tilde l_i = \bv{a}_i^T (\bv{A}_j^T\bv{A}_j)^{-1} \bv{a}_i$
\cite{cohen2015uniform} shows that
if $\bv{A}_j$ is a subset of rows sampled uniformly at random, then the expected leverage score of $\bv{a}_i$ is $d/j$. This simple fact immediately gives a result for online sampling from a \emph{randomly ordered stream}.
If we compute the leverage score of the current row $\bv{a}_i$ against all previously seen rows (or some approximation to these rows), then the expected sum of our overestimates will be bounded by $d + d/2 + ... + ... + d/n = \ensuremath{\mathcal{O}}(d \log n)$. So, sampling $\ensuremath{\mathcal{O}}(d\log d \log n /\epsilon^2)$ rows will be enough obtain a $(1+\epsilon)$ multiplicative factor spectral approximation.
What if we cannot guarantee a randomly ordered input stream? Is there any hope of being able to compute good leverage score estimates in an online manner? Surprisingly the answer to this is yes - we can in fact run nearly the exact same algorithm and be guaranteed that the sum of estimated leverage scores is low, \emph{regardless of stream order}. Roughly, each time we receive a row which has high leverage score with respect to the previous rows, it must compose a significant part of $\bv{A}$'s spectrum. If $\bv{A}$ does not continue to grow unboundedly, there simply cannot be too many of these significant rows.
Specifically, we show that if we sample by the \emph{ridge leverage scores} \cite{alaoui2014fast} over all previously seen rows, which are the leverage scores computed over $\bv{A}_i^T \bv{A}_i + \lambda \bv I$ for some small regularizing factor $\lambda$, then with just $\ensuremath{\mathcal{O}}(d \log d \log (\epsilon\norm{\bv A}_2^2/\delta) / \epsilon^2)$ samples we obtain a $(1+\epsilon)$ multiplicative, $\delta$ additive error spectral approximation. That is, with high probability we sample a matrix $\bv{\tilde A}$ with $(1-\epsilon)\bv{A}^T\bv{A} - \delta\bv{I} \preceq \bv{\tilde{A}}^T\bv{\tilde{A}} \preceq (1+\epsilon)\bv{A}^T\bv{A} + \delta\bv{I}$.
To gain intuition behind this bound, note that we can convert it into a multiplicative one by setting $\delta = \epsilon \sigma_{min}(\bv{A})^2$ (as long as we have some estimate of $\sigma_{min}(\bv{A})$). This setting of $\delta$ will require taking $\ensuremath{\mathcal{O}}(d \log d \log (\kappa(\bv A)) / \epsilon^2)$ samples. If we have a polynomial bound on the condition number of $\bv{A}$, as we do, for instance, for graphs with polynomially bounded edges weights, this becomes $\ensuremath{\mathcal{O}}(d \log^2 d / \epsilon^2)$ -- nearly matching the $\ensuremath{\mathcal{O}}(d \log d / \epsilon^2)$ achievable if sampling by true leverage scores.
Our online sampling algorithm is extremely simple. When each row comes in, we compute the online ridge leverage score, or an estimate of it, and then irrevocably either add the row to our approximation or remove it. As mentioned, it is similar in form to the streaming algorithm of \cite{kelner2013spectral}, except that it does not require pruning previously sampled rows.
This allows us to avoid difficult dependency issues. Additionally, without pruning, we do not even need to store all previously sampled rows. As long as we store a constant factor spectral approximation our previous samples, we can compute good approximations to the online ridge leverage scores. In this way, we can store just $\ensuremath{\mathcal{O}}(d \log d \log (\epsilon \norm{\bv A}_2^2/\delta) )$ rows in working memory ($\ensuremath{\mathcal{O}}(d\log^2 d)$ if we want a spectral graph sparsifier), filtering our input stream into an $\ensuremath{\mathcal{O}}(d \log d \log (\kappa(\bv A)) / \epsilon^2)$ sized output stream. Note that this memory bound in fact \emph{improves} as $\epsilon$ decreases, and regardless, can be significantly smaller than the output size of the algorithm.
In additional to our main sampling result, we use our bounds on online ridge leverage score approximations to show that an algorithm in the style of \cite{batson2012twice} allows us to remove a $\log d$ factor and sample just $\ensuremath{\mathcal{O}}(d \log (\epsilon\norm{\bv A}_2^2/\delta) / \epsilon^2)$ (Theorem \ref{thm:bss}).
This algorithm is more complex and can require $\ensuremath{\mathcal{O}}(d^2)$ working memory. However, in Theorem \ref{thm:lower_bound} we show that it is asymptotically optimal. The $\log (\epsilon\norm{\bv A}_2^2/\delta)$ factor is not an artifact of our analysis, but is truly the cost of restricting ourselves to online sampling. No algorithm can obtain a multiplicative $(1+\epsilon)$ additive $\delta$ spectral approximation taking fewer than $\Omega(d \log (\epsilon\norm{\bv A}_2^2/\delta) / \epsilon^2)$ rows in an online manner.
\section{Matching Lower Bound}
\label{sec:lower}
Here we show that the row count obtained by Theorem \ref{thm:bss} is in fact optimal. While it is possible to obtain a spectral approximation with $\ensuremath{\mathcal{O}}(d/\epsilon^2)$ rows in the offline setting, online sampling always incurs a loss of $\Omega \left (\log(\epsilon\norm{\bv A}^2_2/\delta) \right)$ and must sample $\Omega \left (\frac{d\log(\epsilon\norm{\bv A}^2_2/\delta)}{\epsilon^2} \right)$ rows.
\begin{theorem}
\label{thm:lower_bound}
Assume that $\epsilon \norm{\bv{A}}_2^2 \ge c_1\delta$ and $\epsilon \ge c_2/\sqrt{d}$, for fixed constants $c_1$ and $c_2$.
Then any algorithm that selects rows in an online manner and outputs a spectral approximation to $\bv{A}^T\bv{A}$ with $(1+\epsilon)$ multiplicative error and $\delta$ additive error with probability at least $1/2$ must sample $\Omega \left (\frac{d\log(\epsilon\norm{\bv A}^2_2/\delta)}{\epsilon^2} \right)$ rows of $\bv{A}$ in expectation.
\end{theorem}
Note that the lower bounds we assume on $\epsilon \norm{\bv{A}}_2^2$ and $\epsilon$ are very minor. They just ensure that $\log(\epsilon\norm{\bv A}^2_2/\delta) \ge 1$ and that $\epsilon$ is not so small that we can essentially sample all rows of $\bv{A}$.
\Proof
We apply Yao's minimax principle, constructing, for any large enough $M$, a distribution on inputs $\bv{A}$ with $\norm{\bv{A}}_2^2 \le M$ for which any deterministic online row selection algorithm that succeeds with probability at least $1/2$ must output $\Omega \left (\frac{d\log(\epsilon M/\delta)}{\epsilon^2} \right)$ rows in expectation. The best randomized algorithm that works with probability $1/2$ on any input matrix with $\norm{\bv{A}}_2^2 \le M$ therefore must select at least $\Omega \left (\frac{d\log(\epsilon M/\delta)}{\epsilon^2} \right)$ rows in expectation on the worst case input, giving us the theorem.
Our distribution is as follows. We select an integer $N$ uniformly at random from $[1, \log(M\epsilon/\delta)]$. We then stream in the vertex edge incidence matrices of $N$ complete graphs on $d$ vertices. We double the weight of each successive graph. Intuitively, spectrally approximating a complete graph requires selecting $\Omega(d/\epsilon^2)$ edges \cite{batson2012twice} (as long as $\epsilon \ge c_2/\sqrt{d}$ for some fixed constant $c_2$). Each time we stream in a new graph with double the weight, we force the algorithm to add $\Omega(d/\epsilon^2)$ more edges to its output, eventually forcing it to output $\Omega(d/\epsilon^2 \cdot N)$ edges -- $\Omega(d\log(M\epsilon/\delta)/\epsilon^2)$ in expectation.
Specifically, let $\bv{K}_d$ be the ${d \choose 2} \times d$ vertex edge incidence matrix of the complete graph on $d$ vertices. $\bv{K}_d^T\bv{K}_d$ is the Laplacian matrix of the complete graph on $d$ vertices. We weight the first graph so that its Laplacian has all its nonzero eigenvalues equal to $\delta/\epsilon$. (That is, each edge has weight $\frac{\delta}{d\epsilon}$). In this way, even if we select $N = \lfloor \log(M\epsilon/\delta) \rfloor$ we will have overall $\norm{\bv{A}}^2_2 \le \delta/\epsilon + 2\delta/\epsilon + ... 2^{\lfloor \log(M\epsilon/\delta) \rfloor-1} \delta/\epsilon \le M$.
Even if $N = 1$, all nonzero eigenvalues of $\bv{A}^T\bv{A}$ are at least $\delta/\epsilon$, so achieving $(1+\epsilon)$ multiplicative error and $\delta \bv{I}$ additive error is equivalent to achieving $(1+2\epsilon)$ multiplicative error. $\bv{A}^T\bv{A}$ is a graph Laplacian so has a null space. However, as all rows are orthogonal to the null space, achieving additive error $\delta \bv{I}$ is equivalent to achieving additive error $\delta \bv{I}_r$ where $\bv{I}_r$ is the identity projected to the span of $\bv{A}^T\bv{A}$. $\delta \bv{I}_r \preceq \epsilon \bv{A}^T\bv{A}$ which is why we must achieve $(1+2\epsilon)$ multiplicative error.
In order for a deterministic algorithm to be correct with probability $1/2$ on our distribution, it must be correct for at least $1/2$ of our $\lfloor \log(M\epsilon/\delta) \rfloor$ possible choices of $N$.
Let $i$ be the lowest choice of $N$ for which the algorithm is correct. By the lower bound of \cite{batson2012twice}, the algorithm must output $\Omega(d/\epsilon^2)$ rows of $\bv{A}_i$ to achieve a $(1+2\epsilon)$ multiplicative factor spectral approximation. Here $\bv{A}_i$ is the input consisting of the vertex edge incidence matrices of $i$ increasingly weighted complete graphs. Call the output on this input $\bv{\tilde A}_i$.
Now let $j$ be the second lowest choice of $N$ on which the algorithm is correct. Since the algorithm was correct on $\bv{A}_i$ to within a multiplicative $(1+2\epsilon)$, to be correct on $\bv{A}_j$,
it must output a set of edges $\bv{\tilde A}_j$ such that
\begin{align*}
(\bv{A}_j^T \bv{A}_j- \bv{A}_i^T \bv{A}_i) - 4 \epsilon \bv{A}_j^T \bv{A}_j \preceq \bv{\tilde A}_j^T\bv{\tilde A}_j - \bv{\tilde A}_i^T\bv{\tilde A}_i \preceq (\bv{A}_j^T \bv{A}_j- \bv{A}_i^T \bv{A}_i) + 4 \epsilon \bv{A}_j^T \bv{A}_j.
\end{align*}
Since we double each successive copy of the complete graph, $\bv{A}_j^T \bv{A}_j \preceq 2(\bv{A}_j^T \bv{A}_j- \bv{A}_i^T \bv{A}_i)$. So, $\bv{\tilde A}_j^T\bv{\tilde A}_j - \bv{\tilde A}_i^T\bv{\tilde A}_i$ must be a $1+8\epsilon$ spectral approximation to the true difference $\bv{A}_j^T \bv{A}_j - \bv{A}_i^T \bv{A}_i$. Noting that this difference is itself just a weighting of the complete graph, by the lower bound in \cite{batson2012twice} the algorithm must select $\Omega(d/\epsilon^2)$ additional edges between the $i^{th}$ and $j^{th}$ input graphs. Iterating this argument over all $\lfloor \log(M\epsilon/\delta) \rfloor/2$ inputs on which the algorithm must be correct, it must select a total of $\Omega(d\log(M\epsilon/\delta)/\epsilon^2)$ edges in expectation over all inputs.
\QED
\section{Acknowledgments}
The authors would like to thank Kenneth Clarkson, Jonathan Kelner, Gary Miller, Christopher Musco and Richard Peng for helpful discussions and comments.
Cameron Musco and Jakub Pachocki both acknowledge the Gene Golub SIAM Summer School program on Randomization in Numerical Linear Algebra, where work on this project was initiated.
\bibliographystyle{alpha}
\section{Overview}
\label{sec:overview}
Let $\bv{A}$ be an $n \times d$ matrix with rows $\bv{a}_1, \ldots, \bv{a}_n$.
A natural approach to row sampling from $\bv A$ is picking an \emph{a priori} probability with which each row is kept,
and then deciding whether to keep each row independently.
A common choice is for the sampling probabilities to be proportional to the \emph{leverage scores} of the rows.
The leverage score of the $i$-th row of $\bv{A}$ is defined to be
\begin{align*}
\bv{a}_i^T (\bv{A}^T\bv{A})^{\dagger} \bv{a}_i,
\end{align*}
where the dagger symbol denotes the pseudoinverse.
In this work, we will be interested in approximating $\bv{A}^T\bv{A}$ with some (very) small multiple of the identity added.
Hence, we will be interested in the \emph{$\lambda$-ridge leverage scores} \cite{alaoui2014fast}:
\begin{align*}
\bv{a}_i^T (\bv{A}^T\bv{A} + \lambda \bv{I})^{-1} \bv{a}_i,
\end{align*}
for a parameter $\lambda > 0$.
In many applications, obtaining the (nearly) exact values of $\bv{a}_i^T (\bv{A}^T\bv{A} + \lambda \bv{I})^{-1} \bv{a}_i$ for sampling is difficult or outright impossible.
A key idea is that as long as we have a sequence $l_1, \ldots, l_n$ of \emph{overestimates} of the $\lambda$-ridge leverage scores, that is for $i=1,\ldots,n$
\begin{align*}
l_i \geq \bv{a}_i^T (\bv{A}^T\bv{A} + \lambda \bv{I})^{-1} \bv{a}_i,
\end{align*}
we can sample by these overestimates and obtain rigorous guarantees on the quality of the obtained spectral approximation.
This notion is formalized in Theorem \ref{thm:leverage_sampling}.
\begin{theorem}
\label{thm:leverage_sampling}
Let $\bv{A}$ be an $n \times d$ matrix with rows $\bv{a}_1, \ldots, \bv{a}_n$.
Let $\epsilon \in (0, 1), \delta > 0, \lambda := \delta / \epsilon, c := 8 \log d / \epsilon^2$.
Assume we are given $l_1, \ldots, l_n$ such that for all $i = 1, \ldots, n$,
\begin{align*}
l_i &\geq \bv{a}_i^T (\bv{A}^T \bv{A} + \lambda \bv{I})^{-1} \bv{a}_i.
\end{align*}
For $i = 1, \ldots, n$, let
$
p_i := \min(cl_i, 1).
$
Construct $\bv{\tilde{A}}$ by independently sampling each row $\bv{a}_i$ of $\bv{A}$ with probability $p_i$, and rescaling it by $1 / \sqrt{p_i}$ if it is included in the sample.
Then, with high probability,
\begin{align*}
(1-\epsilon)\bv{A}^T\bv{A} - \delta\bv{I} \preceq \bv{\tilde{A}}^T\bv{\tilde{A}} \preceq (1+\epsilon)\bv{A}^T\bv{A} + \delta\bv{I},
\end{align*}
and the number of rows in $\bv{\tilde{A}}$ is $\ensuremath{\mathcal{O}}\left(\left(\sum_{i=1}^n l_i\right) \log d / \epsilon^2\right)$.
\end{theorem}
\Proof
This sort of guarantee for leverage score sampling is well known. See for example Lemma 4 of \cite{cohen2015uniform}. If we sampled both the rows of $\bv{A}$ and the rows of $\sqrt{\lambda} \bv{I}$ with the leverage scores over $(\bv{A}^T\bv{A} + \lambda \bv I)$, we would have $(1-\epsilon)(\bv{A}^T\bv{A} + \lambda \bv I) \preceq \bv{\tilde A}^T \bv{\tilde A} \preceq(1+\epsilon)(\bv{A}^T\bv{A} + \lambda \bv I)$. However, we do not sample the rows of the identity. Since we could have sampled them each with probability $1$, we can simply subtract $\lambda \bv I = (\delta/\epsilon) \bv I$ from the multiplicative bound and have: $(1-\epsilon)\bv{A}^T\bv{A} - \delta\bv{I} \preceq \bv{\tilde{A}}^T\bv{\tilde{A}} \preceq (1+\epsilon)\bv{A}^T\bv{A} + \delta\bv{I}$.
\QED
The idea of using overestimates of leverage scores to perform row sampling has been applied successfully to various problems (see e.g. \cite{koutis2010approaching,cohen2015uniform}).
However, in these applications, access to the entire matrix is required beforehand.
In the streaming and online settings, we have to rely on partial data to approximate the true leverage scores.
The most natural idea is to just use the portion of the matrix seen thus far as an approximation to $\bv{A}$.
This leads us to introduce the \emph{online $\lambda$-ridge leverage scores}:
\begin{align*}
l_i := \min(\bv{a}_i^T (\bv{A}_{i-1}^T \bv{A}_{i-1} + \lambda \bv{I})^{-1} \bv{a}_i, 1),
\end{align*}
where $\bv{A}_i$ $(i = 0,\ldots,n)$ is defined as the matrix consisting of the first $i$ rows of $\bv{A}$\footnote{We use the proposed scores $l_i$ for simplicity, however note that the following, perhaps more natural, definition of online leverage scores would also be effective:
\begin{align*}
l'_i := \bv{a}_i^T (\bv{A}_{i}^T \bv{A}_{i} + \lambda \bv{I})^{-1} \bv{a}_i.
\end{align*}}.
Since clearly $\bv{A}_i^T\bv{A}_i \preceq \bv{A}^T\bv{A}$ for all $i$, it is not hard to see that $l_i$ does overestimate the true $\lambda$-ridge leverage score for row $\bv{a}_i$.
A more complex question, however, is establishing an upper bound on $\sum_{i=1}^n l_i$ so that we can bound the number of samples needed by Theorem \ref{thm:leverage_sampling}.
A core result of this work, stated in Theorem \ref{thm:online_scores}, is establishing such an upper bound; in fact, this bound is shown to be tight up to constants (Theorem \ref{thm:lower_bound}) and is nearly-linear in most cases.
\begin{theorem}
\label{thm:online_scores}
Let $\bv{A}$ be an $n \times d$ matrix with rows $\bv{a}_1, \ldots, \bv{a}_n$.
Let $\bv{A}_i$ for $i \in \{0, \ldots, n\}$ be the matrix consisting of the first $i$ rows of $\bv{A}$.
For $\lambda > 0$,
let
\begin{align*}
l_i := \min(\bv{a}_i^T (\bv{A}_{i-1}^T \bv{A}_{i-1} + \lambda \bv{I})^{-1} \bv{a}_i, 1).
\end{align*}
be the online $\lambda$-ridge leverage score of the $i^{th}$ row of $\bv{A}$.
Then
\begin{align*}
\sum_{i=1}^n l_i = \ensuremath{\mathcal{O}}(d \log (\norm{\bv{A}}_2^2/\lambda)).
\end{align*}
\end{theorem}
Theorems \ref{thm:online_scores} and \ref{thm:leverage_sampling} suggest a simple algorithm for online row sampling: simply use the online $\lambda$-ridge leverage scores, for $\lambda := \delta / \epsilon$.
This produces a spectral approximation with only $\ensuremath{\mathcal{O}}(d \log d \log(\epsilon\norm{\bv{A}}_2^2/\delta)/\epsilon^2)$ rows.
Unfortunately, computing $l_i$ exactly requires us to store \emph{all} the rows we have seen in memory (or alternatively to store the sum of their outer products, $\bv{A}_i^T\bv{A}_i$).
In many cases, such a requirement would defeat the purpose of streaming row sampling.
A natural idea is to use the sample we have kept thus far as an approximation to $\bv{A}_i$ when computing $l_i$.
It turns out that the approximate online ridge leverage scores $\tilde{l}_i$ computed in this way will not always be good approximations to $l_i$; however, we can still prove that they satisfy the requisite bounds and yield the same row sample size!
We formalize these results in the algorithm $\textsc{Online-Sample}$ (Figure \ref{fig:online-sample}) and Theorem \ref{thm:online-sample}.
\begin{figure}[ht]
\noindent
\centering
\fbox{
\begin{minipage}{6in}
\noindent $\bv{\tilde{A}} = \textsc{Online-Sample} (\bv{A}, \epsilon, \delta)$,
where $\bv{A}$ is an $n \times d$ matrix with rows $\bv{a}_1, \ldots, \bv{a}_n$, $\epsilon \in (0, 1)$, $\delta > 0$.
\begin{enumerate}
\item Set $\lambda := \delta / \epsilon$, $c := 8\log d / \epsilon^2$.
\item Let $\bv{\tilde{A}}_0$ be a $0 \times d$ matrix.
\item For $i = 1, \ldots, n$:
\begin{enumerate}
\item Let $\tilde{l}_i := \min((1+\epsilon)\bv{a}_i^T (\bv{\tilde{A}}_{i-1}^T \bv{\tilde{A}}_{i-1} + \lambda \bv{I})^{-1} \bv{a}_i, 1)$.
\item Let $p_i := \min(c\tilde{l}_i, 1)$.
\item Set
$
\bv{\tilde{A}}_i :=
\begin{cases}
\begin{bmatrix} \bv{\tilde{A}}_{i-1}\\\bv{a}_i / \sqrt{p_i}\end{bmatrix} &\mbox{ with probability $p_i$,} \vspace{0.2cm}\\
\bv{\tilde{A}}_{i-1} &\mbox{ otherwise.}\\
\end{cases}
$
\end{enumerate}
\item Return $\bv{\tilde{A}} := \bv{\tilde{A}}_n$.
\end{enumerate}
\end{minipage}
}
\caption{The basic online sampling algorithm}
\label{fig:online-sample}
\end{figure}
\begin{theorem}
\label{thm:online-sample}
Let $\bv{\tilde{A}}$ be the matrix returned by $\textsc{Online-Sample}(\bv{A}, \epsilon, \delta)$.
With high probability,
\begin{align*}
(1-\epsilon)\bv{A}^T\bv{A} - \delta\bv{I} \preceq \bv{\tilde{A}}^T\bv{\tilde{A}} \preceq (1+\epsilon)\bv{A}^T\bv{A} + \delta\bv{I},
\end{align*}
and the number of rows in $\bv{\tilde{A}}$ is $\ensuremath{\mathcal{O}}(d \log d \log(\epsilon\norm{\bv{A}}_2^2/\delta)/\epsilon^2)$.
\end{theorem}
To save computation, we note that, with a small modification to our analysis, we can run $\textsc{Online-Sample}$ with batch processing of rows. Specifically, say we start from the $i^{th}$ position in the stream. we can store the next $b=\ensuremath{\mathcal{O}}(d)$ rows. We can then compute sampling probabilities for these rows all at once using a system solver for $(\bv{\tilde A}_{i + b}^T \bv{\tilde A}_{i + b} + \lambda \bv{I})$. Using a trick introduced in \cite{spielman2011graph}, by applying a Johnson-Lindenstrauss random projection to the rows whose scores we are computing, we need just $\ensuremath{\mathcal{O}}(\log(1/\delta))$ system solves to compute constant factor approximations to the ridge scores with probability $1-\delta$. If we set $\delta = 1/\poly{n}$ then we can union bound over our whole stream, using this trick with each batch of $\ensuremath{\mathcal{O}}(d)$ input rows.
The batch probabilities will only be closer to the true ridge leverage scores than the non-batch probabilities and we will enjoy the same guarantees as $\textsc{Online-Sample}$.
Additionally, it turns out that with a simple trick, it is possible to reduce the memory usage of the algorithm by a factor of $\epsilon^{-2}$,
bringing it down to $\ensuremath{\mathcal{O}}(d \log d \log(\epsilon\norm{A}_2^2/\delta))$ (assuming the row sample is output to an output stream).
Note that this expression gets \emph{smaller} with $\epsilon$; hence we obtain a row sampling algorithm with memory complexity independent of desired multiplicative precision. The basic idea is that, instead of keeping all previously sampled rows in memory, we store a smaller set of rows that give a constant factor spectral approximation, still enough to give good estimates of the online ridge leverage scores.
This result is presented in the algorithm $\textsc{Slim-Sample}$ (Figure \ref{fig:slim-sample}) and Lemma \ref{lem:slim-sample}.
A particularly interesting consequence for graphs with polynomially bounded edge weights is:
\begin{corollary}
\label{col:graph}
Let $G$ be a simple graph on $d$ vertices, and $\epsilon \in (0,1)$.
We can construct a $(1+\epsilon)$-sparsifier of $G$ of size $\ensuremath{\mathcal{O}}(d \log^2 d / \epsilon^2)$, using only $\ensuremath{\mathcal{O}}(d \log^2 d)$ working memory in the online model.
\end{corollary}
\Proof
This follows simply from applying Theorem \ref{thm:online-sample} with $\delta = \epsilon/\sigma_{min}^2(\bv A)$ and noting that the condition number of a graph on $d$ vertices whose edge weights are within a multiplicative $\poly{d}$ of each other is polynomial in $d$. So $\log(\epsilon \norm{\bv{A}}_2^2/\delta) = \log(\kappa^2(\bv A)) = \ensuremath{\mathcal{O}}(\log d)$.
\QED
We remark that the algorithm of Corollary \ref{col:graph} can be made to run in nearly linear time in the stream size. We combine $\textsc{Slim-Sample}$ with the batch processing idea described above. Because $\bv{A}$ is a graph, our matrix approximation is always a symmetric diagonally dominant matrix, with $\ensuremath{\mathcal{O}}(d)$ nonzero entries. We can solve systems in it in time $\tilde \ensuremath{\mathcal{O}}(d)$. Using the Johnson-Lindenstrauss random projection trick of \cite{spielman2011graph}, we can compute approximate ridge leverage scores for a batch of $\ensuremath{\mathcal{O}}(d)$ rows with failure probability polynomially small in $n$ in $\tilde \ensuremath{\mathcal{O}}(d \log n)$ time. Union bounding over the whole stream, we obtain nearly linear runtime.
To complement the row sampling results discussed above,
we explore the limits of the proposed online setting.
In Section \ref{sec:bss} we present the algorithm $\textsc{Online-BSS}$, which obtains spectral approximations with $\ensuremath{\mathcal{O}}(d\log(\epsilon \norm{\bv{A}}_2^2/\delta)/\epsilon^2)$ rows in the online setting (with larger memory requirements than the simpler sampling algorithms).
Its analysis is given in Theorem \ref{thm:bss}.
In Section \ref{sec:lower}, we show that this number of samples is in fact the best achievable, up to constant factors (Theorem \ref{thm:lower_bound}). The $\log(\epsilon \norm{\bv{A}}_2^2/\delta)$ factor is truly the cost of requiring rows to be selected in an online manner.
\section{Analysis of Sampling Schemes}
\label{sec:sampling}
We begin by bounding the sum of online $\lambda$-ridge leverage scores.
The intuition behind the proof of Theorem \ref{thm:online_scores} is that whenever we add a row with a large online leverage score to a matrix, we increase its determinant significantly, as follows from the matrix determinant lemma (Lemma \ref{lem:matdet}).
Thus we can reduce upper bounding the online leverage scores to bounding the matrix determinant.
\begin{lemma}[Matrix determinant lemma]
\label{lem:matdet}
Assume $\bv{S}$ is an invertible square matrix and $\bv{u}$ is a vector.
Then
\begin{align*}
\det(\bv{S} + \bv{u}\bv{u}^T) = (\det \bv{S})(1 + \bv{u}^T\bv{S}^{-1}\bv{u}).
\end{align*}
\end{lemma}
\Proofof{Theorem \ref{thm:online_scores}}
By Lemma \ref{lem:matdet}, we have
\begin{align*}
\det (\bv{A}_{i+1}^T\bv{A}_{i+1} + \lambda \bv{I}) &= \det (\bv{A}_i^T \bv{A}_i + \lambda \bv{I}) \cdot \left(1 + \bv{a}_{i+1}^T (\bv{A}_i^T \bv{A}_i + \lambda \bv{I})^{-1} \bv{a}_{i+1}\right)\\
&\geq \det (\bv{A}_i^T \bv{A}_i + \lambda \bv{I}) \cdot (1 + l_{i+1})\\
&\geq \det (\bv{A}_i^T \bv{A}_i + \lambda \bv{I}) \cdot e^{l_{i+1} / 2}.
\end{align*}
Hence,
\begin{align*}
\det (\bv{A}^T \bv{A} + \lambda \bv{I}) &= \det (\bv{A}_n^T \bv{A}_n + \lambda \bv{I})\\
&\geq \det (\lambda \bv{I}) \cdot e^{\sum l_i / 2}\\
&= \lambda^d e^{\sum l_i / 2}.
\end{align*}
We have $\det (\bv{A}^T \bv{A} + \lambda \bv{I}) \leq (\|\bv{A}\|_2^2 + \lambda)^d$.
Therefore
\begin{align*}
(\|\bv{A}\|_2^2+\lambda)^d &\geq \lambda^d e^{\sum l_i / 2}.
\end{align*}
Taking logarithms of both sides, we obtain
\begin{align*}
d \log (\|\bv{A}\|_2^2 + \lambda) &\geq d \log \lambda + \sum l_i / 2,\\
\sum l_i &\leq 2 d \log (1 + \|\bv{A}\|_2^2/\lambda).
\end{align*}
\QED
We now turn to analyzing the algorithm $\textsc{Online-Sample}$.
Because the samples taken by the algorithm are \emph{not} independent, we are not able to use a standard matrix Chernoff bound like the one in Theorem \ref{thm:leverage_sampling}.
However, we do know that whether we take row $i$ does not depend on later rows; thus, we are able to analyze the process as a martingale.
We will use a matrix version of the Freedman inequality given by Tropp.
\begin{theorem}[Matrix Freedman inequality \cite{tropp11}]
\label{thm:matrix_freedman}
Let $\bv{Y}_0, \bv{Y}_1, \ldots, \bv{Y}_n$ be a matrix martingale whose values are self-adjoint matrices with dimension $d$, and let $\bv{X}_1, \ldots, \bv{X}_n$ be the difference sequence.
Assume that the difference sequence is uniformly bounded in the sense that
\begin{align*}
\norm{\bv{X}_k}_2 &\leq R\mbox{ almost surely, for } k = 1, \ldots, n.
\end{align*}
Define the predictable quadratic variation process of the martingale:
\begin{align*}
\bv{W}_k := \sum_{j=1}^k \expct{j - 1}{}{\bv{X}_j^2}\mbox{, for }k = 1, \ldots, n.
\end{align*}
Then, for all $\epsilon > 0$ and $\sigma^2 > 0$,
\begin{align*}
\prob{\norm{\bv{Y}_n}_2 \geq \epsilon\mbox{ and } \norm{\bv{W}_n}_2 \leq \sigma^2} &\leq d\cdot\exp\left(-\frac{-\epsilon^2/2}{\sigma^2+R\epsilon/3}\right)
\end{align*}
\end{theorem}
We begin by showing that the output of $\textsc{Online-Sample}$ is in fact an approximation of $\bv{A}$, and that the approximate online leverage scores are lower bounded by the actual online leverage scores.
\begin{lemma}
\label{lem:online-sample}
After running $\textsc{Online-Sample}$, it holds with high probability that
\begin{align*}
(1-\epsilon)\bv{A}^T\bv{A} - \delta\bv{I} \preceq \bv{\tilde{A}}^T\bv{\tilde{A}} \preceq (1+\epsilon)\bv{A}^T\bv{A} + \delta\bv{I},
\end{align*}
and also
\begin{align*}
\tilde{l}_i \geq \bv{a}_i^T (\bv{A}^T\bv{A} + \lambda\bv{I})^{-1} \bv{a}_i
\end{align*}
for $i = 1, \ldots, n$.
\end{lemma}
\Proof
Let
\begin{align*}
\bv{u}_i &:= (\bv{A}^T\bv{A} + \lambda \bv{I})^{-1/2} \bv{a}_i.
\end{align*}
We construct a matrix martingale $\bv{Y}_0, \bv{Y}_1, \ldots, \bv{Y}_n \in \mathbb{R}^{d\times d}$ with the difference sequence $\bv{X}_1, \ldots, \bv{X}_n$.
Set $\bv {Y}_0 = \bv 0$. If $\norm{\bv{Y}_{i-1}}_2 \geq \epsilon$, we set $\bv{X}_i := \bv{0}$.
Otherwise, let
\begin{align*}
\bv{X}_i &:=
\begin{cases}
(1/p_i - 1) \bv{u}_i \bv{u}_i^T &\mbox{ if $\bv{a}_i$ is sampled in $\bv{\tilde A}$,}\\
- \bv{u}_i \bv{u}_i^T &\mbox{ otherwise.}
\end{cases}
\end{align*}
Note that in this case we have
\begin{align*}
\bv{Y}_{i-1} = (\bv{A}^T\bv{A} + \lambda \bv{I})^{-1/2}(\bv{\tilde{A}}_{i-1}^T\bv{\tilde{A}}_{i-1} - \bv{A}_{i-1}^T\bv{A}_{i-1})(\bv{A}^T\bv{A} + \lambda \bv{I})^{-1/2}.
\end{align*}
Hence, since $\norm{\bv{Y}_{i-1}}_2 < \epsilon$, we have
\begin{align*}
\tilde{l}_i &= \min((1+\epsilon)\bv{a}_i^T (\bv{\tilde{A}}_{i-1}^T \bv{\tilde{A}}_{i-1} + \lambda \bv{I})^{-1} \bv{a}_i, 1) \\
&\geq \min((1+\epsilon)\bv{a}_i^T (\bv{A}_{i-1}^T \bv{A}_{i-1} + \lambda \bv{I} + \epsilon (\bv{A}^T\bv{A} + \lambda \bv{I}))^{-1} \bv{a}_i, 1) \\
&\geq \min((1+\epsilon)\bv{a}_i^T ((1+\epsilon)(\bv{A}^T \bv{A} + \lambda \bv{I}))^{-1} \bv{a}_i, 1) \\
&= \bv{a}_i^T (\bv{A}^T\bv{A} + \lambda\bv{I})^{-1} \bv{a}_i \\
&= \bv{u}_i^T\bv{u}_i,
\end{align*}
and so
$
p_i \geq \min(c \bv{u}_i^T \bv{u}_i, 1).
$
If $p_i = 1$, then $\bv{X}_i = 0$.
Otherwise, we have $p_i \geq c \bv{u}_i^T \bv{u}_i$ and so
\begin{align*}
\|\bv{X}_i\|_2 &\leq 1 / c
\end{align*}
and
\begin{align*}
\expct{i-1}{}{\bv{X}_i^2} &\preceq p_i \cdot (1 / p_i - 1)^2 (\bv{u}_i \bv{u}_i^T)^2 + (1 - p_i) \cdot (\bv{u}_i \bv{u}_i^T)^2\\
&= (\bv{u}_i \bv{u}_i^T)^2 / p_i\\
&\preceq \bv{u}_i \bv{u}_i^T / c.
\end{align*}
And so, for the predictable quadratic variation process of the martingale $\{\bv{Y}_i\}$:
\begin{align*}
\bv{W}_i &:= \sum_{k=1}^i \expct{k-1}{}{\bv{X}_k^2},
\end{align*}
we have
\begin{align*}
\norm{\bv{W}_i}_2 \leq \left | \left |\sum_{k=1}^i \bv{u}_i \bv{u}_i^T / c \right | \right |_2 \le 1/c.
\end{align*}
Therefore by, Theorem \ref{thm:matrix_freedman}, we have
\begin{align*}
\prob{\norm{\bv{Y}_n}_2\geq \epsilon} &\leq d\cdot\exp\left(\frac{-\epsilon^2/2}{1/c + \epsilon/(3c)}\right)\\
&\leq d\cdot\exp(-c\epsilon^2/4)\\
&= 1/d.
\end{align*}
This implies that with high probability
\begin{align*}
\norm{(\bv{A}^T\bv{A} + \lambda \bv{I})^{-1/2}(\bv{\tilde{A}}^T\bv{\tilde{A}} + \lambda\bv{I})(\bv{A}^T\bv{A} + \lambda \bv{I})^{-1/2} - \bv{I}}_2 \leq \epsilon
\end{align*}
and so
\begin{align*}
(1-\epsilon)(\bv{A}^T\bv{A} + \lambda \bv{I}) &\preceq \bv{\tilde{A}}^T\bv{\tilde{A}} + \lambda\bv{I} \preceq (1+\epsilon)(\bv{A}^T\bv{A} + \lambda \bv{I}).
\end{align*}
Subtracting $\lambda\bv{I} = (\delta/\epsilon) \bv I$ from all sides, we get
\begin{align*}
(1-\epsilon)\bv{A}^T\bv{A} - \delta\bv{I} &\preceq \bv{\tilde{A}}^T\bv{\tilde{A}} \preceq (1+\epsilon)\bv{A}^T\bv{A} + \delta\bv{I}.
\end{align*}
\QED
If we set $c$ in $\textsc{Online-Sample}$ to be proportional to $\log n$ rather than $\log d$, we would be able to take a union bound over all the rows and guarantee that with high probability all the approximate online leverage scores $\tilde{l}_i$ are close to true online leverage scores $l_i$.
Thus Theorem \ref{thm:online_scores} would imply that $\textsc{Online-Sample}$ only selects $\ensuremath{\mathcal{O}}(d\log n \log(\|\bv{A}\|_2^2 / \lambda)/\epsilon^2)$ rows with high probability.
In order to remove the dependency on $n$, we have to sacrifice achieving close approximations to $l_i$ at every step.
Instead, we show that the \emph{sum} of the computed approximate online leverage scores is still small with high probability, using a custom Chernoff bound.
\begin{lemma}
\label{lem:online-sample-fast}
After running $\textsc{Online-Sample}$, it holds with high probability that
\begin{align*}
\sum_{i=1}^n \tilde{l}_i &= \ensuremath{\mathcal{O}}(d \log(\|\bv{A}\|_2^2 / \lambda)).
\end{align*}
\end{lemma}
\Proof
Define
\begin{align*}
\delta_i &:= \log \det(\bv{\tilde{A}}_i^T \bv{\tilde{A}}_i + \lambda \bv{I}) - \log \det(\bv{\tilde{A}}_{i-1}^T \bv{\tilde{A}}_{i-1} + \lambda \bv{I}).
\end{align*}
The proof closely follows the idea from the proof of Theorem \ref{thm:online_scores}.
We will aim to show that large values of $\tilde{l}_i$ correlate with large values of $\delta_i$.
However, the sum of $\delta_i$ can be bounded by the logarithm of the ratio of the determinants of $\bv{\tilde{A}}^T\bv{\tilde{A}} + \lambda\bv{I}$ and $\lambda\bv{I}$.
First, we will show that $\expct{i-1}{}{\exp(\tilde{l}_i/8 - \delta_i)}$ is always at most $1$.
We begin by an application of Lemma \ref{lem:matdet}.
\begin{align*}
\expct{i-1}{}{\exp(\tilde{l}_i/8 - \delta_i)} &= p_i \cdot e^{l_i/8} (1 + \bv{a}_i^T (\bv{\tilde{A}}_{i-1}^T \bv{\tilde{A}}_{i-1} + \lambda \bv{I})^{-1} \bv{a}_i / p_i)^{-1} + (1 - p_i) e^{l_i/8}\\
&\leq p_i \cdot (1 + l_i/4) (1 + \bv{a}_i^T (\bv{\tilde{A}}_{i-1}^T \bv{\tilde{A}}_{i-1} + \lambda \bv{I})^{-1} \bv{a}_i / p_i)^{-1} + (1 - p_i) (1 + l_i/4).
\end{align*}
If $c\tilde{l}_i < 1$, we have $p_i = c\tilde{l}_i$ and $\tilde{l}_i = (1+\epsilon)\bv{a}_i^T (\bv{\tilde{A}}_{i-1}^T \bv{\tilde{A}}_{i-1} + \lambda \bv{I})^{-1} \bv{a}_i$, and so:
\begin{align*}
\expct{i-1}{}{\exp(\tilde{l}_i/8 - \delta_i)} &\leq c\tilde{l}_i \cdot (1 + l_i/4) (1 + 1 / ((1+\epsilon)c))^{-1} + (1 - c\tilde{l}_i) (1 + l_i/4)\\
&= (1 + l_i/4) (cl_i (1 + 1/((1+\epsilon)c))^{-1} + 1 - cl_i)\\
&\leq (1 + l_i/4) (1 + cl_i (1 - 1/(4c) - 1))\\
&= (1 + l_i/4) (1 - l_i/4)\\
&\leq 1.
\end{align*}
Otherwise, we have $p_i = 1$ and so:
\begin{align*}
\expct{i-1}{}{\exp(\tilde{l}_i/8 - \delta_i)} &\leq (1 + l_i/4) (1 + \bv{a}_i^T (\bv{\tilde{A}}_{i-1}^T \bv{\tilde{A}}_{i-1} + \lambda \bv{I})^{-1} \bv{a}_i)^{-1} \\
&\leq (1 + l_i/4) (1 + l_i)^{-1} \\
&\leq 1.
\end{align*}
We will now analyze the expected product of $\exp(\tilde{l}_i/8 - \delta_i)$ over the first $k$ steps.
We group the expectation over the first $k$ steps into one over the first $k-1$ steps, aggregating the expectation for the last step by using one-way independence.
For $k \geq 1$ we have
\begin{align*}
\expct{}{}{\exp\left(\sum_{i=1}^k \tilde{l}_i/8 - \delta_i\right)} &= \expct{}{\mbox{first $k-1$ steps}}{\exp\left(\sum_{i=1}^{k-1} \tilde{l}_i/8 - \delta_i\right)\expct{k-1}{}{\exp(\tilde{l}_k/8 - \delta_k)}}\\
&\leq \expct{}{}{\exp\left(\sum_{i=1}^{k-1} \tilde{l}_i/8 - \delta_i\right)},
\end{align*}
and so by induction on $k$
\begin{align*}
\expct{}{}{\exp\left(\sum_{i=1}^n \tilde{l}_i/8 - \delta_i\right)} &\leq 1.
\end{align*}
Hence by Markov's inequality
\begin{align*}
\prob{\sum_{i=1}^n \tilde{l}_i > 8d + 8\sum_{i=1}^n \delta_i} &\leq e^{-d}.
\end{align*}
By Lemma \ref{lem:online-sample}, with high probability we have
\begin{align*}
\bv{\tilde{A}}^T\tilde{\bv{A}} + \lambda \bv{I} &\preceq (1+\epsilon)(\bv{A}^T\bv{A} + \lambda \bv{I}).
\end{align*}
We also have with high probability
\begin{align*}
\det (\bv{\tilde{A}}^T\tilde{\bv{A}} + \lambda \bv{I}) &\leq (1+\epsilon)^d(\norm{A}_2^2 + \lambda)^d,\\
\log \det (\bv{\tilde{A}}^T\tilde{\bv{A}} + \lambda \bv{I}) &\leq d(1 + \log(\norm{A}_2^2 + \lambda)).
\end{align*}
Hence, with high probability it holds that
\begin{align*}
\sum_{i=1}^n \delta_i &= \log \det (\bv{\tilde{A}}^T\tilde{\bv{A}} + \lambda \bv{I}) - d \log(\lambda)\\
&\leq d(1 + \log(\norm{A}_2^2 + \lambda) - \log(\lambda))\\
&= d(1 + \log(1 + \norm{A}_2^2 / \lambda)).
\end{align*}
And so, with high probability,
\begin{align*}
\sum_{i=1}^n \tilde{l}_i &\leq 8d + 8\sum_{i=1}^n \delta_i\\
&\leq 9d + 8d\log(1 + \norm{A}_2^2 / \lambda)\\
&= \ensuremath{\mathcal{O}}(d \log(\norm{A}_2^2 / \lambda)).
\end{align*}
\QED
\Proofof{Theorem \ref{thm:online-sample}}
The thesis follows immediately from Lemmas \ref{lem:online-sample} and \ref{lem:online-sample-fast}.
\QED
We now consider a simple modification of $\textsc{Online-Sample}$ that removes dependency on $\epsilon$ from the working memory usage with no additional cost.
\begin{figure}[ht]
\noindent
\centering
\fbox{
\begin{minipage}{6in}
\noindent $\bv{\tilde{A}} = \textsc{Slim-Sample} (\bv{A}, \epsilon, \delta)$,
where $\bv{A}$ is an $n \times d$ matrix with rows $\bv{a}_1, \ldots, \bv{a}_n$, $\epsilon \in (0, 1)$, $\delta > 0$.
\begin{enumerate}
\item Set $\lambda := \delta / \epsilon$, $c := 8\log d / \epsilon^2$.
\item Let $\bv{\tilde{A}}_0$ be a $0 \times d$ matrix.
\item Let $\tilde{l}_1, \ldots, \tilde{l}_n$ be the approximate online leverage scores computed by an independent instance of $\textsc{Online-Sample}(\bv{A}, 1/2, \delta/(2\epsilon))$.
\item For $i = 1, \ldots, n$:
\begin{enumerate}
\item Let $p_i := \min(c\tilde{l}_i, 1)$.
\item Set
$
\bv{\tilde{A}}_i :=
\begin{cases}
\begin{bmatrix} \bv{\tilde{A}}_{i-1}\\\bv{a}_i / \sqrt{p_i}\end{bmatrix} &\mbox{ with probability $p_i$,} \vspace{0.2cm}\\
\bv{\tilde{A}}_{i-1} &\mbox{ otherwise.}\\
\end{cases}
$
\end{enumerate}
\item Return $\bv{\tilde{A}} := \bv{\tilde{A}}_n$.
\end{enumerate}
\end{minipage}
}
\caption{The low-memory online sampling algorithm}
\label{fig:slim-sample}
\end{figure}
\begin{lemma}
\label{lem:slim-sample}
Let $\bv{\tilde{A}}$ be the matrix returned by $\textsc{Slim-Sample}(\bv{A}, \epsilon, \delta)$.
Then, with high probability,
\begin{align*}
(1-\epsilon)\bv{A}^T\bv{A} - \delta\bv{I} \preceq \bv{\tilde{A}}^T\bv{\tilde{A}} \preceq (1+\epsilon)\bv{A}^T\bv{A} + \delta\bv{I},
\end{align*}
and the number of rows in $\bv{\tilde{A}}$ is $\ensuremath{\mathcal{O}}(d \log d \log(\epsilon\norm{\bv{A}}_2^2/\delta)/\epsilon^2)$.
Moreover, with high probability the algorithm $\textsc{Slim-Sample}$'s memory requirement is dominated by storing $\ensuremath{\mathcal{O}}(d \log d \log(\epsilon\norm{\bv{A}}_2^2/\delta))$ rows of $\bv{A}$.
\end{lemma}
\Proof
As the samples are independent, the thesis follows from Theorem \ref{thm:leverage_sampling} and Lemmas \ref{lem:online-sample} and \ref{lem:online-sample-fast}.
\QED
|
\section{Introduction}
Arc algebras were first introduced by Khovanov in~\cite{khovanov02} to extend
his categorification of
Jones' link invariant~\cite{khovanov00} to tangles.
One of his main ingredients is a certain Frobenius
algebra of rank 2, which coincides with the cohomology ring of complex projective space.
In a follow-up paper~\cite{khovanov04}, Khovanov showed that the arc algebra $H^n$ from~\cite{khovanov02}
is closely related to the geometry of Springer varieties. Indeed,
he proved that the center of $H^n$ is isomorphic to the cohomology ring of the $(n,n)$-Springer variety.
Later on, Chen and Khovanov defined in \cite{chen-khovanov} subquotients of $H^n$ with the aim of giving an
explicit categorification of the action of tangles on tensor powers of the fundamental representation of
quantum $\mathfrak{sl}(2)$. To do so, they categorified the $n$-fold tensor power of the fundamental representation of $U_q(\mathfrak{sl}(2))$
together with its weight space decomposition.
Additionally, Khovanov's arc algebra was further studied in the
sequence of papers~\cite{brundan-stroppel2,brundan-stroppel1,brundan-stroppel3,brundan-stroppel5,brundan-stroppel4},
where most of its interesting representation-theoretic properties were revealed.
\medskip
Khovanov's arc algebra was later generalized in several directions by several authors.
It was first generalized to $\mathfrak{sl}_3$-web algebras in~\cite{mackaay-tubbenhauer-pan} and then to
$\mathfrak{sl}_n$-web algebras in~\cite{mackaay1} and further studied in~\cite{tubbenhauer1,tubbenhauer2}.
In~\cite{sartori}, a version of the arc algebra associated with $\mathfrak{gl}(1|1)$ was constructed,
motivated by a representation-theoretic
categorification of the Alexander polynomial.
In~\cite{ehrig-stroppel1,ehrig-stroppel2}, a Khovanov algebra of type $D$ was introduced,
in connection with orthosymplectic Lie algebras.
More recently, a variant of Khovanov's arc algebra based on
Blanchet's version of Khovanov homology~\cite{blanchet} was constructed in~\cite{ehrig-stroppel-tubbenhauer1}. This was extended in~\cite{ehrig-stroppel-tubbenhauer2} to
$\mathfrak{gl}_2$-arc and web algebras associated with the variants of Khovanov homology from~\cite{caprau}
and~\cite{clark-morrison-walker}.
One of the main properties of the arc/web algebras above is that, except
the ones from~\cite{ehrig-stroppel1,ehrig-stroppel2} for which it is not known, they all admit topological constructions using
cobordisms or foams.
\medskip
In~\cite{ors}, Ozsvath, Rasmussen and Szabo used an exterior version of Khovanov's original Frobenius algebra to give an odd version of Khovanov homology.
Odd Khovanov homology agrees with the even (usual) Khovanov homology from~\cite{khovanov00} modulo 2, but they differ over fields
of characteristic other than 2.
Moreover, both categorify the Jones polynomial (see for example~\cite{shumakovitch} for further properties).
Odd Khovanov homology was given a (Bar-Natan style~\cite{barnatan05}) topological set-up by Putyra in~\cite{putyra14}.
He introduced the so-called chronological cobordisms, which are cobordisms together with some extra structure related to
a height function.
\medskip
In this paper we use the set-up from~\cite{ors,putyra14}
and construct an odd version of Khovanov arc algebra from~\cite{khovanov02}.
\subsection{Sketch of the construction and main results}
The first step in the construction of an odd version of Khovanov's arc algebra is to replace the
TQFT obtained by the Frobenius algebra from~\cite{khovanov00} by the chronological TQFT from~\cite{ors,putyra14}.
As explained in~\cite{putyra14}, in order to get a well-defined category of cobordisms one has to choose
an orientation for each of the local Morse moves. It was proved in~\cite{putyra14} (and in~\cite{ors} in an algebraic set-up)
that any consistent choice of orientations gives the same link homology.
This is no longer the case if one tries to extend odd Khovanov homology to tangles.
In particular, a priori there is no reason for two different choices of orientations to result in isomorphic odd arc algebras.
\smallskip
Our first result is that we get a family of odd arc algebras indexed by all possible choices.
We denote $OH^n_C$ the odd arc algebra associated with the choice $C$ of chronological cobordisms.
As a second result, we get that for all $C$ and all $n\geq2$, the odd arc algebra $OH^n_C$ is nonassociative.
This is done in Section~\ref{sect:hnodd}.
\smallskip
In Section~\ref{sec:ospringer}, we prove an odd version of Khovanv's results from~\cite{khovanov04}.
Namely, we prove that the odd center of $OH^n_C$ is isomorphic to the odd cohomology of the $(n,n)$-Springer variety
as given by Lauda and Russell in~\cite{laudarussell14}.
In this paper they constructed an oddification of the cohomology of the Springer variety
associated to any partition, by replacing polynomial rings and symmetric functions by their odd counterparts.
\smallskip
As mentioned above, the algebra $OH^n_C$ is not associative.
But this is not too big a problem, since it is a quasialgebra in the sense of Albuquerque and Majid.
They defined in~\cite{octonions99} the notion of quasialgebra, which is a nonassociative graded algebra
with an associator given by a $3$-cocycle coming from a higher structure, that is, a monoidal category.
In Section~\ref{sec:assoc}, we introduce a grading on $OH^n$ by a groupoid and prove the quasi-associativity of $OH^n_C$,
the associator depending only on $C$.
The idea of looking at an odd version of Khovanov's arc algebra as a quasialgebra goes back to the attempts of
Putyra and Shumakovitch to extend the odd Khovanov homology to tangles.
An extended discussion over generalised quasialgebras can be found in the unpublished work of Putyra~\cite{putyrapreprint}.
We prove that the associator is a coboundary and thus admits a primitive $\tau$.
Twisting the multiplication of $OH^n_C$ by this $\tau$ defines an associative algebra which keeps the odd flavor of $OH^n_C$.
In addition, we prove that all choices of $C$ and of twist lead to isomorphic algebras.
\medskip
\paragraph{Acknowledgments} We would like the thank Krzysztof Putyra for the discussions and ideas leading to the results of Section~\ref{sec:assoc}.
We thank also Daniel Tubbenhauer for comments on a previous version of this paper.
G.N. is a Research Fellow of the Fonds de la Recherche Scientifique - FNRS, under Grant no.~1.A310.16.
P.V. was supported by the Fonds de la Recherche Scientifique - FNRS under Grant no.~J.0135.16.
\section{Reminders}
To begin, we recall the three main constructions we will use: the Khovanov arc algebra, the TQFT from odd Khovanov homology and the oddification of the cohomology of the Springer varieties.
\subsection{Khovanov's arc algebra}
\label{sec:khH}
As the construction in this paper follows Khovanov's original setup from~\cite{khovanov02}, we give below a sketch of the construction of the arc algebra $H^n$.
\paragraph{Crossingless matchings} Let $B^n$ be the set of crossingless matchings of $2n$ points, that is, all ways one can pair $2n$ points on a horizontal line by non-crossing arcs placed below this line. For $b \in B^n$, we denote by $W(b)$ the reflection of $b$ across the horizontal line and by $W(b)a$ the gluing of $W(b)$ on the top of $a \in B^n$. It is clear that $W(b)a$ is a disjoint union of circles.
For example, we have in $B^2$:
\begin{align*}
a &=\ {\Cmda}\ , & b &=\ {\Cmdb}\ , \\
W(b) &=\ {\CmdWb}\ , & W(b)a &=\ {\xy (0,-1.6)*{\Cmda}; (0,1.6)*{\CmdWb}; \endxy}\ .
\end{align*}
We also write $W(d)cW(b)a$ for the concatenation of $W(d)c$ on top of $W(b)a$, which is the disjoint union of $W(d)c$ and $W(b)a$, see for example (\ref{eq:hnpicture}).
\paragraph{Contraction cobordisms} Given a diagram $W(c)bW(b)a$, we construct a cobordism
\begin{equation}
S_{cba} : W(c)bW(b)a \rightarrow W(c)a \label{eq:Scba}
\end{equation}
by contracting the arcs of $b$ with their symmetric counterparts in $W(b)$ by saddles:
\begin{equation*}
{\xy
(0,7.5)*{};
(10,7.5)*{}
**\crv{ (0,2.5) & (10,2.5)};
(0,-7.5)*{};
(10,-7.5)*{}
**\crv{ (0,-2.5) & (10,-2.5)};
(-2.5,7.5)*{}; (12.5,7.5)*{} **\dir{--};
(-2.5,-7.5)*{}; (12.5,-7.5)*{} **\dir{--};
\endxy} \quad \xrightarrow[\text{by a saddle}]{\quad\text{arcs contraction}\quad} \quad
{\xy
(0,-7.5)*{};
(0,7.5)*{}
**\crv{ (0,-7.5) & (0,7.5)};
(10,-7.5)*{};
(10,7.5)*{}
**\crv{ (10,-7.5) & (10,7.5)};
(-2.5,7.5)*{}; (12.5,7.5)*{} **\dir{--};
(-2.5,-7.5)*{}; (12.5,-7.5)*{} **\dir{--};
\endxy}
\quad\overset{\text{rescaling}}{\simeq}\quad
{\xy
(-2.5,0)*{}; (12.5,0)*{} **\dir{--};
\endxy}
\end{equation*}
This gives a surface with one saddle point for each arc in $b$. Therefore, $S_{cba}$ has (minimal) Euler characteristic $-n$, is embedded in $\R^2 \times [0,1]$ and is unique up to isotopy. Indeed, contracting the symmetric arcs in two different orders gives rise to homeomorphic surfaces and thus, the construction does not depend on any choice. Moreover, $S_{cba}$ can be given a canonical orientation. The picture to keep in mind is:
\begin{equation} \label{eq:hnpicture}
{\xy (0,2.5)*{\Cmtb};
(-24,5)*{}; (16,5)*{} **\dir{--};
(-20,10)*{W(c)};
(0,8.25)*{\CmtWa};
(-20,0)*{bW(b)};
(0,-2.5)*{\CmtWb};
(-24,-4.5)*{}; (16,-4.5)*{} **\dir{--};
(-20,-9.5)*{a};
(0,-5.75)*{\Cmtc}; \endxy}
\quad \xrightarrow{\quad S_{cba} \quad} \quad
{\xy (0,8.35)*{\CmtWa}; (0,-5.75)*{\Cmtc};
(-12.5,-4.45)*{}; (-12.5,5)*{} **\dir{-};
(-7.5,-4.45)*{}; (-7.5,5)*{} **\dir{-};
(-2.5,-4.45)*{}; (-2.5,5)*{} **\dir{-};
(2.5,-4.45)*{}; (2.5,5)*{} **\dir{-};
(7.5,-4.45)*{}; (7.5,5)*{} **\dir{-};
(12.5,-4.45)*{}; (12.5,5)*{} **\dir{-};
(-16,5)*{}; (16,5)*{} **\dir{--};
(-16,-4.5)*{}; (16,-4.5)*{} **\dir{--};
\endxy}
\quad \simeq \quad
{\xy (0,3.05)*{\CmtWa};
(20,5)*{W(c)};
(-16,0)*{}; (20,0)*{} **\dir{--};
(20,-5)*{a};
(0,-1.55)*{\Cmtc};
\endxy}
\end{equation}
\paragraph{Frobenius algebra} Let $A := \Z[X]/(X^2)$ be the $\Z$-graded abelian group with grading given by $\deg 1 = -1$ and $\deg X = 1$. This group possesses the structure of a $\Z$-algebra when equipped with the polynomial multiplication. However, notice that this multiplication has degree $1$ and does not gives a graded algebra structure.
We turn $A$ into a Frobenius algebra by defining a trace,
\begin{align*}
\tr : A \rightarrow \Z,\qquad \tr(1) = 0,\quad \tr(X) = 1.
\end{align*}
As the trace is non-degenerate, this defines a TQFT,
$$F : 2Cob \rightarrow \Z\text{-grmod},$$
where $\Z$-grmod is the category of $\Z$-graded free $\Z$-modules with finite rank and $2Cob$ the category of oriented cobordisms between 1-manifolds, see \cite{kock03}. From now on, unless stated otherwise, we will always assume that graded means $\Z$-graded.
Thus, we get $F(W(b)a) \simeq A^{\otimes |W(b)a|}$ for $|W(b)a|$ the number of circle components in $W(b)a$. Moreover, the comultiplication map is explicitly given by
\begin{align*}
\Delta : A \rightarrow A \otimes A,\qquad \Delta(1) = X \otimes 1 + X \otimes 1,\quad \Delta(X) = X \otimes X.
\end{align*}
Applying this TQFT on the cobordism (\ref{eq:Scba}), we get a morphism,
\begin{equation}F(W(c)b) \otimes_\Z F(W(b)a) \simeq F(W(c)bW(b)a) \xrightarrow{F(S_{cba})} F(W(c)a). \label{eq:cobind}\end{equation}
This morphism has degree $n$ since the multiplication and comultiplication maps in $A$ have degree $1$ and $S_{cba}$ possesses $n$ saddle points.
\paragraph{Arc algebra} Define the graded abelian groups
\begin{align*}
H^n &:= \bigoplus_{a,b \in B^n} b(H^n)a, & b(H^n)a &:= F(W(b)a)\{n\},
\end{align*}
where the notation $\{n\}$ means that we shift the degree up by $n$.
Therefore, as the maximal number of components in $W(b)a$ is $n$, every element $x \in F(W(b)a)$ has degree $\deg x \ge -n$ and thus, $H^n$ is a $\Z_+$-graded group.
In order to define a multiplication in $H^n$, we first let the product $d(H^n)c \otimes_\Z b(H^n)a \rightarrow H^n$ be zero whenever $c \ne b$. Then, for the other cases, we define the multiplication such that the diagram
$$\xymatrix{
c(H^n)b \otimes_\Z b(H^n)a \ar[d]_{\simeq} \ar[r] & c(H^n)a \\
F(W(c)b) \otimes_\Z F(W(b)a) \{2n\} \ar[r]_-{(\ref{eq:cobind})}& F(W(c)a)\{n\} \ar[u]_{\simeq}
}$$
commutes. The associativity of the multiplication follows from the fact that $F$ is a TQFT. Moreover, the sum $\sum_{a\in B^n} 1_a$, with $1_a$ the unit in $a(H^n)a \simeq A^{\otimes n}\{n\}$, is a unit for $H^n$. All of this sums up to:
\begin{prop}\emph{(Khovanov, \cite[Proposition 1]{khovanov02})}
The structures above make $H^n$ into a $\Z_+$-graded associative unital $\Z$-algebra.
\end{prop}
\subsection{Odd Khovanov homoloy}\label{sec:oddkh}
Ozsvath, Rasmussen and Szabo constructed in \cite{ors} an odd version of Khovanov homology using some ``projective TQFT" replacing $F$ (projective meaning here that it is well-defined only up to sign). Putyra extended in \cite{putyra14} the work of Bar-Natan for Khovanov homology~\cite{barnatan05} by giving a topological framework for the odd homology: the chronological cobordisms. In addition, Putyra's work allows the construction of the odd Khovanov homology using a well-defined functor. In this subsection, we mainly follow the exposition in \cite{putyra14}.
\paragraph{Chronological cobordims} Recall that a chronological $2$-cobordism is a $2$-cobordism equipped with a chronology, that is, a Morse function with one critical point at each critical level. Moreover, at each critical point, we choose an orientation of the space of unstable directions in the gradient flow induced by the chronology. We write that choice by an arrow.
These chronological $2$-cobordisms, taken up to isotopy which preserves the orientations and the chronology, form a category with composition given by gluing. We denote it by $2ChCob$. Every chronological $2$-cobordism can be built from the six elementary chronological $2$-cobordisms:
\begin{align}\label{eq:elemchcob}
{\cobcup}
&\quad ,&
{\cobmergepos}
&\quad ,&
{\cobsplitpos}
&\quad ,&
{\cobcappos}
&\quad ,&
{\cobcapneg}
&\quad ,&
{\cobtwist}
\end{align}
which are called respectively a birth, a merge, a split, a positive death, a negative death and a twist. As we are only interested in chronological $2$-cobordisms, we will forget the prefix $2$-.
\paragraph{The odd functor} We describe the functor $OF : 2ChCob \rightarrow \Z$-grmod from \cite{ors}. Morally, objects of $2ChCob$ are disjoint unions of circles. For $S$ such a union we denote by $V(S)$ the free abelian group generated by the components of $S$ with a grading such that each generator has degree $2$. We define
$$OF(S) := \Ext^* V(S)\{-|S|\},$$
with $\Ext^* V(S)$ being the exterior algebra generated by the elemens of $V(S)$ and $|S|$ the number of components.
\medskip
We now define the functor on each of the elementary cobordisms (\ref{eq:elemchcob}). Let $S_1$ and $S_2$ be objects of $2ChCob$ with $S_2$ containing one circle more than $S_1$. For a birth of a circle from $S_1$ to $S_2$, there is a canonical inclusion $V(S_1) \subset V(S_2)$ (the new generator being the circle cupped by the birth cobordism). This induces a morphism
$$OF\left(\cobcup\right) : \Ext^* V(S_1) \xrightarrow{\subset} \Ext^* V(S_2), \quad v \mapsto 1 \wedge v.$$
Consider a merge of two circles $a_1,a_2$ in $S_2$ to a single one in $S_1$ with an arrow $a_1 \chemarrow a_2$. The arrow represent one of the two possible choices of orientation of the merge, the other being denoted $a_2 \chemarrow a_1$. There is an isomorphism of groups $V(S_1) \simeq V(S_2)/\{a_1-a_2\}$ and thus the canonical projection $V(S_2) \rightarrow V(S_2)/ \{a_1 - a_2\}$ induces a morphism
$$OF\left(\xy
(0,0)*{\cobmergepos};
(-5,-8)*{a_1};
(5.5,-8)*{a_2};
\endxy\right) : \Ext^* V(S_2) \rightarrow \Ext^* \left(V(S_2)/ \{a_1 - a_2\}\right) \simeq \Ext^* V(S_1). $$
It is not hard to see that the choice of orientation does not change the result in this case and we get
$$OF\left(\xy
(0,0)*{\cobmergepos};
(-5,-8)*{a_1};
(5.5,-8)*{a_2};
\endxy\right) =
OF\left(\xy
(0,0)*{\cobmergeneg};
(-5,-8)*{a_1};
(5.5,-8)*{a_2};
\endxy\right).$$
Now say we have a split sending $a \in S_1$ to $b_1 \chemarrow b_2$ in $S_2$. Again, there is a natural identification $V(S_2) \simeq V(S_1)/\{b_1 - b_2\}$, but now we also use the isomorphism
$$\Ext^* \left(V(S_1)/\{b_1 - b_2\}\right) \simeq (b_1 - b_2) \wedge \Ext^* V(S_1)$$
to get a morphism
$$OF\left(\xy
(0,0)*{\cobsplitpos};
(-5,8)*{b_1};
(5.5,8)*{b_2};
\endxy\right) : \Ext^* V(S_2) \simeq (b_1-b_2)\wedge \Ext^* V(S_1) \xrightarrow{\subset} \Ext^* V(S_1). $$
As a matter of fact, this morphism is easily computable by replacing the occurences of $a$ by $b_1$ (or $b_2$) and multiplying by $(b_1-b_2)$. For example, $1$ is sent to $b_1 - b_2$ and $a$ is sent to $(b_1-b_2)\wedge b_1 = b_1 \wedge b_2$. Notice also that reversing the orientation changes the sign of the morphism
$$OF\left(\xy
(0,0)*{\cobsplitpos};
(-5,8)*{b_1};
(5.5,8)*{b_2};
\endxy\right)
=
-OF\left(\xy
(0,0)*{\cobsplitneg};
(-5,8)*{b_1};
(5.5,8)*{b_2};
\endxy\right).$$
Suppose we have a positive (in other words, anticlockwise oriented) death of $a \in S_2$. We associate to it the morphism given by contraction with the dual of $a_1$
$$OF\left(\xy
(0,0)*{\cobcappos};
(0,-8)*{a_1};
\endxy\right) : \Ext^* V(S_2) \rightarrow \Ext^* V(S_1), \quad v \mapsto a_1^*(v).
$$
The negative one is given by the opposite.
Finally, the twist is given by a the permutation of the corresponding terms
$$OF\left(\xy
(0,0)*{\cobtwist};
(-5,-9)*{a_1};
(5.5,-9)*{a_2};
(-5,9)*{a_1};
(5.5,9)*{a_2};
\endxy\right) : \Ext^* V(S_2) \rightarrow \Ext^* \left(V(S_2)\right) , \quad \begin{cases}
a_1 &\mapsto a_2, \\
a_2 &\mapsto a_1, \\
a_1 \wedge a_2 &\mapsto a_2 \wedge a_1.
\end{cases} $$
\begin{rem}\label{rem:mergeorient}
Since changing the orientations of the merges does not change the result of the functor, we will ignore them in our discussion.
\end{rem}
\subsection{Odd cohomology of the Springer varieties}
First, let us recall the definition of a Springer variety.
\begin{definition}
Let $\lambda = (\lambda_1, \dots , \lambda_m)$ be a partition of $m$, $E_m$ be a complex vector space of dimension $m$ and $z_\lambda : E_m \rightarrow E_m$ be a nilpotent linear endomorphism with $|\lambda|$ nilpotent Jordan blocks of size $\lambda_1, \dots, \lambda_m$. The Springer variety for the partition $\lambda$ is
$$\mathfrak B_{\lambda} := \{\text{complete flags in $E_m$ stabilized by $z_\lambda$}\}.$$
\end{definition}
The cohomology ring of $\mathfrak B_{\lambda}$ can be computed by quotienting the polynomial ring in $m$ variables by the ideal of partially symmetric functions (see \cite{conciniprocesi} for more details). Write $(n,n)$ for the partition $\lambda = (n,n)$ of $2n$.
\begin{thm}\emph{(Khovanov, \cite[Theorem 1.1]{khovanov04})} \label{thm:isozhnhbnn}
There is an isomorphism of graded algebras
$$Z(H^n) \simeq H(\mathfrak B_{n,n}, \Z).$$
\end{thm}
Lauda and Russell constructed in \cite{laudarussell14} an oddification of the cohomology of the Springer varieties, denoted $OH(\mathfrak B_\lambda, \Z)$. Like the usual cohomology is obtained as a quotient of the polynomials by the
partially symmetric functions, they constructed $OH(\mathfrak B_\lambda, \Z)$ as a quotient of the ring $OPol_{m}$ of odd polynomials
\begin{align*}
OPol_{m} &:= \frac{\Z\langle x_1, \dots, x_{m}\rangle}{\langle x_ix_j + x_jx_i = 0 \text{ for all } i\ne j \rangle}, &\deg(x_i) &= 2,
\end{align*}
by some ideal.
Since we only need the case $m=2n$ and $\lambda = (n,n)$ for our discussion, we restrict to this case from now on.
\begin{definition}{(Lauda \& Russell, \cite{laudarussell14})}\label{def:oddcoh}
The odd cohomology of the $(n,n)$-Springer variety is the quotient
$$OH(\mathfrak{B}_{n,n}, \Z) := OPol_{2n}/OI_n$$
where $OI_n$ is the left ideal generated by the set of odd partially symmetric functions
\begin{align*}
OC_n := \left \{ \epsilon_r^S := \sum_{1\le i_1 < \dots < i_r \le 2n} x_{i_1}^S\dots x_{i_r}^S \bigg|
\parbox{15em}{$k\in \{1,2,\dots,n\}$, $|S| = n+k$,\\ \centering{$ r \in \{n-k+1, n-k, \dots,n+k\}$}}
\right\},
\end{align*}
for all $S$ ordered subset of $\{1, \dots, 2n\}$ of cardinality $n+k$ and
$$x_{i_j}^S := \begin{cases}
0, &\text{ if } i_j \notin S,\\
(-1)^{S(i_j)-1}x_{i_j}, &\text{otherwise,}
\end{cases}$$
with $S(i_j)$ the position of $i_j$ in $S$.
\end{definition}
In general, the odd cohomology of a Springer variety is only a module over the odd polynomials. However, in case $\lambda = (n,n)$, it is a graded algebra. This is due to the fact that, thanks to \cite[Lemma~3.6]{laudarussell14}, $x_i^2 \in OI_n$ for all $i$. Thus, $OI^n$ is a $2$-sided ideal. As a matter of fact, $OH(\mathfrak{B}_{n,n}, \Z)$ also possesses the structure of a superalgebra with superdegree given by dividing the degree by~$2$. Finally, by construction, the algebra $OH(\mathfrak{B}_{n,n}, \Z)$ is isomorphic to $H(\mathfrak{B}_{n,n}, \Z)$ modulo $2$.
\begin{example}
$OH(\mathfrak{B}_{2,2}, \Z)$ is given by the odd polynomials in 4 variables $x_1,x_2,x_3,x_4$ quotiented by the (not minimal) relations:
\begin{align*}
x_1 - x_2 + x_3 - x_4 &= 0, \\
-x_ix_j+x_ix_k-x_jx_k &= 0, &\forall i<j<k \in [1,4] , \\
-x_1x_2+x_1x_3-x_1x_4-x_2x_3+x_2x_4-x_3x_4 &= 0, \\
x_ix_jx_k &= 0, &\forall i<j<k \in [1,4] , \\
-x_1x_2x_3+x_1x_2x_4-x_1x_3x_4+x_2x_3x_4 &= 0, \\
x_1x_2x_3x_4 &= 0.
\end{align*}
\end{example}
\section{Odd arc algebra}
\label{sect:hnodd}
In this section, we construct an odd version of the Khovanov arc algebra $H^n$. Therefore, we will closely follow the construction from above, replacing the TQFT $F$ by the odd functor $OF$ from Section~\ref{sec:oddkh}.
First, we define for all $n\ge 0$ the following graded abelian groups,
\begin{align*}
OH^n &:= \bigoplus_{a,b \in B^n} b(OH^n)a, & b(OH^n)a &:= OF(W(b)a)\{n\},
\end{align*}
such that $OH^n$ is $\Z_+$-graded.
The first difficulty we encounter when we try to define a multiplication as in Section~\ref{sec:khH} is that we have to choose a chronology and signs for the splits. In the odd Khovanov homology from \cite{ors}, the signs are forced by the requirement that the cube of resolutions anticommutes (all possible choices leading to isomorphic cubes). However, in our case, there is no condition other than that the cobordisms must be embedded in $\R^2 \times [0,1]$. This means we have to consider all possible choices.
\paragraph{Contraction cobordisms}
For each $a,b,c \in B^n$, there is a canonical cobordism with minimal number of critical points (up to homeomorphism and embedded in $\R^2 \times [0,1]$) from the diagram $W(c)bW(b)a$ to $W(c)a$. This corbordism is given by contracting the arcs of $b$ with their symmetric counterparts in $W(b)$, as in the definition of $H^n$ in Section~\ref{sec:khH}. To be able to apply $OF$, we need to define a chronological cobordism and there are several ways to do so:
\begin{itemize}
\item We have
to choose a chronology, in other words, we have to choose an order in which we contract the symmetric arcs of $bW(b)$, taking
care of never contracting two arcs before the one surrounding them.
This is required to get an embedded surface.
\item We have to give an orientation for the critical points, especially for the splits (we do not need to orient the merges by Remark~\ref{rem:mergeorient}). We express the two possibilities by an arrow:
\begin{align*}
{\xy
(5,-6)*{a};
(0,7.5)*{};
(10,7.5)*{}
**\crv{ (0,2.5) & (10,2.5)};
(5,5.5)*{a};
(0,-7.5)*{};
(10,-7.5)*{}
**\crv{ (0,-2.5) & (10,-2.5)};
(-2.5,7.5)*{}; (12.5,7.5)*{} **\dir{--};
(-2.5,-7.5)*{}; (12.5,-7.5)*{} **\dir{--};
\endxy} &\qquad \longrightarrow \qquad
{\xy
(3,-5.5)*{b_1};
(13,-5.5)*{b_2};
(0,-7.5)*{};
(0,7.5)*{}
**\crv{ (0,-7.5) & (0,7.5)};
(5,0)*{\larrowfill{14pt}};
(10,-7.5)*{};
(10,7.5)*{}
**\crv{ (10,-7.5) & (10,7.5)};
(-2.5,7.5)*{}; (12.5,7.5)*{} **\dir{--};
(-2.5,-7.5)*{}; (12.5,-7.5)*{} **\dir{--};
\endxy}
&\text{ or }&&
{\xy
(5,-6)*{a};
(0,7.5)*{};
(10,7.5)*{}
**\crv{ (0,2.5) & (10,2.5)};
(5,5.5)*{a};
(0,-7.5)*{};
(10,-7.5)*{}
**\crv{ (0,-2.5) & (10,-2.5)};
(-2.5,7.5)*{}; (12.5,7.5)*{} **\dir{--};
(-2.5,-7.5)*{}; (12.5,-7.5)*{} **\dir{--};
\endxy} &\qquad\longrightarrow\qquad
{\xy
(3,-5.5)*{b_1};
(13,-5.5)*{b_2};
(0,-7.5)*{};
(0,7.5)*{}
**\crv{ (0,-7.5) & (0,7.5)};
(5,0)*{\rarrowfill{14pt}};
(10,-7.5)*{};
(10,7.5)*{}
**\crv{ (10,-7.5) & (10,7.5)};
(-2.5,7.5)*{}; (12.5,7.5)*{} **\dir{--};
(-2.5,-7.5)*{}; (12.5,-7.5)*{} **\dir{--};
\endxy}
\end{align*}
meaning that we split the component $a$ in two components $b_1, b_2$ with the orientation $b_2 \chemarrow b_1$ in the first case and $b_1 \chemarrow b_2$ in the second one.
\end{itemize}
\begin{rem}\label{rem:standardchoice}
There is always at least one possible choice: it suffices to go through the end points of $b$ from left to right and contracting whenever we encounter an arc which was not already contracted, then orienting the splits from left to right (i.e. putting an arrow from the component passing through the left point to the one passing through the right point).
\end{rem}
We now assume that for each triplet $a,b,c \in B^n$ we have chosen a chronological cobordism. We write it $C_{cba}$ and denote the collection all of them by $C := \{C_{cba} | a,b,c \in B^n\}$. In addition, we write $\mathcal C^n$ for the set of all possible choices of such a set $C$.
\paragraph{Multiplication} Like in the even case, we let the multiplication
$$d(OH^n)c \otimes_\Z b(OH^n)a \rightarrow \{0\} \subset OH^n$$
be zero for $c \ne b$. We define the multiplication
$c(OH^n)b \otimes_\Z b(OH^n)a \rightarrow c(OH^n)a$
using the morphism $OF(C_{cba})$. More precisely, there is a morphism
\begin{equation}
OF(W(c)b) \otimes_\Z OF(W(b)a) \rightarrow OF(W(c)bW(b)a) : (x,y) \mapsto x \wedge y \label{eq:wedge}
\end{equation}
induced by the inclusions $W(c)b \subset W(c)bW(b)a$ and $W(b)a \subset W(c)bW(b)a$. We compose it with $OF(C_{cba})$ to obtain the multiplication by making the following diagram commutes:
$$ \xymatrix{
c(OH^n)b \otimes_\Z b(OH^n)a \ar[d]_{\simeq} \ar[rrr] &&& c(OH^n)a \\
OF(W(c)b) \otimes_\Z OF(W(b)a) \ar[r]_-{(\ref{eq:wedge})} & OF(W(c)bW(b)a) \ar[rr]_-{OF(C_{cba})} && \ar[u]_{\simeq} OF(W(c)a).
}$$
This map is degree preserving thanks to the minimality hypothesis on the Euler Characteristic of $C_{cba}$ which guarantees that the degree of $OF(C_{cba})$ is $n$.
\paragraph{Unit} We write $1_a$ for the unit
in the exterior algebra $\Ext^* V(W(a)a)$ and we easily check that the sum $\sum_{a \in B^n} 1_a$ is a unit for the multiplication defined in the paragraph above. We also write ${_b1_a}$ for the unit in the exterior algebra $\Ext^* V(W(b)a)$ (notice ${_b1_a}$ is not an idempotent).
\begin{prop}\label{prop:nonassoc}
For $n \ge 2$ and any choice $C \in \mathcal C^n$, the multiplication defined above is not associative.
\end{prop}
\begin{proof}
First, suppose that $n =2$ and
let $C \in \mathcal C^2$ be an arbitrary choice of cobordisms. Consider $a,b \in B^2$ such that
\begin{align*}
a &= {\Cmda}\ , & b &= {\Cmdb}\ .
\end{align*}
Take $x = b_1 \in b(OH^n)b, y = {_b1_a} \in b(OH^n)a$ and $z = {_a1_b} \in a(OH^n)b$, where $b_1$ is the element in the exterior algebra coming from the outer circle in the diagram $W(b)b$. We write $b_2$ for the element generated by the inner circle. Then we compute $x(yz)$ as follows. The cobordism $$C_{bab} : {\xy (0,8)*{}; (0,-8)*{}; (0,4)*{\yo}; (0,-4)*{\xo}\endxy}\rightarrow \bobt$$
is given by a merge followed by a split such that the element $yz$ is
$$yz = \alpha(b_1 - b_2), $$
where $\alpha \in \{\pm 1\}$ depends on the chosen orientation for the unique split in $C_{bab}$. Then, $x(yz)$ is given by
$$C_{bbb} : {\xy (0,8)*{}; (0,-8)*{}; (0,4)*{\bobt}; (0,-4)*{\bobt}\endxy} \rightarrow \bobt, \quad x(yz) = -\alpha b_1 \wedge b_2$$
since $C_{bbb}$ is composed by two merges.
Now, we compute $(xy)z$ by
\begin{align*}
C_{bba} : {\xy (0,8)*{}; (0,-8)*{}; (0,4)*{\bobt}; (0,-4)*{\yo}\endxy} \rightarrow \yo,& \quad xy = c_1,
\intertext{with $c_1$ the element coming from the unique circle in $W(b)a$ and then}
C_{bab} : {\xy (0,8)*{}; (0,-8)*{};(0,4)*{\yo}; (0,-4)*{\xo}\endxy} \rightarrow \bobt,& \quad (xy)z = \alpha b_1 \wedge b_2.
\end{align*}
This means that for every $C \in \mathcal C^2$, $(xy)z = -x(yz)$.
To conclude the proof, we observe that this example can be extended for all $n \ge 2$ by adding the same arcs at right of $a$ and~$b$,
\begin{align*}
a_n &:= {\Cmda} {\xy(0,0)*{}; (5,0)*{}; (10,0)*{} **\crv{ (5,-2.5) & (10,-2.5)};
(2.5,0)*{}; (12.5,0)*{} **\dir{--}; \endxy} \quad\dots\quad {\xy (0,0)*{}; (5,0)*{} **\crv{ (0,-2.5) & (5,-2.5)};(-2.5,0)*{}; (7.5,0)*{} **\dir{--}; \endxy}\ , & b_n &:= {\Cmdb} {\xy(0,0)*{}; (5,0)*{}; (10,0)*{} **\crv{ (5,-2.5) & (10,-2.5)};(2.5,0)*{}; (12.5,0)*{} **\dir{--}; \endxy} \quad\dots\quad {\xy (0,0)*{}; (5,0)*{} **\crv{ (0,-2.5) & (5,-2.5)};(-2.5,0)*{}; (7.5,0)*{} **\dir{--}; \endxy}\ ,
\end{align*}
such that the exact same computation can be done for all $n \ge 2$.
\end{proof}
\begin{definition}
We denote by $OH^n_C$ the $\Z_+$-graded nonassociative, unital $\Z$-algebra given by $OH^n$ with the multiplication obtained from a $C \in \mathcal{C}^n$.
\end{definition}
\begin{rem}
We sometimes write $b(OH^n_C)a$. By this, we mean that we take the elements of the group $b(OH^n)a$, but viewed as elements in $OH^n_C$.
\end{rem}
As for each $C$ we get a family of algebras $OH^n_C$, it is legitimate to ask if on can classify them.
We give some partial answer to this question in Section~\ref{sec:assoc}.
\begin{rem}
From now on, unless otherwise specified, all assertions are valid for all $n \in \N$ and $C \in \mathcal C^n$ and we don't specify them.
\end{rem}
To ensure that $OH^n_C$ is an odd version of $H^n$, we have to check that the two algebras agree modulo $2$.
\begin{prop}
There is an isomorphism of graded algebras
$$H^n \otimes_\Z \Z/2\Z \simeq OH^n \otimes_\Z \Z/2\Z.$$
\end{prop}
\begin{proof}
The result follows directly from the construction since the functor $F$ used to define $H^n$ agrees up to sign with $OF$.
\end{proof}
It is interesting (and it will be useful) to notice that $OH^n_C$ contains a collection of exterior algebras as subalgebras.
\begin{prop} \label{prop:extalg}
There is an inclusion of graded algebras
$$\bigoplus_{a \in B^n} \Ext^* \Z^n \simeq \bigoplus_{a \in B^n} a(OH^n_C)a \subset OH^n_C.$$
\end{prop}
\begin{proof}
It is enough to notice that $OF(W(a)a) \simeq \Ext^* \Z^n$ as $W(a)a$ is a collection of $n$ circles, and to remark that the multiplication
$$a(OH^n_C)a \otimes_\Z a(OH^n_C)a \rightarrow a(OH^n_C)a$$
is the usual product in the exterior algebra. Indeed, the cobordism $C_{aaa}$ consists of $n$ merges and no split and thus gives the exterior product.
\end{proof}
\paragraph{Diagrammatic notation}
To simplify the notation, we propose a way to write the generators of $OH^n$ using diagrams. First, notice that there is an order on the components of $W(b)a$, for $a,b \in B^n$, given by reading the diagram from left to right. More precisely, for $a_1, a_2 \in W(b)a$, we say that $a_1 < a_2$ whenever $a_1$ passes through an end point of $a$ (or equivalently $b$) which is at the left of all end points contained in $a_2$.
An element $x_1 \wedge \dots \wedge x_k$ in $b(OH^n)a$ is written as the diagram $W(b)a$ where we draw the components with a plain line for each $x_i$ and a dashed one for the others. Moreover, we require that $x_1 \wedge \dots \wedge x_k$ is in the order induced by reading the diagram from left to right, otherwise we add a sign to recover this order. Thus, we get for example in $OH^2$:
\begin{align*}
\boid &= b_1 \wedge 1, & -\bobt &= b_2 \wedge b_1,\\ \\
\idab &= {_a1_b}, & \yo &= c_1,
\end{align*}
with $a_i, b_i$ and $c_i$ following the conventions from the proof of Proposition~\ref{prop:nonassoc}.
\subsection{An example : $OH^2_C$} \label{ex:oh2}
We construct explicit multiplication tables for $OH^2_C$, with $C$ the choice from Remark \ref{rem:standardchoice}.
Since the multiplication maps for $*_2(OH^2)a \otimes b(OH^2)*_1 \rightarrow 0$ and $*_2(OH^2)b \otimes a(OH^2)*_1 \rightarrow 0$ are zero for all $*_1, *_2 \in \{a,b\}$, we give the tables only for $*_2(OH^2)a \otimes a(OH^2)*_1 \rightarrow *_2(OH^2)*_1$ and $*_2(OH^2)b \otimes b(OH^2)*_1 \rightarrow *_2(OH^2)*_1$. Moreover, these tables are written with the convention:
$$\begin{tabular}{c|c|}
\cline{2-2}
& y \\\hline
\multicolumn{1}{ |c |} x & xy \\\hline
\end{tabular}$$
By direct computation we get
\begin{center}
\begin{tabular}{|c"c|c|c|c|c|c|}
\hline
$OH^2_C$ & $\idaa$ & $\aoid$ & $\idat$ & $\aoat$ & $\idab$ & $\xo$ \\
\thickhline
$\idaa$ & $\idaa$ & $\aoid$ & $\idat$ & $\aoat$ & $\idab$ & $\xo$ \\
\hline
$\aoid$ & $\aoid$ & $0$ & $\aoat$ & $0$ & $\xo$ & $0$ \\
\hline
$\idat$ & $\idat$ & $-\aoat$ & $0$ & $0$ & $\xo$ & $0$ \\
\hline
$\aoat$ & $\aoat$ & 0 & 0 & 0 & 0 & 0 \\
\hline
${\idba}$ & ${\idba}$ & $\yo$ & $\yo$ & 0 & $\idbt - \boid$ & $-\bobt$ \\
\hline
$\yo$ & $\yo$ & 0 & 0 & 0 & $- \bobt$ & 0 \\
\hline
\end{tabular}
\end{center}
and
\begin{center}
\begin{tabular}{|c"c|c|c|c|c|c|}
\hline
$OH^2_C$ & $\idbb$ & $\boid$ & $\idbt$ & $\bobt$ & $\idba$ & $\yo$ \\
\thickhline
$\idbb$ & $\idbb$ & $\boid$ & $\idbt$ & $\bobt$ & $\idba$ & $\yo$ \\
\hline
$\boid$ & $\boid$ & $0$ & $\bobt$ & $0$ & $\yo$ & $0$ \\
\hline
$\idbt$ & $\idbt$ & $-\bobt$ & $0$ & $0$ & $\yo$ & $0$ \\
\hline
$\bobt$ & $\bobt$ & 0 & 0 & 0 & 0 & 0 \\
\hline
$\idab$ & $\idab$ & $\xo$ & $\xo$ & 0 & $\aoid - \idat$ & $\aoat$ \\
\hline
$\xo$ & $\xo$ & 0 & 0 & 0 & $\aoat$ & 0 \\
\hline
\end{tabular}
\end{center}
\subsection{The odd center of $OH^n_C$}
When talking about exterior algebras (or in general superalgebras), it is common to consider the supercenter which is an extension of the center to the elements that anticommute. In the same spirit, we define the odd center for $OH^n_C$.
\begin{definition}
We define the \emph{parity} of an homogeneous element $z \in a(OH^n)b$ by
$$p(z) := \frac{\deg(z) - \deg({_a1_b})}{2} = \frac{\deg(z) - n + |W(b)a|}{2} \mod 2,$$
with $|W(b)a|$ the number of circle components in $W(b)a$.
\end{definition}
One can easily see that this number counts the factors of $z = a_1 \wedge ... \wedge a_m$, i.e.
$$p(a_1 \wedge \dots \wedge a_m) = m \mod 2.$$
\begin{definition}We call \emph{odd center} of $OH^n_C$ the subset
$$OZ(OH^n_C) := \left\{ z \in OH^n_C | zx = (-1)^{p(x)p(z)}xz, \forall x \in OH^n_C \right\}.$$
\end{definition}
\begin{rem}
The parity does not give a grading on $OH^n_C$, since there are elements $x,y$ in $OH^n_C$ such that
$$p(xy) \ne p(x) + p(y) \mod 2.$$
This means that $OH^n_C$ is not a superalgebra with respect to $p$. As a matter of fact, the parity descends to a degree on the antisymmetric subalgebra $\bigoplus_{a \in B^n} a(OH^n_C)a \subset OH^n_C$, in which the odd center lives.
\end{rem}
\begin{prop}\label{prop:inccenter}
There are inclusions $Z(OH^n_C) \subset OZ(OH^n_C) \subset \bigoplus_{a\in B^n} a(OH^n_C)a$.
\end{prop}
\begin{proof}
The second inclusion is immediate since, for every $z\in OZ(OH^n_C)$, one can decompose $z = \sum_{a,b \in B^n} {_bz_a}$ with ${_bz_a} \in b(OH^n_C)a$ and get ${_bz_a} = 1_bz1_a = z(1_b1_a) = 0$, unless~$b = a$. The first inclusion is obtained by first noticing that $Z(OH^n_C) \subset \bigoplus_{a\in B^n} a(OH^n_C)a$ by the same argument as before, and then observing that every central element has even parity.
\end{proof}
Moreover, one can check that the odd center is an \emph{associative} superalgebra with superdegree given by the parity, and is characterized by the following property.
\begin{prop}\label{prop:caractoddcenter}
An element $z = \sum_{a\in B^n} z_a$ is in $OZ(OH^n_C)$ if and only if $z_b\cdot{_b1_a} = {_b1_a}\cdot z_a$ for all $a,b \in B^n$.
\end{prop}
\begin{proof}
An element $z = \sum_{a\in B^n} z_a$ commutes with $x = \sum_{a,b \in B^n} {_bx_a}$ if and only if $z$ commutes with every~${_bx_a}$. Moreover, $z\cdot{_bx_a} = (z_b\cdot{_b1_a})\wedge{_bx_a} = (-1)^{p(z)p(x)} {_bx_a}\wedge (z_b\cdot{_b1_a})$ and we have ${_bx_a}\wedge (z_b\cdot{_b1_a}) = {_bx_a}\wedge ({_b1_a}\cdot z_a) = {_bx_a}.z$ if and only if $z_b\cdot {_b1_a} = {_b1_a}\cdot z_a$.
\end{proof}
The following result allows us to write $OZ(OH^n)$ with no ambiguity.
\begin{prop}
For all $C, C' \in \mathcal{C}^n$, there is an isomorphism of graded (super)algebras
$$OZ(OH^n_C) \simeq OZ(OH^n_{C'}).$$
\end{prop}
\begin{proof}
The condition $z_b\cdot {_b1_a} = {_b1_a}\cdot z_a$ from Proposition \ref{prop:caractoddcenter} depends only on the multiplication maps
\begin{align*}
b(OH^n)b \otimes_\Z b(OH^n)a &\rightarrow b(OH^n)a& &\text{ and } & b(OH^n)a \otimes_\Z a(OH^n)a \rightarrow b(OH^n)a.
\end{align*}
It is not hard to see that those are defined using only cobordisms without split so that they do not depend on $C$. Moreover, by Proposition~\ref{prop:extalg}, $\bigoplus_{a\in B^n} a(OH^n_C)a$ is isomorphic to a direct sum of exterior algebras and thus, does not depend on $C$.
\end{proof}
\section{The odd center of $OH^n$ and the $(n,n)$-Springer variety}\label{sec:ospringer}
We are now ready to prove one of the main results of this paper, which is to construct an explicit isomorphism between the odd cohomology of the $(n,n)$-Springer variety and the odd center of the odd Khovanov arc algebra.
\begin{thm}\label{thm:iso}
There is an isomorphism of graded (super)algebras between $OZ(OH^n)$ and $OH(\mathfrak B_{n,n}, \Z)$. Moreover, this isomorphism is given by
$$h : OH(\mathfrak{B}_{n,n}, \Z) \rightarrow OZ(OH^n),\qquad x_i \mapsto \sum_{a \in B^n} a_i,$$
where $a_i$ is generated by the circle component of $W(a)a$ passing through the $i^{th}$ end point of~$a$, counting from the left.
\end{thm}
The proof of this theorem will occupy the rest of this section and is split into four steps. Firstly, we define a morphism $h_0 : OPol_{2n} \rightarrow OZ(OH^n_C)$ and prove that $OI_n$ from Definition~\ref{def:oddcoh} lies in the kernel of this map, inducing the map $h$ on $OH(\mathfrak B_{n,n}, \Z)$. Secondly, we show that $h$ is injective using the equivalence up to sign between the odd and the even case together with Theorem~\ref{thm:isozhnhbnn}. Thirdly, we show that the ranks of the two algebras are equal using the cohomology of a geometric construction based on hypertori. Finally, we prove the theorem using all those ingredients.
\subsection*{Existence of $h$}
To construct $h$, we first define the algebra homomorphism
$$h_0 : OPol_{2n} \rightarrow OH^n_C,\qquad x_i \mapsto \sum_{a \in B^n} a_i,$$
where $a_i$ is generated by the circle component of $W(a)a$ passing through the $i^{th}$ end point of $a$, counting from the left.
It is well defined since
$$h_0(x_ix_j) = \sum_{a \in B^n} a_i \wedge a_j = -\sum_{a \in B^n} a_j \wedge a_i = -h_0(x_jx_i).$$
\begin{lem}
The image of $h_0$ lies in the odd center of $OH^n_C$
$$h_0(OPol_{2n}) \subset OZ(OH^n).$$
\end{lem}
\begin{proof}
The proof is straightforward from Proposition~\ref{prop:caractoddcenter} and the fact that for all $a,b \in B^n$, we have
$$(h_0(x_i)) {_b1_a} = b_i .{_b1_a} = {_b1_a}.a_i = {_b1_a}(h_0(x_i)),$$
since $a_i$ and $b_i$ are both sent to the component of $W(b)a$ passing through the $i^{th}$ point.
\end{proof}
Now, we want to show that $\epsilon_r^S$ is in the kernel of $h_0$ for all $S$ and $r$ as in Definition~\ref{def:oddcoh}.
This is equivalent to showing that $\epsilon_r^S$ lies in the kernel of the homomorphism
$$h_a : OPol_{2n} \rightarrow a(OH^n_C)a,\qquad x_i \mapsto a_i,$$
for all $a \in B^n$, since $h = \sum_{a \in B^n} h_a$.
For the sake of simplicity, we fix an element $a \in B^n$. We also denote by $E_{2n}$ the ordered set $\{1, \dots, 2n\}$ and we see it as the set of end points of $a$, from left to right. For $S \subset E_{2n}$, we call a pair of distinct points in $S$ which are linked by an arc in $a$, an \emph{arc} of $S$. The other points of $S$ are called \emph{free points}. For $R = \{i_1, \dots, i_r\} \subset S$, we also write
$$\epsilon_R^S := x_{i_1}^S \dots x_{i_r}^S$$
such that
$$\epsilon_r^S = \sum_{R \subset S, |R| = r} \epsilon_R^S.$$
\begin{lem}\label{lem:freepoints}
Let $S \subset E_{2n}$ be a subset with $|S| = n+k$. Then, $S$ contains at least $k$ arcs and at most $n-k$ free points.
\end{lem}
\begin{proof}
There are at most $n$ free points and we pick $n+k$ points, thus we have to pick at least $k$ arcs. Then, there remain $n+k-2k$ points which can be free.
\end{proof}
\begin{lem}\label{lem:containsarc}
If $R \subset S \subset E_{2n}$ contains an arc of $S$, then
$h_a(\epsilon_R^S) = 0.$
\end{lem}
\begin{proof}
This assertion follows from the fact that if $i$ is connected to $i'$ in $a$, then $a_i = a_{i'}$ and thus,
$h_a(x_i x_{i'}) = a_i \wedge a_{i'} = 0.$
\end{proof}
\begin{lem}\label{lem:rgen}
For all $R \subset S \subset E_{2n}$ with $|R| > n$, one has
$h_a(\epsilon_R^S) = 0.$
\end{lem}
\begin{proof}
There are at most $n$ free points in $S$, but $R$ contains at least $n+1$ points. So, $R$ contains an arc and the result follows from the preceding lemma.
\end{proof}
\begin{lem}\label{lem:arcinR}
For all $R\subset S \subset E_{2n}$ with $|S| = n+k$ and $|R| \ge n - k +1$, there exists an arc $(j,j')$ in $S$ with $j$ or $j'$ in $R$.
\end{lem}
\begin{proof}
We have to choose $n-k+1$ points in $S$, but there are at most $n-k$ free points by the Lemma~\ref{lem:freepoints}.
\end{proof}
\begin{example}\label{ex:hexists}
It is probably time to stop here a bit and look at an example that we will generalize below. So suppose
\begin{small}\begin{align*}
a =
\quad
&{\xy
(0,2.5)*{1}; (5,2.5)*{2}; (10,2.5)*{3}; (15,2.5)*{4}; (20,2.5)*{5}; (25,2.5)*{6};
(30,2.5)*{7}; (35,2.5)*{8}; (40,2.5)*{9}; (45,2.5)*{10}; (50,2.5)*{11}; (55,2.5)*{12};
(0,0)*{};
(35,0)*{}
**\crv{ (0,-15) & (35,-15)};
(5,0)*{};
(10,0)*{}
**\crv{ (5,-3) & (10,-3)};
(15,0)*{};
(30,0)*{}
**\crv{ (15,-9) & (30,-9)};
(20,0)*{};
(25,0)*{}
**\crv{ (20, -3) & (25,-3)};
(40,0)*{};
(55,0)*{}
**\crv{ (40, -9) & (55,-9)};
(45,0)*{};
(50,0)*{}
**\crv{ (45, -3) & (50,-3)};
(-1,0)*{}; (56,0)*{} **\dir{--};
\endxy}\ ,
\intertext{with $n=6$, $r=4$, $k=3$, $S = \{1,2,3,5,7,8,9,11,12\}$ and $R = \{5,8,9,11\} \subset S$. Putting a circle on the end points of $a$ which are in $S$, a bullet on the end points of $R$ and drawing the arcs that are not in $S$ by dotted lines we get}
&{\xy
(0,5)*{\scriptstyle 1}; (5,5)*{2}; (10,5)*{3}; (15,5)*{4}; (20,5)*{5}; (25,5)*{6};
(30,5)*{7}; (35,5)*{8}; (40,5)*{9}; (45,5)*{10}; (50,5)*{11}; (55,5)*{12};
(0,0)*{\circ};
(35,0)*{\bullet}
**\crv{ (0,-15) & (35,-15)};
(5,0)*{\circ};
(10,0)*{\circ}
**\crv{(5,-3) & (10,-3)};
(15,0)*{ };
(30,0)*{ \circ}
**\crv{~*=<4pt>{.} (15,-9) & (30,-9)};
(20,0)*{\bullet};
(25,0)*{}
**\crv{ ~*=<4pt>{.} (20, -3) & (25,-3)};
(40,0)*{\bullet};
(55,0)*{\circ}
**\crv{(40, -9) & (55,-9)};
(45,0)*{};
(50,0)*{\bullet}
**\crv{~*=<4pt>{.} (45, -3) & (50,-3)};
\endxy}\ .
\end{align*}\end{small}
The free points of $S$ are thus $\{5,7,11\}$. We have $\epsilon_R^S = (-x_5)(-x_8)x_9(-x_{11})$ and in $a(OH_C^n)a$ we obtain
$$
h_a(\epsilon^S_R) = (-A_4) \wedge (-A_1) \wedge A_5 \wedge (-A_6),
$$
where $A_i$ corresponds to the $i^{th}$ circle in $W(a)a$ counting from the left (see the paragraph about diagrammatic notation in Section~\ref{sect:hnodd}).
Now we look how $h_a(\epsilon^S_R)$ behaves when we modify $R$. Take $R' = \{5,8,11,12\}$, obtained from $R$ by exchanging the end points of the $5^{th}$ arc. Then we get $\epsilon_{R'}^S = (-x_5)(-x_8)(-x_{11})x_{12}$ and
$$
h_a(\epsilon^S_{R'}) = (-A_4) \wedge (-A_1) \wedge (-A_6) \wedge A_5 = -h_0(\epsilon^S_{R}).
$$
For $R'' = \{1,5,9,11\}$, obtained by replacing $8$ with $1$, we get $\epsilon_{R''}^S = x_1(-x_5)x_9(-x_{11})$ and
$$
h_a(\epsilon^S_{R''}) = A_1 \wedge (-A_4) \wedge A_5 \wedge (-A_6) = h_0(\epsilon^S_{R}).
$$
Fortunately, by replacing $9$ with $12$ in $R''$ we get an element that will cancel with $h_0(\epsilon^S_{R''})$. If we remove $11$ or $9$ from $R$ and take $7$ instead, then exchanging $8$ with $1$ will also create two elements that cancel with each other. If we take $2$ or $3$ instead of $7$, then exchanging those points will also gives $0$. In general, the number of free points in $S$ which are not in $R$ plus the number of points in $R$ that are not free, both between the endpoints of an arc, will say if exchanging those endpoints change the sign in the image of $h_a$. We can observe that for each choice of $R$ there exists a possibility to exchange two points such that they cancel with each other in the image of $h_a$. At the end, we get $h_a(\epsilon_r^S) = 0$.
\end{example}
For $x \in S \subset E_{2n}$, we write
$$F_S(x) := \begin{cases} 1, &\text{ if $x$ is free in $S$, } \\ 0, &\text{ otherwise,} \end{cases}$$
and, for $R \subset S$ and $(j,j')$ an arc of $S$ such that $j\in R$ or $j' \in R$, we define
\begin{align*}
p_{R,S}((j,j')) :&= \sum_{\substack{x \in S \setminus R \\ j < x < j'}} F_S(x) + \sum_{\substack{y \in R \\ j < y < j'}} (1-F_S(y)) \mod 2\\
&\equiv \sum_{\substack{x \in S \\ j < x < j'}} F_S(x) + \sum_{\substack{y \in R \\ j < y < j'}} 1 \mod 2.
\end{align*}
We say that a point $x$ (resp. an arc $(k,k')$) belongs to an arc $(j,j')$ if $j < x < j'$ (resp. $j<k<k'<j'$). Therefore, $p_{R,S}((j,j'))$ counts the number of free points of $S$ belonging to the arc $(j,j')$ and which are not in $R$ plus the number of points of $R$ also belonging ton $(j,j')$ and which are not free. Denote the set of all sub-arcs of $(j,j')$ with an extremity in $R$ by $]j,j'[_R$. Also, write $]j,j[^{max}_R$ for the set of all maximal sub-arcs of $(j,j')$ with an extremity in $R$, that is, the sub-arcs not belonging to any other arcs from $]j,j'[_R$.
\begin{example}
Suppose we take $R$ as in Example~\ref{ex:hexists}. Then we get $p_{R,S}((1,8)) \equiv 3 \mod 2$ and $p_{R,S}((9,12)) \equiv 2 \mod 2$.
\end{example}
\begin{lem}\label{lem:subarccounting}
If no arc of $S \subset E_{2n}$ belongs to some $R \subset S$, then for any arc $(j,j')$ in $S$ with $j$ or $j'$ in $R$ we have
$$p_{R,S}((j,j')) = \sum_{\substack{(k,k') \in ]j,j'[_R^{max}}} (p_{R,S}((k,k'))+1) + \sum_{\substack{x \in S \setminus R \\ x \notin (k,k'), \forall (k,k') \in ]j,j'[_R}} F_S(x) \mod 2, $$
with the right sum on all $x \in S \setminus R$ which are not belonging to any sub-arcs of $]j,j'[_R$.
\end{lem}
\begin{proof}
All points belonging to a maximal sub-arc $(k,k')$ from the left sum belong to $(j,j')$. Moreover, $k$ or $k'$ is a free point in $R$ and thus, has a contribution by $+1$.
\end{proof}
\begin{lem}\label{lem:evenarc}
Let $R \subset S \subset E_{2n}$ be subsets with $|S| = n + k$ and $n-k+1 \le |R| \le n$. If no arc of $S$ belongs to $R$, then there exists an arc $(j,j')$ of $S$ with $j$ or $j'$ in $R$ and such that $p_{R,S}((j,j')) = 0 \mod 2$.
\end{lem}
\begin{proof}
First, by Lemma~\ref{lem:arcinR} there is at least one arc of $S$ with an extremity in $R$. We write $L \ne \emptyset$ for the set of all such arcs. Now, we suppose by contradiction that $p_{R,S} = 1 \mod 2$ on all those arcs. By the Lemma~\ref{lem:subarccounting}, we get for all $(j,j') \in L$,
$$p_{R,S}((j,j')) = \sum_{\substack{x \in S \setminus R \\ x \notin (k,k'), \forall (k,k') \in ]j,j'[_R}}F_S(x) \mod 2,$$
with $x$ not belonging to any sub-arc $(k,k') $ of $(j,j')$ such that $k$ or $k'$ is in $R$. Since by contradiction hypothesis this sum must be equal to $1 \mod 2$, there is at least one such $x$ which is free and thus at least $|L|$ free points in $S \setminus R$. However, we know that there are at least $n-k-|L|+1$ free points in $R$ and by Lemma~\ref{lem:freepoints} there are at most $n-k$ free points in $S$, which is a contradiction.
\end{proof}
\begin{lem}\label{lem:signfromparity}
Let $R \subset S \subset E_{2n}$ be subsets and $(j,j')$ be an arc of $S$ with $j \in R$ (resp. $j' \in R$). We have
$$h_a(\epsilon_{R}^S) = (-1)^{(p_{R,S}((j,j'))+1)} h_a(\epsilon_{R'}^S),$$
with $R'$ obtained by taking $R$ where we replace $j$ by $j'$ (resp. $j'$ by $j$), respecting the order of $S$.
\end{lem}
\begin{proof}
The proof is an induction on the size of $S \cap ]j,j'[$ and $R$. If $S \cap ]j,j'[ = \emptyset$, then clearly $p_R((j,j')) = 0 \mod 2$ and we get the result since $S(j') = S(j) +1$ and thus,
$x_{j}^S = -x_{j'}^S.$
Then, to get the result in the general case, we check how the sign of the exterior product changes when we add free points and arcs to $S$ and points to $R$. We leave the details to the reader.
\end{proof}
\begin{prop}\label{prop:ocnkilled}
For all $a\in B^n$, $k \in \{1, \dots,n\}$, $r \ge n-k+1$ and all $S \subset E_{2n}$ such that $|S| = n+k$, we have
$$h_a(\epsilon_r^S) = 0.$$
\end{prop}
\begin{proof}
This result directly follows from Lemmas~\ref{lem:rgen},~\ref{lem:containsarc},~\ref{lem:evenarc}, and~\ref{lem:signfromparity}.
\end{proof}
\begin{cor}\label{cor:s}
The map $h_0$ induces a homomorphism of graded (super)algebras given by
$$h : OH(\mathfrak{B}_{n,n}, \Z) \rightarrow OZ(OH^n),\qquad x_i \mapsto \sum_{a \in B^n} a_i.$$
\end{cor}
\subsection*{Injectivity of $h$}
We will need the following result.
\begin{lem}
The algebra homomorphism induced by $h$
$$\overline h : OH(\mathfrak{B}_{n,n}, \Z) \otimes_\Z \Z/2\Z \rightarrow OZ(OH^n) \otimes_\Z \Z/2\Z$$
is an isomorphism.
\end{lem}
\begin{proof}
The proof results from the commutativity of the diagram
$$\xymatrix{
OH(\mathfrak{B}_{n,n}, \Z) \otimes_\Z \Z/2\Z \ar[r]^{\overline h} \ar[d]_{\simeq} & OZ(OH^n) \otimes_\Z \Z/2\Z \ar[d]^{\simeq} \\
H(\mathfrak{B}_{n,n}, \Z) \otimes_\Z \Z/2\Z \ar[r]_{\simeq} & Z(H^n) \otimes_\Z \Z/2\Z
}$$
which comes from the equivalence modulo $2$ between the odd and the even cases, knowing that $H(\mathfrak B_{n,n}, \Z)$ is isomorphic to $Z(H^n)$ by a morphism similar to $h$, see \cite[Section~5.3]{khovanov04} for more details.
\end{proof}
\begin{prop}\label{prop:injective}
The homomorphism $h$ is injective.
\end{prop}
\begin{proof}
From \cite[Theorem~3.8]{laudarussell14} we know that $OH(\mathfrak {B}_{n,n}, \Z)$ is a free $\Z$-module.
Hence the injectivity of $h$ follows from the lemma above and the fact that if a $\Z$-linear map between two free $\Z$-modules induces an isomorphism over $\Z/2\Z$ then it is injective.
\end{proof}
\subsection*{Computing the rank of $OZ(OH^n)$}
To show the existence of an isomorphism between $Z(H^n)$ and $H(\mathfrak B_{n,n})$, Khovanov constructed in \cite{khovanov04} a manifold $\widetilde S$ using products of $2$-spheres. We make a similar construction, but using circles instead of spheres. The inspiration from Khovanov's work should be clear. All cohomology groups and rings in this section are supposed to be taken on $\Z$.
\begin{definition}
For an $a \in B^n$, let $T_a \subset T^{2n} := \underset{2n}{\underbrace{S^1 \times \dots \times S^1}}$ be the set of all points $(x_1, \dots, x_{2n}) \in T^{2n}$ such that if $i$ is linked to $j$ by an arc of $a$, then $x_i = x_j$. We also define
$$\widetilde T := \bigcup_{a \in B^n} T_a \subset T^{2n}.$$
\end{definition}
One can notice that $T_a \simeq T^n$ as we equalize $n$ pairs of coordinates. In the same spirit, we have that $T_b \cap T_a \simeq T^{|W(b)a|}$, with $x_k = x_l$ whenever the $k^{th}$ and the $l^{th}$ end points are in the same component of $W(b)a$. As a result, $\widetilde T$ is a collection of hypertori identified to each other on certain subtori.
It is well-know that the cohomology ring of an $n$-torus is the exterior algebra generated by $n$ elements of degree $1$. If we forget the grading, then we get an isomorphism of superrings $a(OH^n_C)a \simeq H(T_a)$ and an isomorphism of abelian groups ${_b}(OH^n_C){_a} \simeq_{ab} H(T_b \cap T_a)$ for all $a,b \in B^n$. Lifting the (super)ring structure from $H(T_b \cap T_a)$, we get a (super)ring structure on ${_b}(OH^n_C){_a}$ and (super)ring morphisms
\begin{align*}
\gamma_{a; b,a} : a(OH^n_C)a \rightarrow b(OH^n_C)a,&\qquad x \mapsto {_b}1_ax, \\
\gamma_{b; b,a} : b(OH^n_C)b \rightarrow b(OH^n_C)a,&\qquad x \mapsto x{_b}{1_a}.
\end{align*}
The inclusions $T_b \cap T_a \subset T_a$ and $T_b \cap T_a \subset T_b$ induce ring morphisms on the cohomology
\begin{align*}
\psi_{a; b,a} &: H(T_a) \rightarrow H(T_b \cap T_a), \\
\psi_{b; b,a} &: H(T_b) \rightarrow H(T_b \cap T_a).
\end{align*}
\begin{lem}\label{lem:diaginclusions}
The morphisms defined above are such that the following diagram of ring morphisms commutes:
$$
\xymatrix{
H(T_b) \ar[r]^-{\psi_{b;b,a}} \ar[d]^{\simeq} & H(T_b \cap T_a) \ar[d]^{\simeq} & H(T_a) \ar[l]_-{\psi_{a;b,a}} \ar[d]^{\simeq} \\
b(OH^n_C)b \ar[r]_{\gamma_{b;b,a}} & b(OH^n_C)a & a(OH^n_C)a \ar[l]^{\gamma_{a;b,a}}.
}
$$
\end{lem}
\begin{proof}
Say $H(T_b) \simeq \Ext^* \{t_1, \dots, t_n\}$. Then the map $\psi_{b;a,b}$ identify $t_k$ with $t_l$ whenever the $k^{th}$ and the $l^{th}$ end points of $b$ are in the same component of $W(b)a$. The map $\gamma_{b;b,a}$ does exactly the same on the generators of $b(OH^n_C)b$ as $C_{bba}$ merges the corresponding components.
\end{proof}
\begin{definition}
Let $I$ and $J$ be finite sets and $A_i, B_j$ be rings for all $i \in I$ and $j \in J$. Moreover, let $\beta_{i,j} : A_i \rightarrow B_j$ be ring morphisms for some pairs $(i,j) \in I \times J$ with
$$\beta := \sum \beta_{i,j} : \prod_{i \in I} A_i \rightarrow \prod_{j \in J} B_j.$$
We define the equalizer $\Eq(\beta)$ of $\beta$ as the subring of $\prod A_i$ such that for $(a_i)_{i\in I} \in \Eq(\beta)$ we have
$$\beta_{i,j} (a_i) = \beta_{k,j} (a_k)$$
whenever $\beta_{i,j}$ and $\beta_{k,j}$ are defined.
\end{definition}
By Proposition~\ref{prop:caractoddcenter}, we get that $OZ(OH^n) = \Eq(\gamma)$ for $\gamma := \sum_{a \ne b} \gamma_{a; b,a} + \gamma_{b; b,a}$. Thus, if we define $\psi := \sum_{a \ne b} \psi_{a;b,a} + \psi_{b; b,a}$, then by Lemma~\ref{lem:diaginclusions}, we get a commutative diagram
\begin{align}
\xymatrix{
H(\widetilde T) \ar[r]^{\kappa} & \Eq(\psi) \ar[r] \ar[d]^{\simeq} & \bigoplus\limits_{a \in B^n} H(T_a) \ar[r]^-{\psi} \ar[d]^-{\simeq} & \bigoplus\limits_{a \ne b \in B^n} H(T_b \cap T_a) \ar[d]^{\simeq} \\
OZ(OH^n) \ar[r]_{\simeq} & \Eq(\gamma) \ar[r] & \bigoplus\limits_{a \in B^n} a(OH^n_C)a \ar[r]_-{\gamma} & \bigoplus\limits_{a \ne b \in B^n} b(OH^n_C)a,
}
\label{eq:tau}
\end{align}
with $\kappa$ coming from the factorization by $\Eq(\psi)$ of the map $\phi : H(\widetilde T) \rightarrow \bigoplus_{a \in B^n} H(T_a)$ induced by the inclusions $T_a \hookrightarrow \widetilde T$. This factorization exists since $\img \phi \subset \Eq(\psi)$. Our goal now is to prove that $\kappa$ is an epimorphism such that $\rank( OZ(OH^n)) \le \rank(H(\widetilde T))$.
\begin{definition}
We say that there is an arrow $a \rightarrow b$ for $a,b \in B^n$ if there exists a quadruplet $1 \le i < j < k < l \le 2n$ such that $(i,j), (k,l) \in a$ and $(i,l), (j,k) \in b$. Visually, we have
$$\xy (0,-1)*++{\Cmda} ; (-7.5,2.5)*{i} ; (-2.5,2.5)*{j} ; (2.5,2.5)*{k} ; (7.5,2.5)*{l} \endxy \longrightarrow \xy (-7.5,2.5)*{i} ; (-2.5,2.5)*{j} ; (2.5,2.5)*{k} ; (7.5,2.5)*{l} ; (0,-2.5)*++{\Cmdb} \endxy.$$
This leads to a partial order $a \prec b$ if there exists a chain $a \rightarrow a_1 \rightarrow \dots \rightarrow a_k \rightarrow b$. We extend (arbitrarily) this partial order to a total order $<$ on $B^n$.
\end{definition}
\begin{lem}
For all $a \in B_n$, we have
$$T_{<a} \cap T_a = \bigcup_{b \rightarrow a} (T_b \cap T_a)$$
with $T_{<a} := \bigcup_{b < a} T_b$.
\end{lem}
\begin{proof}
We use similar arguments as in \cite[Lemma~3.4]{khovanov04}, replacing $S$ by $T$.
\end{proof}
\begin{lem}\label{lem:celldecomp}
There exists a cellular decomposition of $T_a$ which restricts to a decomposition of $T_{<a} \cap T_a$, which itself restricts to the decomposition of $T_b \cap T_a$ for all $b \rightarrow a$. This decomposition is such that there are $\binom{n}{k}$ cells of dimension $k$ in $T_a$.
\end{lem}
\begin{proof}
We construct a similar decomposition as in \cite[Lemma~3.5]{khovanov04}. We stress the fact that the cells are not in even degree only for our case.
\end{proof}
\begin{cor}\label{cor:injdecomp}
The morphism
$$H(T_{<a} \cap T_a) \rightarrow \bigoplus_{b<a} H(T_b \cap T_a),$$
induced by the inclusions $(T_b \cap T_a) \subset (T_{<a} \cap T_a)$ is injective.
\end{cor}
We remark that $T_{\le a} = T_{<a} \cup T_a$ such that there is a Mayer-Vietoris sequence:
\begin{equation}
\xymatrix{
& \dots \ar[r] & H^{m-1}(T_{a} \cap T_{<a}) \ar[dll]_{\delta} \\
H^m(T_{a} \cup T_{<a})\ar[r] & H^m(T_{a}) \oplus H^m(T_{<a})\ar[r] & H^m(T_{a} \cap T_{<a}) \ar[dll]_\delta \\
\ar[r] H^{m+1}(T_{a} \cup T_{<a}) & \dots
}\label{eq:mv}\end{equation}
\begin{prop}\label{prop:exactseq}
The following sequence is exact:
$$\xymatrix{
H(T_{\le a}) \ar[r]^-{\phi} & \bigoplus_{b \le a} H(T_b) \ar[r]^-{\psi^-} & \bigoplus_{b < c \le a} H(T_b \cap T_c),
}$$
where $\phi$ is induced by the morphisms $T_b \hookrightarrow T_{\le a}$, and where we define
$$\psi^- := \sum_{b < c \le a} (\psi_{b,c} - \psi_{c,b}),$$
with
$$\psi_{b,c} = \psi_{b;b,c} : H(T_b) \rightarrow H(T_b \cap T_c),$$
induced by the inclusion $(T_b \cap T_c) \hookrightarrow T_b$.
\end{prop}
\begin{proof}
The proof is an induction on $a$ using Corollary~\ref{cor:injdecomp} like in \cite[Proposition~3.8]{khovanov04} with the only difference that we lose the left part $0 \rightarrow$ in our sequence.
\end{proof}
\begin{prop}\label{prop:epi}
There is an epimorphism of superrings
$$k : H(\widetilde T) \rightarrow OZ(OH^n).$$
\end{prop}
\begin{proof}
We take $a$ maximal in Proposition~\ref{prop:exactseq} giving an exact sequence
$$\xymatrix{
H(\widetilde T) \ar[r]^-{\phi} & \bigoplus_{b} H(T_b) \ar[r]^-{\psi^-} & \bigoplus_{b < c} H(T_b \cap T_c)
}$$
and we observe that by definition
$$\Eq(\psi) = \ker(\psi^-) = \img(\phi)$$
so that $\kappa : H(\widetilde T) \rightarrow \Eq(\psi)$ from diagram (\ref{eq:tau}) is surjective.
\end{proof}
Now we show that the rank of $H(\widetilde T)$ is the same as the rank of $OH(\mathfrak B_{n,n}, \Z)$. By \cite[Corollary~3.9]{laudarussell14}, we already know that $\rank(OH(\mathfrak B_{n,n}, \Z)) = \binom{2n}{n}$.
\begin{lem}\label{lem:hom}
For all $k \ge 0$, the cohomology groups
\begin{align*}
H^k(T_a \cup T_{<a}), H^k(T_a), H^k(T_{<a}) \text{ and } H^k(T_a \cap T_{<a})
\end{align*}
are free of ranks satisfying the relation
$$\rank(H^k(T_a \cup T_{<a})) = \rank(H^k(T_a)) + \rank(H^k(T_{<a})) - \rank(H^k(T_a \cap T_{<a})).$$
\end{lem}
\begin{proof}
The proof is an induction on $B^n$. We consider the Mayer-Vietoris sequence (\ref{eq:mv})
and we claim that the morphisms $H^k(T_a) \rightarrow H^k(T_a \cap T_{<a}) $ are surjective and thus, that the boundary operators $\delta$ are zero. Indeed, the decomposition from Lemma~\ref{lem:celldecomp} has the same number of $k$-cells as the rank of $H^k(T_a)$ so characteristic functions on them are independent generators for the cohomology. Seeing that the cell decomposition restricts to $T_a \cap T_{<a}$, the cohomology groups $H^k(T_a \cap T_{<a})$ are free of rank given by the cell decomposition and the morphisms are surjective.
This claim gives us an isomorphism
$$H^k(T_a \cap T_{<a}) \simeq \frac{H^k(T_a) \oplus H^k(T_{<a})}{H^k(T_a \cup T_{<a})}.$$
If $a$ is minimal, then the lemma is trivial. For the general case, we get the result by induction since $H^k(T_a)$, $H^k(T_{<a})$ and $H^k(T_a \cap T_{<a})$ are free and so is $H^k(T_a \cup T_{<a})$.
\end{proof}
\begin{prop}\label{prop:grouprank}
$H(\widetilde T)$ is a free abelian group of rank
$$\rank(H(\widetilde T)) = \binom{2n}{n}.$$
\end{prop}
\begin{proof}
We obtain a cellular partition of $\widetilde T$ by first taking the cellular decomposition of $T_{a_0}$ from Lemma~\ref{lem:celldecomp}, with $a_0 \in B^n$ the minimal element, and then by adding the cells $T_{a_m} \setminus T_{<a_m}$ for all $a_m \in B^n$ following the total order. We claim that the rank of $H(\widetilde T)$ is given by the number of cells of the partition.
Indeed, all cohomology groups are free and the relation from the Lemma~\ref{lem:hom} gives us the claim since $\rank(H(T_{a_m})) - \rank(H(T_{a_m} \cap T_{<a_m}))$ counts exactly the number of cells of $T_{a_m} \setminus T_{<a_m}$.
Finally, like in \cite{khovanov04} (and proved in \cite[Lemma~3.64]{naisse15}), the number of cells is $\binom{2n}{n}$ and this concludes the proof.
\end{proof}
\begin{cor}\label{cor:ranks}
$OZ(OH^n)$ is a free abelian group and we have
$$\rank(OZ(OH^n)) = \rank(OH(\mathfrak B_{n,n}, \Z)).$$
\end{cor}
\begin{proof}
By Proposition~\ref{prop:injective}, we have
$$\rank(OZ(OH^n)) \ge \rank(OH(\mathfrak B_{n,n}, \Z))$$
and by Propositions~\ref{prop:epi}~and~\ref{prop:grouprank}, and~\cite[Corollary~3.9]{laudarussell14}, we get
$$\rank(OZ(OH^n)) \le \rank(H(\widetilde T)) = \binom{2n}{n} = \rank(OH(\mathfrak B_{n,n}, \Z)).$$
The two inequalities together conclude the proof.
\end{proof}
\subsection*{Proof of Theorem \ref{thm:iso}}
In order to prove Theorem \ref{thm:iso}, we construct a surjective map $\pi : OH(\mathfrak B_{n,n}, \Z) \rightarrow H(\widetilde T)$ such that the following diagram commutes:
\[
\xymatrix{
OH(\mathfrak B_{n,n}, \Z) \ar[rr]^{h} \ar[rd]_\pi && OZ(OH^n) \\
&H(\widetilde T). \ar[ur]_{k} &
}
\]
Following \cite[Section~4]{khovanov04}, let $\imath^* : H(T ^{2n}) \rightarrow H(\widetilde T)$ be the homomorphism induced by the inclusion $\widetilde T \subset T^{2n}$.
\begin{lem}
The map $\imath^* : H(T ^{2n}) \twoheadrightarrow H(\widetilde T)$ is an epimorphism.
\end{lem}
\begin{proof}
The map induced on the homology $H_1(\widetilde T) \hookrightarrow H_1(T^{2n})$ is split injective, and both groups are free. Indeed the cellular decomposition of $\widetilde T$ has all 1-dimensional cells given by linearly independent diagonals in $T^{2n}$, which can be completed into a basis for $H_1(T^{2n})$. Hence, by the universal coefficient theorem the map induced on the cohomology $H^1(T^{2n}) \twoheadrightarrow H^1(\widetilde T)$ is surjective.
The same applies for all $H^k(T^{2n}) \rightarrow H^k(\widetilde T)$, with linearly independent hyperplanes of dimension $k$. Indeed cells of $T_a \setminus T_{<a}$ are given by hyperplanes with at least one generating vector linearly independent to the cells of $T_{<a}$. This concludes the proof.
\end{proof}
We write $X_i \in H(\widetilde T)$ for the image by $\imath^*$ of the generating element in the cohomology of $T^{2n}$ given by the characteristic function on the $i^{th}$ component. In other words, if we say $p_i$ is the projection $T^{2n} \rightarrow S^1$ onto the $i^{th}$ component, then $X_i = i^* \circ p_i^* (X)$, where $H(S^1) = \Ext^* \{X\}$.
Then we define the ring homomorphism $\pi_0 : OPol_{2n} \rightarrow H(\widetilde T)$ by $\pi_0(x_i) = X_i$. Clearly we have an isomorphism $H(T^{2n}) \cong \Ext^*\{p_1^*(X), \dots, p_{2n}^*(X)\}$, and so $\pi_0$ is surjective by the lemma above.
\begin{lem}
The map $k : H(\widetilde T) \rightarrow OZ(OH^n)$ is an isomorphism.
\end{lem}
\begin{proof}
It is an epimorphism between free abelian groups of the same rank.
\end{proof}
From this, we deduce the homomorphism $j^* : H(\widetilde T) \rightarrow \bigoplus_{a \in B^n} H(T_a) \cong \bigoplus_{a \in B^n} a(OH^n_C)a$, induced by the inclusions $T_a \subset \widetilde T$, is an injection.
Moreover, by construction of $\pi_0$, the diagram
\[
\xymatrix{
OPol_{2n} \ar[rr]^{h_0} \ar[rd]_{\pi_0} && \bigoplus_{a \in B^n} a(OH^n_C)a \\
&H(\widetilde T) \ar[ur]_{j^*} &
}
\]
commutes.
\begin{lem}
We have $\pi_0(OC_n) = 0$.
\end{lem}
\begin{proof}
By the commutativity of the diagram above, we get $j^* \circ \pi_0(OC_n) = h_0(OC_n)$, which is zero by Proposition~\ref{prop:ocnkilled}. This concludes the proof since $j^*$ is injective.
\end{proof}
Therefore, there is an induced map $\pi : OH(\mathfrak B_{n,n}, \Z) \rightarrow H(\widetilde T)$ given by $\pi(x_i) = X_i$. Since $\pi_0$ is surjective, so is $\pi$.
\begin{proof}[Proof of Theorem \ref{thm:iso}]
By construction, we have $h = k \circ \pi$. Since $k$ and $\pi$ are both surjective, $h$ is surjective as well. By Proposition~\ref{prop:injective}, it is also injective and thus an isomorphism.
\end{proof}
\section{Turning $OH^n_C$ into an associative algebra}\label{sec:assoc}
In this section, we show that we can twist the multiplication of $OH^n_C$, turning it into an associative $\Z[i]$-algebra. To do so, we begin by proving that $OH^n_C$ is a quasialgebra in the sense of Albuquerque-Majid \cite{octonions99}, when graded by a groupoid as in~\cite{putyrapreprint}. Finally, we give some classification results on all those algebras.
\subsection{The Putyra-Shumakovitch associator}
The material in this subsection is due to Putyra and Shumakovitch \cite{putyrashumakovitch} \footnote{And we would like to thank Krzysztof Putyra for explaining it to us. }.
\paragraph{Grading by a groupoid} A groupoid is a small category with every morphism admitting an inverse. We say that a ring $R$ is graded by a groupoid $\mathcal G$ if
\begin{align*}
R &= \bigoplus_{g \in \Hom(\mathcal G)} R_g& & \text{and}&
R_{g_1} R_{g_2} &\subset R_{g_1 \circ g_2},\end{align*}
whenever $g_1$ and $g_2$ are composable, and $R_{g_1} R_{g_2} = 0$ otherwise.
\paragraph{Arc grading}
Let $\mathcal G^n$ be the groupoid with objects given by the elements of $B^n$ and with a unique morphism $a \rightarrow b$ for all $a,b \in B^n$. By uniqueness of the morphisms, the composition is such that $a \rightarrow b \rightarrow c$ is equal to $a \rightarrow c$. We can view the morphism $a \rightarrow b$ as the diagram $W(b)a$ with the composition defined for all $a,b,c \in B^n$ by $W(c)b \circ W(b)a = W(c)a$. It is clear that $\mathcal G^n$ is a groupoid as every morphism $a \rightarrow b$ possesses an inverse $b \rightarrow a$ and $a \rightarrow a$ is the identity.
\begin{example}
We can put $\mathcal G^n$ in form of a diagram. For example, $\mathcal G^2$ can be pictured as
$$\xymatrix@H=5pc@C=0.1pc{
\ar@{.>}@(dl,ul)^{\xy (0,0)*{\CmdWa}; (0,0)*{\Cmda}; (10,0)*{} \endxy} & {\Cmda} \ar@{.>}@(ur,ul)[rrrrrrrr]^{\xy (0,2)*{\CmdWb}; (0,2)*{\Cmda} \endxy} &&&&&&&& {\Cmdb} \ar@{.>}@(dl,dr)[llllllll]^{\xy (0,-5)*{\CmdWa}; (0,-5)*{\Cmdb} \endxy} & \ar@{.>}@(ur,dr)^{\xy (0,0)*{\CmdWb}; (0,0)*{\Cmdb}; (-10,0)*{} \endxy}
}.$$
\end{example}
The decomposition
$$OH^n = \bigoplus_{a,b \in B^n} b(OH^n)a = \bigoplus_{W(b)a \in \Hom(\mathcal G^n)} b(OH^n)a $$
gives a grading of $OH^n_C$ by $\mathcal G^n$.
The integer degree, coming from the grading of $\Ext^* V(S)$, will be called the \emph{quantum degree} and written $|\_|_q$.
The degree coming from the grading of $\mathcal G^n$ will be called the \emph{arc degree} and denoted $|\_|_B$.
We get a bigrading, written $|\_|$, by the groupoid $\mathcal G^n \times \mathcal Z$, with $\mathcal Z$ being the abelian group $\Z$ viewed as a category with one abstract object $\star$, morphisms given by the integers and composition obtained by taking the sum, i.e.
$$\star \xrightarrow{z_1} \star \xrightarrow{z_2} \star = \star \xrightarrow{z_1 + z_2} \star.$$
Notice that $\Hom(\mathcal G^n \times \mathcal Z) \simeq \Hom(\mathcal G^n) \times \Z$.
\medskip
As a matter of fact, $OH^n_C$ is graded by a subgroupoid with arrows given by diagrams $W(b)a$ and a quantum degree in $2\Z$ or $2\Z+1$ depending on whether $|W(b)a| \equiv n \mod 2$ or~not.
\begin{definition}\label{def:subgroupoid}
We denote by $\mathcal G^n \times \mathcal Z_2$ the groupoid given by the same objects as $\mathcal G^n \times \mathcal Z$ but with hom-spaces defined as
\[
\hom((a, \star), (b, \star)) = \left\{ (W(b)a, n) \left| \begin{array}{ll} n \in 2\Z &\text{ if } |W(b)a| \equiv n \mod 2, \\ n \in 2\Z + 1 &\text{ otherwise.}\end{array}\right. \right\}.
\]
\end{definition}
\paragraph{Quasialgebras} Recall that we proved in Proposition~\ref{prop:nonassoc} that $OH^n_C$ is not an associative algebra. We claim that it is almost one: it is a quasialgebra in the sense of \cite{octonions99} but graded by a groupoid as in \cite{putyrapreprint}.
Before defining quasialgebras, recall that the \emph{nerve} of a category $\mathcal C$ is the simplicial set generated by its morphisms. We denote it $N(\mathcal C)$ and we denote by $N_n(\mathcal C)$ the set of compositions of $n$ morphisms in~$\mathcal C$.
\begin{definition}
A \emph{quasialgebra} $A$ is a (nonassociative) $R$-algebra graded by a groupoid $\mathcal G$ with a 3-cocycle
$$\phi : N_3(\mathcal G) \rightarrow R^*$$
where $R^* \subset R$ are the invertible elements, such that
\begin{equation}\label{eq:assoc}(xy)z = \phi(|x|,|y|,|z|)x(yz)\end{equation}
for all (homogeneous with compatible degrees) $x,y,z \in A$. We call $\phi$ the \emph{associator} of $A$.
\end{definition}
\begin{rem}\label{rem:assoc}
The condition $\phi$ being a 3-cocycle means
$$ \phi(h,k,l) \phi(g,hk,l) \phi(g,h,k) = \phi(gh,k,l)\phi(g,h,kl)$$
for all sequences
$$ e \xleftarrow{g} d \xleftarrow{h} c \xleftarrow{k} b \xleftarrow{l} a \in \mathcal G.$$
We also require $\phi(h,\id_b, l) = 1$ whenever $\id_b$ is an identity morphism.
Hence the following diagram commutes:
\begin{align} \label{diag:pentagon}
\xymatrix{
((A \otimes A ) \otimes A ) \otimes A \ar[rr]^{\phi \otimes \id} \ar[d]_{\phi} && (A\otimes (A \otimes A))\otimes A \ar[d]^{\phi} \\
(A\otimes A) \otimes (A\otimes A) \ar[dr]_{\phi} && A\otimes((A\otimes A)\otimes A) \ar[dl]^{\id\otimes\phi} \\
&A\otimes(A\otimes(A\otimes A)). &
}
\end{align}
Notice if we have a non-associative $R$-algebra $A$ graded by a groupoid $\mathcal G$ and if a $\phi : N_3(\mathcal G) \rightarrow R^*$ respecting~(\ref{eq:assoc}) exists, then its definition on $\deg A := \{ \deg(x) \in \Hom(\mathcal G) | x \in A\}$ is forced by~(\ref{eq:assoc}). Moreover, it will be a 3-cocycle on this subgroupoid since (\ref{diag:pentagon}) must commute for the multiplication in $A$ to be well-defined.
\end{rem}
We can view $\Z^* = \{\pm 1\}$ as the group $\Z/2\Z$ by the isomorphism $x \in \Z/2\Z \mapsto (-1)^x$. Because of that, we will write the associator of $OH^n_C$ as a map with codomain $\Z/2\Z$.
\begin{lem}\label{lem:chchassoc}
There exists a unique map
$$\varphi_C^{ch} : N_3(\mathcal G^n) \rightarrow \Z/2\Z$$
such that for all $a,b,c,d \in B^n$ we have
$$OF(C_{dba} \circ C_{dcb}\Id_{W(b)a}) = (-1)^{\varphi_C^{ch}(W(d)c, W(c)b, W(b)a)} OF(C_{dca} \circ \Id_{W(d)c}C_{cba}).$$
\end{lem}
\begin{proof}
The two cobordisms $C_{dba} \circ C_{dcb}\Id_{W(b)a}$ and $C_{dca} \circ \Id_{W(d)c}C_{cba}$ have the same Euler characteristic and the same source and target, and thus are homeomorphic. This means they are related by changes of chronology and orientations, which induce only potential changes of sign.
\end{proof}
\begin{lem}\label{lem:nbrsplits}
For all choices $C\in \mathcal C^n$, the cobordism $C_{cba}$ is composed by the same number of splits.
\end{lem}
\begin{proof}
This is immediate as all $C_{cba}$ have the same Euler characteristic.
\end{proof}
We define
\begin{align*}
\varphi^{com} : N_3(\mathcal G^n \times \mathcal Z) \subset (\Hom(\mathcal G^n) \times \Z)^3 &\rightarrow \Z/4\Z, \\
\left( (W(d)c, k), (W(c)b, l), (W(b)a, m) \right) &\mapsto
\left(k- n +|W(d)c|\right) s(W(c)b, W(b)a) \mod 4,
\end{align*}
where $s(W(c)b, W(b)a)$ is the number of splits coming from the cobordism $C_{cba}$, which does not depend on $C$ by Lemma~\ref{lem:nbrsplits}. If we take $x,y,z \in OH^n_C$, then
\begin{equation}\label{eq:phicom}
\varphi^{com}(|x|,|y|,|z|) = 2p(x)s(|y|_B, |z|_B).
\end{equation}
In this spirit, we define the parity $p((W(d)c, k)) := \frac{k-n+|W(d)c|}{2} \in \{0,1/2,1,3/2\}$ for any element in $\Hom(\mathcal G^n \times \mathcal Z)$. Notice that the parity gives an integer for every element in the subgroupoid $\mathcal G^n \times \mathcal Z_2$ from Definition~\ref{def:subgroupoid}.
\begin{lem}\label{lem:assoc}
The map
$$\varphi_C := \varphi_C^{ch} + \varphi^{com}/2 : N_3(\mathcal G^n \times \mathcal Z_2) \rightarrow \Z/2\Z$$
is such that
$$(xy)z = (-1)^{\varphi_C(|x|,|y|,|z|)} x(yz)$$
for all $x,y,z \in OH^n_C$.
\end{lem}
\begin{proof}
Suppose we have $x \in d(OH^n_C)c$, $y \in c(OH^n_C)b$ and $z \in b(OH^n_C)a$. We compute
\begin{align*}
xy &= S(d,c,b) \wedge x \wedge y, \\
(xy)z &= S(d,b,a) \wedge S(d,c,b) \wedge x \wedge y \wedge z, \\
yz &= S(c,b,a) \wedge y \wedge z, \\
x(yz) &= S(d,c,a) \wedge x \wedge S(c,b,a) \wedge y \wedge z,
\end{align*}
where $S(d,c,b)$ are the terms coming from the splits of the cobordism $C_{dcb}$. Notice that we abuse the notation by identifying $x,y$ with their images in $d(OH^n_C)b$ in the first line, and so on. \\
This computation means that the non-associativity comes from two phenomena:
\begin{itemize}
\item The commutation between the elements coming from the splits of the product $yz$ and the left term $x$, that is
$$S(d,c,a) \wedge x \wedge S(c,b,a) \wedge y \wedge z = (-1)^{p(x)p(S(c,b,a))} S(d,c,a) \wedge S(c,b,a) \wedge x \wedge y \wedge z.$$
By (\ref{eq:phicom}), we have $p(x)p(S(c,b,a)) = p(x)s(W(c)b,W(b)a) = \varphi^{com}(|x|,|y|,|z|)/2$.
\item The change of chronology and orientations between the cobordisms $C_{dba} \circ C_{dcb}\Id_{W(b)a}$ and $C_{dca} \circ \Id_{W(d)c}C_{cba}$, meaning that
$$S(d,b,a) \wedge S(d,c,b) = (-1)^{\varphi_C^{ch}(|x|_B,|y|_B, |z|_B)} S(d,c,a) \wedge S(c,b,a)$$
by Lemma~\ref{lem:chchassoc}.
\end{itemize}
To conclude, we have
$$(xy)z = (-1)^{\varphi_C^{ch}(|x|_B,|y|_B, |z|_B)+\varphi^{com}(|x|,|y|,|z|)/2} x(yz)$$
and this finishes the proof.
\end{proof}
\begin{lem}\label{lem:cocycle}
The map $\varphi_C$ is a $3$-cocycle. More generally, the map
$$\psi_C := 2 \varphi_C^{ch} + \varphi^{com} : N_3(\mathcal G^n \times \mathcal Z) \rightarrow \Z/4\Z,$$
is a $3$-cocycle.
\end{lem}
\begin{proof}
We mainly use Remark \ref{rem:assoc}. Take $a,b,c,d,e \in B^n$. Substituting ${_e1_d}, {_d1_c}, {_c1_b}$ and ${_b1_a}$ in~(\ref{diag:pentagon}), we get
\begin{equation}\label{eq:cocyclech}
ch_{edcb} + ch_{edba} + ch_{dcba} = s_{edc}s_{cba} + ch_{ecba} + ch_{edca}
\end{equation}
where $s_{cba} = s(W(c)b, W(b)a)$, $ch_{dcba} = \phi_C^{ch}(W(d)c,W(c)b,W(b)a)$, and so on. Now suppose we have a sequence $(e,\star) \xleftarrow{g} (d,\star) \xleftarrow{h} (c, \star) \xleftarrow{k} (b,\star) \xleftarrow{l} (a,\star) \in \mathcal G^n \times \mathcal Z$. We compute
\begin{align*}
\psi_C(h,k,l) + \psi_C(g,hk,l) + \psi_C(g,h,k) &=\\
2(p(h)s_{cba} &+ ch_{dcba} + p(g)s_{dba} + ch_{edba} + p(g)s_{dcb} + ch_{edcb}),
\intertext{and}
\psi_C(gh,k,l) + \psi_C(g,h,kl) &= 2\left(p(gh)s_{cba} + ch_{ecba} + p(g)s_{dca} + ch_{edca}\right).
\end{align*}
It is easy to see that $p(gh) = p(g) + p(h) + s_{edc}$ and
$$s_{dba}+s_{dcb} = s_{cba}+s_{dca} = \#\text{splits in the cobordism } W(d)cW(c)bW(b)a \rightarrow W(d)a,$$
such that by (\ref{eq:cocyclech}) we get
$$\psi_C(h,k,l) + \psi_C(g,hk,l) + \psi_C(g,h,k) = \psi_C(gh,k,l) + \psi_C(g,h,kl),$$
which concludes the proof for $\psi_C$. We get the claim for $\varphi_C$ by seeing that $\varphi_C = \psi_C |_{\mathcal G \times \mathcal Z_2}/2$.
\end{proof}
\begin{thm}
The nonassociative ring $OH^n_C$ is a quasialgebra with associator $\varphi_C$.
\end{thm}
\begin{proof}
This is an immediate consequence from Lemmas~\ref{lem:assoc}~and~\ref{lem:cocycle}.
\end{proof}
We call $\varphi_C$ the Putyra-Shumakovitch associator.
\subsection{Twisting $OH^n_C$}
\paragraph{Twisted multiplication}
The idea of twisting a $\mathcal G$-graded $R$-algebra $A$ by a map
$$\tau : N_2(\mathcal G) \rightarrow R^*$$
is to define a new algebra $A_\tau$ by the same elements as $A$, but with a multiplication given by
$$A_\tau \otimes_R A_\tau \rightarrow A_\tau, \quad (x,y) \mapsto x *_{\tau} y := \tau(|x|,|y|) xy$$
for all $x,y \in A_\tau$, and where $xy$ is the product in $A$.
\begin{prop}\label{prop:cobtwist}
Let $A$ be a quasialgebra graded by $\mathcal G$ with associator $\phi$. If $\phi$ is a coboundary, then there exist a twist $\tau$ such that $A_\tau$ is associative.
\end{prop}
\begin{proof}
By definition of coboundary, there exists a map
$$\tau : N_2(\mathcal G) \rightarrow R^*$$
such that
\begin{equation}\label{eq:2cocycle}\phi(g,h,k) = \tau(g,h)\tau(g,hk)^{-1}\tau(gh,k)\tau(h,k)^{-1}\end{equation}
for all sequence $ d \xleftarrow{g} c \xleftarrow{h} b \xleftarrow{k} a \in \mathcal G$. Let $A_\tau$ be the twisting of $A$ by this $\tau$. Then we have
\begin{align*}
(x *_\tau y)*_\tau z &= \tau(|x|,|y|)\tau(|xy|,|z|) (xy)z, \\
x*_\tau (y*_\tau z) &= \tau(|x|,|yz|)\tau(|y|,|z|) x(yz),
\end{align*}
for all $x,y,z \in A_\tau$ and thus, by (\ref{eq:assoc}) and (\ref{eq:2cocycle}), we conclude that $A_\tau$ is associative.
\end{proof}
The geometric realization of the nerve of a category $\mathcal C$, denoted $|N(\mathcal C)|$, is a topological space constructed by gluing simplexes respecting the simplicial structure of the nerve.
\begin{lem}\label{lem:geomrel}
The geometric realization of $N(\mathcal G^n)$ is a simplex of dimension $C_n-1$, for $C_n$ the $n^{th}$ Catalan number:
$$|N(\mathcal G^n)| \simeq \Delta^{(C_n-1)}.$$
\end{lem}
\begin{proof}
The proof is immediate from the fact that $B^n$ has cardinality $C_n$ and $\mathcal G^n$ possesses one unique morphism between each pair of objects.
\end{proof}
\begin{lem}\label{lem:coh}
The cohomology groups of $\mathcal{G}^n \times \mathcal{Z}$ are
\begin{align*}
H^0(N(\mathcal{G}^n \times \mathcal{Z}), \Z/4\Z) &\simeq \Z/4\Z, \\
H^1(N(\mathcal{G}^n \times \mathcal{Z}), \Z/4\Z) &\simeq \Z/4\Z , \\
H^{\ge 2}(N(\mathcal{G}^n \times \mathcal{Z}), \Z/4\Z) &\simeq 0.
\end{align*}
\end{lem}
\begin{proof}
First, by Lemma~\ref{lem:geomrel}, we get that $|N(\mathcal G^n \times \mathcal Z)| \simeq \Delta^m \times S^1$ for $m = C_n -1$. By the K\"unneth formula, we get
$$H^k(N(\mathcal G^n \times \mathcal Z), \Z/4\Z) \simeq \bigoplus_{i+j=k} H^i(\Delta^{m}, \Z/4\Z) \otimes_{\Z/4\Z} H^j( S^1, \Z/4\Z),$$
which proves the claim.
\end{proof}
Some technicalities still remain to be solved, before being able to apply Proposition~\ref{prop:cobtwist} to $OH^n_C$: we do not know the cohomology of $\mathcal G^n \times \mathcal Z_2$ and thus, we are not able to show that the Putyra-Shumakovitch associator is a coboundary. Therefore, we work with $\psi_C$, which has an image in $\Z/4\Z$ and gives square roots of $-1$. Hence, we must consider the extended version $OH^n_C \otimes_\Z \Z[i]$ to the Gaussian integers, with $\Z[i]^* = \{1,i,-1,-i\} \simeq \Z/4\Z$. By Lemmas~\ref{lem:assoc}~and~\ref{lem:cocycle}, $OH^n_C \otimes_\Z \Z[i]$ is a quasialgebra graded by $\mathcal G^n \times \mathcal Z$ with $\psi_C$ as associator.
\begin{thm}
For all $C \in \mathcal C^n$ there exists a map $\tau_C :N_2(\mathcal G^2 \times \mathcal Z) \rightarrow \Z/4\Z$ such that the twisted algebra $ \left(OH^n_C \otimes_\Z \Z[i] \right)_{\tau_C}$ is associative.
\end{thm}
\begin{proof}
By Lemma \ref{lem:coh}, we have
$$H^3(N(\mathcal{G}^n \times \mathcal{Z}), \Z/4\Z) \simeq 0,$$
and thus every $3$-cocycle is a coboundary. In particular, the associator $\psi_C$ is a coboundary and we can apply Proposition \ref{prop:cobtwist}.
\end{proof}
\begin{rem}
Notice that the twist is not necessarily unique and thus we get potentially a family of associative algebras for each $C \in \mathcal C^n$.
\end{rem}
\begin{example}\label{ex:twistedoh2}
We construct an explicit $\left(OH^2_C\right)_{\tau_C}$ based on the choice $C$ from Remark~\ref{rem:standardchoice}.
We twist this algebra with
\begin{align*}
\tau_C\left((\aoat, 1+4k), (\yo,*)\right) &= i,& \tau_C\left((\yo,2+4k), (\xo,*)\right) = &i, \\
\tau_C\left((\aoat, 2+4k), (\yo,*)\right) &= -1,& \tau_C\left((\yo,3+4k), (\xo,*)\right) = &-1, \\
\tau_C\left((\aoat, 3+4k), (\yo,*)\right) &= -i,& \tau_C\left((\yo,0+4k), (\xo,*)\right) = &-i,
\end{align*}
for every $k \in \Z$ and $\tau_C = 1$ everywhere else. In short, we get the multiplication table:
\begin{center}
\begin{tabular}{|c"c|c|c|c|c|c|}
\hline
$\left(OH^2_C\right)_{\tau_C}$ & $\idaa$ & $\aoid$ & $\idat$ & $\aoat$ & $\idab$ & $\xo$ \\
\thickhline
$\idaa$ & $\idaa$ & $\aoid$ & $\idat$ & $\aoat$ & $\idab$ & $\xo$ \\
\hline
$\aoid$ & $\aoid$ & $0$ & $\aoat$ & $0$ & $\color{red}- \xo$ & $0$ \\
\hline
$\idat$ & $\idat$ & $-\aoat$ & $0$ & $0$ & $\color{red}- \xo$ & $0$ \\
\hline
$\aoat$ & $\aoat$ & 0 & 0 & 0 & 0 & 0 \\
\hline
${\idba}$ & ${\idba}$ & $\yo$ & $\yo$ & 0 & $\idbt - \boid$ & $-\bobt$ \\
\hline
$\yo$ & $\yo$ & 0 & 0 & 0 & $\color{red} \bobt$ & 0 \\
\hline
\end{tabular}
\end{center}
for $a$ and the one for $b$ stays the same. An exhaustive computation (which can easily be done by computer) confirms that $d\tau_C$ gives the associator in this case.
\end{example}
\begin{rem}
For this example, the twisting in $OH^n_C$ results in integer coefficients which is not surprising since the geometric realization of $\mathcal G^2 \times \mathcal Z_2$ has dimension 2 and thus there exists a twist for the associator $\varphi_C$ in this case. In general, this is not true and we lose the property that the algebra agrees modulo 2 with $H^n \otimes_\Z \Z[i]$.
\end{rem}
\subsection{Classification results}
For now, we have a family of quasialgebras $\{OH^n_C\}$ indexed by $\mathcal C^n$ and a family of associative algebras indexed by $\mathcal C^n$ and the twists. In this section, we partially classify these families.
\begin{prop}\label{prop:classodd}
Let $C, C'$ be two choices in $\mathcal C^n$ and $\varphi_C,\varphi_{C'}$ be respectively the associators of $OH^n_C$ and $OH^n_{C'}$. If $\varphi_C = \varphi_{C'}$, then the two quasialgebras are isomorphic, $OH^n_C \simeq OH^n_{C'}$.
\end{prop}
\begin{proof}
Seeing that $C_{cba}$ and $C'_{cba}$ are related by a change of chronology and orientations, there is a map $\eta_{C,C'} :N_2(\mathcal G^n) \rightarrow \Z/2\Z$ such that
$$OF(C_{cba}) = (-1)^{\eta_{C,C'}(W(c)b, W(b)a)} OF(C'_{cba}),$$
for all $a,b,c \in B^n$. Writing $*_C$ for the product in $OH^n_C$ and $*_{C'}$ for the one in $OH^n_{C'}$, this means that $x *_C y = (-1)^{\eta_{C,C'}(|x|_B, |y|_B)} x *_{C'}y$ for all $x,y \in OH^n$. We compute
\begin{align*}
OF(C_{dba} \circ C_{dcb}\Id_{W(b)a}) &= (-1)^{\eta_{C,C'}(W(d)b,W(b)a) + \eta_{C,C'}(W(d)c,W(c)b)} OF(C'_{dba} \circ C'_{dcb}\Id_{W(b)a}), \\
OF(C_{dca} \circ \Id_{W(d)c}C_{cba}) &= (-1)^{\eta_{C,C'}(W(d)c,W(c)a) + \eta_{C,C'}(W(c)b,W(b)a)} OF(C'_{dca} \circ \Id_{W(d)c}C'_{cba}),
\end{align*}
such that by definition of the associators, we get
\begin{equation}\label{eq:deta}
d\eta_{C,C'} = \phi_{C'}^{ch} - \phi_C^{ch},
\end{equation}
and thus, as $\phi_C^{ch} = \phi_{C'}^{ch}$ by hypothesis, $\eta_{C,C'}$ is a $2$-cocycle.
By Lemma~\ref{lem:geomrel}, we know that
$$H^2(N(\mathcal G^n), \Z/2\Z) \simeq 0,$$
and consequently, $\eta_{C,C'}$ is a coboundary. Hence, $\eta_{C,C'} = d\lambda_{C,C'}$ for a $\lambda_{C,C'} : N_1(\mathcal G^n) \rightarrow \Z/2\Z$. This means that the morphism
$$x \mapsto \lambda_{C,C'}(|x|_B) x : OH^n_C \rightarrow OH^n_{C'}, $$
is an isomorphism of quasialgebras.
\end{proof}
For the associative twisted algebra, the case is much simpler and all algebras are isomorphic. This means that the choice of $C$ and of twist $\tau_C$ does not matter in the end.
\begin{prop}
For all choices $C, C' \in \mathcal C^n$ and all choices of twists $\tau_C, \tau_{C'}$, there is an isomorphism
$\left(OH^n_C \otimes_\Z \Z[i]\right)_{\tau_C} \simeq\left(OH^n_{C'}\otimes_\Z \Z[i]\right)_{\tau_{C'}}.$
\end{prop}
\begin{proof}
For all $x, y \in OH^n$, we have
$$x *_{\tau_C} y = i ^{\tau_C(|x|,|y|)+2\eta_{C,C'}(|x|_B,|y|_B)-\tau_{C'}(|x|,|y|)} x *_{\tau_{C'}} y,$$
where $*_{\tau_C}$ is the product in $\left(OH^n_C \otimes_\Z \Z[i]\right)_{\tau_C}$, and we write
$$\theta_{C,C'} := \tau_C+2\eta_{C,C'}-\tau_{C'} : N_2(\mathcal G^n \times \mathcal Z) \rightarrow \Z/4\Z.$$
We compute $d\theta_{C,C'} = \psi_C + 2d\eta_{C,C'} - \psi_{C'} \overset{(\ref{eq:deta})}{=} 0$, and by using similar arguments as in the proof of Proposition~\ref{prop:classodd}, we conclude that the two algebras are isomorphic.
\end{proof}
\begin{cor}
The associative twisted algebra is uniquely determined up to isomorphism. We write it ${OH}^n_{\tau}$.
\end{cor}
\begin{rem}
Finding a twist is not an easy task, which can entail some serious difficulties for the construction of~${OH}^n_{\tau}$.
\end{rem}
\begin{prop}
The following three arc algebras are not isomorphic
$$H^n \otimes \Z[i] \not\simeq {OH}^n_{\tau} \not\simeq OH^n_C \otimes \Z[i].$$
\end{prop}
\begin{proof}
The first two are associative algebras as opposed to the last one. Let us begin with the case $n=2$. We know that the center of $H^2$ has graded rank $1+3q^2+2q^4$. Howewer, by a similar argument as in Proposition~\ref{prop:inccenter}, we have
$$Z({OH}^2_{\tau}) \subset a(OH^2)a \oplus b(OH^2)b.$$
But ${OH}^2_{\tau}$ behaves like an exterior algebra on this subset, implying that elements anticommute and thus the graded rank of the center is $0$ in degree $2$. Therefore, $H^2$ and ${OH}^2_{\tau}$ have non-isomorphic centers and thus cannot be isomorphic as algebras.
\medskip
This can be extended to all $n \ge 2$. Suppose $x,y \in a({OH}^n_{\tau})a$ with $|x|_q=|y|_q = 1$, their products are given by
\begin{align*}
(x,y) &\mapsto \tau(|x|,|y|) x \wedge y,\\
(y,x) &\mapsto \tau(|y|,|x|) y \wedge x,
\end{align*}
since $OF(C_{aaa})$ is the product in the exterior algebra $\Ext^* V(W(a)a)$. However $|x|=|y|$, implying $\tau(|x|,|y|) = \tau(|y|,|x|)$ and thus, $xy = -yx$.
\end{proof}
\begin{prop}
The odd center of ${OH}^2_{\tau}$ is not isomorphic to $OH(\mathfrak B_{2,2}, \Z[i])$.
\end{prop}
\begin{proof}
It is not hard to compute that the odd center of ${OH}^2_{\tau}$ is generated by the elements
$$OZ({OH}^2_{\tau}) = \left\langle \idaa + \idbb, \aoid - \idat, \boid - \idbt, \aoat, \bobt\right\rangle$$
and thus has graded rank $1 + 2q^2 + 2q^4$. However, we know that $OH(\mathfrak B_{2,2}, \Z[i])$ has graded rank $1 + 3q^2 + 2q^4$.
\end{proof}
Despite this result, it is easy to show that ${OH}^n_{\tau}$ contains a subalgebra which is isomorphic to $OH(\mathfrak B_{n,n}, \Z[i])$
by constructing an injective map, say $\tilde h$, similar to $h_0$ from Section~\ref{sec:ospringer}.
As $\bigoplus_a a(OH_\tau^n)a$ is isomorphic to an exterior algebra $\tilde h$ will be well defined.
Moreover the different arc algebras being isomorphic modulo $2$, it is injective by the same
reason as in Proposition~\ref{prop:injective}.
\section{Perspectives}
One natural application of the work in this paper could be the construction of odd Khovanov homology for tangles
(Putyra-Shumako\-vitch's work in progress using the structure of quasialgebras~\cite{putyrashumakovitch}).
The fact that the twist $\tau$ is not explicit may cause several technical difficulties in defining
(and working with) $(OH^n_{\tau},OH^n_{\tau})$-bimodules.
Another possibility consists of working with quasibimodules, that is bimodules with the associativity axiom given by an associator, as in~\cite{putyrapreprint}.
With such a theory at hand, it seems plausible that the braid group action on
the category of complexes of $(OH^n_{C},OH^n_{C})$-quasibimodules up to homotopy descends
to an action of the $(-1)$-Hecke algebra from~\cite[Section 4]{laudarussell14} on its (odd) center,
paralleling the even case (see~\cite[Section 5.3]{khovanov02}).
\smallskip
The fact that the twisted odd arc algebra $OH^n_{\tau}$
is defined over the Gaussian integers was for technical reasons.
One question that we leave open is to find whether it is possible to twist $OH^n_C$ over the integers.
\smallskip
The construction in this paper shares several features with Ehrig-Stroppel's Khovanov arc algebra of type $D$
from~\cite{ehrig-stroppel1}.
It would be interesting to find a connection between these two arc algebras.
\smallskip
In~\cite{brundan-stroppel3}, an action of the 2-Kac-Moody of Rouquier~\cite{rouquier}
(and therefore of Khovanov-Lauda's~\cite{khovanov-lauda1}) on Khovanov's arc algebras was constructed.
The results of Rouquier on strong categorical actions~\cite{rouquier},
together with the fact that our associative arc algebra is
not isomorphic to Khovanov's, imply that the 2-Kac-Moody algebra does not act on it.
It seems plausible to expect that the odd arc algebra admits an action of an algebra akin to
Brundan and Kleshchev's Hecke-Clifford superalgebra
from~\cite{brundan-kleshchev}, which could be seen as a super counterpart of the
cyclotomic KLR algebra.
\smallskip
Another challenging problem we would like to mention is to find the representation-theoretic
context (category $\mathcal{O}$) for the odd arc algebras.
The analogy with~\cite{ehrig-stroppel1} and~\cite{ehrig-stroppel2}, taken together with~\cite{ellis-lauda}
and the results in~\cite[Section 5.1]{ehrig-stroppel-tubbenhauer2} might
suggest a connection to category $\mathcal{O}$ for Lie superalgebras.
\bibliographystyle{habbrv}
|
\section{Introduction}
A significant limitation for underwater robots is their ability to maneuver precisely during complex sensing and tracking tasks. Next generation vehicles require thrusters that can overcome this problem and efficiently provide precise maneuverability at low speeds. Such maneuverability, in turn, requires thrusters that can deliver specific impulses rapidly and efficiently to the vehicle. In these settings, pulsed jets are increasingly used to augment vehicle maneuverability and improve efficiency at low-speeds. Pulsed jets offer many benefits over traditional propeller propulsion, including more precise impulse delivery, more rapid impulse delivery, and the ability to propel a vehicle using zero mass flux \cite{Mohseni:2006eo}.
Pulsed jets also provide opportunities for efficient individual and swarm propulsion as observed in animals such as squid, jellyfish, siphonophores, and salps. For salps and siphonophores uniquely, individual animals form chains where each member can independently control its jetting behavior. By synchronizing jet strength and timing, colonies of these animals can execute precise maneuvers and can reach high speeds efficiently \cite{Bone:1983uo,Madin:1990iy, Sutherland:2010cy, Costello:2015fn}. Taking inspiration from nature, pulsed jets may prove to be an important technology for the development of scalable marine robotic swarms. However, in order for pulsed jets to be used more widely for underwater vehicles, the implications of jet hydrodynamics on vehicle design must be understood.
Individual pulsed jets have become the most widely explored form of underwater jet propulsion since it was shown that a pulsed jet will generate more thrust than an equivalent steady jet \cite{Weihs:1977ty, Krueger:2001vo, Krueger:2003cs}. This phenomenon can be explained by breaking the thrust production process into two contributions - one associated with the inertial momentum transfer of a steady jet, and one associated with a nozzle ``overpressure'' generated by the unsteady starting flow as the pulsed jet rolls into a vortex ring. For pulses shorter than the critical vortex formation time, the nozzle over-pressure contributes as much as half of the total impulse generated by the pulsed jet, suggesting it is more beneficial to ``chop'' the flow into short pulses than to eject a steady jet of long duration \cite{Krueger:2001vo, Krueger:2003cs}. Complementary descriptions of thrust production in pulsed jets can leverage concepts of vortex added mass \cite{Weihs:1977ty, Krueger:2001vo, Krueger:2003cs}, or streamline curvature at the nozzle \cite{Krueger:2005dp}.
To realistically use pulsed jets for marine propulsion, it is necessary to understand how thrust production is affected by different flow conditions. For example, when a pulsed jet is ejected into a background flow parallel to the jet (co-flow), circulation (and hence thrust) production decreases as the speed of the co-flow increases \cite{Krueger:2003fb}. By contrast, if the ambient flow is antiparallel to the jet (counter-flow), the pulsed jet takes longer to separate from the nozzle, increasing the circulation in the vortex ring and the duration of the over-pressure benefit experienced by the jet \cite{Dabiri:2004cf}. Experiments and simulations have further indicated that circulation production can be controlled through the design, and in some cases real-time manipulation, of the nozzle geometry \cite{Allen:2005kk, Dabiri:2005gu, Rosenfeld:2009ii, Ofarrell:2014ko}.
For many applications, interactions between multiple jet pulses will strongly affect propulsion when the pulses are closely spaced in time. For these continuously-pulsed jets, thrust production and propulsive efficiency deviate from the single-pulse behavior. Such effects have been characterized as functions of design parameters including system geometry, jet velocity, and pulse frequency \cite{Moslemi:2010ji, Nichols:2012ku, Krieg:2013bm}. Using experiments on a model vehicle, \citeauthor{Ruiz:2010cd} demonstrated that the vehicle efficiency could be as much as 70\% greater when propelled by pulsed jets than when propelled by steady jets \cite{Ruiz:2010cd}. In follow-up experiments, \citeauthor{Whittlesey:2013fm} used a similar technique to relate such efficiency gains to the wake kinematics \cite{Whittlesey:2013fm}.
In addition to temporally-separated jet pulses, the geometric placement of independent jets can affect propulsive performance. It has been shown that multiple pulsed-jet thrusters can be used to improve the control and maneuverability for underwater vehicles \cite{Mohseni:2006eo, Krieg:2008bl}. However in these cases, the jets were spaced far enough apart that there were no hydrodynamic interactions between them. For other designs -- such as small robots with closely-spaced thrusters, or swarms of independent vehicles operating in close proximity -- the hydrodynamic interactions of multiple pulsed jets could affect the thrust and efficiency of the pulsed-jet propulsion.
Despite the importance of multi-jet interactions for underwater propulsion, there are no existing descriptions of these interactions that can inform thruster design. In this paper, we experimentally investigate how thrust production is affected by multi-jet interactions by visually observing the wakes formed by two parallel pulsed jets. We observe that, as nozzle spacing $\Delta$ decreases, the thrust and efficiency fall according to $1-Co\ (\Delta/D)^{-6},$ where $D$ is the nozzle diameter and $Co$ is a dimensionless ``coupling number'' that describes how strongly the two-nozzle coupling affects thrust and efficiency. We explain this dependence with a model based on vortex interactions and geometric constraints in this problem. Our model predicts the observed wake kinematics and reveals the potential for thrust augmentation under certain conditions.
\section{Experimental Approach}
\subsection{Hardware}
In order to understand the evolution of thrust and efficiency in a two-nozzle pulsed jet, we imaged pulsed jet formation at early times, using the motion of fluorescent dye in the wake to estimate hydrodynamic forces. A schematic of the experimental setup is provided in Fig.~\ref{fig:expt_geom}.
The jets were created in a (30.4cm)$^3$ cubic tank using two $D=6.35$mm inner-diameter stainless steel nozzles submerged in water. The nozzles were mounted on a linear rail so that nozzle spacing could be varied continuously. Flow through each nozzle was driven by an independent pressure reservoir. For these experiments, the pressure reservoir was hydrostatic, consisting of two open 60mL syringes filled to capacity. Between the reservoir and nozzles, additional hardware measured and controlled the flow in the experiment. Volume flux was measured using a low-inertia, positive-displacement flow meter (FCH-m-POM 97478039; BIO-TECH e.k.), and the flow was controlled by an inline solenoid valve (CNYUXI 2W-025-08).
\begin{figure}[t!]
\centering
\includegraphics[width=\narrowfig]{Figure1.eps}
\caption{\label {fig:expt_geom} Experiment schematic depicting all of the system components. The side view reveals the illumination and imaging system, which images the jets from the front.}
\end{figure}
When open, the valve constricted fluid flow to a 2mm diameter, and the associated viscous losses in the tubing and valve brought the effective driving pressure to $p_d\approx114$Pa (see Supplemental Content for details). This driving pressure fixed the jet velocity within the range $u_j=48.2-55.2$mm/s for a 0.3s pulse. Using the nozzle diameter as a length scale, the typical jet Reynolds number was $Re=350.$
In order to follow the evolution of the developing jets, the reservoirs contained water mixed with fluorescein dye ($5\times 10^{-7}$~M fluorescein sodium salt in water). This way, all of the fluid ejected from the nozzle was marked with a fluorescent tracer, while the fluid in the tank was transparent. The jets were illuminated with a blue (462nm) laser diode module (1.5W optical power; Lasertack LDM-462-1400-C). The laser module emits a 4mm-diameter gaussian beam that was expanded into a 4mm thick laser sheet using a cylindrical lens. This laser sheet was centered on the nozzles to illuminate the central plane of both jets, as depicted in Fig.~\ref{fig:expt_geom}. When exposed to the blue laser sheet, the fluorescein dye emits green light (532nm), which was recorded at 400~frames/second using a high-speed camera (Phantom Miro 320s; Vision Research Inc.). This frame rate corresponds to a temporal resolution of 2.5ms.
The hardware was digitally controlled to synchronized the valves, laser, and camera acquisition. Each experiment lasted for 1~second. First, the camera and laser were triggered, then the valves were opened providing a sudden pressure gradient to initiate the pipe flow. After 0.3s, the valves were closed, and the flow was recorded until 1s had passed from the start of the experiment. Finally, the tank was allowed to settle for 3~minutes before beginning the next experiment. This procedure ensured that no residual vorticity or dye would impact the measurements of consecutive experiments.
\subsection{Analytical Framework}
\begin{figure}[b!]
\centering
\includegraphics[width=\narrowfig]{Figure2.eps}
\caption{\label{fig:analysis_CV} Estimating thrust based on wake evolution. \textbf{(a)} The control volumes around a model vehicle and its vortical wake. The thrust experienced by the vehicle is balanced by the rate of momentum change in the wake. \textbf{(b)} A detailed control volume $\Omega_{cv}$ around the vortical wake. Two surfaces contribute to the momentum in the $\mathbf{\hat{z}}$ direction: the nozzle exit surface (red), and the time-varying material boundary between the wake and the external fluid (purple). The contributions to the momentum must balance so that the momentum flux through the nozzle exit surface is compensated by a mixture of control volume growth, motion, and pressure along the surface of the nearly-ellipsoidal control volume. \textbf{(c)} Pressure distribution around an ellipsoid moving steadily in the $+\mathbf{\hat{z}}$ (downward) direction. Red coloring indicates positive pressures, blue indicates negative pressures, and white indicates zero pressure. Streamlines of the associated potential flow are drawn in black.}
\end{figure}
Since the small forces in this experiment ($A_n \rho u_j^2\sim0.1$mN) are difficult to measure directly, we derive an indirect method that allows us to estimate the thrust produced by the jets using a video of the wake kinematics. A similar approach is described by \citet{Ruiz:2010cd}.
The thrust produced by an underwater jet can be calculated from the momentum flux through the nozzle exit plane. Surrounding the nozzle (or thruster) by a control volume (see Fig.~\ref{fig:analysis_CV}a), the forward thrust $T (-\mathbf{\hat{z}})$ is balanced by the inertial momentum transfer out of the nozzle and the pressure on the nozzle exit plane. Assuming that the jet velocity and pressure are roughly constant along the nozzle exit plane (using the `slug model' of vortex formation), the momentum equation can be integrated around the control surface to yield an equation for the thrust produced by the jet:
\begin{align}\label{eq:thruster_thrust}
T &= \int_{S_n} \left[\rho (\mathbf{u}\cdot\mathbf{\hat{z}})^2 + p_n - p_0\right]\ dA \nonumber \\
&\approx A_n (\rho u_j^2 + p_n-p_0),
\end{align}
\noindent where $A_n$ is the nozzle area, $\rho$ is the fluid density, $u_j$ is the fluid velocity exiting the nozzle, $p_n$ is the pressure along the nozzle-exit plane, and $p_0$ is the free-stream pressure. Because a pulsed jet initially rolls up into a vortex ring, for early times the nozzle pressure is not equal to the free-stream pressure, and the `nozzle over-pressure' contributes significantly to thrust production \cite{Krueger:2001vo, Krueger:2003cs}.
To relate the thrust production to wake kinematics, we consider a second control volume, $\Omega_{cv},$ surrounding the wake. This control volume encloses all of the fluid ejected from the nozzle, as well as any fluid entrained into the vortex structures that form. Given this geometry, the leading region within the control volume is often referred to as a ``vortex bubble'' \cite{Maxworthy:1972ty, Krueger:2001vo, Krueger:2003cs, Olcay:2008fc, Ruiz:2010cd}. The bubble's leading edge defines a material surface that separates the ejected fluid from the ambient fluid initially outside of the nozzle. In experiments, this material surface is easily visualized when dyed fluid is ejected into a clear fluid, as shown in Fig.~\ref{fig:analysis_vid}b. Vortex roll-up during early jet formation causes this material surface to resemble an oblate ellipsoid of revolution, enclosing the forming vortex ring and most of the vorticity in the flow (some of the vorticity diffuses beyond the extent of the material surface \cite{Maxworthy:1972ty, Olcay:2008fc}).
Given this second control volume, the thrust can be calculated from the vertical ($\mathbf{\hat{z}}$) momentum conservation equation:
\begin{multline}\label{eq:CV_momentum_z}
0=\underbrace{\frac{d }{d t} \int_{\Omega_{cv}} \rho \mathbf{u}\cdot\mathbf{\hat{z}}\ dV}_\text{term 1 - unsteady flow}
+ \underbrace{\rho \int_{S_{vb}} (\mathbf{u}\cdot\mathbf{\hat{z}}) \mathbf{u}_{rel}\cdot\mathbf{\hat{n}}\ dA}_\text{term 2 - fluid inertia} \\
+ \underbrace{\int_{S_{vb}} (p_{vb} + p_0) \mathbf{\hat{n}}\cdot\mathbf{\hat{z}}\ dA}_\text{term 3 - external pressure}
- \underbrace{\int_{S_n} \rho u_j^2 + p_n\ dA}_\text{term 4 - dynamic pressure at nozzle}
\end{multline}
Here, the domain $\Omega_{cv}$ is the entire control volume (CV) as highlighted in Fig.~\ref{fig:analysis_CV}b, $S_{vb}$ is the control surface bounding the (roughly) ellipsoidal vortex bubble, $S_n$ is the control surface at the nozzle exit plane, $\mathbf{\hat{n}}$ is the unit normal at each point on the control surface, $\mathbf{u}$ is the local velocity at each point, $\mathbf{u}_{rel}$ is the velocity relative to the control surface, and $p_{vb}$ is the pressure on the surface of the vortex bubble. To simplify Eq.~\ref{eq:CV_momentum_z}, we leverage several empirical observations and assumptions.
Focusing first on term 1, an analysis provided in the Supplementary Material reduces the unsteady term to $$\frac{d }{d t} \int_{\Omega_{cv}} \rho \mathbf{u}\cdot\mathbf{\hat{z}}\ dV \approx \rho( \dot{u}_{cm} V_{cv} + 2 u_{cm} \dot{V}_{cv} + z_{cm} \ddot{V}_{cv}).$$ Here, $z_{cm}$ is the $\hat{z}$-position of the control volume's center of mass, $u_{cm} = \dot{z}_{cm}$ is the velocity of that position, and $V_{cv}$ is the CV volume.
Second, observe that on most of the control surface $S_{vb},$ the surface evolves with the fluid so that there is no normal flux, and $\mathbf{u}_{rel}\cdot\mathbf{\hat{n}}=0$ on $S_{vb}.$ Over the region of $S_{vb}$ where entrained fluid enters the control volume, it is often assumed that the entrainment is nearly tangential so that even in this region, $\mathbf{u}_{rel}\cdot\mathbf{\hat{n}}=0$ \cite{Ruiz:2010cd, Olcay:2008fc}. Taken together, these observations indicate that the second term of Eq.~\ref{eq:CV_momentum_z} is negligible.
Third, the two pressures in the third term can be separated. Then, the integrated free-stream pressure around this control surface cancels everywhere except for directly below the nozzle. This reduces the integral to: $$ \int_{S_{vb}} p_0 \mathbf{\hat{n}} \cdot \mathbf{\hat{z}}\ dA = \int_{S_n} p_0\ dA, $$ which can be combined with the fourth term of Eq.~\ref{eq:CV_momentum_z}. This final step reveals the modified fourth term as the thrust that is given in Eq.~\ref{eq:thruster_thrust}.
Leveraging these observations, and rearranging terms in Eq.~\ref{eq:CV_momentum_z}, the thrust produced by the jet can be expressed in terms of the dynamics of the jet wake:
\begin{align}\label{eq:thrust_kinematics}
T & = A_n \rho u_j^2 + A_n p_n \\
& \approx
\underbrace{\vphantom{\int_{A_b}} \rho \dot{u}_{cm} V_{cv}}_\text{I}
+ \underbrace{\vphantom{\int_{A_b}} 2 \rho u_{cm} \dot{V}_{cv}}_\text{II}
+ \underbrace{\vphantom{\int_{A_b}} \rho z_{cm} \ddot{V}_{cv}}_\text{III}
+ \underbrace{\int_{S_{vb}} p_{vb}\, \mathbf{\hat{n}} \cdot \mathbf{\hat{z}}\ dA}_\text{IV}. \nonumber
\end{align}
By applying momentum conservation to the CV surrounding the jet wake (Eq.~\ref{eq:thrust_kinematics}), the thrust generated by the pulsed jet is broken down into four terms that rely on three measurable quantities. Term I represents the force to instantaneously accelerate the CV the $\mathbf{\hat{z}}$ direction. Terms II and III represents the forces required to add mass to the growing CV, both by injection and entrainment. The first three terms can be measured by tracking the CV volume and center of mass position at each time during its growth. Term IV describes the pressure around the growing vortex bubble that resists its growth. This term can be thought of in terms of the added mass associated with the growing vortex bubble as it pushes all of the external fluid out of the way \cite{Krueger:2001vo, Dabiri:2005ce}. Term IV requires knowledge of the pressure field surrounding the vortex bubble, which is not directly measurable from the fluorescence experiments.
In order to circumvent the need to directly measure pressure on the vortex bubble, we estimate the pressure on the vortex bubble using the common approximation that flow outside the vortex bubble is nearly irrotational. Using this assumption and the observation that the vortex bubble is approximately ellipsoidal (see Fig.~\ref{fig:analysis_vid}c), the pressure on the leading surface of $S_{vb}$ can be equated to the pressure on the surface of an ellipsoid that is translating and expanding unsteadily in a potential flow. Such a pressure field is illustrated with the corresponding streamlines in Fig.~\ref{fig:analysis_CV}c. As expected, the pressure field resembles that around a translating sphere, deformed to match the ellipsoidal geometry.
This potential flow model reduces the pressure term (Eq.~\ref{eq:thrust_kinematics}-III) to a function of vortex bubble motion and geometry, which can be tracked as shown in Fig.~\ref{fig:analysis_vid}c. With this simplification, the total thrust can be estimated entirely based on the motion and growth of the wake.
\begin{figure}[t!]
\centering
\includegraphics[width=\narrowfig]{Figure3.eps}
\caption{\label {fig:analysis_vid} Key steps in video processing for thrust estimation. \textbf{(a)} A calibration image of the nozzles is acquired and analyzed to locate nozzle position, size, and camera tilt. These measurements are used to preprocess the images so that they are consistent across imaging runs, and can be analyzed with physical units. \textbf{(b)} Typical image from a two-nozzle experiment ($\widetilde{\Delta}=1.94$). \textbf{(c)} Results of control volume analysis on the experiment images. Left nozzle - the control volume is shaded in blue, with the centerline shown as the vertical dotted line. Right nozzle - the ellipsoid fit to the wake for surface pressure calculations.}
\end{figure}
\subsection{Video Processing for Thrust Estimation}
To estimate forces from the video data, images are acquired and processed to identify the time-dependent control volume surrounding the wake (see Fig.~\ref{fig:analysis_vid}c, left). First, a calibration image (Fig.~\ref{fig:analysis_vid}a) is acquired and used to determine camera orientation and nozzle positions as shown in the figure. These measurements are used to rotate the raw video data to a standard orientation, and to crop the video into two separate frames - one of each nozzle. Next, preprocessed images are filtered and segmented to identify the CV. In addition to the CV, a bounding ellipsoid is identified for each wake (Fig.~\ref{fig:analysis_vid}c). Finally, the (axisymmetric) volume and center of mass position are numerically calculated and used to calculate the different terms in Eq.~\ref{eq:thrust_kinematics}. While the first three terms can be calculated directly from the CV size and motion, the pressure term requires further computation.
The pressure field surrounding the ellipsoid is calculated by using the unsteady Bernoulli equation in conjunction with the velocity potential around a translating ellipsoid. The calculated pressure field is then numerically integrated around the leading edge of the vortex bubble directly below the nozzle (in 3D, assuming axisymmetry) where the pressure field and $\hat{z}$-component of the surface normal are greatest. A more detailed description of the pressure analysis technique can be found in the Supplemental Content and in \citep{Athanassiadis:2016vu}. As will be discussed in Section~\ref{sec:results:two_noz}, the pressure term ultimately contributes less than 10\% of the total thrust at early times, meaning the unsteady flow terms dominate the early-time wake dynamics.
\section{Results}
We track the evolution of jet wakes for individual jets \nobreak{($\Delta\to\infty$)}, and for select nozzle spacings increasing from {$\Delta=1.5 D.$} Because $\widetilde{\Delta} = \Delta/D$ arises as the relevant dimensionless group the new geometry introduces, we use $\widetilde{\Delta}$ instead of $\Delta$. For each value of $\widetilde{\Delta}$ tested, five experiments were performed. Videos from the experiments are available in the Supplemental Content.
For each two-nozzle experiment, the two jet wakes are analyzed independently, providing two measurements of thrust generation for each experiment. Because the nozzles were not perfectly identical, their wakes varied slightly, producing differences in the measurements. To account for this variation in our plots, we eliminate the effects of nozzle fabrication by normalizing each measurement by the average value measured at $\widetilde{\Delta}\to\infty.$ In Figs.~\ref{fig:evolution} and \ref{fig:thrust}, the results from nozzle 1 (left nozzle) are indicated by blue squares, and those from nozzle 2 (right nozzle) are indicated by red triangles. For all experiments, the wake is analyzed between $t=0.08$s and $t=0.30$s. Because the initial startup flow is slow, analysis of the frames before 0.08s did not produce consistent results. The time $t=0$ corresponds to the moment when the solenoid valves are opened and the jets begin to develop.
\subsection{Single Nozzle Dynamics}
To validate the analysis technique described above, we calculate the scale of the force estimated from the wake of individual pulsed jets, corresponding to the limit $\widetilde{\Delta}\to\infty.$ In these experiments, our analysis estimates a time-averaged thrust of $\overline{T}_\infty=0.10 \pm 0.01$mN, which closely matches the expected thrust scale $T_{exp} \sim \rho u_j^2 A_n = 0.07$mN. The measured average thrust is slightly higher than this predicted scale because of the positive nozzle over-pressure during jet formation.
The wake kinematics for a single jet provide insight into the physical mechanisms driving thrust production in the pulsed jet. First, both the CV center of mass $z_{cm}$ and volume $V_{cv}$ grow nearly linearly with time throughout the experiments, so that the higher derivatives $\dot{u}_{cm}=0$ and $\ddot{V}_{cm}=0.$ This experimental observation has important consequences for the thrust calculation process, reducing Eq.~\ref{eq:thrust_kinematics} to just two terms (II and IV). Additionally, the pressure term (term IV) is 10 times smaller than the unsteady term (term II), indicating that flow unsteadiness within the wake is the primary source of thrust during early times. These observations persist in the two-nozzle experiments, supporting the same conclusions across all the experiments (see Supplemental Content for supporting plots).
\begin{figure}[t]
\centering
\includegraphics[width=\narrowfig]{Figure4.eps}
\caption{\label {fig:evolution} Average CV velocity and volume during a 0.3s pulsed jet for different nozzle spacings $\widetilde{\Delta}.$ These results are normalized by the single nozzle values for each nozzle (red triangles represent the right nozzle, blue squares represent the left nozzle). Pulsed-jet interactions cause the wake velocity to drop according to the form $1-C_u\widetilde{\Delta}^{-6}$ (fit shown in figure as solid black line). Error bars on the fit reflect the standard deviation of experimental values $u_{cm,\infty}.$}
\end{figure}
\subsection{Two-Nozzle Dynamics}\label{sec:results:two_noz}
As two nozzles are brought into close proximity, their interactions lead to changes in their kinematics as shown in Fig.~\ref{fig:evolution}. In these plots, the multi-jet interactions manifest as a significant change in the wake velocity, but not in the wake volume. Empirically, the wake velocity drops from the single-nozzle value $u_{cm,\infty}$ according to $u_{cm}(\widetilde{\Delta}) / u_{cm,\infty} = 1 - C_u \widetilde{\Delta}^{-6},$ where $C_u=1.86\pm0.06$ is a dimensionless constant that reflects how strongly the two pulsed jets interact. The solid line in the plot represents the (one-parameter) fit of this form to the data, and the shaded region represents the standard deviation of the experimental values $u_{cm,\infty}$.
When the wake kinematics are used to calculate the thrust as described in Eq.~\ref{eq:thrust_kinematics}, the thrust generated by each nozzle is observed to follow the same form as the wake velocity, as shown in Fig.~\ref{fig:thrust}. As with the single-nozzle experiments, the thrust from the pressure term (Eq.~\ref{eq:thrust_kinematics}-IV) is 10 times smaller than the thrust associated with unsteady motion in the CV, indicating that the flow unsteadiness within the CV dominates thrust production.
As the two jets are brought together, the total thrust is reduced by nearly 10\%. Empirically, the dependence of thrust on nozzle spacing can be described by the same equation as the wake velocity dependence, with a different coupling coefficient. In this case
\begin{equation}\label{eq:EmpericalThrust}
T(\widetilde{\Delta}) / T_\infty = 1 - Co \widetilde{\Delta}^{-6},
\end{equation}
for a thrust coupling coefficient $Co=2.04 \pm 0.11.$
Given the experimental data in Fig.~\ref{fig:thrust}, this model over-predicts the two points where $\widetilde{\Delta}>3.0.$ However, this inconsistency can be seen as arising from the lower volume flow rate in the experiments at those values of $\widetilde{\Delta}$ (see Fig.~\ref{fig:evolution}), and not from the emergence of an unexplained physical phenomenon at those length scales. This explanation is supported by the assumption that the thrust $T$ should smoothly and monotonically asymptote to the single-nozzle value $T_\infty.$
\begin{figure}[t]
\centering
\includegraphics[width=\narrowfig]{Figure5.eps}
\caption{\label {fig:thrust} Time-average thrust produced during a 0.3s pulsed jet for different nozzle spacings $\widetilde{\Delta}.$ Pulsed-jet interactions reduce thrust production for very close spacing, according to $T/T_\infty = 1-Co\widetilde{\Delta}^{-6}$ (fit shown in figure as solid black line). Here, $Co=2.04$ is a dimensionless `coupling coefficient' that describes how strongly jet interactions affect thrust production. As discussed in the text, this plot also identically reflects the behavior of propulsive efficiency $\overline{\eta}(\widetilde{\Delta})/\overline{\eta}_\infty.$\vspace{-2em}}
\end{figure}
\section{Discussion}
Our results indicate that the interactions between two parallel pulsed jets are a very local phenomenon. When nozzles are separated beyond a few nozzle diameters, wake interactions do not measurably affect thrust production. However, when two simultaneously pulsed jets are brought into close proximity, their wakes destructively interfere with each other, causing the thrust produced by each jet to drop by as much as 10\%. In this section, we describe a physical mechanism that accounts for this observed thrust drop with nozzle spacing.
\subsection{Physical Mechanism for Interactions}
To build an intuition for the physics in the two-jet system, it is helpful to recall some of the observations about the kinematics of the control volumes that led to the thrust estimates above. Observation of the wake growth revealed that $z_{cm}$ and $V_{cv}$ both grow linearly with time at early times. This behavior indicates that $u_{cm}$ and $\dot{V}_{cv}$ are the only nonzero derivatives, so that the unsteady terms Eq.~\ref{eq:thrust_kinematics}I and III are zero. Further, observing that the surface pressure (term IV) is 10 times smaller than the unsteady forces within the wake (term II), we can write that to leading order, the wake dynamics should behave according to $T\approx 2\rho u_{cm} \dot{V}_{cv}.$ Since the driving jet velocity, $u_j$ and the wake volume growth rate $\dot{V}_{cv}$ do not change significantly as the nozzles are brought together, the interactions between the nozzles do not affect entrainment in the wakes. Therefore, the scaling behavior of the thrust production is set by the average velocity within the wake, which we have shown is characterized by the wake center of mass velocity $u_{cm}.$ As the nozzles are brought closer together, this velocity decreases, so the thrust should decrease with an identical form. Physically, this picture is consistent with the hypothesis that vortex induction drives the two-jet interaction: if two separate, coplanar vortex rings are established in an infinite fluid, the circulation in each would establish a counter-flow that lowers the average velocity of the other.
Guided by this intuition, we more rigorously derive a mechanism for the thrust reduction observed in our experiments.
\subsubsection{Scaling as $\widetilde{\Delta}^{-6}$}
To predict the thrust scaling, we recall the result from Eq.~\ref{eq:thrust_kinematics} that for a single nozzle $$T \sim A_n (\rho u_j^2 + p_n-p_0). $$ Since $\rho,$ $u_j,$ and $p_0$ remain constant as $\widetilde{\Delta}$ is varied, the thrust should scale as the nozzle pressure $T(\widetilde{\Delta}) \sim p_{n,\infty} (\widetilde{\Delta}).$ When a second pulsed jet is introduced into the problem, the nozzle pressure should be modified by a new pressure scale introduced in the problem. In this case, one new pressure scale introduced is that of the vortex ring formed at the other nozzle. Because the velocity induced by a toroidal vortex ring scales as $u_{ind}(r)\sim r^{-3}$ (derivation in supplemental content), the induced pressure should scale as $p_{ind}(r)\sim\rho u_{ind}^2\sim r^{-6}$. Therefore, we can write the thrust produced by a single pulsed jet when a second is located a distance $\Delta$ away:
\begin{align*}
T(\Delta) &= A_n (\rho u_j^2 + p_n(\Delta)) \\
&= A_n (\rho u_j^2 + p_{n,\infty} \pm p_{ind}(\Delta)) \\
&= T_\infty \pm C \Delta^{-6}.
\end{align*}
Reorganizing and and non-dimensionalizing the right hand side of this equation recovers the functional form observed in our experiments (Eq.~\ref{eq:EmpericalThrust}). From here, the sign of the $T$ vs $\Delta$ relationship remains to be determined, corresponding to an expectation for thrust reduction or augmentation.
\subsubsection{Geometric Argument for Thrust Reduction}
In order to predict whether we should expect thrust to increase or decrease from the two-jet interaction, we consider the analysis by \citeauthor{Krueger:2005dj}, which relates the nozzle pressure to the curvature of streamlines at the nozzle. A higher streamline curvature at the nozzle corresponds to a higher average nozzle pressure, according to the equation \cite{Krueger:2005dj}:
\begin{equation}\label{eq:streamlines}
p_n \sim \rho \int_0^{D/2} u_z \frac{\partial u_r}{\partial z}\Big|_{z=0} dr
\end{equation}
In the case of two simultaneous pulsed jets, we expect that the interaction between the two forming vortex wakes reduces the streamline curvature at the nozzle exit planes. To see this, consider that the presence of a second nozzle introduces a symmetry plane between the nozzles, across which there can be no volume flux. Accordingly, the flow must adjust to satisfy the zero-flux condition at the symmetry plane, which will restrict the radial growth of the wake. Since the radial flow $u_r$ is restricted, the term $\partial u_r/ \partial z$ should be reduced. This effect can be interpreted as forcing the streamlines coming out of the nozzle to straighten, thereby lowering the nozzle over-pressure $p_n$ according to Eq.~\ref{eq:streamlines}. Therefore, the thrust produced by interacting jets should be lower than that of a single pulsed jet.
\subsubsection{Additional comments on this model}
This physical mechanism for thrust reduction is consistent with observations in previous experiments on single-nozzle pulsed jets. When a pulsed jet is exposed to ambient co- or counter-flow, the production of circulation (a proxy for nozzle over-pressure) is decreased or increased (respectively) as described by the results of Krueger, Dabiri, and Gharib \cite{ Krueger:2003fb,Dabiri:2004cf}.
In the case of ambient co-flow (flow parallel to jet), our model would predict that additional stream-wise flow should increase the axial extent of the wake ($a$ decreasing radial growth ($b$) as a result of continuity. Based on these assumptions, our geometric model predicts that the nozzle pressure should decrease, thereby lowering the production of thrust and circulation as is reported by \citet{Krueger:2003fb}. Conversely, in the case of ambient counter-flow (flow antiparallel to the jet), the predictions would reverse, suggesting that the wake should be `flattened' by the counter-flow (i.e. $a$ decreases and $b$ increases by continuity). In that case, the geometric model would predict a higher nozzle pressure, thrust and circulation production, which is reflected in the data from \citet{Dabiri:2004cf}.
\subsection{Efficiency Considerations}
The impact of pulsed-jet interactions on propulsive efficiency can be estimated by considering a conceptual thruster such as the one illustrated in Fig.~\ref{fig:analysis_CV}a. To produce thrust, the jet is driven by a pressure $p_d,$ which ejects fluid at a rate $\dot{Q}.$ Then the power put into generating the thrust is $\dot{E}= p_d \dot{Q}.$
As a measure of useful work, we consider that the goal of these thrusters is maneuverability - rapid bursts of (well-defined) thrust for short periods of time. So a measure of useful work is the thrust produced $T,$ multiplied by the speed at which it can be delivered $u_j$, so that the efficiency of the thruster is given by: $$ \eta = \frac{T u_j}{p_d \dot{Q}} = \frac{T}{p_d A_n}. $$
Since the nozzle area $A_n$ and driving pressure $p_d$ are independent of $\Delta,$ the efficiency of the thruster should scale as the thrust $T$ does. As a result, compared to the efficiency of a single jet $\eta_\infty,$ the efficiency of two jets separated by a distance $\widetilde{\Delta}$ will be given by:
\begin{equation}\label{eq:efficiency_thrust_relation}
\frac{\eta (\widetilde{\Delta})}{\eta_\infty} = \frac{T (\widetilde{\Delta})}{T_\infty} = 1 - Co \widetilde{\Delta}^{-6}.
\end{equation}
\noindent Based on this result, Fig.~\ref{fig:thrust} indicates not only how thrust varies with nozzle spacing, but also how the efficiency should vary with nozzle spacing as well.
\section{Conclusions}
The importance of multiple pulsed jets for underwater propulsion has led us to investigate the effects of multi-jet interactions on thrust production and efficiency for simultaneous pulsed jets. We have developed a control volume approach to estimate thrust production from videos of interacting pulsed jets. This analysis has shown that for intermediate Reynolds number and low stroke length, thrust production is dominated by unsteadiness within the growing wake. When two jets are brought into close proximity, vortex interactions between the jets force the streamlines to straighten, causing the thrust to drop as much as 10\%. However, this effect is highly localized, depending on the nozzle spacing as $\widetilde{\Delta}^{-6},$ so that in practice, a designer should not be concerned with interactions between simultaneous jets separated by more than 2.5 diameters.
Given the data and scaling arguments presented above, the problem remains to determine what sets the coupling coefficient $Co.$ Are there system configurations that can allow $Co<0$? Such behavior would provide a means to increase the pulsed jet's thrust and efficiency through clever system design or control.
One approach to thrust augmentation is suggested by the role of streamline curvature. If instead of being ejected simultaneously, jets are pulsed so that the second jet is ejected as a stopping vortex forms in the first jet, the close proximity of the negative vorticity from the first jet may \emph{help} to curve the streamlines exiting the second nozzle, thereby augmenting the nozzle over-pressure. In this way, well-timed pulses could exploit the stopping vorticity in each others' wakes to roll-up more efficiently and produce more thrust. This behavior is reminiscent of how jellyfish exploit stopping vortices to move more efficiently \cite{Gemmell:2013ve}.
The analysis presented here is not limited to two nozzles. When considering how the nozzle interactions scale to systems with more nozzles (such as a multi-jet vehicle, or a swarm of small vehicles), the geometric and pressure scalings should still apply in a multi-nozzle system. Additional jets will further straighten the flow, and provide additional pressure scales that can be added linearly to the nozzle over-pressure. The most significant difference will be observed because of the 3D nature of the wake development. While the pressure scales can be added linearly, and treated pairwise, the streamlines will develop based on a more complex three-dimensional flow field around the nozzle, and this behavior cannot be predicted by the current analysis.
While we have focused here solely on single-pulsed jets for applications to precision impulse delivery, the interactions between multiple jets can have interesting consequences for continuous pulsed jets. For instance, when our results are related to the analysis and observations by \citeauthor{Ruiz:2010cd} and \citeauthor{Whittlesey:2013fm}, it is reasonable that the changes in wake geometry we observe will have a significant effect on propulsive efficiency in a continuous pulsed jet. Such open questions present exciting opportunities for further research on the interactions of multiple pulsed jets in different operating regimes.
\begin{acknowledgments}
AGA thanks A. Helal, J. Alvarado, A. Nasto, S. Sroka, and C. Wagner for valuable discussions and feedback. This work was supported by the Lincoln Laboratory and the Office of Naval Research through Award No. 7000308296.
\end{acknowledgments}
\section{Additional Experiment Details}
\subsection{Nozzles Fabrication}
The nozzle bores were machined out to the final diameter from stock with a smaller bore. The outer diameter of the nozzles was 9.52mm, and the faces of the nozzles were machined smooth and deburred, but were not tapered to a fine angle or sharp edge as in other experiments \cite{Krueger:2003cs, Krueger:2005dj, Maxworthy:1972ty}. The reservoir was connected to the nozzle with 6.35mm Tygon tubing. Differences in the machining of the two nozzles caused variations in the wakes generated by each nozzle. These differences caused the measurements of wake velocity and thrust to vary slightly between the two nozzles. As described in the text, the effects are accounted for by normalizing the measurements by the single-jet quantity for each jet.
\subsection{Nozzle Positioning and Driving Pressure}
In our experiment, the separation between the tank walls and the nozzle was 120mm, and the nozzle exit plane was submerged 40mm below the free surface. With this configuration, edge-effects are considered negligible during the vortex formation.
During a typical experiment, the height of the fluid in the reservoirs changed by less than 0.5mm, allowing the change in pressure head from surface motion to be ignored relative to the total reservoir height. The reservoirs were elevated $H=570$mm above the water surface for an expected pressure head of $p_h=\rho g H\approx5500$Pa. However, as described in the Supplemental Materials, losses between the reservoir and the nozzle reduced the effective pressure head to $p_{h,eff}\approx114$Pa.
Given the geometry, the hydrostatic pressure at the nozzle exit is $p_{n,g} = \rho g (40\text{mm})=392\text{Pa}.$ Because of the pressure loss through the tubing and valve, the effective driving pressure was estimated using a simplified momentum equation for unsteady pipe flow:
$$p_d - p_{n,g} \approx \frac{\rho u_j H}{t} = 114\text{Pa}.$$
\subsection{Jet Velocity}
To support our claim in the text that the jet velocities were nearly linear over the course of our experiment, Fig.~\ref{fig:supp:FlowMeter}a shows a typical trace of the flow-meter output for two single-nozzle experiments. As the nozzle spacing was varied, we observed no meaningful change in the jet velocity for either nozzle.
\subsection{Optical Filtering}
To prevent scattered laser light from affecting the optical signal, a 495nm optical high-pass filter (Thorlabs FGL495S) was placed in front of the camera.
\begin{figure*}
\centering
\includegraphics[width=0.4\textwidth]{FigureS1.eps}
\caption{\label{fig:supp:FlowMeter} Volume flux through the nozzle as measured by the flow meters attached to each reservoir. The volume flow through the nozzles is approximately linear in time, so to calculate the volume flow rate, the data for each nozzle are fit to a line. Here, the volume flow rate through each nozzle is $\dot{Q}=1.68\pm0.04$mL/s. The vertical offset of the two traces reflects different initial readings in the flow meters, which are irrelevant for the flow rate calculations.}
\end{figure*}
\section{Analysis of Unsteady Thrust Term}
Our dye-based analysis requires a way to calculate the unsteady term in the $\mathbf{\hat{z}}$ momentum equation, which is reproduced here for convenience: $$ \frac{d}{dt} \int_{\Omega_{cv}} \rho \mathbf{u}\cdot\mathbf{\hat{z}}\ dV.$$ As written, this term requires knowledge of the velocity field $\mathbf{u}$ everywhere within the control volume, which is not accessible through our dye measurements.
In order to measure flow unsteadiness from dye motion, we rewrite the integral in a Lagrangian formulation so that the local velocity $u_z$ is the time derivative of the $\mathbf{\hat{z}}$ position of a particle within the CV. Then, we can write:
\begin{align*}
\int_{\Omega_{cv}(t)} u_z\, dV &= \int_{\Omega_{cv}(t)} \frac{\partial Z(\mathbf{r}, t)}{\partial t}\,dV \\
&=\frac{d}{dt}\left[\int_{\Omega_{cv}(t)} Z\ dV\right] - \oint_{S_{cv}(t)} Z\ (\mathbf{u} \cdot\mathbf{\hat{n}})\ dA \\
&= \frac{d}{dt} \left[ z_{cm} V_{cv} \right]
\end{align*}
The jump from the first to second line is made using the Reynolds Transport Theorem. We expect the surface integral to be zero because the control surface tracks the dye motion (this assumption validated below). The final step is completed by observing that the integral in the first term is related to the $\mathbf{\hat{z}}$ position of the center of mass $$z_{cm} = \frac{1}{V_{cv}}\int_{\Omega_{cv}} Z dV.$$ Inserting this derived quantity into the unsteady term above yields the result claimed in the main text:
\begin{equation}\label{eq:supp:unsteadiness}
\frac{d }{d t} \int_{\Omega_{cv}} \rho \mathbf{u}\cdot\mathbf{\hat{z}}\ dV = \frac{d^2}{dt^2}[z_{cm} V_{cv}] = \dot{u}_{cm} V_{cv} + 2 u_{cm} \dot{V}_{cv} + z_{cm} \ddot{V}_{cv}.
\end{equation}
We validated this analysis using 2D finite-volume simulations of a jet exiting a nozzle. By inserting a passive tracer into the nozzle before initiating flow, we can calculate a control volume equivalent to the one used in the experiments. This control volume was then used to calculate the unsteady term by directly integrating the flow field, as well as by using the CV volume and center of mass position. The two independent approaches produced identical results as shown in Fig.~\ref{fig:supp:sim}.
\begin{figure*}[b]
\centering
\includegraphics[width=0.75\textwidth]{FigureS2.eps}
\caption{\label{fig:supp:sim} Results of a 2D simulation used to verify our simplification of the unsteady integral. \textbf{(a)} A passive tracer is convected in the simulated flow to produce an image that can be tracked like the experiments. \textbf{(b)} Using the same analysis technique as for the experiments, the control volume is defined based on the dye location. \textbf{(c)} Comparison between direct integration of the velocity field within the CV and our estimate for the integral \eqref{eq:supp:unsteadiness}. The two calculation techniques overlap exactly within numerical fluctuations that arise from the discrete derivatives required by our simplification.}
\end{figure*}
\newpage
\section{Video Processing}
\subsection{Segmentation and Region Identification}
Once the images are aligned and preprocessed, they are filtered and segmented using a local Otsu threshold \cite{Otsu:1975wu, scikit-image} to identify the regions of dyed fluid. The segmented regions are finally filtered with a horizontal hole-filling algorithm to produce a control volume such as the one shown in Fig.~3c in the main text.
The ellipsoid is identified such that the ellipsoid major radius $b$ is half of the maximum wake width. The minor axis $a$ lies along the nozzle centerline from the major axis to the leading edge of the wake. Since the wake is assumed to be axisymmetric, these two axes uniquely define the bounding ellipsoid as shown on the right of Fig.~3c in the main text.
\section{Leading Edge Pressure Estimation Technique}
This approach was inspired by the work of \citeauthor{Munk:1924tw} on potential flow around ellipsoids for airship design \cite{Munk:1924tw}. In order to compute the last term of the thrust equation (Eq.~3 in the main text), we estimate the pressure on the leading edge of the control surface by evaluate the velocity potential $\phi$ around a translating oblate ellipsoid of revolution. Using the Bernoulli equation:
\begin{equation}\label{eq:bernoulli}
p_{vb} - p_0 = \frac{\rho}{2}(\mathbf{\nabla \phi})^2\Big|_{S_{vb}} + \rho \frac{\partial \phi}{\partial t}\Big|_{S_{vb}}.
\end{equation}
We now derive a technique to analytically evaluate the velocity potential in terms of ellipsoid geometry (principal axes $a$ along $\mathbf{\hat{z}}$ and $b$ along $\mathbf{\hat{r}}$) and velocity ($w \mathbf{\hat{z}}$), which are measured as a function of time based on dye motion.
As given by \citeauthor{Lamb:1945tl}, the velocity potential for a translating ellipsoid with principal radii $a$ in the $\mathbf{\hat{z}}$ direction, $b$ in the $\mathbf{\hat{y}}$ direction, and $c$ in the $\mathbf{\hat{x}}$ direction, moving at velocity $w$ in the $+\mathbf{\hat{z}}$-direction is most efficiently described in ellipsoidal coordinates \cite{Lamb:1945tl}:
\begin{equation}\label{eq:ellipsoid_potential_lamb_1}
\phi(x,y,z,t) = - z\ \cdot\
\underbrace{\frac{w}{2-\alpha_0}}_{C(t)}\ \cdot\
\underbrace{a b c \int_\lambda^\infty \frac{d\lambda'}{(a^2 + \lambda')^{3/2}(b^2+\lambda')^{1/2}(c^2+\lambda')^{1/2}}}_{\alpha(\lambda, a,b,c; t)}.
\end{equation}
In Eq.~\ref{eq:ellipsoid_potential_lamb_1}, $\lambda$ is the ellipsoidal coordinate that grows perpendicular to the ellipsoid with principal radii $(c,b,a)$ along $(x,y,z)$ respectively. As shorthand for the different terms in Eq.~\ref{eq:ellipsoid_potential_lamb_1} I introduce the definitions:
\begin{align}
C(t) &= \frac{w}{2-\alpha_0}, \label{eq:ell_int_C} \\
\alpha_0 &= \alpha(0, a,b,c;\ t),\text{ and} \label{eq:ell_int_alpha0}\\
\alpha(\lambda, a,b,c;\ t) &= abc \int_\lambda^\infty \frac{d\lambda'}{(a^2 + \lambda')^{3/2}(b^2+\lambda')^{1/2}(c^2+\lambda')^{1/2}} \label{eq:ell_int_alpha}
\end{align}
It should be noted that $x,y,z$ and $\lambda$ are defined with respect to the moving ellipsoid center. It therefore makes sense to consider the equivalent problem of the ellipsoid in an unsteady free-stream flow with changing velocity $-w(t)\ \mathbf{\hat{z}}.$
A point in ellipsoidal coordinates is defined by the variables $(\lambda, \mu, \nu)$ such that surfaces of constant $\lambda$ are ellipsoids offset from the primary ellipsoid with principal radii $(c,b,a)$ in the $(x,y,z)$ directions respectively. To convert a point $(x,y,z)$ from cartesian coordinates into ellipsoidal coordinates, the variables $\lambda, \mu, \nu$ can be found as the solutions of the third-order polynomial in $k$ defined by:
$$ \frac{z^2}{a^2 + k} + \frac{y^2}{b^2 + k} + \frac{x^2}{c^2 + k} - 1 = 0. $$
Canonically, the solutions for $k$ are sorted so that $\lambda > \mu > \nu.$ With this sorting, $\lambda$ represents the spatial variable that increases perpendicular to the ellipsoid surface; surfaces of constant $\lambda$ take the form of ellipsoids with scaled principle radii. The surface $\lambda=0$ lays coincident with the surface of the ellipsoid $(c,b,a)$.
For the case of an oblate ellipsoid of revolution (about the $\mathbf{\hat{z}}$ axis) with $a<b,$ $b=c$ and $x^2 + y^2 = r^2,$ this equation reduces to:
\begin{equation} \label{eq:ell_coord_transformation}
\frac{z^2}{a^2 + k} + \frac{r^2}{b^2 + k} - 1 = 0.
\end{equation}
Since we are only interested in the pressure on the surface ($\lambda=0$) of such an oblate, axisymmetric ellipsoid translating along $\mathbf{\hat{z}}$, we can also use these assumptions to simplify Eq.~\ref{eq:ellipsoid_potential_lamb_1}:
\begin{equation}\label{eq:ellipsoid_potential_lamb_simp}
\phi(x,y,z,t) = - z \cdot C(w, a, b; t) \cdot \alpha(\lambda=0, a,b;\,z,r,t).
\end{equation}
\noindent Furthermore, this simplification admits the following closed-form solution to the integral:
\begin{align}
\alpha(&\lambda, a,b;\ t) \nonumber \\
&= a b^2 \int_\lambda^\infty \frac{d\lambda'}{(a^2 + \lambda')^{3/2} (b^2 + \lambda)} \label{eq:alpha_integral}\\
& = ab^2 \left[ -\frac{\pi}{(b^2-a^2)^{3/2}} + \frac{2}{(b^2-a^2)(a^2 + \lambda)^{1/2}} + \frac{2 \tan^{-1}\left(\sqrt{\frac{a^2+\lambda}{b^2-a^2}}\right)}{(b^2-a^2)^{3/2}} \right]. \label{eq:alpha_integral_soln}
\end{align}
The integral $\alpha$ relates to the Bessel added mass $k_z,$ for the ellipsoid moving in the $z$-direction, according to \cite{Lamb:1945tl}:
\begin{equation} \label{eq:added_mass_ellipsoid}
k_z = \frac{\alpha_0}{2-\alpha_0}
\end{equation}
Using this relationship, it can be shown that the closed form solution Eq.~\ref{eq:alpha_integral_soln} approaches the expected limits as $a\to0$ and $a\to b$, corresponding to a flat plate and a sphere respectively.
From these simplifications, it is possible to explicitly calculate the pressure on the surface of the vortex bubble using only the measured motion and growth of the ellipsoid bounding the vortex bubble. To calculate the spatial derivative required by the equation (Eq.~\ref{eq:bernoulli}) I observe that the only spatially dependent terms in Eq.~\ref{eq:ellipsoid_potential_lamb_simp} are $C$ and $\alpha$ so that:
$$ \mathbf{\nabla} \phi(t) \Big|_{S_{vb}} =
-\mathbf{\hat{z}}\cdot C(w, a, b; t)\cdot \alpha_0(a,b,t)
- z\cdot C(w, a, b; t) \cdot \frac{\partial \alpha}{\partial \lambda}\Big|_{\lambda=0}\cdot \left( \mathbf{\hat{z}} \frac{\partial\lambda}{\partial z}\Big|_{\lambda=0} + \mathbf{\hat{r}} \frac{\partial\lambda}{\partial r} \Big|_{\lambda=0}\right).$$
\noindent The derivative $\partial \alpha / \partial \lambda$ can be evaluated by applying the fundamental theorem of calculus to Eq.~\ref{eq:alpha_integral}. The spatial derivatives of $\lambda$ can then be calculated by implicitly differentiating Eq.~\ref{eq:ell_coord_transformation} with $k \equiv \lambda.$ To calculate the time derivative of the velocity potential on the surface of the ellipsoid, the chain rule can be applied to $\phi:$
$$ \frac{\partial \phi}{\partial t} \Big|_{S_{vb}} = -z\cdot\alpha(0,a,b; t)\cdot \left( \frac{\partial C}{\partial w} \dot{u}_{vb} + \frac{\partial C}{\partial a} \dot{a} + \frac{\partial C}{\partial b} \dot{b} \right)
- z\cdot C(w,a,b; t) \cdot \left( \frac{\partial \alpha}{\partial a} \dot{a} + \frac{\partial \alpha}{\partial b} \dot{b} \right)\Big|_{\lambda=0}. $$
\noindent Here, derivatives with respect to $w,a,$ and $b$ can be evaluated using Eqs.~\ref{eq:ell_int_C},\ref{eq:ell_int_alpha0}, and \ref{eq:alpha_integral_soln}.
Once these two derivatives are calculated, they can be inserted into the Bernoulli equation to estimate the pressure on the leading surface of the vortex bubble.
\section{Derivation of Induced Velocity Scaling}
To describe the coupling of two simultaneous jets, we propose that the over-pressure at each nozzle is modified by the ``induced'' pressure, $p_{ind}\sim \rho u_{ind}^2,$ associated with the other jet's developing vortex ring. Here, we derive our scaling for the induced velocity, $u_{ind}\sim r^{-3}.$
\begin{figure*}[h]
\centering
\includegraphics[width=0.4\textwidth]{FigureS3.eps}
\caption{\label{fig:supp:InductionGeom} Geometry for induced velocity derivation.}
\end{figure*}
In 3D, a toroidal vortex ring with circulation $\Gamma$ and radius $a$ is described exactly by the axisymmetric stream function (See \citet{Lamb:1945tl}, Art.~161):
\begin{align}
\psi(r, z) &= - \frac{\Gamma}{2\pi}(r_1 + r_2)\left[K(\lambda) - E(\lambda) \right]
\end{align}
\noindent where
\begin{align*}
\lambda = \frac{r_2 - r_1}{r_2 + r_1}, \qquad
r_1^2 = z^2 + (r-a)^2, \quad \text{and} \quad
r_2^2 = z^2 + (r+a)^2
\end{align*}
\noindent As shown in Fig.~\ref{fig:supp:InductionGeom}, $r_1$ is the closest distance of the vortex ring to the field point $(r,z)$, and $r_2$ is the farthest distance. $K$ and $E$ are complete elliptic integrals of the first and second kind, respectively.
Since the two vortex rings form simultaneously, we expect that their mutual effects are dominant on the plane $z=0$. In this case, the induced velocity is antiparallel to $\mathbf{\hat{z}}$, $r_1=r-a$, $r_2=r+a$, and $\lambda = a/r = D/(2r).$ Differentiating the stream function yields a result for the velocity induced by a developing vortex ring:
\begin{align}\label{eq:supp:induced_vel}
\mathbf{u}_{ind}(r, z=0) &= \frac{1}{r} \frac{\partial \psi}{\partial r} \mathbf{\hat{z}} = - \mathbf{\hat{z}} \left(\frac{2 \Gamma}{\pi D}\right) \lambda \left[K\left(\lambda\right) - \frac{2-\lambda}{2(1-\lambda)}E\left(\lambda \right) \right] .
\end{align}
The complete elliptic integrals can be expanded in a power series about $\lambda=0$ \cite{Friedman:1971cl}:
\begin{align*}
K(\lambda) = \frac{\pi}{2} \left[ 1 + \frac{1}{4} \lambda^2 + O(\lambda^4)\right] \qquad
E(\lambda) = \frac{\pi}{2} \left[ 1 - \frac{1}{4} \lambda^2 - O(\lambda^4) \right].
\end{align*}
Inserting these expressions into Eq.~\ref{eq:supp:induced_vel}, expanding $u_{ind},$ and substituting $\lambda = D/(2r),$ the induced velocity can be written to leading order as
\begin{align}
\mathbf{u}_{ind} (r) \approx (-\mathbf{\hat{z}}) \left(\frac{\Gamma}{2 D} \right) \left(\frac{D}{r}\right)^{3}.
\end{align}
This leading order scaling is valid as long as $\lambda \ll 1$. As $\lambda \to 1,$ $u_{ind}$ diverges to $+\infty$ and higher order terms dominate. In terms of $r/D,$ the higher order terms should be negligible until $r/D \lesssim 1.$ In the regime of our experimental data, $r/D \gtrsim 1.7,$ the next term, of $O(r^{-4})$, is 3 times smaller than the leading order term. Hence we describe the physics of the two-jet coupling using the leading-order approximation, $u_{ind}\sim r^{-3}$.
\section{Additional Measurements}
\subsection{Wake Kinematics}
In our analysis, we asserted that the $z$-position and volume of the control volume grow linearly with time, so that higher order derivatives such as $\dot{u}_{cm}$ and $\ddot{V}_{cv}$ could be ignored. Here we provide typical data to support this claim. As shown in Fig.~\ref{fig:supp:CVMeasurements}, both position and volume are roughly linear across the time range measured. Small sinusoidal oscillations arise in the plots from a shear instability along the jet (visible in the videos). By fitting these quantities to a line, the oscillations are suppressed. While the wake position and volume are roughly linear over the experiment, the vortex bubble shape is not. The major and minor radius of the bounding ellipsoid are tracked during the experiments, and their values are used directly to compute the time-dependent pressure on the leading edge of the wake.
\subsection{Thrust}
In the main text, we showed only normalized thrust measurements, $T(\widetilde{\Delta})/T_\infty,$ but claim that the contribution from the pressure integral is an order of magnitude smaller than the contribution from unsteadiness in the wake. In Fig.~\ref{fig:supp:ThrustContrib}, we provide the data to support this claim.
\begin{figure*}[b]
\centering
\includegraphics[width=\textwidth]{FigureS4.eps}
\caption{\label{fig:supp:ThrustContrib} Contribution of the two different terms, unsteady and surface pressure, to the thrust measurement. Points represent the average of 5 experiments for each value of $\widetilde{\Delta}.$ The dotted lines represent the average of $5$ experiments for the single nozzle experiments - nozzle 2 is indicated by red and nozzle 1 is indicated by blue. The unsteady contribution is typically 10 times stronger than the pressure contribution.}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=0.65\textwidth]{FigureS5.eps}
\caption{\label{fig:supp:CVMeasurements} Results of CV tracking for select experiments. Solid lines represent the median curve from 5 experiments, and the shaded regions represent one standard deviation. As indicated by the small error bars, these experiments produced highly repeatable results.}
\end{figure*}
|
\section{INTRODUCTION}
Planets are expected to form in protoplanetary disks.
Since the disks affect the planet formation processes, it is important to understand the formation and evolution of protoplanetary disks.
Although there are many numerical simulations of the formation of protoplanetary disks (e.g. \cite{1998ApJ...508L..95B,2007ApJ...670.1198M,2010ApJ...714L..58T,2010ApJ...718L..58I,2011MNRAS.416..591T}; see also reviews by \cite{2012PTEP.2012aA307I,2016PASA...33...10T}), long term evolution of the disks is still unclear.
On the other hand, high-angular-resolution direct imaging of protoplanetary disks recently became available.
The observations have revealed that non-axisymmetric structure \cite[e.g.][]{2013Sci...340.1199V,2013Natur.493..191C,2013PASJ...65L..14F}, and ring like structures \cite[e.g.][]{2007A&A...469L..35G,2010ApJ...725.1735I,2012ApJ...747..136I,2013ApJ...775...30I,2011ApJ...732...42A,2011ApJ...729L..17H,2012ApJ...758L..19H,2012ApJ...753...59M,2012ApJ...760L..26M} are formed in protoplanetary disks.
These structures may provide clues to understand an evolution of protoplanetary disks and formation processes of planets.
Recently, the result from the high-resolution observation of the HL Tau region with ALMA was reported by \cite{2015ApJ...808L...3A}.
HL Tau is a well-known T Tauri star and has been extensively observed.
It is deeply embedded in circumstellar gas and observed as a reflection nebula at optical wavelength \cite[]{1995ApJ...449..888S}.
The disk-like infalling envelope whose radius is about 1400 AU and the accretion disk whose radius is about 150 AU are observed around HL Tau in $^{13}$CO and infrared emission \cite[]{1993ApJ...418L..71H,1997ApJ...478..766C}, and the outflow is also observed as the optical image \cite[]{1988ApJ...333L..69M}.
The ALMA observation of HL Tau has revealed that multiple ring structures have been formed in the disk around HL Tau \cite[]{2015ApJ...808L...3A}.
The observed ring structure of the brightness temperature and surface density structures estimated from the observation is shown in Figure 3 in \cite{2015ApJ...808L...3A} and Figure 3 in \cite{2016ApJ...816...25P}, respectively.
This stimulated extensive studies to find physical mechanisms to form the multiple ring structure of HL Tau; gap opening due to planets \cite[]{2015ApJ...806L..15K,2015MNRAS.453L..73D,2016ApJ...818..158A}, dust growth \cite[]{2015IAUGA..2256118Z}, baroclinic instability \cite[]{2015MNRAS.453L..78L}, and the growth and radial drift of dust grains with the effect of sintering \cite[]{2016ApJ...821...82O}, but formation mechanisms of such structures remain unclear.
Prior to the ALMA observation of HL Tau, \cite{2014ApJ...794...55T} (hereafter TI14) proposed a possible formation of multiple ring structures by secular gravitational instability \cite[secular GI, cf.][]{2000orem.book...75W, 2011ApJ...731...99Y, 2011ApJ...738...73S, 2010ApJ...719.1021M,2016ApJ...817..140S}.
Secular GI forms ring structures whose width is about 10AU in timescale $10^{4 \-- 5}$ yr at about 100AU.
Weak turbulence and enhanced dust-to-gas mass ratio are required for secular GI.
TI14 indicates that multiple ring structures may be an indicator of the dust concentration and weak turbulence in the disk.
Detailed analysis of the result of HL Tau observation reveals that weak turbulence (a turbulent viscosity coefficient $\alpha \sim$ a few $10^{-4}$) is required to reproduce the ALMA observation \cite[]{2016ApJ...816...25P}.
\cite{2016ApJ...816...25P} have estimated the parameter $\alpha$ from the dust scaleheight.
The dust scaleheight is given by the balance between the dust settling and diffusion by turbulence in the vertical direction.
Since the disk of HL Tau is inclined, the contrast of the dust continuum emission decreases on the minor axis when the dust scaleheight is large.
Thus, \cite{2016ApJ...816...25P} proposed that
small dust scaleheight that corresponds to $\alpha \approx 3\times 10^{-4}$ is required to reproduce the observed contrast on the minor axis of HL Tau disk\footnote{
\cite{2016ApJ...816...25P} has pointed out that $\alpha$ obtained from the dust scale height is different from that estimated from the mass accretion rate ($\alpha \sim 10^{-2}$).
This discrepancy may suggest that the disk of HL Tau is not steady, or the angular momentum of the disk is transferred by the mechanisms other than the turbulence, for example, magnetic braking \cite[]{2013ApJ...769...76B,2013ApJ...772...96B,2014ApJ...784..121S}.
When the disk is not steady, the accretion rate onto the central star can be different from the accretion rate at $r=100$~AU,
and when the angular momentum is transferred by the mechanisms other than the turbulence, the accretion rate does not reflect the strength of the turbulence.
Therefore, the parameter $\alpha$ estimated from the gas accretion rate is not necessarily related to the turbulence at $r=100$~AU.
On the other hand, $\alpha$ estimated from the dust scaleheight is directly related to the turbulence in the disk.
Since the turbulence that causes the dust diffusion stabilize the secular GI, we use $\alpha$ obtained from the dust scaleheight in this work.}.
Moreover, the disk of HL Tau is gravitationally unstable if we assume typical dust-to-gas mass ratio $\epsilon =0.01$ \cite[]{2011ApJ...741....3K,2015ApJ...808..102K}. When disks become gravitationally unstable, spiral arms are formed in the disk.
However, spiral structures are not observed in HL Tau.
Since observation only gives the dust surface density of the disk, gas surface density decreases as adopted dust-to-gas mass ratio increases.
Therefore, the observation suggests that the disk is gravitationally stable but dust-to-gas mass ratio is larger than typical value $\epsilon =0.01$.
These two features are consistent with what TI14 have anticipated.
Therefore, secular GI is a promising mechanism of ring structure formation in HL Tau.
In this work, we perform a linear stability analysis of secular GI adopting the physical values obtained from observation as the backgrounds state.
We obtain the most unstable wavelength and growth timescale of secular GI and discuss that secular GI can form the ring structures observed in HL Tau.
This paper is organized as follows. Basic equations for secular GI are given in Section \ref{basic_equations}.
In Section \ref{comparison}, we adopt the observational results of HL Tau as background values and derive the dispersion relation and the radial distribution of the maximum wavelength and the growth timescale of secular GI. We compare these results with the observation of ring structures of HL Tau.
We discuss
a possible unified scenario for the formation of both rings and planets
in HL Tau caused by secular GI in Section \ref{discussion}.
A summary is given in Section \ref{conclusion}.
\section{BASIC EQUATIONS}
\label{basic_equations}
In this section, we show basic equations for a linear stability analysis of secular GI.
We use two fluid equations for both gas and dust.
We focus on purely horizontal motions of gas and dust, and use vertically integrated equations of continuity and motion for both gas and dust, and the Poisson equation.
We take into account turbulent viscosity of gas, diffusion of dust caused by turbulence of gas, and velocity dispersion of dust.
\begin{equation}
\frac{\partial \Sigma}{\partial t} + \bm{\nabla}\cdot(\Sigma
\bm{u})=0,
\end{equation}
\begin{eqnarray}
\lefteqn{\Sigma \left(\frac{\partial u_i}{\partial t}+u_k \frac{\partial u_i}{\partial x_k} \right) =}
\nonumber \\
&-&c_s^2\frac{\partial \Sigma}{\partial x_i}-\Sigma\frac{\partial}{\partial x_i}\left(\Phi - \frac{GM_*}{r}\right) + \frac{\Sigma_d(v_i-u_i)}{t_{\rm stop}} \nonumber \\
&+&\frac{\partial}{\partial x_k} \left[\Sigma \nu \left(\frac{\partial u_i}{\partial x_k} +\frac{\partial u_k}{\partial x_i}
-\frac{2}{3}\delta_{ik}\frac{\partial u_l}{\partial x_l}\right)\right].
\end{eqnarray}
\begin{equation}
\frac{\partial \Sigma_d}{\partial t} + \bm{\nabla}\cdot(\Sigma_d
\bm{v})=D\nabla ^2\Sigma,
\end{equation}
\begin{eqnarray}
\lefteqn{\Sigma_{\rm d}\left(\frac{\partial \bm{v}}{\partial
t}+(\bm{v}\cdot\bm{\nabla})\bm{v}\right)} \nonumber\\
&=&-c_{\rm d}^2\nabla\Sigma_{\rm d}
-\Sigma_d\bm{\nabla}\left(\Phi
- \frac{GM_*}{r}\right)
+\frac{\Sigma_d(\bm{u}-\bm{v})}{t_{\rm stop}}.
\end{eqnarray}
\begin{equation}
\nabla^2\Phi=4\pi G(\Sigma+\Sigma_d)\delta(z),
\end{equation}
where $\Sigma$ and $ {\bm u}$ are surface density and velocity of gas,
$\Sigma_{\rm d}$ and ${\bm v}$ are surface density and velocity of dust,
$c_{\rm s}$ is the sound speed of gas,
$\nu$ is the coefficient of kinematic viscosity caused by turbulence,
$D$ is the diffusivity of the dust
due to the gas turbulence,
$c_{\rm d}$ is the velocity dispersion of the dust,
$M_*$ is the central star mass,
$\Phi$ is the gravitational potential of the gas and dust,
and
$t_{\rm stop}$ is the stopping time of a dust particle.
We adopt a local shearing box model
and the local radial and azimuthal coordinates are
$(x,y)$, which rotate with
the Keplerian frequency $\Omega$
\cite[e.g.][]{1965MNRAS.130..125G,1987MNRAS.228....1N}.
Since we investigate the ring structure formation, we assume axisymmetry for simplicity.
We assume a steady state background with uniform surface density;
$\Sigma_0 , \Sigma_{{\rm d}0} = {\rm const}$;
dust-to-gas mass ratio $\epsilon=\Sigma_{{\rm d}0}/\Sigma_0$;
and Keplerian rotation
$u_{x0}=v_{x0}=0,u_{y0}=v_{y0}=(-3/2)\Omega x$.
We decompose the physical quantities into background values and small
perturbations proportional to $\exp[ikx-i\omega t]$.
The linearized equations are given as follows:
\begin{equation}
-i\omega \delta \Sigma + ik \Sigma_0 \delta u_x=0,
\label{eq:eoc_gas_l}
\end{equation}
\begin{eqnarray}
-i\omega \delta u_x -2\Omega \delta u_y &=&
-c_{\rm s}^2\frac{ik\delta\Sigma}{\Sigma_0 } - ik \delta\Phi + \frac{\epsilon
(\delta v_x- \delta u_x)}{t_{\rm stop}} \nonumber\\
&&-\frac{4}{3}\nu k^2\delta u_x,
\label{eq:eom_gas_x_l_vis}
\end{eqnarray}
\begin{equation}
-i\omega \delta u_y+\frac{\Omega}{2} \delta u_x=
\frac{\epsilon (\delta v_y-\delta u_y)}{t_{\rm stop}}
-\nu k^2 \delta u_y -ik\frac{3\nu\Omega}{2\Sigma_0}\delta \Sigma.
\label{eq:eom_gas_y_l_vis}
\end{equation}
\begin{equation}
-i\omega \delta \Sigma_{\rm d} + ik \epsilon \Sigma_0 \delta v_x=-Dk^2 \delta \Sigma_{\rm d},
\label{eq:eoc_dust_l}
\end{equation}
\begin{equation}
-i\omega \delta v_x -2\Omega \delta v_y =
-c_{\rm d}^2 \frac{ik \delta \Sigma_{\rm d}}{\Sigma_0} -ik\delta \Phi
+ \frac{\delta u_x-\delta v_x}
{t_{\rm stop}}.
\label{eq:eom_dust_x_l_p}
\end{equation}
\begin{equation}
-i\omega \delta v_y +\frac{\Omega}{2} \delta v_x =
\frac{\delta u_y-\delta v_y}{t_{\rm stop}},
\label{eq:eom_dust_y_l}
\end{equation}
\begin{equation}
\delta \Phi=- \frac{2\pi G (\delta \Sigma+\delta \Sigma_{\rm d} )}{|k|}.
\label{eq:poisson_l}
\end{equation}
We take into account the effect of the thickness of the gas and dust on
the gravitational potential and rewrite Equation (\ref{eq:poisson_l})
\cite[]{1970ApJ...161...87V, 1984prin.conf..513S}:
\begin{equation}
\delta \Phi = - 2\pi G\left(\frac{\delta \Sigma}{1+kH}+
\frac{\delta \Sigma_{\rm d}}{1+ k H_{\rm d}}\right),
\label{eq:poisson_l_disk_height}
\end{equation}
where $H =c_{\rm s}/\Omega$ and $H_{\rm d}$ are the scale height of the gas and dust.
The relation between $H_{\rm d}$ and $H$ is given as follows
\cite{2007Icar..192..588Y}:
\begin{equation}
H_{\rm d} =
H\left(1+\frac{t_{\rm stop}\Omega}{\alpha}\frac{1+2t_{\rm stop}\Omega}{1+t_{\rm stop}\Omega}\right)
^{-1/2}.
\end{equation}
When $\alpha \ll t_{\rm stop}\Omega \ll 1$,
\begin{equation}
H_{\rm d} \simeq
\left(\frac{t_{\rm stop}\Omega}{\alpha}\right)
^{-1/2}H.
\end{equation}
The relation between the turbulent viscosity and dust diffusivity is given
by \cite{2007Icar..192..588Y} \cite[see also][]{2010ApJ...719.1021M}:
\begin{equation}
{\bar D} = \frac{1+t_{\rm stop}\Omega
+4(t_{\rm stop}\Omega)^2}{[1+(t_{\rm stop}\Omega)^2]^2}\alpha,
\end{equation}
where ${\bar D}\equiv D \Omega/c_{\rm s}^2$ is the normalized dust diffusivity
and $\alpha \equiv \nu \Omega/c_{\rm s}^2$ is the dimensionless measure of
turbulent intensity \cite[]{1973A&A....24..337S}.
In a turbulent disk, the velocity dispersion of dust is estimated by
\cite{2007Icar..192..588Y}:
\begin{equation}
c_{\rm d}^2 = \frac{1+2t_{\rm stop}\Omega +(5/4)(t_{\rm stop}\Omega)^2}{[1+(t_{\rm stop}\Omega)^2]^2} \alphac_{\rm s}^2.
\label{eq:cd}
\end{equation}
Hereafter we use the growth rate of the instability $n\equiv -i\omega$ instead of the frequency $\omega$.
For given values of $\epsilon$, $\alpha$, $t_{\rm stop}\Omega$, and Toomre parameter $Q\equiv c_{\rm s}\Omega/(\pi G \Sigma_0)$ \cite[]{1964ApJ...139.1217T}, we obtain normalized dispersion relations ($n/\Omega$ as a function of $kH$) from Equations (\ref{eq:eoc_gas_l}) to (\ref{eq:cd}).
\section{COMPARISON WITH THE OBSERVATION OF HL TAU}
\label{comparison}
In this section, we calculate the most unstable wavelength and the growth timescale of secular GI.
We adopt physical values obtained from the observation of HL Tau as the background values of linear stability analyses.
We assume that turbulence in HL Tau is not strong and adopt $\alpha = 3\times 10^{-4}$ as observationally suggested by \cite{2016ApJ...816...25P},
This small value of turbulence strength is also suggested by recent theoretical modeling of magnetorotational instability \cite[]{2016ApJ...817...52M}. We discuss this aspect in Section \ref{alpha_distribution}.
We use two models of dust surface density and temperature distributions proposed by previous work.
The gas surface density is not obtained from the observation.
However, if we assume dust-to-gas mass ratio $\epsilon=0.01$, the disk becomes gravitationally unstable since the dust surface density is expected to be very large to reproduce the observed values.
If the disk is gravitationally unstable and $Q\lesssim 2$ is satisfied, the spiral arms should be formed in the disk.
However, the observation of HL Tau shows that no spiral arm is formed in the disk.
In this work, we adopt constant $\epsilon$ for each models so that $Q > 2$ is satisfied in the disk.
The dust radius is also an important parameter since stopping time $t_{\rm stop}$ given by Epstein drag law is proportional to the dust radius: $t_{\rm stop} = \rho_{\rm int} a /(\rho_{\rm g}c_{\rm s})$, where $\rho_{\rm int}$ is the internal density of dust, $a$ is the dust radius, and $\rho_{\rm g}$ is gas density.
In \cite{2016ApJ...816...25P} the maximum and minimum dust sizes are 3~mm and 0.03~${\rm \mu m}$, and the power of the dust size distribution is -3.5 (integrated over the whole disk).
Since the dust mass of the disk is dominated by the largest dust, we simply adopt $a=3$~mm in this work\footnote{
\cite{2016ApJ...816...25P} have mentioned that the power of the dust size distribution in outer $\sim 80$ AU of the disk is about -4.5 and the mm size dust is depleted there.
Although we assume $a=3$~mm for simplicity in this work, further analysis of the observational data and investigation of secular GI including the effect of dust distribution are required \cite[see, e.g.][]{2016ApJ...817..140S}.
}.
The central star mass is obtained by \cite{2015ApJ...808L...3A}: $M_{\rm *} = 0.7 M_{\rm \odot}$.
\subsection{Exponentially Cutoff Disk Model}
\cite{2015ApJ...808..102K} obtain the dust surface density distribution by fitting CARMA data as follows:
\begin{equation}
\Sigma_{\rm d}(r)=0.51 \left(\frac{r}{R_{\rm c}}\right)^{-\gamma}
\exp\left[-\left(\frac{r}{R_{\rm c}}\right)^{2-\gamma}\right]
\ [\rm g \ cm^{-2}],
\label{eq:dust_surface_exp_cutoff}
\end{equation}
where $\gamma = -0.2$, $R_{\rm c} = 80.2$ AU.
\cite{2016ApJ...816...25P} shows that this dust surface density distribution approximately reproduces the ALMA observation data.
The temperature distribution at the midplane obtained in \cite{2015ApJ...808..102K} is about half of the brightness temperature distribution given by \cite{2015ApJ...808L...3A}.
Thus, we adopt the following temperature distribution:
\begin{equation}
T(r) = 30 \left(\frac{r}{20\ [\rm AU]}\right)^{-0.65}\ [\rm K].
\label{eq:Tr_ALMA}
\end{equation}
These dust surface density and temperature distribution give $\Sigma_{\rm d}\approx 0.11 \ {\rm g\ cm^{-2}}$, and $T\approx 11$ K at 100 AU. We assume $\epsilon = 0.02$ for this model.
Hereafter, we call this model ``exponentially cutoff disk model''.
In this case, the disk is gravitationally stable ($Q\approx 2.9$ at 100 AU).
We obtain $t_{\rm stop}\Omega \approx 0.42$ by using dust radius $a=3$ mm, $\rho_{\rm int}=3\,{\rm g\, cm^{-3}}$ and $\rho_{\rm gas} = \Sigma/(\sqrt{2\pi}H)$, where $\Sigma = \Sigma_{\rm d}/\epsilon $ is gas surface density.
Fig. \ref{fig:disp_exp_cut} shows the dispersion relation of secular GI at 100 AU.
\begin{figure}[t]
\epsscale{1}
\includegraphics[width=8cm]{./Fig1.eps}
\caption{
Dispersion relation of secular GI at 100 AU for exponentially cutoff disk model.
The horizontal axis is the normalized wavenumber, $kH$, and the vertical axis is the normalized growth rate of the instability, ${\rm Re}[n]/\Omega$.}
\label{fig:disp_exp_cut}
\end{figure}
The most unstable wavelength is about 11 AU. It agrees with the observation of HL Tau.
The growth timescale of secular GI,
$t_{\rm grow} = ({\rm Re}[n])^{-1}$, is about $9\times 10^4$ yr.
Since HL Tau is thought to be young, $t_{\rm grow} < 10^6$yr is required, but long growth timescale is required to observe the growing unstable mode.
In this sense, the result $t_{\rm grow} \sim 10^5 $ yr is consistent with the scenario that observed ring structures are formed by secular GI.
Fig. \ref{fig:lambda_tgrow_exp_cutoff} shows radial distributions of the most unstable wavelength and the growth timescale for exponentially cutoff disk model.
Secular GI grows in the region where the radius $r$ is larger than about 80 AU. The most unstable wavelength is about 10 AU and the growth time scale is about $10^5$ yr in this region.
Therefore, a few rings observed in the region $r \gtrsim 80$ AU can be created by secular GI.
In the region $r\lesssim 80$ AU, secular GI is stable but viscous overstability grows \cite[cf.][]{1995Icar..115..304S}.
However, since the growth timescale is larger than the typical disk lifetime and the most unstable wavelength is longer than the radius, viscous overstability does not grow in the disk. In this work, we neglect viscous overstability and focus on secular GI.
\begin{figure}[t]
\includegraphics[width=8cm]{Fig2-1.eps}
\includegraphics[width=8cm]{Fig2-2.eps}
\caption{The distribution of the most unstable wavelength (upper panel) and the growth timescale (lower panel) for exponentially cutoff disk model.
The horizontal axis is the radial distance from the central star.
The vertical axis of the upper panel is the most unstable wavelength and the lower panel is the growth timescale.
Secular GI grows in the region $r\gtrsim 80$ AU.
In the region $r\lesssim 80$ AU, secular GI does not grow but viscous overstability grows.
}
\label{fig:lambda_tgrow_exp_cutoff}
\end{figure}
It is difficult to obtain the dust radius and the surface density of gas from the observation.
Thus, there is large uncertainty in the model parameters $a$ and $\epsilon$.
Fig. \ref{fig:e-a-tgrow} shows the growth timescale of the instability for various $\epsilon$ and $a$.
Secular GI grows in the region shown by yellow and red.
This figure shows that $a\gtrsim 2$ mm and $\epsilon \lesssim 0.03$ are required for secular GI in this model.
For $\epsilon \lesssim 4\times 10^{-3}$, gravitational instability of the gas disk grows. The criterion $\epsilon = 4\times 10^{-3}$ corresponds to $Q\approx 0.6$, which is smaller than the critical $Q$ value for the razor-thin disk because of the disk thickness.
In Section \ref{discussion}, we discuss a scenario in which all the rings in the disk can be created by secular GI
\begin{figure}[t]
\includegraphics[width=8cm]{Fig3.eps}
\caption{
Growth timescale of the instability for various $\epsilon$ and $a$.
The horizontal axis is the dust-to-gas mass ratio $\epsilon$ and the vertical axis is dust radius $a$.
Secular GI grows in the region shown by yellow and red.
For $\epsilon \lesssim 4\times 10^{-3}$, gravitational instability of the gas disk grows.
}
\label{fig:e-a-tgrow}
\end{figure}
\subsection{Power-law Disk Model}
Since the dust opacity depends on the details of the dust grain property, for example the shape of dust, the physical values obtained from the observation should have some uncertainties.
Therefore, it is worth investigating the instability with various disk models.
For this purpose, we used the surface density and temperature distributions given by \cite{2016ApJ...820...54K} as an unperturbed state in this section.
In \cite{2016ApJ...820...54K}, ALMA Band 6 continuum along the major axis is fitted by a power-law temperature and dust surface density profile with depression (gaps) at the gap radius.
This model is obtained from only the observed intensity of Band 6 of ALMA and does not fit Band 3 and 7 data.
Although this model may not be adequate for the surface density and the temperature profile of HL Tau,
we use this model without depression as background state to investigate whether secular GI can grow in another background state.
The dust surface density without depression is given by Equation (4) in \cite{2016ApJ...820...54K}: $\Sigma_d = 8.3 (r/1[\rm AU])^{-0.3} {\rm \ g\ cm^{-2}}$.
This surface density represents the observed intensity at the bright rings.
In this work, however, we have to use the surface density before the ring formation as a background state of secular GI.
Thus, we overestimate the dust surface density when we adopt $\Sigma_d = 8.3 (r/1[\rm AU])^{-0.3} {\rm \ g\ cm^{-2}}$ as a background state, because this dust surface density should correspond to the maximum dust surface density after secular GI grows and concentrates the dust.
To avoid overestimating the dust surface density, we halve the dust surface density:
\begin{equation}
\Sigma_{\rm d} = 4.15 \left(\frac{r}{1\ [\rm AU]}\right)^{-0.3} [\rm g\ cm^{-2}]\label{eq:sigma_d_prof}
\end{equation}
\begin{equation}
T = 280 \left(\frac{r}{1\ [\rm AU]}\right)^{-0.3} [\rm K]\label{eq:T_prof}
\end{equation}
We assume $\epsilon = 0.1$ for this model.
Hereafter, we call this model ``power-law disk model''.
In this case, other parameters are given by $Q\approx 3.8$ and $t_{\rm stop}\Omega \approx 0.22$ at 100 AU.
Fig. \ref{fig:disp} shows dispersion relation of Secular GI at 100AU.
\begin{figure}[t]
\includegraphics[width=8cm]{Fig4.eps}
\caption{
Dispersion relation of secular GI at 100 AU for power-law disk model.
The horizontal axis is the normalized wavenumber, $kH$ and the vertical axis is the normalized growth rate of the instability, ${\rm Re}[n]/\Omega$.
}
\label{fig:disp}
\end{figure}
The most unstable wavelength is about 16 AU and growth timescale are about $9 \times 10^3$ yr.
Thus, ring structures formed by secular GI also agree with observation when we adopt the power-law disk model.
Fig. \ref{fig:r-lambda} shows the radial distribution of the most unstable wavelength and growth timescale for power-law disk model.
Secular GI grows in the region $r \gtrsim 80$ AU but cannot grow in $r \lesssim 80$ AU.
Therefore, either in the exponentially cutoff or power-law disk model, rings in the region $r\lesssim 80$ AU are not supposed to be created by in-situ secular GI.
We need other disk models to reproduce all ring structures in the disk by secular GI, which are discussed in a following section.
\begin{figure}[t]
\includegraphics[width=8cm]{Fig5-1.eps}
\includegraphics[width=8cm]{Fig5-2.eps}
\caption{The distribution of the most unstable wavelength (upper panel) and the growth timescale (lower panel) for power-law disk model.
Secular GI grows in the region $r\gtrsim 80$ AU.
}
\label{fig:r-lambda}
\end{figure}
\section{DISCUSSION}
\label{discussion}
\subsection{Strength of Turbulence in the Disk}
\label{alpha_distribution}
One of the most interesting features of HL Tau is that turbulence seems to be weaker than that usually expected.
The ALMA observation of HL Tau indicates that the coefficient of turbulent viscosity $\alpha$ is of order $10^{-4}$ in the disk.
However,
$\alpha$ due to MRI turbulence is usually expected to be large ($\sim 10^{-2}$) \cite[e.g.,][]{1995ApJ...440..742H,2004ApJ...605..321S}.
One candidate of a physical mechanism to suppress the MRI turbulence is electron heating introduced by \cite{2015ApJ...800...47O}. \cite{2016ApJ...817...52M} suggests that MRI is suppressed by electron heating and $\alpha \lesssim 10^{-3}$ can be realized in protoplanetary disks.
When electron heating suppresses the MRI turbulence, $\alpha$ is expected to be small in inner region of the disk.
Figure 9 in \cite{2016ApJ...817...52M} suggests $d\log\alpha/d\log r \gtrsim 2$ at the midplane.
If $\alpha$ is smaller than $3\times 10^{-4}$ in the region $r\lesssim 80$ AU, the region where secular GI grows is larger than that obtained in Section \ref{comparison}.
In this section, we discuss the secular GI in the disk in which $\alpha$ increases with radius as expected from electron heating in contrast to the previous section in which we used the constant $\alpha$ disk model following \cite{2016ApJ...816...25P}.
We try to model small $\alpha$ in inner region by the following ad hoc formula;
\begin{equation}
\alpha_1 = 3\times 10^{-4}\left(\frac{r}{100~{\rm AU}}\right)^2,
\end{equation}
\begin{equation}
\alpha_2= 3\times10^{-4}\left(\frac{r}{100~{\rm AU}}\right)^3.
\end{equation}
Fig. \ref{fig:lambda_tgrow_alpha2} shows the most unstable wavelength and growth timescale for exponentially cutoff and power-law disk model with $\alpha=\alpha_1 =3\times10^{-4}(r/100\,[\rm AU])^2$.
\begin{figure}[t]
\includegraphics[width=8cm]{Fig6-1.eps}
\includegraphics[width=8cm]{Fig6-2.eps}
\caption{The distribution of the most unstable wavelength (upper panel) and the growth timescale (lower panel) for exponentially cutoff disk model (solid line) and power-law disk model (dotted line) with $\alpha = 3\times 10^{-4} (r/100\,[\rm AU])^2$.
Secular GI grows in the region $r\gtrsim 40$ AU for exponentially cutoff disk model and $r\gtrsim 50$ AU for power-law disk model.
}
\label{fig:lambda_tgrow_alpha2}
\end{figure}
Fig. \ref{fig:lambda_tgrow_alpha2} shows that secular GI grows in the region $r\gtrsim 50$ AU, the growth timescale is a few $10^4$ yr and most unstable wavelength is about 10 AU for both models.
As a result, most of ring structures can be created by secular GI.
Moreover, in the case of exponentially cutoff disk model, secular GI grows even in the region around $\sim 40$ AU.
The most unstable wavelength is a few AU and smaller than the observed ring width.
We need detailed analysis of nonlinear growth of secular GI, but this result suggests that the rings whose width is a few AU are formed around $\sim 40$ AU and would be observed by future observations with higher spatial resolution.
Fig. \ref{fig:lambda_tgrow_alpha3} shows the radial distributions of the most unstable wavelength and the growth timescale with $\alpha =\alpha_2= 3\times 10^{-4}(r/100\,[\rm AU])^3$.
\begin{figure}[t]
\includegraphics[width=8cm]{Fig7-1.eps}
\includegraphics[width=8cm]{Fig7-2.eps}
\caption{
The distribution of the most unstable wavelength (upper panel) and the growth timescale (lower panel) for exponentially cutoff disk model (solid line) and power-law disk model (dotted line) with $\alpha = 3\times 10^{-4} (r/100\,[\rm AU])^3$. The secular GI grows in the whole disk.
}
\label{fig:lambda_tgrow_alpha3}
\end{figure}
In this case, secular GI grows in the whole disk.
The growth timescale at the inner region ($r\lesssim 20$ AU) is about an order of magnitude smaller than that of the outer region ($r \sim 100$ AU).
Secular GI, hence, grows faster in the inner region than in the outer region.
The rapid growth of the secular GI in the inner region may cause the planet formation discussed in Section \ref{PlanetFormation}.
Although MRI turbulence in the midplane might be suppressed by electron heating, turbulent motion can be large in the upper region of the disk.
In this case, vertical structure of the disk may affect secular GI and changes the dispersion relation.
In this work, we use vertically integrated equations for linear stability analysis.
To obtain more realistic dispersion relation we should take into account the vertical stratification of gas and dust components.
In general it inevitably includes vertical shear of azimuthal velocity that causes Kelvin-Helmholtz instability (KHI).
A detailed linear stability analysis of the KHI for two component system of gas and dust grains has been done by \cite{2006ApJ...641.1131M} that shows suppressed growth rates of KHI for larger dust grains.
If the non-linear state of the KHI corresponds to a weak turbulence ($\alpha \lesssim 3\times 10^{-4} (r/100 {\rm [AU]})^3$), such effects of vertical stratification may not significantly affect the conclusion in this section.
To clarify the actual outcome, however, we have to do three-dimensional non-linear simulations for such a multi-component system, which is beyond the scope of this paper.
In this section, we discuss the growth of secular GI in the disk in which $\alpha$ changes with radius.
The region where secular GI is expected expands with decreasing $\alpha$.
When $\alpha$ is smaller than $\alpha_2$, the evolution of the disk does not change qualitatively.
\subsection{The Effect of Radial Drift of Dust}
\label{RadialDrift}
In this work, we adopt isothermal and uniform surface density as a background of the linear stability analysis.
As a result, the rotation velocity of gas and dust is the same in a steady state background.
In reality, however, the surface density and temperature are not uniform but depend on radius.
In general, the rotational velocity of the gas decreases because the gas is supported not only by the centrifugal force but also by the pressure gradient force against the gravity.
On the other hand, dust particles rotate with Keplerian velocity even in a nonuniform background.
As a consequence, the dust encounters the headwind, loses angular momentum, and drifts inward.
The drift timescale $t_{\rm dri} = r/v_{\rm dri}$ is $\sim 2\times 10^4$ yr for exponentially cutoff disk model and $\sim 10^4$ yr for power-law disk model.
These timescales are comparable to or smaller than the growth timescales of secular GI.
Therefore, the radial drift of dust will affect the growth of secular GI.
If the dust drift into the region where secular GI cannot grow before sufficient growth of the perturbations, it is difficult to explain the ring structure formation with secular GI.
If the disk extends larger than 100 AU, secular GI may grow with the radial drift in the outer region of the disk.
In this case, secular GI makes pressure bumps, and they prevent the fast radial drift and will support the further growth of secular GI.
The ring structures formed in the outer region ($ r \gtrsim 80 $ AU) move inward with the disk accretion.
If the rings are not destroyed in the secular GI stable region, ring structures can be formed by secular-GI even in the model that secular GI grows only in the outer region shown in Section \ref{comparison}.
This scenario naturally explains the observational result that the separations of all the rings appear to be comparable to the most unstable wavelength of secular GI around the disk radius of 100 AU.
This scenario may predict the existence of the extended outer disk with a much fainter ring-like structure, which can be tested by future ALMA observations.
If the dust radial drift is strong enough, we should also consider the effect of streaming instability \cite[SI,][]{2007ApJ...662..613Y,2005ApJ...620..459Y,2007ApJ...662..627J}.
Even if the non-linear growth of SI results in turbulence, the degree of turbulent mixing should be smaller than $\alpha = 3\times10^{-4}$,
because the value of $\alpha = 3\times10^{-4}$ is inferred from the observation of HL Tau.
Therefore, in the case of HL Tau, SI does not change our conclusion drastically.
Other mechanisms to create ring-like structures have been proposed after the observation of HL Tau.
Actually some of them are not mutually exclusive but may possibly operate with secular GI.
\cite{2016ApJ...821...82O} have proposed ring structure formation by aggregate sintering.
Because sintering suppresses dust growth, the dust radius gets smaller in the sintering zone.
As a result, the radial drift velocity is small in the sintering zone and dust piles up there.
Obviously the dust condensation and piling up may positively affect the growth of secular GI.
Dust condensation due to sintering can support the growth of secular GI.
Therefore, the ring structure of HL Tau may be formed by both secular GI and aggregate sintering.
\cite{2015IAUGA..2256118Z} have proposed the ring structure formation by dust growth around the condensation front.
They have not taken into account the radial drift of dust, but the dust condensation may occur in their scenario in the same manner as in \cite{2016ApJ...821...82O}.
Therefore, this also possibly supports the present scenario.
The numerical simulation of secular GI that include these effects may enable the analysis on the interplay between these mechanisms, which will be our future work.
\subsection{Planet Formation by Secular GI}
\label{PlanetFormation}
\begin{figure}[t]
\includegraphics[width=8cm]{Fig8.eps}
\vspace{-2cm}
\caption{Schematic picture of inner gap (D1 and D2 in \cite{2015ApJ...808L...3A}) formation by secular GI.
Dust condensation occurs in the ring structures by secular GI.
The dust grows in the ring and results in the formation of a planet. Finally, the planet opens the gap in the inner region.
}
\label{fig:Gapformation}
\end{figure}
Since secular GI concentrates the dust, planetesimals could be formed in rings.
Thus, we expect that planet formation might be promoted as a result of secular GI.
The growth timescales of secular GI in the outer regions ($\sim 100$AU) are comparable to the estimated age of the central star and the disk.
Therefore natural interpretation is that the outer region is in the early growth phase of secular GI.
In contrast to the outer region, however, the inner region may have gone to the non-linear stages of secular GI because the growth timescales of secular GI in the inner regions are smaller than those in the outer regions.
Therefore, we speculate that the non-linear growth of secular GI may have resulted in the formation of planets that make gaps observed at $r\sim 10$ AU and $\sim 30$ AU (Fig. \ref{fig:Gapformation}).
The amounts of missing dust in the inner two gaps are estimated by \cite{2016ApJ...816...25P}: $\sim 7.2M_\oplus$ for the inner-most gap (D1 in \cite{2015ApJ...808L...3A}) and $\sim 21.6 M_\oplus$ for the next gap (D2).
We can simply expect the missing dust mass in the gap corresponds to the mass of heavy elements in the hypothetical planet in the gap.
Since we adopt the dust-to-gas mass ratio $\epsilon =0.02$ for our disk model that causes the secular GI, we should expect that the actual total mass of the planet in the gap can be about 50 times larger than the heavy element mass of the planet, if all the gaseous components of the disk in the gap accrete onto the planet.
This argument means that we set the expected ranges of masses of the planets in the inner two gaps are $7.2 M_\oplus \-- 1.1 M_{\rm J}$ and $21.6M_\oplus \-- 3.4 M_{\rm J}$, for D1 and D2, respectively.
When dust grains grow sufficiently in a high density region, the opacity of the ring decreases and the dust thermal radiation may remain weak in the ring.
Even after the formation of planetesimals or protoplanets, the ring may remain as an invisible planetesimal belt.
In this scenario, we expect to have a chance to observe the outcome of planet formation only after the mass of a planet increases sufficiently.
On the other hand, various authors estimated the masses of the hypothetical planets in the gaps by assuming the structure in the disk is created by hypothetical planets in the gap.
For example, \cite{2015ApJ...806L..15K} provided an approximate relation of the gap depth/width and the planet mass.
According to their relation we can obtain the hypothetical planet mass in the gap: $\sim 0.6M_{\rm J}$ for D1 and $\sim 1M_{\rm J}$ for D2.
This may mean that gaseous components in the disk do not perfectly accrete onto the planet in the gap.
Despite the simplification in our argument, the-order-of-magnitude agreement between the maximum planet mass expected from the formation mechanism and the mass estimated from the resultant gap structure is reasonable and may motivate further investigation in this line of research.
One of the differences between the present paper and other papers on HL Tau is in the connection of planet formation mechanism to the ring structure in the disk.
Papers on gap opening dynamics simply assumed the existence of planets and did not discussed their origins \cite[e.g.,][]{2015ApJ...806L..15K,2015MNRAS.453L..73D,2016ApJ...818...76J}.
In contrast, other papers on the mechanism to create multiple rings without embedded planets did not directly link the ring formation to mechanisms of planet formation \cite[e.g.,][]{2015IAUGA..2256118Z,2016ApJ...821...82O} partly because the ring formation in the disk is regarded as a transient phenomenon \cite[]{2016ApJ...821...82O}.
To theoretically verify the discussion in this section, we need to investigate the nonlinear growth of secular GI in the realistic situation taking into account the dust growth and radial drift.
\section{SUMMARY}
In this work, we perform linear stability analyses of secular GI adopting the physical values obtained from observation of HL Tau as background state.
We calculated the radial distributions of the most unstable wavelength and the growth timescale of secular GI and verify that secular GI can form the ring structures observed in HL Tau.
We use two background states: exponentially cutoff disk model and power-law disk model.
Both models give similar results.
We showed that secular GI grows in the outer region $r\gtrsim 80$ AU when we adopt the dust surface density, temperature, and a turbulent viscosity coefficient obtained from the observation.
Around a radius of 100AU, the most unstable wavelength is about 10 AU, and growth timescale is about $10^5$ yr and $10^4$ yr for exponentially cutoff and power-law disk model respectively.
The wavelength is consistent with the observed separations of the rings.
When we adopt a turbulent viscosity coefficient $\alpha =3\times 10^{-4}(r/100\, [\rm AU])^2$, secular GI grows in the region $r \gtrsim 50 $ AU.
Thus, secular GI forms all rings whose width is about 10 AU.
Moreover, in the case that we adopt a turbulent viscosity coefficient $\alpha =3\times 10^{-4}(r/100\, [\rm AU])^3$, secular GI grows in the whole disk.
Since the growth timescale in the inner region $r\lesssim 30$ AU is about an order of magnitude smaller than that of in the outer region $r\sim 100$ AU, secular GI can rapidly grow in the inner region.
If the rings formed in the outer region move inward with the disk accretion, secular GI can naturally form the equi-separation rings if the separations remain the same in the accretion process.
Since secular GI concentrate the dust, it supports the dust growth.
Because the growth timescale of secular GI in the inner region may be smaller than the estimated age of HL Tau,
plant could be formed in rings in the inner region.
In this case, resulting objects are expected to create the gaps at $r\sim 10$ AU and $\sim $ 30 AU.
If the scenario described in this article is indeed realized in HL Tau, it means that we are witnessing both the early stage of planet formation and the stage after the planet formation: early growth of the most unstable mode of secular GI and the gap structure carved by resulting planets.
Therefore, further observation of HL Tau would give more detailed information for various stages of planet formation.
\label{conclusion}
SI thanks Neal Turner, Andrew Youdin, Anders Johansen, and Wladimir Lyra
for visiting him and discussing on various issues for protoplanetary disks.
The authors also thank Hiroshi Kobayashi, Takayuki Muto,
Misato Fukagawa, Munetake Momose, Kazuyuki Omukai, and Shigeo Kimura.
This work was partly supported by a Grant-in-Aid for JSPS Fellows and JSPS KAKENHI Grant Numbers 23244027 and 23103005.
\bibliographystyle{apj}
|
\section{Introduction}
Standard inflation driven by an inflaton field traces back to early
efforts to alleviate the basic problems of the Big-Bang cosmology, namely
horizon, flatness and
monopoles \cite{Guth:1980zm, Albrecht:1982wi}.
The nominal inflationary paradigm contains the slow-roll and the
(P)reheating regimes.
In the slow-roll phase the kinetic energy (which has the canonical form here)
of the scalar field is negligible with respect to the potential energy $V(\phi)$
which implies a deSitter expansion of the Universe.
However, after the slow-roll epoch the kinetic energy
becomes comparable to the potential energy and thus
the inflaton field oscillates around the
minimum and progressively the universe is filled by
radiation \cite{Shtanov:1994ce,Kofman:1997yn}.
Nevertheless, other theoretical patterns suggested
a possible way to treat the physics of the early universe.
For example, in the so-called warm inflationary
scenario the radiation production
occurs during the slow-roll epoch and the
reheating period is avoided \cite{Berera:1995ie, Berera:1996fm}.
The nature of the warm inflationary scenario
is different with respect to that of the
standard cold inflation.
Warm inflation satisfies
the condition $T>H$, where $T$ is the temperature
and $H$ is the Hubble parameter, which implies that
the fluctuations of the inflaton field are thermal instead of quantum.
An obvious consequence of the above inequality is that in the case of warm inflation
density perturbations arise from thermal
fluctuations rather than quantum fluctuations
\cite{Hall:2003zp, Moss:1985wn, Berera:1999ws}.
Specifically, thermal fluctuations are
produced during the warm inflationary epoch and they play
a central role toward describing the CMB anisotropies and thus
providing the initial seeds for the formation of large scale structures.
Of course, after this epoch the universe enters in the radiation
dominated phase as it should \cite{Berera:1995ie, Berera:1996fm}.
In order to achieve warm inflation one may use a tachyon scalar field
for which the kinetic term does not follow the canonical
form (k-inflation \cite{ArmendarizPicon:1999rj}).
It has been found that tachyon fields which are associated with
unstable D-branes \cite{Sen:2002nu} can be responsible
for the cosmic acceleration in early times
\cite{Sen:2002an, Sami:2002fs, ArmendarizPicon:1999rj}.
Notice, that tachyon potentials have the following two properties:
the maximum of the potential occurs when
$\phi\rightarrow 0$ while the corresponding minimum takes place
when $\phi\rightarrow \infty$.
From the dynamical viewpoint one may obtain the equations of motion
using a special Lagrangian
\cite{Gibbons:2002md} which is non-minimally coupled to gravity:
\begin{equation}\label{Lag}
L=\sqrt{-g}\left[\frac{R}{16\pi G}-V(\phi)\sqrt{1-g^{\mu\nu}\partial_{\mu}\phi\partial_{\nu}\phi}\right]\;.
\end{equation}
Considering a spatially flat
Friedmann-Robertson-Walker (hereafter FRW) space-time
the stress-energy tensor is given by
\begin{equation}\label{1.1}
T^{\mu}_{\nu}=\frac{\partial
L}{\partial(\partial_{\mu}\phi)}\partial_{\nu}\phi-g^{\mu}_{\nu}L={\rm diag}(-\rho_{\phi},p_{\phi},p_{\phi},p_{\phi})
\end{equation}{equation}
where $\rho_{\phi}$ and $p_{\phi}$
are the energy density and pressure of the scalar field.
Combining the
above set of equations one can derive
\begin{equation}\label{1.2}
\rho_{\phi}=\frac{V(\phi)}{\sqrt{1-\dot{\phi}^2}}
\end{equation}
and
\begin{equation}\label{1.22}
P_{\phi}=-V(\phi)\sqrt{1-\dot{\phi}^2}
\end{equation}
Where $\phi$ is tachyon scalar field in unite of inverse Planck mass $M_{pl}^{-1}$, and $V(\phi)$ is potential associated with the tachyon field.
In the past few years, there was an intense debate among
cosmologists and particle physicists regarding those
phenomenological models which can be produced in extra dimensions.
For example, the reduction of higher-dimensional
gravitational scale, down to
TeV-scale, could be presented by an
extra dimensional scenario \cite{ArkaniHamed:1998rs,ArkaniHamed:1998nn,Antoniadis:1998ig}. In these scenarios, gravity field
propagates in the bulk while
standard models of particles are confined to the lower-dimensional brane.
In this framework, the extra dimension induces additional terms
in the first Friedmann equation \cite{Binetruy:1999ut,Binetruy:1999hy,Shiromizu:1999wj}.
Especially, if we consider a quadratic term in the energy density
then we can extract an accelerated expansion of the early universe \cite{Maartens:1999hf,Cline:1999ts,Csaki:1999jh,Ida:1999ui,Mohapatra:2000cm}.
In the current study we consider the
tachyon warm inflation model in the framework of
Randall-Sundrum II braneworld which contains
a single, positive tension brane and a non-compact extra dimension.
Following the lines of Ref.\cite{Herrera:2015aja}, we attempt
to study the main properties of the warm inflation
in which the scale factor evolves as $a(t)\propto \exp(At^f)$, where $0<f<1$
("intermediate inflation").
In this case cosmic expansion evolves faster than the power-law inflation
($a \propto t^{p}$, $p>1$)
and slower than the standard deSitter one, $a(t)\propto {\rm exp}(H_{I}t)$
[$H(t)=H_{I}=$const.]. More details regarding the cosmic expansion
in various inflationary solutions can be found in the paper of
Barrow \cite{Barrow:1996bd}.
In the current work, we investigate the possibility
of using the intermediate solution in the case of
warm tachyon inflation. Specifically, the structure of
the article is as follows: In section II we briefly
discuss the main properties of the
warm inflation, while in section III we provide the slow roll parameters.
In section IV we study the
performance of our predictions
against the \textit{Planck 2015} data.
Finally, the main conclusions are
presented in section VI.
\section{Tachyon warm inflation}\label{setup}
Let us assume a flat, homogeneous and isotropic
Friedmann-Robertson-Walker (FRW) universe,
in which the radiation era is endowed with the scalar field
described by the Lagrangian (\ref{Lag})
in the context
of the Randall-Sundrum II (RSII)
brane \cite{Randall:1999vf}
.
Following the notations of
\cite{Shiromizu:1999wj, Binetruy:1999ut, Binetruy:1999hy,Berera:1995ie, Berera:1996fm, Herrera:2015aja}
one may check that
the basic cosmological equations
are
\begin{equation}\label{2.3}
H^2=\frac{8\pi}{3M_{pl}^2}\rho(1+\frac{\rho}{2\lambda})
\end{equation}{equation}
\begin{equation}\label{denn}
{\dot \rho}_{\phi}+3H(\rho_{\phi}+p_{\phi})=-\Gamma{\dot \phi}^2
\end{equation}
\begin{equation}\label{denn1}
{\dot \rho}_{\gamma}+3H(\rho_{\gamma}+p_{\gamma})=\Gamma {\dot \phi}^2
\end{equation}
where $\Gamma$ is the dissipation coefficient, in unit of $M_{pl}^5$.
The latter two equations (\ref{denn}),(\ref{denn1}) imply
the continuity equation, namely ${\dot \rho}+3H(\rho+p)=0$.
Notice, that Eqs.(\ref{denn}) and (\ref{denn1})
have been proposed by various authors such as
\cite{Cai:2010wt,Herrera:2006ck,delCampo:2008fc,Setare:2012fg,Setare:2013dd,Zhang:2013waa}. In these studies the quantity $\Gamma \dot{\phi}^2$ is the dissipation term
which is introduced phenomenologically in order to describe the
nearly-thermal radiation bath that is the outcome of the warm
inflationary scenario.
It is well known that
Tachyon inflation in its standard picture (cold inflation)
suffers from a serious problem. In particular,
reheating and matter creation are both problematic because the tachyon fields
in such theories do not oscillate around the minimum of the potential
\cite{Kofman:2002rh}.
This problem can be alleviated in the context of warm inflation.
In this scenario radiation production occurs during the slow-roll
era which implies that reheating is avoided and thus the universe
heats up and finally it enters in the radiation era \cite{Berera:1995ie} (See Eqs.(\ref{denn}) and (\ref{denn1})).
In the above set of equations, an over-dot denotes
derivative with respect to time,
$\rho=\rho_{\phi}+\rho_{\gamma}$ and $p=p_{\phi}+p_{\gamma}$
($p_{\gamma}=\rho_{\gamma}/3$) are the total density and pressure,
$\rho_{\phi}$ and $\rho_{\gamma}$ are the scalar field and radiation densities,
$H ={\dot {a}}/a$ is
the Hubble parameter.
Notice, that $\lambda$ is the brane tension which obeys the following
restriction $\lambda\geq (10TeV)^4$ \cite{Cline:1999ts, Brax:2003fv, Clifton:2011jh}.
Obviously, substituting
equations (\ref{1.2}),(\ref{1.22}) in Eq.(\ref{denn})
it is easy to derive the modified Klein-Gordon
equation which describes the time evolution
of the tachyon field. This is
\begin{equation} \label{E.O.M}
\frac{\ddot{\phi}}{1-\dot{\phi}^2}+3H\dot{\phi}+\frac{V'}{V}=
-\frac{\Gamma}{V}\dot{\phi}\sqrt{1-\dot{\phi}^2} \;,
\end{equation}
where $V^{\prime}(\phi)=dV/d\phi$.
The above cosmological equations imply that the model is
strongly affected by the quantity $\Gamma$.
This is due to the fact that radiation is exchanging energy
with the tachyon field and this is reflected in the corresponding behavior
of dissipation coefficient $\Gamma$, which is negligible in the classical
inflationary paradigm by definition.
Although, the precise functional form of $\Gamma$ is still
an open issue, a number of different parametrizations
have been proposed in the literature treating the functional
form of $\Gamma$ (see \cite{Hall:2003zp,BasteroGil:2012cm, Bastero-Gil:2014oga, Bartrum:2013fia}). In the current work
we use the well known parametrization of
\begin{equation}
\label{GAM}
\Gamma=\Gamma_{c}\phi^b T^c \;,
\end{equation}
where $T$ is the temperature and $\Gamma_{c}$ is constant.
During the warm inflationary epoch, the energy density of the scalar field
dominates the total fluid (stable regime \cite{BasteroGil:2012zr})
and thus Eq.(\ref{2.3}) becomes
\begin{eqnarray}
H^2=\frac{8\pi}{3M_{pl}^2}\rho_{\phi}(1+\frac{\rho_{\phi}}{2\lambda})
\end{eqnarray}
or
\begin{equation}\label{2.5}
H^{2}=\frac{8\pi}{3M_{pl}^2}\frac{V(\phi)}{\sqrt{1-\dot{\phi}^2}}
\left(1+\frac{V(\phi)}{2\lambda\sqrt{1-\dot{\phi}^2}}\right)\;.
\end{equation}
Another important quantity in this kind of
studies is the dimensionless dissipation parameter
which characterizes the type of inflation
and it is defined as (for more details see appendix)
\begin{eqnarray}\label{12}
R=\frac{\Gamma}{3H\rho_{\phi}}.
\end{eqnarray}
The above definition is presented for warm tachyon inflation in several papers \cite{Herrera:2006ck, Setare:2012fg, Setare:2013ula, Setare:2014gya, Setare:2014uja, Setare:2013dd}. Notice, that for the canonical scalar field model
of warm inflation, the corresponding
dimensionless ratio is defined as $\frac{\Gamma}{3H}$.
In the weak dissipation regime, the ratio $R$
tends to zero
($\Gamma/3H\rho_{\phi} \ll 1$), however, in
the strong dissipation regime, the coefficient $\Gamma$
guides the damped evolution of the scalar field.
Now using Eq.(\ref{E.O.M}) and Eq.(\ref{2.5})
in the high-dissipation
regime ($\Gamma\gg 3H\rho_{\phi}$)
to prove that
\begin{eqnarray}\label{2.7}
\dot{\phi}^2=-\frac{3M_{pl}^2}{4\pi}\frac{(H\dot{H})}{\Gamma}
\left(1+\frac{3M_{pl}^2H^2}{4\pi \lambda}\right)^{-\frac{1}{2}} \;.
\end{eqnarray}
Owing to the fact that during inflation the parameters $H$, $\Gamma$ and $\phi$
are slowly varying functions the production of radiation become
quasi-stable when $\dot{\rho}\ll 4H\rho_{\gamma},$ and $\dot{\rho}_{\gamma}\ll\Gamma\dot{\phi}^2$ \citep{Berera:1995ie, Berera:1996fm, Hall:2003zp}.
Under these conditions, using Eqs.(\ref{E.O.M}) and (\ref{2.7}) we
write the radiation density as follows:
\begin{eqnarray}\label{2.8}
\rho_{\gamma}=\frac{\Gamma \dot{\phi}^2}{4H}=-\frac{3M_{pl}^2 \dot{H}}{16\pi}
\left(1+\frac{3M_{pl}^2H^2}{4\pi\lambda}\right)^{-\frac{1}{2}}\;.
\end{eqnarray}
The latter formula can be identified with the equation relating $\rho_{\gamma}$
with the radiation temperature $T$.
Indeed, under of adiabatic condition we may write
\begin{equation}
\label{Kolb}
\rho_{\gamma}=C_{\gamma}T^4
\end{equation}
where $C_{\gamma}=\frac{\pi^2g_{*}}{30}$ and $g_{*}$
is the degrees of freedom of the created massless modes \citep{Kolb1990}.
Combining Eq.(\ref{2.8}) and (\ref{Kolb})
we obtain the temperature
\begin{eqnarray}\label{}
T=\left(-\frac{3M_{pl}^2 \dot{H}}{16\pi C_{\gamma}}\right)^{\frac{1}{4}}
\left(1+\frac{3M_{pl}^2 H^2}{4\pi\lambda}\right)^{-\frac{1}{8}} \;.
\end{eqnarray}
Lastly, with the aid of Eqs.(\ref{2.5}) and (\ref{2.8}) we obtain the
potential of the scalar field
\begin{eqnarray}\label{}
& V=\lambda\left[-1+\left(1+\frac{3M_{pl}^2H^2}{4\pi\lambda}\right)^{\frac{1}{2}}\right]\\
\nonumber
&\times\left(1+\frac{3M_{pl}^2}{4\pi}\frac{(H\dot{H})}{\Gamma}
\left[1+\frac{3M_{pl}^2H^2}{4\pi \lambda}\right]^{-\frac{1}{2}}\right)^{\frac{1}{2}} \;.
\end{eqnarray}
\section{Slow-roll parameters}
Let us present here the main quantities of the tachyonic inflation.
In particular, the basic slow-roll parameters are given by
\begin{equation}\label{epsilon}
\epsilon=-\frac{\dot{H}}{H^2}
\end{equation}{equation}
\begin{equation}\label{epsilon1}
\eta=-\frac{\ddot{H}}{2H\dot{H}} \;.
\end{equation}{equation}
In this context, the number of e-folds is written as
\begin{equation}
N=\int_{t}^{t_{end}}H dt
\label{efold}%
\end{equation}{equation}
where $t_{end}$ is the value of the cosmic time at the end of inflation,
namely $\epsilon(\phi_{end})\simeq 1$ where $\phi_{end}=\phi(t_{end})$.
Also, the power spectrum of the scalar fluctuations is given by
\citep{Berera:1995ie, Berera:1996fm}
\begin{equation}
\label{PP}
P_{s}=\frac{H^2}{\dot{\phi}^2}\delta\phi^2\;.
\end{equation}{equation}
An important feature of the warm inflationary model is related with the
fact that the origin of $\delta \phi$ is thermal and not quantum as
we consider in the nominal inflationary paradigm.
In the case of warm inflation
it has been found \citep{Berera:1995ie, Berera:1996fm, Hall:2003zp}
that scalar perturbations are written as
\begin{eqnarray}\label{2.12}
\delta\phi^2\simeq\frac{k_F T}{2 M_{pl}^4\pi^2} \;,
\end{eqnarray}
where the wave number
$k_F=\sqrt{\frac{\Gamma H}{V}}=H\sqrt{\frac{\Gamma}{3HV}}\geq H$
corresponds to the freeze-out scale at the special point when,
the dissipation damps out to thermally excited
fluctuations of inflaton
($\frac{V''}{V'}<\frac{\Gamma H}{V}$) \cite{Taylor:2000ze}.
Notice, that Eq.(\ref{2.12}) is valid in the
high-dissipation regime $R\gg 1$.
As we have already mentioned in the previous section
we study our model via Eq.(\ref{2.7}) in the context of
high-dissipation regime, which means that
for scalar perturbations we can utilize Eq.(\ref{2.12}).
Inserting the freeze-out wave-number and Eq.(\ref{2.12}) into
Eq.(\ref{PP}) we find after some simple calculations that the power spectrum of the tachyonic scalar field is given by
\begin{eqnarray}\label{scalar power spectrum}
P_{s}\simeq \frac{H^{\frac{5}{2}}\Gamma^{\frac{1}{2}} T}{2\pi^2 M_{pl}^4 V^{\frac{1}{2}}\dot{\phi}^2} \;.
\end{eqnarray}
We would like to point out that in the case of
canonical scalar fields within the framework of warm inflation
one can find other forms of the power spectrum $P_{s}$
in Refs.\citep{Bartrum:2013fia,Ramos:2013nsa,Bastero-Gil:2014jsa}.
The corresponding spectral index $n_{s}$ is defined in terms of the
$P_{s}$ slow-roll parameters, as usual \cite{LL}, by
\begin{eqnarray}\label{spectral index}
n_s-1=\frac{d\ln P_{s}}{d\ln k} \;.
\end{eqnarray}
On the other hand, it has been found \cite{Langlois:2000ns} that the
power spectrum of the tensor perturbations which are defined on the brane
takes the form
\begin{eqnarray}\label{tensor power spectra}
P_t=\frac{64\pi}{M_{pl}^2}\left(\frac{H}{2\pi}\right)^2
G^2(x)
\end{eqnarray}
where $x\equiv \left[\frac{3HM_{pl}^2}{4\pi\lambda}\right]^{\frac{1}{2}}$
and $G(x)=[\sqrt{1+x^2}-x^2\sinh^{-1}(\frac{1}{x})]^{-\frac{1}{2}}$ arises
from normalization of zero-mod of a graviton \cite{Langlois:2000ns}.
Therefore, using the so called tensor-to-scalar ratio
we arrive at
\begin{eqnarray}\label{Tensor to scalar ratio}
r=32\pi M_{pl}^2\frac{V^{\frac{1}{2}}\dot{\phi}^2}{H^{\frac{1}{2}}\Gamma^{\frac{1}{2}} T}G^2(x)\;.
\end{eqnarray}
In order to proceed with the analysis it would help to know the functional
form of the scale factor $a(t)$.
Barrow \cite{Barrow:1996bd} showed that under of specific conditions
we can have an intermediate inflation in which the scale factor
satisfies the following exponential form:
\begin{eqnarray}\label{scall}
a(t)=a_{I}\exp(At^f)\;,
\end{eqnarray}
which provides
\begin{eqnarray}\label{scal}
H(t)=\frac{{\dot a}(t)}{a(t)}=Aft^{f-1}\;,
\end{eqnarray}
where $f$ satisfies the restriction $0<f<1$.
The above expansion evolves faster than the power-law inflation
($a \propto t^{p}$, $p>1$)
and slower than the standard deSitter one, $a(t)\propto {\rm exp}(H_{I}t)$
[$H(t)=H_{I}=$const.].
Considering the functional form (\ref{GAM}),
one has to deal in general with the following
four parametrizations, which have been considered
within different approaches in the literature.
Depending on the values of $(b,c)$ we have:
(I)- the situation in which the formula is $\Gamma=\Gamma_3 T^3\phi^{-2}$,
$(b,c)=(3,-2)$.
The constant parameter $\Gamma_3$ corresponds to
$0.02 h^2 \mathcal{N}_{Y}$ where
there is generic supersymmetric (SUSY) model
with chiral superfields $X$, $\Phi$ and $Y_i=1,...\mathcal{N}_{Y}$.
This case is mostly used in the low temperature regime
where $m_{\chi}$ ($m_{\chi} $ is the mass of
catalyst field) \citep{Bastero-Gil:2014oga, Bastero-Gil:2013nja};
(II)- for $(b,c)=(2,-1)$ we have
$\Gamma=\Gamma_2\phi^2 T^{-1}$. This parametrization has been used
for non-SUSY models \cite{Berera:1998gx, Yokoyama:1998ju}.
Q5: (III)- the case where
$\Gamma=\Gamma_{0}$ is a positive
constant (hereafter $\Gamma_{0}$-parametrization:
see \cite{Herrera:2006ck, Deshamukhya:2009wc, Setare:2012fg, Setare:2013ula, Setare:2013qfa, Setare:2014gya, Setare:2014uja, Setare:2015cta, Setare:2013dd})
which implies that the pair $(b,c)$ in Eq.(\ref{GAM})
is strictly equal to $(0,0)$
and finally (IV)
we utilize the so called high temperature regime
(hereafter $\Gamma_{1}$-parametrization)
in which we select
$(b,c)=(0,1)$ and thus $\Gamma \propto T$
(see also \cite{Panotopoulos:2015qwa}).
In this paper, we are going to focus
on parameterizations (III) and (IV) in order
to calculate the slow roll parameters.
Lastly, we remind the reader that in the framework of warm inflationary
model thermal fluctuations dominate over the quantum fluctuations.
Combining the latter argument with the fact that
thermal fluctuations are proportional to temperature $T$ while
quantum fluctuations are proportional to $H$, one can easily derive the
condition $T>H$. Obviously,
if we consider our model in the high temperature regime ($\Gamma \propto T$)
then the aforesaid restriction ($T>H$) is satisfied.
For more details we refer the reader the work of \cite{Panotopoulos:2015qwa}.
\subsection{$\Gamma_{0}$-parametrization }\label{}
In this inflationary scenario ($\Gamma=\Gamma_{0}$=const.)
with the aid of Eq.(\ref{scal}) we integrate
Eq.(\ref{2.7}) and we obtain
the evolution of the scalar field in terms of the
hyper-geometric function \cite{Arfken,Abramowitz}
\begin{eqnarray}\label{sola}
\phi(t)-\phi_0=\frac{F(t)}{K}
\end{eqnarray}
where
\begin{eqnarray}\label{}
F(t)&=&t^{\frac{2f-1}{2}} \times \\ \nonumber &_{2}F_1&\left[\frac{1}{4},\frac{1-2f}{4(1-f)},\frac{5-6f}{4(1-f)},-\frac{3M_{pl}^2f^2A^2t^{2f-2}}{4\pi\lambda }\right] \;,
\end{eqnarray}
$$K=-\left(\frac{16\pi\Gamma_0(1-f)}{3M_{pl}^2f^2A^2}\right)^{\frac{1}{2}}\frac{\Gamma(\frac{5-6f}{4(1-f)})}{\Gamma(\frac{1-2f}{4(1-f)})}$$
and $\Gamma(n)$ is the normal
Gamma-function. Notice, that
without losing the generality we have set
$\phi_0=0$.
Now we can derive the Hubble parameter $H$ and the
associated potential $V(\phi)$
in the limit of $\dot{\phi}^2\ll V(\phi)$
\begin{eqnarray}\label{HHA}
&H(\phi)=fA(F^{-1}[K\phi])^{f-1}\\
\nonumber
&V(\phi)\simeq\lambda(-1+\sqrt{1+
\frac{3M_{pl}^2f^2A^2[F^{-1}(K\phi)]^{2f-2}}{4\pi\lambda}})
\end{eqnarray}
where $F^{-1}(\phi)$ is the inverse function of $F(t)$.
Clearly, if we substitute Eq.(\ref{HHA}) in the slow-roll parameters
then we have
\begin{equation}
\label{EEPS}
\epsilon=\frac{(1-f)t^{-f}}{fA}=
\frac{1-f}{fA[F^{-1}(K\phi)]^f}
\end{equation}
\begin{equation}
\label{EETA}
\eta=\frac{(2-f)t^{-f}}{2fA}=
\frac{2-f}{2fA[F^{-1}(K\phi)]^f} \;.
\end{equation}
Notice, that in order to extract the latter equalities in
Eqs(\ref{EEPS}), (\ref{EETA}) we used Eqs.(\ref{scal}), (\ref{sola}).
In our warm intermediate case the condition
$\epsilon=1$, insures the beginning of inflation
\cite{Barrow:2006dh, Barrow:1993zq}.
Therefore, utilizing
Eq.(\ref{efold}),
we can derive the number of e-folds\footnote{In the
literature sometimes we replace $\phi$ by $\phi_{\star}$ which denotes the value
at the horizon crossing.}
\begin{eqnarray}\label{number of efolds}
&N=\int_{t_{in}}^{t_{*}} Hdt=A(t_{*}^f-t_{in}^f)&\\
\nonumber
&=A\left( [F^{-1}(K\phi)]^f-[F^{-1}(K\phi_{in})]^f\right)& \;.
\end{eqnarray}
Plugging $\phi_{in}$ into Eq.(\ref{EEPS}) and using the
constraint $\epsilon(\phi_{in})\simeq 1$ we find
\begin{equation}
\label{ee1}
\phi_{in}=\frac{1}{K}F(y),\;\;\;\;\;
y=\left(\frac{1-f}{fA}\right)^{\frac{1}{f}} \;.
\end{equation}
In order to proceed with the analysis we need to know the values of $N$ and
$\phi_{in}$.
Firstly, it is natural to consider that the number of e-folds
is 50 or 60.
Secondly, using the condition $\epsilon(\phi_{in})=1$ and
Eqs. (\ref{number of efolds}), (\ref{ee1})
we can estimate the slow-roll parameters.
Now we focus on the power spectrum formulas.
Specifically, inserting the appropriate expressions into
Eq.(\ref{scalar power spectrum}) we define
the scalar power spectrum
\begin{eqnarray}\label{}
&P_{s}=p_1I(N)^{\frac{3f}{4}}\left[-1+\left(1+\frac{3M_{pl}^2f^2A^2}
{4\pi\lambda I(N)^{2-2f}}\right)^{\frac{1}{2}}\right]^{-\frac{1}{2}}\\
\nonumber
&\times \left(1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}}\right)^{\frac{3}{8}}
\end{eqnarray}
where $p_1=(\frac{\Gamma_0^6 f^3A^3}{3^3M_{pl}^6\pi^5(1-f)^3C_{\gamma}})^{\frac{1}{4}}$ and $I(N)=\left[\frac{1+f(N-1)}{fA}\right]^{\frac{1}{f}}$.
Combining the definition of the spectral index $n_{s}$
(\ref{spectral index}) and the above equation we obtain
\begin{eqnarray}\label{ns-gama0}
n_s-1=-\frac{3}{4A}I(N)^{-f}+n_1+n_2
\end{eqnarray}
where in the derivation of the above equality we have used
\begin{eqnarray}\label{}
&n_1=\frac{3M_{pl}^4(f-1)fA}{8\pi\lambda} I(N)^{f-2}
\left(1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}}\right)^{-\frac{1}{2}}\\
&\nonumber
\times \left[-1+(1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}})^{\frac{1}{2}}\right]^{-1}\\
\nonumber
&n_2=\frac{9M_{pl}^2(1-f)fA}{16\pi\lambda} I(N)^{f-2}
\left(1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}}\right)^{-1} \;.
\end{eqnarray}
Lastly, based on Eq.(\ref{Tensor to scalar ratio}) we compute
the tensor-to-scalar ratio parameter
\begin{eqnarray}\label{r-gama0}
r=r_1I(N)^{\frac{5f-8}{4}}\left(1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}}\right)^{-\frac{3}{8}}\\
\nonumber
\left[-1+\left(1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}}\right)^{\frac{1}{2}}\right]^{\frac{1}{2}}G^2(N)
\end{eqnarray}
where $r_1=(\frac{2^{16} \lambda^2 3^3 C_{\gamma} \pi (fA)^5 (1-f)^3}{M_{pl}^2\Gamma_0^6})^{\frac{1}{4}}$.
\subsection{$\Gamma_{1}$-parametrization}\label{}
Using the same methodology as in the previous section
we provide the basic slow-roll parameters in the case of
$\Gamma_{1}$-parametrization, namely $\Gamma=\Gamma_1T$, where $\Gamma_{1}$
is constant. In particular, from Eqs.(\ref{scal},\ref{2.7})
the tachyon field is written as
\begin{eqnarray}\label{PPL}
\phi-\phi_0=\frac{{\tilde F}(t)}{\tilde{K}}
\end{eqnarray}
where ${\tilde F}(t)$ and $\tilde{K}$ are given by the following
expressions
\begin{eqnarray}\label{}
&{\tilde F}(t)=t^{\frac{7f-2}{8}}\times \\
\nonumber
&_{2}F_1\left[\frac{3}{16},\frac{7f-2}{16(f-1)},
\frac{23f-18}{16(f-1)},-\frac{3M_{pl}^2f^2A^2}{4\pi\lambda}t^{2f-2}\right]\\
\nonumber
&\tilde{K}=-\left(\frac{2^6\pi\Gamma_1^4 (1-f)^5}{3M_{pl}^2C_{\gamma}f^7A^7}\right)^{\frac{1}{8}}\frac{\Gamma(\frac{23f-18}{16(f-1)})}{\Gamma(\frac{7f-2}{16(f-1)})}\;.
\end{eqnarray}
Also here we have set $\phi_0=0$.
Now the Hubble parameter, the potential and the corresponding slow-roll
parameters are given by
\begin{eqnarray}\label{HUBB}
&H(\phi)=fA({\tilde F}^{-1}[{\tilde K}\phi])^{f-1}\\
\nonumber
&V(\phi)\simeq\lambda(-1+\sqrt{1+
\frac{3M_{pl}^2f^2A^2[{\tilde F}^{-1}({\tilde K}\phi)]^{2f-2}}{4\pi\lambda}})
\end{eqnarray}
\begin{equation}
\epsilon=\frac{1-f}{fA[{\tilde F}^{-1}({\tilde K}\phi)]^f} \;,
\end{equation}
\begin{equation}
\eta=
\frac{2-f}{2fA[{\tilde F}^{-1}({\tilde K}\phi)]^f} \;.
\end{equation}
where ${\tilde F}^{-1}(K\phi)$ is inverse function of ${\tilde F}(t)$.
In the current case the
number of e-folds becomes
\begin{eqnarray}\label{number of efolds2}
&N=\int_{t_{in}}^{t_{*}} Hdt=A(t_{*}^f-t_{in}^f)&\\
\nonumber
&=A\left( [{\tilde F}^{-1}({\tilde K}\phi)]^f-[{\tilde F}^{-1}({\tilde K}\phi_{in})]^f\right)& \;.
\end{eqnarray}
and following standard lines
the end of inflation takes place when
\begin{equation}
\phi_{in}=\frac{1}{\tilde{K}}\tilde{F}(y),\;\;\;\;\;\;
y=\left(\frac{1-f}{fA}\right)^{\frac{1}{f}} \;.
\end{equation}
The scalar power-spectrum can be easily identified by
comparing the current cosmological expressions with
Eq.(\ref{scalar power spectrum}), and we find
\begin{eqnarray}\label{}
&P_{s}=p_2 I(N)^{\frac{9f-6}{8f}}\left[-1+\left(1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}}\right)^{\frac{1}{2}}\right]^{-\frac{1}{2}}\\
\nonumber
&\times\left[1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}}\right]^{\frac{3}{16}}
\end{eqnarray}
where
$p_2=\frac{2(\Gamma_1 fA)^{\frac{3}{2}}}{3M_{pl}^2\lambda^{\frac{1}{2}}}(\frac{16\pi C_{\gamma}}{3M_{pl}^2(1-f)})^{\frac{3}{8}}$.
If we take the aforementioned $P_{s}$ formula we find
the following spectral index
\begin{eqnarray}\label{ns-gama1}
n_s-1=-\frac{9f-6}{8fA} I(N)^{-f}+n_1+n_2
\end{eqnarray}
where
\begin{eqnarray}\label{}
&n_1=\frac{3M_{pl}^4(f-1)fA}{8\pi\lambda} I(N)^{f-2}
\left(1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}}\right)^{-\frac{1}{2}}\\
&\nonumber
\times \left[-1+(1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}})^{\frac{1}{2}}\right]^{-1}\\
\nonumber
&n_2=\frac{9M_{pl}^2(1-f)fA}{32\pi\lambda} I(N)^{f-2}\left(1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}}\right)^{-1}
\end{eqnarray}
Finally, the tensor-to-scalar ratio [see Eq.{\ref{Tensor to scalar ratio})]
takes the form
\begin{eqnarray}\label{r-gama1}
r=r_2 I(N)^{\frac{7f-6}{8}}\left(1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}}\right)^{\frac{-3}{16}}\\
\nonumber
\left[-1+(1+\frac{3M_{pl}^2f^2A^2}{4\pi\lambda I(N)^{2-2f}})^{\frac{1}{2}}\right]^{\frac{1}{2}}G^2(N)
\end{eqnarray}
where $r_2=\frac{24\lambda^{\frac{1}{2}}(fA)^{\frac{7}{8}}}{\Gamma_1^{\frac{3}{2}}}
\left(\frac{16\pi C_{\gamma}(1-f)}{3M_{pl}^2}\right)^{\frac{3}{8}}$.
\section{Comparison with observation}\label{observation}
The analysis of \textit{Planck} \cite{Ade:2015lrj} and
BICEP2/Keck Array \cite{joint} data sets has provided
a new constraint on inflationary scenarios~\cite{Martin:2013tda}.
In particular, the comprehensive analysis
of \textit{Planck} data~\cite{Ade:2015lrj}
indicates that single
scalar-field models of slow-roll inflation have a very low tensor-to-scalar
fluctuation ratio $r=P_{t}/P_{s}\ll1$, a scalar spectral index
$n_{s}=0.968\pm 0.006$ and no appreciable running.
The upper bound set by
the \textit{Planck} team and the joint analysis of
BICEP2/Keck Array/\textit{Planck} \cite{joint}
on tensor-to-scalar fluctuation ratio is $r<0.11$.
In this section we attempt to test the performance of the warm inflationary
model against the above observational results.
Let us now concentrate on our results.
Notice, that in the case of warm inflation
the number of degrees of freedom becomes
$g_{*} \simeq 200$ ($C_{\gamma}\simeq 70$,\cite{BasteroGil:2006vr})
Also, for the rest of the paper
we have set $\lambda=10^{-14}$. Concerning
the number of e-folds, it is natural to consider that $N$
lies in the interval $[50,60]$. Here, we have set it either to
50 or 60.
In figures 1 ($\Gamma_{0}$ parametrization)
and 2 ($\Gamma_{1}$ parametrization)
we present the $A-f$ allowed region
in which our $(n_{s},r)$ results satisfy the above restrictions of
{\em Planck} within 1$\sigma$ uncertainties.
In the case of $\Gamma_0$ model, we observe that for various values
of the dissipation coefficient there is a narrow region in the $A-f$ plane
which is consistent with the observed values of $n_{s}$ and $r$.
The absence of $A-f$ pair solutions and thus of $(n_{s},r)$, appear for
$\Gamma_0\le 10^{-10}$.
For the $\Gamma_1$ parametrization the situation is slightly
different.
Figure (\ref{fig-gama1}) shows
broader $A-f$ regions with respect to those
of $\Gamma_{0}$ parametrization.
Also in this case we verify
that for $\Gamma_1\le 10^{-10}$, there is no $A-f$ pairs
which satisfy the observational criteria.
The theoretical curves of cold intermediate
inflation model in Einstein General Relativity in $n_s-r$ plane are well outside of the $95 \%$
C.L region.
Our aim here is to test the viability of the warm inflation, involving the
latest {\it Planck2015} data. In figure (\ref{fig-planck}) we present the
confidence contours in the $(n_{s},r)$ plane.
On top of figure (\ref{fig-planck}) we provide
the solid stars for the individual
sets of $(n_{s},r)$ which are based on the
$\Gamma_{0}$ parametrization, whereas in the same figure we
display the corresponding solid points
in the case of $\Gamma_{1}$ parametrization.
From the comparison it becomes clear that our
$(n_{s},r)$ results are in excellent agreement
with those of \textit{Planck 2015}.
Indeed, we find:
(a) $\Gamma_{0}$ parametrization: if we use $N=50$ then we find
$n_{s}=0.9675$ and $r=0.0086$, whereas for $N=60$ we have
$n_{s}=0.9627$ and $r=0.0036$.
(b) $\Gamma_{1}$ parametrization: in the case of $N=50$ we obtain
$n_{s}=0.9638$ and $r=0.00857$ and for $N=60$ we have $n_{s}=0.9692$ and
$0.00187$.
Bellow we compare the current predictions with those
of viable literature potentials.
This can help us to understand the variants
of the warm inflationary model
from the observationally viable inflationary scenarios.
\begin{itemize}
\item The chaotic inflation \cite{Linde}:
In this inflationary model the potential is
$V(\phi) \propto \phi^{k}$.
Therefore, the basic slow-roll
parameters are written as $\epsilon=k/4N$, $\eta=(k-1)/2N$
which implies $n_{s}=1-(k+2)/2N$ and $r=4k/N$.
It has been found that
monomial potentials with $k\ge 2$ can not accommodate the Planck
priors \cite{Ade:2015lrj}.
For example, using $k=2$ and $N=50$ we obtain
$n_{s}\simeq 0.96$ and $r\simeq 0.16$. For
$N=60$ we have $n_{s}\simeq 0.967$ and $r\simeq 0.133$.
It is interesting to mention that
the chaotic inflation also
corresponds to the slow-roll regime of intermediate inflation \cite{Barrow:1990vx,Barrow:1993zq,Barrow:2006dh,Barrow:2014fsa} with
Hubble rate during inflation given by $H\propto t^{k/(4-k)}$ with
$n_{s}=1-(k+2)r/8k$ and $k=-2$ gives $n_{s}=1$ exactly to first order.
\item The $R^{2}$ inflation \cite{staro}:
In Starobinsky inflation the asymptotic behavior
of the effective potential becomes
$V(\phi)\propto\lbrack1-2\mathrm{e}%
^{-B\phi/M_{pl}}+\mathcal{O}(\mathrm{e}^{-2B\phi/M_{pl}})]$
which provides the following slow-roll predictions \cite{Muk81,Ellis13}:
$n_{s}\approx 1-2/N$ and $r\approx 8/B^{2}N^{2}$, where $B^{2}=2/3$.
Therefore, if we select $N=50$ then we obtain
$(n_{s},r)\approx(0.96,0.0048)$. For $N=60$ we
find $(n_{s},r)\approx(0.967,0.0033)$.
It has been found that the Planck data \cite{Ade:2015lrj}
favors the Starobinsky inflation. Obviously, our results
(see figure 3) are consistent with those of $R^{2}$ inflation.
\item Hyperbolic inflation \cite{Basilakos:2015sza}:
In hyperbolic inflation the potential is
given by $V(\phi) \propto \mathrm{sinh}^{b}(\phi/f_{1})$.
Initially, Rubano and Barrow \cite{Rubano:2001xi}
proposed this potential in the context of dark energy.
Recently, Basilakos \& Barrow \cite{Basilakos:2015sza} investigated the properties
of this scalar field potential back in the inflationary epoch.
Specifically, the slow-roll parameters are written as
$$
\epsilon=\frac{b^{2}M_{pl}^{2}}{2f_{1}^{2}}\mathrm{coth}^{2}(\phi/f_1), \label{ee1}%
$$
$$
\eta=\frac{bM_{pl}^{2}}{f_{1}^{2}}\left[ (b-1)\mathrm{coth}^{2}(\phi/f_1)+1\right]
$$
and
$$
\phi=f_1\;\mathrm{cosh}^{-1}\left[ e^{NbM_{pl}^{2}/f^{2}}\mathrm{cosh}%
(\phi_{end}/f_1)\right] . \label{efold2}%
$$
where $\phi_{end}\simeq\frac{f}{2}\mathrm{ln}\left( \frac{\theta+1}{\theta
-1}\right)$. Comparing this model with the data
Basilakos \& Barrow \cite{Basilakos:2015sza} found
$n_{s}\simeq 0.968$, $r\simeq 0.075$, $1<b \le 1.5$
and $f_1\ge 11.7M_{pl}$.
\item Other inflationary models:
The origin of brane \cite{Dvali:2001fw,GarciaBellido:2001ky} and exponential \cite{Goncharov:1985yu,Dvali:1998pa}
inflationary models are motivated by the physics of
extra dimentions and supergravity respectively.
It has been found that these models are in agreement
with the Planck data
although the $R^{2}$ inflation
is the winner from the comparison \cite{Ade:2015lrj}.
\end{itemize}
\begin{figure}
\begin{center}
\includegraphics[scale=0.47]{gama0}
\end{center}
\caption{The $A-f$ diagram which coincides
within 1$\sigma$ confidence level of {\em{Planck}} data. The corresponding
values of
$\Gamma_0$ are shown at the top of panels.
The background transparent (foreground opaque) indicates $N=60$
($N=50$). These values of $\Gamma_0$ are consistent with $R\gg1$. The solid black curve shows the boundary $T=H$ and below the curve is consistent with $T>H$.}
\label{fig-gama0}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.47]{gama1}
\end{center}
\caption{The $A-f$ region in the case of $\Gamma_1$ parametrization. These values of $\Gamma_1$ are consistent with $R\gg1$. The solid black curve is same as Fig.(\ref{fig-gama0}). }
\label{fig-gama1}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.47]{planck}
\end{center}
\caption{1$\sigma$ and 2 $\sigma$ confidence regions borrowed from
{\em{Planck}} \cite{Ade:2015lrj}.
Stars (squares) indicate warm inflation with
$\Gamma_0$ ($\Gamma_1$) parametrization. Big and small points correspond to
$N=60$ and $N=50$ respectively. For $\Gamma_0$ we set $f=0.28$, $A=0.35$ and $\Gamma_0=10^{-9}$. For $\Gamma_1$ we set $f=0.13$, $A=3.$ and $\Gamma_1=10^{-3}$.
}
\label{fig-planck}
\end{figure}
At this point we would like to mention that
in the high-dissipation regime $R\gg 1$, there is always a region in $A-f$
plane which is consistent with the warm inflation condition $T>H$.
To clarify this issue we plot in Fig.(\ref{fig:warm})
the diagram of $\log_{10}\frac{T}{H}$ in the $A-f$ plane.
The solid line corresponds to the boundary limit $T=H$.
Clearly, based on the condition $T>H$
we can reduce the
parameter space and thus producing one of the strongest
existing constraints (to our knowledge) on $A$ and $f$.
Note that in order to produce the above diagram we have
fixed the initial values of $T$ and $H$ to those at the beginning of inflation.
After the triggering of inflation the
inflaton/photon interaction takes place which leads to radiation production and thus it guarantees that the above condition holds during the inflationary era.
\begin{figure}
\begin{center}
\includegraphics[scale=0.47]{t-h}
\end{center}
\caption{The value of $\frac{T}{H}$ at beginning of inflation in A-f plane. The solid black curve shows boundary $T=H$.}
\label{fig:warm}
\end{figure}
Finally, we investigate the possibility to treat
$\lambda$ as a free parameter.
In fact there are three main conditions which we need to use in order
to provide a viable limit on the $\lambda$. These are: (a) the
high dissipation regime $R\gg 1$, (b) the warm inflation condition $T>H$
and (c) to recover the {\it Planck2015} $(n_s,r)$ observational constraints.
Our investigation shows that $\lambda$ is correlated
with the $(A,f)$ pair. For example for $(A,f)=(0.5,0.25)$ we find
$\lambda>6\times10^{-16}$ which is consistent with above conditions
while for $(A,f)=(0.4,0.2)$ we obtain $\lambda>1.5\times10^{-20}$.
In general we verify that it is not possible
to find a lower value of $\lambda$ for all pairs of $(A,f)$.
\section{Appendix}\label{Appendix}
In this paper we have studied our model in natural unit ($\frac{h}{2\pi}=c=1$)
therefore we have ($[\rm mass]=M$, $[\rm time]=T$ and $[\rm length]=L$ where $[A]$ means dimension of "$A$")
\begin{eqnarray}\label{}
[c]=LT^{-1}=1~~~~~~~~~~~~~[h]=M L^2 T^{-1}\\
\nonumber
\Rightarrow~~~~T=L=M^{-1}~~~~~~~~~~~~~~~~.
\end{eqnarray}
Using Eq.(\ref{2.3}) we have
\begin{eqnarray}\label{}
[H^2]=[\frac{8\pi}{M^2}\rho_{\phi}(1+\frac{\rho_{\phi}}{2\lambda})]~~~~~~~~~~~~~~~~~~~~~~~\\
\nonumber
\Rightarrow \frac{[a^2]}{[a^2] T^2}=\frac{[\rho_{\phi}]}{[M^2]}\Rightarrow~~~[\rho_{\phi}]=[T_{\mu}^{\nu}]=[P_{s}]=M^4
\end{eqnarray}
where $\rho_{\phi}$ is the scalar field energy density with dimension $M^4$.
From Eq.(\ref{1.2}) we have
\begin{eqnarray}\label{}
[\dot{\phi}]=1~~~\Rightarrow~~~~[\phi]=M^{-1}
\end{eqnarray}
It appears that the tachyon scalar field has dimensions of $M^{-1}$. In Eq.(\ref{denn}) r.h.s and l.h.s have dimension $M^4$
\begin{eqnarray}\label{}
[\dot{\rho_{\phi}}]+[3H\rho_{\phi}]+[3HP_{\phi}]=[\Gamma \dot{\phi}^2]\\
\nonumber
\Rightarrow~~\frac{[\rho_{\phi}]}{T}+\frac{[\rho_{\phi}]}{T}+\frac{[P_{\phi}]}{T}=[\Gamma]\\
\nonumber ~~\Rightarrow [\Gamma]=M^{5}~~~~~~~~~~~~~~~.
\end{eqnarray}
Now based on
Eq.(\ref{12}) we find
\begin{eqnarray}\label{}
[R]=\frac{[\Gamma]}{[H][\rho_{\phi}]}=\frac{M^5}{M M^4}=1 .
\end{eqnarray}
\section{Conclusions}\label{conclusion}
In this article we investigate
the warm inflation for the
Friedmann-Robertson-Walker spatially
flat cosmological model in which the
scale factor of the universe satisfies the form of
Barrow \cite{Barrow:1996bd}, namely
$a(t)=a_I\exp(At^f)$ ($ 0<f<1$).
Within this context, we estimate analytically the
slow-roll parameters and we compare our predictions
with those of other inflationary models as well as
we test the performance of warm inflation against
the observational data.
We find that currently warm
inflationary model
is consistent with the results given by \emph{Planck 2015} within
$1\sigma$ uncertainties.
\begin{acknowledgements}
SB acknowledges support by the Research Center for
Astronomy of the Academy of Athens in the context of the program
\textquotedblleft\textit{Tracing the Cosmic Acceleration}.
\end{acknowledgements}
\bibliographystyle{spphys}
|
\section{Introduction}
Slow evolution is usually observed when glassy and disordered materials or binary systems are
quenched across a phase transition \cite{cuglia,BCKM,ioeleti}. However, while the additional
difficulties related to glassiness and disorder make understanding the formers quite difficult,
the phase-ordering process occurring when a clean magnet is cooled below the critical
point is a simpler context where aging properties can more easily be investigated.
In particular, an important issue concerns
the scaling properties of two-times quantities such as the autocorrelation $C$ and the
response function $\chi$. The mutual relation between them is of paramount importance
for glassy systems because, when suitable conditions are met, it encodes the
still debated structure of the equilibrium states via the Franz-Mezard-Parisi-Peliti theorem \cite{fmpp}. A program towards a full understanding of this relationship might start
very naturally from the paradigmatic case of non-disordered magnets.
Despite their relative simplicity, however, a fully satisfactory
reference analytical theory of aging in these systems does not exist and their behavior
is not fully understood.
In this paper we present a rather complete numerical study of the dynamics of the possible
sub-critical quenches of a bi-dimensional magnet described by the Ising model with
spin-flip dynamics. In particular, we study systems cooled from an initial equilibrium state
at an infinite temperature $T_i=\infty$ and from the critical configuration at the
transition temperature $T_i=T_c$. Starting from these initial states, we consider deep quenches to a final temperature $T_f=0$, or to an intermediate
one $T_f=0.66 \,T_c$, and shallow coolings to $T_f=0.97 \,T_c$.
This allows us to discuss the behavior of the response function in the processes ending at
$T_f=0$ (starting with arbitrary $T_i$) and in those starting from $T_i=T_c$
(ending at any $T_f$), which were never studied before and yield unexpected results.
As expected on the basis of dynamical scaling \cite{bray,furukawa}, we find that observable
quantities take scaling forms regulated by universal exponents.
Our data for quenches to $T_f=0$ show quite unambiguously that, in this case,
the response function exponent $\lambda _\chi$ [see Sec. \ref{Model},
Eqs. (\ref{scalChi},\ref{largeh}), for a definition of these quantities]
takes a value compatible with
$\lambda _\chi =\frac{1}{2}\lambda _C\simeq 5/8$, $\lambda _C $ being the exponent
governing the decay of the autocorrelation function
[see Eqs. (\ref{scalC},\ref{largec})].
This determination of $\lambda _\chi$ is definitely different from the one found
in previous studies \cite{unquarto,algo,henkelrough} which were focused on quenches to final $T_f$.
Since the final temperature
of the quench is
expected to be an irrelevant parameter \cite{bray,brayren}, in the sense of the renormalization group, we conjecture
that $\lambda _\chi=\frac{1}{2}\lambda _C$ is the correct asymptotic value and that, in
the above mentioned previous studies, the correct value was shadowed by finite-$T_f$
preasymptotic corrections (which are indeed observed also in the present study when $T_f$
is chosen finite).
Concerning the role of the initial condition we show that, while the autocorrelation function takes
a radically different value \cite{brayfromcrit} when the quench is made from
$T_i=\infty$ [$\lambda _C(T_i=\infty)=5/4$] or $T_i=T_c$ [$\lambda _C(T_i=T_c)\simeq 1/8$],
the response function exponent remains basically unchanged. By solving the
large-$N$ (or spherical) model we find that the same feature is shared also in this
analytically tractable model of magnetism. Physically, we interpret the different
sensitivity of $C$ and $\chi$ to the initial state as due to the different role played by the
large-scale properties of the system which, in the quench from $T_i=T_c$, keep
memory of the critically correlated initial state.
This Article is organized as follows: In Sec. \ref{Model} we set the notation, define the
model under investigation, the different observables considered throughout the paper
and their scaling behavior.
Sec. \ref{SIMUL} is devoted to the presentation and discussion of our numerical results.
In Sec. \ref{largen}, by exactly solving the large-$N$ model, we discuss
some analogies with the numerical results presented in Sec. \ref{SIMUL}.
In Sec. \ref{concl} we summarize the main findings of this paper, discuss
some open issues, and present the conclusions of the work.
\section{The Ising Model and the observable quantities}
\label{Model}
We consider a system of ${\cal N}$ spins on a lattice with the Ising Hamiltonian
\begin{equation}
H=-J \sum_{<ij>}\sigma_i \sigma_j
\ee
where the sum runs over the nearest neighbors pairs $<ij>$ and $J>0$.
The time evolution occurs through single spin-flip dynamics with Glauber
transition rates \cite{glauber}
\begin{equation}
w_i([\sigma] \to [\sigma'])=\frac{ 1}{2}
\left[ 1-\sigma_i \tanh\left(\frac{h_i^W}{T}\right)\right]
\label{w}
\ee
where $[\sigma]$ and $[\sigma']$ are spin configurations differing
only for the value of the spin on the $i$-th site and
$h_i^W=J\sum_{k\in \{nn_i\}} \sigma_k$, where the sum is restricted to the nearest
neighbors $\{nn_i\}$ of $i$, is the local Weiss field.
In a quenching protocol the system is prepared at $t=0$
in an equilibrium configuration at the initial temperature $T_i$ and
is then evolved with the transition rates (\ref{w}) where $T$ is set equal to
the final temperature $T_f$.
The typical size of the growing ordered domains $L(t)$ will be computed in this
work as the inverse excess energy:
\begin{equation}
L(t)=[E(t)-E_\infty]^{-1} .
\label{lt}
\ee
Here, $E(t)=\langle H(t)\rangle$ is the average energy at time $t$, and $E_\infty $ is the one of the
equilibrium state at the final temperature $T_f$. Here and in the following the average $\langle
\cdots \rangle$ is taken over different realizations of the initial state and of the thermal histories,
namely over the random flip events generated by the transition rates (\ref{w}).
Equation~(\ref{lt}) is often used to determine $L(t)$ \cite{bray} because the excess energy of the coarsening system with respect to the equilibrated one is associated to the density of domain walls which,
in turn, is inversely proportional to the typical domain size.
The two-points/two-times correlation function is defined as
\begin{equation}
{\cal G}(\vec r,t,s)=\langle \sigma _i (t) \sigma _j (s)\rangle - \langle \sigma _i (t)\rangle
\langle \sigma _j (s)\rangle,
\label{defcalG}
\ee
where we assume $t\ge s$,
which depends only on
the distance $\vec r$ between sites $i$ and $j$ due to space homogeneity.
In the phase ordering process this quantity takes the additive structure
\begin{equation}
{\cal G}(\vec r,t,s)={\cal G}_{st}(\vec r,t-s)+{\cal G}_{ag}(\vec r,t,s),
\label{calG}
\ee
where the first contribution describes the fast equilibrium fluctuations in the pure states
which are attained well inside the domains and the second one contains the non-equilibrium aging properties.
Being an equilibrium contribution, the first term in Eq. (\ref{calG}) vanishes over distances larger than the equilibrium coherence length $\xi _{eq}$ and/or time differences longer than the
equilibrium correlation time $\tau _{eq}$. In the present paper we will be interested in the
behavior of the system on distances and time-differences much larger than $\xi _{eq}$
and $\tau _{eq}$, respectively, where the first term of Eq. (\ref{calG}) can be neglected.
We will then focus on the second contribution, the aging term, and drop the suffix $_{ag}$.
Letting $t=s$ in Eq. (\ref{defcalG}) amounts to consider the equal-time correlation function
\begin{equation}
G(\vec r,t)=\langle \sigma _i (t) \sigma _j (t)\rangle,
\label{defG}
\ee
where we have neglected the last term in Eq. (\ref{defcalG}) since in the processes
we will be interested in one always has $\langle \sigma _i(t)\rangle=0$.
We compute numerically this quantity as
\begin{equation}
G(r,t)=\frac{1}{{4\cal N}} \sum _{i,j:|i-j|=r} \left \langle
\sigma _i (t) \sigma _j (t)\right \rangle.
\ee
where the sum runs over all the $4{\cal N}$ couple of sites at distance $r$ on the horizontal
and vertical direction.
This correlation function obeys the scaling form \cite{bray,furukawa}
\begin{equation}
G(r,t)=g\left [\frac{r}{L(t)}\right ],
\label{scalG}
\ee
where $g(y)$ is a scaling function. For completeness, let us mention that
corrections to the form (\ref{scalG}) were reported in \cite{epl}. These corrections,
however, are negligible for the large system-sizes considered in this paper.
The sharp nature of the domains walls imply
a short-distance behavior \cite{bray,porod}
\begin{equation}
g(y)\simeq 1-ay,
\label{porod}
\ee
where $a$ is a constant, in the limit $y\ll 1$.
When the system is quenched from an initially critical state, i.e. with $T_i=T_c$,
the scaling function has the following large-$y$ behavior \cite{brayfromcrit}
\begin{equation}
g(y)\propto y^{-(d-2+\eta)},
\label{critg}
\ee
where $\eta$ is the equilibrium critical exponent, namely $\eta =1/4$ in the two-dimensional case
studied in this paper.
This simply expresses the fact that, for distances $r\gg L(t)$ much larger than those where
ordering has been effective at the current time, the system keeps memory of the equilibrium
critical initial state. As we will see shortly, this bears important consequences, particularly on the
behavior of the autocorrelation function $C$ that is defined by
setting $i=j$ in Eq. (\ref{defcalG})
\begin{equation}
C(t,s)=\langle \sigma _i (t) \sigma _i (s)\rangle.
\label{defC}
\ee
This quantity does not depend on $i$ due to homogeneity.
Enforcing this, $C$ will be computed as
\begin{equation}
C(t,s)=\frac{1}{{\cal N}}\sum _i \left <
\sigma _i (t) \sigma _i (t)\right \rangle,
\label{spataveC}
\ee
in our simulations.
The autocorrelation obeys the scaling form \cite{bray,furukawa}
\begin{equation}
C(t,s)=c\left [\frac{L(t)}{L(s)}\right ],
\label{scalC}
\ee
where $c(x)$ is a scaling function with the large-$x$ behavior
\begin{equation}
c(x)\simeq x^{-\lambda _C}.
\label{largec}
\ee
The autocorrelation exponent is expected to be $\lambda _C(T_i=\infty)=5/4$ \cite{bray,liumaz}
for a quench from $T_i=\infty$ and a much smaller value \cite{brayfromcrit}
$\lambda _C(T_i=T_c)=1/8$ for a quench from the critical state at $T_i=T_c$.
The linear response function is defined as
\begin{equation}
R(t,t')=\left . \frac {\delta \langle \sigma _i (t) \rangle}{\delta h_i (t')} \right \vert _{h=0},
\label{defR}
\ee
where $h_i(t')$ -- the perturbation -- is a magnetic field of amplitude $h$ applied on site
$i$ only at time $t'$. From this quantity the integrated response function,
referred to also as
dynamic susceptibility or zero-field-cooled susceptibility,
is obtained as
\begin{equation}
\chi (t,s)=\int _s^t R(t,t')\,dt'.
\label{defChi}
\ee
Given that also the response function is independent on $i$,
we improve the numerical efficiency by computing this quantity as a spatial
average, similarly to what done for the
autocorrelation function [Eq. (\ref{spataveC})].
We obtain this quantity using the generalization to non-equilibrium states
of the fluctuation-dissipation theorem derived in \cite{algo,algo2}. Similarly to the
well known equilibrium theorem, this amounts to an analytical relation
between $\chi $ and certain correlation functions of the {\it unperturbed} system,
namely the one where the perturbation $h$ is absent. The great advantage of this approach is
the fact that the $h\to 0$ limit in Eq. (\ref{defR}) is dealt with analytically, making the numerical computation of the response function totally reliable and very efficient.
Notice that the use of the non-equilibrium fluctuation-dissipation relation provides
directly the quantity $T_f\chi$.
Being related to correlation functions also the response function can be split in two
terms, similarly to Eq. (\ref{calG}). However, at variance with the other quantities
discussed above, the term $\chi _{st}$ is not negligible and, in order to isolate the
aging part, it has to be subtracted away. This can be done
by computing preliminarily $\chi _{st}$ in the equilibrium state, as described in \cite{unquarto}.
The aging part of the response function obeys the scaling form \cite{revresp}
\begin{equation}
\chi(t,s)=L(s)^{-\alpha}h\left [\frac{L(t)}{L(s)}\right ],
\label{scalChi}
\ee
where $h(x)$ is a scaling function with the large-$x$ behavior
\begin{equation}
h(x)\simeq x^{-\lambda _\chi}.
\label{largeh}
\ee
It is expected that $\chi$ becomes independent on $s$ for large time differences
$t-s\gg s$. Because of Eqs. (\ref{scalChi},\ref{largeh}) this implies that
\begin{equation}
\alpha=\lambda _\chi.
\label{lameqalf}
\ee
This property has been verified several times in the
literature \cite{unquarto}.
As for the value of the response exponent $\alpha$
it was conjectured to be $\alpha =1/2$ and
present numerical determinations set its value in the range $[0.5-0.56]$.
We will discuss further the value of this exponent in the following.
\section{Numerical simulations} \label{SIMUL}
In this Section we present and discuss the results of our simulations
which have been obtained by quenching a two-dimensional system of linear size
$\Lambda =2\cdot 10^3$ (unless differently specified)
with $J=1$ and periodic boundary conditions.
We have evolved the system starting with configurations from the
infinite as well as the critical temperature
from $t=0$ to a final time $t=3\cdot 10^4$.
In the case of quenches from the
critical state, in order to equilibrate the system at $T_c$, we have used
the Wolff algorithm \cite{wolff}. Data are organized in different sections
according to the different choices of the initial and final temperatures
of the various quenching protocols considered.
\subsection{Quenches to $T_f=0$} \label{to0}
\subsubsection{From $T_i=\infty$} \label{TinfT0}
The behavior of $L(t)$ is shown in Fig. \ref{L_to_zero} (lower black curve).
Starting from $t\simeq 10$, the expected power-law $L(t)\sim t^{1/z}$
with $z=2$ sets in (a fit in the last decade provides $z=2.004$).
Data for smaller system sizes ($\Lambda =1.5\cdot 10^3$ and $\Lambda =10^3$, not shown)
superimpose to the plotted data. This, and the power-law behavior of $L$, clearly indicate
that our simulations are finite-size effect free in the present case.
The behavior of the equal-time correlation function is shown in
Fig. \ref{G_to_zero} (lower set of curves) for various times (see key)
on a double-logarithmic plot.
As expected on the basis of Eq. (\ref{scalG}), an excellent data collapse of the
curves at different times is obtained
by plotting $G$ against the rescaled space $y=r/L(t)$ in all the region where
$G$ is significantly larger than zero. As it can be seen in the inset of Fig. \ref{G_to_zero},
where a zoom on the small-$r$ sector is presented on a plot with linear axis,
a linear behavior -- the so called Porod's law, Eq. (\ref{porod}) -- is well obeyed in this regime,
as discussed in Sec. \ref{Model}.
The autocorrelation function is plotted against $x=L(t)/L(s)$
in Fig. \ref{C_to_zero}.
The collapse expected on the basis of Eq. (\ref{scalC}) is excellent in the region of large $x$.
For small values of $x$ the superposition is worse for the smaller
values of $s$ but, also in this region, an excellent
collapse is recovered for sufficiently large values of $s$.
This can be interpreted as due to the presence of preasymptotic corrections to scaling
which are not completely negligible for the smaller values of $s$ considered in our simulations.
Notice that for large $x$ the expected behavior $C(x)\sim x^{-\lambda_C}$ of
Eq. (\ref{largec}) with
$\lambda _C=5/4$ is very well
reproduced (a fit of the curve with $s=10$ for $x\ge 30$ gives $\lambda _C=1.256$).
The data presented insofar show that an excellent scaling behavior is displayed starting from
the region of moderate times. Therefore this case is an optimal playground to assess the scaling properties
of the response function, which in the past have been the subject of some controversies.
The response function is shown in Fig. \ref{Chi_to_zero}.
According to Eq. (\ref{scalChi}) and the discussion thereafter one should get
the collapse of curves for different waiting times $s$
by plotting $L(s)^{\alpha}\chi (t,s)$ against $x=L(t)/L(s)$,
where the exponent $\alpha$ equals the exponent $\lambda _\chi$
regulating the large-$x$ behavior of $\chi$ according to Eq. (\ref{largeh}).
Comparison of the large-$x$ behavior of the curves in Fig. \ref{Chi_to_zero}
with the dark-green line $x^{-5/8}$ shows that $\chi$ is
very well compatible with a value $\lambda _\chi=5/8$.
This value is different from the one found for quenches to finite temperatures
where a value in the range $\alpha \simeq 0.5-0.56$ (according to different determinations)
was reported. This was interpreted \cite{noirough,henkelrough} as the value $\alpha =1/2$ expected on the basis of
an argument associating the properties of the response to the roughening of the interfaces.
The present determination $\lambda_\chi \simeq 5/8$, instead, suggests that the somewhat
different value $\alpha=\lambda_\chi=(1/2)\lambda _C$ could be the asymptotically correct one.
Notice that a value $\alpha \simeq 0.6$, roughly comparable to the one we find here,
was found in \cite{RBIM} in the phase-ordering of a weakly disordered magnet in the
limit of an extremely deep quench.
In order to verify better this conjecture and to have the most
reliable determination of $\alpha$ from our data we plot in Fig. \ref{scal_vs_tw_fix_x}
$T_f\chi (t,s)$ against $L(s)$ for fixed values of $x$ spanning the entire sector $x\in [1.5,30]$.
According to Eq. (\ref{scalChi}) on a double logarithmic scale the slope of the curves
in this figure directly provides the exponent $\alpha $.
The data of Fig. \ref{scal_vs_tw_fix_x} show consistency with an exponent $\alpha =5/8$
(dark-green bold dotted line), while $\alpha=1/2$ does not fit equally well (except, perhaps,
in a region of very small $s$ and $x$ where pre-asymptotic effects may show up,
see discussion below).
Fitting the curves for the different values of $x$, indeed, gives effective exponents
$\alpha _{eff}=0.575,\, 0.600,\, 0.612,\, 0.615,\, 0.615,\, 0.617,\, 0.616,\, 0.616,\, 0.616,\,
0.614,\, 0.612,\, 0.611,\, 0.589$ which (with the exception of the first two and of the last
value, see discussion below) are rather close to $\lambda _C/2=0.625$, while they
seem to rule out the possibility $\alpha= 1/2$.
Let us comment shortly on the somewhat smaller values of $a_{eff}$ found
for small $x$ ($x\le 2$) (namely $\alpha _{eff}=0.575$ for $x=1.5$ and
$\alpha _{eff}=0.600$ for $x=2$)
and for very large $x$ (namely $\alpha _{eff}=0.589$ for $x=30$).
Data for small $x$ are probably affected by pre-asymptotic
effects at small $s$. Indeed, fitting for instance the curve relative to $x=1.5$
only for values of $L(s)>10$ the effective exponent rises to $a_{eff}=0.609$
(from the value $a_{eff}=0575$ when fitted over the whole range of $L(s)$).
The data for $x$ as large as $x=30$ contain only two points and this makes
the determination of $a_{eff}$ in this case probably insecure.
From this analysis we can conclude therefore that, except in the region of very
small or very large $x$ where preasymptotic corrections and other effects make the
determination of the exponent insecure, a value $\alpha =5/8$ is well consistent
with the data.
To provide the most possible reliable determination of the response function exponent
we present a further, alternative analysis of the data in the following.
Let us observe that, from the scaling of $C$ and $\chi$, Eqs. (\ref{scalC},\ref{scalChi}),
in the region of large time differences where the forms of Eqs. (\ref{largec},\ref{largeh})
hold, one has
\begin{equation}
A_\alpha(t,s)\equiv \left [s^{\alpha}\, T_f\,\chi (t,s) \right ]^{\frac{\lambda _C}{\alpha}}=\kappa \,C(t,s),
\hspace{1cm} t-s\gg s,
\label{check}
\ee
where we have used the fact that $\alpha =\lambda _\chi$ and $\kappa $ is a constant.
Eq. (\ref{check}) represents a tool for a stringent test on the value of the exponent
$\alpha $, as we discuss below. In Fig. \ref{Chi2_to_zero} we show the parametric plot of
$A_\alpha(t,s)$ against
$C(t,s)$.
Specifically, for any couple of times $t,s$ we plot $A_\alpha(t,s)$
on the vertical axis and $C(t,s)$ on the horizontal one.
Notice that one can re-parametrize only one (say $t$) of the two-times
appearing in $C(t,s)$ through the value of $C$ itself. In doing that one goes from
$A_\alpha (t,s)$ to a new function $A_\alpha (C,s)$ that, in principle, depends also on $s$
(besides the quantity $C$ on the horizontal axis).
However, according to Eq. (\ref{check}), for large time-differences $t-s\gg s$, meaning small values
of the quantity $C$, the quantity $A_\alpha $ ought to be an $s$-independent linear function of $C$
(i.e. curves for different values of $s$ should collapse), if the value of $\alpha $ is the
appropriate one. On the other hand, for an improper value $\beta\ne \alpha$ of this
exponent, instead of Eq. (\ref{check}) the function $A_\beta$ would behave as
\begin{equation}
A_{\beta}(t,s)\equiv \left [s^{\beta}\, T_f\,\chi (t,s) \right ]^{\frac{\lambda _C}
{\beta}}=\widetilde \kappa s^{\left (1-\frac{\alpha}{\beta}\right )\lambda _c}\,
C(t,s)^{\frac{\alpha}{\beta}},
\hspace{1cm} t-s\gg s,
\label{anticheck}
\ee
where $\widetilde \kappa=\kappa ^{\frac{\alpha}{\beta}}$ is another constant.
Therefore in this case the parametric plot of $A_\beta$ against $C$ would not be linear and
curves for different values of $s$ would not collapse. This qualifies
this kind of plot as a strict
check on the correct value of $\alpha $.
In Fig. \ref{Chi2_to_zero} we compare the performance of the two values
$\alpha = (1/2)\lambda_C=5/8$ (left panel) and $\alpha =1/2$ (right panel).
While in the former case one does observe data collapse (small preasymptotic corrections
are only observed for the smaller values of $s$, as expected) and linear behaviors,
both these feature are clearly lost in the latter case.
Notice also, as a further confirmation of the accuracy of the determination $\alpha = 5/8$,
that in the right panel the curves for small $C$ can be well fitted by the power law
$\sim C^{5/4}$ (bold-dotted green line), as expected on the basis of Eq. (\ref{anticheck})
with $\alpha =5/8$ and $\beta = 1/2$.
As already stated, the value $\alpha =5/8$ is in contrast with the one $\alpha =1/2$
which was argued before.
However, previous computations were always carried out in quenches to finite
final temperatures (indeed, as we will see in the next sections, we recover
a smaller value of $\alpha $ - compatible with the one ($\alpha=0.5-0.56$) found in the aforementioned
literature, when considering quenches to $T_f>0$), whereas to the best of our knowledge
this is the first determination of this exponent in the case with $T_f=0$.
We will comment in Secs. \ref{to15}, \ref{to22} on this new result and we will provide
a possible interpretation of the discrepancy between the present determination of $\alpha $ and
the previous ones.
\begin{figure}
\rotatebox{0}{\resizebox{.95\textwidth}{!}{\includegraphics{L_to_zero.pdf}}}
\caption{(Color on-line). $L(t)$ is plotted against $t$ on a log-log plot
for the quench of
the system to $T_f=0$ starting from the equilibrium state at $T_i=\infty$
(lower black curve with a circle)
and at $T_i=T_c$ (upper partially collapsing three curves, corresponding to three different
system sizes $\Lambda $ as detailed in the key). The bold green lines are the behavior
$L(t)\sim t^{1/2}$ expected for an infinite system for large times.}
\label{L_to_zero}
\end{figure}
\begin{figure}
\rotatebox{0}{\resizebox{.95\textwidth}{!}{\includegraphics{G_to_zero.pdf}}}
\caption{(Color on-line). The correlation function $G(r,t)$ for a quench to $T_f=0$
is plotted against $y=r/L(t)$
on a double-logarithmic plot at the different times indicated in the key
(these are determined as to be exponentially spaced by an automatic routine).
The lower set of collapsing curves (without symbols), correspond to the case
discussed in Sec. \ref{TinfT0} of a quench from an equilibrium state at the initial
temperature $T_i=\infty$. The upper group of curves,
marked with a symbol, correspond to the case discussed in Sec. \ref{TcT0} of a quench
starting from the critical state, i.e. $T_i=T_c$.
The bold green straight line is the power-law $y^{-1/4}$ of Eq. (\ref{critg}).
In the inset a zoom of the same data in the region of small $r/L(t)$ is plotted on a
double-linear plot. The bold-dotted turquoise line is the linear behavior
of Eq. (\ref{porod}), namely the Porod's law.}
\label{G_to_zero}
\end{figure}
\begin{figure}
\rotatebox{0}{\resizebox{.85\textwidth}{!}{\includegraphics{C_to_zero.pdf}}}
\caption{(Color on-line). $C(t,s)$ is plotted against $x=L(t)/L(s)$
for the quench of
the system to $T_f=0$ starting from the equilibrium state at $T_i=\infty$
(lower set of curves)
and at $T_i=T_c$ (upper group of curves) for different times (see key).
The bold dark-green and turquoise lines are the expected
power-laws $x^{-\lambda _C}$ with $\lambda _C=5/4$ and $\lambda _C=1/8$
for the quenches from $T_i=\infty$ and from $T_i=T_c$, respectively.}
\label{C_to_zero}
\end{figure}
\begin{figure}
\rotatebox{0}{\resizebox{.85\textwidth}{!}{\includegraphics{Chi_to_zero.pdf}}}
\caption{(Color on-line). $L(s)^{5/8}T_f\,\chi(t,s)$ is plotted against $x=L(t)/L(s)$
for the quench of
the system to $T_f=0$ starting from the equilibrium state at $T_i=\infty$
(upper set of curves)
and at $T_i=T_c$ (lower group of curves) for different times (see key).
The bold dark-green line is the
power-law $x^{-\lambda _\chi}$ with $\lambda _\chi=5/8$.}
\label{Chi_to_zero}
\end{figure}
\begin{figure}
\rotatebox{0}{\resizebox{.85\textwidth}{!}{\includegraphics{scal_vs_tw_fix_x.pdf}}}
\caption{(Color on-line). $T_f\,\chi(t,s)$ is plotted against $L(s)$ for fixed values of $x$
(indicated in the key) for the quench of
the system to $T_f=0$ starting from the equilibrium state at $T_i=\infty$.
The bold dotted dark-green lines and the dashed magenta one are the
power-laws $x^{-\alpha}$ with $\alpha=5/8$ and $\alpha=1/2$, respectively.}
\label{scal_vs_tw_fix_x}
\end{figure}
\begin{figure}
\vspace{1cm}
\rotatebox{0}{\resizebox{.45\textwidth}{!}{\includegraphics{Chi2_vs_C_inf_to_zero_double_resc.pdf}}}
\rotatebox{0}{\resizebox{.45\textwidth}{!}{\includegraphics{Chi2_vs_C_inf_to_zero_double_resc_025.pdf}}}
\caption{(Color on-line). The function $A_{5/8}(t,s)$ (left panel)
and $A_{1/2}(t,s)$ (right panel)
are plotted against $C(t,s)$ for the quench of
the system to $T_f=0$ starting from the equilibrium state at $T_i=\infty$
for different times (see key).
The bold-dotted turquoise line on the left panel is guideline for the linear behavior
$[L(s)^{5/8}\chi (t,s)]^2\propto C(t,s)$. The bold-dotted green one on the right panel
is the behavior $C(t,s)^{5/4}$.}
\label{Chi2_to_zero}
\end{figure}
\subsubsection{From $T_i=T_c$} \label{TcT0}
Let us now investigate which differences occur in the scaling properties of the system when the
quench is made from the critical point $T_i=T_c$ instead of having $T_i=\infty$.
The behavior of $L(t)$ in this case is shown in Fig. \ref{L_to_zero} (upper set of
curves).
Here it is seen that $L$ is considerably larger than the one computed in the
quench from $T_i=\infty$. This can be understood since the system at infinite
temperature is maximally disordered while the critical state is a coherent one, although
without order.
Larger values of $L(t)$ and -- possibly --
the strong correlations present in the initial critical state, bring in finite-size
effects at earlier times as compared to the quench from $T_i=\infty$.
Indeed also in this case the expected power-law $L(t)\sim t^{1/z}$ with $z=2$
sets in around $t\simeq 10$ but, differently from the case with $T_i=\infty$,
one observes an upward bending of the curves starting from $t\simeq 10^3$ onwards.
The bending is more pronounced for smaller system sizes, confirming that it occurs earlier and
indicating that we are in the presence of important finite-size effects for
$t\gtrsim 10^3$.
Notice that finite-size effects do not produce in this case
an abrupt modification with respect to the behavior in an infinite system but,
rather, a gradual drift (in this case an upward raising) which can be confused with
a genuine effect. For instance, on the basis of Fig. \ref{L_to_zero}
one could erroneously conclude that the asymptotic exponent is smaller than $z=2$.
Clearly, finite-size effects not only modify the behavior of $L(t)$
but affect all observables quantities, as we will discuss below.
Starting from the equal-time correlation function, shown in
Fig. \ref{G_to_zero} (upper group of curves), one finds
data collapse of the curves at different times
by plotting $G$ against the rescaled space $y=r/L(t)$, as expected on the basis of Eq. (\ref{scalG}),
but this is only true
in a small-$r$ region which
shrinks as time increases. The breakdown of dynamical scaling at large $r$ is another
clear manifestation of the finite-size effects.
In an infinite system the curves of Fig. \ref{scalG} would collapse for any
value of $r$ and at any time (except at such early times that scaling has not yet set in).
In our finite system this occur only in a range which gets narrower as $L(t)$ approaches
$\Lambda$. Despite this,
when time is sufficiently small there is room
to observe the typical large-distance power-law behavior $g(y)\sim y^{-\eta}$, with $\eta =1/4$, of
Eq. (\ref{critg}) induced by the reminiscence of the initial critical state.
The autocorrelation function is plotted against $x=L(t)/L(s)$
in Fig. \ref{C_to_zero}.
For all values of $x$ the collapse expected on the basis of Eq. (\ref{scalC}) is worse than
the one observed in the quench from $T_i=\infty$
(the effect is partly masked in Fig. \ref{C_to_zero} because data are compressed).
This is probably due to the combined effect of short-time corrections and finite-size
effects. Nevertheless the behavior $C(x)\sim x^{-\lambda_C}$ expected for large $x$
on the basis of Eq. (\ref{largec}) with
$\lambda _C=1/8$ is well reproduced in the regime of relatively large times but
such that finite-size effects are still negligible (namely, as already noticed
discussing the data for $L(t)$ (i.e. Fig. \ref{L_to_zero}), for $t\lesssim 10^3$
(for the curve with $s=10$ this roughly amounts to $x\lesssim 10$).
For larger values of $x$ the curves for $C(t,s)$ bend upwards, similarly to what
observed for $L(t)$. Needless to say, the behavior of $C$ is profoundly different
when the quench is made from $T_i=\infty$ or from $T_i=T_c$.
Next we consider the response function. To the best of our knowledge the computation
of this quantity
in a quench from criticality was never performed before.
The behavior of $\chi$ in this case is shown in Fig. \ref{Chi_to_zero} (lower set of curves).
The curves can be collapsed, according to Eq. (\ref{scalChi}),
by plotting $L(s)^{\alpha}\chi (t,s)$ against $x=L(t)/L(s)$ where,
recalling the previous discussion, the exponent $\alpha $ is expected
to be equal to $\lambda _\chi$. The data of Fig. \ref{scalChi} show that also in this case
$\alpha $ is well consistent with the
value $\alpha =5/8$, namely the same value obtained in the quench from infinite temperature.
This indicates that, while the behavior of the autocorrelation $C$ is profoundly different in
the two cases with $T_i=\infty$ and $T_i=T_c$, as to give $\lambda _C$ exponents as different
as $\lambda _C=5/4$ and $\lambda _C=1/8$, the response function exponent is insensitive
to the properties of the initial state. An interpretation of this fact will be given in
Sec. \ref{largen} where the exact solution of the large-$N$ model, which shares the same property, will be
discussed.
In the following we will show how thermal fluctuations occurring when $T_f\neq 0$ modify
the scaling picture found above for the zero-temperature quench. We will consider the two cases
with an intermediate temperature $T_f=1.5\simeq 0.66 T_c$ and one, $T_f=2.2\simeq 0.97 T_c$, close to
the critical one.
\subsection{Quenches to $T_f=1.5$} \label{to15}
\subsubsection{From $T_i=\infty$} \label{TinfT15}
As shown in Fig. \ref{L_to_15} the behavior of $L(t)$ in the quench to $T_f=1.5$
is very similar to the case with $T_f=0$ discussed previously.
The expected power-law $L(t)\sim t^{1/z}$
with $z=2$ (a fit in the last decade provides $z=2.031$) sets in quite early
and there is no indication of any finite-size effect.
The behavior of the equal-time correlation function is shown in
Fig. \ref{G_to_15} (lower group of curves) for various times (see key)
on a double-logarithmic plot.
As expected on the basis of Eq. (\ref{scalG}), data collapse of the
curves at different times is obtained
by plotting $G$ against the rescaled space $y=r/L(t)$ in all the region where
$G$ is significantly larger than zero. The collapse is poor at small times but it gets
better moving to larger $t$.
As it can be seen by zooming the small-$r$ region
in the inset of Fig. \ref{G_to_15} with linear axes,
the Porod's law (\ref{porod}) is well reproduced also in this quench.
The autocorrelation function is plotted against $x=L(t)/L(s)$
in Fig. \ref{C_to_15}. Also for this quantity
the collapse expected on the basis of Eq. (\ref{scalC}) is poor for the smaller values
of $s$ but becomes progressively more accurate, particularly in the region of large $x$, as $s$ is increased.
The expected behavior $C(x)\sim x^{-\lambda_C}$ of
Eq. (\ref{largec}) with
$\lambda _C=5/4$ is quite well
observed (a fit of the curve with $s=10$ for $x\ge 30$ gives $\lambda _C=1.236$).
The response function is shown in Fig. \ref{Chi_to_15}.
This quantity was previously computed several times \cite{unquarto} for a quench to the same
final temperature $T_f=1.5$ considered here. The present study therefore
allows us to compare
our results with previous ones and to discuss the discrepancy on the $\alpha $ exponent found
in the previous section \ref{to0}.
According to Eq. (\ref{scalChi}) one should get
the collapse of curves for $\chi $ at different waiting times $s$
by plotting $L(s)^{\alpha}\chi (t,s)$ against $x=L(t)/L(s)$,
where the exponent $\alpha$ equals the exponent $\lambda _\chi$
defined in Eq. (\ref{largeh}).
Comparison of the large-$x$ behavior of the curves in Fig. \ref{Chi_to_15}
with the dark-green line $x^{-5/8}$ shows that $\lambda_\chi$ is
very well compatible with the value $\lambda _\chi=5/8$ found with $T_f=0$
also in this finite-temperature quench.
In previous studies \cite{unquarto,revresp,henkelrough} the scaling of the response function was
usually written in terms of the time variable as
\begin{equation}
\chi(t,s)=s^{-a}\widetilde h\left [\frac{t}{s}\right ].
\label{scalChit}
\ee
Assuming the asymptotic behavior $L(t)\propto t^{\frac{1}{2}}$ this implies
$a =\frac{\alpha}{2}$. In \cite{unquarto} it was found
$a\simeq 0.28$ which implies $\alpha \simeq 0.56$.
Recalling that $\alpha=\lambda _\chi$,
in Fig. \ref{Chi_to_15} it is observed that the exponent $\lambda _\chi =0.56$
describes reasonably well the $x$-dependence of the scaling function $h$ in
Eq. (\ref{scalChi}) up to intermediate values of $x=L(t)/L(s)\simeq 10$.
We will comment later on the occurrence of such intermediate behavior.
For larger values the curves bend slightly and progressively downward
and the exponent $\lambda _\chi=5/8$ looks more consistent with the data
(a fit in the region with $x\ge 10$ provides $\lambda _\chi=0.63$).
Notice that the intermediate exponent $\alpha \simeq 0.56$ was not observed
in the quench to $T_f=0$, signaling that this is due to thermal fluctuations.
The behavior of the response function suggests that the exponent
$\alpha \simeq 0.56$ could be due to a
preasymptotic mechanism associated to thermal fluctuations, while the truly asymptotic behavior
is the one with $\alpha=\lambda_\chi=(1/2)\lambda _C$ as in the quench to $T_f=0$.
Indeed, when trying to obtain data collapse by plotting $L(s)^\alpha \chi(t,s)$
against $L(t)/L(s)$ one finds that curves superimpose better with $\alpha =5/8$
(main figure) than with $\alpha = 0.56$ (inset), although a satisfactory collapse is obtained
in both cases.
This is confirmed in Fig. \ref{Chi2_to_15} where we compare the performance of the two values
$\alpha = 5/8$ (left panel) and $\alpha =0.56$ (right panel) with the method of Sec. \ref{TinfT0},
by using Eq. (\ref{check}).
With $\alpha=5/8$ one observe both data collapse (with small preasymptotic corrections
for small $s$, similarly to the case with $T_f=0$) and linear behavior
$A_{\frac{5}{8}}\sim C$ in the small-$C$ region. On the other hand
with $\alpha=0.56$ both these feature are clearly lost.
Furthermore, in the right panel the curves for small $C$ can be well fitted
by the power law $\sim C^{1.16}$ (bold-dotted green line), as expected on the basis of Eq. (\ref{anticheck})
with $\alpha =5/8$ and $\beta = 0.56$.
Let us now comment on a possible interpretation of the behavior of $\chi $ and,
in particular, of the $\alpha $-exponent. As already mentioned the measured value $\alpha=0.5-0.56$
was interpreted as a result of a putative exponent $\alpha=1/2$ associated to the kinetic roughening of the
domains interfaces whose width is expected to scale as
\begin{equation}
\lambda (t) \simeq A(T_f)L(t)^{\alpha _R},
\label{rough}
\ee
where $\alpha _R$ is the so-called {\it roughness exponent} ($\alpha _R=1/2$ in $d=2$) and
$A(T_f)$ is an increasing function of temperature. It is known that
the extra length $\lambda $ can produce pre-asymptotic corrections to scaling in many observables,
as shown in \cite{interf2d}. However, for large times these corrections can be neglected
because $\lambda (t)$ is eventually dominated by $L(t)$. This is true, for instance, in quantities such as
$G$ or $C$. On the other hand, the statement that the asymptotic
exponent $\alpha $ is the one ($\alpha =1/2$) associated to roughness amounts to assume that
the mechanism whereby the response is built relies only on $\lambda $,
despite the fact that $L\gg \lambda$ is the dominant length. It is interesting to notice that, since
roughness is expected to vanish at zero temperature, i.e. $A(T_f=0)=0$, this mechanism
cannot be sustained in a zero-temperature quench. Indeed we have shown that in this case a different
exponent $\alpha =\lambda _C/2=5/8$ is very neatly observed. The study of the quench to $T_f=1.5$
presented here, upon extending the range of simulated times with respect to the previous ones,
allows us to argue that the value $\alpha =5/8$ is the correct asymptotic one also in a finite-temperature
quench, while a smaller value $\alpha \simeq 0.56$ is only observed pre-asymptotically.
Early-time corrections to the response function are a well known fact and are discussed in \cite{resp2d}.
Notice that
the crossover from the pre-asymptotic roughness-related mechanism to the truly asymptotic one is
regulated by $A(T_f)$: the larger is $T_f$, the larger is $A$ and this makes the pre-asymptotic
behavior with $\alpha =1/2$ to last longer. Finally, let us comment on the fact that the measured
exponent $\alpha$ has been always found larger that $\alpha =1/2$ signaling that also in the
previous determinations the crossover towards $\alpha =5/8$ was very probably already present.
\begin{figure}
\vspace{1cm}
\rotatebox{0}{\resizebox{.85\textwidth}{!}{\includegraphics{L_to_15.pdf}}}
\caption{(Color on-line). $L(t)$ is plotted against $t$ on a log-log plot
for the quench of
the system to $T_f=1.5$ starting from the equilibrium state at $T_i=\infty$
(lower black curve with a circle)
and at $T_i=T_c$ (upper partially collapsing three curves, corresponding to three different
system sizes $\Lambda $ as detailed in the key). The bold green lines are the behavior
$L(t)\sim t^{1/2}$ expected for an infinite system for large times.}
\label{L_to_15}
\end{figure}
\begin{figure}
\rotatebox{0}{\resizebox{.85\textwidth}{!}{\includegraphics{G_to_15.pdf}}}
\caption{(Color on-line). The correlation function $G(r,t)$ for a quench to $T_f=1.5$
is plotted against $y=r/L(t)$
on a double-logarithmic plot at the different times indicated in the key
(these are determined as to be exponentially spaced by an automatic routine ).
The lower set of collapsing curves (without symbols), correspond to the case
discussed in Sec. \ref{TinfT15} of a quench from an equilibrium state at the initial
temperature $T_i=\infty$. The upper group of curves,
marked with a symbol, correspond to the case discussed in Sec. \ref{TcT15} of a quench
starting from the critical state, i.e. $T_i=T_c$.
The bold-green straight line is the power-law $y^{-1/4}$ of Eq. (\ref{critg}).
In the inset a zoom of the same data in the region of small $r/L(t)$ is plotted on a
double-linear plot. The bold-dotted turquoise line is the linear behavior
of Eq. (\ref{porod}), namely the Porod's law.}
\label{G_to_15}
\end{figure}
\begin{figure}
\rotatebox{0}{\resizebox{.85\textwidth}{!}{\includegraphics{C_to_15.pdf}}}
\caption{(Color on-line). $C(t,s)$ is plotted against $x=L(t)/L(s)$
for the quench of
the system to $T_f=1.5$ starting from the equilibrium state at $T_i=\infty$
(lower set of curves)
and at $T_i=T_c$ (upper group of curves) for different times (see key).
The bold dark-green and turquoise lines are the expected
power-laws $x^{-\lambda _C}$ with $\lambda _C=5/4$ and $\lambda _C=1/8$
for the quenches from $T_i=\infty$ and from $T_i=T_c$, respectively.}
\label{C_to_15}
\end{figure}
\begin{figure}
\rotatebox{0}{\resizebox{.85\textwidth}{!}{\includegraphics{Chi_to_15.pdf}}}
\caption{(Color on-line). $L(s)^{\frac{5}{8}}T_f\,\chi(t,s)$ is plotted against $x=L(t)/L(s)$
for the quench of
the system to $T_f=1.5$ starting from the equilibrium state at $T_i=\infty$
(upper set of curves)
and at $T_i=T_c$ (lower group of curves) for different times (see key).
The bold dark-green and dashed-magenta lines are the
power-laws $x^{-\lambda _\chi}$ with $\lambda _\chi=5/8$ and $\lambda _\chi=0.56$.
In the inset, starting from the same data of the main figure for $T_i=\infty$, we plot
$L(s)^{0.56}T_f \chi(t,s)$ against $L(t)/L(s)$.}
\label{Chi_to_15}
\end{figure}
\vspace{2cm}
\begin{figure}
\rotatebox{0}{\resizebox{.45\textwidth}{!}{\includegraphics{Chi2_vs_C_inf_to_15_cinqueottavi.pdf}}}
\rotatebox{0}{\resizebox{.45\textwidth}{!}{\includegraphics{Chi2_vs_C_inf_to_15_056.pdf}}}
\caption{(Color on-line). The function $A_{\frac{5}{8}}(t,s)$ (left panel)
and $A_{0.56}(t,s)$ (right panel)
are plotted against $C(t,s)$ for the quench of
the system to $T_f=1.5$ starting from the equilibrium state at $T_i=\infty$
for different times (see key).
The bold-dotted turquoise line on the left panel is guideline for the linear behavior
$[L(s)^{5/8}\chi (t,s)]^2\propto C(t,s)$. The bold-dotted green one on the right panel
is the behavior $C(t,s)^{1.16}$.}
\label{Chi2_to_15}
\end{figure}
\subsubsection{From $T_i=T_c$} \label{TcT15}
In the case of a quench to the finite final temperature $T_f=1.5$ the differences
between an infinite initial temperature $T_i=\infty$ and a critical one $T_i=T_c$
occur similarly to what observed in the quench to $T_f=0$.
In particular $L(t)$, see Fig. \ref{L_to_15} (upper set of
curves), is considerably larger than the one for $T_i=\infty$
and finite-size effects start occurring around $t\simeq 10^3$ with the
modalities discussed in Sec. \ref{TcT0}.
The behavior of $G(r,t)$, shown in Fig. \ref{G_to_15} (upper group of curves), is also very
similar to the one discussed in Sec. \ref{TcT0}, with scaling obeyed (according to Eq. (\ref{scalG}))
for sufficiently small values of $y=r/L(t)$, a remnant of the critical behavior
of Eq. (\ref{critg}) for an intermediate range of $y$ and marked finite-size effects at large-$y$.
Also the autocorrelation function, plotted against $x=L(t)/L(s)$
in Fig. \ref{C_to_zero} (upper set of curves), closely follows the behavior
already observed in Sec. \ref{TcT0} for a quench to $T_f=0$, with the difference
that the quality of the scaling collapse is worse than before.
Despite this, the behavior $C(x)\sim x^{-\lambda_C}$ expected for large $x$
on the basis of Eq. (\ref{largec}) with
$\lambda _C=1/8$ is well reproduced in an intermediate regime of $x=L(t)/L(s)$ where
finite-size effects are not important.
Analogous considerations apply to the response function, which is shown in Fig. \ref{Chi_to_15}
(lower group of curves). This quantity behaves very similarly to the quench from $T_i=\infty$ to
$T_f=1.5$ discussed in Sec. \ref{TinfT15}. In particular one has a good indication of an exponent
$a=\lambda _\chi=5/8$, while a somewhat smaller value $\lambda _\chi \simeq 0.56$ is compatible with the data at earlier times.
In conclusion, for all the quantities considered we do not find significant differences between
a quench to $T_f=0$ or to $T_f=1.5$ (starting both from $T_i=\infty$ and $T_i=T_c$),
apart from the quality of the scaling which gets worse upon raising $T_f$;
this confirms that $T_f$ is an {\it irrelevant parameter} \cite{brayren} in a renormalization group sense.
Concerning the role of the initial temperature, $\lambda _C$ turns out to be markedly
influenced by $T_i=\infty$ ($\lambda _C=5/4$) or $T_i=T_c$ ($\lambda _C=1/8$),
both with $T_f=0$ and $T_f=1.5$. On the contrary, the response function exponents
are basically independent on $T_i$. Recalling that the memory of the initial condition is
retained at large wavelength this suggests that the contribution of large scales to the response
function is negligible whereas it is important for the autocorrelation.
\subsection{Quenches to $T_f=2.2$} \label{to22}
\subsubsection{From $T_i=\infty$} \label{TinfT22}
When quenching to a temperature as near to the critical one $T_c\simeq 2.26$ as $T_f=2.2$,
preasymptotic effects are so strong to prevent the observation of the expected asymptotic scaling.
This can be seen already from the behavior of the typical length $L(t)$ which is shown in
Fig. \ref{L_to_22}. Here one sees that the growth is slower than the expected one and a fit
for $t\ge 10^4$ yields an effective exponent $1/z_{eff}\simeq 0.44$. Notice that this behavior is
not too far from to the one $L(t)\sim t^{1/z_c}$, with $z_c= 2.1667(5)$ (i.e. $1/z_c\simeq 0.46$),
expected in a critical quench, namely one from $T_i=\infty$ to $T_f=T_c$.
This suggests that the proximity of $T_f$ to the critical point might affect the
behavior of the system at early times. We will confirm that this is the case by studying the
behavior of the autocorrelation function, which is plotted against $x=L(t)/L(s)$
in Fig. \ref{C_to_22} (lower set of curves).
The collapse expected on the basis of Eq. (\ref{scalC}) is
not observed in the range of times accessed in our simulations,
although a tendency of the curves to get closer is observed as time $s$ grows large.
The expected behavior $C(x)\sim x^{-\lambda_C}$ of
Eq. (\ref{largec}) with
$\lambda _C=5/4$ is neither observed. In place of this one has a power-law behavior
$C\sim x^{-1.586}$ for the smaller values of $s$ in an intermediate region of $x$.
This is the expected \cite{noigamba} behavior in a critical
quench, namely the one from $T_i=\infty$ to $T_f=T_c$, for which
\begin{equation}
C= L(s)^{-(d-2+\eta)}F\left [\frac{L(t)}{L(s)}\right ]
\label{critscal}
\ee
with the large-$x$ behavior
\begin{equation}
F(x)\sim x^{(\theta -1)z_c+2-d-\eta}.
\label{cmpl}
\ee
Here $\theta = 0.383(3)$
is the so called {\it initial slip} exponent, $\eta=1/4$, and $L(t)\sim t^{1/z_c}$
(so that the numerical value of the exponent in Eq. (\ref{cmpl}) is $-1.586$).
This suggests quite clearly that, when quenching to a
temperature near to the critical one, data collapse is
delayed because of the influence of the critical point which attracts -- in a
renormalization-group language -- the trajectory of the flow at early times.
Since however $T_f<T_c$, one should observe the same asymptotic
behavior as in a quench to $T_f=0$ if sufficiently large times could be reached.
Indeed, in Fig. \ref{C_to_22} one can notice that the critical behavior $x^{-1.586}$ is
lost for the larger values of $s$ and the curves tend to bend downward at large-$x$,
presumably approaching the expected asymptotic behavior $x^{-5/4}$ at times much
larger than those accessed in our simulations.
A further confirmation that the critical scaling (\ref{critscal}) is obeyed at short
times is given by the analysis of the small-$x$ behavior, which is shown in the two insets
of Fig. \ref{C_to_22}. The upper inset is only a zoom of the main figure and shows that
data collapse is never observed, for any choice of $s$. On the other hand, in the lower inset one sees that,
by plotting $L(s)^{\frac{1}{4}}C$ against $x$,
data collapse is obtained for the smallest values of $s$ in a region of
$x$ which shrinks by increasing $s$,
as it is expected on the basis of the critical scaling (\ref{critscal}).
A behavior similar to the one discussed insofar for the autocorrelation function is displayed
by the equal time correlation function and
by the response function (not shown). Also for these quantities the data collapse expected on the basis of
the scalings (\ref{scalG},\ref{scalChi}) are not observed in the range of times accessed in our simulations
due to important pre-asymptotic effects.
\begin{figure}[h]
\vspace{1cm}
\rotatebox{0}{\resizebox{.85\textwidth}{!}{\includegraphics{L_to_22.pdf}}}
\caption{(Color on-line). $L(t)$ is plotted against $t$ on a log-log plot
for the quench of
the system to $T_f=2.2$ starting from the equilibrium state at $T_i=\infty$
(lower black curve with a circle)
and at $T_i=T_c$ (upper partially collapsing three curves, corresponding to three different
system sizes $\Lambda $ as detailed in the key). The bold green lines are the behavior
$L(t)\sim t^{1/2}$ expected for an infinite system for large times. The dashed-magenta line
is the behavior $t^{}$ expected in a critical quench.}
\label{L_to_22}
\end{figure}
\begin{figure}
\rotatebox{0}{\resizebox{.75\textwidth}{!}{\includegraphics{C_to_22.pdf}}}
\caption{(Color on-line). $C(t,s)$ is plotted against $x=L(t)/L(s)$
for the quench of
the system to $T_f=2.2$ starting from the equilibrium state at $T_i=\infty$
(lower group of curves)
and at $T_i=T_c$ (upper set of curves) for different times (see key).
The bold dark-green and turquoise lines are the expected
power-laws $x^{-\lambda _C}$ with $\lambda _C=5/4$ and $\lambda _C=1/8$
for the quenches from $T_i=\infty$ and from $T_i=T_c$, respectively.
The dashed-indigo line is the power-law $x^{-1.586}$ expected for a quench
from $T_i=\infty$ to the critical point $T_f=T_c$.
In the upper inset a zoom on the small-$x$ part of the same curves of the main figure
is shown. The lower inset contains the same data of the upper one but the rescaled quantity
$L(s)^{1/4}C(t,s)$ is plotted.}
\label{C_to_22}
\end{figure}
\subsubsection{From $T_i=T_c$} \label{TcT22}
The behavior of the characteristic domains size $L(t)$ is shown in Fig. \ref{L_to_22}.
Also in this case long-lasting pre-asymptotic corrections delay the asymptotic
behavior. Indeed, the growth is slower than $L\sim t^{1/2}$
up to times as long as $t\simeq 5\times 10^3$. From $t\simeq 5\times 10^3$ onwards, on the other hand,
finite size effects start to appear, as it can be seen from the fact that curves for different
system sizes start to separate and -- particularly -- because the growth becomes faster than
$L(t)\sim t^{1/2}$. These effects reduce the regime where finite-size effects are absent and scaling
properties can be studied to a very narrow time range (this should compared to the
cases with $T_f=0$ and $T_f=1.5$, where such range is much wider).
As a consequence the data for $G$ and $C$ do not obey the scalings (\ref{scalG},\ref{scalC})
and, when trying to collapse them as in Figs. \ref{G_to_zero},\ref{C_to_zero},\ref{G_to_15},\ref{C_to_15}
the result is poor. This is seen for the autocorrelation function in Fig. \ref{C_to_22}
(the collapse looks better than what really is because the plot is compressed).
Notice however that the exponent $\lambda _c\simeq 1/8$ is quite clearly observed.
It is interesting to observe that, despite the presence of preasymptotic and finite-size effects,
the scaling behavior (\ref{scalChi}) of the response function
is robust. Indeed, as it can be seen in Fig. \ref{Chi_to_22}, data can be collapsed reasonably well
by plotting $L(s)^{\alpha}\chi (t,s)$ against $L(t)/L(s)$. Due to the noisy character of the data
(due to large thermal fluctuations) a precise estimate of the exponent $\alpha $ is not possible.
However, the comparison between the two values $\alpha =5/8$ and $\alpha =0.56$ shows that the latter
provides a slightly better collapse. This is in agreement with the expectation that roughening
effects (which, as discussed above, are associated to an exponent $\alpha =1/2$) are
more pronounced at high $T_f$.
\begin{figure}[h]
\vspace{1cm}
\rotatebox{0}{\resizebox{.45\textwidth}{!}{\includegraphics{Chi_to_22.pdf}}}
\rotatebox{0}{\resizebox{.45\textwidth}{!}{\includegraphics{Chi_to_22_unquarto.pdf}}}
\caption{(Color on-line). $L(s)^{\frac{5}{8}}T_f\,\chi(t,s)$ (left panel) or $L(s)^{0.56}T_f\,\chi(t,s)$
is plotted against $x=L(t)/L(s)$ for the quench of
the system to $T_f=2.2$ starting from the equilibrium state at $T_i=T_c$
for different times (see key).
The bold dark-green line (left panel) is the
power-law $x^{-\lambda _\chi}$ with $\lambda _\chi=5/8$.
The dashed magenta line (right panel) is the
power-law $x^{-\lambda _\chi}$ with $\lambda _\chi=0.56$.}
\label{Chi_to_22}
\end{figure}
\section{The large-N model}
\label{largen}
The dynamics of a classical magnetic system with a vectorial order-parameter with $N$ components
can be exactly solved in the large-$N$ limit \cite{corbcastzan}. This provides an
analytic framework to interpret the behavior of physical systems with finite-$N$ and to
compute the scaling properties at a semi-quantitative level.
In this section, by solving the large-$N$ model for growth kinetics and computing correlation and response
functions we show how some of the features observed in the numerical simulations
can be interpreted in this analytical framework. The present solution closely follows and
generalizes the
one contained in \cite{corberi2002}, to which we generally refer the reader for specific details,
extending the results to the case of a quench from an initial critical state.
\subsection{Model definition}
We consider a $d$-dimensional magnetic system with a vectorial order parameter
$\vec{\phi}=\left( \phi_1,...,\phi_N \right)$
and a Ginzburg-Landau Hamiltonian
\begin{equation}
\mathcal{H}\left[\vec{\phi}\right] = \int_V d^dx \, \left[ \frac{1}{2}\left(\nabla\vec{\phi}\right)^2 + \frac{r}{2}\vec{\phi}^2 + \frac{g}{4N}\left(\vec{\phi}^2\right)^2 \right], \label{hamiltonian-general}
\end{equation}
where $V$ is the volume and $r$ and $g$ are constants ($r<0$, $g>0$).
In the case of a non-conserved order parameter we are considering here the Langevin equation of motion
reads
\begin{equation}
\frac{\partial \vec{\phi}\left(\vec{x},t\right)}{\partial t} = -\frac{\delta \mathcal{H}\left[\vec{\phi}\right]}{\delta \vec{\phi}\left(\vec{x},t\right) } + \vec{\eta}\left(\vec{x},t\right), \label{langevin}
\end{equation}
where $\vec \eta\left(\vec{x},t\right)$ is a Gaussian white noise with expectations
\begin{eqnarray}
\overline {\eta_\alpha \left(\vec{x}\right) } &=& 0 \\
\overline {\eta_\alpha\left(\vec{x}\right) \eta_\beta\left(\vec{x}'\right) } &=& 2T \delta_{\alpha\beta} \delta\left(\vec{x}-\vec{x}'\right) \delta\left( t-t' \right)
\label{gaussian-noise}
\end{eqnarray}
Here $\eta _\alpha $ is the $\alpha$ component of the vector $\vec \eta $, $T$ is the temperature of the thermal bath, and $\overline {\cdots}$ denotes an average over thermal fluctuations, namely over
different realizations of $\vec \eta$.
In the large-$N$ limit the substitution $\vec{\phi}^2(\vec x,t)\to \langle \vec \phi ^2 (\vec x,t)\rangle$, where
$\langle\cdot\rangle$ denotes an ensemble average, namely over thermal noise and initial conditions,
becomes exact. Notice that the quantity $S(t)=\frac{1}{N}\langle \vec \phi ^2 (\vec x,t)\rangle$ does not depend on
$\vec x$ due to space homogeneity.
In terms of the Fourier transform
$\vec \phi \left(\vec{k},t\right) = \int_V d\vec x \, \vec{\phi}\left(\vec{x},t\right) \exp\left(i\vec{k}\cdot \vec{x}\right)$ of $\vec \phi (\vec x,t)$ the evolution (\ref{langevin}) then reads
\begin{equation}
\frac{\partial \vec{\phi}\left(\vec{k},t\right)}{\partial t} = -\left[k^2 + I(t)\right]\vec{\phi}\left(\vec{k},t\right) + \vec{\eta}\left(\vec{k},t\right), \label{foutrans-eq-motion}
\end{equation}
where $I(t)$ is the self-consistent function
\begin{equation}
I(t)= r + gS(t).
\label{self-consist-func}
\end{equation}
\subsection{Non-equilibrium dynamics}
The general solution of the formally linear Eq.~(\ref{foutrans-eq-motion}) is
\begin{equation}
\phi(\vec{k},t)={\cal R}(k,t,0)\,\phi (\vec{k},0)+\int\limits_0^t d t' {\cal R}(k,t,t')\,\eta(\vec{k},t'), \label{generalsolution}
\end{equation}
where we denote by $\phi $ one of the equivalent components
$\vec \phi _\alpha $ of the order parameter and the response function ${\cal R}$,
which only depends on the modulus $k=\vert \vec k \vert$ of the wave-vector, is given by
\begin{equation}
{\cal R}(k,t,t') = \frac{Y(t')}{Y(t)} e^{-k^2 (t-t')}, \label{response-function}
\end{equation}
with $Y(t) = \exp\left[ \int_0^t d s I(s) \right]$, $I(t)$ being the self-consistent function defined
in~(\ref{self-consist-func}). The squared quantity $Y^2(t)$ obeys the following differential equation
\begin{equation}
\frac{d Y^2(t)}{d t} = 2 I(t) Y^2(t). \label{diff-eq-selfconsisfac}
\end{equation}
This equation contains, hidden in $I(t)$, the unknown $S(t)=\langle \phi^2 \rangle $ which is related to
the structure factor (the Fourier transform of the equal time correlation function $G(r,t)$)
\begin{equation}
{\cal C}(k,t)=\langle \phi (\vec k,t) \phi (-\vec k,t)\rangle
\label{strucfac}
\ee
by
\begin{equation}
\langle \phi^2 \left( \vec{x},t \right) \rangle =\frac{1}{(2\pi)^d}\int d\vec k \, {\cal C}(k,t)
e^{-\frac{k^2}{\Lambda ^2}},
\ee
where the smooth cut-off around $|\vec k| \sim \Lambda$ mimics the presence of
a lattice and regularizes the theory in the ultraviolet sector, and $d$ is the spatial dimension.
Using Eq. (\ref{generalsolution}) to build products of order parameter fields and
plugging them into Eq. (\ref{strucfac}), using the expectations of the noise (\ref{gaussian-noise})
one arrives at the following expression
\begin{equation}
{\cal C}(k,t) = {\cal R}^2\left(\vec{k},t,0\right) {\cal C}(k,0) + 2 T \int\limits_0^t d t' {\cal R}^2 \left(\vec{k},t,t'\right). \label{strfac-respfunc}
\end{equation}
We now specify the initial configuration of the order parameter as
\begin{eqnarray}
\langle \phi(\vec{k},0)\rangle &=& 0 \nonumber \\
\langle \phi(\vec{k},0)\phi(\vec{k}',0)\rangle &=& (2\pi)^d
\frac{\Delta}{k^\mu} \delta(\vec{k}+\vec{k}')
\label{initial condition}
\end{eqnarray}
This form contains, as special cases, an initial equilibrium state at $T_i=\infty$,
with the choice $\mu =0$, and that of a critical state at $T_c$
for $\mu = 2$, since in the large-$N$ model the critical exponent $\eta $ is
$\eta =0$ \cite{corberi2002}. The solution presented below, however,
is valid also for correlated initial states with different values of $\mu $.
We consider a quench from such an initial condition to
the final temperature $T_f$.
Using the above initial conditions in Eq. (\ref{strfac-respfunc}) one has
\begin{equation}
{\cal C}(\vec{k},t) = {\cal R}^2\left(\vec{k},t,0\right) \frac{\Delta}{k^\mu} + 2 T_f \int\limits_0^t d t' {\cal R}^2 \left(\vec{k},t,t'\right)
\label{strfac-respfunc2}
\end{equation}
and inserting this form into Eq. (\ref{diff-eq-selfconsisfac}) one is left with the following
integro-differential equation
\begin{equation}
\frac{d Y^2(t)}{d t} = 2 r Y^2(t) + 2 g \Delta f\left(t+\frac{1}{2 \Lambda^2} ; \mu \right) + 4 g T_f \int\limits_0^t d t' f\left( t-t'+\frac{1}{2 \Lambda^2} ; 0 \right) Y^2(t) \label{diff-eq-ysqr}
\end{equation}
where
\begin{equation}
f(x;\mu)=\frac{1}{(2\pi)^d}\int d \vec k \,k^{-\mu} e^{ -2k^2 x } = 2^{-\frac{3d-\mu}{2}}\pi^{-\frac{d}{2}} x^{-\frac{d-\mu}{2}} \frac{\Gamma\left( \frac{d-\mu}{2} \right)}{\Gamma\left( \frac{d}{2} \right)} \label{g-f-definitions}
\end{equation}
and $\Gamma $ is the Euler special function.
Eq. (\ref{diff-eq-ysqr}) is a closed equation for the quantity $Y$. Its solution will be discussed in Appendix \ref{appA},\ref{appB}. As we will show in Sec. \ref{largeNcorr}, any observable can be expressed in terms of $Y$.
\subsection{Observables} \label{largeNcorr} \label{secstrfac-respfunc}
From the knowledge of $Y(t)$ it is possible to compute the two-time quantities considered
in this paper. Extending the definition (\ref{strucfac}) to a two-time correlation
\begin{equation}
{\cal C}(k,t,s)=\langle \phi (\vec k,t) \phi (-\vec k,s)\rangle
\label{strucfac2}
\ee
and using the expression (\ref{generalsolution}) to build the $\phi$-product one arrives at
\begin{equation}
{\cal C}\left(\vec{k},t,s\right) = {\cal R}\left(\vec{k},t,0\right) {\cal R}\left(\vec{k},s,0\right) \frac{\Delta}{ k^{\mu}} + 2T_f \int\limits_0^{s} dt' {\cal R}\left(\vec{k},t,t'\right) {\cal R}\left(\vec{k},s,t'\right),
\label{two-time-structure-factor2}
\end{equation}
from which the autocorrelation function is readily obtained as
\begin{equation}
C(t,s)=\frac{1}{(2\pi)^d}\int d\vec k\,
{\cal C}(k,t,s)e^{-\frac{k^2}{\Lambda ^2}}.
\ee
Using the expression (\ref{response-function}) one arrives at
\begin{equation}
C(t,s)=\frac{1}{Y(t)Y(s)}\left [ f \left (\frac{t+s}{2}+\frac{1}{2\Lambda ^2};\mu\right ) \Delta
+2T_f \int _0 ^s dt' f\left (\frac{t+s}{2}-t'+\frac{1}{2\Lambda ^2};0\right ) Y^2(t') \right ].
\ee
The first term on the r.h.s. is responsible for the aging properties \cite{corberi2002}. Using
the expression (\ref{g-f-definitions}) for $f$ and the expressions for $Y$ derived
in Appendix \ref{appA}, focusing on the large time sector
$t+s\gg \frac{1}{\Lambda ^2}$ one has
\begin{equation}
C(t,s)=M^2 \left [\frac{4x}{x^2+1}\right ]^{\frac{(d-\mu)}{2}}
\ee
where $M^2=-\frac{r}{g} \frac{T_c-T_f}{T_c}$ is a constant
[the critical temperature of the large-$N$ model is
$T_c=\frac{-r (4\pi)^{d/2}}{2g\Lambda ^{d-2}}(d-2)$],
$x=L(t)/L(s)$ and $L(t)\sim t^{1/2}$.
This result shows that the autocorrelation function takes the general scaling form of
Eq. (\ref{scalC}) and that
\begin{equation}
\lambda _C=\frac{d-\mu}{2}.
\ee
Therefore, there is a memory of the initial condition -- through the value of $\mu $ --
in the exponent $\lambda _C$. Notice that going from $T_i=\infty$ (i.e. $\mu =0$)
to $T_i=T_c$ (i.e. $\mu =2$) reduces the autocorrelation exponent, as it also true in the
Ising model (see Sec. \ref{SIMUL}). The actual value of this exponent in the
large-$N$ model is different from the one observed in the scalar case,
as expected.
Let us now consider the response function. It is easy to show \cite{corberi2002} that
the impulsive auto-response $R(t,t')$ defined in Eq. (\ref{defR}) is related to
${\cal R}(k,t,t')$ by
\begin{equation}
R(t,t')=\frac{1}{(2\pi )^d}\int d\vec k \,{\cal R}(k,t,t')\,e^{-\frac{k^2}{\Lambda ^2}}.
\ee
Using the expressions (\ref{response-function},\ref{g-f-definitions}) and the behavior
of $Y(t)$ derived in Appendix \ref{appA} to compute this
quantity and plugging the result in the definition (\ref{defChi}) of the integrated response
(the quantity measured in the numerical simulations of Sec. \ref{SIMUL}) one
obtains, for large $s$, the general scaling form (\ref{scalChi}) with
\begin{equation}
\alpha = d-2
\label{alfaninf}
\ee
and
\begin{equation}
h(x)=(4\pi )^{-\frac{d}{2}}x^{2-d}\int _{x^{-2}}^1 dz \,z^{-\frac{d-\mu}{4}}\,
(1-z)^{-\frac{d}{2}}.
\label{scalhninf}
\ee
Eq. (\ref{alfaninf}) shows that the response exponent $\alpha$ is independent
of $\mu $, therefore it is not touched by changing the initial condition.
This is indeed what we found in Sec. \ref{SIMUL} also in the simulations
of the Ising model. Notice that for $2<d\le 4+\mu$ (a range where all physically
relevant cases are included), since the integral in Eq. (\ref{scalhninf}) converges,
one finds the general behavior (\ref{largeh}) and the constraint (\ref{lameqalf}),
so that also $\lambda _\chi$ does not change with $T_i$. The shape of the scaling
function $h$, instead, changes. In particular, the effect of raising $\mu$ is to lower $h(x)$,
as indeed it was found also in the Ising model (see Sec. \ref{SIMUL}) since the response function
is smaller with $T_i=T_c$ than with $T_i=\infty$.
In the large-$N$ model the different sensitivity of the correlation and of the response function
exponents have a clear mathematical origin.
We have seen that the initial condition plays a role in determining the
time-behavior of the self-consistent quantity $Y$, Eq. (\ref{mainresy}).
Eq. (\ref{response-function}) shows that this different time-behavior is the only
effect of the initial condition on the wave-vector-resolved response function ${\cal R}(k,t,s)$.
The situation is different for the autocorrelation function (\ref{strucfac2}).
Indeed, Eq. (\ref{two-time-structure-factor2}) shows that the different spatial organization
of the correlation is explicitly determined by the one in the
initial state through the factor $\frac{\Delta }{k^\mu}$. This makes the effect of a different
initial condition much more pronounced than in the response function, producing a different
exponent $\lambda _C$. It can be observed that the role of the factor $\frac{\Delta}{k^\mu}$ in
Eq. (\ref{two-time-structure-factor2}) is such to weight more the contribution of
the small wave-vectors if the initial condition is critical than in the case of a disordered ones.
We have already noticed in Sec. \ref{SIMUL} that in the scalar case
the different behavior of the system at
large distances, where memory of the initial condition is retained, might be the origin of the different value
of the exponent $\lambda _C$ in the quenches from $T_i=\infty$ or from $T_i=T_c$.
We see here that a similar property is shared by the analytically tractable large-$N$ model.
\section{Summary and conclusions} \label{concl}
In this paper we have discussed the results of a rather general investigation of the
phase-ordering process observed in a ferromagnetic system described by the
Ising model with Glauber single-spin flip dynamics quenched from equilibrium
states at infinite temperature or at the critical one $T_c$. We have considered three values of
the final quench temperature $T_f$ in order to scan the region $0\le T_f <T_c$.
When a deep quench is made from an infinite initial temperature $T_i$ to a vanishing
one $T_f=0$, all the quantities considered show an excellent agreement with the expected
dynamical scaling forms. Similar results are also found in quenches from
the critical state, but severe finite-size effects restrict the region where scaling is observed
to a much smaller time/space region than in the quench from $T_i=\infty$, as an effect
of the correlated initial state. Upon raising $T_f$ (for any choice of the initial state),
the quality of the data collapse predicted by dynamical scaling gets progressively poorer
until, in a shallow quench at $T_f=2.2$, scaling is basically lost. This is due partly to the
stationary term of Eq. (\ref{calG}) becoming more important and to
the relevance of pre-asymptotic effects due to the proximity of the critical point.
Besides providing a comparative study of the effects of changing the initial and the final
quench temperature, our study includes the first determination of the response function
in quenches from $T_i=T_c$ and in those to $T_f=0$. We have shown that, while
starting from $T_i=T_c$ instead of $T_i=\infty$
does change the aging properties of the process,
as witnessed by the markedly different behavior of the autocorrelation function,
the universal properties of the response function are basically insensitive to the different initial conditions.
The same effect is found in the behavior of the exactly solvable large-$N$ model
and has been interpreted as due to the large sensitivity of the autocorrelation
function to the large-scale properties of the system, which are reminiscent of the
correlated initial configuration, whereas the largest contributions to the response
are provided at small scales.
The computation of the dynamical susceptibility in a quench to a vanishing final temperature
allows a rather accurate determination of the response function exponent which turns out
to be consistent with $\alpha=\frac{1}{2}\lambda _C=5/8$, at variance with previous
determinations lying in the range $[0.5-0.56]$ obtained at
finite $T_f$. This seems to rule out the
conjecture that the value $a=\frac{1}{2}$ is the asymptotic one at $T_f=0$.
Instead, our data at $T_f>0$ suggest that the smaller exponent in this case might be a pre-asymptotic
effect associated to the thermal roughening of interfaces.
The observed relation $\alpha = \frac{1}{2} \lambda _C$ has presently no physical
interpretation.
We hope that the results of this paper will refresh the attention on the non-equilibrium
response exponent possibly providing a thorough understanding.
\vspace{1cm}
{\bf Acknowledgments}
We thank Eugenio Lippiello and Marco Zannetti for discussions.
We acknowledge financial support by MURST PRIN 2010HXAW77\_005.
|
\section{$\tmin$-plots for moving pions}\label{a:pi}
All $\tmin$-plots for fits to single-pion correlation functions at various momenta are shown in Fig.~\ref{f:pi_tmins}. These energies are used in Strategy 1
discussed in Sec.~\ref{s:ens} to determine the renormalized anisotropy $\xi$.
\begin{figure}[!htb]
\includegraphics[width=0.32\textwidth]{level_00_tminplot_fitval.pdf}
\includegraphics[width=0.32\textwidth]{level_01_tminplot_fitval.pdf}
\includegraphics[width=0.32\textwidth]{level_02_tminplot_fitval.pdf}
\includegraphics[width=0.32\textwidth]{level_03_tminplot_fitval.pdf}
\includegraphics[width=0.32\textwidth]{level_04_tminplot_fitval.pdf}
\includegraphics[width=0.32\textwidth]{level_05_tminplot_fitval.pdf}
\includegraphics[width=0.32\textwidth]{level_06_tminplot_fitval.pdf}
\includegraphics[width=0.32\textwidth]{level_08_tminplot_fitval.pdf}
\includegraphics[width=0.32\textwidth]{level_09_tminplot_fitval.pdf}
\caption{\label{f:pi_tmins}$\tmin$-plots of single-pion energies
for all total momenta $0\le\boldsymbol{d}^2\le9$. The chosen
$t_{\mathrm{min}}$ is indicated by the solid line (central value) and dotted lines (1$\sigma$ errors) as well as the black square.}
\end{figure}
\section{$\tmin$-plots for all $I = 1$ levels}\label{s:app1}
All $t_{\mathrm{min}}$-plots for finite volume energies
used in the determination of the $I=1$, $\ell=1$ elastic scattering amplitude are shown in
Figs.~\ref{f:i1d0},~\ref{f:i1d1},~\ref{f:i1d2},~\ref{f:i1d3}, and~\ref{f:i1d4}. The ratio
fits of Eq.~\ref{e:rat} are employed and the dimensionless center-of-mass momentum $u^2$
is shown.
\begin{figure}[!htp]
\centering
\includegraphics[width=0.32\textwidth]{isotriplet/mom_000_t1up_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_000_t1up_usq_level_1_tmin_final.pdf}
\caption{\label{f:i1d0} $t_{\mathrm{min}}$-plots of
the dimensionless center-of-mass momentum $u^2$ for $I=1$,
$\boldsymbol{d}^2=0$. The chosen $\tmin$ is indicated by the black square and the lines.}
\end{figure}
\begin{figure}[!htb]
\includegraphics[width=0.32\textwidth]{isotriplet/mom_001_a1p_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_001_a1p_usq_level_1_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_001_ep_usq_level_0_tmin_final.pdf}
\caption{\label{f:i1d1}Same as Fig.~\ref{f:i1d0} but for $I=1$,
$\boldsymbol{d}^2=1$.}
\end{figure}
\begin{figure}[!htb]
\includegraphics[width=0.32\textwidth]{isotriplet/mom_011_a1p_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_011_a1p_usq_level_1_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_011_b1p_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_011_b2p_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_011_b2p_usq_level_1_tmin_final.pdf}
\caption{\label{f:i1d2}Same as Fig.~\ref{f:i1d0} but for $I=1$,
$\boldsymbol{d}^2=2$.}
\end{figure}
\begin{figure}[!htb]
\includegraphics[width=0.32\textwidth]{isotriplet/mom_111_a1p_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_111_a1p_usq_level_1_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_111_a1p_usq_level_2_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_111_ep_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_111_ep_usq_level_1_tmin_final.pdf}
\caption{\label{f:i1d3}Same as Fig.~\ref{f:i1d0} for $I=1$, $\boldsymbol{d}^2=3$.}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.32\textwidth]{isotriplet/mom_002_a1p_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_002_a1p_usq_level_1_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_002_ep_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isotriplet/mom_002_ep_usq_level_1_tmin_final.pdf}
\caption{\label{f:i1d4}Same as Fig.~\ref{f:i1d0} but for
$I=1$, $\boldsymbol{d}^2=4$.}
\end{figure}
\section{$\tmin$-plots for all $I = 2$ levels}\label{s:app2}
As in App.~\ref{s:app1}, this appendix contains $t_{\mathrm{min}}$-plots for all finite volume energies
used in the determination of the $I=2$, $\ell=0$ scattering amplitude. They are shown in
Figs.~\ref{f:i2d0},~\ref{f:i2d1},~\ref{f:i2d2}, and~\ref{f:i2d3}.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.32\textwidth]{isoquintet/mom_000_a1gp_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isoquintet/mom_000_a1gp_usq_level_1_tmin_final.pdf}
\caption{\label{f:i2d0}$t_{\mathrm{min}}$-plots of
the dimensionless center-of-mass momentum $u^2$ for $I=2$,
$\boldsymbol{d}^2=0$. The chosen $\tmin$ is indicated by the black square and the lines.}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.32\textwidth]{isoquintet/mom_001_a1p_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isoquintet/mom_001_a1p_usq_level_1_tmin_final.pdf}
\caption{\label{f:i2d1}Same as Fig.~\ref{f:i2d0} but for
$I=2$, $\boldsymbol{d}^2=1$.}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.32\textwidth]{isoquintet/mom_011_a1p_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isoquintet/mom_011_a1p_usq_level_1_tmin_final.pdf}
\caption{\label{f:i2d2}Same as Fig.~\ref{f:i2d0} but for
$I=2$, $\boldsymbol{d}^2=2$.}
\end{figure}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.32\textwidth]{isoquintet/mom_111_a1p_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isoquintet/mom_111_a1p_usq_level_1_tmin_final.pdf}
\caption{\label{f:i2d3}Same as Fig.~\ref{f:i2d0} but for
$I=2$, $\boldsymbol{d}^2=3$.}
\end{figure}
\begin{figure}[!htb]
\includegraphics[width=0.32\textwidth]{isoquintet/mom_002_a1p_usq_level_0_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isoquintet/mom_002_a1p_usq_level_1_tmin_final.pdf}
\includegraphics[width=0.32\textwidth]{isoquintet/mom_002_a1p_usq_level_2_tmin_final.pdf}
\caption{\label{f:i2d4}Plots showing the $t_{\mathrm{min}}$-dependence of
the dimensionless center-of-mass momentum $u^2$ for $I=2$,
$\boldsymbol{d}^2=4$.}
\end{figure}
\section{Conclusions}\label{s:concl}
The elastic $I=1$ and $I=2$ $\pi-\pi$ scattering phase shifts are determined from a $\nf=2+1$ dynamical lattice QCD simulation in a
large spatial volume with a light pion mass. In particular, the stochastic
estimation scheme employed here performs efficiently, and determines the
correlation functions with sufficient precision to extract the finite-volume energies and
scattering phase shifts.
This suggests that larger volumes and lighter pions are possible due to the
favorable scaling of the stochastic LapH method.
After extracting finite-volume energy levels, the L\"{u}scher method is employed to
calculate elastic scattering phase shifts.
The $I=1$, $\ell=1$ partial wave is well described by a Breit-Wigner form and
exhibits rapid phase motion indicative of a resonance.
Our main results are Fig.~\ref{f:i1scat} and Eq.~\ref{e:rhofit}.
We have compiled recent published calculations of the $\rho$-resonance in
Fig.~\ref{f:rho_sum}
\begin{figure}
\includegraphics[width=0.49\textwidth]{mrho_summary_final.pdf}
\includegraphics[width=0.49\textwidth]{g_summary_final.pdf}
\caption{\label{f:rho_sum}Summary of recent published work for $m_{\rho}$ and
$g_{\rho\pi\pi}$. The legend denotes Ref.~\cite{Dudek:2012xn} `Dudek et al.',
Ref.~\cite{Feng:2010es} `ETMC', Ref.~\cite{Lang:2011mn} `Lang, et al.',
Ref.~\cite{Wilson:2015dqa} `Wilson et al.', Ref.~\cite{Feng:2014gba} `Feng, et al.', Ref.~\cite{Pelissier:2012pi} `GWU', and Ref.~\cite{Aoki:2011yj}
`PACS-CS'.}
\end{figure}
indicating that this calculation (together with Ref.~\cite{Wilson:2015dqa}) is
the closest to the physical quark masses achieved so far. Fig.~\ref{f:rho_sum}
compares $m_{\rho}/m_{\pi}$ to reduce scale uncertainties, as none of the
results are extrapolated to the continuum limit.
The results for
the mass are generally in good agreement, but $g_{\rho\pi\pi}$ is known
with considerably less precision.
Due to our light quark
masses, the lowest inelastic threshold (due to four pions) is close to the
resonance region limiting
the applicability of the elastic L\"{u}scher formulae. Hopefully, existing
work on extending the L\"{u}scher formulae to three-particle
scattering~\cite{Polejaeva:2012ut,Hansen:2014eka,Hansen:2015zga} can
be adapted to treat these thresholds in the future. Of course, the problem
worsens as the quark masses are lowered to their physical values as
experimentally $m_{\rho} > 4m_{\pi}$. Once this threshold
can be treated quantitatively its effect may be small, as the
experimental branching fraction for $\rho\rightarrow4\pi$ is below the percent
level.
We have less points below inelastic threshold for the $I=2$, $\ell=0$ partial
wave, as there are fewer lattice irreps in which it appears. Still,
our data below the $t$-channel cut $q_{{\mathrm {cm}}} = m_{\pi}$ is well-described by the
first two terms in the effective range expansion and provides a determination of the scattering length to about 20\%. Our results for $I=2$
are shown in Fig.~\ref{f:i2scat} and Eq.~\ref{e:i2fit}. Calculations of the $I=2$ $s$-wave scattering length\footnote{For a recent review of
these calculations see Ref.~\cite{Helmes:2015gla} and the references quoted therein.}
are considerably more advanced than in $I=1$, so a single-ensemble result is not fit for direct comparison.
However, the $\approx 20\%$ error on $a_{0}$ is somewhat remarkable given our stochastic estimation of the all-to-all quark propagators
and the precise calculation of small energy shifts required to obtain a signal.
As mentioned in the introduction, Ref.~\cite{Wilson:2015dqa} appeared
during the preparation of this manuscript which uses the full
distillation method to treat the required all-to-all propagators and
can be viewed as the maximal dilution limit of our approach.
We compare results in Tab.~\ref{t:comp} for a selection of published
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|c|}
\hline
Ref. & $N_{D}$ & $a_{t}m_{\pi}$ & $T_{1u}^{+}$ $E_0$ &
$T_{1u}^{+}$ $E_1$ & $a_tm_{\rho}$ & $g_{\rho\pi\pi}$ \\
\hline
This work & 2304 & 0.03939(19) & 0.12625(94) & 0.1470(16) &
0.13190(87) & 5.99(26) \\
\hline
Ref.~\cite{Wilson:2015dqa} & 393216 & 0.03928(18) & 0.12488(40)
& 0.14534(52) & 0.13175(35) & 5.688(70)\\
\hline
\end{tabular}
\caption{\label{t:comp}Comparison of the number of Dirac matrix inversions per
configuration $N_D$, the pion mass, the first two energies in the $I=1$,
$\boldsymbol{d}^2=0$, $T_{1u}^{+}$ channel, and the $\rho$ resonance parameters
between this work and Ref.~\cite{Wilson:2015dqa}. We see that while the
precision on the pion mass is comparable, the distillation method is
2-3 times more precise for the other quantities while requiring about
170 times more Dirac matrix inversions.}
\end{table}
numbers as well as the required number of Dirac matrix inversions per
configuration. Although $a_tm_{\rho}$ and $g_{\rho\pi\pi}$ are also obtained
in Ref.~\cite{Wilson:2015dqa} from a Breit-Wigner ansatz, their fitting method
constructs a correlated-$\chi^2$ directly from the finite-volume energies
rather than $(q_{{\mathrm {cm}}}/m_{\pi})^{3}\cot \delta_1$. However, the errors on
$m_{\pi}$ and $\xi$ (which are comparable to those on the energies) are not
taken into account in their fit procedure. It is unclear what effect this has
on the resultant fit parameters and their errors. Our methods for extracting
finite-volume energies are also different from those employed in
Ref.~\cite{Wilson:2015dqa}.
We see that the distillation results are comparable in precision for the pion,
but have roughly half the statistical error for two-pion states,
while requiring about 170 times more Dirac matrix inversions per configuration. The Dirac matrix inversion cost for the distillation method is significantly
larger than the cost for the gauge generation and does
not include the (sizeable) cost of constructing correlation functions from the
sources and solutions, which also scales poorly with
the volume. Even so, presumably a 170-fold increase in computational effort
would reduce the error on our results by more than an order of
magnitude, making them significantly more precise than Ref.~\cite{Wilson:2015dqa}.
Overall, this first large-volume scattering calculation using stochastic LapH
is promising for future work.
As it is clear
that scattering calculations are entering a new era of increased
statistical precision, it is important to quantify the remaining systematic
errors. These include exponential finite-volume effects,
the effect of higher partial waves, the presence of inelastic thresholds, and
lattice spacing effects.
To this end, work has progressed~\cite{Bulava:2015qjz} in applying the stochastic LapH method to state-of-the-art ensembles generated by
the Coordinate Lattice Simulations (CLS) consortium~\cite{Bruno:2014jqa}. Apart from the elastic scattering phase shifts presented here, these
isotropic ensembles simplify the renormalization and $\mathrm{O}(a)$-improvement pattern of composite operators, enabling the determination
of resonance matrix elements. Preliminary work on the simplest such matrix element, the timelike pion form factor, is also reported in Ref.~\cite{Bulava:2015qjz}.
Finally, pushing to lighter pions would be desired. While this can be done
using these CLS ensembles, the lower inelastic thresholds limit the
applicability of the
L\"{u}scher formula. More theoretical work is required to rigorously treat
these thresholds.
\section{Introduction}
Hadron-hadron scattering amplitudes are of central importance in the
phenomenology of QCD and confining scenarios of Beyond-the-Standard Model (BSM)
physics. While Euclidean lattice gauge simulations are a
proven first-principles approach for these theories,
the calculation of hadron-hadron scattering on the lattice has long
been a challenge. First and foremost, the Maiani-Testa No-Go Theorem demonstrates that
on-shell amplitudes cannot (in general) be directly obtained from
Euclidean space matrix elements~\cite{Maiani:1990ca}. This
difficulty was overcome by L\"{u}scher's
relation between elastic scattering
phase shifts and the deviation of finite-volume two-hadron energy spectra
from their non-interacting values~\cite{Luscher:1990ux}.
While this relation has been known since the early 90's, only recently are
lattice QCD calculations of scattering amplitudes starting to have sufficient
statistical precision and energy resolution to clearly identify resonance
features. This delay is mostly due to the difficulty in precisely calculating
temporal correlation functions
\begin{align}\label{e:cor}
C_{ij}(t-t_0) = \langle \mathcal{O}_{i}(t) \bar{\mathcal{O}}_j(t_0) \rangle =
\sum_{n} A_{ni}A^{*}_{nj} \mathrm{e}^{-E_n (t-t_0)},
\end{align}
where $\hat{\mathcal{O}}_i$ and $\hat{\mathcal{O}}_j$ are suitable
interpolating operators with the quantum numbers of interest and the sum is over all finite-volume energy eigenstates. After calculating such correlation functions on a gauge field
ensemble, the finite-volume energies $\{E_n\}$ are extracted from their
temporal fall-off.
To obtain finite-volume two-hadron energies,
correlation functions between two-hadron interpolating operators are
required. These two-hadron correlation functions in turn typically
require the evaluation of valence-quark-line-disconnected Wick contractions\footnote{`Disconnected' Wick contractions
are those in which quark fields at the same time are contracted.} and
contain interpolating operators which annihilate states with definite momentum.
After integration over the Grassmann-valued quark fields, this requires
the quark propagator from all space-time points to all
space-time points. As the quark propagator is the inverse of the
large-dimension and ill-conditioned Dirac matrix, these `all-to-all' propagators (and
thus multi-hadron correlation functions) are
naively intractable. Inversion of the Dirac matrix $M$ is performed by solving
the linear system $M\phi = \eta$ for multiple right-hand sides and is
typically the dominant cost in calculating fermionic correlation
functions. The solution of this system for each spacetime point is not
feasible, preventing the naive approach to all-to-all quark propagators.
However, substantial progress has been made by treating quark propagation only
in the subspace spanned by the lowest-lying eigenmodes of the
three-dimensional gauge-covariant Laplace operator~\cite{Peardon:2009gh}.
Apart from facilitating the evaluation of these correlation functions, this
`distillation' procedure has the added benefit of reducing the contamination
of unwanted excited states. It can thus be viewed as a form of `quark smearing', a
common procedure used in lattice QCD to reduce the contribution
of higher terms in Eq.~\ref{e:cor} by suppressing their overlaps. The spatial profile of this smearing
wavefunction is approximately gaussian with a width controlled by the number
of low-lying Laplacian eigenmodes retained in the projection ($N_v$).
The cutoff
eigenvalue therefore defines the smearing wavefunction and must be
fixed in physical units.
Unfortunately, if the cutoff eigenvalue is held fixed the number of eigenmodes
in this subspace increases proportionally to the spatial volume. The
distillation approach requires a number of Dirac matrix inversions $N_D \propto N_v$ which results in an
unfavorable volume scaling, hindering the application
of this method to large physical volumes. Nonetheless, it has been applied
successfully in smaller volumes~\cite{Wilson:2014cna,Lang:2014yfa,
Prelovsek:2013ela,Lang:2011mn,Dudek:2012xn,Dudek:2012gj,Lang:2012db}.
Based on this idea, the stochastic LapH method was proposed in
Ref.~\cite{Morningstar:2011ka} and achieves an improved scaling with the
physical volume by introducing stochastic estimators in the low-dimensional
subspace spanned by the Laplacian eigenmodes. The variance of these
stochastic estimators is reduced by `dilution'~\cite{Foley:2005ac}, which
partitions the space using $N_{\mathrm{dil}}$ complete, orthogonal projectors.
Ref.~\cite{Morningstar:2011ka} demonstrates that the efficiency of these
modified stochastic estimators remains constant for fixed (sufficiently
large) $N_{\mathrm{dil}}$ as the spatial volume is increased. Since
$N_D \propto N_{\mathrm{dil}}$ in this approach, the volume scaling is significantly
improved.
This scaling is tested further in this work by applying the
stochastic LapH method for the first time to lattices with
$L= 3.7\mathrm{fm}$, while it has been successful in smaller
volumes~\cite{Lang:2014yfa,Helmes:2015gla}. Although the stochastic LapH
method was designed to enable exploratory calculations of finite-volume
spectra, we demonstrate here that it can resolve
these energies with a sufficient precision to determine elastic scattering
phase shifts in a large spatial volume.
As a first large-volume application we treat $\pi-\pi$ scattering in the $I=1$
and $I=2$ channels. The lowest-lying hadronic resonance, the $\rho$-meson,
occurs in the $\ell=1$ partial wave
of the $I=1$ channel, resulting in significant shifts of finite-volume
energies from their non-interacting values. In contrast
the $I=2$, $\ell=0$ partial
wave is considerably more weakly interacting and well-described by the
effective range expansion. Therefore, this channel presents a
more stringent test of the stochastic LapH method as deviations from
non-interacting energies are generally much smaller. For example, in large
volume the
difference between the ground-state energy in the $I=2$ $A_{1g}^{+}$ channel (relevant for the $\ell=0$ partial wave)
and $2m_{\pi}$ is given by
\begin{align}
\Delta E = E_{2\pi} - 2\mpi = -\frac{4\pi a_0}{m_{\pi}L^3} +
\mathrm{O}(L^{-4})
\end{align}
where $a_0$ is the $I=2$ $s$-wave scattering length. Although additional
statistics are accrued by summing over a large spatial
volume, the signal in this channel also decreases with the spatial volume,
complicating the determination of $a_0$.
Although these two systems are benchmark tests of the efficacy of our methods,
they are
also interesting in their own right. The quark mass
dependence of the $\rho$-resonance pole position is an
important input to Unitarized Chiral Perturbation Theory (see e.g. Refs.~\cite{Hanhart:2014ssa,Bolton:2015psa}) while
the $\ell=0$ scattering length in the $I=2$ channel provides another important
test of Chiral Perturbation Theory.
This work is part of an ongoing effort to investigate the low-lying resonance
spectrum of QCD.
Preliminary work using only single-hadron interpolating operators has been
reported in
Refs.~\cite{Basak:2005aq,Bulava:2009jb,Bulava:2010yg}
while development of the all-to-all propagator algorithms discussed above is detailed in
Refs.~\cite{Peardon:2009gh,Morningstar:2011ka}. First results with multi-hadron operators are given in
Ref.~\cite{Morningstar:2013bda} and a preliminary account of the results shown here is given in
Refs.~\cite{Fahy:2014jxa,Bulava:2015qjz}.
During the preparation of this manuscript, a calculation of the $I=1$ $p$-wave
scattering phase shift appeared~\cite{Wilson:2015dqa} using the same ensemble
of gauge configurations. Rather than stochastic LapH,
Ref.~\cite{Wilson:2015dqa} employs the full distillation method of Ref.~\cite{Peardon:2009gh}.
Comparison of results and computational cost with Ref.~\cite{Wilson:2015dqa} is made in Sec.~\ref{s:concl}.
\newpage
The main results of this work are Figs.~\ref{f:i1scat} and~\ref{f:i2scat}, which show the $I=1$ $p$-wave and $I=2$ $s$-wave scattering
phase shifts (respectively) as well as Eqs.~\ref{e:rhofit} and~\ref{e:i2fit}, which describe fits to those scattering phase shifts.
Our methodology is described in Sec.~\ref{s:meth}, which provides details of the
gauge field ensemble, the stochastic LapH method discussed above,
our procedure for extracting finite-volume energies from temporal correlation functions, and the L\"{u}scher method for obtaining
scattering phase shifts from those energies. Finally, results are described in
Sec.~\ref{s:res} while conclusions and a comparison with previous work are in
Sec.~\ref{s:concl}. Additional details concerning the determination of
finite-volume energies are relegated to an appendix.
\section{Methodology}\label{s:meth}
Here we detail technical aspects of the methods used in this work.
For this exploratory
large-volume calculation, an anisotropic lattice regularization is employed
to achieve a large spatial volume and a good temporal resolution at moderate
computational cost. On this anisotropic lattice the ratio of the spatial and temporal lattice spacings (the renormalized anisotropy) appears in the pion dispersion relation
and must be determined precisely.
The required temporal correlation matrices are measured on these gauge configurations using the stochastic
LapH method, while ground and low-lying excited-state energies are extracted from
them using solutions of a generalized eigenvalue problem (GEVP). Finally,
these energies are used in L\"{u}scher formulae to obtain elastic
scattering phase shifts.
\subsection{Ensemble details}\label{s:ens}
In order to suppress unwanted (exponential) finite-volume effects in lattice
QCD simulations with light pions, large spatial volumes are required. These
large volumes also increase the density of states in two-hadron channels,
improving the energy resolution of scattering phase shifts.
A large \emph{temporal} extent is additionally required to suppress thermal effects in
correlation functions with periodic temporal boundary conditions. Finally, a
good temporal resolution is needed to accurately extract
finite-volume energies from the fall-off of temporal correlation functions.
In order to satisfy these requirements with a moderate computational cost,
we employ an anisotropic lattice regularization in which the spatial and
temporal lattice spacings differ.
Our ensemble of gauge configurations is covered in detail in
Ref.~\cite{Lin:2008pr} and reviewed here briefly. Basic details are listed in
Tab.~\ref{t:ens}, where the temporal lattice spacing ($a_t$) is
determined
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|}
\hline
$(L/a_{s})^3\times(T/a_{t})$ & $a_tm_{K}$ & $a_tm_{\pi}$ & $m_{\pi}$ $(\mathrm{MeV})$ & $a_t (\mathrm{fm})$ & $m_{\pi}L$ & $N_{\mathrm{cfg}}$ \\
\hline
$32^3 \times 256$ & $0.08354(15)$ & $0.03938(19)$ & $233.0(1.2)$ & $0.033357(59)$ & $4.3$ & $412$ \\
\hline
\end{tabular}
\caption{\label{t:ens}Ensemble details for our $\nf = 2+1$ dynamical gauge
configurations. More details on the ensemble generation
can be found in Ref.~\cite{Lin:2008pr}, while the scale is determined using
$m_{K}$ as discussed in the text.}
\end{table}
by setting the mass of the kaon to $m_{K,\mathrm{phys}} = 494.2\mathrm{MeV}$.
This physical value was obtained in Ref.~\cite{Colangelo:2010et} by taking
the
isospin-symmetric limit and removing QED effects. We prefer scale setting
with $m_K$ to the method of
Ref.~\cite{Lin:2008pr}, which uses the mass of the Omega baryon ($m_{\Omega}$),
due to difficulties in determining $m_{\Omega}$. Still, this scale
should be viewed as indicative as the kaon mass was not extrapolated to
the physical light quark masses but taken on this single ensemble only.
However, our results for dimensionful quantities are naturally
expressed in terms of $m_{\pi}$ so that the lattice spacing enters only in
comparison with the literature in Fig.~\ref{f:rho_sum}. The determination of
$a_tm_{\pi}$, $a_tm_{K}$, and the renormalized anisotropy $\xi$ will be
discussed shortly.
Although these 412 configurations are separated by $20$
Hybrid Monte Carlo (HMC) molecular
dynamics trajectories of length $\tau = 1.0$, there is a small amount of
residual autocorrelation evident in the measured correlation
functions.\footnote{In lattice QCD the largest autocorrelations are typically
observed for `smoothed' observables such as the topological charge and smoothed action~\cite{Schaefer:2010hu}, which are not examined here.}
In
order to mitigate effects of this autocorrelation on estimates of
statistical uncertainties, we average measurements on pairs of subsequent
configurations.
Statistical errors are estimated using the bootstrap
technique~\cite{efron1986} on this rebinned ensemble with
$N_{B} = 800$ bootstrap samples.
In this anisotropic setup the lattice regulator is fully
specified by the temporal lattice spacing $a_t$ and renormalized anisotropy
$\xi = a_{s}/a_{t}$. The anisotropy is determined from the (continuum) pion
dispersion relation
\begin{align}\label{e:disp}
\left[a_{t}E_{\pi}(\boldsymbol{d}^2)\right]^2 = (a_tm_{\pi})^2 + \left(\frac{2\pi a_s}{\xi L}\right)^2 \boldsymbol{d}^2,
\end{align}
where $\boldsymbol{d} \in \mathbb{Z}^3$ is the
quantized finite-volume three-momentum of the pion.
Determination of $\xi$ requires the single-pion energies $a_{t}E_{\pi}(\boldsymbol{d}^2)$ in
Eq.~\ref{e:disp}. Periodic temporal boundary conditions are used for this ensemble, potentially
complicating the extraction of finite-volume energies from the
fall-off of temporal correlation functions.
In particular, a single zero-momentum
pion correlation function has the `cosh' form in the limit
of ground-state saturation
\begin{align}\label{e:teff1}
\lim_{ {t \gg 1/E_1,} \atop {T-t\gg 1/E_1} } C_{\pi}(t) = A\mathrm{e}^{-m_{\pi}t}\left(1 + \mathrm{e}^{-m_{\pi}(T-2t)}\right),
\end{align}
where $E_1$ is the relevant first excited-state energy.
Two-pion correlation functions with zero total momentum have the more
complicated form (ignoring small energy shifts due to pion interactions)
\begin{align}\label{e:teff2}
\lim_{ {t \gg 1/E_1,} \atop {T-t\gg 1/E_1} } C_{2\pi}(t) = A\mathrm{e}^{-2m_{\pi}t}\left(1 +
\mathrm{e}^{-2m_{\pi}(T-2t)} + B\mathrm{e}^{-m_{\pi}(T-2t)}\right),
\end{align}
while two-pion correlation functions with non-zero total momenta
have a similar but more complicated additional exponential term.
Since our finite-volume energies are extracted from fits of temporal
correlation functions to an exponential form, these additional terms
add potentially significant complication as has been discussed in (e.g.)
Ref.~\cite{Dudek:2012gj}. Fortunately, the large temporal extent of our
lattice ($m_{\pi}T \approx 10$) suppresses such terms below the statistical accuracy
of the energy levels. This can be demonstrated by performing two-parameter
correlated-$\chi^2$
fits of the single zero-momentum pion correlation function to both a single
exponential and the cosh of Eq.~\ref{e:teff1}. The
second exponential in Eq.~\ref{e:teff1} is larger than or equal to the
additional problematic exponential terms which appear in two-pion correlation
functions, apart from small hadronic interaction effects.
As the single pion at rest is our most precisely determined
correlation function, it is most sensitive to these thermal effects.
The absence of these effects, such as the second exponential in
Eq.~\ref{e:teff1}, indicates that additional exponentials in
two-pion correlation functions may be neglected. The comparison of
single-exponential and cosh fits for various fitting ranges is shown in
Fig.~\ref{f:cexp}. Clearly, no effect from the finite temporal extent is
evident for these temporal separations. All subsequent correlated-$\chi^2$ fits to temporal correlation
functions will thus ignore finite-$T$ effects. The extraction of
these energies will be discussed in more detail in Sec.~\ref{s:ener}.
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{tmin_kaon_err.pdf}
\includegraphics[width=0.49\textwidth]{pion_comp.pdf}
\caption{\label{f:cexp} \textbf{Left}: $\tmin$-plot (defined in the text)
for two-parameter single-exponential correlated-$\chi^2$ fits used in
the determination of $a_tm_{K}$ from which $a_t$ is set. The solid and dashed lines
show the mean value and $1\sigma$ errors (respectively) for the chosen fit
range, which is also indicated by a black square.
\textbf{Right}: Comparison of $t_{\mathrm{min}}$-plots for
single-exponential fits (denoted `exp') and
cosh fits according to Eq.~\ref{e:teff1}
for the single-pion correlation function with zero total momentum. Fit ranges
are shown with $t_{\mathrm{max}}=38a_t$ and varying $t_{\mathrm{min}}$. The
consistency of these two fit forms for our most precisely determined
correlation function demonstrates that `thermal' effects due to
the finite temporal extent may be neglected. }
\end{figure}
The fits to the single-pion zero-momentum correlation function shown in
Fig.~\ref{f:cexp}, as well as all other fits to correlated data in this
work, minimize a correlated-$\chi^2$ to properly treat the covariance between
observables measured on the same ensemble of gauge configurations. The covariance matrix is obtained using the bootstrap estimator
\begin{align}\label{e:cov}
\mathrm{Cov}(t,t') &= \frac{1}{N_B - 1} \sum_{n=1}^{N_B} (\langle C(t) \rangle_n - \langle\langle C(t) \rangle\rangle) (\langle C(t') \rangle_n - \langle\langle C(t') \rangle\rangle),
\end{align}
where $N_B = 800$ is the number of bootstrap samples,
$\langle C(t) \rangle_n$ is the bootstrap replicum of $C(t)$ on the $n$th
sample, and $\langle\langle C(t) \rangle\rangle$ is the average over all
bootstrap replica.
The covariance matrix is taken as identical across each bootstrap sample's
determination of the fit parameters.
Apart from effects due to the finite temporal extent, the
range of timeslices $[\tmin,\tmax ]$ over which the fit is performed is
another source of systematic error. In particular, fitted values exhibit a
marked sensitivity to $\tmin$ due to the influence of
higher-lying exponentials in Eq.~\ref{e:cor}. In this work we employ
`$\tmin$-plots' to ensure that
this systematic error is smaller than the statistical error on the fit
parameters. The guidelines for selecting a $\tmin$ satisfying this criterion
are given in Eq.~\ref{e:tmin}. These plots show the fitted values for many $\tmin$
with a fixed $\tmax$, and are exemplified in Fig.~\ref{f:cexp} which shows
$\tmin$-plots for $a_tm_{K}$ and $a_tm_{\pi}$.
With stochastically-estimated correlation functions, $\tmin$-plots are
preferable to effective masses $m_{\mathrm{eff}}(t) = \ln [C(t)/C(t+1)]$ for
determining the range of times over which a single exponential
dominates. Stochastically-estimated effective masses
typically have larger error than the corresponding fitted
energies and are thus not useful to assess systematic errors from the choice
of $\tmin$.
We now discuss two determinations of $\xi$. In the first, single-pion
energies at various total momenta\footnote{Pion correlation functions are
averaged over
equivalent momenta before fitting.} are first obtained from this fitting
procedure and then used in Eq.~$\ref{e:disp}$.
The $t_{\mathrm{min}}$-plots for single-exponential
fits to these correlation functions together with the fitted energies used in
our analysis are given in Fig.~\ref{f:pi_tmins} of App.~\ref{a:pi}.
Generally, these energies are chosen somewhat conservatively so that
systematic errors due to excited-state contamination are small in comparison to the statistical errors.
They are summarized in Fig.~\ref{f:fart} together with a fit to Eq.~\ref{e:disp} for $\boldsymbol{d}^2 \le 6$.
Correlation exists among the fitted energies;
their covariance is estimated using the bootstrap method of Eq.~\ref{e:cov}
and fixed on each bootstrap sample.
Evidently
the continuum dispersion relation describes the single-pion energies
up to large total momenta, suggesting that lattice spacing effects are
under control here. The pion
mass and $\xi$ determined from this linear fit (denoted `Strategy 1') are
given in Tab.~\ref{t:xi}.
\begin{figure}
\includegraphics[width=0.49\textwidth]{disp_rel.pdf}
\includegraphics[width=0.49\textwidth]{tmin_xi_err_final.pdf}
\caption{\label{f:fart}Two strategies to determine $\xi$.
\textbf{Left}: (Strategy 1) Single pion energies at various
momenta together with a linear fit to Eq.~\ref{e:disp}.
\textbf{Right}: (Strategy 2) $t_{\mathrm{min}}$-plot for
a simultaneous fit to all pion correlation functions to
Eq.~\ref{e:xi2} together with the chosen fit range indicated by the solid
and dotted lines.}
\end{figure}
An alternative determination (denoted `Strategy 2') fits all single-pion
correlation functions simultaneously to the ansatz
\begin{align}\label{e:xi2}
C_{\boldsymbol{d}^2}(t) = A_{\boldsymbol{d}^2} \times \mathrm{e}^{-\frac{t}{a_{t}}\sqrt{(a_tm_{\pi})^2 + \left(\frac{2\pi a_s}{\xi L}\right)^2 \boldsymbol{d}^2}}
\end{align}
where the $\{A_{\boldsymbol{d}^2}\}$, $m_{\pi}$ and $\xi$ are free
parameters. The covariance between all correlation functions at all time
separations is explicitly taken into account in these correlated-$\chi^2$ fits.
The results for $\xi$ from this fit are shown in Fig.~\ref{f:fart} for
various $t_{\mathrm{min}}$ (identical for all correlation functions) together with the chosen fit range. This fit is
also given in Tab.~\ref{t:xi}, where it is denoted `Strategy 2'.
Although the continuum dispersion relation fits the data well, we additionally
perform fits like Strategy 2 but using the lattice-modified dispersion relation
\begin{align}\label{e:lat}
(a_tE_{\pi})^2 = (a_tm_\pi)^2 + 4\sum_{i=1}^{3} \sin^2 \left(\frac{\pi a_s}{\xi L} d_i\right).
\end{align}
The results of this fit are also consistent and shown in Tab.~\ref{t:xi} as
`Strategy 3'.
For this
work we take $\xi$ from Strategy 1 as it is the most conservative estimate,
although the final results have little dependence on this choice.
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Strategy & $a_{t}m_{\pi}$ & $\xi$ & $\chi^2/d.o.f$ \\
\hline
1 & 0.03938(19) & 3.451(11) & 1.4 \\
2 & 0.03978(19) & 3.4654(98) & 1.19 \\
3 & 0.03978(19) & 3.4649(98) & 1.20 \\
\hline
\end{tabular}
\caption{\label{t:xi}Results for $\xi$ from linear fits of moving pion
energies to Eq.~\ref{e:disp} (Strategy 1), a simultaneous fit
of all pion correlation functions to Eq.~\ref{e:xi2} (Strategy 2), and a
simultaneous fit to Eq.~\ref{e:lat} (Strategy 3).}
\end{table}
\subsection{Correlation function calculation}\label{s:corr}
Because of the finite spatial extent and lattice spacing, the
symmetry group of lattice QCD is $O^{D}_h$, the double
cubic point group. Irreducible representations of this group (or the relevant
little group for a particular momentum) together with
total isospin and $G$-parity fully specify the quantum numbers of our energy
eigenstates. Therefore, operators which transform irreducibly under these symmetries are employed.
The procedure for constructing such operators is well
known. Here we are concerned only with the ground state and $2-3$
low-lying excited states in the relevant irreducible representations
(irreps).
However, this work is part of a broader program to explore many higher-lying
resonances in QCD. Interpolating operators with large overlap onto these
higher-lying resonances are more complicated and require non-trivial
spatial structures as in Refs.~\cite{Basak:2005aq,Morningstar:2013bda}. Such operators are not used here, but rather only
(smeared) single-site interpolators for each hadron.
Correlation matrices are typically required to obtain excited-state energies.
In order to build correlation matrices in each of the irreps, we
examine the expected non-interacting single-$\rho$ and two-pion levels.
Generally, an interpolator for each of these levels below the
inelastic threshold $E_{{\mathrm {cm}}}/m_{\pi}=4$ is included while additional
two-pion operators are used as a check of systematic effects.
As discussed above, these multi-hadron correlation matrices
require all-to-all quark propagators. We use the method of
Ref.~\cite{Morningstar:2011ka} which
introduces noise in the subspace spanned by low-lying eigenmodes of the
gauge-covariant Laplace operator. This noise can be diluted~\cite{Foley:2005ac} in time (T), spin (S), and
Laplacian eigenvector (L) indices, each of
which can be fully diluted (`F') or have some number of dilution projectors
`interlaced' (`I$n$') uniformly throughout the space. Note that the distillation method
of Ref.~\cite{Peardon:2009gh} is recovered in the maximal dilution limit (TF,
SF, LF).
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
$N_{v}$ & line type & $N_{r}$ & scheme & $N_{t_0}$ & $N_{D}$ \\
\hline
264 & fixed & 5 & (TF, SF, LI8) & 8 & 1280 \\
& relative & 2 & (TI16, SF, LI8) & - & 1024 \\
\hline
\end{tabular}
\caption{\label{t:dil} The number of eigenvectors ($N_v$), noise sources ($N_{r}$),
source times ($N_{t_0}$), and Dirac
matrix inversions per configuration ($N_D$), together with the
dilution schemes for fixed and relative quark lines.}
\end{table}
On this anisotropic ensemble it is beneficial to choose different dilution schemes for quark
propagators between different times (so-called `fixed'
quark lines) and for quark propagators starting and ending at the same time
(`relative' quark lines).
These different dilution schemes are specified in
Tab.~\ref{t:dil} together with the number of required Dirac matrix inversions ($N_D$)
and low-lying Laplacian eigenvectors defining the LapH subspace ($N_v$).
In order to ensure an unbiased estimate of correlation functions, each quark line requires an independent stochastic source. The total number of such sources
used per configuration ($N_r$) is shown in Tab.~\ref{t:dil} together with the
number of source times ($N_{t_0}$) used to reduce statistical errors.
It should be noted that only $N_{r} = 4$ fixed lines and
$N_r=1$ relative lines (for a minimum $N_D=640$)
are required to ensure unbiased estimates of the required correlation functions.
However, additional source times and
noise sources are employed here to increase statistics. While the additional
noise sources are required for
other systems, different noise combinations provide additional stochastic
estimates and are thus averaged over. The required Wick contractions are
enumerated in Ref.~\cite{Morningstar:2011ka}.
\subsection{Finite-volume energies}\label{s:ener}
After constructing the correlation functions as described in Sec.~\ref{s:corr}, the
method for extracting finite-volume energies from them is now
discussed. For this work we aim to utilize not only the ground state in each
irreducible representation, but several excited states as well. In order to
reliably extract these excited-state energies, solutions of a generalized
eigenvalue problem are employed.
In each channel, a correlation matrix is formed
consisting of a single-site $\rho$ interpolating operator
(if present)
together with the relevant two-pion operators. These two-pion operators are
chosen to match the expected non-interacting states and all such operators
below inelastic threshold are included.
For each of these correlation matrices ($C(t)$) we solve the generalized eigenvalue
problem
\begin{align}\label{e:gevp}
C(t_d)v(t_0,t_d) = \lambda(t_0,t_d)C(t_0)v(t_0,t_d)
\end{align}
for a particular set of $(t_0,t_d)$. The eigenvectors $\{v_{n}(t_0,t_d)\}$ are
used to define correlation functions between
`optimal' interpolators~\cite{Michael:1982gb}
\begin{align}\label{e:rcor}
\hat{C}_{ij}(t) = \left( v_{i}(t_0,t_d), C(t)v_j(t_0,t_d)\right)
\end{align}
where the outer parentheses denote an inner product over GEVP indices. Although
these optimal interpolators are constructed to have maximal overlap with a
single Hamiltonian eigenstate, the off-diagonal elements of
$\hat{C}_{ij}(t)$ are not exactly zero resulting in a
source of systematic error that
must be assessed. It should be noted that this is a different approach
to Refs.~\cite{Luscher:1990ck,Blossier:2009kd}
which require the solution of the GEVP at different $(t_0,t_d)$, possibly
introducing ambiguities between closely spaced levels at different times,
but guaranteeing that the eigenvalues approach the desired exponential fall-off.
To extract energies in a particular channel we solve the GEVP of
Eq.~\ref{e:gevp} and form the rotated correlation matrix of
Eq.~\ref{e:rcor}. The GEVP diagonalization is not performed on each bootstrap sample, due to similar ambiguities identifying closely spaced levels on
different bootstrap samples.
We first perform
two-parameter correlated-$\chi^2$ fits with a
single-exponential ansatz on the diagonal elements of the rotated
correlation matrix to obtain a preliminary determination of the finite-volume
spectra. These preliminary energies are used in Fig.~\ref{f:i1box} and with
Eq.~\ref{e:ov} to obtain a qualitative picture of the spectrum and nature
of the states.
For our final analysis we employ
a different approach which exploits the similarity (and correlation) between
two-pion and single-pion correlation functions. As in
Ref.~\cite{Helmes:2015gla} but here generalized to arbitrary momenta, for an
optimized two-pion operator with pion momenta $\boldsymbol{d}_1$ and $\boldsymbol{d}_2$ ($\mathcal{O}_{\boldsymbol{d}_1,\boldsymbol{d}_2}$), we define the ratio
\begin{align}\label{e:rat}
R(t) = \frac{\langle \mathcal{O}_{\boldsymbol{d}_1,\boldsymbol{d}_2} (t)
\bar{\mathcal{O}}_{\boldsymbol{d}_1,\boldsymbol{d}_2}(0) \rangle}{
\langle \mathcal{O}_{\boldsymbol{d}_1}(t)
\bar{\mathcal{O}}_{\boldsymbol{d}_1}(0)\rangle
\langle \mathcal{O}_{\boldsymbol{d}_2}(t)
\bar{\mathcal{O}}_{\boldsymbol{d}_2}(0)\rangle}
\end{align}
which is constructed on each bootstrap sample and fit in a fully correlated manner to the ansatz
$R(t) = A\mathrm{e}^{-\Delta Et}$. The energy shift $\Delta E$ is used to
reconstruct the desired energy via
\begin{align}
a_tE = a_t\Delta E + \sqrt{(a_tm_{\pi})^2 + \left(\frac{2\pi a_s}{\xi L}\right)^2\boldsymbol{d}_1^2} + \sqrt{(a_tm_{\pi})^2 + \left(\frac{2\pi a_s}{\xi L}\right)^2\boldsymbol{d}_2^2},
\end{align}
where $m_{\pi}$ is obtained from the single-pion fits.
In the $I=1$ channel, these two-hadron states mix with the $\rho$-meson. For
such mixed states these ratio fits are still beneficial, but exhibit an
increased amount of excited-state contamination, which will be discussed
shortly.
Several sources of systematic error in this procedure must be addressed. First,
the fitting range $\left[\tmin, \tmax\right]$ is varied, in particular $\tmin$. Second, systematic errors due to the small but
non-zero off-diagonal elements of $\hat{C}_{ij}(t)$ must be assessed.
\begin{figure}
\includegraphics[width=\textwidth]{final_diagcomp.pdf}
\caption{\label{f:exp_tmin}Variation of $\tmin$, $(t_0,t_d)$, and the
number of operators included in the GEVP for three representative energy
levels in $I=1$. Each row corresponds to three different GEVP's for a single
energy level. We have chosen a representative sample of three energy levels
consisting of a ground state below the resonance region (top), a first excited
state near the resonance energy (middle) and a first excited state
somewhat above the resonance energy (bottom). The dimensionless
center-of-mass momentum $u^2$ (defined in Eq.~\ref{e:kin}) is shown, as it determines the scattering phase shift.}
\end{figure}
To this end, we not only vary the fitting range $\left[\tmin, \tmax\right]$
but also $(t_0,t_d)$ and the operators included in the GEVP. The variation
of these systematics for a selection of energy levels is shown in
Fig.~\ref{f:exp_tmin}. There the dimensionless center-of-mass momentum $u^2$
is shown, which is defined in Eq.~\ref{e:kin}.
Generally, systematic effects due to $\tmin$ are the largest and must be
treated with care. To this end we fix $(t_0,t_d) = (12a_t,24a_t)$ and choose
$\tmin$ conservatively.
As minimum requirements we demand that the chosen $\tmin$ gives a suitable
correlated $\chi^2/d.o.f. < 1.7$ and that
\begin{align}\label{e:tmin}
\Delta E_{\mathrm{fit}}(\tmin) - \Delta E_{\mathrm{fit}}(\tmin - \delta_t) < \sigma(\tmin)
\end{align}
where $\sigma(\tmin)$ is the bootstrap error on $\Delta E_{\mathrm{fit}}(\tmin)$ and
$\delta_t = 4a_t$.
While these ratio fits have the advantage of directly determining the energy
shifts, their excited-state contamination may have a non-standard form. This
can be seen by examining the leading excited-state corrections for the
ratio directly
\begin{align}
\lim_{t\rightarrow \infty} R(t) = A\mathrm{e}^{-\Delta E t} \left[
1 + B_{\boldsymbol{d}_1,\boldsymbol{d}_2}
\mathrm{e}^{-\Delta E_{\boldsymbol{d}_1,\boldsymbol{d}_2} t} -
B_{\boldsymbol{d}_1}\mathrm{e}^{-\Delta E_{\boldsymbol{d}_1} t}
- B_{\boldsymbol{d}_2}\mathrm{e}^{-\Delta E_{\boldsymbol{d}_2} t} \right],
\end{align}
where $\Delta E_{\boldsymbol{d}_1,\boldsymbol{d}_2}$ is the energy gap from
the two-pion correlator in the numerator
and $B_{\boldsymbol{d}_1,\boldsymbol{d}_2}$ the relevant interpolator-dependent
prefactor, while $B_{\boldsymbol{d}_{1,2}}$ and
$\Delta E_{\boldsymbol{d}_{1,2}}$ are the analogous
quantities for each of the single-pion correlators in the denominator.
If the first two excited states in the numerator effectively consist of
one pion in the ground state and the other in an excited state, the
overall excited-state contamination in $R(t)$ will be very small.
However, in general the excited-state contamination from the denominator
enters with different sign,
possibly causing a non-monotonically decreasing `bump'-type behavior
in $\tmin$-plots. Such
bumps must be taken into account when choosing fit ranges for the
strongly-interacting $I=1$ states.
Apart from Fig.~\ref{f:exp_tmin}, $\tmin$-plots for ratio fits performed to
all correlation functions used in the phase shift analysis are shown in
App.~\ref{s:app1} and App.~\ref{s:app2}. Although bumps are evident for
some levels, we choose conservative fit ranges in those cases to
ensure systematic effects from excited states are smaller than the
statistical error.
\subsection{Scattering phase shifts}\label{s:phase}
After discussing the procedure for extracting finite-volume energies, we
now turn to using them to calculate elastic scattering phase shifts.
The relation between finite-volume energy spectra and infinite-volume
elastic scattering amplitudes is derived in Ref.~\cite{Luscher:1990ux} and generalized to
non-zero total momentum in Ref.~\cite{Rummukainen:1995vs}. A useful summary of
the method for several different situations may be found in
Ref.~\cite{Gockeler:2012yj}, while generalizations to asymmetric spatial
volumes~\cite{Feng:2004ua}, multiple coupled two-particle
channels~\cite{He:2005ey} and three-particle
scattering~\cite{Polejaeva:2012ut,Hansen:2014eka,Hansen:2015zga} have been developed.
For scattering between two identical particles of mass $m$, we denote by $E$
the energy measured in the lattice (`lab') frame in a particular
irrep with total momentum $\boldsymbol{P} = \frac{2\pi\boldsymbol{d}}{L}$. We
define the kinematical variables
\begin{align}\label{e:kin}
E_{{\mathrm {cm}}} &= \sqrt{E^2 - \boldsymbol{P}^2}, \qquad \gamma=\frac{E}{E_{{\mathrm {cm}}}},
\qquad \boldsymbol{q}_{{\mathrm {cm}}}^2 = \frac{1}{4}E_{{\mathrm {cm}}}^2 - m^2, \qquad
u^2 = \frac{L^2\boldsymbol{q}^2_{{\mathrm {cm}}}}{(2\pi)^2}.
\end{align}
Up to exponentially suppressed finite-volume effects, the elastic scattering
matrix is related to the finite-volume energy spectra via the well-known
quantization condition, which is a matrix equation of the form
\begin{align}
\mathrm{det} \left[ 1 + F^{(\boldsymbol{d}, \gamma,u)} (S-1)\right] = 0
\end{align}
where $S$ is the infinite-volume scattering matrix and the determinant is taken over the indices $(J,m_{J},L,\sigma)$
corresponding to total angular momentum, its projection along some axis,
orbital angular momentum, and spin, respectively.
Note that the matrix $F$ in general mixes different partial waves.
For elastic
scattering between identical spin-zero particles $\sigma=0$ and $J=L$. In this case
the matrix $F$ is given by
\begin{align}
F^{(\boldsymbol{d},\gamma,u)}_{L'm_{L'};Lm_L} &= \frac{1}{2}\left(
\delta_{L'L}\delta_{m_{L'}m_L} + W_{L'm_{L'};Lm_L}\right), \\
W_{L'm_{L'};Lm_L} &= \frac{2i}{\pi\gamma u^{\ell+1}}Z_{\ell m}(\boldsymbol{d},
\gamma, u^2) \int d^2\Omega Y^{*}_{L'm_{L'}}(\Omega) Y_{\ell m}^{*}(\Omega)
Y_{Lm_L}(\Omega),
\end{align}
where $\ell$ and $m$ are summed over and we have introduced the L\"{u}scher
zeta functions $Z_{\ell m}(\boldsymbol{d}, \gamma, u^2)$. We use a
representation of the zeta functions given in App. A of
Ref.~\cite{Gockeler:2012yj} for their numerical evaluation, which is
consistent with an independent implementation based on an alternative representation discussed in
Ref.~\cite{Fahy:2014jxa}.
While we have expressed $F$ in the $Lm$ basis, it is more convenient to
express the relation in terms of finite-volume irreps, as both $F$ and $S$
become block diagonal, facilitating the evaluation of the determinant.
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|}
\hline
$\ell$ & $\boldsymbol{d}_{\mathrm{ref}}$ & irrep & $q_{{\mathrm {cm}}}^{2\ell + 1}\cot \delta_{\ell}$ \\
\hline
0 & $(0,0,0)$ & $A_{1g}$ & $w^{(0)}_{00}$ \\
\hline
& $(0,0,n)$ & $A_{1}$ & $w^{(0)}_{00}$ \\
\hline
& $(0,n,n)$ & $A_{1}$ & $w^{(0)}_{00}$ \\
\hline
& $(n,n,n)$ & $A_{1}$ & $w^{(0)}_{00}$ \\
\hline
\hline
1 & $(0,0,0)$ & $T_{1u}$ & $w^{(1)}_{00}$ \\
\hline
& $(0,0,n)$ & $A_{1}$ & $w^{(1)}_{00} + \frac{2}{\sqrt{5}}w^{(1)}_{20}$ \\
& & $E$ & $w^{(1)}_{00} - \frac{1}{\sqrt{5}}w^{(1)}_{20}$ \\
\hline
& $(0,n,n)$ & $A_{1}$ & $w^{(1)}_{00} + \frac{1}{2\sqrt{5}}w^{(1)}_{20} -
i\sqrt{\frac{6}{5}}w^{(1)}_{21} - \sqrt{\frac{3}{10}}w^{(1)}_{22}$ \\
& & $B_{1}$ & $w^{(1)}_{00} - \frac{1}{\sqrt{5}}w^{(1)}_{20} +
\sqrt{\frac{6}{5}}w^{(1)}_{22}$ \\
& & $B_2$ & $w^{(1)}_{00} + \frac{1}{2\sqrt{5}}w^{(1)}_{20} +
i\sqrt{\frac{6}{5}}w^{(1)}_{21} -\sqrt{\frac{3}{10}}w^{(1)}_{22}$ \\
\hline
& $(n,n,n)$ & $A_{1}$ & $w^{(1)}_{00} + 2i\sqrt{\frac{6}{5}}w^{(1)}_{22}$ \\
& & $E$ & $w^{(1)}_{00} - i\sqrt{\frac{6}{5}}w^{(1)}_{22}$ \\
\hline
\end{tabular}
\caption{\label{t:phase} Expressions for the scattering phase shifts in each
irreducible representation for both the $\ell=0$ and $\ell=1$ partial waves in terms of the quantities defined in Eq.~\ref{e:w}.}
\end{table}
After performing this block diagonalization and neglecting the contribution of higher partial waves, the relationship between the scattering phase shifts and
\begin{align}\label{e:w}
w^{(\ell)}_{lm} = \left(\frac{2\pi}{L}\right)^{2\ell+1} \frac{u^{2\ell-l}}{\gamma\pi^{3/2}} Z_{lm}(\boldsymbol{d},\gamma, u^2)
\end{align}
for each irrep is shown in Tab.~\ref{t:phase}.
One advantage of employing expressions relating the real part of the inverse
scattering amplitude to the
$w_{l m}^{(\ell)}$ is that the analyticity of
$q_{{\mathrm {cm}}}^{2\ell+1}\cot \delta_{\ell}$ near threshold is explicit. For weakly
interacting
channels such as the $I=2$ $A_{1g}^{+}$, this enables a smooth
behavior between positive and negative $q_{{\mathrm {cm}}}^2$.
As we treat identical-particle scattering,
particle-exchange symmetry prevents mixing between successive partial waves
in moving frames.
Neglecting the remaining partial wave mixing amounts to neglecting
the $I=1$, $\ell = 3$
and $I=2$, $\ell =2$ partial waves.
\section{Results}\label{s:res}
This section contains our results for elastic scattering phase shifts. We
neglect exponential finite-volume corrections and, as discussed in
Sec.~\ref{s:phase}, treat only the lowest partial wave which contributes to
each lattice irrep. Our results are interpreted in
terms of the effective range expansion, which provides the correct
threshold behavior of the scattering amplitude while also accommodating
resonances. Finite-volume energy levels near or above the inelastic threshold
$E_{{\mathrm {cm}}}/m_{\pi}=4$ are
not described by the elastic L\"{u}scher formulae of Sec.~\ref{s:phase} and
thus not used.
\subsection{$I=1$}\label{s:res1}
The $I=1$, $\ell=1$ partial wave contains the $\rho$-resonance. Not only is
this evident in the scattering phase shifts, but it is also
suggested by examining the overlaps of interpolating operators onto finite-volume Hamiltonian eigenstates. Specifically, we estimate
$Z_{in} = |\langle 0 | \hat{\mathcal{O}}_i | n \rangle|^2$ by forming the ratio
\begin{align}\label{e:ov}
Z_{in}(t) = \left|\frac{\sum_{j} C_{ij}(t)v_{nj}(t_0,t_d)}{\mathrm{e}^{-\frac{E_n}{2}t}\sqrt{\hat{C}_{nn}(t)}}\right|^2,
\end{align}
where $E_n$ is the fitted energy,
and taking $t=20a_t$. For each interpolating operator the overlaps onto the
Hamiltonian eigenstates are plotted in Fig.~\ref{f:i1box} together with the
\begin{figure}
\includegraphics[width=\textwidth]{isotriplet/box_plot_isotriplet.pdf}
\caption{\label{f:i1box}(color online) $I=1$ center-of-mass energies (upper panel) for
each irrep together with the overlaps of each interpolator.
Each column (across both the upper and lower panels) corresponds to a
single irrep and the colors are consistent between the energy levels and
overlap plots. States significantly below the resonance mass
(located at $E_{{\mathrm {cm}}}/m_{\pi}\approx 3.4$)
have significant overlap with two-pion operators only, while those near the
resonance region overlap with both two-pion and single-$\rho$ interpolators.}
\end{figure}
energies extracted from single-exponential fits. Center-of-mass energies
are shown in that figure to facilitate comparison between channels with
different total momenta.
As expected, local $\rho$-meson interpolating operators have
significant overlap with energy eigenstates near the resonance mass
$E_{{\mathrm {cm}}}/m_{\pi} \approx 3.4$, where mixing with two-pion operators can be
observed. However, only two-pion interpolating operators have significant
overlap with energy eigenstates outside this resonance region. For
states which have significant overlap onto two-pion interpolators only, the
ratio fits described previously have very little excited state contamination.
Clearly, there are a number of states near or above the four-pion threshold
$E_{{\mathrm {cm}}}/m_{\pi} = 4$. While these states can be extracted with suitable
statistical precision, their interpretation in terms of infinite-volume
scattering amplitudes is unknown.
\begin{table}
\centering
\input{table_i1.tex}
\caption{\label{t:i1dat}Results for the center-of-mass energies and
scattering phase shifts in $I=1$. For each total momentum
($\boldsymbol{d}^2$), lattice irrep and energy level, the two single-pion
correlation functions used in the ratio fits are denoted by $(\boldsymbol{d}^2_1,\boldsymbol{d}^2_2)$. The minimum time included in the fit range, the
correlated-$\chi^2$, fitted energy shift, reconstructed center-of-mass energy,
and scattering phase shift are also given for each energy level.
}
\end{table}
Numerical results for our final analysis using ratio fits are listed in Tab.~\ref{t:i1dat},
where $(q_{{\mathrm {cm}}}/m_{\pi})^3\cot{\delta_{1}}$ is obtained by applying the formulae of
Tab.~\ref{t:phase}.
\begin{figure}
\includegraphics[width=0.49\textwidth]{isotriplet/p3cot_isotriplet.pdf}
\includegraphics[width=0.49\textwidth]{isotriplet/phase_isotriplet.pdf}
\includegraphics[width=0.49\textwidth]{isotriplet/argand_isotriplet.pdf}
\includegraphics[width=0.49\textwidth]{isotriplet/sinsq_isotriplet.pdf}
\caption{\label{f:i1scat} The real part of the inverse scattering
amplitude (top left), phase shift (top right), Argand plot (lower left), and partial wave amplitude (lower right)
for the $I=1$, $\ell =1$ partial wave. The dotted lines indicate a
fit to data in the upper left plot.
Points from the first excited state in the $\boldsymbol{d}^2=3$, $A_{1}^{+}$ and $\boldsymbol{d}^2=4$, $E^{+}$ channels are omitted from the
phase shift, Argand, and amplitude plots due to their large errors.}
\end{figure}
This particular quantity is the real part of the inverse scattering
amplitude and is thus analytic in the complex momentum plane
near the two-pion threshold $E_{{\mathrm {cm}}}/m_{\pi} = 2$, making it a natural choice
for fits of the amplitude's energy dependence.
For this resonant $\ell=1$ partial wave, $(q_{{\mathrm {cm}}}/m_{\pi})^{3}\cot \delta_{1}$ can be
described by the Breit-Wigner parametrization
\begin{align}\label{e:bw}
\left(\frac{q_{{\mathrm {cm}}}}{m_{\pi}}\right)^{3}\cot \delta_1 =
\left(\frac{m^2_{\rho}}{m^2_{\pi}} - \frac{E^2_{{\mathrm {cm}}}}{m_{\pi}^2}\right)
\frac{6\pi E_{{\mathrm {cm}}}}{g_{\rho\pi\pi}^2m_{\pi}}
\end{align}
which also has the correct threshold behavior dictated by the effective range expansion.
A two-parameter (fully-correlated) $\chi^2$-fit to Eq.~\ref{e:bw} is performed.
This fit must not only take into account the correlation between
different data points, but also the correlation between $E_{{\mathrm {cm}}}/m_{\pi}$
and $ \left(q_{{\mathrm {cm}}}/m_{\pi}\right)^{3}\cot \delta_1 $ for each data
point. In order to do this, we employ the correlated-$\chi^2$ which is
the maximum likelihood estimator for the distribution of the residuals $d_i$
\begin{align}
\chi^2 &= \sum_{i,j} d_i \, \mathrm{Cov}^{-1}(i,j) \,d_j,
\\\nonumber d_i(m_{\rho}/m_{\pi}, g_{\rho\pi\pi}) &= \left[\left(\frac{q_{{\mathrm {cm}}}}{m_{\pi}}\right)^{3}\cot \delta_1\right]_i - \left[\left(\frac{m^2_{\rho}}{m^2_{\pi}} - \frac{E^2_{{\mathrm {cm}}}}{m_{\pi}^2}\right)
\frac{6\pi E_{{\mathrm {cm}}}}{g_{\rho\pi\pi}^2m_{\pi}}\right]_i.
\end{align}
As with the other fits in this work, the bootstrap estimator is used to
estimate the covariance between the $\{d_i\}$. However, each $d_i$ depends
nontrivially on the fit parameters $m_{\rho}/m_{\pi}$ and $g_{\rho\pi\pi}$
so the bootstrap estimate of the covariance must be recalculated on each
call to the correlated-$\chi^2$ function. In other words, on each bootstrap sample, each call to the correlated-$\chi^2$ employs all bootstrap samples to
estimate the covariance. While this method may seem cumbersome, it ensures that
all correlations among the data are taken into account.
The results of this fit are
\begin{align}\label{e:rhofit}
\frac{m_{\rho}}{m_{\pi}} = 3.350(24), \qquad g_{\rho\pi\pi} = 5.99(26) , \qquad \chi^2/d.o.f. = 1.04.
\end{align}
While other resonance parametrizations have been applied
(in e.g. Ref.~\cite{Dudek:2012xn}) which maintain unitarity above the resonance
region, given the proximity of the four-pion threshold such parametrizations
seem poorly motivated here. However, we test the dependence of these resonance
parameters on the Breit-Wigner fit form by employing a non-relativistic
ansatz to $\left(q_{{\mathrm {cm}}}/m_{\pi}\right)^{3}\cot \delta_1$
\begin{align}
\tan \delta_1 = \frac{\Gamma/2}{m_{\rho} - E_{{\mathrm {cm}}}} + A,
\quad \Gamma = \frac{g_{\rho\pi\pi}^2}{48\pi m_{\rho}^2}(m_{\rho}^2 - 4m_{\pi}^2)^{3/2},
\end{align}
where $\Gamma$ is an energy-independent width and $A$ parametrizes a
slowly-varying background. This three-parameter fit gives
\begin{align}\label{e:rhofit2}
\frac{m_{\rho}}{m_{\pi}} = 3.352(23), \qquad g_{\rho\pi\pi} = 5.84(34) , \qquad A = -0.160(26) , \qquad \chi^2/d.o.f. = 1.03.
\end{align}
A
summary of our data as well as the fit of Eq.~\ref{e:rhofit} are shown in Fig.~\ref{f:i1scat}.
Several different representations of the data are shown in that figure. First,
the $ (q_{{\mathrm {cm}}}/m_{\pi})^{3}\cot \delta_1 $ data points are
shown with the corresponding fit to them. Then, $\delta_1$ is shown (in
$[0,\pi]$) with the fit to $(q_{{\mathrm {cm}}}/m_{\pi})^{3}\cot \delta_1 $. The rapid variation of the phase shift is clear in this plot. Further
evidence of this rapid variation is seen in an Argand plot showing the real
and imaginary parts of the partial wave amplitude (following the conventions
of Ref.~\cite{taylor2012scattering})
\begin{align}
q_{{\mathrm {cm}}} f_{1} = \mathrm{e}^{i\delta_1}\sin \delta_1.
\end{align}
Finally, a plot of the
partial wave cross section
\begin{align}
m_{\pi}^2 \sigma_1 = 12\pi m_{\pi}^2 \frac{\sin^2 \delta_1}{q_{{\mathrm {cm}}}^2}
\end{align}
shows a clear enhancement due to the resonance.
Due to the singular nature of the L\"{u}scher zeta functions at
non-interacting energies, the distribution of bootstrap samples of the
quantities shown in Fig.~\ref{f:i1scat} can show significant asymmetry. In that
figure we therefore display asymmetric $1\sigma$ bootstrap error bars.
Displaying the points in this manner indicates the level of asymmetry but
ignores the correlation between the horizontal and vertical error bars.
\subsection{$I=2$}
\begin{table}
\centering
\input{table_i2.tex}
\caption{\label{t:i2dat}The same as Tab.~\ref{t:i1dat} but $I=2$ data for the $\ell=0$ partial wave.}
\end{table}
The $I=2$ channel is weakly interacting and thus a good test
of the stochastic LapH method. As in the $I=1$ case, we examine the real part
of the inverse scattering amplitude, which is analytic near the
two-pion threshold. Our fitted energies and
resultant phase shifts for the $\ell =0$ partial wave
are shown in Tab.~\ref{t:i2dat}.
\begin{figure}
\includegraphics[width=0.49\textwidth]{isoquintet/pcot_final.pdf}
\includegraphics[width=0.49\textwidth]{isoquintet/phase_final.pdf}
\caption{\label{f:i2scat} The real part of the inverse scattering amplitude
(left) and the scattering phase shift (right) for the $I=2$, $\ell=0$
partial wave.}
\end{figure}
The weakly interacting nature of this channel motivates its description by the lowest few terms of the effective range expansion
\begin{align}\label{e:er}
\left(\frac{q_{{\mathrm {cm}}}}{m_{\pi}}\right)\cot \delta_0 = \frac{1}{m_{\pi}a_{0}} +
\frac{1}{2}(m_{\pi}r)\left( \frac{q_{{\mathrm {cm}}}}{m_{\pi}}\right)^2.
\end{align}
This parametrization is expected to be valid for momenta below the
$t$-channel cut $q_{{\mathrm {cm}}} \ll m_{\pi}$~\cite{Beane:2011sc}.
Our results are collected in Fig.~\ref{f:i2scat}.
Due to the smaller number of finite-volume irreps in which the $\ell =0$
partial wave appears, there are only four points in this low-momentum region.
A two-parameter fit to the effective range ansatz of Eq.~\ref{e:er} yields
\begin{align}\label{e:i2fit}
m_{\pi} a^{I=2}_{0} = -0.064(12) ,\qquad m_{\pi} r = 18.1(8.4), \qquad \chi^2/d.o.f. = 0.19.
\end{align}
The small number of points in this channel suggests that not much can be
gained by adding the next term in the effective range expansion which contains
the shape parameter. The scattering length is determined with about $20\%$
precision and is consistent with the (continuum) $\chi$PT extrapolation of
(e.g.) Ref.~\cite{Helmes:2015gla} but the pion mass used in this work is
lighter than those employed there.
|
\section{Introduction}
This paper deals with the notion of mixed Hodge structure associated to the intersection spaces of a complex projective variety $X$ of complex dimension $n$ with only isolated singularities and simply connected links.
Intersection spaces were defined by Markus Banagl in \cite{Banagl2010} as a way to spatialize Poincaré duality for singular spaces. Suppose given a compact, connected pseudomanifold of dimension $n$ with only isolated singularities and simply connected links. We assign to this space a family of topological spaces $\I{p}{X}$, its intersection spaces, where $\overline{p}$ is an element called a perversity varying in a poset $\Pos{n}$ called the poset of perversities. We then have for complementary perversities a generalized Poincaré duality isomorphism
\[
\redHI{p}{k}{X} \cong \D{\redhI{q}{n-k}{X}}.
\]
with $\D{\redhI{q}{n-k}{X}}=\hom(\redhI{q}{n-k}{X}, \mathbf{Q}) $
The theory of intersection spaces can be seen as an enrichment of intersection homology since they both gives complementary informations about $X$.
The aim of this paper is twofold. First we want to get a better understanding of the family of cohomology algebras $\lbrace \HI{\bullet}{\ast}{X} \rbrace_{\overline{p} \in \Pos{n}}$ when we take all the spaces into consideration. We then want to put a mixed Hodge structure on these algebras and get results about formality of intersections spaces.
Formality is a notion tied to the rational homotopy theory of topological spaces. The rational homotopy type of a topological space $X$ is given by the commutative differential graded algebra $\Apl{X}$ in the homotopy category $\mathsf{Ho}(\CDGA{Q})$ defined by formally inverting quasi-isomorphisms and where $\Apl{-} \col \Top \rightarrow \CDGA{Q}$ is the polynomial De Rham functor defined by Sullivan. The space $X$ is then formal if there is a string of quasi-isomorphisms from the cdga $\Apl{X}$ to its cohomology with rational coefficients $H^{\ast}(\Apl{X}) \cong H^{\ast}(X, \mathbf{Q})$ seen as a cdga with trivial differential. In particular is $X$ is formal then its rational homotopy type is a formal consequence of its cohomology ring, its higher order Massey products vanish.
The combination of rational homotopy theory and Hodge theory has already been showed to be fruitful. Using Hodge theory, Deligne, Griffiths, Morgan and Sullivan proved in \cite{Deligne1975} that compact Kähler manifolds, in particular smooth projective varieties, are formal. It was also shown by Simpson in \cite{Simpson2011} that every finitely presented group $G$ is the fundamental group of a singular projective variety $X$ and then Kapovich and Koll\'ar showed in \cite{Kapovich2014} that this $X$ could be chosen to be complex projective with only simple normal crossing singularities. More recently, Chataur and Cirici proved in \cite{Chataur2015} that every complex projective variety of dimension $n$ with only isolated singularities $\Sigma = \lbrace\sigma_{1}, \dots, \sigma_{\nu} \rbrace$ such that the link $L_{i}$ of each singularities $\sigma_{i}$ is $(n-2)$-connected is then a formal topological space.
The intersection spaces $\I{p}{X}$ of $X$ are not complex nor algebraic varieties, even if $X$ is. Thus at first glance there should be no reasons the cohomology of these spaces carry a mixed Hodge structure. On second thought, when $X$ is a complex projective variety of complex dimension $n$ with only isolated singularities and that we look at the rational cohomology of their intersection spaces
\[
\HI{p}{k}{X} =
\begin{cases}
\mathbf{Q} & k=0 \\
H^{k}(X) & 1 \leq k \leq p \\
H^{k}(X) \oplus \im{H^{k}(X_{reg}) \rightarrow H^{k}(L)} & k = p+1 \\
H^{k}(X_{reg}) & k > p+1
\end{cases}
\]
it becomes a bit more natural to think that there is a mixed Hodge structure since each part of their rational cohomology can be endowed with a natural mixed Hodge structure coming from $X$. We show here that in fact all these structures naturally come from a mixed Hodge structure at the algebraic models level and that this structure is compatible with the different operations defined on intersection spaces. Note that our definition of intersection spaces \ref{def:intersectionspace2} differs slightly from the original definition given in \cite{Banagl2010}.
It must be pointed out here that the question of a Hodge structure on the intersection spaces as already been looked at in the work of Banagl and Hunsicker \cite{Banagl2015} where they use $L^{2}$-cohomology to provide a Hodge theoretic structure. We do not follow this path here and rather modify the rational homotopy theory tools developed in \cite{Chataur} for the mixed Hodge structures in intersection cohomology.
We explain the contents of this paper.
The section \ref{section:background} is devoted to collect the different definitions needed. We recall what we call a perversity, the definition of the intersection spaces and the convention we use to construct them. We also introduce the notion of a coperverse cdga which is the main tool for the rational algebraic models of the intersection spaces. We then define a model category structure on the category of coperverse cdga's \ref{thm:model_structure}.
The section \ref{section:coperv_algebras} is a direct application of the previous section. We define the notion of a coperverse cdga associated to a morphism of cdga's. As a result we show that the whole family of algebraic model $\AI{\bullet}{X}$ computing the rational cohomology of intersection spaces carry a structure of coperverse algebra and that we have a external product on that family, extending the cup product that each $\I{p}{X}$ naturally has as a topological space.
The section \ref{section:Hodge_th} is the main section of this paper, we extend our notion of coperverse cdga to the notion of coperverse mixed Hodge cdga. These coperverse mixed Hodge algebras carry a mixed Hodge structure which is compatible the differential, product and poset maps of the underlying coperverse cdga. After developing their algebraic definitions we show in theorem \ref{thm:coperverse_MI} that given a complex projective variety $X$ of complex dimension $n$ with only isolated singularities and simply connected links, there is a coperverse mixed Hodge cdga $\MI{\bullet}{X}$ quasi-isomorphic to the coperverse cdga $\AI{\bullet}{X}$. As a result the whole family $\HI{\bullet}{\ast}{X}$ carry a well defined mixed Hodge structure defined at the algebraic models level.
The section \ref{section:Weight_SS} is devoted to the computation of the associated weight spectral sequence. If $X$ is a complex projective algebraic variety with only isolated singularities and such that $X$ admits a resolution of singularities where the exceptional divisor is smooth, we are able to compute the weight spectral sequence associated to the mixed Hodge structure. We then use this spectral sequence to show a result of "purity implies formality" in theorem \ref{thm:pure_is_formal}.
The section \ref{sec:formality_3folds} is completely devoted to the proof of the theorem \ref{thm:3fold_formal} : suppose $X$ to be a complex projective algebraic threefold with isolated singularities such that there exist a resolution of singularities with a smooth exceptional divisor, then if the links are simply connected the intersection spaces $\I{p}{X}$ are formal topological spaces for any perversity $\overline{p}$. The proof being rather long and intricate, we made the choice of giving it its own section. This result goes well with the result of \cite[Theorem E p.76]{Chataur2012} stating that any nodal hypersurface in $\mathbf{C}P^{4}$ is intersection-formal.
The last section \ref{section:Eg} deals with computations, with for instance the computations for the Calabi-Yau generic quintic 3-fold \ref{subsec:CY_generic_quintic} and the Calabi-Yau quintic 3-fold \ref{subsec:CY_quintic} where we are able to retrieve the cohomology of the associated smooth deformation as stated in \cite{Banagl2012}.
\section{Background, intersection spaces and coperverse algebras}
\label{section:background}
\subsection{Perversities and intersection spaces}
Unless stated otherwise, all cohomology groups will be considered with rational coefficients and they will be omitted.
Since we are concerned about complex algebraic varieties of complex dimension $n$ with only isolated singularities we use the following definition of a perversity. A perversity $\overline{p}$ is determined by a integer $0 \leq p \leq 2n-2$, we then denote by $\Pos{n}^{op}$ the poset $\lbrace 0, \dots, 2n-2 ; \leq \rbrace$ with the reverse order and $\widehat{\Pos{n}}^{op} := \Pos{n}^{op} \cup \lbrace \infty \rbrace$. The posets $\Pos{n}^{op}$ and $\widehat{\Pos{n}}^{op}$ are then totally ordered and look like
\[
2n-2 \rightarrow 2n-3 \rightarrow \cdots \rightarrow 2 \rightarrow 1 \rightarrow 0.
\]
\[
\overline{\infty} \rightarrow 2n-2 \rightarrow 2n-3 \rightarrow \cdots \rightarrow 2 \rightarrow 1 \rightarrow 0.
\]
The maximal element is the zero perversity $\overline{0}=0$, the minimal element is the top perversity $\overline{t}=2n-2$ for $\Pos{n}^{op}$ and $\infty$ for $\widehat{\Pos{n}}^{op}$. The partial addition $\oplus$ is just the classical addition and we put $\overline{p} \oplus \overline{q} := \overline{p+q}$ if $p+q \leq 2n-2$ for $\Pos{n}^{op}$ and $\widehat{\Pos{n}}^{op}$. The complementary perversity $\overline{q}$ of $\overline{p}$ is then $\overline{q} = \overline{t}-\overline{p} = \overline{t-p}$.
If we do not consider complex varieties but just pseudomanifold of dimension $n$ with only isolated singularities, we will still use a linear poset
\[
n-2 \rightarrow n-3 \rightarrow \cdots \rightarrow 2 \rightarrow 1 \rightarrow 0.
\]
Throughout this paper, every equation involving perversities will be considered in $\mathcal{P}^{op}$. For example $\max(\overline{p}, \overline{0}) = \overline{0}$ for all $\overline{p}$ and if $\overline{p}=2$ and $\overline{q}=1$, then $\overline{p} < \overline{q}$.
Intersection spaces were defined by Markus Banagl in \cite{Banagl2010} in an attempt to spatialize Poincaré duality for singular spaces. The construction of these spaces rely on the notion of \textit{spatial homology truncation} also introduced in \cite{Banagl2010}.
\begin{defi}
Given a simply connected CW-complex $K$ of dimension $n$ and an integer $k \leq n$. A spatial homology truncation of cut-off degree $k$ of $K$ is a CW-complex $\tr{k}{K}$ together with a comparison map
\[
f \col \tr{k}{K} \longrightarrow K
\]
such that
\begin{equation}
H_{r}(\tr{k}{K}) \cong
\begin{cases}
H_{r}(K) & r < k,\\
0 & r \geq k.
\end{cases}
\end{equation}
The integer $k$ is called the cut off degree of the homological truncation.
\end{defi}
\begin{remark}
Such a truncation always exists provided that $K$ is simply connected and this truncation is in fact defined on $\mathbf{Z}$ and not just on $\mathbf{Q}$, see \cite{Banagl2010}.
\end{remark}
\begin{defi}
Let $X$ be a compact, connected, oriented pseudomanifold of dimension $n$ and denote by $\Sigma = \lbrace \sigma_{1}, \dots, \sigma_{\nu} \rbrace$ the singular locus of $X$. The pseudomanifold $X$ is called supernormal if the link $L_{i}$ of each singularity $\sigma_{i} \in \Sigma$ is simply connected.
We denote by $\super$ the category of supernormal complex projective varieties with only isolated singularities together with the morphisms $f \col X \rightarrow Y$ such that $f(X_{reg}) \subset Y_{reg}$.
\end{defi}
For the rest of this paper, we assume that the definition of a supernormal pseudomanifold $X$ includes the fact that $X$ is a connected pseudomanifold of dimension $n$ (the compacity and orientability assumptions being automatic since we work in projective spaces $\mathbf{C}P^{n}$).
Before giving our definition of intersection spaces, let us define which cut off degree we use with respect to the perversities for the spatial homological truncation. This definition will be different from the one in \cite{Banagl2010} and will be more suited to our notion of coperverse cdga we will introduce in definition \ref{def:coperverse_cdga}.
Let $K$ be a simply connected CW-complex of dimension $n$ and suppose given a perversity $\overline{p}$. We set that the cut-off degree is directly given by the perversity $\overline{p}$ and we denote it by $\tr{\overline{p}}{K}$. That is
\begin{equation}
H_{r}(\tr{\overline{p}}{K}) \cong
\begin{cases}
H_{r}(K) & \text{if } r \leq p\\
0 & \text{if } r > p.
\end{cases}
\end{equation}
Note that we also swap the strict and large inequalities in the definition. We will use this convention for the rest of this paper.
By convention we also define $\tr{\overline{\infty}}{K} = K$.
Given a supernormal pseudomanifold $X$ with isolated singularities,
\[
L(\Sigma, X) := \sqcup_{\sigma_{i}} L_{i}
\]
is then the disjoint union of simply connected topological manifold of dimension $n-1$. Denote by $X_{reg} := X - \Sigma$ the regular part of $X$. We denote by $\cotr{\overline{p}}{L_{i}}$ the homotopy cofiber of the map
\[
f_{i} \col \tr{\overline{p}}{L_{i}} \rightarrow L_{i}.
\]
We have maps
\[
f^{i} \col L_{i} \longrightarrow \cotr{\overline{p}}{L_{i}}.
\]
\begin{defi}
\label{def:intersectionspace2}
The intersection space $\I{p}{X}$ of the space $X$ is defined by the following homotopy pushout diagram
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=2em, column sep=5em, text height=1.5ex, text depth=0.25ex]
{ L(\Sigma,X) & X_{reg} \\
\bigsqcup_{i} \cotr{\overline{p}}{L_{i}} & \I{p}{X} \\};
\path[->]
(m-1-1) edge node[] {} (m-1-2);
\path[dashed,->]
(m-2-1) edge node[] {} (m-2-2);
\path[->]
(m-1-1) edge node[] {} (m-2-1);
\path[dashed,->]
(m-1-2) edge node[] {} (m-2-2);
\end{tikzpicture}
\]
\end{defi}
We shall use this definition of intersection spaces for the rest of the paper. Note that with this definition we have $\I{\infty}{X} = \overline{X}$ which is the normalization of $X$. We will denote by $\HI{p}{\ast}{X} := H^{\ast}(\I{p}{X})$ and by $\redHI{p}{\ast}{X}$ the reduced cohomology. We then have
\[
\HI{p}{r}{X} =
\begin{cases}
\mathbf{Q} & r=0 \\
H^{r}(X) & 1 \leq r \leq p \\
H^{r}(X) \oplus \im{H^{r}(X_{reg}) \rightarrow H^{r}(L)} & r = p+1 \\
H^{r}(X_{reg}) & r > p+1
\end{cases}
\]
In particular, we have $\HI{0}{\ast}{X} = H^{\ast}(X_{reg})$ and $\HI{\infty}{\ast}{X} = H^{\ast}(\overline{X})$.
\begin{remark}
\begin{enumerate}
\item Our intersection spaces $\I{p}{X}$ are different from the intersection spaces originally defined in \cite{Banagl2010} since they are not defined as a homotopy cofiber. When there is only one isolated singularity, there is no difference between the two definitions. Differences arise only for the first cohomology group when there is more than one isolated singularity.
\item This convention also has to be compared at the level of algebraic models with \cite{Chataur2012}, where a $\overline{p}$-perverse rational model of a cone $cL$ on a topological space $L$ of dimension $n$ is given by a truncation in degree $\overline{p}(n)$ of the rational model of $L$. In our case, a rational model of the intersection space $\I{p}{cL}$ is then given by a unital cotruncation in degree $\overline{p}(n)$ of the rational model of $L$.
\end{enumerate}
\end{remark}
Let's compute the bounds of the different weight filtrations involved in $\HI{p}{r}{X}$ for a general perversity $\overline{p}$. Denote by $R^{r}(X_{reg},L) := \im{H^{r}(X_{reg}) \rightarrow H^{r}(L)}$.
\begin{lem}
For $r < n$, $R^{r}(X_{reg},L)$ is pure of weight $r$. For $r \geq n$, we have
\[
0 = W_{r} \subset W_{r+1} \subset \cdots \subset W_{2r} = R^{r}(X_{reg},L).
\]
\end{lem}
\begin{proof}
This follows from the semi purity of the link, see \cite{Steenbrink1983}. Since $\dim(\Sigma) =0$, the weight filtration on the cohomology of the link is semi-pure, meaning :
\begin{itemize}
\item the weights on $H^{r}(L)$ are less than or equal to $r$ for $r <n$,
\item the weights on $H^{r}(L)$ are greater or equal to $r+1$ for $r \geq n$.
\end{itemize}
Combined with the two following facts
\begin{itemize}
\item The filtration $0 \subset W_{r} \subset \cdots \subset W_{2r} = H^{r}(X_{reg})$.
\item $H^{r}(X_{reg}) \rightarrow H^{r}(L)$ is a morphism of mixed Hodge structures.
\end{itemize}
\end{proof}
We have three cases
\noindent\makebox[\textwidth]{
\(
\begin{array}{|c||c | c c c| c : c|}
\hline
\multicolumn{7}{|c|}{\text{First case : \,}\overline{p} < \overline{m} =n-1} \\
\hline
& 1 \leq r \leq \overline{p} & \multicolumn{3}{c|}{r=p+1} & \overline{p}+1 <r<n & n \leq r \\
\hline
-1 & 0 & 0 & & & & \\
0 & W_{0} & W_{0} & & & & \\
1 & W_{1} & W_{1} & & & & \\
\vdots & \vdots & \vdots & & & & \\
r-1 & W_{r-1} & W_{r-1} & & W_{r-1}=0 & W_{r-1}=0 & W_{r-1}=0 \\
\hline
r & W_{r} & W_{r} & \oplus & W_{r} & W_{r} & W_{r} \\
\hline
r+1 & & & & & & W_{r+1} \\
\vdots & & & & & & \vdots \\
2r-1 & & & & & & W_{2r-1} \\
2r & & & & & & W_{2r} \\
\hline
& H^{r}(X) & H^{r}(X) & \oplus & R^{r}(X_{reg},L) & H^{r}(X_{reg}) & H^{r}(X_{reg}) \\
\hline
\end{array}
\)}
\noindent\makebox[\textwidth]{
\(
\begin{array}{|c||c | c c c| c|}
\hline
\multicolumn{6}{|c|}{\text{Second case : \,}\overline{p} = \overline{m} =n-1} \\
\hline
& 1 \leq r \leq n-1 & \multicolumn{3}{c|}{r=p+1=n} & n \leq r \\
\hline
-1 & 0 & 0 & & & \\
0 & W_{0} & W_{0} & & & \\
1 & W_{1} & W_{1} & & & \\
\vdots & \vdots & \vdots & & & \\
r-1 & W_{r-1} & W_{r-1} & & & W_{r-1}=0 \\
\hline
r & W_{r} & W_{r} & \oplus & W_{r}=0 & W_{r} \\
\hline
r+1 & & & & W_{r+1} & W_{r+1} \\
\vdots & & & & \vdots & \vdots \\
2r-1 & & & & W_{2r-1} & W_{2r-1} \\
2r & & & & W_{2r} & W_{2r} \\
\hline
& H^{r}(X) & H^{r}(X) & \oplus & R^{r}(X_{reg},L) & H^{r}(X_{reg}) \\
\hline
\end{array}
\)}
\noindent\makebox[\textwidth]{
\(
\begin{array}{|c||c : c | c c c| c|}
\hline
\multicolumn{7}{|c|}{\text{Third case : \,}\overline{p} > \overline{m} = n-1} \\
\hline
& 1 \leq r \leq n & n < r \leq p & \multicolumn{3}{c|}{ r=p+1} & p+1<r \\
\hline
-1 & 0 & & & & & \\
0 & W_{0} & & & & & \\
1 & W_{1} & & & & & \\
\vdots & \vdots & & & & & \\
r-1 & W_{r-1} & W_{r-1}=0 & W_{r-1}=0& & & W_{r-1}=0 \\
\hline
r & W_{r} & W_{r} & W_{r} & \oplus & W_{r}=0 & W_{r} \\
\hline
r+1 & & & & & W_{r+1} & W_{r+1} \\
\vdots & & & & & \vdots & \vdots \\
2r-1 & & & & & W_{2r-1} & W_{2r-1} \\
2r & & & & & W_{2r} & W_{2r} \\
\hline
& H^{r}(X) & H^{r}(X) & H^{r}(X) & \oplus & R^{r}(X_{reg},L) & H^{r}(X_{reg}) \\
\hline
\end{array}
\)}
\subsection{Coperverse algebras and their homotopy theory}
\subsubsection{Coperverse algebras}
Let $\mathbf{k}$ be a fixed field of characteristic zero.
\begin{defi}
\label{def:coperverse_cdga}
A $n$-coperverse commutative differential graded algebra over $\mathbf{k}$, coperverse cdga for short, is a functor
\[
A_{\overline{\bullet}} \col \Pos{n}^{op} \longrightarrow \mathsf{CDGA}_{\mathbf{k}}.
\]
That is for all perversities $\overline{p} \in \Pos{n}^{op}$, $A_{\overline{p}}$ is a bigraded $\mathbf{k}$-algebra $(A_{\overline{p}}^{k})_{k \in \mathbf{N}}$, together with a linear differential $d \col A_{\overline{p}}^{k} \rightarrow A_{\overline{p}}^{k+1}$ and an associative product $\mu \col A_{\overline{p}}^{i} \times A_{\overline{p}}^{j} \rightarrow A_{\overline{p}}^{i+j}$.
We assume that products and differentials satisfy graded commutativity, Leibniz rules, and are compatible with poset maps. That is for every $\overline{p} \leq \overline{q}$ in $\Pos{n}^{op}$ we have the following commutative diagrams.
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ A_{\overline{p}} \times A_{\overline{p}} & A_{\overline{p}} \\
A_{\overline{q}} \times A_{\overline{q}} & A_{\overline{q}} \\};
\path[->]
(m-1-1) edge node[auto] {$\mu$} (m-1-2);
\path[->]
(m-2-1) edge node[auto,swap] {$\mu$} (m-2-2);
\path[->]
(m-1-1) edge node[auto,swap] {$(\pomap{p}{q}, \pomap{p}{q})$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$\pomap{p}{q}$} (m-2-2);
\end{tikzpicture}
\quad
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ A_{\overline{p}} & A_{\overline{p}} \\
A_{\overline{q}} & A_{\overline{q}} \\};
\path[->]
(m-1-1) edge node[auto] {$d$} (m-1-2);
\path[->]
(m-2-1) edge node[auto,swap] {$d$} (m-2-2);
\path[->]
(m-1-1) edge node[auto,swap] {$\pomap{p}{q}$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$\pomap{p}{q}$} (m-2-2);
\end{tikzpicture}
\]
We denote by $H_{\overline{\bullet}}(A, \mathbf{k}) := H(A_{\overline{\bullet}},d)$.
\end{defi}
We denote by $\cdga{k}$ the category of coperverse cdga's over $\mathbf{k}$.
Note that with this definition, we have an extended product over the whole family $(A_{\overline{p}})_{\overline{p} \in \Pos{n}^{op}}$. Indeed, for every $\overline{p} \leq \overline{q}$ in $\Pos{n}^{op}$, denote by $\mu_{\overline{p},\overline{q}}$ the following composition
\[
\mu_{\overline{p},\overline{q}} \col A_{\overline{p}} \times A_{\overline{q}} \overset{(\pomap{p}{q}, \mathrm{id})}{\longrightarrow} A_{\overline{q}} \times A_{\overline{q}} \overset{\mu}{\longrightarrow} A_{\overline{q}}.
\]
\begin{defi}
The map $\mu_{\overline{\bullet},\overline{\bullet}}$ defined for all $\overline{p} \leq \overline{q}$ in $\mathcal{P}^{op}$ by the above composition is called the extended product over the family $(A_{\overline{p}})_{\overline{p} \in \mathcal{P}^{op}}$.
\end{defi}
\begin{remark}
\begin{enumerate}
\item
The following diagram, where $T$ is the twist isomorphism $T(a,b) := (-1)^{|a|\cdot|b|}(b,a)$, commutes. Because of that and for the sake of simplicity, we will then adopt the following convention. Each time a product $A_{\overline{p}} \times \cdots \times A_{\overline{q}}$ will appear, we will consider that the perversities are put in order, that is $\overline{p} \leq \cdots \leq \overline{q}$ in $\Pos{n}^{op}$.
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ A_{\overline{p}} \times A_{\overline{q}} & A_{\overline{q}} \times A_{\overline{q}} & A_{\overline{q}}\\
A_{\overline{q}} \times A_{\overline{p}} & A_{\overline{q}} \times A_{\overline{q}} & \\};
\path[->]
(m-1-1) edge node[auto] {$(\pomap{p}{q}, \mathrm{id})$} (m-1-2);
\path[->]
(m-2-1) edge node[auto] {$(\mathrm{id},\pomap{p}{q})$} (m-2-2);
\path[->]
(m-1-1) edge node[auto,swap] {$T$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$T$} (m-2-2);
\path[->]
(m-1-2) edge node[auto] {$\mu$} (m-1-3);
\path[->]
(m-2-2) edge node[auto,swap] {$\mu$} (m-1-3);
\end{tikzpicture}
\]
\item The extended product $\mu_{\overline{\bullet},\overline{\bullet}}$ verifies Leibniz rule, is associative and compatible with poset maps and morphisms of coperverse algebras. That is all $\overline{p} \leq \overline{q} \leq \overline{r}$ in $\Pos{n}^{op}$ we have the commutative diagram,
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ A_{\overline{p}} \times A_{\overline{q}} \times A_{\overline{r}} & A_{\overline{p}} \times A_{\overline{r}} \\
A_{\overline{q}} \times A_{\overline{r}} & A_{\overline{r}} \\};
\path[->]
(m-1-1) edge node[auto] {$(\mathrm{id}, \mu_{\overline{q},\overline{r}})$} (m-1-2);
\path[->]
(m-1-1) edge node[auto,swap] {$(\mu_{\overline{p},\overline{q}}, \mathrm{id})$} (m-2-1);
\path[->]
(m-2-1) edge node[auto,swap] {$\mu_{\overline{q},\overline{r}}$} (m-2-2);
\path[->]
(m-1-2) edge node[auto] {$\mu_{\overline{p},\overline{r}}$} (m-2-2);
\end{tikzpicture}
\]
and for all $\overline{p_{1}} \leq \overline{p_{2}} \leq \overline{q_{1}} \leq \overline{q_{2}}$ in $\Pos{n}^{op}$ we have the commutative diagram.
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ A_{\overline{p_{1}}} \times A_{\overline{q_{1}}} & A_{\overline{q_{1}}} \\
A_{\overline{p_{2}}} \times A_{\overline{q_{2}}} & A_{\overline{q_{2}}} \\};
\path[->]
(m-1-1) edge node[auto] {$\mu_{\overline{p_{1}},\overline{q_{1}}}$} (m-1-2);
\path[->]
(m-1-1) edge node[auto,swap] {$\pomap{p_{1}}{p_{2}} \times \pomap{q_{1}}{q_{2}}$} (m-2-1);
\path[->]
(m-2-1) edge node[auto,swap] {$\mu_{\overline{p_{2}},\overline{q_{2}}}$} (m-2-2);
\path[->]
(m-1-2) edge node[auto] {$\pomap{q_{1}}{q_{2}}$} (m-2-2);
\end{tikzpicture}
\]
\end{enumerate}
\end{remark}
Since $\mu_{\overline{p},\overline{p}} = \mu$ for all $\overline{p}$ we will always consider the family $(A_{\overline{p}})_{\overline{p} \in \Pos{n}^{op}}$ endowed with the extended product. We then denote a coperverse cdga by $(A_{\overline{\bullet}}, \mu_{\overline{\bullet},\overline{\bullet}})$.
\subsubsection{Homotopy theory of coperverse algebras}
We now define a model structure on the category of coperverse cdga's by using the formalism of Reedy categories. The definitions and results involving Reedy categories can be found in \cite{Hovey1999}.
First, recall the model structure of $\mathsf{CDGA}_{\mathbf{k}}$. The projective model structure on $\mathsf{CDGA}_{\mathbf{k}}$ is given by the following
\begin{itemize}
\item the weak equivalences are the quasi-isomorphims,
\item the fibrations are the degreewise surjections,
\item the cofibrations are the retracts of relative Sullivan algebras.
\end{itemize}
For $n \in \mathbf{N}$, consider the semifree dga's
\[
S(n) := (\wedge \mathbf{k}[n],d=0)
\]
where $\mathbf{k}[n]$ denotes the graded vector space which is $\mathbf{k}$ in degree $n$ and $0$ otherwise. For $n \geq 1$, consider the semifree dga's
\[
D(n) :=
\begin{cases}
0 & n=0, \\
(\wedge(\mathbf{k}[n+1] \oplus \mathbf{k}[n]),d=0) & n>0
\end{cases}
\]
and write
\[
i_{n} \col S(n) \rightarrow D(n)
\]
for the morphism that send the generator of degree $n$ to the generator of degree $n$. If $n=0$ then this is the unique morphism $0 \rightarrow 0$, and for $n>0$
\[
j_{n} \col 0 \rightarrow D(n).
\]
\begin{propo}
\label{prop:gen_cofib_cdga}
The sets $I := \lbrace i_{n} \rbrace_{n} \cup \lbrace S(0) \rightarrow 0\rbrace$, and $J:= \lbrace j_{n} \rbrace_{n>0}$ are the sets of generating cofibrations and acyclic cofibrations, respectively, of $\mathsf{CDGA}_{\mathbf{k}}$. The category $\mathsf{CDGA}_{\mathbf{k}}$ is then cofibrantly generated.
\end{propo}
Before talking about Reedy categories, note that we have an exact evaluation functor
\[
Ev_{\overline{p}} \col \cdga{k} \longrightarrow \mathsf{CDGA}_{\mathbf{k}}
\]
that send $A_{\overline{\bullet}}$ to $A_{\overline{p}}$, this functor admits an exact left adjoint $F_{\overline{p}}$ defined by $F_{\overline{p}}(A)_{\overline{q}} = A$ if $\overline{p} \leq \overline{q}$ and zero otherwise.
\begin{defi}
Let $\mathcal{C}$ be a small category and $\mathcal{C}' \subset \mathcal{C}$ a subcategory. The subcategory $\mathcal{C}'$ is said to be a lluf subcategory if the objects of $\mathcal{C}'$ and $\mathcal{C}$ are the same.
\end{defi}
\begin{defi}[Reedy category]
Let $\mathcal{C}$ be a small category together with a degree function $\deg \col \mathcal{C} \longrightarrow \mathbf{N}$ defined on the objects and suppose that we have two lluf subcategories $\overrightarrow{\mathcal{C}}$ and $\overleftarrow{\mathcal{C}}$. We say that $(\mathcal{C}, \overrightarrow{\mathcal{C}}, \overleftarrow{\mathcal{C}})$ is a Reedy category if the two following conditions are satisfied.
\begin{enumerate}
\item If $\alpha \col c \rightarrow c'$ is a non-identity map in $\overrightarrow{\mathcal{C}}$ (resp. in $\overleftarrow{\mathcal{C}}$) then $\deg (c) < \deg(c')$ (resp. $\deg (c) > \deg(c')$).
\item Every map $\alpha$ in $\mathcal{C}$ has a unique factorization
\[
\begin{cases}
\alpha & = \overrightarrow{\alpha} \circ \overleftarrow{\alpha}, \\
\overrightarrow{\alpha} & \in \overrightarrow{\mathcal{C}}, \\
\overleftarrow{\alpha} & \in \overleftarrow{\mathcal{C}}.
\end{cases}
\]
\end{enumerate}
\end{defi}
\begin{exmp}
\begin{enumerate}
\item A discrete category $\mathcal{C}$, that is a category where $\mathcal{C}(x,y) = \lbrace \mathrm{id}_{x} \rbrace$ if and only if $x=y$ and the empty set otherwise, is a Reedy category where all the objects are of degree 0.
\item \label{exmp:Poset_reedy_cat} Let $\mathcal{P}$ be a finite poset. We define every minimal element to be of degree 0 and we define the degree of an element $p \in \mathcal{P}$ to be the length of the longest path of non-identity maps from an element of degree zero to $p$. If we have $p \rightarrow p'$ with $p \neq p'$ then necessarily we have $\deg p < \deg p'$. The poset $\mathcal{P}$ is then endowed with a structure of Reedy category with
\[
\begin{cases}
\overrightarrow{\mathcal{P}}& = \mathcal{P}, \\
\overleftarrow{\mathcal{P}} & = \mathrm{Disc}(\mathcal{P}).
\end{cases}
\]
where $\mathrm{Disc}(\mathcal{P})$ is the discrete category underlying the poset $\mathcal{P}$, every elements of $\mathrm{Disc}(\mathcal{P})$ are of degree 0.
\end{enumerate}
\end{exmp}
For every Reedy category $\mathcal{C}$ there exist subcategories $\mathcal{C}_{<k}$ of objects of degree strictly inferior to $n$. Let then $F \col \mathcal{C} \rightarrow \mathcal{M}$ a functor which we suppose covariant, consider $c \in \mathcal{C}$ with $\deg c=n$, we have the two objects and maps
\[
L^{c}F \overset{\ell_{c}}{\longrightarrow} F(c) \overset{m_{c}}{\longrightarrow} M^{c}X,
\]
where
\[
L^{c}F := \mathrm{colim}(\partial(\overrightarrow{\mathcal{C}}_{<k}/c) \overset{U_{c}}{\longrightarrow} \mathcal{C} \overset{F}{\longrightarrow} \mathcal{M}),
\]
\[
M^{c}F := \lim(\partial(c/\overleftarrow{\mathcal{C}}_{<k}) \overset{U_{c}}{\longrightarrow} \mathcal{C} \overset{F}{\longrightarrow} \mathcal{M}).
\]
with $\partial(\overrightarrow{\mathcal{C}}_{<k}/c)$ and $\partial(c/\overleftarrow{\mathcal{C}}_{<k})$ are the two full subcategories of respectively $\overrightarrow{\mathcal{C}}_{<k}/c$ and $c/\overleftarrow{\mathcal{C}}_{<k}$ where we have removed the identity object $c \rightarrow c$.
\begin{defi}
The objects $L^{c}F$ and $M^{c}F$ are respectively called the $c$-th latching and $c$-th matching objects. The maps $\ell_{c}$ and $m_{c}$ are then the $c$-th latching and $c$-th matching maps.
\end{defi}
Given a map $F \rightarrow G$ in $\mathsf{Fun}(\mathcal{C}, \mathcal{M})$, we define the $c$-th relative latching map by the following diagram of pushout
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ L^{c}F & F(c) & \\
L^{c}G & \cdot & \\
& & G(c) \\};
\path[->]
(m-1-1) edge node[auto] {} (m-1-2);
\path[->]
(m-1-1) edge node[auto] {} (m-2-1);
\path[->]
(m-1-2) edge node[below=1em, left=1em] {$\lrcorner$} (m-2-2);
\path[->]
(m-2-1) edge node[auto] {} (m-2-2);
\path[->]
(m-1-2) edge node[auto] {} (m-3-3);
\path[->]
(m-2-1) edge node[auto] {} (m-3-3);
\path[->,dashed]
(m-2-2) edge node[auto] {} (m-3-3);
\end{tikzpicture}
\]
\noindent and the $c$-th relative matching map by the following diagram of pullback
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ F(c) & & \\
& \cdot & M^{c}F \\
& G(c) & M^{c}G \\};
\path[->,dashed]
(m-1-1) edge node[auto] {} (m-2-2);
\path[->]
(m-1-1) edge node[auto] {} (m-2-3);
\path[->]
(m-1-1) edge node[auto] {} (m-3-2);
\path[->]
(m-2-2) edge node[auto] {} (m-2-3);
\path[->]
(m-2-2) edge node[above=1em, right=1em] {$\ulcorner$} (m-3-2);
\path[->]
(m-3-2) edge node[auto] {} (m-3-3);
\path[->]
(m-2-3) edge node[auto] {} (m-3-3);
\end{tikzpicture}
\]
\begin{thm}[\cite{Hovey1999}, 5.2.5]
Let $\mathcal{M}$ be a model category et let $\mathcal{C}$ be a Reedy category. Then there is a model category on $\mathsf{Fun}(\mathcal{C}, \mathcal{M})$ such that :
\begin{enumerate}
\item the weak equivalences are defined pointwise,
\item the cofibrations are the maps $F \rightarrow G$ such that each relative latching map
\[
L^{c}G \coprod_{L^{c}F} F(c) \longrightarrow G(c)
\]
is a cofibration in $\mathcal{M}$,
\item the fibrations are the maps $F \rightarrow G$ such that each relative matching map
\[
F(c) \longrightarrow G(c) \times_{M^{c}G} M^{c}F
\]
is a fibration in $\mathcal{M}$.
\end{enumerate}
\end{thm}
We now apply this result to our context. We endow $\Pos{n}^{op}$ with the structure of a Reedy category defined in the item \ref{exmp:Poset_reedy_cat} of the last example.
Let $A_{\overline{\bullet}} \col \Pos{n}^{op} \rightarrow \mathsf{CDGA}_{\mathbf{k}}$ be a coperverse cdga and $\overline{p} \in \Pos{n}^{op}$ such that $\deg \overline{p} = k$. We have
\[
L^{\overline{p}}A_{\overline{\bullet}} := \mathrm{colim}(\partial(\mathcal{P}^{op}_{<k}/\overline{p}) \overset{U_{\overline{p}}}{\longrightarrow} \mathcal{P}^{op} \overset{A_{\overline{\bullet}}}{\longrightarrow} \mathcal{M}) = \mathrm{colim}_{\overline{p} < \overline{q}} A_{\overline{q}}
\]
and
\[
M^{\overline{p}}A_{\overline{\bullet}} := \lim(\partial(\overline{p}/\mathrm{Disc}(\mathcal{P})) \overset{U_{\overline{p}}}{\longrightarrow} \mathcal{P}^{op} \overset{A_{\overline{\bullet}}}{\longrightarrow} \mathcal{M}) = 0.
\]
Computing the relative latching and matching map we get the following result
\begin{thm}
\label{thm:model_structure}
The category $\cdga{k}$ has a structure of a cofibrantly generated model category which we call the projective model structure. In this model category, the weak equivalences are the quasi-isomorphisms and the fibrations are the surjections.
\end{thm}
\begin{proof}
The computations of weak equivalences and fibrations are clear.
The fact that $\cdga{k}$ is cofibrantly generated comes from \cite[Remark 5.1.8]{Hovey1999}, the generating cofibrations are the $\lbrace F_{\overline{p}}(i) \rbrace_{i \in I, \overline{p} \in \Pos{n}^{op}}$ and the generating acyclic cofibrations are the $\lbrace F_{\overline{p}}(j) \rbrace_{i \in J, \overline{p} \in \Pos{n}^{op}}$ where $I$ and $J$ are the sets defined in the proposition \ref{prop:gen_cofib_cdga}.
\end{proof}
For clarity, we give the following definition as a result of the previous theorem.
\begin{defi}
Let $f_{\overline{\bullet}} \col A_{\overline{\bullet}} \rightarrow B_{\overline{\bullet}}$ be a morphism of coperverse algebras. The morphism $f_{\overline{\bullet}}$ is
\begin{enumerate}
\item A quasi-isomorphism if, for every perversity $\overline{p} \in \Pos{n}^{op}$, the induced map $H^{\ast}_{\overline{p}}(A) \rightarrow H^{\ast}_{\overline{p}}(B)$ is an isomorphism.
\item A fibration if, for every perversity $\overline{p} \in \Pos{n}^{op}$, the induced map $f_{\overline{p}} \col A_{\overline{p}} \rightarrow B_{\overline{p}}$ is a degreewise surjection.
\end{enumerate}
\end{defi}
We denote by $\Ho{\cdga{\mathbf{k}}}$ the homotopy category associated to the model category structure on $\cdga{\mathbf{k}}$. That is the category defined by formally inverting quasi-isomorphisms.
\begin{remark}
There are many ways to put a model structures on $\cdga{k}$. Indeed the category $\mathsf{CDGA}_{\mathbf{k}}$ also has an injective model structure where the weak equivalences are the quasi-isomorphisms and the cofibrations are the injections and we could have choose this model structure to do the computations.
On the other hand we could have chose the projective or injective model structure on $\cdga{k}$ coming from $\mathsf{CDGA}_{\mathbf{k}}$ rather than doing computations using Reedy categories. But since $\mathsf{CDGA}_{\mathbf{k}}$ is a combinatorial model category all the ways mentioned above are guaranteed to be Quillen equivalent to the projective model structure on $\cdga{k}$.
By the way, all these model structures share the same weak equivalences.
\end{remark}
\section{Coperverse rational models}
\label{section:coperv_algebras}
\subsection{Coperverse cdga's associated with a morphism of cdga's}
The tools in this section are modified versions of the one appearing the work of Chataur and Cirici \cite{Chataur} on the interactions between intersection cohomology and mixed Hodge structures.
Let $(A,d) \in \mathsf{CDGA}_{\mathbf{k}}$. We denote by $\mathbf{k}(t, dt) := \wedge (t,dt)$ the free cdga generated by $t$ and $dt$ with $\deg t =0$, $\deg dt =1$ and $d(t)=dt$.
\begin{defi}
\label{def:At,dt}
We denote by $A(t,dt) := A \otimes_{\mathbf{k}} \mathbf{k}(t, dt)$. For $\lambda \in \mathbf{k}$ we also define the evaluation map
\[
\delta_{\lambda} \col A(t,dt) \longrightarrow A
\]
by $\delta_{\lambda}(t) = \lambda$ and $\delta_{\lambda}(dt)=0$.
\end{defi}
For all $r \geq 0$, we have the following short exact sequence
\[
0 \longrightarrow \ker d^{r} \longrightarrow A^{r} \longrightarrow \mathrm{Coim\,}d^{r} \longrightarrow 0
\]
where $\mathrm{Coim\,}d^{r} := A^{r} / \ker d^{r}$. Denote by $s_{r} \col \mathrm{Coim\,}d^{r} \rightarrow A^{r}$ a choice of section. For all $r \geq 0$, we denote by $\coim{r} := \im{s_{r}}$, the differential $d^{r}$ induces the isomorphism $\coim{r} \rightarrow \im{d^{r}}$.
\begin{defi}
The unital $\overline{p}$-cotruncation of $A(t,dt)$ is defined by
\[
\dgaucotr{p}{A(t,dt)} := A^{0} \oplus \dgacotr{p}{A(t,dt)}.
\]
where $\dgacotr{p}{A(t,dt)}$ is defined by
\[
\dgacotr{p}{A(t,dt)}^{r} :=
\begin{cases}
A^{r}\otimes \mathbf{k}[t]t \oplus A^{r-1}\otimes \mathbf{k}[t]dt & r < p \\
A^{p-1}\otimes \mathbf{k}[t]dt \oplus A^{p} \otimes \mathbf{k}[t]t \oplus \coim{p} & r=p \\
A^{r-1}\otimes \mathbf{k}[t]dt \oplus A^{r}\otimes \mathbf{k}[t] & r> p
\end{cases}
\]
\end{defi}
\begin{lem}
\label{lem:coperv_alg}
$\dgaucotr{\bullet}{A(t,dt)}$ is a coperverse cdga.
\end{lem}
\begin{proof}
Consider first $\dgacotr{p}{A(t,dt)}$.
The compatibility of $\dgacotr{\bullet}{A(t,dt)}$ with the differential $d(\dgacotr{p}{A(t,dt)}) \subset \dgacotr{p}{A(t,dt)}$ and product $\dgacotr{p}{A(t,dt)} \times \dgacotr{p}{A(t,dt)} \rightarrow \dgacotr{p}{A(t,dt)}$ is clear by construction. We detail the compatibility with the poset maps. By unicity of the maps $\pomap{p}{q}$, every $\pomap{p}{q}$ is a composition of poset maps $\pomap{k+1}{k}$ so we only detail these ones. We have
\[
\dgacotr{k+1}{A(t,dt)}^{r} :=
\begin{cases}
A^{r} \otimes \mathbf{k}[t]t \oplus A^{r-1} \otimes \mathbf{k}[t]dt & r < k+1 \\
A^{k} \otimes \mathbf{k}[t]dt \oplus A^{k+1} \otimes \mathbf{k}[t]t \oplus \coim{k+1} & r=k+1 \\
A^{r-1}\otimes \mathbf{k}[t]dt \oplus A^{r} \otimes \mathbf{k}[t] & r> k+1
\end{cases}
\]
and
\[
\dgacotr{k}{A(t,dt)}^{r} :=
\begin{cases}
A^{r} \otimes \mathbf{k}[t]t \oplus A^{r-1} \otimes \mathbf{k}[t]dt & r < k \\
A^{k-1}\otimes \mathbf{k}[t]dt \oplus A^{k} \otimes \mathbf{k}[t]t \oplus \coim{k} & r=k \\
A^{r-1}\otimes \mathbf{k}[t]dt \oplus A^{r} \otimes \mathbf{k}[t] & r> k.
\end{cases}
\]
For $r \leq k$ or $r >k+1$, $\pomap{k+1}{k}$ is the identity map. For $r=k+1$, since $A^{k+1}\otimes \mathbf{k}[t] = A^{k+1} \oplus A^{k+1}\otimes \mathbf{k}[t]t$, $\pomap{k+1}{k}$ is an injection.
Now for $\dgaucotr{p}{A(t,dt)} := A^{0} \oplus \dgacotr{p}{A(t,dt)}$ the compatibility with the differential and the poset maps is clear by the same arguments than above. The product $\dgaucotr{p}{A(t,dt)} \times \dgaucotr{p}{A(t,dt)} \rightarrow \dgaucotr{p}{A(t,dt)}$ is also clear by construction.
\end{proof}
Let now $f \col A \longrightarrow B$ be a morphism of cdga's. Given a perversity $\overline{p} \in \Pos{n}^{op}$, we consider the following pull-back diagram in the category $\CDGA{k}$.
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \pb{p}{f} & \dgaucotr{p}{B(t,dt)} \\
A & B \\};
\path[->]
(m-1-1) edge node[auto] {} (m-1-2);
\path[->]
(m-2-1) edge node[auto,swap] {$f$} (m-2-2);
\path[->]
(m-1-1) edge node[above=1em, right=1em] {$\ulcorner$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$\delta_{1}$} (m-2-2);
\end{tikzpicture}
\]
With the product, the differential defined component-wise and the compatibility with poset maps.
\begin{propo}
The pull-back $\pb{\bullet}{f}$ is a coperverse cdga.
\end{propo}
\begin{defi}
\label{def:pb_coperv_cdga}
$\pb{\bullet}{f}$ is the coperverse cdga associated to the morphism of cdga's $f \col A \longrightarrow B$.
\end{defi}
\begin{defi}\label{def:sharp_pcdga}
Let $(A_{\overline{\bullet}}, \mu_{\overline{\bullet},\overline{\bullet}})$ be a coperverse cdga and $r \in \mathbf{Z}$. We say that $(A_{\overline{\bullet}}, \mu_{\overline{\bullet},\overline{\bullet}})$ is a $r$-sharp coperverse cdga if the product satisfies the two following conditions
\begin{enumerate}
\item \textbf{Unity} For $A_{\overline{p}}^{i} \times A_{\overline{0}}^{j} \rightarrow A_{\overline{0}}^{i+j}$ the product lifts to
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ & A_{\overline{p}}^{i+j} \\
A_{\overline{p}}^{i} \times A_{\overline{0}}^{j} & A_{\overline{0}}^{i+j} \\};
\path[->]
(m-2-1) edge node[auto] {$\mu_{\overline{p},\overline{0}}$} (m-2-2);
\path[->,dashed]
(m-2-1) edge node[auto] {} (m-1-2);
\path[->]
(m-1-2) edge node[auto] {$\pomap{p}{0}$} (m-2-2);
\end{tikzpicture}
\]
\item \textbf{Factorization} For $\overline{p}, \overline{q} \neq \overline{0}$ and $i,j \neq 0$ the product lifts to
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ & A_{\overline{p+q+r}}^{i+j} \\
A_{\overline{p}}^{i} \times A_{\overline{q}}^{j} & A_{\overline{q}}^{i+j} \\};
\path[->]
(m-2-1) edge node[auto] {$\mu_{\overline{p},\overline{q}}$} (m-2-2);
\path[->,dashed]
(m-2-1) edge node[auto] {} (m-1-2);
\path[->]
(m-1-2) edge node[auto] {$\pomap{p+q+r}{q}$} (m-2-2);
\end{tikzpicture}
\]
\end{enumerate}
We assume that this lift satisfies all the properties of the product $\mu$. That is Leibniz rule with respect to the differential, graded commutativity and compatibility with poset maps and morphisms of cdga's.
\end{defi}
\begin{lem}
$\dgaucotr{\bullet}{A(t,dt)}$ is a $(-1)$-sharp coperverse cdga.
\end{lem}
\begin{cor}\label{cor:J(f)_-1_pcdga}
Let $f \col A \longrightarrow B$ be a morphism of cdga's, then $\pb{\bullet}{f}$ is a $(-1)$-sharp coperverse cdga
\end{cor}
\begin{remark}
\begin{enumerate}
\item The first condition means that the final cdga $A_{\overline{0}}$, since $\overline{0}$ is the maximal element of $\Pos{n}^{op}$, plays the role of the unit for the family $(A_{\overline{p}})_{\overline{p} \in \Pos{n}^{op}}$ and in particular for the unit $\eta_{\overline{0}} \col \mathbf{k} \rightarrow A_{\overline{0}}^{0}$ we have $A_{\overline{p}}^{i} \times A_{\overline{0}}^{0} \rightarrow A_{\overline{p}}^{i}$ for every $\overline{p}$ and every $i \geq 0$.
\item coperverse cdga's are meant to model the rational cohomology of intersection spaces $\HI{p}{k}{X}$. Since the $\I{p}{X}$ are topological spaces their cohomology bear an inner cup-product which is reflected in the definition of the coperverse cdga's. The lift is here to show the interactions between the different $\HI{p}{k}{X}$.
\end{enumerate}
\end{remark}
\subsection{Coperverse rational model of intersection spaces}
Let $X \in \super$ of complex dimension $n$, we denote by $\Sigma$ the singular locus of $X$.
Let $T$ be a closed algebraic neighbourhood of the singular locus in $X$ such that the inclusion $\Sigma \subset T$ is a homotopy equivalence. Such a neighbourhood exists and is constructed with "rug functions", see \cite[p.144]{Peters2007} or \cite{Durfee1983}.
The link $L:=L(\Sigma,X)$ of $\Sigma$ in $X$ is defined by $L := \partial T \simeq T^{\ast} := T - \Sigma$. The inclusion $i \col L \hookrightarrow X_{reg}$ of the link into the regular part of $X$ induces a morphism of cdga's over $\mathbf{Q}$
\[
i^{\ast} \col \Apl{X_{reg}} \longrightarrow \Apl{L}.
\]
Let $\overline{p} \in \Pos{n}^{op}$ be a perversity, the rational model of the intersection space $\I{p}{X}$ is given by $\AI{p}{X} := \pb{p}{i^{\ast}}$, which is the following pull-back diagram, see \cite{Klimczak2015}.
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \pb{p}{i^{\ast}} & \dgaucotr{p}{\Apl{L}(t,dt)} \\
\Apl{X_{reg}} & \Apl{L} \\};
\path[->]
(m-1-1) edge node[auto] {} (m-1-2);
\path[->]
(m-2-1) edge node[auto] {$i^{\ast}$} (m-2-2);
\path[->]
(m-1-1) edge node[above=1em, right=1em] {$\ulcorner$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$\delta_{1}$} (m-2-2);
\end{tikzpicture}
\]
\begin{defi}
The coperverse cdga $\AI{\bullet}{X}$ is called the coperverse rational model of the intersection spaces $\I{\bullet}{X}$.
\end{defi}
If $A_{\overline{\bullet}}$ is a coperverse cdga, its cohomology is also a coperverse cdga. We then have the following proposition.
\begin{propo}
$\HI{\bullet}{\ast}{X}$ is a coperverse cdga.
\end{propo}
We have an isomorphism of coperverse cdga $H^{\ast}(\AI{p}{X}) \cong \HI{\bullet}{\ast}{X}$.
This then defines a functor
\[
AI_{\overline{\bullet}} \col \super \longrightarrow \Ho{\cdga{k}}.
\]
If we only consider the coperverse rational model of $X \in \super$, we then have that $AI_{\overline{\bullet}}(X)$ is a $(-1)$-sharp coperverse cdga by corollary \ref{cor:J(f)_-1_pcdga}. But if we only want to consider the cohomology coperverse algebra $\HI{\bullet}{\ast}{X}$, we can have an even sharper result.
\begin{propo}
Let $X \in \super$ with only isolated singularities. Then $(\HI{\bullet}{\ast}{X},0)$ is a 1-sharp coperverse cdga. That is we have
\[
\begin{cases}
\HI{0}{i}{X} \otimes \redHI{p}{j}{X} \longrightarrow \redHI{p}{i+j}{X} & \\
\redHI{p}{i}{X} \otimes \redHI{q}{j}{X} \longrightarrow \redHI{p+q+1}{i+j}{X} & p+q+1 \leq 2n-2.
\end{cases}
\]
\end{propo}
\begin{remark}
It is important to make a difference between the extended product $\mu_{\overline{\bullet}, \overline{\bullet}}$ and the property of sharpness. The existence of the extended product is a consequence of the definition \ref{def:coperverse_cdga} and as such every coperverse cdga defined in the same way naturally has an extended product.
The property of sharpness of our coperverse algebras defined in \ref{def:pb_coperv_cdga} is a consequence of our methods of construction. There might be coperverse algebras which do not have any property of sharpness, but still have an extended product.
\end{remark}
\section{Hodge Theory}
\label{section:Hodge_th}
\subsection{Coperverse mixed Hodge algebras}
We now put a mixed Hodge structure on the coperverse rational model of $X \in \super$.
\begin{defi}
A coperverse filtered cdga $(A_{\overline{\bullet}}, W)$ is a coperverse cdga $A_{\overline{\bullet}}$ together with a filtration $\lbrace W_{m}A_{\overline{\bullet}} \rbrace_{m \in \mathbf{Z}}$ such that
\begin{enumerate}
\item $W_{m-1}A_{\overline{p}} \subset W_{m}A_{\overline{p}}$ and $d(W_{m}A_{\overline{p}}) \subset W_{m}A_{\overline{p}}$, for all $m \in \mathbf{Z}$ and all $\overline{p} \in \Pos{n}$,
\item $W_{m}A_{\overline{p}}.W_{n}A_{\overline{p}} \subset W_{m+n}A_{\overline{p}}$,
\item $W_{m}A_{\overline{p}} \subset W_{m}A_{\overline{q}}$ for all $\overline{p} \leq \overline{q}$ in $\Pos{n}^{op}$,
\item The filtration $W$ is exhaustive and biregular : for all $n \geq 0$ and all $\overline{p} \in \mathcal{P}^{op}$ there exist integers $m$ and $l$ such that $W_{m}A^{n}_{\overline{p}} =0$ and $W_{l}A^{n}_{\overline{p}} = A^{n}_{\overline{p}}$.
\end{enumerate}
\end{defi}
\begin{defi}
A coperverse mixed Hodge cdga over $\mathbf{Q}$ is a coperverse filtered cdga $(A_{\overline{\bullet}}, W)$ with a filtration $F$ on $A_{\overline{\bullet}} \otimes \mathbf{C}$ such that for all $n \geq 0$ and all $\overline{p} \in \Pos{n}^{op}$,
\begin{enumerate}
\item the triple $(A^{n}_{\overline{p}}, \Dec{W}, F)$ is a mixed Hodge structure,
\item the differential $d \col A_{\overline{p}}^{k} \rightarrow A_{\overline{p}}^{k+1}$, the product $\mu \col A_{\overline{p}}^{i} \times A_{\overline{p}}^{j} \rightarrow A_{\overline{p}}^{i+j}$ and the poset maps $\pomap{p}{q} \col A_{\overline{p}}^{k} \rightarrow A_{\overline{q}}^{k}$ are morphisms of mixed Hodge structures.
\end{enumerate}
The filtration $W$ is called the weight filtration and the filtration $F$ is called the Hodge filtration.
\end{defi}
We will denote, by an abuse of notations, such a mixed Hodge cdga by the triple $(A_{\overline{\bullet}}, W, F)$ with in mind the fact that $F$ is not defined on $A_{\overline{\bullet}}$ but on its complexification $A_{\overline{\bullet}} \otimes \mathbf{C}$. The filtration $\Dec{W}$ is the Deligne's décalage of the weight filtration defined in \cite[15]{Deligne1971} which is given by
\[
\Dec{W_{p}}A_{\overline{\bullet}}^{n} := W_{p-n}A_{\overline{\bullet}}^{n} \cap d^{-1}(W_{p-n-1}A_{\overline{\bullet}}^{n+1}).
\]
We denote by $\mhcdga{Q}$ the category of coperverse mixed Hodge cdga's over $\mathbf{Q}$.
\begin{lem}
Let $(A_{\overline{\bullet}}, W, F)$ be a coperverse mixed Hodge cdga, then the extended product $\mu_{\overline{\bullet},\overline{\bullet}}$ is a morphism of mixed Hodge structure.
\end{lem}
\begin{defi}
A coperverse filtered cdga $(A_{\overline{\bullet}}, W)$ is said to be $r$-sharp if $A_{\overline{\bullet}}$ is a filtered coperverse cdga such that the lift is compatible with the filtration $\lbrace W_{m}A_{\overline{\bullet}} \rbrace_{m \in \mathbf{Z}}$. That is we have the two following conditions
\begin{enumerate}
\item \textbf{Filtered unity} For $W_{m}A_{\overline{p}}^{i} \times W_{n}A_{\overline{0}}^{j} \rightarrow W_{m+n}A_{\overline{0}}^{i+j}$ the product lifts to
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ & W_{m+n}A_{\overline{p}}^{i+j} \\
W_{m}A_{\overline{p}}^{i} \times W_{n}A_{\overline{0}}^{j} & W_{m+n}A_{\overline{0}}^{i+j} \\};
\path[->]
(m-2-1) edge node[auto] {$\mu_{\overline{p},\overline{0}}$} (m-2-2);
\path[->,dashed]
(m-2-1) edge node[auto] {} (m-1-2.south west);
\path[->]
(m-1-2) edge node[auto] {$\pomap{p}{0}$} (m-2-2);
\end{tikzpicture}
\]
\item \textbf{Filtered factorization} For $\overline{p}, \overline{q} \neq \overline{0}$ and $i,j \neq 0$ the product lifts to
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ & W_{m+n}A_{\overline{p+q+r}}^{i+j} \\
W_{m}A_{\overline{p}}^{i} \times W_{n}A_{\overline{q}}^{j} & W_{m+n}A_{\overline{q}}^{i+j} \\};
\path[->]
(m-2-1) edge node[auto] {$\mu_{\overline{p},\overline{q}}$} (m-2-2);
\path[->,dashed]
(m-2-1) edge node[auto] {} (m-1-2.south west);
\path[->]
(m-1-2) edge node[auto] {$\pomap{p+q+r}{q}$} (m-2-2);
\end{tikzpicture}
\]
\end{enumerate}
\end{defi}
\begin{defi}
A $r$-sharp coperverse mixed Hodge cdga over $\mathbf{Q}$ is a coperverse mixed Hodge cdga $(A_{\overline{\bullet}}, W,F)$ such that the lift is a morphism of mixed Hodge structure.
\end{defi}
Consider $\mathbf{Q}(t,dt)$ together with the \textit{bête} filtration $\sigma$, that is the multiplicative filtration with $t$ of weight $0$ and $dt$ of weight $-1$. We endow $\mathbf{C}(t,dt) := \mathbf{Q}(t,dt) \otimes \mathbf{C}$ with the bête filtration $\sigma$ and the trivial filtration $t$, that is decreasing filtration given by
\[
0=t^{1}\mathbf{C}(t,dt) \subset t^{0}\mathbf{C}(t,dt) =\mathbf{C}(t,dt).
\]
Since $\Dec{\sigma} = t$ the triple $(\mathbf{Q}(t,dt), \sigma, t)$ is a mixed Hodge cdga.
Given another mixed Hogde cdga $(A,W,F)$, since the category of mixed Hodge structure is abelian the triple
\[
(A(t,dt), W \ast \sigma, F \ast t)
\]
is again a mixed Hodge cdga where the filtrations are defined by convolution. That is we have
\[
(W \ast \sigma)_{m}A(t,dt)^{n} := W_{m}A^{n} \otimes \mathbf{Q}[t] \oplus W_{m+1}A^{n-1} \otimes \mathbf{Q}[t]dt
\]
and
\[
(F \ast t)^{k}A(t,dt) := F^{k}A \otimes \mathbf{C}(t,dt).
\]
The evaluation map $\delta_{1}$ is strictly compatible with filtrations.
\begin{lem}
Let $(A,W,F)$ be a mixed Hodge cdga. Then for all $\overline{p} \in \Pos{n}^{op}$, $\dgaucotr{p}{A(t,dt)}$ is a $(-1)$-sharp mixed Hodge cdga.
\end{lem}
\begin{proof}
The triple $(A(t,dt), W \ast \sigma, F \ast t)$ is a mixed Hodge cdga, for all $\overline{p} \in \mathcal{P}^{op}$, $\dgaucotr{p}{A(t,dt)}$ is a sub-algebra with the filtrations induced by restriction.
The differential is a morphism of mixed Hodge structure since the differential on $(A(t,dt), W \ast \sigma, F \ast t)$ is and $d(\dgaucotr{p}{A(t,dt)}) \subset \dgaucotr{p}{A(t,dt)}$.
The poset maps $\pomap{k+1}{k}$, $k \geq 0$, are the identity everywhere but at the cut-off degree $k+1$ where they are canonical inclusions, $\pomap{k+1}{k}$ in then compatible with both filtrations and by composition so are the $\pomap{p}{q}$.
The extended product $\dgaucotr{p}{A(t,dt)}^{i} \times \dgaucotr{q}{A(t,dt)}^{j} \rightarrow \dgaucotr{q}{A(t,dt)}^{i+j}$ being defined as the composition of $\mu$ with poset maps $\pomap{p}{q}$, it is a morphism of mixed Hodge structure.
The sharpness comes from the fact that $\dgaucotr{p}{A(t,dt)}$ is $(-1)$-sharp and that the product is a morphism of mixed Hodge structure.
\end{proof}
Let then $f \col (A, W, F) \rightarrow (B, W, F)$ be a morphism of mixed Hodge cdga. Since the category of mixed Hodge structures is abelian, see \cite[Theorem 2.3.5]{Deligne1971}, we have the following proposition.
\begin{propo}
The coperverse cdga $\pb{\bullet}{f}$ is a coperverse mixed Hodge cdga.
\end{propo}
\subsection{Mixed Hodge structure on the coperverse rational model of the intersection spaces \texorpdfstring{$\I{\bullet}{X}$}{IpX}}
\begin{defi}[\cite{Cirici2014}]
A mixed Hodge diagram of cdga's over $\mathbf{Q}$ consists of a filtered cdga $(A_{\mathbf{Q}}, W)$ over $\mathbf{Q}$, a bifiltered cdga $(A_{\mathbf{C}}, W, F)$ over $\mathbf{C}$, together with a string of filtered $E_{1}$-quasi-isomorphisms from $(A_{\mathbf{Q}}, W)\otimes \mathbf{C}$ to $(A_{\mathbf{C}}, W)$. in addition, the following axioms must hold :
\begin{itemize}
\item The weight filtrations $W$ are regular and exhaustive. The Hodge filtration $F$ is biregular. The cohomology $H(A_{\mathbf{Q}})$ has finite type.
\item For all $p \in \mathbf{Z}$, the differential of $\gr{p}{W}{A_{\mathbf{C}}}$ is strictly compatible with $F$.
\item For all $n \geq 0$ and all $p \in \mathbf{Z}$, the filtration $F$ induced on $H^{n}(\gr{p}{W}{A_{\mathbf{C}}})$ defines a pure Hodge structure of weight $p+n$ on $H^{n}(\gr{p}{W}{A_{\mathbf{Q}}})$.
\end{itemize}
\end{defi}
Morphisms of mixed Hodge diagrams are defined by level-wise morphisms of bifiltered cdga's such that the associated diagram is strictly commutative. Forgetting the multiplicative structure gives back the notion of mixed Hodge complex defined by Deligne in \cite[section 8.1]{Deligne1974}.
\begin{defi}
Let $X$ be a topological space. A mixed Hodge diagram for $X$ is a mixed Hodge diagram $M(X)$ such that $M(X)_{\mathbf{Q}} \simeq \Apl{X}$, that is its rational component is quasi-isomorphic to the rational algebra of piecewise linear forms on $X$.
\end{defi}
The following theorem is a modified version of a theorem appearing in \cite[theorem 3.10]{Chataur} stating that the intersection homotopy type of a complex variety $X$ with only isolated singularities carries well-defined mixed Hodge structures.
\begin{thm}
\label{thm:coperverse_MI}
Let $X \in \super$ of complex dimension $n$. There exist a coperverse mixed Hodge cdga $\MI{\bullet}{X}$ together with a string of quasi-isomorphisms
\[
\MI{\bullet}{X} \leftarrow \ast \rightarrow AI_{\overline{\bullet}}(X)
\]
such that :
\begin{enumerate}
\item $\MI{\bullet}{X} = \pb{\bullet}{\tilde{\iota}}$ where $\tilde{\iota} \col M(X_{reg}) \rightarrow M(L)$ is a model of mixed Hodge cdga's for the rational homotopy type of the inclusion $i \col L \hookrightarrow X_{reg}$.
\item there is an isomorphism of coperverse mixed Hodge cdga's
\[
H^{\ast}(\MI{\bullet}{X}) \cong \HI{\bullet}{\ast}{X}.
\]
\item The mixed Hodge cdga's $\MI{0}{X}$ and $\MI{\infty}{X}$ defines respectively the mixed Hodge structure on the rational homotopy type of the regular part $X_{reg}$ of $X$ and on the normalisation $\overline{X}$ of $X$.
\item The differential of $\MI{\bullet}{X}$ satisfies $d(W_{p}\MI{\bullet}{X}) \subset W_{p-1}\MI{\bullet}{X}$.
\end{enumerate}
This defines a functor
\[
MI_{\overline{\bullet}} \col \super \longrightarrow \Ho{\mhcdga{Q}}.
\]
\end{thm}
\begin{proof}
The proof is similar to \cite[theorem 3.10]{Chataur}.
By \cite[theorem 3.2.1]{Durfee1988}, there is a morphism of mixed Hodge diagrams $M(X_{reg}) \rightarrow M(L)$ induced by the inclusion $i \col L \hookrightarrow X_{reg}$. The rational component of this morphism is the morphism $i^{\ast} \col \Apl{X_{reg}} \rightarrow \Apl{L}$ of rational piecewise linear forms induced by the inclusion $i \col L \hookrightarrow X_{reg}$. By \cite[theorem 3.19]{Cirici2014}, there is a commutative diagram of mixed Hodge diagrams
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=2.5em, text height=1.5ex, text depth=0.25ex]
{ \Apl{X_{reg}} & \Apl{L} \\
\ast & \ast \\
M(X_{reg}) & M(L) \\};
\path[->]
(m-1-1) edge node[auto] {$i^{\ast}$} (m-1-2);
\path[->]
(m-2-1) edge node[style] {} (m-2-2);
\path[->]
(m-3-1) edge node[auto] {$\tilde{\iota}$} (m-3-2);
\path[->]
(m-2-1) edge node[auto] {} (m-1-1);
\path[->]
(m-2-2) edge node[auto] {} (m-1-2);
\path[->]
(m-2-1) edge node[auto] {} (m-3-1);
\path[->]
(m-2-2) edge node[auto] {} (m-3-2);
\end{tikzpicture}
\]
where the vertical maps are quasi-isomorphisms and $\tilde{\iota}$ is a map of mixed Hodge cdga's whose differential satisfies $d(W_{p}) \subset W_{p-1}$. We then let $\MI{\bullet}{X} := \pb{\bullet}{\tilde{\iota}}$. This construction is functorial for stratified morphisms. The above commutative diagram defines a string of quasi-isomorphisms from $\MI{\bullet}{X}$ to $\AI{\bullet}{X}$.
Let now show that $\MI{\bullet}{X}$ is a coperverse mixed Hodge cdga. Consider the mixed Hodge cdga $M(L)(t,dt)$ defined as in definition \ref{def:At,dt}. Then $\dgaucotr{p}{M(L)(t,dt)}$ is a complex of mixed Hodge structure for every perversities $\overline{p} \in \Pos{n}^{op}$. The product
\[
\dgaucotr{p}{M(L)(t,dt)} \times \dgaucotr{q}{M(L)(t,dt)} \longrightarrow \dgaucotr{q}{M(L)(t,dt)}
\]
and the poset maps
\[
\dgaucotr{p}{M(L)(t,dt)} \longrightarrow \dgaucotr{q}{M(L)(t,dt)}
\]
for $\overline{p} \leq \overline{q} \in \Pos{n}^{op}$ are strictly compatible with filtrations. Since the category of mixed Hodge structures is abelian, for each $n \geq 0$ and each $\overline{p} \in \Pos{n}^{op}$, the vector space $\MI{p}{X}^{n}$ carries a mixed Hodge structure. The compatibility with product and poset maps is a matter of verifications. This proves the first three properties.
The differential on $\MI{p}{X}$ being defined via the pull-back of cdga's whose differential satisfies $d(W_{p}) \subset W_{p-1}$, this also holds for $\MI{p}{X}$.
Functoriality follows by construction.
\end{proof}
From this result we can deduce the two following product structure.
\begin{cor}
Let $X \in \super$ with only isolated singularities, then the family $\lbrace \MI{p}{X} \rbrace_{(\overline{p})}$ is a $(-1)$-sharp mixed Hodge coperverse cdga.
\end{cor}
\begin{cor}
Let $X \in \super$ with only isolated singularities, then the family of algebras $\lbrace \HI{0}{\ast}{X},\redHI{1}{\ast}{X}, \dots, \redHI{2n-2}{\ast}{X} \rbrace$ is endowed with a product
\[
\begin{cases}
\HI{0}{i}{X} \otimes \redHI{p}{j}{X} \longrightarrow \redHI{p}{i+j}{X} & \\
\redHI{p}{i}{X} \otimes \redHI{q}{j}{X} \longrightarrow \redHI{p+q+1}{i+j}{X} & p+q+1 \leq 2n-2.
\end{cases}
\]
This product is a morphism of mixed Hodge structure.
\end{cor}
Due to the method of construction of the coperverse mixed Hodge cdga $\MI{\bullet}{X}$, we have the following commutative diagram of mixed Hodge cdga's.
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ M(X_{reg}) & M(L) & \dgaucotr{k+1}{M(L)(t,dt)} \\
M(X_{reg}) & M(L) & \dgaucotr{k}{M(L)(t,dt)}\\
0 & 0 & M(L,k)\\};
\path[->]
(m-1-1) edge node[auto] {$\tilde{\iota}$} (m-1-2);
\path[->]
(m-1-3) edge node[auto,swap] {$\delta_{1}$} (m-1-2);
\path[->]
(m-2-1) edge node[auto] {$\tilde{\iota}$} (m-2-2);
\path[->]
(m-2-3) edge node[auto,swap] {$\delta_{1}$} (m-2-2);
\path[->]
(m-3-1) edge node[auto] {} (m-3-2);
\path[->]
(m-3-3) edge node[auto] {} (m-3-2);
\path[->]
(m-1-1) edge node[auto] {} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {} (m-2-2);
\path[injection]
(m-1-3) edge node[auto] {$\pomap{k+1}{k}$} (m-2-3);
\path[->]
(m-2-1) edge node[auto] {} (m-3-1);
\path[->]
(m-2-2) edge node[auto] {} (m-3-2);
\path[->>]
(m-2-3) edge node[auto] {} (m-3-3);
\end{tikzpicture}
\]
Where each elements of the last row is the quotient of the previous elements in the same column. That is $M(L,k)$ is the mixed Hodge cdga quotient such that $H^{i}(M(L,k)) = H^{k}(L)$ for $i =k$ and zero otherwise. Taking the pullback on each rows we then have a short exact sequence of mixed Hodge structure
\[
0 \longrightarrow \MI{k+1}{X} \longrightarrow \MI{k}{X} \longrightarrow M(L,k) \longrightarrow 0.
\]
This short exact sequence induces a long exact sequence of mixed Hodge structure and extends to arbitrary perversities. That is we have
\begin{cor}
Let $X \in \super$ with only isolated singularities and two perversities $\overline{p} \leq \overline{q} \in \widehat{\Pos{n}}^{op}$. We have a long exact sequence of mixed Hodge structures
\[
\cdots \rightarrow \HI{p}{i}{X} \rightarrow \HI{q}{i}{X} \rightarrow H^{i}(M(L,q,p)) \rightarrow \HI{p}{i+1}{X} \rightarrow \cdots
\]
where
\[
H^{i}(M(L,q,p)) =
\begin{cases}
H^{i}(L) & q \leq i < p, \\
0 & \text{otherwise.}
\end{cases}
\]
\end{cor}
\section{Weight spectral sequence}
\label{section:Weight_SS}
Let $(B,W,F)$ a mixed Hodge cdga, then $(B(t,dt),W \ast \sigma,F \ast t)$ is again a mixed Hodge cdga where the filtrations are given by
\[
(W \ast \sigma)_{m}B(t,dt)^{n} := W_{m}B^{n} \otimes \mathbf{Q}[t] \oplus W_{m+1}B^{n-1} \otimes \mathbf{Q}[t]dt
\]
and
\[
(F \ast t)^{k}B(t,dt) := F^{k}B \otimes \mathbf{C}(t,dt).
\]
The graded subspace associated to the the weight filtration is then given by
\[
\gr{m}{W \ast \sigma}{B(t,dt)^{n}} = \gr{m}{W}{B^{n}} \otimes \mathbf{Q}[t] \oplus \gr{m+1}{W}{B^{n-1}} \otimes \mathbf{Q}[t]dt.
\]
Given a mixed Hodge cdga $(B,W,F)$, we then have a cohomological weight spectral sequence $E(B,W)$ whose $E_{1}$ page is defined by
\[
\E{W}{1}{p}{q}{B} := H^{p+q}(\gr{-p}{W}{B^{p+q}}).
\]
The spectral sequence associated to a coperverse filtered cdga $(A_{\overline{\bullet}}, W)$ is compatible with the multiplicative structure. Thus, for all $r \geq 0$, The term $E_{r}(A_{\overline{\bullet}}, W)$ is a coperverse bigraded algebra with differential $d_{r}$ of degree $(r,1-r)$.
\begin{lem}
\label{lem:E_tdt_commutes}
Let $(B,W,F)$ a mixed Hodge cdga, we have a canonical isomorphism of differential bigraded algebras
\[
E_{1}(B(t,dt), W \ast \sigma) \cong E_{1}(B,W)(t,dt)
\]
\end{lem}
\begin{lem}
\label{lem:pb_E1_commutes}
Let $f \col (A,W,F) \rightarrow (B,W,F)$ be a morphism of coperverse mixed Hodge cdga's. There is a quasi-isomorphism of coperverse differential bigraded algebras
\[
E_{1}(\pb{\bullet}{f},W) \overset{\sim}{\longrightarrow} \pb{\bullet}{(E_{1}(f,W))}.
\]
\end{lem}
\begin{proof}
The proof is similar to \cite[lemma 3.7]{Chataur} unless for the map
\[
E_{1}(\dgaucotr{\bullet}{B(t,dt)}, W \ast \sigma) \overset{\sim}{\rightarrow} \dgaucotr{\bullet}{E_{1}(B,W)(t,dt)}
\]
which is not an isomorphism but a quasi-isomorphism.
\end{proof}
\begin{lem}
\label{lem:A_to_E1}
Let $(A_{\overline{\bullet}}, W, F)$ be a coperverse mixed Hodge cdga such that
\[
d(W_{p}A_{\overline{\bullet}}) \subset W_{p-1}A_{\overline{\bullet}}.
\]
There is an isomorphism of complex coperverse cdga's
\[
A_{\overline{\bullet}} \otimes \mathbf{C} \cong E_{1}(A_{\overline{\bullet}} \otimes \mathbf{C}, W).
\]
\end{lem}
\begin{proof}
The proof is is the same as the proof of \cite[lemma 3.4]{Chataur} for perverse mixed Hodge cdga's.
\end{proof}
Let $(A,W)$ be a filtered cdga of finite type over a field $\mathbf{k}$ and $\mathbf{k} \subset \mathbf{K}$ a field extension. By \cite[theorem 2.26]{Cirici2014} we have that $A \cong E_{r}(A,W)$ if and only if $A \otimes_{\mathbf{k}} \mathbf{K} \cong E_{r}(A \otimes_{\mathbf{k}} \mathbf{K},W)$. For a coperverse cdga of finite type the same proof is valid. This implies the isomorphism of lemma \ref{lem:A_to_E1} descends to an isomorphism over $\mathbf{Q}$.
Let $X \in \super$ of complex dimension $n$. The inclusion $i \col L \hookrightarrow X_{reg}$ of the link into the regular part induces a morphism of multiplicative weight spectral sequence $E_{1}(i^{\ast}) \col E_{1}(X_{reg}) \rightarrow E_{1}(L)$. We define
\[
\EI{1}{\bullet}{X} := \pb{\bullet}{E_{1}(i^{\ast})}.
\]
This is a coperverse differential bigraded algebra whose cohomology satifies
\[
\EIbidg{2}{p}{r,s}{X} := H^{r,s}(\EI{1}{p}{X}) \cong \gr{s}{W}{\HI{p}{r+s}{X}}
\]
\begin{defi}
Let $X \in \super$ of complex dimension $n$. The spectral sequence $\EI{1}{\bullet}{X}$ defined by
\[
\EI{1}{\bullet}{X} := \pb{\bullet}{E_{1}(i^{\ast})}
\]
is called the coperverse weight spectral sequence associated to $X$.
\end{defi}
In \cite[theorem 3.12]{Chataur}, Chataur and Cirici prove the existence of a quasi-isomorphism between the rational perverse model $\IA{\bullet}{X}$ of a complex projective variety with only isolated singularities and the first term of its perverse weight spectral sequence $\IE{1}{\bullet}{X}$. This theorem can be modified to get the following one.
\begin{thm}
\label{thm:AI_to_EI1}
Let $X \in \super$ with only isolated singularities. There is a string of quasi-isomorphisms of coperverse cdga's from $\MI{\bullet}{X} \otimes \mathbf{C}$ to $\EI{1}{\bullet}{X} \otimes \mathbf{C}$. In particular, there is an isomorphism in $\mathrm{Ho}(\cdga{C})$ from $AI_{\overline{\bullet}}(X) \otimes \mathbf{C}$ to $\EI{1}{\bullet}{X} \otimes \mathbf{C}$.
\end{thm}
\begin{proof}
The proof is similar to \cite[theorem 3.12]{Chataur}.
Let $(\MI{\bullet}{X}, W,F)$ be the coperverse mixed Hodge cdga given by the theorem \ref{thm:coperverse_MI}. Since the differential satisfies
\[
d(W_{p}\MI{\bullet}{X}) \subset W_{p-1}\MI{\bullet}{X}
\]
by the lemma \ref{lem:A_to_E1} we have an isomorphism of complex coperverse cdga's $\MI{\bullet}{X} \otimes \mathbf{C} \cong E_{1}(\MI{\bullet}{X} \otimes \mathbf{C},W)$.
By construction, we have $\MI{\bullet}{X} := \pb{\bullet}{\tilde{\iota}}$, where
\[
\tilde{\iota} \col (M(X_{reg}),W,F) \rightarrow (M(L),W,F)
\]
is a morphism of mixed Hodge cdga's which computes the rational homotopy type of $\iota \col L \rightarrow X_{reg}$. Thus by lemma \ref{lem:pb_E1_commutes} we have a quasi-isomorphism of coperverse cdga's $E_{1}(\MI{\bullet}{X},W) \longrightarrow \pb{\bullet}{E_{1}(\tilde{\iota},W)}$. It remains to note that we have a string of quasi-isomorphisms from $\pb{\bullet}{{E_{1}(\tilde{\iota})}}$ to $\EI{1}{\bullet}{X} := \pb{\bullet}{E_{1}(i^{\ast})}$
\end{proof}
\begin{remark}
Suppose we have a topological space $X$ such that its rational model is endowed with an increasing filtration $W$, then one can consider the associated spectral sequence $E_{1}(X,W)$. The existence of a string of quasi-isomorphisms between the rational model of $X$ and the first page $E_{1}(X,W)$ is called the $E_{1}$-formality and is a property of complex algebraic varieties, see \cite{Cirici2014} and \cite{Chataur2015}. It is an interesting result that the intersection spaces of complex projective varieties have this property although they are not algebraic varieties.
\end{remark}
\begin{defi}
Let $X$ be a compact, connected oriented pseudomanifold of dimension $n$ with only isolated singularities. We say that $X$ is a $EI_{r,\overline{\bullet}}$-formal topological space if its coperverse rational model $\AI{\bullet}{X}$ can be endowed with an increasing bounded filtration $W$ such that there exists a string of quasi-isomorphisms between $\AI{\bullet}{X}$ and the $r$-th term of its associated spectral sequence $\EI{r}{\bullet}{X,W}$.
\end{defi}
With this definition, the theorem \ref{thm:AI_to_EI1} can be rephrased in the following corollary.
\begin{cor}
Let $X \in \super$ with only isolated singularities. The space $X$ is $EI_{1,\overline{\bullet}}$-formal with respect to the weight filtration.
\end{cor}
\subsection{The case of a smooth exceptional divisor}
\label{subsec:smooth_div}
\subsubsection{Notations}
Let $X$ be a complex projective variety of complex dimension $n$ with only normal isolated singularities. We denote by $\Sigma = \lbrace \sigma_{1},\dots, \sigma_{\nu} \rbrace
$ the singular locus of $X$ and by $X_{reg} := X - \Sigma$ its regular part. We also denote by $L := L(\Sigma, X)$ the link of $\Sigma$ in $X$ and by $i \col L \hookrightarrow X_{reg}$ the natural inclusion of the link into the regular part.
Since $\Sigma$ is discrete, we can write $L$ as a disjoint union $L= \sqcup_{\sigma_{i}} L_{i}$ where $L_{i}:= L(\sigma_{i}, X)$ is the link of $\sigma_{i} \in \Sigma$ in $X$. The assumption that $X$ is normal implies that $L_{i}$ is connected for all $\sigma_{i} \in \Sigma$.
From now on, we will always assume $X$ admits a resolution of singularities
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ D & \widetilde{X} \\
\Sigma & X \\};
\path[injection]
(m-1-1) edge node[auto] {$j$} (m-1-2);
\path[injection]
(m-2-1) edge node[auto,swap] {} (m-2-2);
\path[->]
(m-1-1) edge node[auto,swap] {} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$f$} (m-2-2);
\end{tikzpicture}
\]
such that the exceptional divisor $D := f^{-1}(\Sigma)$ is smooth.
We denote by
\[
j^{k} \col H^{k}(\widetilde{X}) \longrightarrow H^{k}(D) \text{ and } \gamma^{k} \col H^{k-2}(D) \longrightarrow H^{k}(\widetilde{X})
\]
the restriction maps and the Gysin maps induced by the inclusion $j$.
For all $k \geq 2$ we also denote by
\[
j^{k}_{\sharp} \col H^{k-2}(D) \overset{\gamma^{k}}{\longrightarrow} H^{k}(\widetilde{X}) \overset{j^{k}}{\longrightarrow} H^{k}(D)
\]
the composition of the two maps.
The morphism $E_{1}(i^{\ast}) \col E_{1}^{\ast,\ast}(X_{reg}) \rightarrow E_{1}^{\ast,\ast}(L)$ of weight spectral sequence induced by the inclusion $i \col L \hookrightarrow X_{reg}$ is defined by
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=3em, text height=2ex, text depth=0.25ex]
{E_{1}^{-1,s}(X_{reg}) & E_{1}^{0,s}(X_{reg}) & H^{s-2}(D) & H^{s}(\widetilde{X}) \\
E_{1}^{-1,s}(L) & E_{1}^{0,s}(L) & H^{s-2}(D) & H^{s}(D) \\};
\path[->]
(m-1-1) edge node[auto,swap] {$E_{1}^{-1,s}(i^{\ast})$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$E_{1}^{0,s}(i^{\ast}) \, =$} (m-2-2);
\path[->]
(m-1-3) edge node[auto,swap] {$\mathrm{id}$} (m-2-3);
\path[->]
(m-1-4) edge node[auto] {$j^{s}$} (m-2-4);
\path[->]
(m-1-1) edge node[auto] {$d$} (m-1-2);
\path[->]
(m-2-1) edge node[auto,swap] {$d$} (m-2-2);
\path[->]
(m-1-3) edge node[auto] {$\gamma^{s}$} (m-1-4);
\path[->]
(m-2-3) edge node[auto,swap] {$j^{s}_{\sharp}$} (m-2-4);
\end{tikzpicture}
\]
The algebra structure on $E_{1}^{\ast,\ast}(X_{reg})$ is given by the cup product of $H^{\ast}(\widetilde{X})$, together with the map
\[
\begin{array}{ccc}
H^{s}(\widetilde{X}) \times H^{s'}(D) & \longrightarrow & H^{s+s'}(D)\\
(x,a) & \longmapsto & j^{s}(x)\cdot a.\\
\end{array}
\]
This algebra structure is compatible with the differential $\gamma$ because $\gamma(j^{s}(x)\cdot a) = x \cdot \gamma(a)$.
The non-trivial products on $E_{1}^{\ast,\ast}(L)$ are the maps
\[
E_{1}^{0,s}(L) \times E_{1}^{r,s'}(L) \longrightarrow E_{1}^{r,s+s'}(L) \quad r \in \lbrace 0,1 \rbrace, s,s' \geq 0
\]
induced by the cup-product on $H^{\ast}(D)$.
The coperverse weight spectral sequence $\EI{1}{\bullet}{X} := \pb{\bullet}{E_{1}(i^{\ast})}$ for $X$ is then given by
\noindent\makebox[\textwidth]{
\(
\begin{array}{| c || c c c c c |}
\hline
s > p+1 & H^{s-2}(D)\otimes \mathbf{Q}[t] & \rightarrow & \mathcal{I}^{s}_{0} \oplus H^{s-2}(D)\otimes \mathbf{Q}[t]dt & \rightarrow & H^{s}(D)\otimes \mathbf{Q}[t]dt \\
\hline
s = p+1 & \coim{p} \oplus H^{s-2}(D)\otimes \mathbf{Q}[t]t & \rightarrow & \mathcal{I}^{s}_{0} \oplus H^{s-2}(D)\otimes \mathbf{Q}[t]dt & \rightarrow & H^{s}(D)\otimes \mathbf{Q}[t]dt \\
\hline
1 \leq s < p+1 & H^{s-2}(D)\otimes \mathbf{Q}[t]t & \rightarrow & \mathcal{I}^{s}_{1} \oplus H^{s-2}(D)\otimes \mathbf{Q}[t]dt & \rightarrow & H^{s}(D)\otimes \mathbf{Q}[t]dt \\
\hline
\hline
s = 0 & 0 & & \mathcal{I}^{0}_{0} & \rightarrow & H^{0}(D)\otimes \mathbf{Q}[t]dt \\
\hline
\hline
\EIbidg{1}{p}{r,s}{X} & r=-1 & & r=0 & & r=1 \\
\hline
\end{array}
\)}
Where
\begin{enumerate}
\item $\coim{p}$ is the image of the section of $d^{-1,s}_{1} \col E^{-1,s}_{1}(L) \rightarrow E^{0,s}_{1}(L)$, ie a section of $j^{s}_{\sharp} \col H^{s-2}(D) \rightarrow H^{s}(D)$. Note that $\coim{p}$ is just a computational tool and does not impact the value of the $EI_{2}$ term since it has been shown in \cite[theorem 2.18]{Banagl2010} that the values of $\HI{p}{k}{X}$ for rational coefficients are independent of the choices made during the construction.
\item $\mathcal{I}^{s}_{k}$, $k \in \lbrace 0,1 \rbrace$, is the vector space given by the following pullback square.
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \mathcal{I}^{s}_{k} & H^{s}(D)\otimes \mathbf{Q}[t]t^{k} \\
H^{s}(\widetilde{X}) & H^{s}(D) \\};
\path[->]
(m-1-1) edge node[auto] {} (m-1-2);
\path[->]
(m-2-1) edge node[auto] {$j^{s}$} (m-2-2);
\path[->]
(m-1-1) edge node[above=1em, right=1em] {$\ulcorner$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$\delta_{1}$} (m-2-2);
\end{tikzpicture}
\]
\item The differential $\difb{p}{-1,s} \col \EIbidg{1}{p}{-1,s}{X} \rightarrow \EIbidg{1}{p}{0,s}{X}$ is defined by
\[
\sum a_{i}t^{i} \mapsto \left((\sum \gamma^{s}(a_{i}), \sum j^{s}_{\sharp}(a_{i})t^{i}), \sum i a_{i} t^{i-1}dt \right) \quad a_{i} \in H^{s-2}(D).
\]
\item The differential $\difb{p}{0,s} \col \EIbidg{1}{p}{0,s}{X} \rightarrow \EIbidg{1}{p}{1,s}{X}$ is defined by
\[
\left((x, \sum a_{i}t^{i}), \sum b_{i} t^{i}dt \right) \mapsto \sum ia_{i}t^{i-1}dt + \sum j^{s}_{\sharp}(b_{i})t^{i}dt
\]
with
\[
\begin{cases}
a_{i} \in H^{s}(D), b_{i} \in H^{s-2}(D), & \\
x \in H^{s}(\widetilde{X}), j^{s}(x) = \sum a_{i}. & \\
\end{cases}
\]
\end{enumerate}
We describe the internal algebra structure of the coperverse weight spectral sequence $\EIbidg{1}{p}{r,s}{X}$. Due to the method of construction, this algebra structure is similar to the external one on the perverse weight spectral sequence for intersection cohomology in \cite{Chataur}.
The algebra structure is described by the following maps. We set $x,x' \in H^{\ast}(\widetilde{X})$ and $a,a',b,b' \in H^{\ast}(D)\otimes \mathbf{Q}[t]$.
\[
\begin{array}{ccc}
\EIbidg{1}{p}{0,s}{X} \times \EIbidg{1}{p}{0,s'}{X} & \longrightarrow & \EIbidg{1}{p}{0,s+s'}{X} \\
((x,a+b\cdot dt),(x',a' + b' \cdot dt)) & \longmapsto & (xx', aa' + (a'b + b'a)dt) \\
\end{array}
\]
\[
\begin{array}{ccc}
\EIbidg{1}{p}{0,s}{X} \times \EIbidg{1}{p}{1,s'}{X} & \longrightarrow & \EIbidg{1}{p}{1,s+s'}{X} \\
((x,a+b\cdot dt),(a' \cdot dt)) & \longmapsto & aa' \cdot dt \\
\end{array}
\]
\[
\begin{array}{ccc}
\EIbidg{1}{p}{-1,s}{X} \times \EIbidg{1}{p}{1,s'}{X} & \longrightarrow & \EIbidg{1}{p}{0,s+s'}{X} \\
(a,a'\cdot dt) & \longmapsto & aa' \cdot dt \\
\end{array}
\]
\[
\begin{array}{ccc}
\EIbidg{1}{p}{-1,s}{X} \times \EIbidg{1}{p}{0,s'}{X} & \longrightarrow & \EIbidg{1}{p}{-1,s+s'}{X} \\
(a,(x,a' + b' \cdot dt)) & \longmapsto & aa' \\
\end{array}
\]
Note that since $\coim{p} \subset H^{s-2}(D)$ and $\mathcal{I}^{s}_{1} \subset \mathcal{I}^{s}_{0}$, $\pomap{k+1}{k}$ induces a morphism of spectral sequences of bidegree $(0,0)$
\[
EI_{1}(\pomap{k+1}{k}) \col \EI{1}{k+1}{X} \rightarrow \EI{1}{k}{X}.
\]
This poset map extends the internal structure structure into an external one, meaning we then have an extended product
\[
\EIbidg{1}{p}{r,s}{X} \times \EIbidg{1}{q}{r',s'}{X} \longrightarrow \EIbidg{1}{q}{r+r',s+s'}{X}
\]
defined with the same map as before for the internal structure and following the same rules for $r,r',s,s'$.
By computing the cohomology of $\EI{1}{p}{X}$ we have
\[
\begin{array}{| c || c c c |}
\hline
s > p+1 & \ker \gamma^{s} & \coker{\gamma^{s}} & 0 \\
\hline
s = p+1 & 0 & \coker{\gamma^{s}_{|\coim{p}}} & 0 \\
\hline
1 \leq s < p+1 & 0 & \ker j^{s} & \coker{j^{s}} \\
\hline
\hline
s = 0 & 0 & H^{0}(\widetilde{X}) & 0 \\
\hline
\hline
\EIbidg{2}{p}{r,s}{X} & r=-1 & r=0 & r=1 \\
\hline
\end{array}
\]
Where $\gamma^{s}_{|\coim{p}}$ is the restriction of $\gamma^{s}$ to
\[
\coim{p} \rightarrow H^{s}(\widetilde{X}).
\]
We then have the following isomorphisms
\noindent\makebox[\textwidth]{
\(
\HI{p}{k}{X} =
\begin{cases}
H^{0}(\widetilde{X}) = \mathbf{Q} & k=0 \\
H^{k}(X) \cong \ker j^{k} \oplus \coker{j^{k-1}} & 1 \leq k < p+1 \\
H^{k}(X) \oplus \im{H^{k}(X_{reg}) \rightarrow H^{k}(L)} \cong \coker{\gamma^{k}_{|\coim{p}}} \oplus \coker{j^{k-1}} \oplus \ker \gamma^{k+1} & k = p+1 \\
H^{k}(X_{reg}) \cong \ker \gamma^{k+1} \oplus \coker{\gamma^{k}} & k > p+1
\end{cases}
\)}
\subsubsection{Remark on \texorpdfstring{$\coker{j^{0}}$}{coker j0}}
\label{subsubsec:coker}
It is important to note here that the values of $\ker j^{s}$ and $\coker{j^{s}}$ recorded in the array of the $EI_{2}$ term above start with $s=1$, meaning we don't take into account $\ker j^{0}$ and $\coker{j^{0}}$, this is intended.
Indeed, $\coker{j^{0}}$ accounts for the number of loops created when the intersection spaces are defined as a homotopy pushout over a single point, like in the original definition of \cite{Banagl2010}, this not the definition we use.
As a consequence, when we have multiple isolated singularities, the generalised Poincaré duality of the intersection spaces fails for $\redHI{p}{1}{X} \cong \redHI{q}{n-1}{X}$.
This is also one of the reason we modified the definition of intersection spaces. If we used the original definition of \cite{Banagl2010}, the mixed Hodge structure on $\HI{p}{k}{X}$ would never be pure unless if there is only one isolated singularity, which is the case where $\coker{j^{0}}=0$.
\subsubsection{Remark on the zero perversity}
The intersection space for the zero perversity is by definition \ref{def:intersectionspace2} the regular part $X_{reg}$ of the complex projective variety $X \in \super$ involved. The isomorphism given above by the $EI_{2}$ term gives
\[
\HI{0}{1}{X} = \coker{\gamma^{1}_{|\coim{0}}} \oplus \ker \gamma^{2}.
\]
Let's see that this coincides with $H^{1}(X_{reg})$.
Consider the term $\coker{\gamma^{1}_{|\coim{0}}}$, by definition $\coim{0}$ is defined as the image of a section of $j^{0}_{\sharp} \col H^{-1}(D)=0 \rightarrow H^{1}(D)$. So we have $\coim{0}=0$, and we then have $\coker{\gamma^{1}_{|\coim{0}}} \cong \coker{\gamma^{1}}.$
We then have what we wanted
\[
\HI{0}{1}{X} = \coker{\gamma^{1}} \oplus \ker \gamma^{2} = H^{1}(X_{reg}).
\]
\subsection{\texorpdfstring{$(\overline{p},r)$}{(p,r)}-purity implies \texorpdfstring{$(\overline{p},r)$}{(p,r)}-formality}
\begin{defi}
Let $0 \leq r \leq \infty$ be an integer and $\overline{p}$ a perversity. A morphism of coperverse cdga's $f_{\overline{\bullet}} \col A_{\overline{\bullet}} \rightarrow B_{\overline{\bullet}}$ is a $(\overline{p},r)$-quasi-isomorphism if for all perversities $\overline{s} \leq \overline{p}$ in $\mathcal{P}^{op}$ the induced morphism
\[
H_{\overline{s}}^{i}(f) \col H_{\overline{s}}^{i}(A) \longrightarrow H_{\overline{s}}^{i}(B)
\]
is an isomorphism for all $i \leq r$ and a monomorphism for $i = r+1$.
\end{defi}
\begin{defi}
\begin{enumerate}
\item A coperverse cdga $(A_{\overline{\bullet}},d)$ over $\mathbf{k}$ is said to be $(\overline{p},r)$-formal if there exist a string of $(\overline{p},r)$-quasi-isomorphisms from $(A_{\overline{\bullet}},d)$ to its cohomology $(H_{\overline{\bullet}}(A,\mathbf{k}),0)$ seen as a coperverse cdga with zero differential.
\item Let $X \in \super$, $\I{\bullet}{X}$ is said to be $(\overline{p},r)$-formal if its coperverse rational model $AI_{\overline{\bullet}}(X)$ is $(\overline{p},r)$-formal.
\item Let $X \in \super$, $\I{\bullet}{X}$ is said to be $(\overline{p},r)$-pure if the weight filtration $\HI{s}{k}{X}$ is pure of weight $k$ for all $k \leq r$ and for all perversities $\overline{s} \leq \overline{p}$ in $\mathcal{P}^{op}$.
\end{enumerate}
\end{defi}
\begin{thm}\label{thm:pure_is_formal}
Let $X \in \super$ of dimension $n$ with only isolated singularities. Let $r \geq 0$ be an integer and $\overline{p}$ a perversity. Suppose that $\I{\bullet}{X}$ is $(\overline{p},r)$-pure, then $\I{\bullet}{X}$ is $(\overline{p},r)$-formal.
\end{thm}
\begin{proof}
By theorem \ref{thm:AI_to_EI1}, we need to define a string of $(\overline{p},r)$-quasi-isomorphisms of differential bigraded algebras from $(\EIbidg{1}{s}{i,j}{X}, \difb{s}{i,j})$ to $(\EIbidg{2}{s}{i,j}{X}, 0)$ for $i+j \leq r$ and $\overline{s} \leq \overline{p}$ in $\Pos{n}^{op}$.
Given $X \in \super$ of dimension $n$ with only isolated singularities, the terms $EI_{1}$ and $EI_{2}$ of the spectral sequence look like.
\[
\begin{array}{| c || c : c : c |}
\hline
j=5 & \vdots & \vdots & \vdots \\
\hline
j=4 & \EIbidg{1}{\bullet}{-1,4}{X} & \EIbidg{1}{\bullet}{0,4}{X} & \EIbidg{1}{\bullet}{1,4}{X} \\
\hline
j=3 & \EIbidg{1}{\bullet}{-1,3}{X} & \EIbidg{1}{\bullet}{0,3}{X} & \EIbidg{1}{\bullet}{1,3}{X} \\
\hline
j=2 & \EIbidg{1}{\bullet}{-1,2}{X} & \EIbidg{1}{\bullet}{0,2}{X} & \EIbidg{1}{\bullet}{1,2}{X} \\
\hline
j=1 & \EIbidg{1}{\bullet}{-1,1}{X} & \EIbidg{1}{\bullet}{0,1}{X} & \EIbidg{1}{\bullet}{1,1}{X} \\
\hline
j=0 & 0 & \EIbidg{1}{\bullet}{0,0}{X} & \EIbidg{1}{\bullet}{1,0}{X} \\
\hline
\hline
\EIbidg{1}{\bullet}{i,j}{X} & i=-1 & i=0 & i=1 \\
\hline
\end{array}
\]
\[
\begin{array}{| c || c : c : c |}
\hline
j=5 & \vdots & \vdots & \vdots \\
\hline
j=4 & \gr{4}{W}{\HI{\bullet}{3}{X}} & \gr{4}{W}{\HI{\bullet}{4}{X}} & \gr{4}{W}{\HI{\bullet}{5}{X}} \\
\hline
j=3 & \gr{3}{W}{\HI{\bullet}{2}{X}} & \gr{3}{W}{\HI{\bullet}{3}{X}} & \gr{3}{W}{\HI{\bullet}{4}{X}} \\
\hline
j=2 & \gr{2}{W}{\HI{\bullet}{1}{X}} & \gr{2}{W}{\HI{\bullet}{2}{X}} & \gr{2}{W}{\HI{\bullet}{3}{X}} \\
\hline
j=1 & \gr{1}{W}{\HI{\bullet}{0}{X}} & \gr{1}{W}{\HI{\bullet}{1}{X}} & \gr{1}{W}{\HI{\bullet}{2}{X}} \\
\hline
j=0 & 0 & \gr{0}{W}{\HI{\bullet}{0}{X}} & \gr{0}{W}{\HI{\bullet}{1}{X}} \\
\hline
\hline
\EIbidg{2}{\bullet}{i,j}{X} & i=-1 & i=0 & i=1 \\
\hline
\end{array}
\]
The $(\overline{p},r)$-purity assumption implies that $\gr{j}{W}{\HI{s}{j-1}{X}}=0$ for all $j \leq r+1$ and $\gr{j}{W}{\HI{s}{j+1}{X}}=0$ for all $j \leq r-1$. This means that $\ker \difb{s}{-1,j} =0$ for all $j \leq r+1$ and $\im{\difb{s}{0,j}} = \EIbidg{1}{s}{1,j}{X}$ for all $j \leq r-1$.
Denote by $\FI{s}{i,j}{X}$ the bigraded differential algebra defined by, for all $\overline{s} \leq \overline{p}$ in $\Pos{n}^{op}$
\[
\begin{cases}
\FI{s}{-1,j}{X} := \EIbidg{1}{s}{-1,j}{X} & j \leq r+1, \\
\FI{s}{-1,j}{X} := 0 & j > r+1, \\
\FI{s}{0,j}{X} := \ker \difb{s}{0,j} & \forall \,j, \\
\FI{s}{1,j}{X} := 0 & \forall \, j.
\end{cases}
\]
The differential being $\difb{s}{i,j}$.
The bigraded differential algebra $\FI{s}{\ast,\ast}{X}$ has the following product structure
\[
\begin{cases}
\FI{s}{-1,j}{X} \times \FI{s}{-1,j}{X} \longrightarrow 0 & \forall \, j, \\
\FI{s}{-1,j}{X} \times \FI{s}{0,j'}{X} \longrightarrow \FI{s}{-1,j+j'}{X} & \forall \, j,j', \\
\FI{s}{0,j}{X} \times \FI{s}{0,j'}{X} \longrightarrow \FI{s}{0,j+j'}{X} & \forall \, j,j'.
\end{cases}
\]
which is well defined and is compatible with $\difb{s}{i,j}$ and poset maps $EI_{1}(\pomap{s+1}{s})$ for all $\overline{s} \leq \overline{p}$.
We then clearly have a inclusion $(\FI{s}{i,j}{X}, \difb{s}{i,j}) \hookrightarrow (\EIbidg{1}{s}{i,j}{X}, \difb{s}{i,j})$, the map $(\FI{s}{i,j}{X}, \difb{s}{i,j}) \rightarrow (\EIbidg{2}{s}{i,j}{X},0)$ is defined by the following commutative diagram where the dashed arrows are the zero map.
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=-1em, text height=2ex, text depth=0.25ex]
{\FI{s+1}{-1,j}{X} & & \ker d^{0,j}_{1, \overline{s+1}} & \\
& \FI{s}{-1,j}{X} & & \ker d^{0,j}_{1, \overline{s}} \\
\gr{j}{W}{\HI{s+1}{j-1}{X}} & & \gr{j}{W}{\HI{s+1}{j}{X}} & \\
& \gr{j}{W}{\HI{s}{j-1}{X}} & & \gr{j}{W}{\HI{s}{j}{X}}\\};
\path[injection]
(m-1-1) edge node[auto] {$\difb{s+1}{-1,j}$} (m-1-3);
\path[injection]
(m-2-2) edge node[above left] {$\difb{s}{-1,j}$} (m-2-4);
\path[dashed,->]
(m-3-1) edge node[auto,swap] {} (m-3-3);
\path[dashed,->]
(m-4-2) edge node[auto,swap] {} (m-4-4);
\path[dashed,->]
(m-1-1) edge node[auto] {} (m-3-1);
\path[dashed,->]
(m-2-2) edge node[auto] {} (m-4-2);
\path[->>]
(m-1-3) edge node[below right] {$p$} (m-3-3);
\path[->>]
(m-2-4) edge node[auto] {$p$} (m-4-4);
\path[->]
(m-1-1) edge node[auto] {$EI_{1}(\pomap{s+1}{s})$} (m-2-2);
\path[->]
(m-3-1) edge node[auto,swap] {$EI_{2}(\pomap{s+1}{s})$} (m-4-2);
\path[->]
(m-1-3) edge node[auto] {$EI_{1}(\pomap{s+1}{s})$} (m-2-4);
\path[->]
(m-3-3) edge node[auto,swap] {$EI_{2}(\pomap{s+1}{s})$} (m-4-4);
\end{tikzpicture}
\]
The string $(\EIbidg{1}{s}{i,j}{X}, \difb{s}{i,j}) \longleftarrow (\FI{s}{i,j}{X}, \difb{s}{i,j}) \rightarrow (\EIbidg{2}{s}{i,j}{X},0)$ then defines a $(\overline{p},r)$-quasi-isomorphism.
\end{proof}
Regardless of the perversity. The two cases of special interest here are the cases where $r=1$ and $r= \infty$.
The case $r=1$, the 1-formality, implies that the rational Malcev completion of $\pi_{1}(\I{p}{X})$ can be computed directly from the cohomology group $\HI{p}{1}{X}$, together with the cup product $\HI{p}{1}{X} \otimes \HI{p}{1}{X} \rightarrow \HI{p}{2}{X}$. We then say that $\pi_{1}(\I{p}{X})$ is 1-formal.
The case $r=\infty$ implies the formality of $\I{p}{X}$ in the usual sense, which in the cases where $\I{p}{X}$ is simply-connected or nilpotent implies that the rational homotopy groups $\pi_{i}(\I{p}{X})\otimes \mathbf{Q}$ can be directly computed from the cohomology ring $\HI{p}{\ast}{X}$. We note that formality implies 1-formality.
Suppose now $X \in \super$ with only normal isolated singularities, that is
\[
\HI{\infty}{k}{X}=H^{k}(\overline{X})=H^{k}(X)
\]
then by the Van-Kampen theorem and by definition \ref{def:intersectionspace2} for any perversity $\overline{p}$ we have
\[
\pi_{1}(X) = \pi_{1}(\I{p}{X}) = \pi_{1}(X_{reg}).
\]
Morevover, whether $\overline{p}=\overline{0}$ or $\overline{p} \neq \overline{0}$ we have the two following commutative diagrams.
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=3em, text height=2ex, text depth=0.25ex]
{ \HI{0}{1}{X} \otimes \HI{0}{1}{X} & \HI{0}{2}{X} \\
H^{1}(X_{reg}) \otimes H^{1}(X_{reg}) & H^{2}(X_{reg}) \\};
\path[->]
(m-1-1) edge node[auto] {$- \cup -$} (m-1-2);
\path[->]
(m-2-1) edge node[auto,swap] {$- \cup -$} (m-2-2);
\path[->]
(m-1-1) edge node[auto,swap] {=} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {=} (m-2-2);
\end{tikzpicture}
\]
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=3em, text height=2ex, text depth=0.25ex]
{ H^{1}(X) \otimes H^{1}(X) & H^{2}(X) \\
\HI{p}{1}{X} \otimes \HI{p}{1}{X} & \HI{p}{2}{X} \\};
\path[->]
(m-1-1) edge node[auto] {$- \cup -$} (m-1-2);
\path[->]
(m-2-1) edge node[auto,swap] {$- \cup -$} (m-2-2);
\path[->]
(m-1-1) edge node[auto,swap] {$\cong$} (m-2-1);
\path[injection]
(m-1-2) edge node[auto] {} (m-2-2);
\end{tikzpicture}
\]
Which means that if $X$ is 1-formal then we can compute the rational Malcev completion of $\pi_{1}(\I{p}{X})$ by computing the one from $\pi_{1}(X)$. It is a result from \cite{Arapura2015} that when considering normal projective varieties the fundamental group is always 1-formal, see also \cite[Corollary 3.8]{Chataur2015} for the isolated singularities case. We can then deduce the following result
\begin{propo}
Let $X \in \super$ with only normal isolated singularities. Then for any perversity $\overline{p}$ $\pi_{1}(\I{p}{X})$ is formal.
\end{propo}
We also highlight the case $r=\infty$.
\begin{cor}
Let $X \in \super$ with only isolated singularities. If $\I{\bullet}{X}$ is $(\overline{p},\infty)$-pure then $\I{s}{X}$ is formal for any $\overline{s} \leq \overline{p}$ in $\mathcal{P}^{op}$.
\end{cor}
\begin{remark}
The question of the purity of the weight filtration is also considered in intersection cohomology, where a similar result of "purity implies formality" exists \cite[corollary 3.13]{Chataur}. It must be pointed out that the purity of $X \in \super$ in intersection cohomology does not imply the purity of $\I{\bullet}{X}$. For example the Kummer surface of section \ref{subsec:Kummer}, it is a $\mathbf{Q}$-homology manifold and as such $\IH{p}{k}{X}$ is pure of weight $k$ for any perversities and then is intersection formal. This is not the case of the corresponding intersection space for the middle perversity $\I{1}{X}$ since $\gr{4}{W}{\HI{1}{3}{X}} \neq 0$.
Another and more involved example. It is a consequence of Gabber's purity theorem and the decomposition theorem of intersection homology (see \cite{Steenbrink1983}) that for projective varieties $X$ with isolated singularities and for the middle perversity, the weight filtration $W$ on $\IH{m}{k}{X}$ is pure of weight $k$ for all $k \geq 0$, this is not the case for the Calabi-Yau 3-folds treated in the last part as we see that the weight filtration $W$ on $\HI{m}{k}{X}$ isn't pure.
\end{remark}
\section{Formality of intersection spaces for 3-folds}
\label{sec:formality_3folds}
\subsection{Preparatory work}
Let $X$ be a complex projective algebraic 3-fold with isolated singularities and denote by $\Sigma= \lbrace \sigma_{1}, \dots, \sigma_{\nu} \rbrace$ the singular locus of $X$. Assume that there is a resolution of singularities $f \col \widetilde{X} \rightarrow X$ such that the exceptional divisor $D := f^{-1}(\Sigma)$ is smooth and the the link $L_{i}$ of $\sigma_{i}$ in $X$, for all $\sigma_{i} \in \Sigma$ is simply connected.
First we recall and collect the different properties we will need. We state them in the case of a space of complex dimension 3 but they are completely general and holds for any complex projective variety of complex dimension $n$ with only isolated singularities by replacing $3$ by $n$. The proofs can be found in \cite{Chataur}.
\begin{lem}
\label{lem:DP_ker_coker}
We have the following Poincaré duality isomorphisms for all $0 \leq s \leq 3$,
\[
\coker{\gamma^{3+s}} \cong (\ker j^{3-s})^{\vee} \quad \ker \gamma^{3+s} \cong (\coker{j^{3-s}})^{\vee}
\]
\end{lem}
Recall that since $\dim(\Sigma) =0$, the weight filtration on the cohomology of the link is semi-pure, meaning :
\begin{itemize}
\item the weights on $H^{k}(L)$ are less than or equal to $k$ for $k <3$,
\item the weights on $H^{k}(L)$ are greater or equal to $k+1$ for $k \geq 3$.
\end{itemize}
We than have the following results.
\begin{lem}
\label{lem:properties_1}
With the previous notations we have :
\begin{enumerate}
\item The map $j_{\sharp}^{k} \col H^{k-2}(D) \rightarrow H^{k}(D)$ is injective for $k \leq 3$ and surjective for $k \geq 3$.
\item The Gysin map $\gamma^{k} \col H^{k-2}(D) \rightarrow H^{k}(\widetilde{X})$ is injective for $k \leq 3$ and $\gamma^{6}$ is surjective.
\item The restriction morphism $j^{k} \col H^{k}(\widetilde{X}) \rightarrow H^{k}(D)$ is surjective for $k \geq 3$.
\end{enumerate}
\end{lem}
\begin{lem}
\label{lem:properties_2}
With the assumption on the links $L$, we have the following :
\begin{enumerate}
\item The map $j_{\sharp}^{2} \col H^{0}(D) \rightarrow H^{2}(D)$ in injective, the map $j_{\sharp}^{4} \col H^{2}(D) \rightarrow H^{4}(D)$ is surjective, $j_{\sharp}^{k} \col H^{k-2}(D) \rightarrow H^{k}(D)$ is an isomorphism for $k =1,3,5$.
\item The map $\gamma^{k} \col H^{k-2}(D) \rightarrow H^{k}(\widetilde{X})$ is injective for all $k \neq 4,6$ and $j^{k} \col H^{k}(\widetilde{X}) \rightarrow H^{k}(D)$ is surjective for all $k \neq 0,2$.
\end{enumerate}
\end{lem}
\begin{lem}
\label{lem:properties_3}
With the above assumptions we have the following :
\begin{enumerate}
\item $H^{k}(\widetilde{X}) \cong \ker j^{k} \oplus \im{\gamma^{k}}$ for $k=1,3,5$.
\item $\ker j^{2} \cap \im{\gamma^{2}} = 0$.
\end{enumerate}
\end{lem}
With the lemmas above the second term of the spectral sequences for the regular part and the links are given by
\[
\begin{array}{| c || c | c |}
\hline
\multicolumn{3}{|c|}{E_{2}^{r,s}(X_{reg})} \\
\hline
s=6 & \ker \gamma^{6} & 0 \\
\hline
s=5 & 0 & \coker{\gamma^{5}} \\
\hline
s=4 & \ker \gamma^{4} & \coker{\gamma^{4}} \\
\hline
s=3 & 0 & \coker{\gamma^{3}} \\
\hline
s=2 & 0 & \coker{\gamma^{2}} \\
\hline
s=1 & 0 & \coker{\gamma^{1}} \\
\hline
s=0 & 0 & H^{0}(\widetilde{X}) \\
\hline
\hline
& r=-1 & r=0 \\
\hline
\end{array}
\quad
\begin{array}{| c || c | c |}
\hline
\multicolumn{3}{|c|}{E_{2}^{r,s}(L)} \\
\hline
s=6 & H^{4}(D) & 0 \\
\hline
s=5 & 0 & 0 \\
\hline
s=4 & \ker j^{4}_{\sharp} & 0 \\
\hline
s=3 & 0 & 0 \\
\hline
s=2 & 0 & \coker{j^{2}_{\sharp}} \\
\hline
s=1 & 0 & 0 \\
\hline
s=0 & 0 & H^{0}(D) \\
\hline
\hline
& r=-1 & r=0 \\
\hline
\end{array}
\]
The computation of the cohomology of the intersection spaces involve a choice of complementary subspace $\coim{p}$, we detail here the choice we make.
\begin{itemize}
\item For the perversity $\overline{1}$, the map $j_{\sharp}^{2}$ is injective by lemma \ref{lem:properties_1}, we then have $\coim{1} = H^{0}(D)$ and $\coker{\gamma^{2}_{|\coim{1}}} = \coker{\gamma^{2}}$.
\item For the perversity $\overline{2}$, the map $j_{\sharp}^{3}$ is an isomorphism by lemma \ref{lem:properties_2}, we then also have $\coim{2} = H^{1}(D)$ and $\coker{\gamma^{3}_{|\coim{2}}} = \coker{\gamma^{3}}$.
\item For the perversity $\overline{3}$, there is no assumption on $j_{\sharp}^{4}$ and we chose a complementary subspace of $\ker j_{\sharp}^{4}$ which we denote by $\coim{3}$.
\item For the perversity $\overline{4}$, the map $j_{\sharp}^{5}$ is an isomorphism by lemma \ref{lem:properties_2}, we then also have $\coim{4} = H^{3}(D)$ and $\coker{\gamma^{5}_{|\coim{4}}} = \coker{\gamma^{5}}$.
\end{itemize}
Since the links of the singularities are simply connected five dimensional manifolds, by definition of the intersection spaces we have $\I{0}{X} \simeq \I{1}{X}$ and $\I{3}{X} \simeq \I{4}{X}$. Thus the second terms of the corresponding spectral sequences must be isomorphic, for now the corresponding second term for the associated spectral sequences are the following.
\[
\begin{array}{| c || c c c || c c c|}
\hline
& \multicolumn{3}{c||}{\EIbidg{2}{0}{r,s}{X}} & \multicolumn{3}{c|}{\EIbidg{2}{1}{r,s}{X}} \\
\hline
s = 6 & \ker \gamma^{6} & 0 & 0 & \ker \gamma^{6} & 0 & 0\\
\hline
s = 5 & 0 & \coker{\gamma^{5}} & 0 & 0 & \coker{\gamma^{5}} & 0 \\
\hline
s = 4 & \ker \gamma^{4} & \coker{\gamma^{4}} & 0 & \ker \gamma^{4} & \coker{\gamma^{4}} & 0\\
\hline
s = 3 & 0 & \coker{\gamma^{3}} & 0 & 0 & \coker{\gamma^{3}} & 0\\
\hline
s = 2 & 0 & \coker{\gamma^{2}} & 0 & 0 & \coker{\gamma^{2}} & 0\\
\hline
s=1 & 0 & \coker{\gamma^{1}} & 0 & 0 & \ker j^{1} & 0\\
\hline
\hline
s = 0 & 0 & H^{0}(\widetilde{X}) & 0 & 0 & H^{0}(\widetilde{X}) & 0\\
\hline
\hline
& r=-1 & r=0 & r=1 & r=1 & r=0 & r=1\\
\hline
\end{array}
\]
\[
\begin{array}{| c || c c c || c c c|}
\hline
& \multicolumn{3}{c||}{\EIbidg{2}{3}{r,s}{X}} & \multicolumn{3}{c|}{\EIbidg{2}{4}{r,s}{X}} \\
\hline
s = 6 & \ker \gamma^{6} & 0 & 0 & \ker \gamma^{6} & 0 & 0\\
\hline
s = 5 & 0 & \coker{\gamma^{5}} & 0 & 0 & \coker{\gamma^{5}} & 0 \\
\hline
s = 4 & 0 & \coker{\gamma^{4}_{|\coim{3}}} & 0 & 0 & \ker j^{4} & 0\\
\hline
s = 3 & 0 & \ker j^{3} & 0 & 0 & \ker j^{3} & 0\\
\hline
s = 2 & 0 & \ker j^{2} & \coker{j^{2}} & 0 & \ker j^{2} & \coker{j^{2}}\\
\hline
s=1 & 0 & \ker j^{1} & 0 & 0 & \ker j^{1} & 0\\
\hline
\hline
s = 0 & 0 & H^{0}(\widetilde{X}) & 0 & 0 & H^{0}(\widetilde{X}) & 0\\
\hline
\hline
& r=-1 & r=0 & r=1 & r=1 & r=0 & r=1\\
\hline
\end{array}
\]
We then need to show that $\EIbidg{2}{0}{r,s}{X} \cong \EIbidg{2}{1}{r,s}{X}$ and $\EIbidg{2}{3}{r,s}{X} \cong \EIbidg{2}{4}{r,s}{X}$. The first isomorphism is given by the isomorphism
\[
H^{1}(\widetilde{X}) \cong \ker j^{1} \oplus \im{\gamma^{1}}
\]
from the lemma \ref{lem:properties_3}, we then have $\coker{\gamma^{1}} \cong \ker j^{1}$.
For the second isomorphism we need to show that
\[
\coker{\gamma^{4}_{|\coim{3}}} \cong \ker j^{4}.
\]
Which is given by the following lemma
\begin{lem}
\label{lem:properties_4}
We have the following isomorphism
\[
H^{4}(\widetilde{X}) \cong \ker j^{4} \oplus \im{\gamma^{4}_{|\coim{3}}}.
\]
\end{lem}
\begin{proof}
Denote by $(\ker j^{4})^{\perp}$ a complementary subspace of $\ker j^{4} \subset H^{4}(\widetilde{X})$. The maps $j_{\sharp}^{4}$ and $j^{4}$ are surjective by lemma \ref{lem:properties_1}. We then have the following commutative diagram
\[
\begin{tikzpicture}[description/.style={fill=white,inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ H^{2}(D) \cong \ker j_{\sharp}^{4} \oplus \coim{3} & H^{4}(\widetilde{X}) \cong\ker j^{4} \oplus (\ker j^{4})^{\perp} \\
& H^{4}(D) \\};
\path[->]
(m-1-1) edge node[auto] {$\gamma^{4}$} (m-1-2);
\path[->>]
(m-1-2) edge node[auto] {$j^{4}$} (m-2-2);
\path[->>]
(m-1-1) edge node[description] {$j_{\sharp}^{4}$} (m-2-2);
\end{tikzpicture}
\]
By definition of $\coim{3}$ we have $\gamma^{4}_{|} \col \coim{3} \rightarrow (\ker j^{4})^{\perp}$. The commutative diagram restricts then to the following commutative diagram where the restrictions $j_{\sharp |}^{4}$ and $j^{4}_{|}$ are isomorphisms. Which finishes the proof.
\[
\begin{tikzpicture}[description/.style={fill=white,inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \coim{3} & (\ker j^{4})^{\perp} \\
& H^{4}(D) \\};
\path[->]
(m-1-1) edge node[auto] {$\gamma^{4}_{|}$} (m-1-2);
\path[->]
(m-1-2) edge node[auto] {$j^{4}_{|}$} (m-2-2);
\path[->]
(m-1-1) edge node[description] {$j_{\sharp |}^{4}$} (m-2-2);
\end{tikzpicture}
\]
\end{proof}
The second terms of the spectral sequences of $\EIbidg{2}{p}{r,s}{X}$ for $\overline{p} \in \lbrace \overline{0}, \overline{2}, \overline{4} \rbrace$ are finally.
\[
\hspace{-4.25em}
\begin{array}{| c || c c c || c c c|| c c c |}
\hline
& \multicolumn{3}{c||}{\EIbidg{2}{0}{r,s}{X}} & \multicolumn{3}{c||}{\EIbidg{2}{2}{r,s}{X}} & \multicolumn{3}{c|}{\EIbidg{2}{4}{r,s}{X}} \\
\hline
s = 6 & \ker \gamma^{6} & 0 & 0 & \ker \gamma^{6} & 0 & 0 & \ker \gamma^{6} & 0 & 0 \\
\hline
s = 5 & 0 & \coker{\gamma^{5}} & 0 & 0 & \coker{\gamma^{5}} & 0 & 0 & \coker{\gamma^{5}} & 0\\
\hline
s = 4 & \ker \gamma^{4} & \coker{\gamma^{4}} & 0 & \ker \gamma^{4} & \coker{\gamma^{4}} & 0 & 0 & \ker j^{4} & 0 \\
\hline
s = 3 & 0 & \coker{\gamma^{3}} & 0 & 0 & \coker{\gamma^{3}} & 0 & 0 & \ker j^{3} & 0\\
\hline
s = 2 & 0 & \coker{\gamma^{2}} & 0 & 0 & \ker j^{2} & \coker{j^{2}} & 0 & \ker j^{2} & \coker{j^{2}}\\
\hline
s=1 & 0 & \coker{\gamma^{1}} & 0 & 0 & \ker j^{1} & 0 & 0 & \ker j^{1} & 0 \\
\hline
\hline
s = 0 & 0 & H^{0}(\widetilde{X}) & 0 & 0 & H^{0}(\widetilde{X}) & 0 & 0 & H^{0}(\widetilde{X}) & 0 \\
\hline
\hline
& r=-1 & r=0 & r=1 & r=-1 & r=0 & r=1 & r=-1 & r=0 & r=1\\
\hline
\end{array}
\]
We are now ready to state the following theorem.
\subsection{Statement and proof}
In \cite[theorem E]{Chataur2012} it is proved that any nodal hypersurface $X$ in $\mathbf{C}P^{4}$ is GM-intersection-formal, meaning that their perverse rational models $\IA{\bullet}{X}$ is quasi-isomorphic to their intersection cohomology algebras $\IH{\bullet}{\ast}{X}$. This result is extended in \cite[theorem 4.5]{Chataur} to the case of complex projective varieties of dimension $n$ with only isolated singularities and $(n-2)$-connected links using mixed Hodge structures. We show, using the same ideas, that for $X$ a complex projective algebraic 3-fold with isolated singularities and simply connected links, the intersection spaces are formal topological spaces.
\begin{thm}
\label{thm:3fold_formal}
Let $X$ be a complex projective algebraic 3-fold with isolated singularities and denote by $\Sigma = \lbrace \sigma_{1}, \dots, \sigma_{\nu}\rbrace$ the singular locus of $X$. Assume that there is a resolution of singularities $f \col \widetilde{X} \rightarrow X$ such that the exceptional divisor $D := f^{-1}(\Sigma)$ is smooth and the link $L_{i}$ of $\sigma_{i}$ in $X$, for all $\sigma_{i} \in \Sigma$, is simply connected. Then $\I{p}{X}$ is formal over $\mathbf{C}$ for all $\overline{p} \in \lbrace \overline{\infty}, \overline{0}, \dots, \overline{4} \rbrace$.
\end{thm}
By theorem \ref{thm:AI_to_EI1} there is a string of quasi-isomorphisms of coperverse cdga's from $\AI{\bullet}{X}\otimes \mathbf{C}$ to $\EI{1}{\bullet}{X}\otimes \mathbf{C}$. Moreover we have $\EIbidg{2}{\bullet}{\ast,\ast}{X} \cong \HI{\bullet}{\ast}{X}$. We follow this pattern
\begin{enumerate}
\item We define a bigraded differential algebra $(\FI{p}{r,s}{X}, \diff{p}{r,s})$ step by step for the perversities $\overline{4}$, $\overline{2}$ and $\overline{0}$.
\begin{itemize}
\item When needed, we then define the poset map $\pomap{p}{q} \col \FI{p}{r,s}{X} \rightarrow \FI{q}{r,s}{X}$ and show its compatibility with the product and the differential.
\end{itemize}
\item We define the quasi-isomorphisms
\[
(\EIbidg{1}{p}{r,s}{X},\difb{p}{r,s}) \overset{\psip{p}{r,s}}{\longleftarrow} (\FI{p}{r,s}{X},\diff{p}{r,s}) \overset{\phip{p}{r,s}}{\longrightarrow} (\EIbidg{2}{p}{r,s}{X},0)
\]
and check their compatibility with the products and differentials.
\begin{itemize}
\item When needed, we then check the compatibility of the maps $\psip{\bullet}{\ast,\ast}$ and $\phip{\bullet}{\ast,\ast}$ with the poset map $\pomap{p}{q} \col \FI{p}{r,s}{X} \rightarrow \FI{q}{r,s}{X}$.
\end{itemize}
\end{enumerate}
\subsubsection{The top perversity}
We begin with the top perversity $\overline{t} = \overline{4}$. We define the bigraded differential algebra $(\FI{4}{r,s}{X}, \diff{4}{r,s})$.
\[
\begin{array}{| c || c c c |}
\hline
s = 6 & H^{4}(D) & H^{6}(\widetilde{X}) & 0 \\
\hline
s = 5 & H^{3}(D) & H^{5}(\widetilde{X})& 0 \\
\hline
s = 4 & 0 & \ker j^{4} & 0 \\
\hline
s = 3 & 0 & \ker j^{3} & 0 \\
\hline
s = 2 & 0 & \ker j^{2} & \D{(\ker \gamma^{4})} \otimes dt \\
\hline
s=1 & 0 & \ker j^{1} & 0 \\
\hline
\hline
s = 0 & 0 & H^{0}(\widetilde{X}) & 0 \\
\hline
\hline
\FI{4}{r,s}{X} & r=-1 & r=0 & r=1 \\
\hline
\end{array}
\]
The only non-trivial differentials are $\diff{4}{-1,s} \col H^{s-2}(D) \rightarrow H^{s}(\widetilde{X})$ given by $\diff{4}{-1,s} = \gamma^{s}$ for $s=5,6$. The algebra structure is defined by $\FI{4}{0,s}{X} \times \FI{4}{0,s'}{X} \rightarrow \FI{4}{0,s+s'}{X}$.
Let's now define the map $\psip{4}{\ast,\ast} \col \FI{4}{\ast,\ast}{X} \rightarrow \EIbidg{1}{4}{\ast,\ast}{X}$. Recall that we have the following first term for the weight spectral sequence.
\noindent\makebox[\textwidth]{
\(
\begin{array}{| c || c c c c c |}
\hline
s \geq 5 & H^{s-2}(D)\otimes \mathbf{Q}[t] & \rightarrow & \mathcal{I}^{s}_{0} \oplus H^{s-2}(D)\otimes \mathbf{Q}[t]dt & \rightarrow & H^{s}(D)\otimes \mathbf{Q}[t]dt \\
\hline
1 \leq s \leq 4 & H^{s-2}(D)\otimes \mathbf{Q}[t]t & \rightarrow & \mathcal{I}^{s}_{1} \oplus H^{s-2}(D)\otimes \mathbf{Q}[t]dt & \rightarrow & H^{s}(D)\otimes \mathbf{Q}[t]dt \\
\hline
\hline
s = 0 & 0 & & \mathcal{I}^{0}_{0} & \rightarrow & H^{0}(D)\otimes \mathbf{Q}[t]dt \\
\hline
\hline
\EIbidg{1}{4}{r,s}{X} & r=-1 & & r=0 & & r=1 \\
\hline
\end{array}
\)}
For $r=s=0$, the map $\psip{4}{0,0}$ is the identity map.For $r=0$, $s>0$, the map $\psip{4}{0,s} \col \FI{4}{0,s}{X} \rightarrow \EIbidg{1}{4}{0,s}{X}$ is defined to be
\[
\psip{4}{0,s}(x) := (x, j^{s}(x)).
\]
For $r=-1$, $\psip{4}{-1,s}$ is defined to be the canonical inclusion.
By lemma \ref{lem:DP_ker_coker} we have $\D{(\ker \gamma^{4})} \cong \coker{j^{2}} \subset H^{2}(D)$, we then define $\psip{4}{1,2}$ to be the injective map
\[
\psip{4}{1,2} \col \D{(\ker \gamma^{4})} \otimes dt \longrightarrow \EIbidg{1}{4}{1,2}{X} = H^{2}(D)\otimes \mathbf{Q}[t]dt.
\]
By definition $\mathcal{I}^{s}_{k}$, $k \in \lbrace 0,1 \rbrace$, is the vector space given by the following pullback square.
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \mathcal{I}^{s}_{k} & H^{s}(D)\otimes \mathbf{Q}[t]t^{k} \\
H^{s}(\widetilde{X}) & H^{s}(D) \\};
\path[->]
(m-1-1) edge node[auto] {} (m-1-2);
\path[->]
(m-2-1) edge node[auto] {$j^{s}$} (m-2-2);
\path[->]
(m-1-1) edge node[above=1em, right=1em] {$\ulcorner$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$\delta_{1}$} (m-2-2);
\end{tikzpicture}
\]
We have $\mathcal{I}^{s}_{1} \subset \mathcal{I}^{s}_{0}$, the map $\psip{4}{0,s}(x) := (x, j^{s}(x))$ is then compatible with the algebra structure of $\FI{4}{\ast,\ast}{X}$. The commutativity of the following diagrams
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \FI{4}{-1,s}{X} & \FI{4}{0,s}{X} \\
\EIbidg{1}{4}{-1,s}{X} & \EIbidg{1}{4}{0,s}{X} \\};
\path[->]
(m-1-1) edge node[auto] {$\diff{4}{-1,s} = \gamma^{s}$} (m-1-2);
\path[->]
(m-2-1) edge node[auto,swap] {$\difb{4}{-1,s}$} (m-2-2);
\path[injection]
(m-1-1) edge node[auto,swap] {$\psip{4}{-1,s}$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$\psip{4}{0,s}$} (m-2-2);
\end{tikzpicture}
\quad
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \ker j^{2} & \D{(\ker \gamma^{4})} \otimes dt \\
\EIbidg{1}{4}{0,2}{X} & \EIbidg{1}{4}{1,2}{X} \\};
\path[->]
(m-1-1) edge node[auto] {$0$} (m-1-2);
\path[->]
(m-2-1) edge node[auto,swap] {$\difb{4}{-1,s}$} (m-2-2);
\path[->]
(m-1-1) edge node[auto,swap] {$\psip{4}{0,2}$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$\psip{4}{1,2}$} (m-2-2);
\end{tikzpicture}
\]
concludes that we have a quasi-isomorphism $\psip{4}{\ast,\ast} \col \FI{4}{\ast,\ast}{X} \rightarrow \EIbidg{1}{4}{\ast,\ast}{X}$.
We now detail the map $\phip{4}{\ast,\ast} \col \FI{4}{\ast,\ast}{X} \rightarrow \EIbidg{2}{4}{\ast,\ast}{X}$.
For $r=-1$, $\phip{4}{-1,s}$ is non zero only for $s=6$ where it is the projection $H^{4}(D) \twoheadrightarrow \ker \gamma^{6}$.
For $r=0$, since $\FI{4}{0,s}{X} = \ker \difb{4}{0,s}$ for all $s$, we define the map $\phip{4}{0,s}$ to be the surjection $\phip{4}{0,s} \col \ker \difb{4}{0,s} \twoheadrightarrow \EIbidg{2}{4}{0,s}{X}$.
For $r=1$, the assignation $\D{(\ker \gamma^{4})} \otimes dt \mapsto \coker{j^{2}}$ defines $\phip{4}{1,2}$ and $\phip{4}{1,s}$ is zero for any other $s$.
Since we have $\ker \difb{4}{0,s} \times \ker \difb{4}{0,s'} \rightarrow \ker \difb{4}{0,s+s'}$, the map $\phip{4}{\ast,\ast}$ is compatible with the algebra structure of $\FI{4}{\ast,\ast}{X}$.
The map $\phip{4}{\ast,\ast}$ is also compatible with the two non zero differentials of $\FI{4}{\ast,\ast}{X}$ since the two following diagrams are commutative.
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ H^{4}(D) & H^{6}(\widetilde{X}) \\
\ker \gamma^{6} & 0 \\};
\path[->]
(m-1-1) edge node[auto] {$\gamma^{6}$} (m-1-2);
\path[->]
(m-2-1) edge (m-2-2);
\path[->>]
(m-1-1) edge node[auto,swap] {$\phip{4}{-1,6}$} (m-2-1);
\path[->]
(m-1-2) edge (m-2-2);
\end{tikzpicture}
\quad
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ H^{3}(D) & H^{5}(\widetilde{X}) \\
0 & \EIbidg{2}{4}{0,5}{X} = \coker{\gamma^{5}} \\};
\path[injection]
(m-1-1) edge node[auto] {$\gamma^{5}$} (m-1-2);
\path[->>]
(m-1-2) edge node[auto] {$\phip{4}{0,5}$} (m-2-2);
\path[->]
(m-2-1) edge (m-2-2);
\path[->]
(m-1-1) edge (m-2-1);
\end{tikzpicture}
\]
We then have a quasi-isomorphism of algebras
\[
(\EIbidg{1}{4}{r,s}{X},\difb{4}{r,s}) \overset{\psip{4}{r,s}}{\longleftarrow} (\FI{4}{r,s}{X},\diff{4}{r,s}) \overset{\phip{4}{r,s}}{\longrightarrow} (\EIbidg{2}{4}{r,s}{X},0).
\]
\subsubsection{The middle perversity}
We define the bigraded differential algebra $(\FI{2}{r,s}{X},\diff{2}{r,s})$ as the sub-algebra of $(\EI{1}{2}{X},\difb{2}{r,s})$ given by
\[
\hspace{-1em}
\begin{array}{| c || c c c |}
\hline
s = 6 & H^{4}(D) & H^{6}(\widetilde{X}) & 0 \\
\hline
s = 5 & H^{3}(D) & H^{5}(\widetilde{X})& 0 \\
\hline
s = 4 & H^{2}(D) & H^{4}(\widetilde{X})& 0 \\
\hline
s = 3 & 0 & \ker j^{3} & 0 \\
\hline
s = 2 & 0 & \ker j^{2} & \D{(\ker \gamma^{4})} \otimes dt \\
\hline
s=1 & 0 & \ker j^{1} & 0 \\
\hline
\hline
s = 0 & 0 & H^{0}(\widetilde{X}) & 0 \\
\hline
\hline
\FI{2}{r,s}{X} & r=-1 & r=0 & r=1 \\
\hline
\end{array}
\]
Where $(\EI{1}{2}{X},\difb{2}{r,s})$ is given by
\noindent\makebox[\textwidth]{
\(
\begin{array}{| c || c c c c c |}
\hline
s \geq 3 & H^{s-2}(D)\otimes \mathbf{Q}[t] & \rightarrow & \mathcal{I}^{s}_{0} \oplus H^{s-2}(D)\otimes \mathbf{Q}[t]dt & \rightarrow & H^{s}(D)\otimes \mathbf{Q}[t]dt \\
\hline
1 \leq s \leq 2 & H^{s-2}(D)\otimes \mathbf{Q}[t]t & \rightarrow & \mathcal{I}^{s}_{1} \oplus H^{s-2}(D)\otimes \mathbf{Q}[t]dt & \rightarrow & H^{s}(D)\otimes \mathbf{Q}[t]dt \\
\hline
\hline
s = 0 & 0 & & \mathcal{I}^{0}_{0} & \rightarrow & H^{0}(D)\otimes \mathbf{Q}[t]dt \\
\hline
\hline
\EIbidg{1}{2}{r,s}{X} & r=-1 & & r=0 & & r=1 \\
\hline
\end{array}
\)}
Compared to $\FI{4}{\ast,\ast}{X}$, we added $H^{2}(D)$ in bidegree $(-1,4)$ and replaced $\ker j^{4}$ by $H^{4}(\widetilde{X})$ in bidegree $(0,4)$, both are related by a new non-trivial differential $\diff{2}{-1,4} = \gamma^{4}$.
The algebra structure is still non-trivial only for $r=0$, with
\[
\FI{2}{0,s}{X} \times \FI{2}{0,s'}{X} \rightarrow \FI{2}{0,s+s'}{X}.
\]
The map $\pomap{4}{2} \col \FI{4}{\ast,\ast}{X} \rightarrow \FI{2}{\ast,\ast}{X}$ is then the canonical inclusion, which is clearly compatible with the differential and the algebra structure.
To construct $\psip{2}{\ast,\ast} \col \FI{2}{\ast,\ast}{X} \rightarrow \EIbidg{1}{2}{\ast,\ast}{X}$, we extend $\psip{4}{\ast,\ast}$, meaning that $\psip{2}{-1,s}$ is the inclusion, $\psip{2}{0,s}(x) = (x, j^{s}(x))$ and $\psip{2}{1,s} = \psip{4}{1,s}$. The algebra structure is preserved by $\psip{2}{0,s}$ and the following diagram commutes
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ H^{2}(D) & H^{4}(\widetilde{X}) \\
\EIbidg{1}{2}{-1,4}{X} & \EIbidg{1}{2}{0,4}{X} \\};
\path[->]
(m-1-1) edge node[auto] {$\diff{2}{-1,4} = \gamma^{4}$} (m-1-2);
\path[->]
(m-2-1) edge node[auto,swap] {$\difb{2}{-1,4}$} (m-2-2);
\path[injection]
(m-1-1) edge node[auto,swap] {$\psip{2}{-1,4}$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$\psip{2}{0,4}$} (m-2-2);
\end{tikzpicture}
\]
The rest being the same as for the top perversity, we have a quasi-isomorphism
\[
\psip{2}{\ast,\ast} \col \FI{2}{\ast,\ast}{X} \longrightarrow \EIbidg{1}{2}{\ast,\ast}{X}.
\]
We now construct $\phip{2}{\ast,\ast} \col \FI{2}{\ast,\ast}{X} \rightarrow \EIbidg{2}{2}{\ast,\ast}{X}$.
First of all nothing changes for $r=1$ and $\phip{2}{1,s}=\phip{4}{1,s}$.
For $r=-1$, $\phip{2}{-1,s}$ is non zero only for $s=4,6$ where it is the projection $H^{s-2}(D) \twoheadrightarrow \ker \gamma^{s}$.
For $r=0$, since $\FI{2}{0,s}{X} = \ker \difb{2}{0,s}$ for all $s \neq 3$, we define the map $\phip{2}{0,s}$ to be the surjection $\phip{2}{0,s} \col \ker \difb{2}{0,s} \twoheadrightarrow \EIbidg{2}{2}{0,s}{X}$. For $s=3$, by lemma \ref{lem:properties_3} we have $\ker j^{3} \cong \coker{\gamma^{3}}$, this isomorphism defines $\phip{2}{0,3}$.
For $s=4,6$ or $s=5$, the following diagrams commute
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ H^{s-2}(D) & H^{s}(\widetilde{X}) \\
\ker \gamma^{s} & \coker{\gamma^{s}} \\};
\path[->]
(m-1-1) edge node[auto] {$\gamma^{s}$} (m-1-2);
\path[->]
(m-2-1) edge node[auto,swap] {$0$} (m-2-2);
\path[->>]
(m-1-1) edge node[auto,swap] {$\phip{2}{-1,s}$} (m-2-1);
\path[->>]
(m-1-2) edge node[auto] {$\phip{2}{0,s}$} (m-2-2);
\end{tikzpicture}
\quad
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ H^{3}(D) & H^{5}(\widetilde{X}) \\
0 & \EIbidg{2}{2}{0,5}{X} = \coker{\gamma^{5}} \\};
\path[injection]
(m-1-1) edge node[auto] {$\gamma^{5}$} (m-1-2);
\path[->>]
(m-1-2) edge node[auto] {$\phip{2}{0,5}$} (m-2-2);
\path[->]
(m-2-1) edge (m-2-2);
\path[->]
(m-1-1) edge (m-2-1);
\end{tikzpicture}
\]
So $\phip{2}{\ast,\ast}$ is compatible with the differential.
To see its compatibility with the algebra structure of $\FI{2}{\ast,\ast}{X}$ we have to check the commutativity of the following diagram
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \ker j^{1} \times \ker j^{2} & \ker j^{3} \\
\coker{\gamma^{1}} \times \coker{\gamma^{2}} & \coker{\gamma^{3}} \\};
\path[->]
(m-1-1) edge node[auto] {} (m-1-2);
\path[->]
(m-2-1) edge node[auto,swap] {} (m-2-2);
\path[->]
(m-1-1) edge node[auto,swap] {} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$\cong$} (m-2-2);
\end{tikzpicture}
\]
We then have a quasi-isomorphism of algebras
\[
(\EIbidg{1}{2}{r,s}{X},\difb{2}{r,s}) \overset{\psip{2}{r,s}}{\longleftarrow} (\FI{2}{r,s}{X},\diff{2}{r,s}) \overset{\phip{2}{r,s}}{\longrightarrow} (\EIbidg{2}{2}{r,s}{X},0).
\]
We now check the commutativity of the following diagram
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ (\EIbidg{1}{4}{r,s}{X},\difb{4}{r,s}) & (\FI{4}{r,s}{X},\diff{4}{r,s}) & (\EIbidg{2}{4}{r,s}{X},0) \\
(\EIbidg{1}{2}{r,s}{X},\difb{2}{r,s}) & (\FI{2}{r,s}{X},\diff{2}{r,s}) & (\EIbidg{2}{2}{r,s}{X},0)\\};
\path[->]
(m-1-2) edge node[auto,swap] {$\psip{4}{r,s}$} (m-1-1);
\path[->]
(m-1-2) edge node[auto] {$\phip{4}{r,s}$} (m-1-3);
\path[->]
(m-2-2) edge node[auto] {$\psip{2}{r,s}$} (m-2-1);
\path[->]
(m-2-2) edge node[auto,swap] {$\phip{2}{r,s}$} (m-2-3);
\path[->]
(m-1-1) edge node[auto,swap] {$EI_{1}(\pomap{4}{2})$} (m-2-1);
\path[injection]
(m-1-2) edge node[auto] {$\pomap{4}{2}$} (m-2-2);
\path[->]
(m-1-3) edge node[auto] {$EI_{2}(\pomap{4}{2})$} (m-2-3);
\end{tikzpicture}
\]
The only differences between $\EIbidg{i}{4}{r,s}{X}$ and $\EIbidg{i}{2}{r,s}{X}$, $i=1,2$, arise for $s=3,4$. We then only check these cases.
The only square that does not trivially commutes for $s=3$ is the following
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \mathcal{I}^{3}_{1} \oplus H^{1}(D)\otimes \mathbf{Q}[t]dt & \ker j^{3} & \ker j^{3} \\
\mathcal{I}^{3}_{0} \oplus H^{1}(D)\otimes \mathbf{Q}[t]dt & \ker j^{3} & \coker{\gamma^{3}}\\};
\path[->]
(m-1-2) edge node[auto,swap] {$\psip{4}{0,3}$} (m-1-1);
\path[->]
(m-1-2) edge node[auto] {$\phip{4}{0,3}$} (m-1-3);
\path[->]
(m-2-2) edge node[auto] {$\psip{2}{0,3}$} (m-2-1);
\path[->]
(m-2-2) edge node[auto,swap] {$\phip{2}{0,3}$} (m-2-3);
\path[injection]
(m-1-1) edge node[auto,swap] {$EI_{1}(\pomap{4}{2})$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$\pomap{4}{2}$} node[auto,swap] {=} (m-2-2);
\path[->]
(m-1-3) edge node[auto] {$EI_{2}(\pomap{4}{2})$} (m-2-3);
\end{tikzpicture}
\]
The left hand square commutes because $\im{\psip{4}{0,3}} \subset \mathcal{I}^{3}_{1}$, $\im{\psip{2}{0,3}} \subset \mathcal{I}^{3}_{0}$ and the fact that $\mathcal{I}^{3}_{1} \subset \mathcal{I}^{3}_{0}$. The right hand square commutes because of the isomorphism $\ker j^{3} \cong \coker{\gamma^{3}}$.
For $s=4$, the only square that does not trivially commutes is the following
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \mathcal{I}^{4}_{1} \oplus H^{2}(D)\otimes \mathbf{Q}[t]dt & \ker j^{4} & \ker j^{4} \\
\mathcal{I}^{4}_{0} \oplus H^{2}(D)\otimes \mathbf{Q}[t]dt & H^{4}(\widetilde{X}) & \coker{\gamma^{4}}\\};
\path[->]
(m-1-2) edge node[auto,swap] {$\psip{4}{0,4}$} (m-1-1);
\path[->]
(m-1-2) edge node[auto] {$\phip{4}{0,4}$} (m-1-3);
\path[->]
(m-2-2) edge node[auto] {$\psip{2}{0,4}$} (m-2-1);
\path[->]
(m-2-2) edge node[auto,swap] {$\phip{2}{0,4}$} (m-2-3);
\path[injection]
(m-1-1) edge node[auto,swap] {$EI_{1}(\pomap{4}{2})$} (m-2-1);
\path[injection]
(m-1-2) edge node[auto] {$\pomap{4}{2}$} (m-2-2);
\path[->]
(m-1-3) edge node[auto] {$EI_{2}(\pomap{4}{2})$} (m-2-3);
\end{tikzpicture}
\]
The left hand square commutes for the same reason that for $s=3$. We then consider the right hand square. By lemma \ref{lem:properties_4} we have $H^{4}(\widetilde{X}) \cong \ker j^{4} \oplus \im{\gamma^{4}_{|\coim{3}}}$, moreover we have $\im{\gamma^{4}_{|\coim{3}}} \subset \im{\gamma^{4}}$, this implies that
\[
\ker j^{4} \cap \im{\gamma^{4}} \neq \lbrace 0 \rbrace.
\]
We may then find a direct sum decomposition
\[
\ker j^{4} = (\ker j^{4} \cap \im{\gamma^{4}}) \oplus C
\]
and defines a map $\ker j^{4} \rightarrow C$ by projection on the second summand. We then have
\[
H^{4}(\widetilde{X}) \cong (\ker j^{4} \cap \im{\gamma^{4}}) \oplus C \oplus \im{\gamma^{4}_{|\coim{3}}},
\]
the maps $EI_{2}(\pomap{4}{2})$ and $\phip{2}{0,4}$ then send the summand $(\ker j^{4} \cap \im{\gamma^{4}}) \oplus \im{\gamma^{4}_{|\coim{3}}}$ to zero and $C$ to its class in $\coker{\gamma^{4}}$. Which makes the right hand square commute.
\subsubsection{The zero perversity}
We define the bigraded differential algebra $(\FI{0}{r,s}{X},\diff{0}{r,s})$ as the sub-algebra of $(\EI{1}{0}{X},\difb{0}{r,s})$ given by
\[
\begin{array}{| c || c c c |}
\hline
s = 6 & H^{4}(D) & H^{6}(\widetilde{X}) & 0 \\
\hline
s = 5 & H^{3}(D) & H^{5}(\widetilde{X})& 0 \\
\hline
s = 4 & H^{2}(D) & H^{4}(\widetilde{X})& 0 \\
\hline
s = 3 & 0 & \ker j^{3} & 0 \\
\hline
s = 2 & 0 & \D{(\ker j^{4})} \oplus \D{(\ker \gamma^{4})} \otimes t & \D{(\ker \gamma^{4})} \otimes dt \\
\hline
s=1 & 0 & \ker j^{1} & 0 \\
\hline
\hline
s = 0 & 0 & H^{0}(\widetilde{X}) & 0 \\
\hline
\hline
\FI{0}{r,s}{X} & r=-1 & r=0 & r=1 \\
\hline
\end{array}
\]
Where $(\EI{1}{0}{X},\difb{0}{r,s})$ is given by
\noindent\makebox[\textwidth]{
\(
\begin{array}{| c || c c c c c |}
\hline
s \geq 1 & H^{s-2}(D)\otimes \mathbf{Q}[t] & \rightarrow & \mathcal{I}^{s}_{0} \oplus H^{s-2}(D)\otimes \mathbf{Q}[t]dt & \rightarrow & H^{s}(D)\otimes \mathbf{Q}[t]dt \\
\hline
\hline
s = 0 & 0 & & \mathcal{I}^{0}_{0} & \rightarrow & H^{0}(D)\otimes \mathbf{Q}[t]dt \\
\hline
\hline
\EIbidg{1}{0}{r,s}{X} & r=-1 & & r=0 & & r=1 \\
\hline
\end{array}
\)}
Compared to $\FI{2}{\ast,\ast}{X}$, we added $\D{(\ker \gamma^{4})} \otimes t$ and replaced $\ker j^{2}$ by $\D{(\ker j^{4})}$ in bidegree $(0,2)$. There is also a new differential $\diff{0}{0,2} \col \D{(\ker \gamma^{4})} \otimes t \rightarrow \D{(\ker \gamma^{4})} \otimes dt$ which is differentiation with respect to $t$.
The algebra structure is non trivial only for $r=0$ where we have
\[
\begin{cases}
(\D{(\ker \gamma^{4})} \otimes t) \times \FI{0}{0,s}{X} \longrightarrow 0 & \forall \, s, \\
\FI{0}{0,s}{X} \times \FI{0}{0,s'}{X} \longrightarrow \FI{0}{0,s+s'}{X} & \text{otherwise.} \\
\end{cases}
\]
We now define $\pomap{2}{0} \col \FI{2}{\ast,\ast}{X} \rightarrow \FI{2}{\ast,\ast}{X}$. For $s \geq 3$, there is no changes and $\pomap{2}{0}$ is the identity, same if $s=0,1$. For $s=2$, by lemma \ref{lem:properties_3} we have $\ker j^{2} \cap \im{\gamma^{2}} = 0$ so we have the inclusion $\ker j^{2} \rightarrow \D{(\ker j^{4})}$. The map $\pomap{2}{0}$ is then an inclusion and is compatible with the differential and the algebra structure.
We now construct $\psip{0}{\ast,\ast} \col \FI{0}{\ast,\ast}{X} \rightarrow \EIbidg{1}{0}{\ast,\ast}{X}$. Since we have $\D{(\ker j^{4})} \oplus \D{(\ker \gamma^{4})} \otimes t \subset \mathcal{I}^{2}_{0}$ there is no difference between $\psip{0}{\ast,\ast}$ and $\psip{2}{\ast,\ast}$ and the definition is the same. We then have a quasi-isomorphism
\[
\psip{0}{\ast,\ast} \col \FI{0}{\ast,\ast}{X} \rightarrow \EIbidg{1}{0}{\ast,\ast}{X}.
\]
We define $\phip{0}{\ast,\ast} \col \FI{0}{\ast,\ast}{X} \rightarrow \EIbidg{2}{0}{\ast,\ast}{X}$, for $s \geq 3$ there is no difference with the middle perversity. If $s=2$ then we define $\phip{0}{0,2}$ by $\D{(\ker j^{4})} \mapsto \coker{\gamma^{2}}$ and $ \D{(\ker \gamma^{4})} \mapsto 0$, we then have the following commutative diagram.
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \D{(\ker j^{4})} \oplus \D{(\ker \gamma^{4})} \otimes t & \D{(\ker \gamma^{4})} \otimes dt \\
\EIbidg{2}{0}{0,2}{X} = \coker{\gamma^{2}} & 0 \\};
\path[->]
(m-1-1) edge node[auto] {$\diff{0}{0,2}$} (m-1-2);
\path[->]
(m-2-1) edge node[auto,swap] {} (m-2-2);
\path[->]
(m-1-1) edge node[auto,swap] {$\phip{0}{0,2}$} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {} (m-2-2);
\end{tikzpicture}
\]
If $s=1$, the isomorphism $\ker j^{1} \cong \coker{\gamma^{1}}$ defines $\phip{0}{0,1}$.
We then have a quasi-isomorphism of algebras
\[
(\EIbidg{1}{0}{r,s}{X},\difb{0}{r,s}) \overset{\psip{0}{r,s}}{\longleftarrow} (\FI{0}{r,s}{X},\diff{0}{r,s}) \overset{\phip{0}{r,s}}{\longrightarrow} (\EIbidg{2}{0}{r,s}{X},0).
\]
We now check the commutativity of the following diagram
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ (\EIbidg{1}{2}{r,s}{X},\difb{2}{r,s}) & (\FI{2}{r,s}{X},\diff{2}{r,s}) & (\EIbidg{2}{2}{r,s}{X},0) \\
(\EIbidg{1}{0}{r,s}{X},\difb{0}{r,s}) & (\FI{0}{r,s}{X},\diff{0}{r,s}) & (\EIbidg{2}{0}{r,s}{X},0)\\};
\path[->]
(m-1-2) edge node[auto,swap] {$\psip{2}{r,s}$} (m-1-1);
\path[->]
(m-1-2) edge node[auto] {$\phip{2}{r,s}$} (m-1-3);
\path[->]
(m-2-2) edge node[auto] {$\psip{0}{r,s}$} (m-2-1);
\path[->]
(m-2-2) edge node[auto,swap] {$\phip{0}{r,s}$} (m-2-3);
\path[->]
(m-1-1) edge node[auto,swap] {$EI_{1}(\pomap{2}{0})$} (m-2-1);
\path[injection]
(m-1-2) edge node[auto] {$\pomap{2}{0}$} (m-2-2);
\path[->]
(m-1-3) edge node[auto] {$EI_{2}(\pomap{2}{0})$} (m-2-3);
\end{tikzpicture}
\]
The only differences between $\EIbidg{i}{2}{r,s}{X}$ and $\EIbidg{i}{0}{r,s}{X}$, $i=1,2$, arise for $s=1,2$. We then only check these cases.
For $s=1$, there is nothing to check and everything commutes. For $s=2$, the only thing to check is the commutativity of the square
\[
\hspace{-1em}
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \mathcal{I}^{2}_{1} \oplus H^{0}(D)\otimes \mathbf{Q}[t]dt & \ker j^{2} & \ker j^{2} \\
\mathcal{I}^{2}_{0} \oplus H^{0}(D)\otimes \mathbf{Q}[t]dt & \D{(\ker j^{4})} \oplus \D{(\ker \gamma^{4})} \otimes t & \coker{\gamma^{2}}\\};
\path[->]
(m-1-2) edge node[auto,swap] {$\psip{2}{0,2}$} (m-1-1);
\path[->]
(m-1-2) edge node[auto] {$\phip{2}{0,2}$} (m-1-3);
\path[->]
(m-2-2) edge node[auto] {$\psip{0}{0,2}$} (m-2-1);
\path[->]
(m-2-2) edge node[auto,swap] {$\phip{0}{0,2}$} (m-2-3);
\path[injection]
(m-1-1) edge node[auto,swap] {$EI_{1}(\pomap{2}{0})$} (m-2-1);
\path[injection]
(m-1-2) edge node[auto] {$\pomap{2}{0}$} (m-2-2);
\path[injection]
(m-1-3) edge node[auto] {$EI_{2}(\pomap{2}{0})$} (m-2-3);
\end{tikzpicture}
\]
Which is clear by the previous computations.
\subsubsection{The infinite perversity}
We finish with the perversity $\overline{\infty}$. We define the bigraded differential algebra $(\FI{\infty}{r,s}{X},\diff{\infty}{r,s})$ as the sub-algebra of $(\EI{1}{\infty}{X},\difb{\infty}{r,s})$ given by
\[
\begin{array}{| c || c c c |}
\hline
s = 6 & 0 & H^{6}(\widetilde{X}) & 0 \\
\hline
s = 5 & 0 & H^{5}(\widetilde{X}) & 0 \\
\hline
s = 4 & 0 & \ker j^{4} & 0 \\
\hline
s = 3 & 0 & \ker j^{3} & 0 \\
\hline
s = 2 & 0 & \ker j^{2} & \D{(\ker \gamma^{4})} \otimes dt \\
\hline
s=1 & 0 & \ker j^{1} & 0 \\
\hline
\hline
s = 0 & 0 & H^{0}(\widetilde{X}) & 0 \\
\hline
\hline
\FI{\infty}{r,s}{X} & r=-1 & r=0 & r=1 \\
\hline
\end{array}
\]
There is no non trivial differentials. The algebra structure is as always concentrated in $r=0$. The map $\pomap{\infty}{4}$ is the canonical inclusion and is compatible the algebra structure.
The maps $\psip{\infty}{\ast,\ast}$ and $\phip{\infty}{\ast,\ast}$ are clear from the previous computations for the top perversity.
We then have a quasi-isomorphism of algebras
\[
(\EIbidg{1}{\infty}{r,s}{X},\difb{\infty}{r,s}) \overset{\psip{\infty}{r,s}}{\longleftarrow} (\FI{\infty}{r,s}{X},\diff{\infty}{r,s}) \overset{\phip{\infty}{r,s}}{\longrightarrow} (\EIbidg{2}{\infty}{r,s}{X},0).
\]
We then define the coperverse cdga $\FI{\bullet}{\ast,\ast}{X}$ to be
\[
\FI{\bullet}{\ast,\ast}{X} =
\begin{cases}
\FI{\infty}{\ast,\ast}{X} & \overline{p} = \infty, \\
\FI{4}{\ast,\ast}{X} & \overline{p} \in \lbrace \overline{3} , \overline{4} \rbrace, \\
\FI{2}{\ast,\ast}{X} & \overline{p} = \overline{2}, \\
\FI{0}{\ast,\ast}{X} & \overline{p} \in \lbrace \overline{0} , \overline{1} \rbrace. \\
\end{cases}
\]
We then have a quasi-isomorphism of coperverse cdga's.
\[
(\EIbidg{1}{\bullet}{\ast,\ast}{X},\difb{\bullet}{\ast,\ast}) \overset{\psip{\bullet}{\ast,\ast}}{\longleftarrow} (\FI{\bullet}{\ast,\ast}{X},\diff{\bullet}{\ast,\ast}) \overset{\phip{\bullet}{\ast,\ast}}{\longrightarrow} (\EIbidg{2}{\bullet}{\ast,\ast}{X},0).
\]
Then $\I{\bullet}{X}$ is formal.
\section{Examples and Applications}
\label{section:Eg}
We use the following conventions in the rest of this section :
\begin{itemize}
\item When needed, we will denote by $\lbrace 1_{i}, E_{i} \rbrace$ a basis of $H^{\ast}(\mathbf{C}P^{1}_{(i)})$, we complete it into a basis $\lbrace 1_{i}, E_{i}, \mathcal{E}_{i}, \Lambda_{i} \rbrace$ of $H^{\ast}(\mathbf{C}P^{1}_{(i)}\times \mathbf{C}P^{1}_{(i)})$ with $|E_{i}|=|\mathcal{E}_{i}|=2$, $|\Lambda_{i}|=4$ and where $\mathcal{E}_{i} E_{i} = \Lambda_{i}$.
\item even if we do not take into account the loops in the first cohomology group (see subsection \ref{subsubsec:coker}), we mark them in red
\end{itemize}
\subsection{Projective cone over a K3 surface}
\begin{defi}
A K3 surface $S$ is a simply connected compact smooth complex surface such that its canonical bundle $K_{S}$ is trivial.
\end{defi}
Denote by $S$ a K3 surface, for example a nonsingular degree 4 hypersurface in $\mathbf{C}P^{3}$, such as the Fermat quartic
\[
S = \lbrace [z_{0}:z_{1}:z_{2}:z_{3}] \in \mathbf{C}P^{3} \, : \, z_{0}^{4}+z_{1}^{4}+z_{2}^{4}+z_{3}^{4}=0 \rbrace.
\]
In fact every K3 surface over $\mathbf{C}$ is diffeomorphic to this example, see \cite{1964}. The Hodge diamond of a K3 surface is completely determined and is given by the following.
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=0.5em, column sep=1em, text height=2ex, text depth=0.25ex]
{ & & h^{2,2} & & \\
& h^{2,1} & & h^{1,2} & \\
h^{2,0} & & h^{1,1} & & h^{0,2} \quad =\\
& h^{1,0} & & h^{0,1} & \\
& & h^{0,0} & & \\};
\end{tikzpicture}
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=0.5em, column sep=1.5em, text height=2ex, text depth=0.25ex]
{ & & 1 & & \\
& 0 & & 0 & \\
1 & & 20 & & 1 \\
& 0 & & 0 & \\
& & 1 & & \\};
\end{tikzpicture}
\]
Which means that we have the following cohomology.
\[
\begin{array}{| c || c | c | c | c | c |}
\hline
s & 0 & 1 & 2 & 3 & 4 \\
\hline
\hline
H^{s}(S) & \mathbf{Q} & 0 & \mathbf{Q}^{22} & 0 & \mathbf{Q} \\
\hline
\end{array}
\]
Denote by $\mathbb{P}_{\mathbf{C}}S \subset \mathbf{C}P^{4}$ the projective cone over the K3 surface. This is a simply connected hypersurface of complex dimension 3 with only one isolated singularity which is the cone point and defined by the same equation but in $\mathbf{C}P^{4}$
\[
\mathbb{P}_{\mathbf{C}}S = \lbrace [z_{0}:z_{1}:z_{2}:z_{3}:z_{4}] \in \mathbf{C}P^{4} \, : \, z_{0}^{4}+z_{1}^{4}+z_{2}^{4}+z_{3}^{4}=0 \rbrace.
\]
The cohomology of $\mathbb{P}_{\mathbf{C}}S$ is given by (see \cite[p.169]{Dimca1992})
\[
H^{k}(\mathbb{P}_{\mathbf{C}}S) = H^{k-2}(S) \, \forall \, k \geq 2.
\]
By Hironaka's Theorem on resolution of singularities there exists a cartesian diagram
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ S & \widetilde{P} \\
\ast & \mathbb{P}_{\mathbf{C}}S \\};
\path[->]
(m-1-1) edge node[auto] {} (m-1-2);
\path[injection]
(m-2-1) edge node[auto,swap] {} (m-2-2);
\path[->]
(m-1-1) edge node[auto,swap] {} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$f$} (m-2-2);
\end{tikzpicture}
\]
where the exceptional divisor is the K3 surface $S$ and $\widetilde{P}$ is a smooth projective variety of complex dimension 3. We then have the following Mayer-Vietoris sequence
\[
\cdots \rightarrow H^{k}(\mathbb{P}_{\mathbf{C}}S) \rightarrow H^{k}(\widetilde{P}) \oplus H^{k}(\ast) \rightarrow H^{k}(S) \rightarrow \cdots
\]
which gives the following cohomology for $\widetilde{P}$.
\[
\begin{array}{| c || c | c | c | c | c | c | c |}
\hline
s & 0 & 1 & 2 & 3 & 4 & 5 & 6 \\
\hline
\hline
H^{s}(\widetilde{P}) & \mathbf{Q} & 0 & \mathbf{Q} \oplus \mathbf{Q}^{22} & 0 & \mathbf{Q} \oplus \mathbf{Q}^{22} & 0 & \mathbf{Q} \\
\hline
\end{array}
\]
We compute the intersection space for the perversities $\lbrace \overline{0}, \overline{1}, \overline{2}, \overline{3}, \overline{4} \rbrace$.
First of all The intersection space for the zero perversity is by definition the regular part, which is computed by the following spectral sequence
\[
\begin{array}{| c || c | c || c | c |}
\hline
\multicolumn{3}{|c||}{E_{1}^{r,s}((\mathbb{P}_{\mathbf{C}}S)_{reg})} & \multicolumn{2}{c|}{E_{2}^{r,s}((\mathbb{P}_{\mathbf{C}}S)_{reg})} \\
\hline
s=6 & \mathbf{Q} & \mathbf{Q} & 0 &0\\
\hline
s=5 & 0 & 0 & 0 &0\\
\hline
s=4 & \mathbf{Q}^{22} & \mathbf{Q} \oplus \mathbf{Q}^{22} & 0 & \mathbf{Q}\\
\hline
s=3 & 0 & 0 & 0 &0\\
\hline
s=2 & \mathbf{Q} & \mathbf{Q} \oplus \mathbf{Q}^{22} & 0 & \mathbf{Q}^{22}\\
\hline
s=1 & 0 & 0 & 0 &0\\
\hline
s=0 & 0 & \mathbf{Q} & 0 &\mathbf{Q}\\
\hline
\hline
& r=-1 & r=0 & \ker \gamma^{s} & \coker{\gamma^{s}} \\
\hline
\end{array}
\]
Now we need the cohomology of the link, which is given by the spectral sequence defined by $j^{s}_{\sharp} \col H^{s-2}(D) \rightarrow H^{s}(D)$, as in the section \ref{subsec:smooth_div}.
\[
\begin{array}{| c || c | c || c | c |}
\hline
\multicolumn{3}{|c||}{E_{1}^{r,s}(L)} & \multicolumn{2}{c|}{E_{2}^{r,s}(L)} \\
\hline
s=6 & \mathbf{Q} & 0 & \mathbf{Q} &0\\
\hline
s=5 & 0 & 0 & 0 &0\\
\hline
s=4 & \mathbf{Q}^{22} & \mathbf{Q} & \mathbf{Q}^{21} & 0\\
\hline
s=3 & 0 & 0 & 0 &0\\
\hline
s=2 & \mathbf{Q} & \mathbf{Q}^{22} & 0 & \mathbf{Q}^{21}\\
\hline
s=1 & 0 & 0 & 0 &0\\
\hline
s=0 & 0 & \mathbf{Q} & 0 &\mathbf{Q}\\
\hline
\hline
& r=-1 & r=0 & \ker j_{\sharp}^{s} & \coker{j_{\sharp}^{s}} \\
\hline
\end{array}
\]
We then have
\[
\begin{cases}
H^{0}(L) & = H^{5}(L) = \mathbf{Q}, \\
H^{1}(L) & = H^{4}(L) = 0, \\
H^{2}(L) & = H^{3}(L) = \mathbf{Q}^{21}. \\
\end{cases}
\]
By the $E_{2}$ term of the previous spectral sequence we see that the only sections of $j^{s}_{\sharp}$ for which the image won't be zero correspond to the perversities $\overline{1}$ and $\overline{3}$. Each times the image of the section is equal to $\mathbf{Q}$, we then have the two following map
\[
\gamma^{2}_{|\coim{1}} \col \coim{1}=\mathbf{Q} \longrightarrow H^{2}(\widetilde{P}) = \mathbf{Q} \oplus \mathbf{Q}^{22},
\]
\[
\gamma^{4}_{|\coim{3}} \col \coim{3}=\mathbf{Q} \longrightarrow H^{4}(\widetilde{P}) = \mathbf{Q} \oplus \mathbf{Q}^{22}.
\]
and $\coker{\gamma^{2}_{|\coim{1}}} \cong \coker{\gamma^{4}_{|\coim{3}}} \cong \mathbf{Q}^{22}$.
The last map we need to know is $j^{s} \col H^{s}(\widetilde{P}) \rightarrow H^{s}(S)$, the map induced by the inclusion $S \hookrightarrow \widetilde{P}$.
\[
\begin{array}{| c || c | c |}
\hline
s=6 & \mathbf{Q} & 0 \\
\hline
s=5 & 0 & 0 \\
\hline
s=4 & \mathbf{Q}^{22} & 0 \\
\hline
s=3 & 0 & 0 \\
\hline
s=2 & \mathbf{Q} & 0 \\
\hline
s=1 & 0 & 0 \\
\hline
s=0 & 0 & 0 \\
\hline
\hline
& \ker j^{s} & \coker{j^{s}} \\
\hline
\end{array}
\]
We recall the $EI_{2}$ term of the spectral sequence of $\I{p}{X}$.
\[
\begin{array}{| c || c c c |}
\hline
s > p+1 & \ker \gamma^{s} & \coker{\gamma^{s}} & 0 \\
\hline
s = p+1 & 0 & \coker{\gamma^{s}_{|\coim{p}}} & 0 \\
\hline
1 \leq s < p+1 & 0 & \ker j^{s} & \coker{j^{s}} \\
\hline
\hline
s = 0 & 0 & H^{0}(\widetilde{P}) & 0 \\
\hline
\hline
\EIbidg{2}{p}{r,s}{X} & r=-1 & r=0 & r=1 \\
\hline
\end{array}
\]
We then have the following results.
\[
\begin{array}{| c || c |}
\hline
\multicolumn{2}{|c|}{\EIbidg{2}{1}{r,s}{\mathbb{P}_{\mathbf{C}}S}} \\
\hline
s \geq 5 & 0 \\
\hline
s = 4 & \mathbf{Q} \\
\hline
s = 3 & 0 \\
\hline
s = 2 & \mathbf{Q}^{22} \\
\hline
s = 1 & 0 \\
\hline
\hline
s = 0 & H^{0}(\widetilde{P}) \\
\hline
\hline
& r=0 \\
\hline
\end{array}
\quad
\begin{array}{| c || c |}
\hline
\multicolumn{2}{|c|}{\EIbidg{2}{2}{r,s}{\mathbb{P}_{\mathbf{C}}S}} \\
\hline
s \geq 5 & 0 \\
\hline
s = 4 & \mathbf{Q} \\
\hline
s = 3 & 0 \\
\hline
s = 2 & \mathbf{Q} \\
\hline
s = 1 & 0 \\
\hline
\hline
s = 0 & H^{0}(\widetilde{P}) \\
\hline
\hline
& r=0 \\
\hline
\end{array}
\quad
\begin{array}{| c || c|}
\hline
\multicolumn{2}{|c|}{\EIbidg{2}{3}{r,s}{\mathbb{P}_{\mathbf{C}}S}} \\
\hline
s \geq 5 & 0 \\
\hline
s = 4 & \mathbf{Q}^{22} \\
\hline
s = 3 & 0 \\
\hline
s = 2 & \mathbf{Q} \\
\hline
s = 1 & 0 \\
\hline
\hline
s = 0 & H^{0}(\widetilde{P}) \\
\hline
\hline
& r=0 \\
\hline
\end{array}
\quad
\begin{array}{| c || c |}
\hline
\multicolumn{2}{|c|}{\EIbidg{2}{4}{r,s}{\mathbb{P}_{\mathbf{C}}S}} \\
\hline
s \geq 5 & 0 \\
\hline
s = 4 & \mathbf{Q}^{22} \\
\hline
s = 3 & 0 \\
\hline
s = 2 & \mathbf{Q} \\
\hline
s = 1 & 0 \\
\hline
\hline
s = 0 & H^{0}(\widetilde{P}) \\
\hline
\hline
& r=0 \\
\hline
\end{array}
\]
Note that for complementary perversities, such as $\overline{1}$ and $\overline{3}$ or $\overline{0}$ and $\overline{4}$, and for $s \neq 0$ the $EI_{2}$ term gives back the generalized Poincaré duality between the various intersection spaces such as proved in \cite[theorem 2.12]{Banagl2010}. The middle perversity here is $\overline{2}$ and we also get back the self-duality of the space $\I{2}{\mathbb{P}_{\mathbf{C}}S}$.
For any perversity $\overline{p}$ the weight filtration is pure, so by the theorem \ref{thm:pure_is_formal} we get the following proposition.
\begin{propo}
Given any perversity $\overline{p}$, the intersection space $\I{p}{\mathbb{P}_{\mathbf{C}}S}$ is a formal topological space.
\end{propo}
\subsection{Kummer quartic surface}
\label{subsec:Kummer}
Let $K$ be a Kummer quartic surface, that is an irreducible surface of degree 4 in $\mathbf{C}P^{3}$ with 16 ordinary double points, which is the maximum for such surfaces.
From the algebraic topologist point of view, a Kummer surface is constructed in the following way. Let's consider a 4-dimensional torus
\[
\mathbf{T} = S^{1} \times S^{1} \times S^{1} \times S^{1}
\]
endowed with the complex involution $\tau \col z \mapsto \bar{z}$ action. This action has 16 fixed point and we define the Kummer surface to be the quotient complex surface
\[
K := \mathbf{T}/\tau.
\]
We have the following cohomology for $K$.
\[
\begin{array}{| c || c | c | c | c | c |}
\hline
s & 0 & 1 & 2 & 3 & 4 \\
\hline
\hline
H^{s}(K) & \mathbf{Q} & 0 & \mathbf{Q}^{6} & 0 & \mathbf{Q} \\
\hline
\end{array}
\]
The link of each singularity is then a projective space $\mathbf{R}P^{3}$. These singularities are quotients singularities so by \cite{Durfee1979} $K$ admits a resolution where the exceptional set consists of curves of genus zero and self-intersection $-2$. Which means we have the following resolution diagram
\[
\begin{tikzpicture}[injection/.style={right hook->,fill=white, inner sep=2pt}]
\matrix (m)[matrix of math nodes, row sep=3em, column sep=5em, text height=2ex, text depth=0.25ex]
{ \bigsqcup_{i=1}^{16} \mathbf{C}P^{1}_{(i)} & \widetilde{K} \\
\bigsqcup_{i=1}^{16} \ast_{(i)} & K \\};
\path[->]
(m-1-1) edge node[auto] {} (m-1-2);
\path[injection]
(m-2-1) edge node[auto,swap] {} (m-2-2);
\path[->]
(m-1-1) edge node[auto,swap] {} (m-2-1);
\path[->]
(m-1-2) edge node[auto] {$f$} (m-2-2);
\end{tikzpicture}
\]
The Mayer-Vietoris sequence gives the following cohomology for $\widetilde{K}$.
\[
\begin{array}{| c || c | c | c | c | c |}
\hline
s & 0 & 1 & 2 & 3 & 4 \\
\hline
\hline
H^{s}(\widetilde{K}) & \mathbf{Q} & 0 & \mathbf{Q}^{6}\oplus \bigoplus_{i=1}^{16}\mathbf{Q}E_{i} & 0 & \mathbf{Q} \\
\hline
\end{array}
\]
We have the fairly easy following spectral sequence for the links.
\[
\begin{array}{| c || c | c || c | c |}
\hline
\multicolumn{3}{|c||}{E_{1}^{r,s}(L)} & \multicolumn{2}{c|}{E_{2}^{r,s}(L)} \\
\hline
s=4 & \bigoplus_{i=1}^{16}\mathbf{Q}E_{i} & 0 & \bigoplus_{i=1}^{16}\mathbf{Q}E_{i} & 0\\
\hline
s=3 & 0 & 0 & 0 &0\\
\hline
s=2 & \bigoplus_{i=1}^{16}\mathbf{Q}1_{i} & \bigoplus_{i=1}^{16}\mathbf{Q}E_{i} & 0 & 0\\
\hline
s=1 & 0 & 0 & 0 &0\\
\hline
s=0 & 0 & \bigoplus_{i=1}^{16}\mathbf{Q}1_{i} & 0 &\bigoplus_{i=1}^{16}\mathbf{Q}1_{i}\\
\hline
\hline
& r=-1 & r=0 & \ker j_{\sharp}^{s} & \coker{j_{\sharp}^{s}} \\
\hline
\end{array}
\]
The rational cohomology of link of each singularities is then a 3-sphere, which is the rationalization of $\mathbf{R}P^{3}$.
The only interesting perversity here is the middle perversity $\overline{1}$. We need a $\coim{1}$ for the computation, we have here
\[
\coim{1} = \bigoplus_{i=1}^{16}\mathbf{Q}1_{i}
\]
and $\gamma^{2}_{|\coim{1}} = \gamma^{2}$.
The following spectral sequence computes the regular part and the second array is the restriction map $j^{s}$.
\[
\begin{array}{| c || c | c || c | c |}
\hline
\multicolumn{3}{|c||}{E_{1}^{r,s}(K_{reg}) \quad \gamma^{s} \col H^{s-2}(D) \longrightarrow H^{s}(\widetilde{K})} & \multicolumn{2}{c|}{E_{2}^{r,s}(K_{reg})} \\
\hline
s=4 & \bigoplus_{i=1}^{16}\mathbf{Q}E_{i} & \mathbf{Q} & \bigoplus_{i=1}^{15}\mathbf{Q}E_{i} & 0\\
\hline
s=3 & 0 & 0 & 0 &0\\
\hline
s=2 & \bigoplus_{i=1}^{16}\mathbf{Q}1_{i} &\mathbf{Q}^{6}\oplus \bigoplus_{i=1}^{16}\mathbf{Q}E_{i} & 0 & \mathbf{Q}^{6}\\
\hline
s=1 & 0 & 0 & 0 &0\\
\hline
s=0 & 0 & \mathbf{Q} & 0 &\mathbf{Q}\\
\hline
\hline
& r=-1 & r=0 & \ker \gamma^{s} & \coker{\gamma^{s}} \\
\hline
\end{array}
\]
\[
\begin{array}{| c || c | c || c | c |}
\hline
s=4 & \mathbf{Q} & 0 & \mathbf{Q} & 0\\
\hline
s=3 & 0 & 0 & 0 &0\\
\hline
s=2 & \mathbf{Q}^{6} \oplus \bigoplus_{i=1}^{16}\mathbf{Q}E_{i} &\bigoplus_{i=1}^{16}\mathbf{Q}E_{i} & \mathbf{Q}^{6} & 0\\
\hline
s=1 & 0 & 0 & 0 &0\\
\hline
s=0 & \mathbf{Q} & \bigoplus_{i=1}^{16} \mathbf{Q}1_{i} & 0 &\bigoplus_{i=1}^{15} \mathbf{Q}1_{i}\\
\hline
\hline
& H^{s}(\widetilde{K}) & H^{s}(D) & \ker j^{s} & \coker{j^{s}} \\
\hline
\end{array}
\]
The cohomology of the middle perversity intersection space of a Kummer surface is then given by the following array. Note that the cohomology obtained isn't pure.
\[
\begin{array}{| c || c c c || c |}
\hline
s = 4 & \bigoplus_{i=1}^{15} \mathbf{Q}E_{i} & 0 & 0 & 0\\
\hline
s = 3 & 0 & 0 & 0 & \bigoplus_{i=1}^{15} \mathbf{Q}E_{i}\\
\hline
s = 2 & 0 & \mathbf{Q}^{6} & 0 & \mathbf{Q}^{6} \\
\hline
s = 1 & 0 & 0 & 0 & \textcolor{red}{\bigoplus_{i=1}^{15} \mathbf{Q}1_{i}} \\
\hline
\hline
s = 0 & 0 & H^{0}(\widetilde{K}) & \textcolor{red}{\bigoplus_{i=1}^{15} \mathbf{Q}1_{i}} &H^{0}(\widetilde{K}) \\
\hline
\hline
\EIbidg{2}{1}{r,s}{K} & r=-1 & r=0 & r=1 &\HI{1}{s}{K} \\
\hline
\end{array}
\]
\subsection{The Calabi-Yau generic quintic 3-fold}
\label{subsec:CY_generic_quintic}
Let $Y \subset \mathbf{C}P^{4}$ the singular hypersurface given by the equation
\[
Y := \lbrace [z_{0}:z_{1}:z_{2}:z_{3}:z_{4}] \in \mathbf{C}P^{4} \, : \, z_{3}g(z_{0}, \dots, z_{4}) + z_{4}h(z_{0}, \dots, z_{4}) =0 \rbrace
\]
where $g$ and $h$ are generic homogeneous polynomials of degree 4. $Y$ is the Calabi-Yau generic quintic 3-fold containing the plane
\[
\pi := \lbrace z_{3}=z_{4}=0 \rbrace \cong \mathbf{C}P^{2}.
\]
The singular locus
\[
\Sigma := \lbrace [x] \in \mathbf{C}P^{4} \, : \, z_{3}=z_{4}=g(z)=h(z)=0 \rbrace \subset \mathbf{C}P^{2}
\]
is given by 16 ordinary double points. That is the link of each singularity $\sigma \in \Sigma$ is topologically equal to $L_{\sigma} = S^{2} \times S^{3}$.
We have the following cohomology for $Y$.
\[
\begin{array}{| c || c | c | c | c | c | c | c |}
\hline
s & 0 & 1 & 2 & 3 & 4 & 5 & 6 \\
\hline
\hline
H^{s}(Y) & \mathbf{Q} & 0 & \mathbf{Q} & \mathbf{Q}^{189} & \mathbf{Q}^{2} & 0 & \mathbf{Q} \\
\hline
\end{array}
\]
We consider the following commutative diagram of resolutions
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=2.5em, text height=1.5ex, text depth=0.25ex]
{\bigsqcup_{i=1}^{16} \mathbf{C}P^{1}_{(i)} \times \mathbf{C}P^{1}_{(i)} & \bigsqcup_{i=1}^{16} \mathbf{C}P^{1}_{(i)} & \bigsqcup_{i=1}^{16} \ast_{(i)} \\
\overline{Y} & \widetilde{Y} & Y \\};
\path[->]
(m-1-1) edge node[right=2.5em] {Blow up} (m-2-1);
\path[->]
(m-1-2) edge node[right=1.25em] {small res.} (m-2-2);
\path[->]
(m-1-3) edge node[auto] {} (m-2-3);
\path[->]
(m-1-1) edge node[auto] {} (m-1-2);
\path[->]
(m-1-2) edge node[auto] {} (m-1-3);
\path[->]
(m-2-1) edge node[auto,swap] {$\mathcal{B}\ell$} (m-2-2);
\path[->]
(m-2-2) edge node[auto,swap] {$f$} (m-2-3);
\end{tikzpicture}
\]
The first square is a simultaneous small resolution of the 16 singularities obtained by blowing up $\mathbf{C}P^{4}$ along the plane $\pi \cong \mathbf{C}P^{2}$. The exceptionnal divisor of this blow-up is a $\mathbf{C}P^{1}$-bundle over $\pi \cong \mathbf{C}P^{2}$.
For the second square $\mathcal{B}\ell$ is a blow-up along the $\mathbf{C}P^{1}_{(i)}$'s.
Denote by $\Psi$ the generator of $H^{2}(Y)$.
By using twice the Mayer-Vietoris long exact sequence, we get the following cohomology for $\overline{Y}$.
\[
\begin{cases}
H^{0}(\overline{Y}) = H^{6}(\overline{Y}) = \mathbf{Q}, & \\
H^{1}(\overline{Y}) = H^{5}(\overline{Y}) = 0, &\\
H^{2}(\overline{Y}) = \mathbf{Q}\Psi \oplus \mathbf{Q}E_{1} \oplus \bigoplus_{i=1}^{16} \mathbf{Q}\D{\Lambda_{i}}, & \\
H^{4}(\overline{Y}) = \mathbf{Q}\D{\Psi} \oplus \mathbf{Q}\D{E_{1}} \oplus \bigoplus_{i=1}^{16} \mathbf{Q}\Lambda_{i}, & \\
H^{3}(\overline{Y}) = \mathbf{Q}^{174}.
\end{cases}
\]
The cohomology of the links of the singularities is given by the spectral sequence
\[
\begin{array}{| c || c | c || c | c |}
\hline
\multicolumn{3}{|c||}{E_{1}^{r,s}(L)} & \multicolumn{2}{c|}{E_{2}^{r,s}(L)} \\
\hline
s=6 & \bigoplus_{i=1}^{16} \mathbf{Q}\Lambda_{i} & 0 &\bigoplus_{i=1}^{16} \mathbf{Q}\Lambda_{i} & 0 \\
\hline
s=5 & 0 & 0 & 0 & 0\\
\hline
s=4 & \bigoplus_{i=1}^{16} (\mathbf{Q}E_{i} \oplus \mathbf{Q}\mathcal{E}_{i}) & \bigoplus_{i=1}^{16} \mathbf{Q}\Lambda_{i} &\bigoplus_{i=1}^{16} \mathbf{Q}\mathcal{E}_{i} & 0\\
\hline
s=3 & 0 & 0 & 0 & 0\\
\hline
s=2 & \bigoplus_{i=1}^{16} \mathbf{Q}1_{i} & \bigoplus_{i=1}^{16} (\mathbf{Q}E_{i} \oplus \mathbf{Q}\mathcal{E}_{i}) & 0 & \bigoplus_{i=1}^{16} \mathbf{Q}E_{i} \\
\hline
s=1 & 0 & 0 & 0 & 0 \\
\hline
s=0 & 0 & \bigoplus_{i=1}^{16} \mathbf{Q}1_{i} & 0 & \bigoplus_{i=1}^{16} \mathbf{Q}1_{i} \\
\hline
\hline
& r=-1 & r=0 &\ker j^{s}_{\sharp} & \coker{j^{s}_{\sharp}} \\
\hline
\end{array}
\]
We here follow the section \ref{sec:formality_3folds} and do the computations for the top, middle and zero perversity.
The spectral sequence of the regular part is given by
\[
\hspace{-0.5em}
\begin{array}{| c || c | c || c | c |}
\hline
\multicolumn{3}{|c||}{E_{1}^{r,s}(Y_{reg}) \quad \gamma^{s} \col H^{s-2}(D) \longrightarrow H^{s}(\overline{Y})} & \multicolumn{2}{c|}{E_{2}^{r,s}(Y_{reg})} \\
\hline
s=6 & \bigoplus_{i=1}^{16} \mathbf{Q}\Lambda_{i} & \mathbf{Q} & \bigoplus_{i=1}^{15} \mathbf{Q}\Lambda_{i} & 0 \\
\hline
s=5 & 0 & 0 & 0 & 0 \\
\hline
s=4 & \bigoplus_{i=1}^{16} (\mathbf{Q}E_{i} \oplus \mathbf{Q}\mathcal{E}_{i}) & \mathbf{Q}\D{\Psi} \oplus \mathbf{Q}\D{E_{1}} \oplus \bigoplus_{i=1}^{16} \mathbf{Q}\Lambda_{i} & \bigoplus_{i=1}^{15} \mathbf{Q}\mathcal{E}_{i} & \mathbf{Q}\D{\Psi} \\
\hline
s=3 & 0 & \mathbf{Q}^{174} & 0 & \mathbf{Q}^{174} \\
\hline
s=2 & \bigoplus_{i=1}^{16} \mathbf{Q}1_{i} & \mathbf{Q}\Psi \oplus \mathbf{Q}E_{1} \oplus \bigoplus_{i=1}^{16} \mathbf{Q}\D{\Lambda_{i}} & 0 & \mathbf{Q}\Psi \oplus \mathbf{Q}E_{1} \\
\hline
s=1 & 0 & 0 & 0 & 0 \\
\hline
s=0 & 0 & \mathbf{Q} & 0 & \mathbf{Q} \\
\hline
\hline
& r=-1 & r=0 & \ker \gamma^{s} &\coker{\gamma^{s}} \\
\hline
\end{array}
\]
Finally we also need the restriction morphism $j^{s}$.
\[
\hspace{-0.9em}
\begin{array}{| c || c | c || c | c |}
\hline
s=6 & \mathbf{Q} & 0 & \mathbf{Q} & 0 \\
\hline
s=5 & 0 & 0 & 0 & 0 \\
\hline
s=4 & \mathbf{Q}\D{\Psi} \oplus \mathbf{Q}\D{E_{1}} \oplus \bigoplus_{i=1}^{16} \mathbf{Q}\Lambda_{i} & \bigoplus_{i=1}^{16} \mathbf{Q}\Lambda_{i} & \mathbf{Q}\D{\Psi} \oplus \mathbf{Q}\D{E_{1}} & 0 \\
\hline
s=3 & \mathbf{Q}^{174} & 0 & \mathbf{Q}^{174} & 0 \\
\hline
s=2 & \mathbf{Q}\Psi \oplus \mathbf{Q}E_{1} \oplus \bigoplus_{i=1}^{16} \mathbf{Q}\D{\Lambda_{i}} & \bigoplus_{i=1}^{16} (\mathbf{Q}E_{i} \oplus \mathbf{Q}\mathcal{E}_{i}) & \mathbf{Q}\Psi & \bigoplus_{i=1}^{15}\mathbf{Q}E_{i} \\
\hline
s=1 & 0 & 0 & 0 & 0 \\
\hline
s=0 & \mathbf{Q} & \bigoplus_{i=1}^{16} \mathbf{Q}1_{i} & 0 & \bigoplus_{i=1}^{15} \mathbf{Q}1_{i} \\
\hline
\hline
& H^{s}(\overline{Y}) & H^{s}(D) & \ker j^{s} & \coker{j^{s}} \\
\hline
\end{array}
\]
We then get the following tables for the perversities $\overline{0}, \overline{2}, \overline{4}$. Note here that the generalized Poincaré duality is only partial as we explained in the subsection \ref{subsubsec:coker} since we do not take into accounts the loops of $\coker{j^{0}}$ (marked in red in the arrays).
\[
\begin{array}{| c || c c c || c |}
\hline
s = 6 & \bigoplus_{i=1}^{15} \mathbf{Q}\Lambda_{i} & 0 & 0 & 0 \\
\hline
s = 5 & 0 & 0 & 0 & \bigoplus_{i=1}^{15} \mathbf{Q}\Lambda_{i} \\
\hline
s = 4 & \bigoplus_{i=1}^{15}\mathbf{Q}\mathcal{E}_{i} & \mathbf{Q}\D{\Psi} & 0 & \mathbf{Q}\D{\Psi} \\
\hline
s = 3 & 0 & \mathbf{Q}^{174} & 0 & \mathbf{Q}^{189} \\
\hline
s = 2 & 0 & \mathbf{Q}\Psi \oplus \mathbf{Q}E_{1} & 0 & \mathbf{Q}\Psi \oplus \mathbf{Q}E_{1} \\
\hline
s = 1 & 0 & 0 & 0 & \textcolor{red}{\bigoplus_{i=1}^{15} \mathbf{Q}1_{i}} \\
\hline
\hline
s = 0 & 0 & H^{0}(\overline{Y}) & \textcolor{red}{\bigoplus_{i=1}^{15} \mathbf{Q}1_{i}} & H^{0}(\overline{Y}) \\
\hline
\hline
\EIbidg{2}{0}{r,s}{Y} & r=-1 & r=0 & r=1 & \HI{0}{s}{Y} \\
\hline
\end{array}
\]
Note here the partial duality for the values $s=2,3,4$ for the perversities $\overline{0}$ and $\overline{4}$.
\[
\begin{array}{| c || c c c || c |}
\hline
s = 6 & \bigoplus_{i=1}^{15} \mathbf{Q}\Lambda_{i} & 0 & 0 & 0 \\
\hline
s = 5 & 0 & 0 & 0 & \bigoplus_{i=1}^{15} \mathbf{Q}\Lambda_{i} \\
\hline
s = 4 & 0 & \mathbf{Q}\D{\Psi} \oplus \mathbf{Q}\D{E_{1}} & 0 & \mathbf{Q}\D{\Psi} \oplus \mathbf{Q}\D{E_{1}}\\
\hline
s = 3 & 0 & \mathbf{Q}^{174} & 0 & \mathbf{Q}^{189}\\
\hline
s = 2 & 0 & \mathbf{Q}\Psi & \bigoplus_{i=1}^{15}\mathbf{Q}E_{i} & \mathbf{Q}\Psi \\
\hline
s = 1 & 0 & 0 & 0 & \textcolor{red}{\bigoplus_{i=1}^{15} \mathbf{Q}1_{i}} \\
\hline
\hline
s = 0 & 0 & H^{0}(\overline{Y}) & \textcolor{red}{\bigoplus_{i=1}^{15} \mathbf{Q}1_{i}} & H^{0}(\overline{Y})\\
\hline
\hline
\EIbidg{2}{4}{r,s}{Y} & r=-1 & r=0 & r=1 & \HI{4}{s}{Y} \\
\hline
\end{array}
\]
For the perversity $\overline{2}$ we retrieve the values of the smooth deformation as in \cite{Banagl2012}, unless for $s=1$.
\[
\begin{array}{| c || c c c || c |}
\hline
s = 6 & \bigoplus_{i=1}^{15} \mathbf{Q}\Lambda_{i} & 0 & 0 &0 \\
\hline
s = 5 & 0 & 0 & 0 &\bigoplus_{i=1}^{15} \mathbf{Q}\Lambda_{i} \\
\hline
s = 4 & \bigoplus_{i=1}^{15}\mathbf{Q}\mathcal{E}_{i} & \mathbf{Q}\D{\Psi} & 0 &\mathbf{Q}\D{\Psi}\\
\hline
s = 3 & 0 & \mathbf{Q}^{174} & 0 &\mathbf{Q}^{204}\\
\hline
s = 2 & 0 & \mathbf{Q}\Psi & \bigoplus_{i=1}^{15}\mathbf{Q}E_{i} &\mathbf{Q}\Psi \\
\hline
s = 1 & 0 & 0 & 0 &\textcolor{red}{\bigoplus_{i=1}^{15} \mathbf{Q}1_{i}}\\
\hline
\hline
s = 0 & 0 & H^{0}(\overline{Y}) & \textcolor{red}{\bigoplus_{i=1}^{15} \mathbf{Q}1_{i}} & H^{0}(\overline{Y}) \\
\hline
\hline
\EIbidg{2}{2}{r,s}{Y} & r=-1 & r=0 & r=1 & \HI{2}{s}{Y}\\
\hline
\end{array}
\]
\subsection{The Quintic}
\label{subsec:CY_quintic}
Let $\psi$ be a complex number and consider the variety
\[
X_{\psi} := \left\lbrace [z_{0}:z_{1}:z_{2}:z_{3}:z_{4}] \in \mathbf{C}P^{4} \, : \, z_{0}^{5} + z_{1}^{5} + z_{2}^{5} + z_{3}^{5} + z_{4}^{5} -5\psi z_{0}z_{1}z_{2}z_{3}z_{4}=0 \right\rbrace,
\]
which is Calabi-Yau. It is smooth for small $\psi \neq 1$ and becomes singular when $\psi =1$, denote by $X$ the singular degeneration $X_{\psi=1}$.
The singular locus $\Sigma$ of $X$ is here composed of 125 ordinary double points. That is the link of each singularity $\sigma \in \Sigma$ is topologically equal to $L_{\sigma} = S^{2} \times S^{3}$, just like before.
We get the following cohomology for $X$.
\[
\begin{array}{| c || c | c | c | c | c | c | c |}
\hline
s & 0 & 1 & 2 & 3 & 4 & 5 & 6 \\
\hline
\hline
H^{s}(X) & \mathbf{Q} & 0 & \mathbf{Q} & \mathbf{Q}^{103} & \mathbf{Q}^{25} & 0 & \mathbf{Q} \\
\hline
\end{array}
\]
Using the same method of resolution that before
\[
\begin{tikzpicture}
\matrix (m)[matrix of math nodes, row sep=3em, column sep=2.5em, text height=1.5ex, text depth=0.25ex]
{\bigsqcup_{i=1}^{125} \mathbf{C}P^{1}_{(i)} \times \mathbf{C}P^{1}_{(i)} & \bigsqcup_{i=1}^{125} \mathbf{C}P^{1}_{(i)} & \bigsqcup_{i=1}^{125} \ast_{(i)} \\
\overline{X} & \widetilde{X} & X \\};
\path[->]
(m-1-1) edge node[right=2.5em] {Blow up} (m-2-1);
\path[->]
(m-1-2) edge node[right=1.25em] {small res.} (m-2-2);
\path[->]
(m-1-3) edge node[auto] {} (m-2-3);
\path[->]
(m-1-1) edge node[auto] {} (m-1-2);
\path[->]
(m-1-2) edge node[auto] {} (m-1-3);
\path[->]
(m-2-1) edge node[auto,swap] {$\mathcal{B}\ell$} (m-2-2);
\path[->]
(m-2-2) edge node[auto,swap] {$f$} (m-2-3);
\end{tikzpicture}
\]
With the Mayer-Vietoris long exact sequence, we get the following cohomology for $\overline{X}$, we still denote by $\Psi$ the generator of $H^{2}(X)$.
\[
\begin{array}{| c || c |}
\hline
6 & \mathbf{Q} \\
\hline
5 & 0 \\
\hline
4 & \mathbf{Q}\D{\Psi} \oplus \bigoplus_{i=1}^{24}\D{E_{i}} \oplus \bigoplus_{i=1}^{125}\mathbf{Q}\Lambda_{i} \\
\hline
3 & \mathbf{Q}^{2} \\
\hline
2 & \mathbf{Q}\Psi \oplus \bigoplus_{i=1}^{24}\mathbf{Q}E_{i} \oplus \bigoplus_{i=1}^{125}\mathbf{Q}\D{\Lambda_{i}} \\
\hline
1 & 0 \\
\hline
0 & \mathbf{Q} \\
\hline
\hline
s & H^{s}(\overline{X}) \\
\hline
\end{array}
\]
The spectral sequences of the regular part if given by
\[
\hspace{-2.7em}
\begin{array}{| c || c | c || c | c |}
\hline
\multicolumn{3}{|c||}{E_{1}^{r,s}(X_{reg}) \quad \gamma^{s} \col H^{s-2}(D) \longrightarrow H^{s}(\overline{X})} & \multicolumn{2}{c|}{E_{2}^{r,s}(X_{reg})} \\
\hline
s=6 & \bigoplus_{i=1}^{125} \mathbf{Q}\Lambda_{i} & \mathbf{Q} & \bigoplus_{i=1}^{124} \mathbf{Q}\Lambda_{i} & 0 \\
\hline
s=5 & 0 & 0 & 0 & 0 \\
\hline
s=4 & \bigoplus_{i=1}^{125} (\mathbf{Q}E_{i} \oplus \mathbf{Q}\mathcal{E}_{i}) & \mathbf{Q}\D{\Psi} \oplus \bigoplus_{i=1}^{24}\D{E_{i}} \oplus \bigoplus_{i=1}^{125}\mathbf{Q}\Lambda_{i} & \bigoplus_{i=1}^{101} \mathbf{Q}\mathcal{E}_{i} & \mathbf{Q}\D{\Psi}\\
\hline
s=3 & 0 & \mathbf{Q}^{2} & 0 & \mathbf{Q}^{2} \\
\hline
s=2 & \bigoplus_{i=1}^{125} \mathbf{Q}1_{i} & \mathbf{Q}\Psi \oplus \bigoplus_{i=1}^{24}\mathbf{Q}E_{i} \oplus \bigoplus_{i=1}^{125}\mathbf{Q}\D{\Lambda_{i}} & 0 & \mathbf{Q}\Psi \oplus \bigoplus_{i=1}^{24}\mathbf{Q}E_{i}\\
\hline
s=1 & 0 & 0 &0 & 0 \\
\hline
s=0 & 0 & \mathbf{Q} & 0 & \mathbf{Q} \\
\hline
\hline
& r=-1 & r=0 & \ker\gamma^{s} &\coker{\gamma^{s}} \\
\hline
\end{array}
\]
The formulas for the restriction morphism are
\[
\hspace{-3.2em}
\begin{array}{| c || c | c || c | c |}
\hline
s=6 & \mathbf{Q} & 0 & \mathbf{Q} & 0 \\
\hline
s=5 & 0 & 0 & 0 & 0 \\
\hline
s=4 & \mathbf{Q}\D{\Psi} \oplus \bigoplus_{i=1}^{24}\D{E_{i}} \oplus \bigoplus_{i=1}^{125}\mathbf{Q}\Lambda_{i} & \bigoplus_{i=1}^{125} \mathbf{Q}\Lambda_{i} & \mathbf{Q}\D{\Psi} \oplus \bigoplus_{i=1}^{24}\mathbf{Q}\D{E_{i}} & 0 \\
\hline
s=3 & \mathbf{Q}^{2} & 0 & \mathbf{Q}^{2} & 0 \\
\hline
s=2 & \mathbf{Q}\Psi \oplus \bigoplus_{i=1}^{24}\mathbf{Q}E_{i} \oplus \bigoplus_{i=1}^{125}\mathbf{Q}\D{\Lambda_{i}} & \bigoplus_{i=1}^{125} (\mathbf{Q}E_{i} \oplus \mathbf{Q}\mathcal{E}_{i}) & \mathbf{Q}\Psi & \bigoplus_{i=1}^{101}\mathbf{Q}E_{i} \\
\hline
s=1 & 0 & 0 & 0 & 0 \\
\hline
s=0 & \mathbf{Q} & \bigoplus_{i=1}^{125} \mathbf{Q}1_{i} & 0 & \bigoplus_{i=1}^{124} \mathbf{Q}1_{i} \\
\hline
\hline
& H^{s}(\overline{X}) & H^{s}(D) & \ker j^{s} & \coker{j^{s}} \\
\hline
\end{array}
\]
We let the reader fill in the arrays for the top and zero perversities, we here give the result for the middle perversity $\overline{2}$.
\[
\begin{array}{| c || c c c || c |}
\hline
s = 6 & \bigoplus_{i=1}^{124} \mathbf{Q}\Lambda_{i} & 0 & 0 & 0 \\
\hline
s = 5 & 0 & 0 & 0 & \bigoplus_{i=1}^{124} \mathbf{Q}\Lambda_{i} \\
\hline
s = 4 & \bigoplus_{i=1}^{101} \mathbf{Q}\mathcal{E}_{i} & \mathbf{Q}\D{\Psi} & 0 & \mathbf{Q}\D{\Psi} \\
\hline
s = 3 & 0 & \mathbf{Q}^{2} & 0 & \mathbf{Q}^{204}\\
\hline
s = 2 & 0 & \mathbf{Q}\Psi & \bigoplus_{i=1}^{101}\mathbf{Q}E_{i} & \mathbf{Q}\Psi \\
\hline
s = 1 & 0 & 0 & 0 &\textcolor{red}{\bigoplus_{i=1}^{124} \mathbf{Q}1_{i}} \\
\hline
\hline
s = 0 & 0 & H^{0}(\overline{X})& \textcolor{red}{\bigoplus_{i=1}^{124} \mathbf{Q}1_{i}} & H^{0}(\overline{X})\\
\hline
\hline
\EIbidg{2}{2}{r,s}{X} & r=-1 & r=0 & r=1 & \HI{2}{s}{X} \\
\hline
\end{array}
\]
\printbibliography
\end{document}
|
\section{Introduction}
\label{sec:introduction}
\noindent Multimedia information retrieval on a large scale database has to address
at the same time both issues related to effectiveness and efficiency.
Search results should be pertinent to the submitted queries, and
should be obtained quickly, even in presence of very large multimedia
archives and simultaneous query load.
Vectors of Locally Aggregated Descriptors (VLAD) \cite{jegou12}
were recently proposed as a way of producing compact
representation of local visual descriptors, as for instance SIFT
\cite{lowe04}, while still retaining high level of accuracy. In
fact, experiments, demonstrated that VLAD accuracy is higher than
Bag of Words (BoW) \cite{sivic03}. The advantage of BoW representation is that
it is very sparse and allows using inverted files to also achieve
high efficiency. VLAD representation is not sparse, so general
indexing methods for similarity searching \cite{metric-book} must
be used, which are typically less efficient than inverted files.
One of the best performing generic methods for similarity
searching, is the use of permutation based indexes
\cite{Navarro2008,mtap12}. Permutation based indexes rely on
the assumption that to objects that are very similar, ``see'' the
space around them in a similar way. This assumption is exploited
by representing the objects as the ordering of a fixed set of
reference objects (or pivots), according to their distance from
the objects themselves. If two objects are very similar, the two
corresponding ordering of the reference objects will be similar as well.
However, measuring the similarity between objects using the
similarity between permutations is a coarse approximation. In
fact, in order to achieve also high accuracy, similarity between
permutations is used just to identify an appropriate set of
candidates, which is then reordered according to the original
similarity function to obtain the final result. This reordering
phase, contributes to the overall search cost.
Given that objects are represented as ordering (permutations) of
reference objects, permutation based indexes offer the possibility of using
inverted files, in every similarity searching problem, where
distance functions are metric functions. In fact, \cite{ecdl10}
presents an approach where the Lucene text search engines, was
used to index and retrieve objects by similarity. The technique is
based on an encoding of the permutations by means of a
\emph{Surrogate Text Representation} (STR). In this respect, VLAD
can be easily indexed using this technique, as discussed in
\cite{Amato:2014:IVL:2578726.2578788} so that efficient and
effective image search engines can be built on top of a standard text
search engine.
In this paper, we propose an advancement on this basic techniques,
which exploits the internal structure of VLAD. Specifically, the
STR technique is applied, independently, to portions of the entire
VLAD. This leads, at the same time, to higher efficiency and
accuracy without the need of executing the reordering of the set of
candidates, which was mentioned above. The final result is
obtained by directly using the similarity between the permutations
(the textual representation), so saving both time in the searching
algorithms, and space, since the original VLAD vectors no longer
need to be stored.
The paper is organized as follows. Section \ref{sec:related-work}
makes a survey of the related works. Section \ref{sec:vlad} provides a brief introduction to the VLAD approach. Section \ref{sec:proposed
approach} introduces the proposed approach. Section
\ref{sec:experiments} discusses the validation tests. Section
\ref{sec:conclusions} concludes.
\section{Related Work}
\label{sec:related-work} \noindent
\noindent
In the last two decades, the breakthroughs in the field of image retrieval have been mainly based on the use of the local features.
Local features, as SIFT \cite{lowe04} and SURF
\cite{bay06}, are visual descriptors of selected interest points of an image. Their use allows one to effectively match local structures between images.
However, the costs of comparison of the local features lay some limits on large scale, since each image is represented by typically thousands
of local descriptors.
Therefore, various methods for the
aggregation of local features have been proposed.
One of the most popular aggregation method is the \emph{Bag-of-Word} (BoW), initially proposed in \cite{sivic03,csurka04} for matching object in videos. BoW uses a \textit{visual vocabulary} to quantize the local descriptors extracted from images; each image is then represented by a histogram of occurrences of visual words.
The BoW approach used in computer vision is very similar to the BoW used in natural language processing and information retrieval \cite{salton86}, thus many text indexing techniques, as inverted files \cite{witten99}, have been applied for image search.
From the very beginning \cite{sivic03} words reductions techniques have been used and images have been ranked using the standard \emph{term frequency-inverse document frequency} ({tf-idf}) \cite{salton86} weighting.
In order to improve the efficiency of BoW, several approaches for the reduction of visual words have been investigated \cite{thomee10,amato13:onReducing}.
Search results obtained using BoW in CBIR (Content Based Image Retrieval) has also been improved by exploiting additional geometrical information \cite{philbin07,perdoch09,tolias11:SpeededUp,zhao13} and applying re-ranking approaches \cite{philbin07,jegou08,chum07,tolias13:queryExp}.
The baseline BoW encoding is affected by the loss of information about the original descriptors due to the quantization process. For example, corresponding descriptors in two images may be assigned to different visual words. To overcome the quantization loss, more accurate representation of the original descriptors and alternative encoding techniques have been used, such as \emph{Hamming Embedding} \cite{jegou08,Jegou2010}, \emph{soft-assignment} \cite{philbin08,vanGemert08,vanGemert10}, \emph{multiple assignment} \cite{Jegou2010,jegou10:AccurateImage}, \emph{locality-constrained linear coding} \cite{wang10}, \emph{sparse coding} \cite{yang09,boureau10} and the use of \emph{spatial pyramids} \cite{lazebnik06}.
Recently, other aggregation schemes, such as the \emph{Fisher Vector} (FV) \cite{perronnin07,jaakkola98} and the \emph{Vector of Locally Aggregated Descriptors} (VLAD) \cite{JegouCVPR2010}, have attracted much attention because of their effectiveness in both image classification and large-scale image search. Both FV and VLAD use some statistics about the distribution of the local descriptors in order to transform an incoming set of descriptors into a fixed-size vector representation.
The basic idea of FV is to characterize how a sample of descriptors deviates from an average distribution that is modeled by a parametric generative model. The Gaussian Mixture Model (GMM) \cite{mclachlan2000}, estimated on a training set, is typically used as generative model and might be understood as a ``probabilistic visual vocabulary''.
While BoW counts the occurrences of visual words and so takes in account just 0-order statistics, the VLAD approach, similarly to BoW, uses a visual vocabulary to quantize the local descriptors of an image. The visual vocabulary is learned using a clustering algorithm, as for example the $k$-means. Compared to BOW, VLAD exploits more aspects of the distribution of the descriptors assigned to a visual word. In fact, VLAD encodes the accumulated difference between the visual words and the associated descriptors, rather than just the number of descriptors assigned to each visual word.
As common post-processing step VLAD is power and L2 normalized \cite{jegou12,Perronnin2010}. Furthermore, PCA
dimensionality reduction and product quantization have been
applied and several enhancements to the basic VLAD have been
proposed \cite{arandjelovic13:allAbVALD,Perronnin2010,chen11,delhumeau13,zhao13}
In this work, we will focus on VLAD which is very similar to FV. In fact VLAD has been proved to be a simplified non-probabilistic version of FV that performs very similar to FV \cite{jegou12}.
However, while BoW is a sparse vector of occurrence, VLAD is not. Thus, inverted files cannot be directly applied for indexing and Euclidean Locality-Sensitive Hashing \cite{Datar:2004} is, as far as we know, the only technique tested with VLAD.
Many other similarity search indexing techniques \cite{metric-book} could be applied to VLAD.
A very promising direction is Permutation-Based Indexing \cite{Navarro2008,mtap12,MiPai}. In particular the MI-File
allows one to use inverted files to perform similarity search with an
arbitrary similarity function.
Moreover, in \cite{ecdl10,AmatoCBMI2011}
a Surrogate Text Representation (STR) derived from the MI-File has been proposed.
The conversion of the image description in textual form enables us to exploit
the off-the-shelf search engine features with a little implementation effort.
In this paper, we extend the STR approach to deal with the VLAD descriptions
comparing both effectiveness and efficiency with the STR baseline approach, which has been studied in \cite{amato13}.
The experimentation was carried out on the same hardware and software infrastructure using
a publicly available INRIA Holidays \cite{jegou08} dataset and comparing the effectiveness with the sequential scan.
\section{Vector of Locally Aggregated Descriptors (VLAD)}\label{sec:vlad}
\noindent
The VLAD representation was proposed in \cite{Jegou2010}. As for
the BoW, a visual vocabulary, here called \emph{codebook},
$\{\bm{\mu}_1, \ldots,
\bm{\mu}_{\mathcal{K}}\}$\footnote{Throughout the paper bold
letters denote row vectors.} is first learned using a cluster
algorithm (e.g. $k$-means). Each local descriptor $\mathbf{x}_t$
is then associated with its nearest visual word (or
\emph{codeword}) $NN(\mathbf{x}_t)$ in the codebook. For each
codeword the differences between the sub-vectors $\mathbf{x}_t$
assigned to $\bm\mu_i$ are accumulated:
$$
\mathbf{v}_i=\sum_{\mathbf x_t:NN(\mathbf x_t)=i} \mathbf x_t - \bm{\mu}_i
$$
The VLAD is the concatenation of the accumulated sub-vectors, i.e.
$\mathbf V=(\mathbf v_1, \ldots, \mathbf v_\mathcal{K})$. Throughout the paper, we refer to the accumulated sub-vectors $\mathbf{v}_i$ simply as ``sub-vectors''.
Two normalization are performed: first, a power normalization with power $0.5$; second, a
L2 normalization. After this process two descriptions can be compared using the inner product.
The observation that descriptors are relatively sparse and very
structured suggests performing a principal component analysis
(PCA) to reduce the dimensionality of the VLAD. In this work, we
decide not to use dimensionality reduction techniques because we
will show that our space transformation approach is independent
from the original dimensionality of the description. In fact, the
STR approach that we propose, transforms the VLAD description in a
set of words from a vocabulary that is independent from the
original VLAD dimensionality.
\section{Surrogate Text Representation for VLAD Vectors}
\label{sec:proposed approach} \noindent
In this paper, we propose to index VLAD using
a text encoding that allows using any text retrieval engine to
perform image similarity search.
As discussed later, we implemented this idea on top of
the Lucene text retrieval engine\footnote{http://lucene.apache.org}.
To this end, we extend the permutation-based approach developed by Chavez et al. \cite{Navarro2008} to deal with the internal representation of the VLAD vectors. In this section, we first introduce the basic principle of the permutation-based approach and then describe the generalization to VLAD vectors.
\subsection{Baseline Permutation-based Approach and Surrogate Text Descriptor}
The key idea of the Permutation-based approach relies on the
observation that if two objects are near one another, they have a
similar view of the objects around them. This means that the
orderings (permutations) of the surrounding objects, according to
the distances from the two objects, should be similar as well.
Let $\mathcal{D}$ be a domain of objects (features, points, etc.),
and $d:\mathcal{D}\times\mathcal{D}\rightarrow \mathbb{R}$ a
distance function able to assess the dissimilarity between two
objects of $\mathcal{D}$. Let $R\subset\mathcal{D}$, be a set of
$m$ \emph{distinct} objects (reference objects), i.e.,
$R=\{r_1,\ldots,r_m\}$. Given any object $o \in \mathcal{D}$, we
denote the vector of rank positions of the reference objects,
ordered by increasing distance from $o$, as $\mathbf{p}(o) =
(p_1(o),\ldots,p_m(o))$. For instance, if $p_3(o) = 2$ then $r_3$
is the 2nd nearest object to $o$ among those in $R$. The essence
of the permutation-based approach is to allow executing similarity
searching exploiting distances between permutations in place of
original objects' distance. This, as discussed in the following, has the
advantage of allowing using a standard text retrieval engine to
execute similarity searching.
There are several standard methods for comparing two ordered lists, such as Kendall's tau distance, Spearman Footrule distance, and Spearman Rho distance. In this paper, we concentrate our attention on the latter distance, which is also used in \cite{Navarro2008}. The reason of this choice (explained later on) is tied to the way standard search engines process the similarity between documents and query.
In particular, we exploit a generalization of the Spearman Rho distance that allows us to compare two top-$k$ ranked lists. Top-$k$ list is a particular case of a partial ranked list, which is a list that contains rankings for only a subset of items. For top-$k$ lists, we can use a generalization of the Spearman Rho distance $\tilde{d}(o,q)$, called \textit{location parameter distance} \cite{FAGIN2003b}, which assigns a rank $k+1$ for all items of the list that have rank greater than $k$.
In particular, let $k$ be an integer less or equal than $m$, and $\mathbf{p}^k(o)=(p^k_1(o),\ldots,p^k_m(o))$ the vector defined as follows:
\begin{equation}
{{p}_i^{k}}(o)=\left\{ \begin{matrix}
p_i(o)\text{ if }p_i(o)\le k\text{ } \\
k+1\text{ if }p_i(o)>k \\
\end{matrix}. \right.\label{eq:pk}
\end{equation}
Given two top-$k$ ranked lists with $k=k_q$ and $k=k_x$, we define the approximate distance function $\tilde{d}(o,q)$ as follows:
\begin{equation} \label{eq:d} \tilde{d}(o,q)=||\mathbf{p}^{k_x}(o)-\mathbf{p}^{k_q}(q)||_2, \end{equation}
where $k_q$ is used for queries and $k_x$ for indexing. The reason for using two different $k$ relies on the fact the performance of the inverted files is optimal when the size of the queries are much smaller than the size of documents. Therefore, we will typically require that $k_q \leq k_x$.
Since, the square root in Eq. (\ref{eq:d}) is monotonous, it does not affect the ordering \cite{FAGIN2003b}, so we can safely use $\tilde{d}(o,q)^2$ instead of its square-root counterpart:
\begin{equation} \begin{array}{ll}\label{eq:d2} \tilde{d}(o,q)^2 = \sum^m_{i=1}{{\left(p^{k_x}_i(o)-p^{k_q}_i(q)\right)}^{2}} =\\
||\mathbf{p}^{k_x}(o)||_2^2 + ||\mathbf{p}^{k_q}(q)||_2^2 - 2 \,\mathbf{p}^{k_x}(o)\cdot \mathbf{p}^{k_q}(q) \end{array}\end{equation}
Figure \ref{Fig:fig-trans-examp} exemplifies the transformation
process. Figure \ref{Fig:fig-trans-examp}a sketches a number of
reference objects (black points), objects (white points), and
a query object (gray point). Figure \ref{Fig:fig-trans-examp}b
shows the encoding of the data objects in the transformed space.
We will use this illustration as a running example throughout the remainder of
the paper.
\begin{figure*}[t] \centering
\includegraphics[ width=14cm]{fig-trans-examp_mod_bold.pdf}
\caption{Example of perspective based space transformation. a)
Black points are reference objects; white points are data objects;
the grey point is a query. b) Encoding of the data objects in the
transformed space. c) Encoding of the data objects in textual form.
}
\label{Fig:fig-trans-examp}
\end{figure*}
So far, we have presented a method for approximating the function $d$. However, our primary objective is to implement the function $\tilde{d}(o,q)$ in an efficient way by exploiting the built-in cosine similarity measure of standard text-based search engines based on vector space model. For this purpose, we associate each element $r_i\in R$ with a unique key $\tau_i$. The set of keys $\{\tau_1,\ldots,\tau_m\}$ represents our so-called ``reference-dictionary''.
Then, we define a function $t^k(o)$ that returns a space-separated concatenation of zero or more repetitions of ${\tau }_i$ keywords, as follows:
\({{t}^{k}}(o)=\bigcup\limits_{i=1}^{m}{\bigcup\limits_{j=1}^{k+1-p_{i}^{k}(o)\;}{{{\tau }_{i}}}}\)
where, by abuse of notation, we denote the space-separated
concatenation of keywords with the union operator $\cup$. The
function $t^k(o)$ is used to generate the Surrogate Text Representation for both indexing
and querying purposes. $k$ assumes in general the values $k_x$ for
indexing and $k_q$ for querying. For instance, consider the case
exemplified in Figure \ref{Fig:fig-trans-examp}c, and let us assume
$\tau_1=$A, $\tau_2=$B, etc. The function ${t}^{k}$ with
$k_x=3$ and $k_q=2$, will generate the following outputs
\begin{align*}
{t}^{k_x}(o_1) = & \text{ ``E E E B B A''}\\
{t}^{k_x}(o_2) = & \text{ ``D D D C C E''}\\
{t}^{k_q}(q) \,\,\,\,= & \text{ ``E E A''}
\end{align*}
As can be seen intuitively, strings corresponding to $o_1$ and $q$ are more similar to those corresponding to $o_2$ e $q$, this reflects the behavior of the distance $\tilde{d}$. However, this should not mislead the reader: our proposal is not a heuristic, the distance between the strings corresponds exactly to the distance $\tilde{d}$ between the objects, as we will prove below.
As explained above, the objective now is to force a standard text-based search engine to generate the approximate distance function $\tilde{d}$. How this objective is obtained becomes obvious by the following considerations. A text based search engine will generate a vector representation of STRs generated with $t^{k_x}(o)$ and $t^{k_q}(q)$ containing the number of occurrences of words in texts. This is the case of the simple term-frequency weighting scheme. This means that, if for instance keyword ${\tau }_i$ corresponding to the reference object $r_i$ ($1\le i\le m$) appears $n$ times, the $i$-th element of the vector will assume the value $n$, and whenever ${\tau }_i$ does not appear it will be 0.
Let $\mathbf{k_x}$ and $\mathbf{k_q}$ be respectively the constant $m$-dimensional vectors, $(k_x+1,\ldots,k_x+1)$ and $(k_q+1,\ldots,k_q+1)$, then
\begin{equation}
\begin{array}{ll}
\mathbf{\widehat{p}}^{k_x}(o)=\mathbf{k_x}-\mathbf{p}^{k_x}(o)\\
\mathbf{\widehat{p}}^{k_q}(q)=\mathbf{k_q}-\mathbf{p}^{k_q}(q)
\end{array} \label{eq:minusvects}
\end{equation}
It is easy to see that the vectors corresponding to $t^{k_x}(o)$ and $t^{k_q}(q)$, are the same of $\mathbf{\widehat{p}}^{k_x}(o)$ and $\mathbf{\widehat{p}}^{k_q}(q)$, respectively.
The cosine similarity is typically adopted to determine the similarity of the query vector and a vector in the database of the search engine, and it is defined as:
\begin{equation}
si{{m}_{cos}}\left( o,q \right)=\frac{{{{\mathbf{\widehat{p}}}}^{{{k}_{x}}}(o)}\cdot{{{\mathbf{\widehat{p}}}}^{{{k}_{q}}}(q)}\text{ }\!\!~\!\!\text{ }}{\left\| {{{\mathbf{\widehat{p}}}}^{{{k}_{x}}}(o)} \right\|\text{ }\!\!~\!\!\text{ }\left\| {{{\mathbf{\widehat{p}}}}^{{{k}_{q}}}(q)} \right\|} \propto {\mathbf{\widehat{p}}}^{{k}_{x}}(o)\cdot{\mathbf{\widehat{p}}}^{{k}_{q}}(q). \label{eq:simcos}
\end{equation}
It is worth noting that ${\mathbf{\widehat{p}}}^{{k}}$ is a permutation of the $m$-dimensional vector $(1,2,\dots, k, 0,\dots, 0)$, thus its norm equals $\sqrt{k(k+1)(2k+1)/6}$. Since $k_x$ and $k_q$ are constants, the norms of vectors ${\mathbf{\widehat{p}}}^{{k}_{x}}$ and ${\mathbf{\widehat{p}}}^{{k}_{q}}$ are constants too, therefore can be neglected during the cosine evaluation (they do not affect the final ranking of the search result).
What we are now to show is that $si{{m}_{cos}}$ can be used as a function for evaluating a similarity of two objects in place of the distance $\tilde{d}$ and it possible to prove that the first one is a order reversing monotonic transformation of the second one (they are equivalent for practical aspects). This means that if we use $\tilde{d}(o,q)$ and we take the first $k$ nearest objects from a dataset $X \subset \mathcal{D}$ (i.e, from the shortest distance to the highest) we obtain exactly the same objects in the same order if we use ${sim}_{cos}\left(o,q\right)$ and take the first $k$ similar objects (i.e., from the greater values to the smaller ones).
By substituting Eq. \eqref{eq:minusvects} into Eq. \eqref{eq:simcos}, we obtain:
\begin{equation}
\begin{array}{ll}
\displaystyle{sim}_{cos}(o,q) \propto (\mathbf{k_x}-\mathbf{p}^{k_x}(o)) \cdot (\mathbf{k_q}-\mathbf{p}^{k_q}(q)) = \\
\ \\
= \mathbf{k_x} \cdot \mathbf{k_q} - \mathbf{k_x}\cdot\mathbf{p}^{k_q}(q) - \mathbf{k_q}\cdot\mathbf{p}^{k_x}(o) + \mathbf{p}^{k_x}(o)\cdot\mathbf{p}^{k_q}(q)\\
\end{array}\label{eq:simcos2}
\end{equation}
since $\mathbf{p}^{k_x}(o)$ ($\mathbf{p}^{k_q}(q)$) include all integers numbers from $1$ to $k_x$ ($k_q$) and the remaining assumes $k_x+1$ ($k_q+1$) values, the scalar product $\mathbf{k_x}\cdot\mathbf{p}^{k_q}(q)$ ($\mathbf{k_q}\cdot\mathbf{p}^{k_x}(o)$) is constant. We can substitute the first three member in Eq. (\ref{eq:simcos2}) with a constant $L(m,k_x,k_q)$, which depends only on $m$, $k_x$, and $k_q$ as follows:
\begin{equation}
{sim}_{cos}(o,q) \propto L(m,k_x,k_q) + \mathbf{p}^{k_x}(o)\cdot\mathbf{p}^{k_q}(q) \label{eq:simcos3}.
\end{equation}
Finally, combining Eq. (\ref{eq:simcos3}) with Eq. (\ref{eq:d2}), we obtain:
\begin{equation}
\begin{array}{ll}
{sim}_{cos}(o,q) \propto L(m,k_x,k_q) + \frac{1}{2}||\mathbf{p}^{k_x}(o)||_2^2 +\\
\\
\qquad + \frac{1}{2}||\mathbf{p}^{k_q}(q)||_2^2 - \frac{1}{2}\tilde{d}(o,q)^2. \\
\end{array}\label{eq:simcos4}
\end{equation}
Since $||\mathbf{p}^{k_x}(o)||$ and
$||\mathbf{p}^{k_q}(q)||$ depend only on the constants $m$, $k_x$, and
$k_q$, the Eq. (\ref{eq:simcos4}) proves that
${sim}_{cos}\left(o,q\right)$ is a monotonic transformation of
${\tilde{d}(o,q)}^2$ in the form $sim_{cos}=\alpha - \beta
\tilde{d}^2$.
To summarize, given a distance function $d$, we were
able to determine an approximate distance function $\tilde{d}$, which we transformed in a similarity measure.
We proved that this similarity measure can be obtained using the STR
and that it is equivalent from the point of view of the result ranking to $\tilde{d}$.
Note, however, that searching using directly the distance from
permutations suffers of low precision. To improve effectiveness,
\cite{mtap12} proposes to reorder the results set according to
original distance function $d$. Suppose we are searching for the
$k$ most similar (nearest neighbors) descriptors to the query. The
quality of the approximation is improved by reordering, using the
original distance function $d$, the first $c$ ($c \geq k$)
descriptors from the approximate result set at the cost of $c$
additional distance computations.
\subsection{Blockwise Permutation-based Approach}
The idea described so far uses a textual/permutation
representation of the object as whole, however, in our particular
scenario, we can exploit the fact that VLAD vector is the result
of concatenation of sub-vectors. In short, we apply
and compare the textual/permutation representation for each
sub-vector $\mathbf{v}_i$ of the whole VLAD, independently.
We refer to this approach as Blockwise Permutation-based approach.
As we will see, this approach has the advantage of providing a finer
representation of objects, in terms of permutations, so that no
reordering is needed to guarantee the quality of the search result.
In order to decrease the complexity of the approach and since
sub-vectors $\mathbf{v}_i$ are homogeneous, we use the same set of
reference objects $R=\{r_1,\ldots,r_m\}$ to represent them as
permutations taken at random from the dataset of VLAD vectors. Let
$\mathbf{v}_i$ be the $i$-st sub-vector of a VLAD sub-vector
$\mathbf{V}$, we denote by $\mathbf{p}^{k_x}(\mathbf{v}_i)$ the
corresponding permutation vector. Given two VLAD vectors
$\mathbf{V}=(\mathbf{v}_1, \ldots, \mathbf{v}_\mathcal{K})$ and
$\mathbf{W} = (\mathbf{w}_1, \ldots, \mathbf{w}_\mathcal{K})$,
and their corresponding concatenated permutation vectors
$\mathbf{O} = (\mathbf{p}^{k_x}(\mathbf{v}_1), \ldots,
\mathbf{p}^{k_x}(\mathbf{v}_\mathcal{K}))$ and
$\mathbf{Q}= (\mathbf{p}^{k_q}(w_1), \ldots,
\mathbf{p}^{k_q}(w_\mathcal{K}))$, we generalize the
Spearman Rho distance for two vectors $\mathbf{V}$ and
$\mathbf{W}$ as follows:
\begin{equation} \label{eq:d2vlad}
\begin{array}{ll}
\tilde{d}(\mathbf{V},\mathbf{W})^2 =\sum^\mathcal{K}_{i=1}\tilde{d}(\mathbf{v}_i,\mathbf{w}_i)^2 =\\
\sum^\mathcal{K}_{i=1} ||\mathbf{p}^{k_x}(\mathbf{v}_i)-\mathbf{p}^{k_q}(\mathbf{w}_i)||_2^{2} = ||\mathbf{O}-\mathbf{Q}||_2^2\\
\end{array}
\end{equation}
This generalization has the advantage of being faster to compute since it treats the concatenated permutation vector as a whole. Moreover, it does not
require square roots and it can be evaluated using the cosine.
Defining in the same way as above:
\begin{equation}
\begin{array}{ll}
\mathbf{\widehat{p}}^{k_x}(\mathbf{v}_i)=\mathbf{k_x}-\mathbf{p}^{k_x}(\mathbf{v}_i)\\
\mathbf{\widehat{p}}^{k_q}(\mathbf{w}_i)=\mathbf{k_q}-\mathbf{p}^{k_q}(\mathbf{w}_i).
\end{array} \label{eq:minusvectsvlad}
\end{equation}
By a similar procedure shown above, it is possible to prove that also in this case $sim_{cos} (\mathbf{V}, \mathbf{W})\propto \alpha - \beta\, \tilde{d}^2(\mathbf{V}, \mathbf{W}) $ holds.
In order to correctly match the transformed blockwise vectors, we need to extended the reference dictionary to distinguish the key produced from sub-vectors $\mathbf{v}_i$ with different subscript $i$. There for a set $m$ of reference objects, and $\mathcal{K}$ element in the VLAD codebook, we employ dictionary including a set of $m \times \mathcal{K}$ keys $\tau_{i,j}$ ($1\leq i \leq m$, $1\leq j \leq \mathcal{K}$).
For example, we associate, say, the set of keys A$_1$, B$_1$,... to the sub-vector $\mathbf{v}_1$, A$_2$, B$_2$,... to the sub-vector $\mathbf{v}_2$, and so on.
\subsection{Dealing with VLAD Ambiguities}
One of the well-known problems of VLAD happens when no local descriptor is assigned to a codeword \cite{peng2014}. A simple approach to this problem is produce a sub-vector of all zeros ($\mathbf{v}_i=\mathbf{0}$) but this has the disadvantage to be ambiguous since it is identical to the case in which the mean of the local descriptors assigned to a codeword is equal to the codeword itself.
Moreover, as pointed out by \cite{spyromitros2014comprehensive}, given two images and the corresponding VLAD vectors $\mathbf{V}$ and $\mathbf{W}$, and assuming that $\mathbf{v}_i=\mathbf{0}$, the contribution of codeword $\bm{\mu}_i$ to the cosine similarity of $\mathbf{V}$ and $\mathbf{W}$ will be the same when either $\mathbf{w}_i=\mathbf{0}$ or $\mathbf{w}_i\neq\mathbf{0}$. Therefore, this under-estimates the importance of jointly zero components, which gives some limited yet important evidence on visual similarity \cite{jegou2012negative}. In \cite{jegou2012negative}, this problem was treated by measuring the cosine between vectors $\mathbf{V}$ and $\mathbf{W}$ at different point from the origin.
This technique, however, did not lead to significant improvement of our experiments. To tackle this problem, we simply get rid of the sub-vectors $\mathbf{v}_i=\mathbf{0}$ and omit to transform them in text. Mathematically, this means that we assume $\mathbf{\widehat{p}}^{k_x}(\mathbf{0}) = \mathbf{0}$.
\section{Experiments}
\label{sec:experiments} \noindent
\subsection{Setup}
\label{setup}
INRIA Holidays \cite{JegouCVPR2010,jegou12} is a collection of 1,491 holiday images. The authors selected 500 queries and for each of them a list of positive results.
As in \cite{Jegou2009,Jegou2010,jegou12}, to evaluate the approaches on a large scale, we merged the Holidays dataset with the Flickr1M collection\footnote{\url{http://press.liacs.nl/mirflickr/}}.
SIFT features have been extracted by Jegou et al. for both the Holidays and the Flickr1M datasets\footnote{\url{http://lear.inrialpes.fr/~jegou/data.php}}.
For representing the images using the VLAD approach, we selected 64 reference features using \emph{k-means} over a subset of the Flickr1M dataset.
All experiments were conducted on a Intel Core i7 CPU, 2.67 GHz with 12.0 GB of RAM a 2TB 7200 RPM HD for the Lucene index. We used Lucene v4.7 running on Java 6 64 bit.
The quality of the retrieved images is typically evaluated by means of precision and recall measures. As in many other papers \cite{Jegou2009,Perronnin2010,jegou12}, we combined this information by means of the mean Average Precision (mAP), which represents the area below the precision and recall curve.
\subsection{Results}
In a first experimental analysis, we compared the performance of blockwise approach versus the baseline approach (with and without reordering) that threats the VLAD vectors as whole-objects, which was studied in \cite{Amato:2014:IVL:2578726.2578788}. In this latter approach, as explained Section \ref{sec:proposed approach}, since the performance was low, we had to reorder the best results using the actual distance between the VLAD descriptors. With this experiment, we want to show that with the blockwise approach this phase is no longer necessary, and the search is only based on the result provided by text-search engine Lucene. For the baseline approach, we used $m=$4,000 reference objects while for blockwise, 20,000. In both cases, we set $k_x = 50$, which, we recall, is the number of closest reference objects used during indexing.
Figure \ref{fig:basevsblock} shows this comparison in terms of mAP. We refer to baseline approach as STR, the baseline approach with reordering as rSTR, and to blockwise approach as BSTR. For the rSTR approach, we reordered the first 1,000 objects of the results set. The horizontal line at the top represents the performance obtained matching the original VLAD descriptors with the inner product, performing a sequential scan of the dataset, which exhibits a mAP of 0.55. The graph in the middle shows the mAP of our approach (BSTR) versus the baseline approach without reordering (STR) and with reordering (rSTR). The graphs show also how the performance changes varying $k_q$ (the number of closest reference objects for the query) from 10 to 50.
An interesting by-product of the experiment is that, we obtain a little improvement of the mAP for the BSTR approach when the number of reference objects used for the query is 20.
\begin{figure}[t]
\centering
\includegraphics[ trim=27mm 15mm 27mm 20mm, width=0.98\columnwidth]{blocwise_vs_baseline.pdf}
\caption{Effectiveness (mAP) of the various approach for the INRIA Holidays dataset, using $k_x=50$ for STR, rSTR, BSTR, and BSTR {tfidf} (higher
values mean better reults).}
\label{fig:basevsblock}
\end{figure}
A quite intuitive way of generalizing the idea of reducing the size of the query is to exploit the knowledge of the \emph{tf*idf} (i.e., term frequency * inverse document frequency) statistic of the BSTR textual representation. Instead of simply reducing the $k_q$ of the query, i.e., the top-$k_q$ element nearest to the query, we can retain the elements that exhibit greater values of \emph{tf*idf} starting from the document generated with $k_q=50$ and eliminate the others. Therefore, we take, for instance, the first 40 elements that have best \emph{tf*idf}, the first 30 elements, and so on. Figure \ref{fig:basevsblock} shows the performance of this approach, with the name `BSTR tfidf'. It is interesting to note that we had not only an important improvement of the mAP for increasing reduction of the queries but also that this approach outperforms the performance of the inner product on the original VLAD dataset.
In order to ascertain the soundness of the proposed approach, we tested it on the larger and challenger Flickr1M dataset.
\begin{figure}[t]
\centering
\includegraphics[ trim=27mm 15mm 27mm 20mm, width=0.98\columnwidth]{blocwise_vs_baseline_1M.pdf}
\caption{Effectiveness (mAP) of the various approach for the INRIA Holidays + Flickr1M dataset, using $k_x=50$ for STR, rSTR, BSTR, and BSTR tfidf. While for BSTR tfidf$^2$, we set $k_x=k_q$ (higher
values mean better reults).}
\label{fig:basevsblock1M}
\end{figure}
The results are shown in Figure \ref{fig:basevsblock1M}. We can see that BSTR tfidf is still the winner in terms of mAP. However, in this case all the techniques exhibit lower performance with respect the inner product on the original VLAD dataset. The latter test is performed as a sequential scan of the entire dataset obtaining a mAP of 0.34. The results presented in this figure also show the performance of the approach called BSTR tfidf$^2$, which consists in applying the reduction of the blockwise textual representation using \emph{tf*idf} also for the indexed document (in addition to the queries), setting $k_x = k_q$ for all the experiments. The mAPs values in this case are slightly lower than BSTR tfidf, however, as we are going to see in the next experiment there is a great advance in terms of space occupation.
In order to assess which approach is most promising, we have also evaluated the efficiency in terms of space and time overhead. Figure \ref{fig:basevsblock1Mtimes} shows the average time for a query for the proposed approaches. The rSTR approach considers also the time for reordering the result set, however, its average time is obtained using a solid state disk (SSD) disk in which the original VLAD vectors are available for the reordering. The SSD is necessary to guarantee fast random I/O, while using a standard disk the seek time would affect the query time of more than one order of magnitude.
Figure \ref{fig:basevsblock1Mspace} presents the index occupation expressed in GB. The rSTR approach occupies 16.8 GB on the disk, including the overhead for the storage of the VLAD vectors used for the reordering of the results. The BSTR tfidf$^2$ solution has great impact of the space occupation: just for a reduction of the 20\% of the documents ( i.e., from $k_x=50$ to $k_x=40$) we get a reduction of the 80\% for the inverted file.
Considering all the alternatives seen so far, an optimal choice could be BSTR tfidf$^2$ with $k_x=k_q=20$, which is efficient in term of both time and space overheads and still maintains satisfactory mAP.
\begin{figure}[t]
\centering
\includegraphics[ trim=27mm 15mm 27mm 20mm, width=0.98\columnwidth]{blocwise_vs_baseline_1M_times.pdf}
\caption{Average time per query in seconds of the various approaches for the INRIA Holidays + Flickr1M dataset, using $k_x=50$ for rSTR, BSTR, and BSTR tfidf. While for BSTR tfidf$^2$, we set $k_x=k_q$ (higher values mean worse performance).}
\label{fig:basevsblock1Mtimes}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[ trim=26mm 15mm 26mm 20mm, width=0.98\columnwidth]{blocwise_vs_baseline_1M_space.pdf}
\caption{Space occupation of the index for the different type of solutions, using the same value of $k_x=50$ for BSTR and rSTR, and varing $k_x$ for BSTR tfidf$^2$. Note that for the rSTR, we consider also the overhead for the storage of the VLAD vectors used for the reordering of the results (higher values mean greater occupations).}
\label{fig:basevsblock1Mspace}
\end{figure}
\section{Conclusions and Future Work}
\label{sec:conclusions} \noindent
In this work, we proposed a `blockwise' extension of surrogate text representation, which is in principle applicable not only to VLAD but also to any other vector or compound metric objects. The main advantage of this approach is the elimination for the need of the reordering phase.
Using the same hardware and text search engine (i.e., Lucene), we were able to compare with the state-of-the-art baseline STR approach exploiting the reordering phase.
The experimental evaluation on the blockwise extension revealed very promising performance in terms of mAP and response time. However, the drawback of it resides in the expansion of the number of terms in the textual representation of the VLADs. This produces an inverted index that, using Lucene, is one order of magnitude greater than the baseline STR. To alleviate this problem, we propose to shrink the index reducing the document, as we did for the query, by eliminating the terms associated with a low value of \emph{tf*idf} weight.
This approach is very effective but has the disadvantage that need a double indexing phase or at least a pre-analysis of the dataset in order to calculate the \emph{tf*idf} weight of the terms. Future work will investigate this aspect in more detail.
\section*{\uppercase{Acknowledgements}}
This work was partially supported by EAGLE, Europeana network of Ancient Greek and Latin Epigraphy, co-founded by the European Commision, CIP-ICT-PSP.2012.2.1 - Europeana and creativity, Grant Agreement n. 325122.
\bibliographystyle{apalike}
{\small
|
\section{Introduction}
This paper is concerned with the numerical solution of initial-value problems for a class of partial differential equations, subject to Dirac initial datum, which have the form
\begin{alignat}{2}
\label{generalpde}
\frac{\partial u}{\partial t} + c(x,t)\frac{\partial u}{\partial y} &= \mathcal{L} u, &&\qquad (x,y,t) \in \mathbb{R}^n\times\mathbb{R}\times (0,T], \\
u(x,y,0)&=\delta (x-x_0)\otimes \delta (y-y_0), &&\qquad (x,y) \in \mathbb{R}^n\times\mathbb{R},
\label{generaldelta}
\end{alignat}
where $n \geq 1$, $T>0$, $x_0 \in \mathbb{R}^n$, $y_0 \in \mathbb{R}$, the symbol $\otimes$ signifies the (associative) binary operation of tensor product of distributions, the drift coefficient $c=c(x,t)$ is independent of the variable $y \in \mathbb{R}$, $\nabla_x c(x,t) \neq 0$ for all $(x,t) \in \mathbb{R}^n \times (0,T]$, the elliptic differential operator $\mathcal{L}$ does not include any $y$-derivatives and its coefficients only depend on $x=(x_1,\dots,x_n) \in \mathbb{R}^n$ and $t \in [0,T]$; in other words, $\mathcal{L}$ is assumed to be of the form
\begin{eqnarray}
\label{Lop}
\label{Lop-gen}
\mathcal{L} = \sum_{i,j=1}^n a_{ij}(x,t) \frac{\partial^2}{\partial x_i \partial x_j} + \sum_{i=1}^n b_{i}(x,t) \frac{\partial}{\partial x_i}
+ d(x,t),
\end{eqnarray}
where $a_{ij}$, $i,j=1,\dots,n$, $b_i$, $i=1,\dots n$, and $d$ are continuous functions of $(x,t) \in \mathbb{R}^n \times [0,T]$, and there exists a constant $c_0>0$ such that
\[ \sum_{i,j=1}^n a_{ij}(x,t) \xi_i \xi_j \geq c_0
|\xi|^2 \qquad \forall\, \xi=(\xi_1, \dots, \xi_n) \in \mathbb{R}^n\quad \forall\, (x,t)
\in \mathbb{R}^n \times [0,T].
\]
Hypoelliptic problems of this kind arise naturally from statistical physics, stochastic analysis, and from mathematical finance in particular, as Kolmogorov equations that describe the evolution of the probability density function of stochastic processes, and the initial datum in such problems is frequently a point source, which is modelled by a Dirac measure concentrated at a point. Such initial conditions are clearly also relevant in the construction of Green's functions.
For the definition of hypoellipticity and sufficient and necessary conditions for $C^\infty$ regularity see \cite{hormander1967hypoelliptic}.
Contemporary applications and extensions to nonlinear problems are found in
\cite{villani2006hypocoercive,dric2009hypocoercivity}.
Given the interest in this class of equations, methods have recently been put forward for their numerical approximation,
with a focus on preserving the long-time behaviour of solutions to the original equation.
In \cite{foster2014structure}, a self-similar change of variables was performed and convergence of the numerical solution to the steady state under these new variables was established; furthermore, an operator splitting scheme based on decomposing the hypoelliptic operator
into coercive and convective terms was proposed.
In contrast, in \cite{PORRETTA} asymptotic properties of standard central finite difference schemes are analyzed and decay rates of difference quotients were proved in the case of $L^2$ initial data. The analysis in those papers does not cover the case of Dirac initial data, which are important in a number of applications, and indeed the numerical experiments at the end of this section show that in the case of Dirac initial datum convergence of a finite difference approximation to the problem based on central differences is not guaranteed in the discrete maximum norm.
The semidiscrete Fourier scheme that we propose here for the numerical approximation of problem \eqref{generalpde} involves the application of a Fourier transform to (\ref{generalpde}) in the $y$-direction ($y$-FT) to reduce the dimension of the problem by transforming it into a one-parameter family of parabolic problems, and it then applies finite difference discretization in the $x$-direction to this parametrized family of parabolic problems, followed by the application of an inverse Fourier transform. The observed exponential convergence of the numerical approximation to the inverse Fourier transform reduces the computational complexity of the proposed scheme to that of a finite difference approximation of a problem that has no dependence on $y$, as long as the solution is required for a single value of $y$ only. The application of the $y$-FT avoids the use of lower-order stable upwind or semi-Lagrangian discretizations
of the $y$-derivative. Moreover, it transforms the Dirac initial datum into a constant function in the $y$-direction, which is easier
to handle numerically. To the best of our knowledge, the numerical scheme proposed in this paper is the first provably convergent numerical method for hypoelliptic problems of this type, with Dirac initial datum.
We shall study in-depth the stylized problem
\begin{alignat}{2}
\label{toypde}
\frac{\partial u}{\partial t}+x\frac{\partial u}{\partial y}&=\frac{\partial^2 u}{\partial x^2}, &&\qquad (x,y,t) \in \mathbb{R}\times\mathbb{R}\times (0,T], \\
u(x,y,0) &= \delta (x)\otimes \delta (y), &&\qquad (x,y) \in \mathbb{R} \times\mathbb{R}.
\label{toydelta}
\end{alignat}
This is the partial differential equation originally considered by Kolmogorov in his 1934 paper \cite{KOLM34}. The significance of this simple model problem stems from the fact that it incorporates two important features: hypoellipticity and Dirac initial datum. The equation (\ref{toypde}) results
from (\ref{Lop-gen}) for $d\equiv 0$ by shifting the point $(x_0,y_0)$ at which the initial Dirac datum is concentrated to the origin, linearization of the coefficient $c$ with respect to the $y$ variable around $y=0$, translation in the $y$-direction with $c(0,0) t$, and finally freezing the coefficients $c$, $b$ and $a$ with respect to $t$. Since the final term in \eqref{Lop-gen} does not affect the ideas presented herein, for the sake of simplicity of the exposition we shall confine ourselves to the case of $d\equiv 0$.
Subject to sign change, our model equation \eqref{toypde} coincides with the one studied in \cite{PORRETTA}.
Because of the special structure of the equation \eqref{toypde}, we shall apply Fourier methods in the construction of its numerical approximation and also in the convergence analysis of the proposed numerical algorithms. We shall explore the behaviour of two numerical schemes for the solution of our model problem: a semidiscrete Fourier method and a fully-discrete Fourier method, which will be described below.
Whereas the proposed numerical techniques apply to the more general model problem \eqref{generalpde}, \eqref{generaldelta},
and, in fact, with more general probability measures as initial data than the Dirac measure considered herein,
for the sake of simplicity and clarity of the exposition the mathematical analysis of the proposed numerical methods is restricted to the
simplified model \eqref{toypde}, \eqref{toydelta}, which is a special case of the Cauchy problem \eqref{generalpde}, \eqref{generaldelta} above.
For this toy model, we shall prove the convergence of the semidiscrete Fourier method and derive expressions for the leading order terms for the global discretization error.
In particular, we shall analyze the behaviour of the error between the analytical solution and its numerical approximation and will establish the rate of convergence of
the scheme as the spatial discretization parameter $\Delta x \rightarrow 0$. It should be noted that this analysis only relates to the analytical solution of the
semidiscrete scheme (\ref{eq:SEMIDSCHEME}) and its subsequent exact $y$-FT inversion. We shall discuss these additional approximations in Section \ref{sec:fourier-sol}.
We approach the task of error analysis
by applying the inverse Fourier transform to the error between $W$ and $w$, where $W$ is the solution of the equation resulting from Fourier
transforming \eqref{toypde} with respect to $y$, discretizing with respect to $x$, and applying a discrete Fourier transform with respect to $x$, whereas $w$ is the
solution of the equation resulting from Fourier transforming \eqref{toypde} with respect to both $x$ and $y$. We then perform a wavenumber analysis of the resulting
expressions to establish convergence. This method is based on similar ideas to those in
\cite{GC} and \cite{CRAW}, where time-stepping schemes for the one-dimensional heat
equation with Dirac initial datum were analyzed. While in the cited papers the Fourier transform was used purely as a mathematical tool in the analysis of the
discretization error in the original space-time co-ordinates, here we use a partial Fourier transform (i.e., we transform in the $y$-variable only) in the construction
of the actual numerical method and use a double-Fourier transform (i.e., the Fourier transform in both the $x$ and the $y$ variable) to quantify the error of this
approximation. An interesting feature of the present analysis is the intricate interplay between the $x$- and $y$-Fourier modes, due to the hypoellipticity of the
equation.
In order to motivate the numerical method proposed in the next section, we illustrate the smoothing and convergence properties of the central difference scheme with implicit Euler time stepping,
\begin{eqnarray*}
\frac{U_{j,k}^{n+1} - U_{j,k}^{n}}{\Delta t}
+
x_j \frac{U_{j,k+1}^{n+1} - U_{j,k-1}^{n+1}}{2 \Delta y}
= \frac{U_{j+1,k}^{n+1} - 2 U_{j,k}^{n+1} + U_{j-1,k}^{n+1}}{\Delta x^2}, \qquad \mbox{with $(j,k) \in \mathbb{Z}^2$ and $n \geq 0$.}
\end{eqnarray*}
This is the scheme studied in \cite{PORRETTA}; it is shown there in particular that, for $\ell^2$ initial data, the $\ell^2$ norms of the first-order difference quotients in the $x$- and $y$-directions decay as $t^{-1}$ and $t^{-3}$, respectively, as $t \rightarrow \infty$. Closer to the situation in the present paper,
let us consider, instead, a Dirac delta concentrated at the origin in the $(x,y)$-plane. We approximate the Dirac initial datum by
\begin{eqnarray*}
U_{j,k}^0 =
\left\{
\begin{array}{cl}
\frac{1}{\Delta x\, \Delta y} & \mbox{for $(j,k)=(0,0)$,} \\
0 & \mbox{for $(j,k) \neq (0,0)$},
\end{array}
\right.
\qquad (j,k) \in \mathbb{Z}^2,
\end{eqnarray*}
which can be viewed as the mollification of the Dirac measure concentrated at the origin through convolution, in the sense of distributions, with the scaled characteristic function $\frac{1}{\Delta x\, \Delta y} \chi_{[-\Delta x/2, \Delta x/2]\times[-\Delta y/2, \Delta y/2]}$ with unit $L^1$ norm (cf. \cite{JS}). We shall consider the problem on a sufficiently large square domain $(-L,L) \times (-L,L)$ in the $(x,y)$-plane, with zero Dirichlet boundary condition along
$(\pm L, y)$ for all $y \in [-L,L]$ and along $(x,-L)$ for all $x \in [-L,L]$; in our numerical experiment below we took $L=10$, which ensures that the Dirichlet boundary condition has negligible influence on the values of the numerical solution at the final time of interest, $T=1$ in our case, close to the centre of the square where the initial Dirac delta is concentrated.
Fig.~\ref{fig:nonsmoothcentral}, left, shows the numerical solution at $T=1$, which exhibits large oscillations in the $x$-direction, but is smooth in the
$y$-direction. Indeed, good approximation to the analytical solution is visible between the oscillations. The discrete Fourier transform of the numerical solution,
with wavenumbers $s$ and $p$ (described in more detail later), is depicted in the right panel in Fig.~\ref{fig:nonsmoothcentral}. It shows low wavenumber
components in both $s$ and $p$ near the origin in Fourier space, which approximate well the Fourier transform of the analytical solution, and low $s$-/high $p$-wavenumber
components concentrated at $(s,p) = (0,\pm \pi/\Delta y) = (0,\pm 20\pi) \approx (0,\pm 63)$, which trigger the spurious oscillations in the numerical approximation of the analytical solution.
\begin{figure}[H]
\caption{Central difference scheme with Dirac initial datum; the computational domain is $x\in [-10,10]$, $y\in [-10,10]$ (smaller plot range); the numerical
solution at $T=1$ with $n_x= n_y=n_t=400$ grid spacings in the $x$-, $y$-, and $t$-directions, respectively;
left the numerical solution, right its discrete Fourier transform, with $s$ and $p$ signifying wavenumbers corresponding to the $x$ and $y$ co-ordinate
directions, respectively.
}
\begin{center}
\includegraphics[trim={0.5cm 1 1 1},clip, width=3.4in,height=2.2in]{NonSmoothNoUpwindNumSol400TimeSteps} \hfill
\includegraphics[trim={0.5cm 1 1 1},clip, width=3.4in,height=2.2in]{NonSmoothNoUpwindDFT400TimeSteps}
\end{center}
\label{fig:nonsmoothcentral}
\end{figure}
Extending the techniques from \cite{GC}, the time-discrete evolution of the discrete Fourier transform $W$ of the numerical solution $U$ is found to be governed by the recursion (see also Sections \ref{subsec:evol} and especially \ref{subsec:fullydiscr})
\begin{align*}
\left(1+4 \frac{\Delta t}{\Delta x^2} \sin^2\left(\frac{s \Delta x}{2}\right)\right) W^{n+1}(s,p) &= W^n(s,p) + \frac{\Delta t}{\Delta y} \sin(p \Delta y)
\frac{\partial W^{n+1}}{\partial s}(s,p),\qquad n \geq 0,\\
W^0(s,p) &= 1,
\end{align*}
for all $(s,p) \in [-\frac{\pi}{\Delta x}, \frac{\pi}{\Delta x}] \times [-\frac{\pi}{\Delta y}, \frac{\pi}{\Delta y}]$. Setting $s=0$ and $p=\pm \pi/\Delta y$ we deduce that
\begin{eqnarray*}
W^n(0,\pm \pi/\Delta y) = 1 \qquad \forall\, n\ge 1.
\end{eqnarray*}
This finding is in line with our numerical simulations, and suggests that the numerical solution will not converge to the analytical solution
as $\Delta x, \Delta y \rightarrow 0$. To reconcile this evidence with
the results in \cite{PORRETTA}, we compute a numerical approximation to the solution at $T=2$, starting with the exact (smooth) solution (see (\ref{toyexact})) as initial datum at $t=1$.
The numerical solution and its discrete Fourier transform are shown in Fig.~\ref{fig:smoothcentral}.
As the initial datum in this case does not have high wavenumber components, the numerical solution approximates the analytical solution well. Indeed, the maximum error is around
$1.8 \times 10^{-4}$ (solution $\approx 0.07$), compared to $0.28$ (solution $\approx 0.3$) for the Dirac case.
Replacing the central $y$-difference with an upwind $y$-difference is observed to produce a convergent sequence of numerical approximations to the analytical solution in the case of a Dirac initial datum, just as for smooth
initial data, but such a finite difference scheme is only of first-order accuracy with respect to $\Delta y$. We shall therefore propose in the next section a numerical scheme applied to the analytical $y$-Fourier transform, which does not suffer from this shortcoming.
\begin{figure}[H]
\caption{Central difference scheme with smooth initial datum: $x\in [-10,10]$, $y\in [-10,10]$, $t \in [1,2]$, $n_x = n_y = n_t = 400$; the numerical solution at $T=2$ (left) and
its discrete Fourier transform (right).
}
\begin{center}
\includegraphics[trim={0.5cm 1 1 1},clip, width=3.4in,height=2.2in]{SmoothNoUpwindNumSolAt2nt400}\hfill
\includegraphics[trim={0.5cm 1 1 1},clip, width=3.4in,height=2.2in]{SmoothNoUpwindNumDFTAt2nt400}
\end{center}
\label{fig:smoothcentral}
\end{figure}
The remainder of the article is structured as follows.
In Section \ref{sec:scheme} we introduce the semidiscrete Fourier scheme and formulate the discretization of the model problem (\ref{toypde}), (\ref{toydelta}); in Section \ref{sec:analysis} we prove that the scheme is second-order convergent in $\Delta x$ in the space-time $\ell^\infty$ norm.
Next, in Section \ref{sec:fourier-sol}, we describe aspects of the numerical implementation including
the exponential convergence in $\Delta p$, the mesh size of the Fourier variable in the $y$-direction,
and present numerical results for the model problem (\ref{toypde}), (\ref{toydelta}).
In Section \ref{sec:hedge}, we discuss the application of the method to a financial hedging problem, while Section \ref{sec:multi} is concerned
with the convergence analysis of the extension of the scheme to the case of $n=2$, i.e. the diffusion operator acts in two spatial directions,
$x=(x_1,x_2)$, while in the third spatial direction, $y$, there is only a transport term present in the equation; in other words, there is no diffusion in the $y$-direction. Our findings are summarized in the concluding section.
\section{The semidiscrete Fourier scheme}
\label{sec:scheme}
The Fourier transform in the $y$-direction (briefly, $y$-FT) is defined by
\[v(x,p,t) := \int^{\infty}_{-\infty}u(x,y,t)\,{\rm e}^{i py}\,{\rm d} y, \qquad x \in \mathbb{R},\; p \in \mathbb{R}, \; t> 0.\]
where it is tacitly understood that the function $y \in \mathbb{R} \mapsto u(x,y,t) \in \mathbb{R}$
is an element of $L^1(\mathbb{R})$ for all $(x,t) \in \mathbb{R}^n\times (0,\infty)$. The Fourier transform in the $y$-direction
of the initial Dirac measure $\delta(x-x_0) \otimes \delta(y-y_0)$ at $t=0$ is to be understood in the sense of distributions, and is equal to $\delta(x-x_0) \otimes {\rm e}^{ipy_0}$.
Application of the Fourier transform to \eqref{generalpde} yields
\[ \frac{\partial v}{\partial t} -i pc(x,t) v =\mathcal{L} v \qquad x \in \mathbb{R},\; p \in \mathbb{R}, \; t> 0,\]
which is a family of PDEs in $x$ and $t$, parametrized by $p$, for the $y$-FT: $v(x,p,t)$. We then discretize the operator $\mathcal{L}$ from (\ref{Lop})
in the $x$-direction(s) by means of a standard finite difference scheme, using equally spaced grid points, with spacing $\Delta x$, but we keep the time variable continuous
for the moment at least. This leads to a system of ordinary differential equations (ODEs) indexed by the $x$-grid points, $x_j$, and also by the Fourier
wavenumber in the $y$-direction, which we are denoting by $p$, for the function $V_j(p,t)$. After solving this system of ODEs (in practice numerically, for a finite set of grid points $x_j$ subject to an artificial/numerical Dirichlet boundary condition at the `far-field', and a finite set of $y$-wave numbers, $\{p_l\}_{l=l_{min}}^{l_{max}})$, we use the inverse $y$-FT to obtain an approximate solution, $U(x_j,y,t)$, in the original co-ordinates as
\begin{eqnarray}
\label{inv-trunc}
U(x_j,y,t) = \frac{1}{2\pi} \int^{2 t/(\Delta x)^r}_{-2 t/(\Delta x)^r}
V_j(p,t)\, {\rm e}^{-i y p} \,{\rm d} p.
\end{eqnarray}
The truncation of the $p$-integration range in the inversion (\ref{inv-trunc}) from $p \in (-\infty,\infty)$ to the truncated range $p \in (-2 t/(\Delta x)^r, 2 t/(\Delta x)^r)$, for
a suitable $r>0$, is
dictated by the practical requirement to carry out numerical integration over a finite range.
The scaling with $2t$ simplifies the numerical analysis, but of course in practice any suitable scaling would be chosen empirically.
To find the approximation $U_{j,k}(t)$ to $u(x_j,y_k,t)$ numerically, we apply a uniformly spaced and equally weighted quadrature rule to (\ref{inv-trunc}) and obtain
\begin{eqnarray}
\label{numinvft}
U_{j,k}(t) := \frac{\Delta p}{2\pi} \sum^{l_{max}}_{l=l_{min}} V_{j}(p_l,t)\,{\rm e}^{-i y_k p_l}
\end{eqnarray}
\noindent for a given $k$, $l_{max}>0$, $l_{min}<0$,
$l_{max} \Delta p = - l_{min} \Delta p = 2 t \Delta x^{-r}$. We will also denote $n_p := l_{max}-l_{min}+1$.
The numerical results will be seen to exhibit exponential convergence of the $p$-quadrature (see below and \cite{EXTTRAP}).
For an efficient implementation of (\ref{numinvft}), if the solution is needed for several values of $k$, the Fast Fourier Transform (FFT) can be used.
\medskip
We illustrate the method by applying it to the Cauchy problem
\begin{alignat}{2}
\label{toy2}
\frac{\partial u}{\partial t}+x\frac{\partial u}{\partial y}&=\frac{\partial^2 u}{\partial x^2}, &&\qquad (x,y,t) \in \mathbb{R}\times\mathbb{R}\times (0,T], \\
u(x,y,0) &= \delta (x)\otimes \delta (y), &&\qquad (x,y) \in \mathbb{R} \times\mathbb{R},
\label{toyic2}
\end{alignat}
with the aim to analyze the stability and accuracy of the numerical scheme we develop below.
Applying the $y$-FT to (\ref{toy2}) we get
\begin{alignat}{2}
\label{toyftpde}
\frac{\partial v}{\partial t}-i pxv &=\frac{\partial^2 v}{\partial x^2},
&&\qquad (x,p,t) \in \mathbb{R}\times\mathbb{R}\times (0,T], \\
u(x,p,0) &= \delta (x), &&\qquad (x,p) \in \mathbb{R} \times\mathbb{R},
\label{toyicft}
\end{alignat}
and we then discretize this one-parameter family of Cauchy problems in the $x$-direction only, using central differencing with spacing $\Delta x>0$, to obtain, for $x_j = j \Delta x$, $j \in \mathbb{Z}$,
\begin{align} \label{eq:SEMIDSCHEME}
\frac{\partial V_{j}(p,t)}{\partial t}-i p x_jV_{j}(p,t)=\frac{V_{j+1}(p,t)-2V_{j}(p,t)+V_{j-1}(p,t)}{\Delta x ^2}, \qquad p \in \mathbb{R}, \; t \in (0,T],
\end{align}
so that the function $V_{j}(p,t)$, which approximates $v(x_j,p,t)$, satisfies this equation for each $j\in \mathbb{Z}$ and for all $(p,t)
\in \mathbb{R}\times (0,T]$. The initial condition we use for this parametrized ODE system is
\begin{eqnarray}
\label{ic-semift}
V_{j}(p,0) = \left\{
\begin{array}{cl}
0 & \text{ for } j \ne 0, \\
\frac{1}{\Delta x} & \text{ for } j = 0,
\end{array}\right.
\end{eqnarray}
for all $p \in \mathbb{R}$, which approximates (\ref{toyicft}).
\section{Analysis of the numerical method for the stylized problem}
\label{sec:analysis}
In this section, we will prove the following theorem, which is one of our main results.
\begin{theorem}
\label{mainthm}
Let $u$ be the solution to the Cauchy problem \eqref{toy2}, \eqref{toyic2}, and let $U$ be given by \eqref{inv-trunc}, \eqref{eq:SEMIDSCHEME}, \eqref{ic-semift}. Then, for any $r>0$ in \eqref{inv-trunc},
\[
U(x_j,y,t) - u(x_j,y,t) \;= \; C(x_j,y,t) \, \Delta x^2 + o(\Delta x^2),\qquad j \in \mathbb{Z},\; y \in \mathbb{R},\; t>0,
\]
where
\[
C(x_j,y,t) :=
\left[\frac{t}{2 \pi^2} \left( \frac{1}{4!} \left(\frac{\partial}{\partial x}+ \frac{t}{2} \frac{\partial}{\partial y} \right)^4 +
\frac{1}{5!} \left(\frac{t}{2} \frac{\partial}{\partial y} \right)^4 \right) u\right]\Bigg|_{(x_j,y,t)}.
\]
\end{theorem}
\subsection{The time evolution of the numerical double transform}
\label{subsec:evol}
To investigate the stability and accuracy of the numerical scheme, we use techniques motivated by those in \cite{GC} and \cite{CRAW}.
Thus, given a set of values $\{f_j\}_{j \in \mathbb{Z}} \in \ell^1(\mathbb{Z})$, on a uniformly-spaced grid $\{x_j\}_{j \in \mathbb{Z}}$ of spacing $\Delta x>0$
on $\mathbb{R}$, we consider the (semi-)discrete Fourier transform
\[
\widehat{f}(s) := \Delta x\sum^{\infty}_{j=-\infty} f_j\,{\rm e}^{i sx_j},\qquad s \in \bigg[-\frac{\pi}{\Delta x}, \frac{\pi}{\Delta x}\bigg],
\]
and the inverse of this transform,
\[
f_j=\frac{1}{2\pi}\int^{\frac{\pi}{\Delta x}}_{-\frac{\pi}{\Delta x}} \widehat{f}(s)\,{\rm e}^{i sx_j} \,{\rm d} s, \qquad j \in \mathbb{Z},
\]
(see, e.g.\ \cite{larsson2008partial}).
We note that the method of analysis here is specific to equation (\ref{toy2}) and its higher-dimensional variants (see Section \ref{sec:multi}),
but the numerical algorithm itself is not, as we demonstrate in the numerical example of Section \ref{sec:hedge}.
Then, by applying the semidiscrete $x$-FT to the system of ODEs (\ref{eq:SEMIDSCHEME}), we find
\begin{align*}
\frac{\partial W(s,p,t) }{\partial t} -p\frac{\partial W(s,p,t) }{\partial s}
&= \sum^{\infty}_{j=-\infty} \frac{\partial V_{j}(p,t)}{\partial t}\, {\rm e}^{i sx_j} \Delta x - i p \sum^{\infty}_{j=-\infty}x_jV_{j}(p,t)\,{\rm e}^{i sx_j} \Delta x \\
&= \sum_{j=-\infty}^{\infty}\frac{V_{j+1}(p,t)\,{\rm e}^{i sx_j}-2V_{j}(p,t)\,{\rm e}^{i sx_j}+V_{j-1}(p,t)\,{\rm e}^{i sx_j}}{\Delta x ^2}\, \Delta x \\
&=W(s,p,t)\,\frac{4}{\Delta x ^2}\sin^2{\bigg(\frac{s\Delta x}{2}\bigg)},\qquad (s,p) \in \bigg[-\frac{\pi}{\Delta x}, \frac{\pi}{\Delta x}\bigg] \times \mathbb{R},\; t>0.
\end{align*}
As $W(s,p,0)=1$, it follows by the method of characteristics that $W(s,p,t)>0$ for all $t>0$, and therefore
\[
\frac{\partial \log{W} }{\partial t} -p\frac{\partial\log{W} }{\partial s} =\frac{4}{\Delta x ^2}\sin^2{\bigg(\frac{s\Delta x}{2}\bigg)},
\qquad (s,p) \in \bigg[-\frac{\pi}{\Delta x}, \frac{\pi}{\Delta x}\bigg] \times \mathbb{R},\; t>0.
\]
In contrast, taking the $x$-FT of (\ref{toyftpde}), the true double Fourier transform satisfies
\begin{equation*} \label{eq:w}
\frac{\partial w }{\partial t}-p\frac{\partial w}{\partial s}=-s^2w, \qquad (s,p) \in \mathbb{R}^2,\; t>0.
\end{equation*}
Since $w(s,p,0)=1$, and therefore, by the method of characteristics $w(s,p,t)>0$ for all $t>0$, we have that
\begin{equation} \label{eq:logw}
\begin{aligned}
\frac{\partial \log{w} }{\partial t}-p\frac{\partial \log{w}}{\partial s}=-s^2, \qquad (s,p) \in \mathbb{R}^2,\; t>0.
\end{aligned}
\end{equation}
Then, by defining
\[
Z(s,p,t)=\log{\left( \frac{W(s,p,t)}{w(s,p,t)} \right )},\qquad (s,p) \in \bigg[-\frac{\pi}{\Delta x}, \frac{\pi}{\Delta x}\bigg] \times \mathbb{R},\; t\geq 0,
\]
we find that
\[
\frac{\partial Z(s,p,t) } {\partial t} -p\frac{\partial Z(s,p,t) } {\partial s} = s^2- \frac{4}{\Delta x ^2}\sin^2{\bigg(\frac{s\Delta x}{2}\bigg)}, \qquad (s,p) \in \bigg[-\frac{\pi}{\Delta x}, \frac{\pi}{\Delta x}\bigg] \times \mathbb{R},\; t>0,
\]
where
\[
g(s) = s^2- \frac{4}{\Delta x ^2}\sin^2{\bigg(\frac{s\Delta x}{2}\bigg)}, \qquad s \in \bigg[-\frac{\pi}{\Delta x}, \frac{\pi}{\Delta x}\bigg],
\]
and $Z(s,p,0)=0$. We can solve for $Z(s,p,t)$ to obtain
\[
Z(s,p,t)= \frac{1}{p} \int^{s+pt}_{s} g(\sigma) \,{\rm d}\sigma, \qquad (s,p) \in \bigg[-\frac{\pi}{\Delta x}, \frac{\pi}{\Delta x}\bigg] \times \mathbb{R},\; t \geq 0.
\]
Finally, we have
\begin{equation}
\label{eq:Ww}
W(s,p,t) =w(s,p,t) \exp{\left(\frac{1}{p} \int^{s+pt}_{s} g(\sigma) \,{\rm d}\sigma\right )},\qquad (s,p) \in \bigg[-\frac{\pi}{\Delta x}, \frac{\pi}{\Delta x}\bigg] \times \mathbb{R},\; t\geq 0,
\end{equation}
an expression for the numerical double transform, $W$, in terms of the true double transform, $w$, with the exponential factor
on the right-hand side of \eqref{eq:Ww} reflecting the error introduced by the finite difference approximation in the $x$-direction.
In fact, one can solve (\ref{eq:logw}) with initial datum $w(s,p,0)=1$ to obtain
\begin{eqnarray}
\label{toydoubleft}
w(s,p,t)=\exp{\left(-s^2t-spt^2-\frac{1}{3}p^2t^3\right)}, \qquad (s,p) \in \mathbb{R}^2,\; t \geq 0.
\end{eqnarray}
A key observation is that it is more convenient to restate the solution in terms of the variables
\begin{eqnarray*}
\eta \;=\; s+\frac{pt}{2},\qquad
\xi \;=\; \frac{pt}{2},
\end{eqnarray*}
in Fourier space, a manifestation of the mixing of Fourier modes due to the lack of commutativity of the differential operators in $x$ and $y$ in (\ref{toy2}).
We will refer to these variables in Fourier space as wave numbers, and
it is only in these new variables that suitable wave number regimes can be defined. Indeed, in these new variables, we get
\begin{align}
\nonumber
W(s,p,t)
&=w(s,p,t) \exp{\left(\frac{1}{p} \int^{s+pt}_{s} \sigma^2 \,{\rm d}\sigma\right )} \exp{\left(-\frac{1}{p} \int^{s+pt}_{s} \frac{4}{\Delta x^2} \sin^2{\bigg(\frac{\sigma\Delta x}{2}\bigg)} \,{\rm d}\sigma\right )} \notag\\
\nonumber
&=w(s,p,t) \,w(s,p,t)^{-1} \exp{\left(-\frac{2}{p\Delta x^2} \int^{s+pt}_{s}(1- \cos{(\sigma\Delta x)}) \,{\rm d}\sigma\right )}\notag \\
&=\exp{ \bigg(-\frac{ 2t}{\Delta x^2}(1-\sinc{(\xi\Delta x)}\cos{(\eta\Delta x)}) \bigg)},
\qquad (s,p) \in \bigg[-\frac{\pi}{\Delta x}, \frac{\pi}{\Delta x}\bigg] \times \mathbb{R},\; t \geq 0,
\label{numdouble}
\end{align}
where sinc is the unnormalized sinc function, defined as follows (see, for example, \cite{OLVER}):
\[
\sinc x := \frac{\sin x}{x} \qquad \mbox{for $x \in \mathbb{R}\setminus\{0\}$ and $\sinc 0 :=1$}.
\]
We also have
\[
w(\eta,\xi,t)=\exp{\Big(-\eta^2t-\frac{1}{3}\xi^2t\Big)}, \qquad (\eta,\xi) \in \mathbb{R}^2,\; t \geq 0,
\]
in the new variables, $\eta$ and $\xi$.
\subsection{The wave number regimes} \label{sect:SEMIDANAL}
We decompose $\mathbb{R}^2$ into suitable wave number regimes in $\eta$ and $\xi$, for some $0<q<1$ and $r>0$,
\begin{align}
\label{omega1}
\Omega_1 &:= [ -\Delta x^{q-1},\Delta x^{q-1} ] \times [ -\Delta x^{q-1},\Delta x^{q-1} ], \\
\Omega_2 &:=
\left\{ (\eta,\xi): -\Delta x^{q-1} \le \xi \le \Delta x^{q-1},\; -\frac{\pi}{\Delta x}+\xi \le \eta \le -\Delta x^{q-1} \vee
\Delta x^{q-1} \le \eta \le \frac{\pi}{\Delta x}+\xi \right\}, \\
\Omega_3 &:=
\left\{ (\eta,\xi):
-\frac{\pi}{\Delta x}+\xi \le \eta \le \frac{\pi}{\Delta x}+\xi,\;
-\Delta x^{-r} \le \xi \le -\Delta x^{q-1} \vee \Delta x^{q-1} \le \xi \le \Delta x^{-r}
\right\}, \\
\Omega_4 &:= \mathbb{R}^2 \backslash (\Omega_1\cup \Omega_2\cup \Omega_3),
\label{omega4}
\end{align}
which are also illustrated in Fig.~\ref{fig-regimes}.
\begin{figure}
\begin{center}
\setlength{\unitlength}{0.5cm}
\begin{picture}(20,15)(0,0)
\put(15,7.5){\vector(1,0){4}}
\put(19.75,7.4){$\eta$}
\put(10,11.5){\vector(0,1){3}}
\put(10,15){\makebox(0,0){$\xi$}}
\linethickness{.4mm}
\put(9,11.5){\line(1,0){10}}
\put(6,8.5){\line(1,0){10}}
\put(4,6.5){\line(1,0){10}}
\put(1,3.5){\line(1,0){10}}
\put(9,6.5){\makebox(2,2){$\Omega_1$}}
\put(9,6.5){\line(0,1){2}}
\put(11,6.5){\line(0,1){2}}
\put(6,6.5){\makebox(2,2){$\Omega_2$}}
\put(12,6.5){\makebox(2,2){$\Omega_2$}}
\put(11.5,9){\makebox(2,2){$\Omega_3$}}
\put(6.5,4){\makebox(2,2){$\Omega_3$}}
\put(4,10){\makebox(2,2){$\Omega_4$}}
\put(15,3){\makebox(2,2){$\Omega_4$}}
\linethickness{1mm}
\put(1,3.5){\line(1,1){8}}
\put(19,11.5){\line(-1,-1){8}}
\end{picture}
\vspace{-2 cm}
\end{center}
\caption{Schematic representation of the wave number regimes defined by (\ref{omega1}) to (\ref{omega4}).}
\label{fig-regimes}
\end{figure}
By applying the inverse transforms to the error term we obtain an equation for the error in the $(x,y)$-plane for $t>0$, i.e.,
\begin{align*}
\setbraceratio{1}{3}
(2\pi)^2(U(x,y,t)-u(x,y,t)) &= \int^{2 t \Delta x^{-r}}_{-2 t \Delta x^{-r}} \int^{\frac{\pi}{\Delta x}}_{-\frac{\pi}{\Delta x}}
(W(s,p,t)-w(s,p,t))\, {\rm e}^{-i sx}\,{\rm e}^{-i py} \,{\rm d} s \,{\rm d} p \\
& \qquad - \int_{\Omega_4} w(s,p,t)\, {\rm e}^{-i sx}\, {\rm e}^{-i py} \,{\rm d} s \,{\rm d} p \\
&=
\frac{2}{t}
\int^{\Delta x^{-r}}_{-\Delta x^{-r}} \int^{\frac{\pi}{\Delta x}+\xi}_{-\frac{\pi}{\Delta x}+\xi} (W(\eta,\xi,t)-w(\eta,\xi,t))\,
{\rm e}^{-i (\eta-\xi)x}\, {\rm e}^{-i \frac{2\xi y}{t}} \,{\rm d} \eta \,{\rm d} \xi
\\
& \qquad - \frac{2}{t} \int_{\Omega_4} w(\eta,\xi,t)\, {\rm e}^{-i(\eta-\xi)x}\, {\rm e}^{-i \frac{2\xi y}{t}} \,{\rm d}\eta \,{\rm d}\xi.
\end{align*}
Here $x$ is any $x$-grid node. We now define
\begin{align*}
I_1(x,y,t) &:= \int_{\Omega_1} (W(\eta,\xi,t)-w(\eta,\xi,t))\,
{\rm e}^{-i(\eta-\xi)x}\, {\rm e}^{-i\frac{2\xi y}{t}} \,{\rm d}\eta \,{\rm d}\xi, \\
I_k(x,y,t) &:= \int_{\Omega_k} W(\eta,\xi,t)\, {\rm e}^{-i(\eta-\xi)x}\, {\rm e}^{-i\frac{2\xi y}{t}} \,{\rm d}\eta \,{\rm d}\xi, \qquad k =2,3, \\
I_4(x,y,t) &:= \int_{\mathbb{R}^2 \backslash \Omega_1} w(\eta,\xi,t)\, {\rm e}^{-i(\eta-\xi)x}\, {\rm e}^{-i\frac{2\xi y}{t}} \,{\rm d}\eta \,{\rm d}\xi.
\end{align*}
Except for the joint low wavenumber regime, we will perform separate calculations on $W$ and $w$.
We will find that all but the first term can be made exponentially small, while the first term is $O(\Delta x^2)$.
We begin by considering the joint low wavenumber analysis for $W$ and $w$.
\subsection{Joint low wavenumbers (region $\boldsymbol{\Omega}_{\mathbf{1}}$)}
We have to determine an exponent $q \in (0,1)$ such that, in $\Omega_1$,
\[
|\eta \Delta x|<\Delta x^q \rightarrow 0 \qquad \text{ and } \qquad
|\xi \Delta x|<\Delta x^q \rightarrow 0,
\]
and certain expansions can be usefully truncated. By straightforward Taylor expansion,
\begin{align*}
-\frac{2t}{\Delta x^2}(1-\sinc{(\xi\Delta x)}\cos{(\eta\Delta x)})
&= -\eta^2t-\frac{1}{3} \xi^2t + \frac{2t}{4!} \eta^4\Delta x^2 +\frac{2t}{5!} \xi^4\Delta x^2 \\
& \quad\, +\, O(\eta^6\Delta x^4)+O(\xi^6\Delta x^4).
\end{align*}
Our objective is to choose $q \in (0,1)$ so that the remainder terms, resulting from approximating the integrand of $I_1$ by its Taylor series expansion, are $o(\Delta x^2)$. To this end, we write
\begin{align}
\nonumber
I_1(x,y,t)
&= \int_{\Omega_1}
w\, \bigg (\exp{\left(\frac{2t}{4!} \eta^4\Delta x^2 +\frac{2t}{5!} \xi^4\Delta x^2+o(\Delta x^2)\right)} -1 \bigg)\, {\rm e}^{-i(\eta-\xi)x}
\,{\rm e}^{-i\frac{2\xi y}{t}} \,{\rm d} \eta \,{\rm d} \xi \\
\label{extension}
&= \Delta x^2 \int_{\Omega_1}
w
\left (\frac{2t}{4!} \eta^4\Delta x^2 +\frac{2t}{5!} \xi^4\Delta x^2\right) {\rm e}^{-i(\eta-\xi)x}\, {\rm e}^{-i\frac{2 \xi y}{t}} \,{\rm d} \eta \,{\rm d}\xi +
o(\Delta x^2) \\
\nonumber
&= \Delta x^2 \int_{\mathbb{R}^2}
w
\left (\frac{2t}{4!} \eta^4\Delta x^2 +\frac{2t}{5!} \xi^4\Delta x^2\right) {\rm e}^{-i(\eta-\xi)x} \,{\rm e}^{-i\frac{2 \xi y}{t}} \,{\rm d} \eta \,{\rm d}\xi
+o(\Delta x^2) \\
&= \Delta x^2\, F(x,y,t) +o(\Delta x^2),
\nonumber
\end{align}
where
\begin{align*}
F(x,y,t) &:= \int_{\mathbb{R}^2}
w \left (\frac{2t}{4!} \eta^4 +\frac{2t}{5!} \xi^4\right) \,{\rm e}^{-i(\eta-\xi)x} \,{\rm e}^{-i\frac{2 \xi y}{t}} \,{\rm d} \eta \,{\rm d}\xi \\
&= \frac{t}{2} \int_{\mathbb{R}^2}
w \left (\frac{2t}{4!} \left(s+\frac{p t}{2} \right)^4 +\frac{2t}{5!} \left(\frac{p t}{2} \right)^4\right) \,{\rm e}^{-i s x} \,{\rm e}^{-i p y} \,{\rm d} s \,{\rm d} p \\
&= t^2 \left( \frac{1}{4!} \left(\frac{\partial}{\partial x}+ \frac{t}{2} \frac{\partial}{\partial y} \right)^4 +
\frac{1}{5!} \left(\frac{t}{2} \frac{\partial}{\partial y} \right)^4 \right) u.
\end{align*}
We are able to replace $\Omega_1$ by $\mathbb{R}^2$ in (\ref{extension}) to $o(\Delta x^2)$ because $w$ decays exponentially
in $\eta$ and $\xi$ (see also Section \ref{sec:loc}). The last step uses the relation between the Fourier transform of a smooth function and the transforms of its derivatives.
We also require
\[ \Delta x^2( \eta^8+\xi^8 ) \rightarrow 0 \]
for the remainders in the Taylor expansion of the exponential to be $o(\Delta x^2)$. We can therefore take $q \in (\frac{3}{4}, 1)$ to define the joint low wavenumber regime.
Having dealt with the region $\Omega_1$ in the $(s,p)$-plane, we move on to consider the remaining terms involving $W$. There are two cases to discuss, corresponding
to low $\xi$-wavenumbers and high $\eta$-wavenumbers (region $\Omega_2$), and to high $\xi$-wavenumbers (region $\Omega_3$).
\subsection{Low $\boldsymbol{\xi}$-wavenumbers and high $\boldsymbol{\eta}$-wavenumbers (region $\boldsymbol{\Omega}_{\mathbf{2}}$)}
\label{low-high}
For these wavenumbers we have
$\eta\Delta x \in [-\pi+\xi\Delta x,\pi+\xi\Delta x]$ and in this interval $\eta=0$ is the only solution to $\cos{(\eta\Delta x)}=1$
since $|\xi \Delta x| \le \Delta x^{q}$. The following inequality is valid for $\theta \in [-\pi, \pi]$:
\[
\sin^2{ \bigg(\frac{\theta}{2} \bigg)} \ge \bigg(\frac{\theta}{\pi} \bigg)^2.
\]
However since we wish to take $\theta = \eta\Delta x$ with $\theta \in [-\pi+\xi\Delta x,\pi+\xi\Delta x]$ and $|\xi \Delta x| \le \Delta x^{q}$, we shall use instead
the weaker inequality
\[
\sin^2{ \bigg(\frac{\theta}{2} \bigg)} \ge \frac{1}{4}\bigg(\frac{\theta}{\pi} \bigg)^2,
\]
which is valid for all such $\theta$, provided that $\Delta x$ is sufficiently small (whereby also $|\xi \Delta x| \le \Delta x^{q}$ is sufficiently small).
Hence,
\[
\cos{(\eta\Delta x)} = 1- 2\sin^2{ \bigg(\frac{\eta\Delta x}{2} \bigg)} \le 1-\frac{1}{2} \bigg (\frac{\eta\Delta x }{\pi} \bigg)^2 = 1-\frac{1}{2\pi^2} (\eta\Delta x )^2.
\]
We can then choose $\alpha<\frac{1}{2\pi^2}$ such that for small enough $\Delta x$ we have
\[
\cos{(\eta\Delta x)} \leq 1-\alpha(\eta \Delta x) ^2
\]
for $\eta\Delta x \in [-\pi + \xi \Delta x, \pi + \xi \Delta x]$, since $|\xi\Delta x|\leq \Delta x^{q} \rightarrow 0$ as $\Delta x\rightarrow 0$.
Also, for $\xi\Delta x$ small enough, we can ensure that
\[
\frac{1}{2}<\sinc{(\xi\Delta x)} \le 1.
\]
So then, for $\cos(\eta\Delta x) \geq 0$,
\begin{eqnarray}
\label{high-eta-est}
\alpha(\eta\Delta x)^2 \le 1-\sinc{(\xi\Delta x)} \cos{(\eta\Delta x)}.
\end{eqnarray}
If, on the other hand, $\cos(\eta \Delta x)\leq 0$, then, because we
always have $\alpha (\eta \Delta x)^2 \le \frac{1}{2\pi^2} (\pi + \Delta x^q)^2 \le 1$ for $\Delta x$ sufficiently small (more precisely, for
$\Delta x \leq [(\sqrt{2}-1)\pi]^{1/q}$), while the
right-hand side of \eqref{high-eta-est} is $\ge 1$ for $\cos(\eta \Delta x)\leq 0$, it once again follows that \eqref{high-eta-est} holds.
Hence,
\begin{eqnarray*}
\exp{ \bigg(-\frac{2t}{\Delta x^2}( 1-\sinc{(\xi\Delta x)} \cos{(\eta\Delta x)}) \bigg)} \le
\exp{\bigg(-\frac{2t}{\Delta x^2}\alpha (\eta\Delta x)^2\bigg)}
\le \exp{(-2\alpha t\Delta x^{2(q-1)})}
\end{eqnarray*}
as $\eta\Delta x >\Delta x ^{q}$ and $\eta^2 >\Delta x ^{2(q-1)}$. Thus, the contribution to the integral satisfies
\begin{align*}
|I_2|&\le\int_{\Omega_2}W(\eta,\xi,t) \,{\rm d} \eta \,{\rm d} \xi
= o(\Delta x^m)\qquad \mbox{as $\Delta x \rightarrow 0$} \notag
\end{align*}
for all $m>0$. We therefore deduce that this contribution to the integral is exponentially small as $\Delta x \rightarrow 0$.
\subsection{High $\boldsymbol{\xi}$-wavenumbers (region $\boldsymbol{\Omega}_{\mathbf{3}}$)}
\label{sec:high}
We observe that $\xi \mapsto \sinc{(\xi\Delta x)}$ is a decreasing function
from $\xi=\Delta x ^{q-1}$ to some $\xi=\xi_0$, if we make $\Delta x$ small enough.
At that first local minimum, $\xi_0$, one then has $\tan{(\xi_0\Delta x)}=\xi_0\Delta x$ and then
$|\sinc{(\xi_0\Delta x)}|=\alpha$ with $\alpha<0.3$, say, and for values of $\xi>\xi_0$, we have $|\sinc{(\xi\Delta x)}|<\alpha$.
This is illustrated in Figure \ref{fig:SINCPLOT}, with the symbol {\footnotesize{\sf{x}}} in the figure signifying $\xi \Delta x$.
\begin{figure}[H]
\caption{Plot of $\sinc x=\frac{\sin{x}}{x}$ and its relation to other functions needed in
Section \ref{sec:high}.}
\centering
\includegraphics[width=5.5in,height=2.5in]{SINCPLOT}
\label{fig:SINCPLOT}
\end{figure}
We see that for $0<\xi\le \xi_0$ we also have (see again Figure \ref{fig:SINCPLOT})
\[
\sinc(\xi\Delta x) \cos(\eta \Delta x) \le |\sinc{(\xi\Delta x)}| <1-\frac{1}{100} (\xi\Delta x)^2,
\]
while for $\xi \ge \xi_0$ we have
\[
\sinc(\xi\Delta x) \cos(\eta \Delta x) \le | \sinc(\xi\Delta x) | \le \alpha\, ( < 0.3).
\]
So, for $\xi\Delta x \ge \Delta x^{q}$, the following inequality holds:
\[
1- \sinc(\xi\Delta x)\cos(\eta \Delta x) \ge \frac{1}{100}\min\left(\Delta x^{2q}, 100(1-\alpha) \right)\!.
\]
Therefore,
\begin{align*}
\exp{\bigg(-\frac{2t}{\Delta x^2}( 1-\sinc{(\xi\Delta x)} \cos{(\eta\Delta x)} ) \bigg)}
&\le \exp{ \bigg( -\frac{t}{50\Delta x^2}\min{(\Delta x^{2q},100 (1-\alpha))} \bigg )} \\
&\le \exp{ \bigg( -\frac{t}{50\Delta x^{2(1-q)}} \bigg)}
\end{align*}
for small enough $\Delta x$. Hence, similarly as before (cf. Section \ref{low-high}),
\begin{align*}
| I_3|
\le &( 2\Delta x^{-r} ) ( 2\pi\Delta x^{-1} ) \exp{ \bigg( -\frac{t}{50} \Delta x^{-2(1-q)} \bigg )} =o(\Delta x ^m)
\qquad \mbox{as $\Delta x \rightarrow 0$}
\end{align*}
for all $m>0$. So this contribution to the integral is also exponentially small as $\Delta x \rightarrow 0$.\\
\begin{remark}
We note that for the high $\xi$-wavenumber integration range to be nonempty, we need that
\[
\Delta x^{-(1-q)}<\xi<\Delta x^{-r} ,
\]
i.e.\ $r>1-q$. If $r\le 1-q$, there is only a low $\xi$-wavenumber regime, but this is irrelevant for the computations, and
quadratic convergence is still guaranteed for any $r \in (0,1)$ because the contributions from outside $(-\Delta x^{-r}, \Delta x^{-r})$
decay exponentially for all $r$ as $\Delta x \rightarrow 0$.
\end{remark}
\begin{remark}
For any fixed $\Delta x$, the numerical double Fourier transform $W$ in (\ref{numdouble}) approaches a constant value of $\exp(-2t/\Delta x^2)$ as
$\xi \rightarrow \pm \infty$ (equivalently, $p \rightarrow \pm \infty$) independent of $\eta$ (or $s$), hence
\[ \lim_{p \rightarrow \pm \infty} V_j(p,t) = \lim_{p \rightarrow \pm \infty} \frac{1}{2\pi} \int_{-\frac{\pi}{\Delta x}}^{\frac{\pi}{\Delta x}}
W(s,p,t)\,{\rm e}^{i s x_j} \,{\rm d} s = \left\{ \begin{array}{cc} \frac{1}{\Delta x}\,\exp(-2t/\Delta x^2) & \mbox{if $j=0$},\\
0 & \mbox{if $j \neq 0$}. \end{array} \right.\]
This implies that for $j=0$, the integrand appearing in (\ref{inv-trunc}), as a function of $p$, does not belong to $L^1(\mathbb{R})$,
which is yet another reason why instead of using the actual inverse Fourier transform in (\ref{inv-trunc}) to define $U(x_j,y,t)$, with
$\mathbb{R}$ as integration range, we have integrated over the compact interval
$[-2t \Delta x^{-r}, 2t \Delta x^{-r}]$. However, because $V_0(p,t)$ rapidly decays to zero with $\Delta x \rightarrow 0$ for $t>0$ as $p \rightarrow \pm \infty$, the sequence of integrals over $[-2t\Delta x^{-r},2t\Delta x^{-r}]$ converges to the true inverse Fourier transform, as $\Delta x \rightarrow 0$.
\end{remark}
\subsection{Localization of the exact Fourier transform (region $\boldsymbol{\Omega}_{\mathbf{4}}$)}
\label{sec:loc}
Finally, we need to estimate the error contribution for the remaining terms, which involve $w$.
We get
\begin{eqnarray*}
|I_4|\le \int^{\infty}_{-\infty} \int^{-\Delta x^{q-1}}_{-\infty} \!\! w \, \,{\rm d}\eta \,{\rm d}\xi
+ \int^{\infty}_{-\infty} \int_{\Delta x^{q-1}}^{\infty} \!\! w \, \,{\rm d}\eta \,{\rm d}\xi
+ \int^{\infty}_{\Delta x^{q-1}}\int^{\infty}_{-\infty} \!\! w \, \,{\rm d}\eta \,{\rm d}\xi
+ \int_{-\infty}^{-\Delta x^{q-1}} \!\!\! \int^{\infty}_{-\infty} \!\! w \, \,{\rm d}\eta \,{\rm d}\xi.
\end{eqnarray*}
As we have
\[
\int^{\infty}_{-\infty} w(\eta,\xi,t)\,{\rm d}\eta =\exp{ \bigg(-\frac{1}{3} \xi^2 \bigg)} \int^{\infty}_{-\infty} \exp{(-
\eta^2)} \,{\rm d}\eta = \sqrt{\pi} \exp{ \bigg(-\frac{1}{3} \xi^2 \bigg)}
\]
and
\[ \int^{\infty}_{-\infty} w(\eta,\xi,t)\,{\rm d}\xi =\exp{(-
\eta^2)} \int^{\infty}_{-\infty} \exp{ \bigg(-\frac{1}{3}
\xi^2 \bigg)} \,{\rm d}\xi = \sqrt{3\pi} \exp{(-
\eta^2)},
\]
it follows that
\begin{align*}
|I_4| &\le \sqrt{3\pi} \int^{-\Delta x^{q-1}}_{-\infty} \exp{(-
\eta^2)} \,{\rm d}\eta + \sqrt{3\pi}\int_{\Delta x^{q-1}}^{\infty} \exp{(-
\eta^2)} \,{\rm d}\eta \\
& \quad\, + \sqrt{\pi} \int^{\infty}_{\Delta x^{q-1}} \exp{ \bigg(-\frac{1}{3}
\xi^2 \bigg)} \,{\rm d}\xi + \sqrt{\pi} \int_{-\infty}^{-\Delta x^{q-1}} \exp{ \bigg(-\frac{1}{3}
\xi^2 \bigg)} \,{\rm d}\xi .
\end{align*}
Lemma 3 in the Appendix of \cite{GILES} implies that each of these integrals is $o(\Delta x^m)$,
for any $m>0$, and so $|I_4|$ is exponentially small as $\Delta x \rightarrow 0$.
Collecting the above results we find that
\begin{align*}
U(x,y,t)-u(x,y,t)
&= \frac{1}{2 t \pi^2} \left( I_1(x,y,t) + I_2 + I_3 - I_4 \right) \\
&= \frac{1}{2 t \pi^2} \left( I_1(x,y,t) + o(\Delta x^m) \right) \\
&= \Delta x^2 \left[\frac{t}{2 \pi^2} \left( \frac{1}{4!} \left(\frac{\partial}{\partial x}+ \frac{t}{2} \frac{\partial}{\partial y} \right)^4 +
\frac{1}{5!} \left(\frac{t}{2} \frac{\partial}{\partial y} \right)^4 \right) u\right]\Bigg|_{(x,y,t)} + o(\Delta x^2) \quad \mbox{as $\Delta x \rightarrow 0$},
\end{align*}
which completes the proof of Theorem \ref{mainthm}.
\section{Computation of the semidiscrete Fourier solution}
\label{sec:fourier-sol}
In this section we present the results of applying the semidiscrete Fourier scheme to the toy model \eqref{toypde}, \eqref{toydelta}.
We compute the solution to (\ref{eq:SEMIDSCHEME}), (\ref{ic-semift}) by using the matrix exponential
and we then use the trapezium rule
and the Fast Fourier Transform (FFT) to compute the values of $U(x_j,y_k,T)$ from (\ref{inv-trunc}).
\subsection{Solving the ODEs}
The ODE system (\ref{eq:SEMIDSCHEME}) can be written as
\begin{align} \label{eq:SEMID}
\frac{\rm{d}}{{\rm d}t} {V}(p,t) =M {V}(p,t),
\end{align}
where $V=(V_j)_{j}$, $M=M_1+ipM_2$ is a bi-infinite matrix, with
\[M_1=\frac{1}{\Delta x^2}
\arraycolsep=1.6pt\def1{1}
\left(\begin{matrix}
\ldots & \ldots & \ldots & \ldots & \ldots & \ldots & \ldots &\\
\ldots & -2 & ~~1 & ~~0 & ~~0 & ~~0 & \ldots\\
\ldots & ~~1 & -2& ~~1 & ~~0 & ~~0 & \ldots\\
\ldots & ~~0 & ~~1 & -2& ~~1 & ~~0 & \ldots\\
\ldots & ~~0 & ~~0 &~~1 & -2 & ~~1 & \ldots\\
\ldots & ~~0 & ~~0 & ~~0 & ~~1 & -2 & \ldots \\
\ldots & \ldots & \ldots & \ldots & \ldots & \ldots & \ldots & \\
\end{matrix}\right)
}, \qquad
M_2=\Delta x
\arraycolsep=1.6pt\def1{1}
\left(\begin{matrix}
\ldots & \ldots & \ldots & \ldots & \ldots & \ldots & \ldots &\\
\ldots & -2& ~~0 & ~~0 & ~~0 & ~~0 & \ldots\\
\ldots & ~~0 & -1 & ~~0 & ~~0 & ~~0 & \ldots\\
\ldots & ~~0 & ~~0 & ~~0 & ~~0& ~~0 & \ldots\\
\ldots & ~~0 & ~~0 & ~~0 & ~~1 & ~~0 & \ldots\\
\ldots & ~~0 & ~~0 & ~~0 & ~~0 & ~~2 & \ldots \\
\ldots & \ldots & \ldots & \ldots & \ldots & \ldots & \ldots & \\
\end{matrix}\right).
}
\]
In our implementation, the problem is considered over a sufficiently large square domain $x \in (-L, L)$ with zero Dirichlet boundary
condition at $x=\pm L$; in our numerical experiment below we took $L$ large enough ($L=10$ or $L=20$) to ensure that the Dirichlet boundary condition has negligible influence on the values of the numerical solution at the final time of interest, $T>0$. Hence, the bi-infinite matrices $M_1$, $M_2$ and
$M$ are truncated to square matrices $\widetilde{M_1}$, $\widetilde{M}_2$ and $\widetilde{M}$ of a certain finite size (depending on the choice of $L$ and $\Delta x$). Since the matrix $\widetilde{M}$ is independent of $t$, the truncated counterpart of (\ref{eq:SEMID}) has the obvious solution:
\begin{align}
\label{matexp}
\widetilde{V}(p,t)= {V}(p,0)\,{\rm e}^{\widetilde{M}t},\qquad p \in \mathbb{R}, \; t>0,
\end{align}
where $V_j(p,0)=0$ for $j \ne 0$ and $V_0(p,0)=\frac{1}{\Delta x}$.
We use the Matlab {\tt expm} function for the matrix exponential, which is based on the scaling and squaring method (cf. \cite{higham2005scaling}).
To improve this part of the algorithm, one could exploit that the eigenvalues of the symmetric matrix $\Delta x^2 \widetilde{M}$ are contained in the Gershgorin discs with centers $-2 + i j p \Delta x^3$, and radii $\leq 2$, where $|j p \Delta x^3| \le x_{\max} \Delta x^{2-r}$.
We do not pursue this further here.
The exact solution (\ref{matexp}) in the form of a matrix exponential is only available because the coefficients of \eqref{toypde}
do not depend on time.
We present results for a time stepping method in Section \ref{subsec:fullydiscr}.
The solution of (\ref{eq:SEMID}) can, in principle, be found for any value of $p$ but, in practice we will only be able to calculate a finite number of values, $\widetilde{V}_{j}(p_k,T)$.
To compute the solution in the original variables, we numerically invert the $y$-FT, using the grid values, $\widetilde{V}_{j}(p_k,T)$. By using this method, we have avoided discretizing the $y$-partial derivative: in effect, this is replaced by a discretization in the $p$-direction.
\subsection{Convergence of the $\mathbf{p}$-discretization}
When we solve the Kolmogorov forward equation (KFE) using the semidiscrete Fourier method, we need to consider the effect of the numerical
inversion of the $y$-FT in the final step of the solution procedure. As the inversion is only approximate, we have to decide how to set the
parameters for the inversion, and, more specifically, how to choose the range of $p$-wave numbers and the $p$-step.
According to the
Euler--Maclaurin expansion of the error of the composite trapezium rule applied to a sufficiently smooth function $f$ over an interval $[a,b]$
(see, for example, \cite{suli}, pp.\ 213, Theorem 7.4),
\begin{eqnarray}\label{eml}
\sum_{j=1}^k c_j h^{2j} \left[f^{(2k-1)}(b) - f^{(2k-1)}(a) \right] - \left(\frac{h}{2}\right)^{2k} \int_a^b q_{2k}(\tau(p)) f^{(2k)}(p) \ {\rm d}p,
\end{eqnarray}
where $h = (b-a)/m$ is the uniform mesh size, $\tau$ is a piecewise linear `saw tooth' function with values in $[-1,1]$, and $q_{2k}$ and $c_j$, $j=1,\dots,k$, are computable polynomials and constants, respectively.
The integrand in \eqref{inv-trunc}, up to a $p$-integrable $O(h^2)$ term, is, for $x$, $y$ and $t$ fixed,
\begin{equation}
\label{toysingleft}
f(p)=V(x,p,t) \exp(-ipy) = \frac{1}{2\sqrt{\pi t}} \exp\left(- \frac{t^3}{12} p^2 + i \frac{xt}{2} p - \frac{x^2}{4t} \right) \exp(-ipy),
\end{equation}
as one finds by taking the inverse Fourier transform (in $p$) of (\ref{toydoubleft}).
In our case, the integrand (\ref{toysingleft}) and its derivatives (in $p$) vanish so rapidly for large $p$ that by taking the integration limits to be
$a = -2t\Delta x^{-r}$ and $b = 2t\Delta x^{-r}$ the first term in \eqref{eml} becomes exponentially small in $\Delta x$ for any fixed $k$.
Then, taking $m=n_p:=l_{max}-l_{min} + 1$ in \eqref{numinvft} such that $h = \Delta p = 4t \Delta x^{-r}/n_p \rightarrow 0$, say $h=\Delta x^{r}$, the second term is $O(h^{2k})$ for any $k$.
Hence, the total error can be made exponentially small in $\Delta x$.
See \cite{EXTTRAP} for a detailed discussion of the behaviour of the trapezium rule for analytic integrands in a neighbourhood of the real line.
\subsection{Using the inverse FFT}
When inverting the $y$-FT, we use the formula (\ref{numinvft}) and we can take advantage of the low computational complexity
of the Inverse Fast Fourier Transform (IFFT). Specifically, we use the MATLAB procedure {\tt ifft}. In the MATLAB documentation \cite{MATLAB}, the IFFT
is defined by the discrete inverse Fourier transform as
\[
X_j=\frac{1}{N}\sum^N_{l=1}Y_l \omega_N^{-(j-1)(k-1)},
\]
where
\[
\omega_N= {\rm e}^{-2i\pi/N},
\]
and where both $X$ and $Y$ have length $N$. We have to relate our inversion formula to the definition of the inversion formula in MATLAB;
the details of this are described in the Appendix \ref{app:ifft}.
\subsection{Numerical experiments}
By taking the inverse Fourier transform (in $s$) of $V$ in (\ref{toysingleft}), one finds the exact solution to (\ref{toypde}),
\begin{eqnarray}
\label{toyexact}
u(x,y,t) = \frac{\sqrt{3} }{2\pi t^2 } \exp{\left(-\frac{1}{4t} x^2 \right ) } \exp{\left(-\frac{3 }{ t^3 } \left(y-\frac{t}{2} x\right)^2\right)}, \qquad (x,y) \in \mathbb{R}^2,\; t>0,
\end{eqnarray}
(see also \cite{KOLM34}), with which the numerical solution can be compared.
As $\Delta x$ is reduced we will find that we also need to reduce $\Delta p$ in order to maintain quadratic convergence. This behaviour is shown in
Figure \ref{fig:SemiDLevelPlot}.
\begin{figure}[H]
\caption{Toy model: maximum error in the semi-discrete Fourier method against the number of spatial steps. We show the effect of the choice of $n_p$ on the
accuracy of the $y$-FT and the resulting convergence in $\Delta x \rightarrow 0$.}
\centering
\includegraphics[width=5.5in,height=2.5in]{SemiDLevelPlot}
\label{fig:SemiDLevelPlot}
\end{figure}
Here, we have chosen the $x$-range as $[-10,10]$ and the $p$-range as $[-20,20]$ and we then monitor the convergence as
$\Delta x \rightarrow 0$ for a range of values of $\Delta p$, which correspond to $n_p=36, 40, 44, 48, 52$.
For large enough $n_p$, e.g.\ $n_p=52$, the plot shows quadratic convergence to the true solutions.
However, for a given value of $\Delta p$, quadratic convergence only continues to hold, as $\Delta x \rightarrow 0$, up to a critical value of $\Delta x$,
beyond which the decay of the discretization error stops.
Consequently, we need to reduce $\Delta p$ appropriately in order to achieve a satisfactory degree of accuracy.
We can also see that as $n_p$ increases towards the value where the error approaches the log-log line of quadratic convergence in $x$, the convergence in $p$ is very fast. Indeed, the horizontal asymptotes for large $n_x$, are roughly equally spaced on a log-scale when $n_p$ increases by a constant step, in line with the theoretical exponential convergence in $n_p$.
In the application in Section \ref{sec:hedge} we will study a setting where $u$ is interpreted as a joint probability density of two variables and we will be interested in the marginal density in $y$. We study this next.
The numerical solution, $U$, at time $T>0$, is calculated at grid points $(x_j,y_k)$ and is then numerically integrated over $x$ to give the marginal probability distribution for $y$.
Figure \ref{fig:SemiFComp2560}, left, plots the true and numerical marginal densities, while Figure \ref{fig:SemiFDiff2560}, right, shows the difference between the marginal
densities of the true and numerical solutions. We took $x\in[-20,20]$ and $p\in[-20,20]$ with 80 grid points in the $p$-direction and 2560 grid points in the $x$-direction.
\begin{figure}[H]
\caption{Toy model approximated by the semidiscrete Fourier method using the matrix exponential. Left, a comparison of marginal densities for $x\in[-20,20]$, $p\in[-20,20]$, $n_p=80$, $n_x=2560$. Right, the difference between solutions.}
\centering
\includegraphics[width=3.4in,height=2in]{SemiFComp2560} \hfill
\hspace{-0.5 cm}
\includegraphics[width=3.4in,height=2in]{SemiFDiff2560}
\label{fig:SemiFComp2560}
\label{fig:SemiFDiff2560}
\end{figure}
\subsection{A fully-discrete Fourier scheme}
\label{subsec:fullydiscr}
The fully-discrete Fourier scheme again uses the $y$-FT to reduce the dimension of the problem, but it then applies discretization in both the
$x$- and $t$-direction.
After solving the linear system we again invert the $y$-FT to obtain
our approximate solution, $U(x_j,y,t_l)$, for the original problem. Here $t_l=l \Delta t$ are equally spaced time steps.
We use the toy model described above to study this scheme.
We discretize (\ref{toyftpde}) in the $x$-direction, using equally spaced
grid points, to obtain (\ref{eq:SEMIDSCHEME}), and then in the $t$-direction, using the (semi)implicit Euler scheme,
to obtain
\begin{align*}
\frac{ V^{n+1}_{j}(p,t)- V^n_{j}(p,t)}{\Delta t}-ipx_jV^{n'}_{j}(p,t)=\frac{V^{n+1}_{j+1}(p,t)-2V^{n+1}_{j}(p,t)+V^{n+1}_{j-1}(p,t)}{\Delta x ^2} ,
\end{align*}
where $n'=n$ if the drift term is treated explicitly and $n'=n+1$ if it is treated implicitly. Obviously, other time discretizations are also possible, such
as second or higher order BDF, but for the sake of simplicity of the analysis to be presented we shall focus here on time stepping via the
Euler scheme.
\begin{remark}
For the purposes of an error analysis along the lines of Section \ref{sec:analysis}, we would apply the semidiscrete $x$-FT to obtain
\begin{equation*}
\begin{aligned}
\frac{W^{n+1}(s,p,t) -W^{n}(s,p,t) }{\Delta t} -p\frac{\partial W^{n'}(s,p,t) }{\partial s} =-W^{n+1}(s,p,t) \frac{4\sin^2{(\frac{s\Delta x}{2})}}{\Delta x ^2},
\end{aligned}
\end{equation*}
which can then be written as
\begin{equation}
\begin{aligned}
f(s)W^{n+1}(s,p,t) -p\Delta t\frac{\partial W^{n'}(s,p,t) }{\partial s} = W^{n}(s,p,t),
\end{aligned}
\end{equation}
where $f(s)=1+4\lambda\sin^2{(\frac{s\Delta x}{2})}$ and $\lambda=\frac{\Delta t}{\Delta x^2}$. This equation is a `differential recursion' for the double-FT in the explicit case ($n'=n$) and an `integral recursion' in the implicit case ($n'=n+1$).
Although it is possible to identify leading order terms heuristically by an expansion, the rigorous analysis of these fully discrete schemes has proved elusive due to the complexity of the error propagation over the time steps across wave-numbers. We shall therefore explore these schemes by way of numerical tests.
\end{remark}
We present the results obtained by applying the two methods above to the toy model PDE. As with the semidiscrete Fourier method, the solution for $U$ is first integrated over $x$, before being plotted against $y$. We took $x\in[-20,20]$, $p\in[-20,20]$, $t \in [0,1]$, with $n_x=1800$ grid points in the $x$-direction, $n_p=80$ grid points in the $p$-direction, and $n_t=10240$ grid points in the $t$-direction.
We first present the results obtained by treating the drift term explicitly.
The left plot of Figure \ref{fig:DiffSolutionExplicitT10240} shows the difference between the true and numerical solutions. The numerical solution itself closely resembles the result in Figure \ref{fig:SemiFComp2560}, left. \\
\begin{figure}[H]
\caption{Toy model approximated by the discrete Fourier method and an explicit approximation for the drift term. Left, the difference between the exact and approximate
solutions, with $x\in[-20,20]$, $p\in[-20,20]$, $t \in [0,1]$, $n_x=1800$, $n_p=80$, $n_t=10240$.
Right, with the other parameters the same, the error for decreasing timestep $\Delta t$ with $\Delta t/\Delta x^2$ held constant.}
\includegraphics[width=3.4in,height=2in]{DiffSolutionExplicitT10240} \hfill
\hspace{-0.5 cm}
\includegraphics[width=3.4in,height=2in]{ConvExplicitSep26}
\label{fig:DiffSolutionExplicitT10240}
\label{fig:ConvExplicitSep26}
\end{figure}
Finally, the right plot of Figure \ref{fig:ConvExplicitSep26} shows the convergence rate for the solution as the grid is refined with $\lambda=\frac{\Delta t}{\Delta x^2}$ held constant. We observe first-order convergence in time where the error is evaluated in the maximum norm. The slope of the last two points of the log-log plot is 1.00. \\
For this method, it is interesting to note that the maximum error in this case is approximately four times the corresponding error of the semidiscrete method (see Figures \ref{fig:SemiFDiff2560} and \ref{fig:DiffSolutionExplicitT10240}) and we observe upwards `bumps' at around $\pm 2$ in the latter case.
Next we present the results based on treating the drift implicitly.
The left plot in Figure \ref{fig:DiffSolutionImplicitT10240} shows the difference between the true and numerical solutions. Once again, the numerical solution itself closely resembles the result in Figure \ref{fig:SemiFComp2560}. \\
\begin{figure}[H]
\caption{Toy model approximated by the discrete Fourier method with an implicit approximation for the drift term. Left, the difference between the exact and approximate solutions,
with $x\in[-20,20]$, $p\in[-20,20]$, $t \in [0,1]$, $n_x=1800$, $n_p=80$, $n_t=10240$.
Right, with the other parameters the same, the error for decreasing timestep $\Delta t$ with $\Delta t/\Delta x^2$ held constant.
}
\includegraphics[width=3.4in,height=2in]{DiffSolutionImplicitT10240}
\hspace{-0.5 cm}
\includegraphics[width=3.4in,height=2in]{ConvImplicitSep26}
\label{fig:ConvImplicitSep26}
\label{fig:DiffSolutionImplicitT10240}
\end{figure}
Finally, the right plot in Figure \ref{fig:ConvImplicitSep26} shows the convergence rate for the solution as the grid is refined. We observe first-order convergence in time where the error is evaluated in the maximum norm. The slope of the last two points of the log-log plot is 1.00. \\
We observe that the maximum error in this case is approximately twice the corresponding error of the explicit drift case (see Figures \ref{fig:DiffSolutionExplicitT10240} and \ref{fig:DiffSolutionImplicitT10240}).
\section{Application to a financial hedging problem}
\label{sec:hedge}
In this section, we apply the Fourier method to a hypoelliptic problem from mathematical finance, namely the computation of hedging errors under misspecification of the market model. If the true model governing an underlying stock is known to the trader and the market is ``complete'', they can perfectly hedge a position in an option by dynamic trading in the stock. If the model is unknown and the hedging strategy is based on a misspecified model, a certain profit or loss will materialize and one can ask what the distribution of this hedging error is.
Our example follows a simplified version of the analysis in El Karoui \emph{et al}., \cite{ELK}, where the true dynamics of a single underlying asset satisfy the SDE
\begin{equation}
\frac{\,{\rm d} S_t}{S_t}=\mu_t \,{\rm d} t + \sigma_t \,{\rm d} W_t,
\end{equation}
where $\sigma$ is a volatility process, $\mu$ a drift, and $W$ a standard Brownian motion.
We analyze the situation where the trader instead assumes
that the underlying follows a different process, i.e. mis-specifies the model for $S$ as
\begin{equation} \label{eq:DiffVol}
\frac{\,{\rm d} S_t}{S_t}=\widehat{\mu}_t \,{\rm d} t + \widehat{\sigma} \,{\rm d} W_t
\end{equation}
for some
constant $\widehat{\sigma}$.
We define $\widehat{V}(S_t,t)$ to be the (Black-Scholes) price of the European option with payoff $F(S_T)$ based on the process defined in (\ref{eq:DiffVol}), which satisfies the PDE
\begin{eqnarray}
\label{bspde}
\mathcal{L}^{\widehat{\sigma}}\widehat{V} \equiv \frac{\partial \widehat{V}}{\partial t}+\frac{1}{2}\widehat{\sigma}^2S^2 \frac{\partial^2 \widehat{V}}{\partial S^2}+r S\frac{\partial \widehat{V}}{\partial S}-r\widehat{V} = 0.
\end{eqnarray}
If the trader then uses the sensitivity $\frac{\partial \widehat{V}}{\partial S}$ of $\widehat{V}$ as the hedge ratio,
it is shown in \cite{ELK} that the hedging ``error'' $Y_t$, i.e, the difference in time $t$ value between the option and the portfolio set up to hedge it,
discounted with the risk-free rate $r$ to the present time,
is governed by the stochastic differential equation
\begin{equation}
\,{\rm d} Y_t= \frac{1}{2} \mbox{e}^{-rt}(\widehat{\sigma}^2-\sigma_t^2)S_t^2 \frac{\partial^2\widehat{V}}{\partial S^2}\,{\rm d} t.
\end{equation}
We notice that the dynamics of $Y_t$ only consists of a drift term, i.e.\ there is no Brownian component.
In the following, we assume that $\sigma_t=\sigma$ constant, but different from $\widehat{\sigma}$.
Therefore, if we write down the Kolmogorov Forward Equation (KFE) for the probability density function $P$ of the pair $(S_t,Y_t)$ at time $t$,
we find
\begin{eqnarray}\nonumber
\frac{\partial P}{\partial t} &=&
\frac{1}{2} \sigma^2 \frac{\partial^2}{\partial S^2} \left(S^2 P\right)
- \mu \frac{\partial}{\partial S} \left(S P\right) - c \frac{\partial P}{\partial y} \\
&=& \frac{1}{2}\sigma^2 S^2 \frac{\partial^2P}{\partial S^2}+(2\sigma^2-\mu) S \frac{\partial P}{\partial S}+(\sigma^2-\mu)P -c\frac{\partial P}{\partial y},
\label{eq:KFEBlackScholes}
\end{eqnarray}
where
\begin{eqnarray}\label{eq:KFEBlackScholesDrift}
c(S,t)=\frac{1}{2}\mbox{e}^{-rt}(\widehat{\sigma}^2-\sigma^2)S^2 \frac{\partial^2\widehat{V}}{\partial S^2}.
\end{eqnarray}
The equation hence falls into the class of hypoelliptic PDEs studied in the preceding sections.
We solve the KFE numerically using the following parameters, $\sigma=0.1$, $\widehat{\sigma}=0.2$, $\mu=0.05$, and $r=0.05$. We take $T=1.0$, $S_0=125$, and the option being hedged is a so-called spread with payoff defined as
\begin{align*}
F(S_T)=\max(S_T-100,0)-\max(S_T-150,0).
\end{align*}
We note that the convexity of this option payoff changes as $S$ varies, and hence also the sign of the function $c$, which contains the second derivative of $\widehat{V}$. The function $\widehat{V}$ is known in closed form in this case as the solution to (\ref{bspde}), and so is its second derivative, referred to as the ``gamma'' in the financial community.
To find the univariate density of the discounted terminal hedging error, the joint density is integrated over the $S$-direction.
The univariate distribution of $y$ has been calculated using the fully-discrete Fourier method from Section \ref{subsec:fullydiscr}.
A Monte Carlo hedging simulation has been also performed where the asset price was simulated according to the true asset model (which used the true volatility) while the hedged portfolio was controlled by delta-hedging using the Black--Scholes equation with the erroneous volatility.
Figure~\ref{fig:ElKarouiCompNewNov2015} shows how the approximate distributions compare.
The shape of the hedging error distribution can be rationalized by noting that the positive drift, $\mu$, implies that the stock price is more likely to move into a regime where the option gamma is negative, resulting in a preponderance of negative hedging errors. Note that in these calculations, we assume that the option is sold for a price consistent with $\widehat{\sigma}$ so that $Y$ starts from zero. If some other price were initially realized, it would result in the densities being shifted by this amount.
For the fully-discrete Fourier method, we used $S \in [5,245]$, $y \in [-20,20]$ (not plotted over whole range) and $p \in [-20,20]$, $n_p=n_y=220$. For the Monte Carlo results, we used 100 time steps and 100000 paths as well as 50 bins for the approximation of the density.
\begin{figure}[H]
\caption{Black--Scholes model with mis-specified volatility. Comparison of the KFE solution, integrated over $x$, for an increasing number of spatial steps $n_S$ and keeping $n_t/n_S^2$ fixed. Shown also is the empirical density from a Monte Carlo simulation as detailed in the text.
}
\begin{center}
\includegraphics[width=5.5in,height=2.5in]{TidyPlot220Edited.eps}
\end{center}
\label{fig:ElKarouiCompNewNov2015}
\end{figure}
Having derived an approximation to the hedging error distribution, one can compute quantities of financial interest, such as the Value-at-Risk, a quantile-based risk-measure. For instance, the 10\%-quantile of the hedging error for the data set above is found to be $-4.58$. This compares to an initial option value of 29.31 and 26.67, respectively, for the low and high volatility.
\section{Multi-dimensional problems}
\label{ch:OTHER}
\label{sec:multi}
It is straightforward to extend the toy model (\ref{toypde}), (\ref{toydelta}) to higher dimensions by introducing multiple diffusive terms and/or multiple drift terms.
For the semidiscrete case, we expect that the analysis will be similar to that carried out for the one-dimensional case. We present this analysis in the remainder of this section.
\subsection{Multiple diffusive terms} \label{app:HIGHERDIMTOY}
We consider the case of a PDE with a single drift coefficient which is a linear combination of the components of $x$,
\begin{alignat}{2}
\label{general-2d}
\frac{\partial u}{\partial t}+(\gamma_1 x_1 +\gamma_2 x_2 ) \frac{\partial u}{\partial y} &=
a_{11} \frac{\partial^2 u}{\partial x_1^2} + 2 a_{12} \frac{\partial^2 u}{\partial x_1 \partial x_2} +a_{22} \frac{\partial^2 u}{\partial x_2^2},
&&\quad (x_1,x_2,y,t) \in \mathbb{R}^
\times (0,T], \\
u(x_1,x_2,y,0)&=\delta (x_1-x_{1,0})\otimes \delta (x_2-x_{2,0})\otimes \delta (y-y_0), &&\quad (x_1,x_2,y) \in \mathbb{R}^3,
\nonumber
\end{alignat}
which extends the toy model to one containing two independent variables $x_1$ and $x_2$.
We consider the hypoelliptic case, i.e., where $(\gamma_1,\gamma_2) \neq (0,0)$ and $(a_{ij})_{1\le i,j,\le 2}$ is strictly positive definite,
i.e., $a_{11}, a_{22}>0$ and $a_{11} a_{22}> a_{12}^2$.
We apply the $y$-FT to get
\begin{eqnarray}
\label{2dft}
\frac{\partial v}{\partial t}-ip(\gamma_1 x_1 +\gamma_2 x_2) v =a_{11} \frac{\partial^2 v}{\partial x_1^2}
+ 2a_{12} \frac{\partial^2 v}{\partial x_1 \partial x_2}+a_{22} \frac{\partial^2 v}{\partial x_2^2}, \quad (x_1,x_2,p,t) \in \mathbb{R}^3 \times (0,T].
\end{eqnarray}
Now for the analytical solution we next apply the $x_1$-FT and the $x_2$-FT in turn to get
\begin{eqnarray}
\label{true-double-ft}
\frac{\partial w}{\partial t}-p \bigg(\gamma_1 \frac{\partial w}{\partial s_1} +\gamma_2 \frac{\partial w}{\partial s_2} \bigg) =-(a_{11} s_1^2+2a_{12} s_1s_2+a_{22} s_2^2)w,\quad (s_1,s_2,p,t) \in \mathbb{R}^3 \times (0,T],
\end{eqnarray}
in conjunction with the initial condition
\[
w(s_1,s_2,p,0)=1.
\]
We can now solve this first-order hyperbolic initial-value problem directly, but we can also use an ansatz by seeking a solution of the form
\[
w(s_1,s_2,p,t)=\exp{(-a_{11}s_1^2t -2a_{12}s_1s_2t -a_{22}s_2^2t xs- B_1s_1pt^2-B_2s_2pt^2-Cp^2t^3)}
\]
and then finding the coefficients that fit the PDE.
By insertion we get
\begin{align*}
B_1 &= \gamma_1a_{11}+\gamma_2a_{12}, \\
B_2 &= \gamma_1a_{12}+\gamma_2a_{22}, \\
C &= \frac{1}{3}( \gamma_1(a_{11} +a_{12} )+\gamma_2(a_{12} +a_{22}) ),
\end{align*}
and thus we have the analytical solution in Fourier space. We could, if necessary, invert all these transforms but we will instead concentrate on the effect of discretization in the $x_1$- and $x_2$-directions.
So we now start with the PDE (\ref{2dft})
and discretize this as follows by writing
\begin{eqnarray*}
\frac{\partial V_{j,k}}{\partial t}-ip(\gamma_1 x_{1,j} +\gamma_2 x_{2,k})V_{j,k} &=&
a_{11} \frac{V_{j+1,k}-2V_{j,k}+V_{j-1,k}}{\Delta x_1^2} + \, a_{22} \frac{V_{j,k+1}-2V_{j,k}+V_{j,k-1}}{\Delta x_2^2} \\
&&\qquad+\, 2a_{12} \frac{V_{j+1,k+1}-V_{j+1,k-1}-V_{j-1,k+1}+V_{j-1,k-1} }{4\Delta x_1\Delta x_2},
\end{eqnarray*}
where we have omitted $p$ and $t$ as arguments for the sake of brevity, and where $V_{j,k}(p,t)$ is an approximation to $v(x_{1,j},x_{2,k},p,t)$ on a uniform
two-dimensional mesh of widths $\Delta x_1, \Delta x_2>0$.
We can then apply the Fourier transforms in the $x_1$ and $x_2$ variables and after steps similar to the one-dimensional case
obtain the solution for the double-transformed equation,
\begin{equation}
W(s_1,s_2,p,t)
=
w(s_1,s_2,p,t)
\exp{\bigg(\frac{1}{p} \int^{s_1+pt}_{s_1} ( a_{11}g_{11} (\sigma,s_2)+2a_{12}g_{12} (\sigma,s_2)+a_{22}g_{22} (\sigma,s_2) )\,{\rm d}\sigma \bigg)},
\end{equation}
where
\begin{align*}
g_{11} (s_1,s_2) &= s_1^2-\frac{4}{\Delta x_1^2}\sin^2{\bigg(\frac{s_1\Delta x_1}{2}\bigg)}, \\
g_{12} (s_1,s_2) &= s_1s_2- \frac{1}{\Delta x_1\Delta x_2}\sin{(s_1\Delta x_1)}\sin{(s_2\Delta x_2)}, \\
g_{22} (s_1,s_2) &= s_2^2-\frac{4}{\Delta x_2^2}\sin^2{\bigg(\frac{s_2\Delta x_2}{2}\bigg)}.
\end{align*}
We can now investigate the low and high wavenumber behaviour in terms of the new variables.
For the sake of simplicity of the exposition, we shall do this for the special case $\gamma_1 = 1$, $\gamma_2 = 0$, $a_{11} = a_{22} = 1$, $a_{12} = \rho \in (-1,1)$, i.e., we consider
\begin{eqnarray}
\label{special-2d}
\frac{\partial u}{\partial t} + x_1 \frac{\partial u}{\partial y} = \frac{\partial^2 u}{\partial x_1^2} + 2 \rho \frac{\partial^2 u}{\partial x_1 \partial x_2} +
\frac{\partial^2 u}{\partial x_2^2}.
\end{eqnarray}
\begin{remark}
Equation (\ref{special-2d}) is a simplified case of the general equation (\ref{general-2d}), where the term $x_2$ has been dropped
from the drift and the diffusion is normalized. This can be achieved by a rotation of the original co-ordinates to align the drift
$(\gamma_1,\gamma_2)$ with the $x_1$-axis, and by subsequent scaling of the new $y$, $x_1$ and $x_2$ co-ordinates.
\end{remark}
In this case, the numerical solution is
\begin{eqnarray*}
W(s_1,s_2,p,t)
=w(s_1,s_2,p,t)
\exp{\bigg(\frac{1}{p} \int^{s_1+p t}_{s_1} (g_{11} (\sigma,s_2)+ 2 \rho g_{12} (\sigma,s_2) + g_{22} (\sigma,s_2) )
\,{\rm d}\sigma \bigg)}.
\end{eqnarray*}
Similarly to the one-dimensional case, this leads us to investigate
\[
W(s_1,s_2,p,t)
=
W_1(s_1,p,t) \ \exp \bigg(- \frac{1}{p} \int^{s_1+p t}_{s_1}
\left(
2 \rho \frac{\sin{(\sigma\Delta x_1)}\sin{(s_2\Delta x_2)}}{\Delta x_1\Delta x_2}
+ \frac{4}{\Delta x_2^2}\sin^2{\bigg(\frac{s_2\Delta x_2}{2}\bigg)} \right) \,{\rm d}\sigma \bigg),
\]
where $W_1$ is the solution from the one-dimensional case.
By standard integration and the change of variables $\xi = pt/2$ and $\eta_1 = s_1 + pt/2$ we find this to be
\begin{eqnarray*}
W(\eta_1,s_2, \xi,t) = W_1(\eta_1,\xi,t) \, \exp \bigg(- 2 t \rho \sinc(\xi \Delta x_1) \frac{\sin(\eta_1 \Delta x_1)}{\Delta x_1} \frac{\sin(s_2 \Delta x_2)}{\Delta x_2} - \frac{4t}{\Delta x_2^2} \sin^2(s_2 \Delta x_2/2) \bigg),
\end{eqnarray*}
where
\begin{eqnarray*}
W_1(\eta_1,\xi,t) =
\exp \bigg(-
\frac{2t}{\Delta x_1^2} \left(1-\sinc(\xi \Delta x_1) \cos(\eta_1 \Delta x_1)\right) \bigg).
\end{eqnarray*}
The exact solution in these variables is
\begin{eqnarray}
\label{2d-exact}
w(\eta_1,s_2,\xi,t)=\exp{\left(- t \left\{ \eta_1^2 + \frac{\xi^2}{3} + s_2^2 + 2 \rho \eta_1 s_2 \right\}\right)}
\end{eqnarray}
and we recognize in the first two terms the one-dimensional solution
\begin{eqnarray}
\label{1d-exact}
w_1(\eta_1,\xi,t)=\exp{\left(- t \left\{ \eta_1^2 + \frac{\xi^2}{3} \right\}\right)}.
\end{eqnarray}
We proceed by a wavenumber analysis broadly similar to the one before, but made somewhat more complicated by the
presence of an extra variable $s_2$ and the fact that the problem degenerates as $|\rho|\rightarrow 1$, necessitating a
more careful estimation for $\rho$ close to 1.
\subsection{Joint low wavenumbers}
We write
\begin{eqnarray*}
\log W = \log W_1
- t \left\{
2 \rho \sinc(\xi \Delta x_1) \frac{\sin(\eta_1 \Delta x_1)}{\Delta x_1} \frac{\sin\left(s_2 \Delta x_2\right)}{ \Delta x_2}
+ \frac{4}{\Delta x_2^2} \sin^2\left(\frac{s_2 \Delta x_2}{2}\right)
\right\}
\end{eqnarray*}
and compare this to
\begin{eqnarray*}
\log w = \log w_1 - t \{s_2^2 + 2 \rho \eta_1 s_2 \},
\end{eqnarray*}
where $w$ is the exact solution to the two-dimensional problem from (\ref{2d-exact}) and $w_1$ is the solution to the one-dimensional problem
from (\ref{1d-exact}). By standard Taylor expansion we find
\begin{align*}
\log(W/w) & = \log(W_1/w_1) + 2 \rho t \left(
\frac{1}{12} \eta_1 s_2^3 \Delta x_2^2 + \frac{1}{6} \eta_1^3 s_2^2 \Delta x_1^2 + \eta_1 s_2^2 \xi^2 \Delta x_1^2
\right) + \frac{t}{12} s_2^4 \Delta x_2^2 + o(\Delta x_1^2) + o(\Delta x_2^2) \\
&= \log(W_1/w_1) + \left(\frac{\rho t}{3} \eta_1^3 s_2^2 + 2 \rho t \eta_1 s_2^2 \xi^2 \right) \Delta x_1^2 + \left(\frac{\rho t}{6} \eta_1 s_2^3 + \frac{t}{12} s_2^4 \right) \Delta x_2^2 + o(\Delta x_1^2) + o(\Delta x_2^2)\\
&=
\left(
\frac{ 2t}{4!} \eta_1^4 + \frac{2 t}{5!} \xi^4 +
\frac{\rho t}{3} \eta_1^3 s_2^2 + 2 \rho t \eta_1 s_2^2 \xi^2 \right) \Delta x_1^2 + \left(\frac{\rho t}{6} \eta_1 s_2^3 + \frac{t}{12} s_2^4 \right) \Delta x_2^2
+ o(\Delta x_1^2) + o(\Delta x_2^2).
\end{align*}
Therefore the numerical error contribution from the low wavenumber regime is
\begin{align}
\label{2derr}
I_1(x_1,x_2,y,t) &=
\Delta x_1^2 F_1(x_1,x_2,y,t)
+
\Delta x_2^2 F_2(x_1,x_2,y,t) + o(\Delta x_1^2) + o(\Delta x_2^2),
\end{align}
where
\begin{align*}
F_1 &=
t^2 \left(
\frac{1}{4!} \left(\frac{\partial }{\partial x_1} + \frac{t}{2} \frac{\partial }{\partial y} \right)^4 + \frac{1}{5!} \left(\frac{t}{2} \frac{\partial }{\partial y} \right)^4 \right) \ u\\
& \hspace{2 cm} + \ t^2 \left(
\frac{\rho}{3!} \left(\frac{\partial }{\partial x_1} + \frac{t}{2} \frac{\partial }{\partial y} \right)^3 \frac{\partial^2 }{\partial x_2^2} +
\rho \left(\frac{\partial }{\partial x_1} + \frac{t}{2} \frac{\partial }{\partial y} \right) \frac{\partial^2 }{\partial x_2^2} \left(\frac{t}{2} \frac{\partial }{\partial y} \right)^2
\right) \ u \\
F_2 &= t^2 \left(
\frac{2 \rho}{4!} \left(\frac{\partial }{\partial x_1} + \frac{t}{2} \frac{\partial }{\partial y} \right) \frac{\partial^3 }{\partial x_2^3}
+ \frac{1}{4!} \frac{\partial^4 }{\partial x_2^4}
\right) \ u.
\end{align*}
The point is less the explicit form, but that the error can be expressed in terms of up to fifth mixed partial derivatives, or
up to fourth if $\rho=0$. As in Section \ref{sec:analysis}, we will find that this is the only wavenumber range that contributes to the leading order error, and therefore, up to higher order terms, (\ref{2derr}) fully describes the discretization error.
\subsection{Low $\boldsymbol{\xi}$-wavenumbers and high $\boldsymbol{\eta_1}$- or $\bf s_2$-wavenumbers}
We require two simple inequalities. The first one states that for $|\eta_1 \Delta x_1| \le \pi/2$, and since
$\sinc(\xi \Delta x_1)>0$ (in the present small $\xi$ regime),
\[
1 - \sinc(\xi \Delta x_1) \cos(\eta_1 \Delta x_1) \ge 1 - \cos(\eta_1 \Delta x_1) = 2 \sin^2(\eta_1 \Delta x_1/2).\]
The second elementary inequality used in the argument below, in the transition from the second-to-last inequality to the last
inequality, is that
\[
\sin \alpha \ge \frac{2}{\pi} \alpha\qquad \mbox{for}\qquad 0\le \alpha \le \frac{\pi}{2}.
\]
Therefore
\begin{align*}
W(\eta_1,s_2) &\le\exp \bigg(
- \frac{4 t}{\Delta x_1^2} \sin^2(\eta_1 \Delta x_1/2) + 2 t |\rho| \frac{|\sin(\eta_1 \Delta x_1)|}{\Delta x_1}
\frac{|\sin(s_2 \Delta x_2)|}{\Delta x_2} - \frac{4 t}{\Delta x_2^2} \sin^2(s_2 \Delta x_2/2)
\bigg) \\
&\le
\exp \bigg(
- \frac{4 t}{\Delta x_1^2} \sin^2(\eta_1 \Delta x_1/2) + 8 t |\rho| \frac{|\sin(\eta_1 \Delta x_1/2)|}{\Delta x_1}
\frac{|\sin(s_2 \Delta x_2/2)|}{\Delta x_2} - \frac{4 t}{\Delta x_2^2} \sin^2(s_2 \Delta x_2/2)
\bigg) \\
&=
\exp \bigg(
- 4 t (1-|\rho|)
\left(
\frac{1}{\Delta x_1^2} \sin^2(\eta_1 \Delta x_1/2) + \frac{1}{\Delta x_2^2} \sin^2(s_2 \Delta x_2/2)
\right)
\bigg)\\
&
\hspace{5 cm} \cdot \; \exp \bigg(
- 4t |\rho| \left(
\frac{\sin(\eta_1 \Delta x_1/2) }{\Delta x_1} - \frac{\sin^2(s_2 \Delta x_2)}{\Delta x_2}
\right)^2
\bigg) \\
&\leq \exp \bigg(- 4 t (1-|\rho|) \left(
\frac{1}{\Delta x_1^2} \sin^2(\eta_1 \Delta x_1/2) + \frac{1}{\Delta x_2^2}
\sin^2(s_2 \Delta x_2/2)
\right)
\bigg) \\
&\le
\exp \bigg(
-\frac{4 \, t \, (1-|\rho|)}{\pi^2}
\left(\eta_1^2 + s_2^2 \right) \bigg) \\
&= o(\Delta x^r)\qquad \mbox{as $\Delta x \rightarrow 0$}
\end{align*}
for any $r>0$ if either $\nu_1 \ge \Delta x_1^{-q}$ or $s_2\ge \Delta x_2^{-q}$ for any $q>0$.
For $|\eta_1 \Delta x_1| > \pi/2$, we have that
\[
1 - \sinc(\xi \Delta x_1) \cos(\eta_1 \Delta x_1) \ge 1,
\]
and therefore
\begin{align*}
W(\eta_1,s_2)
&\le
\exp \bigg(
- \frac{2 t}{\Delta x_1^2} + 2 t |\rho| \frac{1}{\Delta x_1}
\frac{|\sin(s_2 \Delta x_2)|}{\Delta x_2} - \frac{4 t}{\Delta x_2^2} \sin^2(s_2 \Delta x_2/2)
\bigg) \\
&\le
\exp \bigg(
- \frac{2 t}{\Delta x_1^2} + \frac{4 t}{\Delta x_1}
\frac{|\sin(s_2 \Delta x_2/2)|}{\Delta x_2} - \frac{4 t}{\Delta x_2^2} \sin^2(s_2 \Delta x_2/2)
\bigg) \\
&\le
\exp \bigg(
- \frac{2 t}{\Delta x_1^2} + \frac{\sqrt{21} t}{\Delta x_1}
\frac{|\sin(s_2 \Delta x_2/2)|}{\Delta x_2} - \frac{4 t}{\Delta x_2^2} \sin^2(s_2 \Delta x_2/2)
\bigg) \\
&=
\exp \bigg(
- \frac{t}{2}
\left(
\frac{1}{\Delta x_1^2} + \frac{1}{\Delta x_2^2} \sin^2(s_2 \Delta x_2/2)
\right)
\bigg) \cdot \\
&
\hspace{5 cm} \cdot \; \exp \bigg(
- t \left(
\sqrt{\frac{3}{2}} \frac{1}{\Delta x_1} - \sqrt{\frac{7}{2}} \frac{|\sin(s_2 \Delta x_2/2)|}{\Delta x_2}
\right)^2
\bigg) \\
&\le
\exp \bigg(
- \frac{t}{2}
\left(
\frac{1}{\Delta x_1^2} + \frac{1}{\Delta x_2^2} \sin^2(s_2 \Delta x_2/2)
\right)
\bigg) \\
& = o(\Delta x^r)
\end{align*}
as before.
\subsection{High $\boldsymbol{\xi}$-wavenumbers}
Completing squares, we write
\begin{align*}
W(\eta_1,s_2,\xi) &=
W_1(\eta_1,\xi) \, \cdot \,
\exp\left(
\frac{t}{\Delta x_1^2} \cos^2(s_2\Delta x_2/2) \sinc^2(\xi \Delta x_1) \sin^2(\eta_1 \Delta x_1) \rho^2
\right) \, \cdot \\
&\qquad \cdot \, \exp\left(
- t \left(
\frac{\rho}{\Delta x_1}
\cos(s_2 \Delta x_2/2) \sinc(\xi \Delta x_1) \sin(\eta_1 \Delta x_1)
+ \frac{2}{\Delta x_2}
\sin(s_2 \Delta x_2/2)
\right)^2
\right),
\end{align*}
and so, neglecting the last factor, which is $\le 1$,
\begin{align*}
\log W &\le - \frac{2 t}{\Delta x_1^2}
\left(1-\sinc(\xi \Delta x_1) \cos(\eta_1 \Delta x_1)\right)
+ \rho^2 \frac{t}{\Delta x_1^2} \cos^2(s_2\Delta x_2/2) \sinc^2(\xi \Delta x_1) \sin^2(\eta_1 \Delta x_1) \\
&\le \frac{t}{\Delta x_1^2}
\left(-2 + 2 \sinc(\xi \Delta x_1) \cos(\eta_1 \Delta x_1) + \sinc^2(\xi \Delta x_1) \sin^2(\eta_1 \Delta x_1) \right) \\
&\le \frac{t}{\Delta x_1^2} \left(-2 + |\sinc(\xi \Delta x_1)|
\left(2 |\cos(\eta_1 \Delta x_1)| + \sin^2(\eta_1 \Delta x_1) \right) \right) \\
&= \frac{t}{\Delta x_1^2} \left(-2 + |\sinc(\xi \Delta x_1)|
\left(2 - (1-|\cos(\eta_1 \Delta x_1)|)^2 \right)
\right) \\
&\le - \frac{2 t}{\Delta x_1^2} (1-|\sinc(\xi \Delta x_1)|),
\end{align*}
and following the remainder of the argument in the one-dimensional case we can deduce that the contribution from
this range is also exponentially small.
\subsection{Multiple drift terms} \label{app:MOREDRIFTTOY}
Another possible extension is a model of the form
\begin{equation} \label{eq:MULTDRIFT}
\frac{\partial u}{\partial t} +a_1 x \frac{\partial u}{\partial y_1} +a_2 x \frac{\partial u}{\partial y_2} = \frac{\partial^2 u}{\partial x^2},
\end{equation}
where $(a_1,a_2) \neq (0,0)$ so that we have two drift terms and neither of the drift coefficients depend on $y$.
For this, we can apply the $y_{1}$-FT and the $y_{2}$-FT to get
\[ \frac{\partial v}{\partial t}-ia_1 p_1xv-ia_2 p_2xv= \frac{\partial^2 v}{\partial x^2}, \]
where $p_1$ and $p_2$ are wavenumbers.
To find the analytical solution we apply the $x$-FT to get
\[ \frac{\partial w}{\partial t} -(a_1p_1+a_2 p_2)\frac{\partial w}{\partial s}= -s^2w .\]
The analysis would then continue as before but with $a_1 p_1+a_2 p_2$ as a joint wavenumber.
In this situation however the wavenumbers have to be handled differently. For example, we can have separately
large values of $|p_1|$ and $|p_2|$ but the value of $a_1p_1+a_2p_2$ can be small.
\section{Conclusions}
The numerical analysis of hypoelliptic PDEs with variable coefficients and Dirac initial datum is notoriously difficult because standard approaches to convergence analysis are not directly applicable. In contrast with parabolic initial-value problems, in the case of hypoelliptic PDEs the situation
is complicated by the fact that
diffusion generally acts only in some, but not all, co-ordinate directions. The approach of \cite{GC} for the one-dimensional heat equation is based on Fourier analysis to show approximation of a certain order for the low wavenumber components and exponential decay of the high wavenumber components for smoothing schemes. In the present case of hypoelliptic equations the analysis is more intricate because of the interplay of wavenumbers in the different co-ordinate directions as a consequence of variable coefficients and the resulting noncommutativity of the spatial differential operators in the different directions.
We exploit the special structure of the Kolmogorov equations under consideration by performing a Fourier transform in the direction with the pure transport term, and discretize in the diffusive direction(s) by a finite difference scheme. This results in a parametrized system of semi-discrete equations, which can be solved by standard methods, and the resulting solution is then transformed back from Fourier space using exponentially convergent numerical quadrature, thus leading to an effective technique for dimensional reduction. The numerical analysis sheds light on the interaction of different wavenumbers and allows us to derive second-order convergence for the proposed numerical scheme in the absence of time-discretization.
An open question at present is how to replicate the analysis herein for the fully discrete counterpart of the method, which also includes time-discretization. Although unconditional stability of the backward Euler scheme for initial data in $L^2$, can be proved by standard energy estimates, a similar argument does not seem to apply in the case of Dirac initial datum, as the use of a discrete negative-order Sobolev norm, which would be the natural choice of norm for the discrete representation of the Dirac function on the finite difference grid, in conjunction with discrete counterparts of parabolic smoothing estimates in the diffusive directions, is obstructed by
the particular form of the variable coefficients in the Kolmogorov equations under consideration, and the complex nature of error propagation for the fully-discrete scheme has not so far allowed a rigorous convergence analysis of the fully discrete scheme.
\bibliographystyle{plain}
|
\section{Introduction}
For some time now, we have been able to compute the low-energy
effective action of ${\cal N}=2$ supersymmetric gauge theories in four
dimensions. In \cite{Seiberg:1994rs,Seiberg:1994aj}, the solution for
the low-energy theory was given in terms of an algebraic curve and an
associated differential. Subsequent works have simplified and
clarified many aspects of the Seiberg-Witten solution. The
Seiberg-Witten curves may be intuitively pictured in terms of M-theory
five-branes \cite{Witten:1997sc}, and this geometric picture has inspired
a description of class $\mathcal{S}$ theories in terms of punctured
Riemann surfaces \cite{Gaiotto:2009we}. In a parallel development, it
has also become possible to compute instanton contributions by
invoking the powerful machinery of equivariant localization
\cite{Nekrasov:2002qd}. Of particular note, the calculation of the
gauge theory partition function on $S^4$ via localization
naturally incorporates these instanton sums \cite{Pestun:2007rz}. All
these developments were key to writing a dictionary between
observables in four-dimensional gauge theories and those in two-dimensional
conformal field theories: the 2d/4d correspondence
\cite{Alday:2009aq}.
The 2d/4d correspondence makes it possible to use the technology of
conformal field theory to gain deeper insights into the behavior of
${\cal N}=2$ gauge theories. For instance, the $\Omega$-deformed gauge
theory partition function with a surface operator insertion maps to
the meromorphic solution of a null vector decoupling equation
\cite{Drukker:2009id,Alday:2009fs}. Thus, an analysis of conformal
blocks in two-dimensional conformal field theory yields information
about surface operators in gauge theories. These conformal blocks can
be viewed as solutions to Riemann-Hilbert problems specified by a
differential equation with singularities and associated monodromies \cite{Vartanov:2013ima}. We
expect this exact picture to be valid in gauge theory (and the field
theory limit of topological string theory) \cite{Awata:2010bz}.
In this paper, we study quantum chromodynamics with ${\cal N}=2$ supersymmetry and gauge group SU$(2)$, and the
corresponding Virasoro conformal blocks.
In particular, we study the differential equation
that the instanton partition function with surface operator insertion
satisfies. This corresponds to an analysis of null vector decoupling equations in the presence
of irregular blocks. The differential equations satisfied by
correlators involving irregular blocks were described in \cite{Awata:2010bz,AwataYamada,Gaiotto:2012sf}. The
equations are exact in the $\Omega$-deformation parameters
$\left(\epsilon_{1},\epsilon_2\right)$, and provide for a map to standard gauge theory
expressions for the Seiberg-Witten curve, including $\epsilon_i$
corrections.
We then concentrate on the limit $\epsilon_2/\epsilon_1 \rightarrow 0$ \cite{NSlimit}, which is a large central charge limit in the conformal field
theory. It has been shown in e.g. \cite{MironovMorozov,Fateev:2009aw,TroostToroidal} that a WKB analysis of the null vector decoupling equations in this semi-classical limit reproduces the non-convergent $\epsilon_1$-expansion of the
instanton partition function of the gauge theory.\footnote{For non-perturbative results in the context of topological strings we refer to \cite{Marino,Marino1308}.} There is a rich
literature
\cite{Voros,Delabaere1,Jidoumou,Delabaere2,Delabaere3,ZinnJustin,ZinnJustin2,book,Costin,NakanishiandIwaki} on methods which may be used to enhance these results non-perturbatively.
Using the exact WKB analysis, we study the resulting differential equations satisfied by
the (ir)regular conformal blocks (equivalently, the
$\epsilon_1$-deformed surface operator partition function). This allows us to compute the monodromy group of
each of the differential equations as a function of (i) the parameters
of the differential equations, and (ii) the Borel resummed monodromies
that are properties of individual solutions. The monodromy group
contains information about the instanton partition function with
surface operator insertion, which is non-perturbative in $\epsilon_1$.
In doing so, we provide the underlying exact picture
\cite{Vartanov:2013ima} with a detailed description of how these
beautiful and abstract mathematical constructs reduce to the more
hands-on limiting analysis of ${\cal N}=2$ gauge theories to which we have become accustomed.
In this physical set-up, we apply the theorems of
\cite{NakanishiandIwaki}, thereby drawing on intuition from both gauge
theory and the mathematical study of singular perturbation theory
\cite{book}. As a by-product, we add details to the WKB analysis and
provide a calculation of the monodromy group of the differential
equation in terms of deformed gauge theory data. For instance, we
analyze the occurrence of a double flip, consisting of simultaneous
single flips. Two different ways of splitting the double flip into two
single flips give the same monodromy group and Stokes automorphism.
Although we demonstrate this result in the context of $N_f=4$ theory,
this is a new result in the exact WKB method and we believe it is
valid in a more general context.
In \cite{GMN}, a WKB analysis of the Hitchin systems corresponding to
circle compactifications of undeformed SU$(2)$ gauge theories was
undertaken. Our work may be viewed as an alternative route to the WKB analysis, which is closely
related to \cite{GMN} at zeroth order in $\epsilon_1$.
Our broader goal is to communicate the extreme generality of the correspondence between $\epsilon_1$-deformed ${\cal N}=2$
gauge theories --- specifically, their instanton partition functions with surface operator insertions ---
and certain Schr\"odinger equations amenable to exact WKB analysis. As a first step, we show the extent
to which the program applied to pure ${\cal N}=2$ super Yang-Mills in \cite{Kashani-Poor:2015pca} generalizes to theories with matter.
We will now briefly present the structure of our paper. In section
\ref{CFT}, we present a derivation of the null vector decoupling
equation satisfied by the five-point conformal block with a light
degenerate insertion, which has a null vector at level two. We apply
the collision procedure of \cite{Gaiotto:2009ma} to produce
irregular conformal blocks and derive the null vector decoupling
equations satisfied by the limit blocks. We then consider the
semi-classical limit (of infinite central charge) of these
differential equations. These equations will be the starting point for
the exact WKB analysis of section \ref{exactWKB}. In this section, we
briefly review the exact WKB approach, and in section
\ref{monodromygroup} apply it to the calculation of the monodromy
groups of our differential equations.
We make contact with the standard undeformed Seiberg-Witten perspective in
section \ref{gaugetheory} and end with comments and future directions
for work in section \ref{conclusions}. The appendices collect details
regarding the derivation of the $\epsilon_2$-exact differential
equations for the asymptotically free theories, and an independent
check of the semi-classical differential equations via the
saddle-point analyses of Nekrasov partition functions
\cite{NPS13}.
\section{The Conformal Field Theory Perspective}
\label{CFT}
In this section, we present the null vector decoupling equation
satisfied by the five-point conformal block with one degenerate
operator insertion. We then list the corresponding equations satisfied
by irregular blocks that arise when punctures collide \cite{Gaiotto:2009ma}. We study these equations within the framework of conformal field theory, and finally, exploit the fact that these conformal
blocks also capture the $\epsilon_i$-deformed instanton partition function of ${\cal N}=2$ supersymmetric gauge theories in four dimensions with SU$(2)$ gauge group and a varying number of flavours \cite{Nekrasov:2002qd}. We thus lay the groundwork for further analysis of these partition functions, which will be non-perturbative in the deformation parameter $\epsilon_1$.
For completeness, we provide the details of the derivation of all these equations in appendix \ref{diffeqns}.
We start our analysis by considering regular conformal blocks with four
ordinary primary operator insertions on the sphere and one degenerate operator insertion with a null vector at level two, which remains light in the limit of large central charge. On the gauge theory side of the $2$d/$4$d correspondence, this set-up corresponds to the conformal $N_f=4$ case.
To get asymptotically free (lower $N_f$) theories, we sequentially collide primary operators on the sphere in such a way that they
generate irregular conformal blocks \cite{Gaiotto:2009ma}. The case of three flavours will correspond to one irregular block, the case
of two flavours can correspond to either one or two irregular blocks, while a lower number of flavours corresponds to
two irregular blocks in the conformal field theory. For all these collision limits, we give the corresponding
null vector decoupling equations.
\subsection{The Five-Point Block}
We study a conformal field theory with central charge
\begin{equation}
c = 1 + 6 Q^2 \, , \qquad \mbox{where} \qquad Q = b + b^{-1} \qquad \mbox{and} \qquad b=\sqrt{\frac{\epsilon_2}{\epsilon_1}} \, .
\label{Q}
\end{equation}
We consider a five-point chiral conformal block $\Psi$ with four primary operator insertions $V_{\alpha_i}$
and an insertion of a degenerate field $\Phi_{2,1}(z)$ of the Virasoro algebra \cite{Alday:2009fs}:
\begin{equation}
\Psi(z_i, z) = \Big\langle
\Phi_{2,1}(z)\!:\prod_{i=1}^{4}V_{\alpha_i}(z_i) \!: \Big\rangle \, .
\label{psi}
\end{equation}
The degenerate field $\Phi_{2,1}$ has conformal dimension $\Delta_{2,1}$
\begin{equation}
\Delta_{2,1}=-\frac{1}{2} - \frac{3}{4}\, \frac{\epsilon_2}{\epsilon_1} \, ,
\label{del}
\end{equation}
while the conformal dimensions of the generic primaries are denoted $\Delta_{\alpha_i}$. We have chosen the degenerate insertion
such that it remains light in the limit of large central charge $\epsilon_2/\epsilon_1 \rightarrow 0$.
The degenerate field $\Phi_{2,1}$ has a null vector at level two, and consequently satisfies the null vector condition
\begin{equation}
\frac{\epsilon_1}{\epsilon_2}\, \partial^2
\Phi_{2,1}(z)
+\, :\! T(z) \Phi_{2,1}(z)\!:\,\ = 0~ ,
\label{nuv}
\end{equation}
where the operator $T(z)$ is the holomorphic stress tensor of the conformal field theory. Using the operator
product expansion between the stress tensor and the primary fields, the second term can be written as:
\begin{equation}
\begin{aligned}
\Big\langle\!\! :\!T(z) \Phi_{2,1}(z)\!:\prod_{i=1}^{4}V_{\alpha_i}(z_i)\Big\rangle &=
\sum_{i=1}^{4}\left(\frac{\Delta_{\alpha_i}}{(z-z_i)^2}+\frac{1}{z-z_i}
\frac{\partial}{\partial z_i} \right)\,\Big\langle \Phi_{2,1}(z)\prod_{i=1}^{4}V_{\alpha_i}(z_i)\Big\rangle~.
\end{aligned}
\label{tpsi}
\end{equation}
Imposing global conformal invariance allows us to express the
derivatives with respect to $z_1, z_3$ and $z_4$ in terms of the derivatives at $z_2$ and $z$.
Then, setting the insertions to be at $(z,0,q,1,\infty)$, the null vector decoupling equation takes the form
\begin{align}
\label{tpsi2}
\nonumber \left[\frac{\epsilon_1}{\epsilon_2}\,\frac{\partial^2}{\partial z^2}\, + \left(\frac{\Delta_{\alpha_2}}{(z-q)^2}+\frac{q(q-1)}{z(z-1)(z-q)}\frac{\partial}{\partial q}\right) - \frac{2z-1}{z(z-1)}\frac{\partial}{\partial z} +\frac{\Delta_{\alpha_1}}{z^2}+ \frac{\Delta_{\alpha_{3}}}{(z-1)^2} \right. &\\
\left. -\frac{\Delta_{2,1}+\Delta_{\alpha_1} +\Delta_{\alpha_2} + \Delta_{\alpha_{3}} -\Delta_{\alpha_{4}} }{z(z-1)} \right] & \Psi(z, q) = 0
\end{align}
The null vector decoupling on the five point conformal block was also studied in \cite{Alday:2009fs,Kashani-Poor:2013oza}.
The conformal dimensions $\Delta_i$ of the primary fields
$V_{\alpha_i}$ can be written in terms of the momenta $\alpha_i$ as
\begin{equation}
\Delta_{\alpha_i}=\alpha_i(Q-\alpha_i)~.
\label{deltaalpha}
\end{equation}
We further parameterize the momenta $\alpha_i$ in terms of the four masses $m_i$:
\begin{equation}
\begin{aligned}
\alpha_1&=\frac{Q}{2}+\frac{m_1-m_2}{2\sqrt{\epsilon_1\epsilon_2}}~,\qquad
\alpha_2=\frac{Q}{2}+\frac{m_1+m_2}{2\sqrt{\epsilon_1\epsilon_2}}~,\\
\alpha_{3}&=\frac{Q}{2}-\frac{m_3+m_4}{2\sqrt{\epsilon_1\epsilon_2}}~,\qquad
\alpha_{4}=\frac{Q}{2}-\frac{m_3-m_4}{2\sqrt{\epsilon_1\epsilon_2}}~.
\end{aligned}
\label{alphas}
\end{equation}
As a function of the masses, the conformal dimensions are
\begin{equation}
\begin{aligned}
\Delta_{\alpha_1}&=\frac{(\epsilon_1+\epsilon_2)^2-(m_1-m_2)^2}{4\epsilon_1\epsilon_2}~,\qquad
\Delta_{\alpha_2}=\frac{(\epsilon_1+\epsilon_2)^2-(m_1+m_2)^2}{4\epsilon_1\epsilon_2}~,\\
\Delta_{\alpha_{3}}&=\frac{(\epsilon_1+\epsilon_2)^2-(m_3+m_4)^2}{4\epsilon_1\epsilon_2}~,\qquad
\Delta_{\alpha_{4}}=\frac{(\epsilon_1+\epsilon_2)^2-(m_3-m_4)^2}{4\epsilon_1\epsilon_2}~.
\end{aligned}
\label{Deltas}
\end{equation}
In terms of these variables that are appropriate for comparison to the four dimensional gauge theory, the null vector decoupling equation for the $N_f=4$ theory takes the following form:
\begin{align}
\nonumber &\left[ -\epsilon_1^2\frac{\partial^2}{\partial z^2} +\frac{(m_1-m_2)^2}{4z^2} + \frac{(m_1+m_2)^2}{4(z-q)^2} +\frac{(m_3+m_4)^2}{4(z-1)^2} + \frac{m_1^2+m_2^2+2m_3m_4}{2z(1-z)} \right. \\
\nonumber &\quad -\epsilon_1^2 \left( \frac{q^2-2 q z+z^2 \left(z^2-2 z+2\right)}{4 (z-1)^2 z^2 (q-z)^2} \right) + \epsilon_1\epsilon_2\left(\frac{q(1-q)}{z(z-1)(z-q)}\frac{\partial}{\partial q} +\frac{2z-1}{z(z-1)}\frac{\partial}{\partial z}\right.\\
\nonumber &\quad \quad \left. +\frac{ q^2 \left(-z^2+z-1\right)+2 q z \left(z^2-z+1\right)+z^2 \left(-2 z^2+3 z-2\right)}{2 (z-1)^2 z^2 (q-z)^2}\right) \\
&\quad \left. +\epsilon_2^2 \left(\frac{q^2 \left(-3 z^2+3 z-1\right)+2 q z \left(3 z^2-3 z+1\right)+z^2 \left(-4 z^2+5 z-2\right)}{4 (z-1)^2 z^2 (q-z)^2}\right) \right] \Psi(z,q) = 0\,.
\end{align}
\subsection{The Null Vector Decoupling Equations for Irregular Blocks}
We now take limits of the five-point null vector decoupling equation
(\ref{tpsi2}) in which various primary operators $V_{\alpha_i}$
collide to form irregular conformal blocks of order one
\cite{Gaiotto:2009ma}. These limiting configurations are in direct
correspondence with the $\epsilon_i$-deformed SU$(2)$ gauge theories
with $N_f <4$. We list below the null vector decoupling equations for each
of these cases and refer to appendix \ref{diffeqns} for a
detailed derivation. A summary of these equations can also be found
in \cite{Awata:2010bz}.
\paragraph{$N_f=3:$} In this case, we have one irregular block of order
one with a fourth order pole at $z=0$. In the gauge theory variables,
we take $q\rightarrow 0$ and $m_2\rightarrow\infty$, keeping the dynamical scale
$\Lambda_3 = q \, m_2$ finite. The resulting differential equation is:
\begin{align}
\nonumber &\left[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} +\frac{(m_3+m_4)^2}{4(z-1)^2}+ \frac{m_3 m_4}{z(1-z)} +\frac{m_1\Lambda_3}{z^3}+\frac{\Lambda_3^2}{4z^4} +\epsilon_1\epsilon_2\left(\frac{1-2z}{z\left(1-z\right)}\frac{\partial}{\partial z} + \frac{1-2z}{2z(z-1)^2} \right)\right. \\
\nonumber &\left. \ +\frac{1}{z^2\left(1-z\right)}\left(-\epsilon_1\epsilon_2\Lambda_3\frac{\partial}{\partial\Lambda_3}+ m_1^2+m_1(\epsilon_1+\epsilon_2) \right) -\frac{\epsilon_1^2}{4(z-1)^2} +\epsilon_2^2\frac{(3-4z)}{4z(z-1)^2} \right]\Psi_3(z,\Lambda_3) = 0 \, .
\end{align}
\paragraph{$N_f=2:$} There are two ways to reach the case
with two flavours from the case with three flavours. One could
decouple either the flavour with mass $m_1$ or one of those with
masses $m_{3,4}$. As shown in \cite{GMN}, these lead to inequivalent
Hitchin systems and give rise to distinct differential equations.
Let us first consider the irregular block of order one with a third
order pole at $z=0$. This corresponds to decoupling $m_1$. We refer to this as the
asymmetric configuration and the associated null vector decoupling equation becomes:
\begin{multline}
\left[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} +\frac{(m_3+m_4)^2}{4(z-1)^2}+ \frac{m_3 m_4}{z(1-z)} +\frac{\Lambda_2^2}{z^3} -\frac{\epsilon_1\epsilon_2}{2z^2(1-z)}\Lambda_2\frac{\partial}{\partial\Lambda_2}
\right.\cr
\left.+\epsilon_1\epsilon_2\left(\frac{1-2z}{z(1-z)}\frac{\partial}{\partial z} + \frac{1-2z}{2z(z-1)^2} \right)-\frac{\epsilon_1^2}{4(z-1)^2} +\epsilon_2^2\frac{(3-4z)}{4z(z-1)^2} \right]\Psi_{2,A}(z,\Lambda_2) = 0
\end{multline}
Alternatively, one can consider two irregular blocks of order one,
with equal fourth order poles. This corresponds to decoupling $m_3$
while keeping $m_1$ and $m_4$ finite. We refer to this as the
symmetric configuration and the associated null vector decoupling
equation reads:
\begin{align}
\nonumber &\left[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} +\frac{\Lambda_2^2}{4z^4} + \frac{\Lambda_2m_1}{z^3(z-1)^2}-\frac{\Lambda_2 m_4}{z(z-1)^3} +\frac{\Lambda_2^2}{4(z-1)^4} +\frac{2-3z}{4z(z-1)^2}(2\epsilon_1\epsilon_2+3\epsilon_2^2)\right. \\
\nonumber & \left. \ +\frac{1}{z^2(z-1)^2}\left(-\epsilon_1\epsilon_2\Lambda_2\frac{\partial}{\partial\Lambda_2} -2\Lambda_2m_1+ m_1^2+m_1(\epsilon_1+\epsilon_2)\right) +\epsilon_1\epsilon_2 \frac{3z-1}{z(z-1)}\frac{\partial}{\partial z} \right]\Psi_{2,S}(z, \Lambda_2) = 0\,.
\end{align}
\paragraph{$N_f=1:$} We consider two irregular blocks of order one with one fourth order pole and one third order pole. This corresponds to decoupling $m_4$ and the null vector decoupling equation takes the form
\begin{align}
\nonumber \Bigg[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} &+\frac{\Lambda_1^2}{4z^4} + \frac{\Lambda_1m_1}{z^3(z-1)^2}-\frac{\Lambda_1^2}{4z(z-1)^3}
+\frac{2-3z}{4z(z-1)^2}(2\epsilon_1\epsilon_2+3\epsilon_2^2) +\epsilon_1\epsilon_2 \frac{3z-1}{z(z-1)}\frac{\partial}{\partial z}
\\
& \quad +\frac{1}{z^2(z-1)^2}(-\epsilon_1\epsilon_2\Lambda_1\frac{\partial}{\partial\Lambda_1} -2\Lambda_1m_1+ m_1^2+m_1(\epsilon_1+\epsilon_2))
\Bigg]\Psi_1(z, \Lambda_1) = 0\,.
\end{align}
\paragraph{$N_f=0:$} Finally, we consider the case with two irregular blocks of order one with equal third order poles. All masses have been decoupled and the null vector decoupling equation becomes
\begin{multline}
\left[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} + \frac{\Lambda_0^2}{z^3(z-1)^2} +\frac{1}{z^2(z-1)^2} \left(-\frac{1}{2}\epsilon_1\epsilon_2\Lambda_0\frac{\partial}{\partial\Lambda_0 }- 2\Lambda_0^2\right)
+ \frac{\Lambda_0^2}{z(z-1)^3}
\right.\cr
\left. +\epsilon_1\epsilon_2 \frac{3z-1}{z(z-1)}\frac{\partial}{\partial z} +\frac{2-3z}{4z(z-1)^2}(2\epsilon_1\epsilon_2+3\epsilon_2^2)
\right]\Psi_0(z, \Lambda_0) = 0\,.
\end{multline}
This completes the list of six differential equations that we refer to throughout.
\subsection{The Semi-Classical Limit}
\label{semiclassical:subsec}
In the rest of our paper, we will concentrate on the limit $\epsilon_2/\epsilon_1 \rightarrow 0$, which is a large central charge
limit. We keep the ratio of the mass parameters $m_i$ and the deformation parameter $\epsilon_1$
fixed. In this limit, the primary insertions $V_{\alpha_i}$ are heavy, while the degenerate insertion $\Phi_{2,1}$ is light. Thus, in this limit,
the differential equation (\ref{tpsi2}) simplifies, and we can drop the term proportional to $\partial_z$, while the terms proportional
to the conformal dimensions $\Delta_{\alpha_i}$ grow large.
To simplify the equation further, we must specify the leading
dependence of the $q$-derivative of the five-point block on $\epsilon_2$. To that end, we make the semi-classical $\epsilon_2\rightarrow 0$ ansatz
\begin{equation}\label{semiclassicalF}
\Psi(z,q) =
\text{exp}\left(-\frac{\widetilde{F}(q, m_i,\epsilon_i)}{\epsilon_1\epsilon_2}\right) \, \psi(z,q) \, .
\end{equation}
We suppose that the $q$-derivative of the remaining function $\psi(z,q)$ is sub-dominant in the small $\epsilon_2/\epsilon_1$ limit,
and observe that the leading dependence in $\epsilon_2$ is only on the cross-ratio $q$ of the heavy operators.
We then define the quantity
\begin{eqnarray}
\widetilde{u} &=& q (1-q) \partial_q \widetilde{F} \, .
\end{eqnarray}
The parameter $\widetilde{u}$ is identified with the Coulomb modulus of the gauge
theory up to shifts that depend on the masses. Substituting this parameterization
into the null vector decoupling equation and taking the semi-classical
limit $\epsilon_2 \rightarrow 0$ leads to the Schr\"odinger
equation
\begin{equation}
\label{diffeqngeneral1}
\left( -\epsilon_1^2 \frac{\text{d}^2}{\text{d}z^2} + Q(z,\epsilon_1) \right) \psi(z,q) = 0 \, ,
\end{equation}
where the potential function $Q$ has an $\epsilon_1$ expansion which
terminates at second order
\begin{equation}
\label{diffeqngeneral2}
Q(z) = Q_0 (z) + \epsilon_1 \ Q_1 (z) + \epsilon_1^2 \ Q_2 (z) \, .
\end{equation}
The coefficient functions are
\begin{align}\label{nullNf4Qfunctions}
Q_0(z) &= -\frac{\widetilde{u}}{z(z-1)(z-q)} + \frac{(m_1-m_2)^2}{4z^2} + \frac{(m_1+m_2)^2}{4(z-q)^2} +\frac{(m_3+m_4)^2}{4(z-1)^2}+ \frac{m_1^2 + m_2^2 + 2 m_3 m_4}{2 z (1- z)} \, ,
\cr
Q_1(z)&= 0 \, , \cr
Q_2(z) &= -\frac{1}{4z^2}-\frac{1}{4(z-1)^2}-\frac{1}{4(z-q)^2}+\frac{1}{2z(z-1)} \, .
\end{align}
\subsection{The Semi-Classical Irregular Blocks}
\label{semiclassicalIrregular:subsec}
The same type of ansatz \eqref{semiclassicalF} can be used in order to obtain the differential equations for the irregular blocks in the
semi-classical $\epsilon_2\rightarrow 0$ limit. The variable parameterizing the Coulomb modulus is now defined as
\begin{equation}
\widetilde{u} = \Lambda_{N_f}\frac{\partial \widetilde{F}}{\partial \Lambda_{N_f}} \,,
\end{equation}
where $\Lambda_{N_f}$ is the corresponding strong coupling scale of the $N_f < 4$ gauge theory. As in the conformal case,
the prepotential of the gauge theory will differ mildly from $\widetilde{F}$.
However, what is of importance to us is the pole structure of the functions $Q_k(z)$, and we choose a parameterization
that descends naturally from the conformal theory and that allows for a simple presentation of the differential equations.
In the following, we present all the asymptotically free cases:
\begin{itemize}
\item {$N_f=3$}: The Schr\"odinger equation which governs the $\epsilon_1$-deformed gauge theory is given by
\begin{align}
\label{Nf=3 semiclassical}
\left[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} +\frac{(m_3+m_4)^2}{4(z-1)^2}+ \frac{m_3 m_4}{z(1-z)} +\frac{m_1\Lambda_3}{z^3}+\frac{\Lambda_3^2}{4z^4} +\frac{\widetilde{u}}{z^2(1-z)}
-\frac{\epsilon_1^2}{4(z-1)^2}
\right]\psi_3(z,\Lambda_3)
= 0 \,.
\end{align}
\item {$N_f=2$ (asymmetric realization)}: The differential equation in the semi-classical limit takes the form
\begin{equation}
\label{Nf=2 asymmetric semiclassical}
\left[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} +\frac{(m_3+m_4)^2}{4(z-1)^2}+ \frac{m_3 m_4}{z(1-z)} +\frac{\Lambda_2^2}{z^3} +\frac{\widetilde{u}}{z^2(1-z)}
-\frac{\epsilon_1^2}{4(z-1)^2}
\right]\psi_{2,A}(z,\Lambda_2) = 0
\end{equation}
\item {$N_f=2$ (symmetric realization)}:
\begin{equation}
\label{Nf=2 symmetric semiclassical}
\left[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} +\frac{\Lambda_2^2}{4z^4} + \frac{\Lambda_2m_1}{z^3(z-1)^2}-\frac{\Lambda_2 m_4}{z(z-1)^3} +\frac{\Lambda_2^2}{4(z-1)^4}
+\frac{\widetilde{u}}{z^2(z-1)^2}
\right]\psi_{2,S}(z, \Lambda_2) = 0\,.
\end{equation}
\item {$N_f=1$}:
\begin{equation}
\label{Nf=1 semiclassical}
\left[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} +\frac{\Lambda_1^2}{4z^4} + \frac{\Lambda_1m_1}{z^3(z-1)^2}-\frac{\Lambda_1^2}{4z(z-1)^3}
+\frac{\widetilde{u}}{z^2(z-1)^2}
\right]\psi_1(z, \Lambda_1) = 0\,.
\end{equation}
\item {$N_f=0$}: Finally, for the pure super Yang-Mills theory, the equation reads
\begin{equation}
\label{PureNVD}
\left[-\epsilon_1^2 \frac{\partial^2}{\partial z^2}
+\frac{\Lambda_0^2}{z^3(z-1)^2} +\frac{\widetilde{u}}{z^2(z-1)^2}
+ \frac{\Lambda_0^2}{z(z-1)^3}
\right]\psi_0(z, \Lambda_0) = 0\,.
\end{equation}
\end{itemize}
We have thus obtained the differential equations which we analyze in detail in section \ref{monodromygroup}.
\section{The Exact WKB Analysis of Differential Equations}
\label{sec:IntroExactWKB}
\label{exactWKB}
In this section, we review the exact WKB approach to the analysis of
differential equations and apply it to the
null vector decoupling equations in the semi-classical
limit. We will carry out the exact WKB analysis
with respect to the small parameter $\epsilon_1$. Our analysis
will therefore be valid to zeroth order in $\epsilon_2$ and non-perturbatively
in $\epsilon_1$. Below, we briefly review the salient features of the exact WKB analysis and refer the reader to \cite{book, NakanishiandIwaki} for a more comprehensive treatment
of the same.
\subsection{The Exact WKB Method}
The differential equations that we study can be written in the form of
a Schr\"odinger equation:
\begin{equation}
\left( -\epsilon_1^2 \frac{\text{d}^2}{\text{d}x^2} + Q(x) \right)
\psi(x,\epsilon_1) = 0 \, .
\end{equation}
We allow the function $Q$ to have an expansion of the form
\begin{equation}
Q(x) = Q_0 (x) + \epsilon_1 \ Q_1 (x) + \epsilon_1^2 \ Q_2 (x) + \cdots \, .
\end{equation}
For the null vector
decoupling equations that we study, the only non-zero coefficient functions are
$Q_0,Q_1$ and $Q_2$. We choose a WKB ansatz for the solution to this
differential equation, which takes the form
\begin{equation}
\psi (x,\epsilon_1) = \text{exp} \left( \int_{x_0}^{x} \text{d}x' \ S(x',\epsilon_1) \right) \, ,
\end{equation}
with $S(x,\epsilon_1)$ expanded as a formal power series in
$\epsilon_1$ as
\begin{equation}
S(x,\epsilon_1) = \frac{1}{\epsilon_1} \ S_{-1} (x) + S_0 (x) + \epsilon_1 \ S_1 (x) + \cdots \, .
\end{equation}
Substituting this ansatz into the differential equation, we get
recursion relations governing the coefficients $S_k$
\begin{align}
S_{-1}^2 &= Q_0 \, , \\
2 S_{-1} S_{n+1} + \sum_{k+l = n} S_k S_l +
\frac{\text{d}S_n}{\text{d}x} &= Q_{n+2} \quad \text{for} \quad n
\geq -1 \, .
\end{align}
We see that the initial conditions governing the system of recursion
relations allow for two possible sets of solutions to these recursion
relations, as $S_{-1} = \pm \sqrt{ Q_0} $. We also note the crucial
feature that the zeroes of $Q_0$, which we call turning points,
introduce branch cuts on the Riemann surface $\Sigma$ on which our
differential equation and its exact solutions live. Thus, in our exact
WKB treatment, we introduce a new manifold $\hat{\Sigma}$, which is a
double cover of the Riemann surface, and we move between sheets as we
pass branch cuts that emanate from turning points, or odd order poles.
From hereon, we will distinguish the choice of WKB solution by
attaching to it the subscript ($\pm$). We also observe that in the $\epsilon_1$-expansion of
$S(x,\epsilon_1)$, the sets of odd and even coefficients are
dependent. If we define
\begin{equation}
S_{\text{odd}} = \sum_{j \geq 0} S_{2j-1} \; \epsilon_1^{2j-1} \quad
\text{and} \quad S_{\text{even}} = \sum_{j \geq 0} S_{2j} \;
\epsilon_1^{2j} \, ,
\end{equation}
we have the relation
\begin{equation}
S_{\text{even}} = -\frac{1}{2} \frac{\text{d}}{\text{d}x} \log S_{\text{odd}} \, .
\end{equation}
Putting all this together, we can write down a formal expression for
the two linearly independent solutions to our differential equation:
\begin{equation}
\label{eq:WKBSoln}
\psi_{\pm} = \frac{1}{\sqrt{S_{\text{odd}}}} \ \text{exp} \left\lbrace
\pm \int_{x_0}^{x} \text{d}x' \ S_{\text{odd}} \right\rbrace \, .
\end{equation}
This formal expression should be understood as an analytic function of
$x$ multiplying an asymptotic series in $\epsilon_1$\ :
\begin{equation}
\psi_{\pm} = \text{exp} \left\lbrace \pm \frac{1}{\epsilon_1}
\int_{x_0}^{x} \text{d}x' \ \sqrt{Q_0 (x')} \right\rbrace
\epsilon_1^{1/2} \sum_{k=0}^{\infty} \epsilon_1^k \ \psi_{\pm,k} (x)
\, .
\end{equation}
\subsubsection*{Borel Resummation}
In the exact WKB approach, it is convenient to normalize
wave-functions at distinguished points of the differential
equation. As mentioned earlier, in addition to the
singularities of the coefficient functions of the differential
equations, their zeros (turning points) also play an important role. We will normalize our solutions with respect to
the turning points, i.e. choose the starting point $x_0$ of the
integration path to be a turning point $t$,
\begin{equation}
\psi_{\pm} = \frac{1}{\sqrt{S_{\text{odd}}}} \ \text{exp} \left\lbrace
\pm \int_{t}^{x} \text{d}x' \ S_{\text{odd}} \right\rbrace \, .
\end{equation}
Formal WKB solutions are generically
divergent. To remedy this, we invoke Borel resummation: a technique
that constructs an analytic function whose asymptotic expansion
matches the formal WKB series. The Borel transformed series is defined as
\begin{equation}
\psi (\epsilon_1) = \sum_{k=0}^{\infty} \psi_k \ \epsilon_1^k \quad
\xrightarrow{\text{Borel transform}} \quad \widetilde{\psi}(y) =
\sum_{k=1}^{\infty} \psi_k \frac{y^{k-1}}{(k-1)!} \, .
\end{equation}
Next, define the function \cite{NakanishiandIwaki}
\begin{equation}
\Psi(\epsilon_1) = \psi_0 + \int_{\ell_\theta}
\text{d}y \ e^{-y/\epsilon_1} \widetilde{\psi}(y) \, ,
\end{equation}
where $\ell_\theta$ is the line connecting a point at which the series
$\widetilde{\psi}(y)$ converges\footnote{To be precise, this is true
for Gevrey-1 series, which in our context corresponds to the
following statement. If $\psi_k$ is the $k$th coefficient of the
asymptotic series then the series is Gevrey-1 type if growth of
$\psi_k$ is bounded by $\psi_k \leq A B^k k!$ for some constants $A$
and $B$. If $\psi_k$ is a function of a continuous variable, say,
$x\in {\mathbb C}$ then
this condition applies to the supremum of $\psi_k(x)$ in a compact
subset of ${\mathbb C}$.} --- typically, a turning point --- to the point at infinity at an angle $\theta$. If this integral exists, $\Psi(\epsilon_1)$ is the requisite analytic function, called the Borel sum.
Notice that the Borel sum contains an angular dependence. In order to
understand this better, one must appreciate that Borel sums are
typically defined only in regions of the complex $\epsilon_1$-plane,
and not throughout. These regions are bounded by Stokes
lines, defined by the condition
\begin{equation}
\text{Im} \left[ \int_{x_0}^{x} \text{d}x' \sqrt{Q_0 (x')} \right] = 0 \, ,
\end{equation}
and different Stokes regions are assigned different linear
combinations of a given basis of analytic solutions to the
differential equation, arrived at via Borel resummation. One of the
key components of the exact WKB analysis is understanding how
solutions in different Stokes regions are related by analytic continuation; these often go by
the name of ``connection formulae''. However, before we address this
transition behaviour, we will find it necessary to endow Stokes lines
with an orientation. To this end, we adopt the convention that Stokes
lines are oriented \emph{away} (i.e. the arrow on the Stokes
line is pointing away from a turning point) if
\begin{equation}
\text{Re} \left[ \int_{x_0}^{x} \text{d}x' \sqrt{Q_0 (x')} \right] > 0 \,
\end{equation}
along the Stokes line. Else, the arrow points towards the turning point.
Three Stokes lines emanate from a first order zero of $Q_0(x)$, which
is also referred to as a simple turning point. Thus one end of any
Stokes line is at a turning point. The other end can either be at a
singularity or at a turning point. When both the end points of a
Stokes line in a given Stokes graph terminate at turning points then
the corresponding Stokes graph is called ``critical''.\footnote{The general behaviour of Stokes lines is discussed in \cite{NakanishiandIwaki}. We restrict ourselves to situations that are relevant
in this work.}
\subsubsection*{Connection Formulae}
We are now in a position to state the connection formulae. For a Stokes graph which is not critical, consider two
regions $U_{1}$ and $U_{2}$ separated by a Stokes curve $\Gamma$, and
consider $\Psi_{\pm}^{j}$ to be the Borel sums of WKB solutions in
each of the regions $U_j$. The connection formulae for the Borel sums
in different Stokes domains are given by:
\begin{align}
\text{if} \ \text{Re} \left[ \int_{x_0}^{x} \text{d}x' \sqrt{Q_0 (x')} \right] < 0 \ \text{on} \ \Gamma \quad : \quad
\begin{cases}
\Psi_+^1 &= \Psi_+^2 \, , \\
\Psi_-^1 &= \Psi_-^2 \pm \text{i} \Psi_+^2 \, ,
\end{cases}
\end{align}
\begin{align}
\text{if} \ \text{Re} \left[ \int_{x_0}^{x} \text{d}x' \sqrt{Q_0 (x')} \right] > 0 \ \text{on} \ \Gamma \quad : \quad
\begin{cases}
\Psi_+^1 &= \Psi_+^2 \pm \text{i} \Psi_-^2 \, , \\
\Psi_-^1 &= \Psi_-^2 \, .
\end{cases}
\end{align}
In the above connection formulae, there is an ambiguity ($\pm$) that
is fixed by noting that the turning point that $\Gamma$ originates
from serves as a point of reference. If the path of analytic
continuation crosses $\Gamma$ counter-clockwise as seen from the
turning point, we pick the ($+$) sign, and if this path crosses
$\Gamma$ clockwise, we pick the ($-$) sign. Later in this
section, we will write down the Stokes matrices that multiply
wave-functions; these are equivalent to the above result.
The global properties of solutions to the differential equations we consider are governed by the monodromy group and the
Stokes phenomena around singular points. The monodromy group of these
differential equations can be expressed entirely in terms
of two sets of quantities: (a) the characteristic exponents at each
singular point $s_k$, and (b) the contour integrals of
$S_{\text{odd}}$ around branch cuts. We now parameterize the
characteristic exponents conveniently.
As a system of solutions to our differential equation, we consider the
WKB solutions \eqref{eq:WKBSoln}, and define the characteristic
exponents as residues of the differential:
\begin{equation}
M_k = \text{Res} \ \sqrt{Q_0(x)} \Big\vert_{x=s_k} \, .
\end{equation}
From the null vector decoupling equations we derived in the
previous section, one can check that the residues $M_k$ are linear
combinations of the mass parameters of the gauge theory. As the
monodromy group computations will use WKB wave-functions \eqref{eq:WKBSoln}, we relate the
residues of $S_{\text{odd}}$ to our characteristic exponents
as\footnote{This is true under the assumptions that $\text{Re} \ M_k
\neq 0$.}
\begin{equation}
\text{Res} \ S_{\text{odd}} (x,\epsilon) \Big\vert_{x=s_k} = \frac{M_k}{\epsilon_1} \sqrt{1+\frac{\epsilon_1^2}{4M_k^2}} \, .
\end{equation}
Finally, upon exponentiating this contribution, we get the multiplier that affects WKB wave-functions:
\begin{equation}
\nu_k^\pm = \text{exp} \left[ \text{i}\pi \left( 1 \pm \sqrt{\frac{4M_k^2}{\epsilon_1^2}+1}\right)\right] \, .
\end{equation}
Notice that $\nu^+_k = 1/\nu_k^-$, a fact that we will use repeatedly.
Since the base point $x_0$ will not always be a turning point, the modified connection formulae can be obtained by a
composition of the contour integrals. We find it convenient to use a
matrix notation to exhibit the connection formulae. As an example, let us consider
analytically continuing the Borel resummed wave-functions from Stokes
region $U_1$ to Stokes region $U_2$. As shown in figure \ref{analytic_continuation}, there are two distinct possibilities.
\begin{figure}[ht!]
\centering
\includegraphics[width=100mm]{small_diagrams_together1.pdf}
\caption{Analytic continuation of wave-functions from $U_1$ to $U_2$}
\label{analytic_continuation}
\end{figure}
If the contour crosses a Stokes line that is directed inwards to a turning point as in figure \ref{analytic_continuation}\ (A), we find the connection formula:
\begin{equation}
\begin{pmatrix}
\Psi_+^1\ , & \Psi_-^1
\end{pmatrix} \Longrightarrow
\begin{pmatrix}
\Psi_+^2\ , & \Psi_-^2
\end{pmatrix}
\begin{pmatrix}
1 &\pm \mathrm{i} u_i^{-1} \\
0& 1
\end{pmatrix} \,.
\end{equation}
In the above equation, we use the notation,
\begin{equation}
u_j=\mbox{exp}\left(2\int_{\gamma_j} \text{d}x \ S_{\text{odd}}\right)
\, ,
\end{equation}
where $\gamma_j$ is an oriented curve from the base point to the
turning point $t_j$. Along a contour that crosses a Stokes line which
is directed outwards from a turning point as in figure \ref{analytic_continuation}\ (B), we have the
connection formula:
\begin{equation}
\begin{pmatrix}
\Psi_+^1\ , & \Psi_-^1
\end{pmatrix} \Longrightarrow
\begin{pmatrix}
\Psi_+^2\ , & \Psi_-^2
\end{pmatrix}
\begin{pmatrix}
1 &0 \\
\pm \mathrm{i} u_i & 1
\end{pmatrix} \, .
\end{equation}
In the above, the $+(-)$ sign is chosen for counter-clockwise (clockwise)
crossing of the contour from one Stokes region to the other, with respect to the turning point.
For more complicated contours, it is important to take into account
contributions from any branch cuts and/or singularities enclosed along
the closed contour from the base point to the intersection point,
the turning point and then back to the base point. As a simple example of this phenomenon, let us suppose the
contour chosen happens to encircle a branch cut --- say between $t_j$
and $t_i$ as in figure \ref{branchcutcontour} -- counter-clockwise.
\begin{figure}[ht!]
\centering
\includegraphics[width=85mm]{small_diagram4.pdf}
\caption{Encircling branch cuts}
\label{branchcutcontour}
\end{figure}
Here, the curves $\gamma_i$ are those that define the parameter $u_i$.
The closed contour $\gamma_{ji}$ that encircles the branch cut has a contribution of the form
\begin{equation}
u_{ji} = \text{exp} \left( \int_{\gamma_{ji}} \text{d}x' \ S_{\text{odd}} \right) \, ,
\end{equation}
where from the figure it is clear that
\begin{equation}
u_{ji} = u_j^{-1} u_i \, .
\end{equation}
One can see that although the $u_i$ by itself is dependent on the base point, the contour integral is independent of this choice.
\subsubsection*{Contour Encircling a Turning Point}
Let us make another
important preliminary point regarding the choice of cycles. In order to define the
monodromy group, we first choose a base point and define a basis of
closed loops that encircle just the singularities. In some of the
cases we encounter, there are branch cuts between turning
points and singularities. In such cases, we choose the contours to
also include these turning points.
In order to prove that this is consistent with the usual definition of the monodromy group, let us consider a contour that only encircles the turning point, as shown in figure \ref{tpath}.
\begin{figure}[ht!]
\centering
\includegraphics[width=50mm]{diagram1.pdf}
\caption{Contour with base point $x_0$ encircling a turning point $t$}
\label{tpath}
\end{figure}
If we choose to normalize the wave-functions at $x_0$, the wave-functions undergo the following transformation as we travel along the path:
\begin{align}
M_{x_0,\text{path}} &=
\begin{pmatrix}
1 & 0 \\
\frac{\mathrm{i}}{u_1} &1
\end{pmatrix}
\begin{pmatrix}
0 & -\mathrm{i} \\
-\mathrm{i} &0
\end{pmatrix}
\begin{pmatrix}
1 & 0 \\
\mathrm{i} u_1 &1
\end{pmatrix}
\begin{pmatrix}
1 & \frac{\mathrm{i}}{u_1} \\
0 &1
\end{pmatrix}\\
&=
\begin{pmatrix}
u_1 & 0 \\
0 &\frac{1}{u_1}
\end{pmatrix}
\end{align}
\iffalse
If we choose to normalize the wave-functions at $t$ itself then as we travel along the path, the wave-functions undergo the transformation:
\begin{align}
M_{x_0,\text{path}} &=
\begin{pmatrix}
1 & \mathrm{i} \\
0 &1
\end{pmatrix}
\cdot
\begin{pmatrix}
1 & 0 \\
\mathrm{i} &1
\end{pmatrix}
\cdot
\begin{pmatrix}
0 & -\mathrm{i} \\
-\mathrm{i} &0
\end{pmatrix}
\cdot
\begin{pmatrix}
1 & 0 \\
\mathrm{i} &1
\end{pmatrix}\\
&=
\begin{pmatrix}
1 & 0 \\
0 &1
\end{pmatrix}
\end{align}
\fi
Here we have associated the matrix $-\mathrm{i}\sigma_1$ to the branch-cut crossing, which ensures that we remain on the same
sheet of the Riemann surface.
It can be easily shown that for any base point that one may choose, the answer
is trivial as above. If we
chose the turning point itself to be the base-point, $u_1=1$ and the
matrix reduces to the identity matrix. Since the net result is simply
the identity matrix, in order to calculate the monodromy matrix for
the contour that encircles the singularity $s$, one may just as well
compute the monodromy of the wave-functions around the cycle that
encircles both the turning point $t$ and the singularity $s$.
We will make use of this
repeatedly in those cases in which the branch cut extends between a
turning point and a singularity.
It is instructive to square this situation with the solution of a
differential equation near an ordinary point. It is known that any
solution of a differential equation can be written as a Taylor series
in the neighbourhood of an ordinary point. The radius of convergence
of this solution is at least as much as the distance from the chosen point
to the nearest singularity. The Taylor series solution will clearly have trivial
monodromy property. Although the WKB analysis assigns a special status to
turning points, from the differential equation point of view the turning
point is an ordinary point. Clearly, the branch cut and the Stokes
lines emanating from a turning point are artefacts of the WKB
approximation and the insertion of the matrix $-\mathrm{i}\sigma_1$ restores the
fact that the turning point is an ordinary point of the differential
equation.
\subsubsection*{Contours Encircling a Singular Point}
Let us now consider the toy example, as shown in figure \ref{Stokes_example}, where the contour encloses
a singularity.\footnote{This example will illustrate the manner in which the Stokes matrices at each intersection are written down. The Stokes lines here don't end at turning points or singularities; the reader is encouraged to think of the figure as a part of a complete Stokes graph that has been zoomed into.}
\begin{figure}[ht!]
\centering
\includegraphics[width=75mm]{small_diagram3.pdf}
\caption{Evaluation of Stokes matrices: effect of singularities}
\label{Stokes_example}
\end{figure}
In figure \ref{Stokes_example}, at the first intersection point $A$, the contour crosses counter-clockwise a Stokes line emanating from $t_i$.
Thus, the Stokes matrix is
\begin{equation}
\begin{pmatrix}
1 & 0 \\
+ \text{i} u_i & 0
\end{pmatrix} \, .
\end{equation}
In order to determine the Stokes matrix at $B$, we need to know to which turning point the Stokes line is connected.
Since this is irrelevant to the present discussion, we move on to consider the third intersection point $C$.
This time the contour crosses a Stokes line going into $t_i$, and the crossing is clockwise as seen from $t_i$.
Further, when this contour is completed using $\gamma_j$, we see that a singularity is encircled counter-clockwise.
Taking this into account, the Stokes matrix is
\begin{equation}
\begin{pmatrix}
1 & -\text{i} u_i^{-1} \nu_k^{-2} \\
0 & 1
\end{pmatrix} \, .
\end{equation}
Finally, at the fourth intersection point $D$, the contour crosses the Stokes line clockwise. In fact, it is very similar to the first intersection, except that now there is a singularity encircled. Consequently, the Stokes matrix is
\begin{equation}
\begin{pmatrix}
1 & 0 \\
- \text{i} u_j \nu_k^2 & 0
\end{pmatrix} \, .
\end{equation}
This concludes our brief review of the exact WKB analysis. We refer the reader to \cite{book, NakanishiandIwaki} for a more detailed discussion and further references.
\subsection{The Applicability of the Exact WKB Analysis}
The application of the exact WKB techniques depends on the precise
differential equation under consideration.
Before we apply the exact WKB method to the equations derived in the
previous section, it is important to point out the subtleties in the
applicability of this analysis.
In the Schr\"odinger type differential equations listed in sections \ref{semiclassical:subsec} and \ref{semiclassicalIrregular:subsec}, the parameter $\epsilon_1$ functions as the
Planck's constant $\hbar$ in the WKB approximation scheme. For our
null vector decoupling equations, the potential has
zeroth, first and second order terms in $\epsilon_1$. In order to apply the exact WKB
techniques to the solution of the differential equation,
the $\epsilon_1$-deformed potential must satisfy certain
conditions. These consistency conditions not only ensure normalizability of the
wave-functions at singularities but also are useful in proving Borel
summability of the WKB wave-functions.
The necessary conditions (eq.~(2.8) and (2.9) in \cite{NakanishiandIwaki}) are:
\begin{itemize}
\item If the leading coefficient $Q_0$ has a pole of order $m \ge 3$, then the order of $Q_{n \ge1}$ at
that pole should be smaller than $1+m/2$.
\item If the pole of $Q_0$ (at, say $z=z_0$) is of order $m=2$, then $Q_{n \neq 2}$ may have at most
a simple pole there and $Q_2$ should have a double pole\ :
\begin{eqnarray}
Q_2 &=& - \frac{1}{4 (z-z_0)^2} (1+ O(z-z_0)) \quad \mbox{as} \quad z \rightarrow z_.
\label{condition2}
\end{eqnarray}
\end{itemize}
It is easily checked that the potentials that appear in the various
Schr\"odinger type differential equations in sections
\ref{semiclassical:subsec} and \ref{semiclassicalIrregular:subsec} satisfy these conditions.
\subsection{Theorems on Stokes Automorphisms}
Since all the equations listed in sections \ref{semiclassical:subsec}
and \ref{semiclassicalIrregular:subsec} satisfy the necessary
conditions, the theorems proved in \cite{NakanishiandIwaki} using
these conditions can be directly applied to our equations. There is
however, an interesting exception and we will comment on it
momentarily. In particular, the results of \cite{NakanishiandIwaki} include theorems
on the Stokes automorphisms that relate WKB resummed monodromies with
a given Borel resummation angle, to monodromies with another Borel
resummation angle.
We will now list the relevant results from these theorems. Consider a
closed curve $\gamma$ on the double cover $\hat{\Sigma}$ of the
Riemann surface $\Sigma$ encircling either a singularity or a turning
point. We then define the Voros symbol $ e^{V_\gamma}$ as a formal
power series using the integral
\begin{equation}
\label{eq:voros}
V_\gamma(\epsilon_1)=\oint_\gamma dz\ S_{\text{odd}}(z, \epsilon_1) \ .
\end{equation}
The Borel sums of the Voros symbol are then defined as
$S_\pm [ e^{V_\gamma}]$. They satisfy the Stokes automorphism
formula
\begin{eqnarray}
S_-[e^{V_\gamma}] &=& S_+ [ e^{V_\gamma}] ( 1 + S_+ [e^{V_{\gamma_0}}])^{-(\gamma_0,\gamma)}
\end{eqnarray}
whereby we suppose a simple flip, with the critical Stokes cycle being denoted by
$\gamma_0$, and $(\gamma_0,\gamma)$ is the intersection
number of the critical cycle with the cycle $\gamma$ defining the
Voros symbol. The resummations $S_\pm$ are the Borel resummations of
the Voros symbol on either side of (and close enough to) the critical
graph. The intersection numbers are defined using the
convention that, if the cycle $\gamma_1$ has the arrow pointing outwards in
the positive $x$ direction and the cycle $\gamma_2$, which crosses
$\gamma_1$, with the arrow pointing towards the upper half-plane, then
$(\gamma_1,\gamma_2)= +1$. When we have Borel sums on either side of a
pop rather than a flip, the Voros symbols (importantly, associated to closed cycles) are
trivially related
\begin{eqnarray}
\label{stokes pops}
S_-[e^{V_\gamma}] &=& S_+ [ e^{V_\gamma}] \, .
\end{eqnarray}
These two theorems govern the transformation of Voros symbols
associated to closed cycles. In the next section we will frequently
use results of these theorems to study global properties of our
differential equations.
In the case of the conformal SU$(2)$ gauge theory (with $N_f=4$
flavours) however, the extra assumptions of \cite{NakanishiandIwaki} are
not always fully satisfied. In particular, in this case we find that pairs of Stokes
graphs that are related by a simultaneous or double flip, excluded in
\cite{NakanishiandIwaki}. When such a double flip occurs, we show
that the formulae for the Stokes automorphisms derived
for single flips compose without change to give
the Stokes automorphism for the double flip. This is an extension of
the results of \cite{NakanishiandIwaki}. We will discuss this case
in detail in the next section.
\section{The Monodromy Group}
\label{monodromygroup}
In this section, we study global properties of the differential equations derived in section \ref{CFT}. The differential
equations are second order and hence have two linearly independent
global solutions. The solutions undergo a monodromy as we analytically continue them around a singular point. The monodromies,
defined up to a change of basis, form a group called the monodromy group. The monodromy
group of the differential equations we consider can be expressed
entirely in terms of two sets of quantities: (i) the characteristic
exponents $\nu_k$ at the singular points $s_k$
and (ii) the Borel resummed contour integrals of the WKB differential $S_{\text{odd}}$ around branch cuts,
which we denote by $u_{ij}$.
The connection formulae which relate the Borel
resummed wave functions in the various Stokes regions are sufficient
to completely determine the monodromy group associated to the relevant
null vector decoupling equation. The Borel resummed exact WKB contour integrals depend on the Borel resummation angle, (equivalently,
on the phase of $\epsilon_1$) and undergo Stokes automorphisms as a function of these parameters. Thus,
the expression of the monodromy group in terms of the resummed integrals varies, and we determine the explicit
transformation rules as we pass through a critical graph.
In this section, we calculate the monodromy groups, starting with the
simplest case of zero flavours, with no regular singular points
in the differential equation, and we end with the
conformal case ($N_f=4$) which has four regular singular points.
We stress the fact that there is a dictionary between the Borel resummation
angle $\theta$, and the phase of the zeroth order differential which is determined by the phase of $\epsilon_1$ in
our set-up. (See e.g. \cite{NakanishiandIwaki} for the details, which follow from the definition of the Borel
sum.)
We see that this dictionary is given a natural home in $\epsilon_1$-deformed ${\cal N}=2$ gauge theories.
The formal dependence on the Borel resummation angle that induces the Stokes automorphism, has a physical counterpart in the dependence of all non-perturbatively resummed monodromies
on the phase of the deformation parameter $\epsilon_1$.
\subsubsection*{A Brief Summary of our Analysis}
Throughout this section, we perform the calculation of the monodromy group in a strong coupling regime. In all the examples, we will plot the Stokes graphs emphasizing
the connectivity of the graphs and the choice of branch
cuts; we refer to \cite{GMN} for various possible sequences of
Stokes graphs. To illustrate the detailed coding of the monodromy group in terms of the
characteristic exponents and the resummed monodromies, as well as the ambiguity of their formal
expression in terms of the monodromies, we calculate the
monodromy groups associated to two distinct Stokes graphs. Equating the invariants constructed
from the monodromy groups of the two graphs gives us the Stokes
automorphism relating the variables in each description. We will thus find concrete descriptions of the monodromy group,
as well as the Stokes automorphisms that the exact WKB parameters undergo. The Stokes automorphisms must satisfy the theorems
of \cite{NakanishiandIwaki} and this fact serves as a consistency check of our analysis.
\subsection{Pure Super Yang-Mills}
\label{pureSYM}
The semi-classical null vector decoupling equation corresponding to the case of pure super Yang-Mills theory has been discussed in detail in \cite{Kashani-Poor:2015pca}.
The description was mostly in terms of variables that resulted after mapping the sphere onto a cylinder,
such that the differential equation became the Mathieu equation, and the monodromy group was coded
in the Floquet exponent. Below, we perform an equivalent analysis on the sphere, which will prepare
us to include flavours. A WKB analysis of the Mathieu equation can be found in \cite{HeMiao1,HeMiao2}
and further in \cite{Dunne} in the context of exact WKB and the 2d/4d correspondence.
The Stokes graph only depends on the leading potential term $Q_0(z)$. For the pure ${\cal N}=2$
super Yang-Mills theory, the zeroth order term is given by \eqref{PureNVD}
\begin{equation}
Q_0(z) =
\frac{\Lambda_0^2}{z^3(z-1)^2} +\frac{\widetilde{u}}{z^2(z-1)^2}
+ \frac{\Lambda_0^2}{z(z-1)^3} \, .
\end{equation}
\begin{figure}[ht!]
\centering
\includegraphics[width=170mm]{diagram2.pdf}
\caption{The two Stokes graphs of the $N_f=0$ case that are related by a simple flip. We also exhibit the contour used to calculate the monodromy matrix. These graphs were obtained with the parameters $\Lambda_0 = e^{-\mathrm{i}\frac{\pi}{4}}$ and $\widetilde{u} = -1+\mathrm{i}$, with the critical graph observed at $\theta = \pi$.}
\label{Nf=0graphs}
\end{figure}
In figure \ref{Nf=0graphs}, we exhibit various Stokes graphs in the
strong coupling region of the pure super Yang-Mills
theory.\footnote{Using the form of the $N_f = 0$ differential as in
\cite{GMN}, we are within the strong coupling region if we make the
choice $\Lambda = 1$ and $u = 1/2$. A series of conformal
transformations and rescalings relate the differential presented
here and the one presented in \cite{GMN}. At the end of this series of transformations, we are led to the choice of parameters presented in the caption of figure \ref{Nf=0graphs}.} We
first draw the critical graph \ref{Nf=0graphs} that has a finite
WKB line connecting the turning points $t_1$ and $t_2$. The Stokes
graphs we work with are related by a flip \cite{GMN} about this finite
WKB line.
\subsubsection*{The Monodromy Group}
In this case, there is a single independent generator
of the monodromy group and we choose the contour enclosing the
singularity $s_1$ and the turning point $t_1$ to be it.
Consider first the Stokes graph \ref{Nf=0graphs}(A). The
contour intersects two Stokes lines; the monodromy matrix is given by
\begin{align}
\begin{split}
M_{A, s_1} &= \left( \begin{array}{cc}
x &\ 0\\
0 &\ x^{-1} \end{array} \right)
\left( \begin{array}{cc}
1 &\ 0\\
-\mathrm{i} u_2 &\ 1 \end{array} \right)
\left( \begin{array}{cc}
1 &\ +\mathrm{i} u_1^{-1}\\
0 &\ 1 \end{array} \right) \, . \\
\end{split}
\end{align}
Note that the matrices are written from right to left as we go around the branch cut. The final matrix encodes the overall normalization factor as we return to the base point $x_0$. The variable $x$ which appears there is identified with the overall monodromy around the branch cut connecting $s_1$ and $t_1$.
We now turn to the second Stokes graph \ref{Nf=0graphs}(B). We see that the contour intersects four Stokes lines, including two lines arising from the flip. The monodromy matrix is given by
\begin{align}
\begin{split}
M_{B, s_1} &=
\left( \begin{array}{cc}
\tilde x &\ 0\\
0 &\ \tilde{x}^{-1} \end{array} \right)
\left( \begin{array}{cc}
1 &\ 0\\
-\mathrm{i}\tilde u_2 &\ 1 \end{array} \right)
\left( \begin{array}{cc}
1 &\ -\mathrm{i}\tilde u_2^{-1}\\
0 &\ 1 \end{array} \right)
\left( \begin{array}{cc}
1 &\ 0\\
+\mathrm{i}\tilde u_1 &\ 1 \end{array} \right)
\left( \begin{array}{cc}
1 &\ +\mathrm{i}\tilde u_1^{-1}\\
0 &\ 1 \end{array} \right)\\
&=\left( \begin{array}{cc}
\frac{\tilde{x}(\tilde u_1+\tilde u_2)}{\tilde u_2} &\ \frac{\mathrm{i} \tilde{x}}{\tilde u_1}\\
-i\frac{\tilde u_2}{\tilde{x}} &\ \frac{\tilde u_2}{\tilde{x}\tilde
u_1} \end{array} \right)\,.
\end{split}
\end{align}
We have denoted the variables in Stokes graph \ref{Nf=0graphs}(B) by variables with tildes since they correspond to a different
Borel resummation. The monodromy matrix $M_{A,s_1}$ must be equivalent to the monodromy matrix calculated on the basis of
graph (B), since the monodromy (equivalence class) is a property of the exact solutions on the Riemann surface $\Sigma$.
\subsubsection*{The Stokes Automorphism}
Above, we have the explicit expressions for the monodromy matrices for the two Stokes graphs.
The independent Stokes variables are given by $x$ and $u_{21}$ in graph \ref{Nf=0graphs}(A) and the tilde-variables in graph \ref{Nf=0graphs}(B).
Using this notation, we calculate the conjugation invariant traces of the two monodromy matrices:
\begin{align}
\mathrm{Tr}\,\left(M_{A, s_1} \right)&=x+\frac{1}{x}+\frac{1}{u_{21}x}\\
\mathrm{Tr}\,\left(M_{B, s_1} \right)&=\tilde{x}+\tilde{u}_{21}\tilde{x}+\frac{1}{\tilde{u}_{21}\tilde{x}}\, .
\end{align}
Requiring that the traces of the two monodromy matrices match leads to the map between the parameters appearing in the two graphs:
\begin{align}
u_{21}&=\tilde{u}_{21}\nonumber\\
x&=\tilde{x}(1+\tilde{u}_{21}) \, .
\end{align}
This agrees with the Stokes automorphisms derived in \cite{Kashani-Poor:2015pca}.
This is also consistent with the general analysis in \cite{NakanishiandIwaki}. Let us expand on this briefly:
the two Stokes graphs lie on either side of the $t_1-t_2$ flip in the critical graph \ref{Nf=0graphs}. Since the $t_1-t_2$ cycle
corresponding to $u_{12}$ has zero intersection number with itself, the variable $u_{12}$ is unaffected by the flip. However, the $x$ variable changes because the contour around the branch cut has intersection number $1$ with the $t_1-t_2$ cycle.
\subsection{One Flavour}
The Stokes graphs corresponding to the differential in the case of one flavour are determined by the corresponding zeroth order differential \eqref{Nf=1 semiclassical}:
\begin{equation}
Q_0(z) =\frac{\Lambda_1^2}{4z^4} + \frac{\Lambda_1m_1}{z^3(z-1)^2}-\frac{\Lambda_1^2}{4z(z-1)^3}
+\frac{\widetilde{u}}{z^2(z-1)^2} \, .
\end{equation}
{From} the form of the differential,
one can see that in the $z$-plane, there are three turning points and
two irregular singularities (at $z=0$ and $z=1$). The WKB triangulations are
given in figure 61 of \cite{GMN}. We consider a particular pair that
are separated by a flip\footnote{These are the first and the third
out of the six triangulations given in figure 61 of \cite{GMN}.} and
draw only the corresponding Stokes graphs. The two Stokes graphs
correspond to a flip about the $t_2-t_3$ finite Stokes line in the
critical graph (see figure \ref{Nf=1graphs}).
\subsubsection*{The Monodromy Group}
\begin{figure}[ht!]
\centering
\includegraphics[width=170mm]{diagram3.pdf}
\caption{The critical graph, the pair of Stokes graphs related by a flip and the contours that define the monodromy group for the $N_f=1$ case. The parameters chosen were $\Lambda_1 = 2$, $\widetilde{u} = -1/2$, and $m_1 = 1$, and the critical graph was observed at $\theta = \pi$.}
\label{Nf=1graphs}
\end{figure}
We proceed to calculate the monodromy group for both the Stokes
graphs. There are two irregular singularities in the graphs and one
expects two independent generators of the monodromy group. We choose
the two corresponding generators of the monodromy group as shown in
figure \ref{Nf=1graphs}. The contour around the singularity $s_2$ is
treated in much the same way as the irregular singular point in the
$N_f=0$ case, while the singularity $s_1$ behaves slightly differently.
Let us first consider Stokes graph \ref{Nf=1graphs}(A) and calculate the monodromy matrices; we find
\begin{align}
\begin{split}
M_{A, s_1} &=
\begin{pmatrix}
\nu_1 &0\\
0&\frac{1}{\nu_1}
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} u_1&1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_1}\\
0&1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_2}\\
0&1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_3}\\
0&1
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} u_3&1
\end{pmatrix} \, ,\\
M_{A, s_2}&=\left( \begin{array}{cc}
x &\ 0\\
0 &\frac{1}{x} \end{array} \right)
\left( \begin{array}{cc}
1 &\ 0\\
\mathrm{i} u_3x^2 &1 \end{array} \right)
\left( \begin{array}{cc}
1 &\ \frac{\mathrm{i}}{u_3x^2}\\
0 &1 \end{array} \right)
\left( \begin{array}{cc}
1 &\ 0\\
-\mathrm{i} u_2x^2 &1 \end{array} \right)
\left( \begin{array}{cc}
1 &\ -\frac{\mathrm{i}}{u_2x^2}\\
0 &\ 1 \end{array} \right)
\left( \begin{array}{cc}
1 &\ -\frac{\mathrm{i}}{u_1u_{12}^2x^2}\\
0 &1 \end{array} \right) \, .
\end{split}
\end{align}
The matrix element on the extreme left in the second monodromy matrix is the naive WKB monodromy around the branch cut connecting $t_3$ and $s_2$. This contribution $x$ satisfies the relation,
\begin{equation}
x\, u_{12}\nu_1=1\,.
\end{equation}
Let us now turn to the Stokes graph \ref{Nf=1graphs}(B).
The monodromy matrices are given by
\begin{align}
\begin{split}
M_{B, s_1} &=
\begin{pmatrix}
\nu_1 &0\\
0&\frac{1}{\nu_1}
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} \tilde{u}_1&1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{\tilde{u}_1}\\
0&1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{\tilde{u}_2}\\
0&1
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i}\tilde{u}_2&1
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} \tilde{u}_3&1
\end{pmatrix} \, ,\\
M_{B,s_2}&=\left( \begin{array}{cc}
\tilde{x} &\ 0\\
0 & \frac{1}{\tilde{x}}\end{array} \right)
\left( \begin{array}{cc}
1 &\ 0\\
\mathrm{i}\tilde{u}_3\tilde{x}^2 &1 \end{array} \right)
\left( \begin{array}{cc}
1 &\ \frac{-\mathrm{i}}{\tilde{u}_2\tilde{x}^2}\\
0 &1 \end{array} \right)
\left( \begin{array}{cc}
1 &\ \frac{-\mathrm{i}}{\tilde{u}_1\tilde{u}_{12}^2\tilde{x}^2}\\
0 &1 \end{array} \right) \,.
\end{split}
\end{align}
As before, we define $\tilde x=\frac{1}{\tilde u_{12}\nu_1}$.
\subsubsection*{The Stokes Automorphism}
Now that we have the two sets of monodromy matrices, we calculate the traces of the two sets and obtain
\begin{align}
\begin{split}
\mathrm{Tr}\,(M_{A,s_1})&= -\nu_1 \left(\frac{u_3}{u_1}+\frac{u_3}{u_2} \right)-\frac{1}{\nu_1}\left(\frac{u_1}{u_2}+\frac{u_1}{u_3}\right) \, ,\\
\mathrm{Tr}\,(M_{A,s_2})&=\frac{1}{\nu_1}(u_{31}+u_{21})+(u_{13}+u_{23})\nu_1 \, ,\\
\mathrm{Tr}\,(M_{B,s_1})&=-\nu_1\left(\frac{\tilde{u}_2}{\tilde{u}_1}+\frac{\tilde{u}_3}{\tilde{u}_2}+\frac{\tilde{u}_3}{\tilde{u}_1} \right)-\frac{1}{\nu_1}\frac{u_1}{u_2} \, ,\\
\mathrm{Tr}\,(M_{B,s_2})&=\frac{1}{\nu_1}(\tilde{u}_{21})+(\tilde{u}_{12}+\tilde{u}_{13}+\tilde{u}_{23})\nu_1 \, .
\end{split}
\end{align}
Substituting $u_{13}=u_{12}u_{23}$, and equating the expressions for
the traces in powers of $\nu_1$ (where we use the fact that the
characteristic exponents are true invariants of the differential
equation), we can extract the Stokes automorphism formulae for the
independent contour integrals $u_{21}$ and $u_{23}$, namely:
\begin{align}
\tilde{u}_{23}&=u_{23} \, ,\\
\tilde{u}_{21}&=u_{21}(1+u_{32}) \, .
\end{align}
Since there is more than one generator of the monodromy group, one can
calculate higher-order invariants by calculating traces of products of
the matrices. Using the Stokes automorphism, one can check that the
trace of the products also coincide, thus confirming the
identification of the monodromy group.
\subsection{Two Flavours}
In this section, we consider the SU$(2)$ gauge theory with two flavours.
We concentrate on the asymmetric configuration. The zeroth order potential function is given by \eqref{Nf=2 asymmetric semiclassical}
\begin{equation}
\label{Nf2Potential}
Q_0(z) = \frac{(m_3+m_4)^2}{4(z-1)^2}+ \frac{m_3 m_4}{z(1-z)} +\frac{\Lambda_2^2}{z^3} +\frac{\widetilde{u}}{z^2-z^3}
\, .
\end{equation}
In the $z$-plane, the quadratic differential has three singularities, and three turning points. One of these is an irregular singularity
at $z=0$.
As before, we work in a strong coupling limit, where $\widetilde{u}\ll\Lambda_2^2$. We consider the critical graph (see figure \ref{Nf=2graphs}),
and by a flip about the $t_1-t_3$ finite WKB line, obtain the two Stokes graphs, as shown in the figure
An important difference from the earlier cases is that we have regular singularities at $s_1$ and $s_2$.
\begin{figure}[ht!]
\centering
\includegraphics[width=150mm]{diagram4.pdf}
\caption{The critical graph and the Stokes graphs for the $N_f=2$ case. While plotting the figures, we used a potential that is
conformally equivalent to \eqref{Nf2Potential}. We set $\Lambda_2\rightarrow \mathrm{i}, \widetilde{u}\rightarrow \frac{1}{2},m_3\rightarrow 0,m_4\rightarrow -2$. The
two Stokes graphs presented were observed at $\theta=\frac{2\pi}{3}$ and $\theta=\frac{3\pi}{4}$.}
\label{Nf=2graphs}
\end{figure}
\subsubsection*{The Monodromy Group}
In order to calculate the monodromy group, we
first consider the Stokes graph \ref{Nf=2graphs}(A) and determine the generators
\begin{align}
\begin{split}
M_{A,s_1}& =\begin{pmatrix}
\nu_1 &0\\
0 &\frac{1}{\nu_1}
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_2\nu_1^2}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} u_2\nu_1^2 &1
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} u_3\nu_1^2 &1
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} u_1\nu_1^2 &1
\end{pmatrix
\begin{pmatrix}
1 &\frac{\mathrm{i}}{u_2}\\
0 &1
\end{pmatrix} \, ,\\
M_{A, s_2} &= \begin{pmatrix}
\nu_2 &0\\
0 &\frac{1}{\nu_2}
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} u_3u_{32}^2 &1
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} u_2 &1
\end{pmatrix} \, , \\
M_{A, s_3} &= \begin{pmatrix}
x &0\\
0 &\frac{1}{x}
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_2x^2}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &0\\
\mathrm{i} u_1\nu_1^2x^2 &1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_3\nu_1^2x^2}\\
0 &1
\end{pmatrix} \, .
\end{split}
\end{align}
In the above, $x$ is the naive WKB monodromy around the branch cut connecting $t_1$ and $s_3$. This contribution satisfies the relation,
\begin{equation}
u_{23}\, \nu_1\, \nu_2\, x=1 \,.
\end{equation}
A similar calculation for Stokes graph \ref{Nf=2graphs}(B), gives us the following monodromy matrices for circling the singularities
\begin{align}
\begin{split}
M_{B, s_1} &= \begin{pmatrix}
\nu_1 &0\\
0 &\frac{1}{\nu_1}
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{\tilde{u}_2\nu_1^2}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i}\tilde{u}_2\nu_1^2 &1
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i}\tilde{u}_3\nu_1^2 &1
\end{pmatrix
\begin{pmatrix}
1 &\frac{\mathrm{i}}{\tilde{u}_2}\\
0 &1
\end{pmatrix} \, , \\
M_{B, s_2} &= \begin{pmatrix}
\nu_2 &0\\
0 &\frac{1}{\nu_2}
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} \tilde{u}_1\tilde{u}_{32}^2 &1
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} \tilde{u}_3\tilde{u}_{32}^2 &1
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i}\tilde{u}_2 &1
\end{pmatrix} \, , \\
M_{B, s_3} &= \begin{pmatrix}
\tilde{x} &0\\
0 &\frac{1}{\tilde{x}}
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{\tilde{u}_2\tilde{x}^2}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{\tilde{u}_3\nu_1^2\tilde{x}^2}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &0\\
\mathrm{i}\tilde{u}_1\nu_1^2\tilde{x}^2 &1
\end{pmatrix}\, .
\end{split}
\end{align}
Again we have the relation,
\begin{equation}
u_{23} \, \nu_1\, \nu_2\, \tilde{x} = 1\,.
\end{equation}
\subsubsection*{Stokes Automorphisms}
We now compare the traces of the generators of the monodromy group:
\begin{align}
\begin{split}
\mathrm{Tr}\, M_{A,s_1} &= \mathrm{Tr}\, M_{B,s_1} = \nu_1 + \frac{1}{\nu_1} \, ,\\
\mathrm{Tr}\, M_{A,s_2} &= \mathrm{Tr}\, M_{B,s_2} = \nu_2 + \frac{1}{\nu_2} \, ,\\
\mathrm{Tr}\, M_{A,s_3} &=\frac{1}{\nu_1\nu_2}u_{32}+\frac{\nu_1}{\nu_2}u_{21}u_{32}+\nu_1\nu_2(u_{21}+\frac{1}{u_{32}}) \, ,\\
\mathrm{Tr}\, M_{B,s_3}
&=\frac{1}{\nu_1\nu_2}\tilde{u}_{32}(\tilde{u}_{21}\tilde{u}_{32}+1)+\frac{\nu_1}{\nu_2}\tilde{u}_{21}\tilde{u}_{32}+\nu_1\nu_2\frac{1}{\tilde{u}_{32}}
\,.
\end{split}
\end{align}
These equations illustrate a recurring feature: the
traces of the monodromy matrices around regular singular points will always be given by the critical exponents, with no $u_{ij}$ monodromy factors entering the expression. This is
because the Stokes lines are either
all going in or coming out at such regular singular points. As a result, the relevant Stokes
matrices are all either upper triangular or lower triangular,
respectively. This leads to the trivial nature of the trace. The
irregular singularity, on the other hand, has non-trivial structure
even at the level of the simple traces.
\\
Matching the traces between the graphs \ref{Nf=2graphs}(A) and \ref{Nf=2graphs}(B) leads to the Stokes automorphism relations,
\begin{align}\label{stokesautoNf=2}
\begin{split}
u_{31} &=\tilde{u}_{31} \, ,\\
u_{32} &=\tilde{u}_{32}(1+\tilde{u}_{31}) \, ,\\
u_{21} &=\tilde{u}_{21}(1+\tilde{u}_{31})^{-1} \,.
\end{split}
\end{align}
This is once again as expected from the general results of \cite{NakanishiandIwaki} and the intersection numbers between the various cycles.
As a consistency check on the monodromy matrices, we have also
computed the traces of products of matrices, and a
similar analysis as above confirms the Stokes automorphisms \eqref{stokesautoNf=2}.
\subsection{Three Flavours}
We move on to the SU$(2)$ theory with three flavours. The Seiberg-Witten differential is
\begin{equation}
\label{Nf3Potential}
Q_0(z) = \frac{(m_3+m_4)^2}{4(z-1)^2}+ \frac{m_3 m_4}{z(1-z)} +\frac{m_1\Lambda_3}{z^3}+\frac{\Lambda_3^2}{4z^4} +\frac{\widetilde{u}}{z^2-z^3}\,.
\end{equation}
There are four turning points and three singularities on the $z$-plane. The two Stokes graphs in
figure \ref{Nf3stokesgraphs} are related by a flip about the $t_2- t_3$ finite line in the critical graph.
\begin{figure}[ht!]
\centering
\includegraphics[width=150mm]{diagram5.pdf}
\caption{The critical graph and the Stokes graphs for the $N_f=3$ case. While plotting the figures, we used a potential that is
conformally equivalent to \eqref{Nf3Potential}. We set $\Lambda_3\rightarrow 1, \widetilde{u}\rightarrow 2,m_1\rightarrow -1,m_3\rightarrow 0,m_4\rightarrow -2$. The
two Stokes graphs presented were observed at $\theta=\frac{\pi}{2}$ and $\theta=\frac{7\pi}{12}$.}
\label{Nf3stokesgraphs}
\end{figure}
\subsubsection*{The Monodromy Group}
For Stokes graph \ref{Nf3stokesgraphs}(A), we find the generators of the monodromy group:
\begin{align}
\begin{split}
M_{A, s_1} &=
\begin{pmatrix}
\nu_1& 0\\
0 &\frac{1}{\nu_1}
\end{pmatrix
\begin{pmatrix}
1 &0 \\
\mathrm{i} u_3\nu_1^2 &1
\end{pmatrix
\begin{pmatrix}
1 & -\frac{\mathrm{i}}{u_4\nu_1^2\nu_2^2}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_1} \\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_2} \\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_3} \\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &0 \\
-\mathrm{i} u_3 &1
\end{pmatrix} \, ,\\
M_{A, s_2} &=
\begin{pmatrix}
\nu_2& 0\\
0 &\frac{1}{\nu_2}
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} u_4 \nu_2^2 & 1\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} u_3 & 1\end{pmatrix} \, ,
\\
M_{A, s_3} &=
\begin{pmatrix}
\nu_3 & 0\\
0 &\frac{1}{\nu_3}
\end{pmatrix}
\begin{pmatrix}
1 &0 \\
\mathrm{i} u_3\nu_3^2 &1
\end{pmatrix}
\begin{pmatrix}
1 &\frac{\mathrm{i}}{u_3\nu_3^2} \\
0 &1
\end{pmatrix}
\begin{pmatrix}
1 &0 \\
-\mathrm{i} u_2\nu_3^2 &1
\end{pmatrix}
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_2\nu_3^2} \\
0 &1
\end{pmatrix}
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_1\nu_3^2 u_{12}^2} \\
0 &1
\end{pmatrix}\\
&\hspace{3cm}\times
\begin{pmatrix}
1 &0 \\
-\mathrm{i} u_1\nu_3^2 u_{12}^2&1
\end{pmatrix}
\begin{pmatrix}
1 &0 \\
-\mathrm{i} u_4 u_{43}^2 &1
\end{pmatrix}
\begin{pmatrix}
1 &0 \\
-\mathrm{i} u_3 &1
\end{pmatrix}
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_3}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &0 \\
-\mathrm{i} u_3 &1
\end{pmatrix} \, .
\end{split}
\end{align}
For the Stokes graph \ref{Nf3stokesgraphs}(B), a similar calculation yields:
\begin{align}
\begin{split}
M_{B, s_1} &=
\begin{pmatrix}
\nu_1& 0\\
0 &\frac{1}{\nu_1}
\end{pmatrix
\begin{pmatrix}
1 &0 \\
\mathrm{i} \tilde{u_3}\nu_1^2 &1
\end{pmatrix
\begin{pmatrix}
1 &0 \\
\mathrm{i}\tilde{u}_2\nu_1^2 &1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{\tilde{u}_4\nu_1^2\nu_2^2} \\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{\tilde{u}_1} \\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{\tilde{u}_2} \\
0 &1
\end{pmatrix}\\
&\hspace{2cm}\times
\begin{pmatrix}
1 &0\\
-\mathrm{i}\tilde{u}_2 &1
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i}\tilde{u}_3 &1
\end{pmatrix}\, ,\\
M_{B,s_2} &=
\begin{pmatrix}
\nu_2 &0\\
0 &\frac{1}{\nu_2}
\end{pmatrix
\begin{pmatrix}
1 &0\\
-\mathrm{i} \tilde{u}_4\nu_2^2 & 1
\end{pmatrix
\begin{pmatrix}
1 &0 \\
-\mathrm{i} \tilde{u}_2 & 1
\end{pmatrix
\begin{pmatrix}
1 &0 \\
-\mathrm{i} \tilde{u}_3 &1
\end{pmatrix}\, ,
\\
M_{B, s_3} &=
\begin{pmatrix}
\nu_3 & 0\\
0 &\frac{1}{\nu_3}
\end{pmatrix}
\begin{pmatrix}
1 &0 \\
\mathrm{i} \tilde{u}_3\nu_3^2 &1 \\
\end{pmatrix}
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{\tilde{u}_2\nu_3^2} \\
0 &1 \\
\end{pmatrix}
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{\tilde{u}_1\nu_3^2 \tilde{u}_{12}^2} \\
0 &1
\end{pmatrix}\\
&\hspace{2cm}\times
\begin{pmatrix}
1 &0 \\
-\mathrm{i} \tilde{u}_1\nu_3^2 \tilde{u}_{12}^2&1
\end{pmatrix}
\begin{pmatrix}
1 &0 \\
-\mathrm{i} \tilde{u}_4 \tilde{u}_{43}^2 &1
\end{pmatrix}
\begin{pmatrix}
1 &0 \\
-\mathrm{i} \tilde{u}_3 &1
\end{pmatrix}
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{\tilde{u}_3}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &0 \\
-\mathrm{i} \tilde{u}_3 &1
\end{pmatrix}\, .
\end{split}
\end{align}
\subsubsection*{The Stokes Automorphism}
As in the previous examples, the Stokes automorphism can be obtained by comparing the invariants built out of the monodromy matrices.
The trace of the monodromy around the irregular singular point $s_3$ is non-trivial and it is important that we express it in terms of independent Stokes variables. The variables are constrained by the relation
\begin{equation}
u_{12}u_{34} \nu_1\nu_2\nu_3 = 1\,,
\end{equation}
and similarly for the $\tilde{u}$ variables. We solve for $u_{34}$ using this relation and choose the independent variables to be $u_{12}$ and $u_{23}$. In terms of these variables, we find that
\begin{align}
\mathrm{Tr}\, M_{A,s_3} &= -\nu_3 u_{12} - \nu_1\nu_2 u_{23}(1+u_{12}) - \frac{1}{\nu_3\, u_{12}} (1+u_{23}+u_{12}u_{23}) \,.
\end{align}
Similarly, from the monodromy around $s_3$ in Stokes graph \ref{Nf3stokesgraphs}(B) we find
\begin{equation}
\mathrm{Tr}\, M_{B,s_3} = -\nu_3\tilde{u}_{12}(\frac{1}{\tilde{u}_{23}}+1) - \nu_1\nu_2(\tilde{u}_{12}+\tilde{u}_{23}+\tilde{u}_{12}\tilde{u}_{23})-\frac{\tilde{u}_{23}}{\nu_3\tilde{u}_{12}}(1+\tilde{u}_{12}) \, .
\end{equation}
Matching the traces leads to the Stokes automorphisms:
\begin{align}
u_{23} &= \tilde{u}_{23} \, , \\
u_{12} &= \tilde{u}_{12}\left(1+\frac{1}{\tilde{u}_{23}} \right) \, .
\end{align}
As a check of our monodromy matrices, we computed the traces of the products of the monodromy matrices. These imply the same Stokes automorphism as above.
\subsection{The Conformal Theory}\label{ConfTh}
We consider Stokes graphs in a strong coupling region of the conformal SU$(2)$ theory. The zeroth order
potential has four regular singular points and four turning points.
\begin{figure}[ht!]
\centering
\includegraphics[width=165mm]{diagram6plusCG}
\caption{The critical graph and a pair of Stokes graphs for the conformal SU$(2)$ theory. A double flip relates one graph to the other. We refer the reader to \ref{BDoubleFlip} for details.}
\label{Nf=4graphs}
\end{figure}
In particular, we consider the Stokes graphs corresponding to two out
of the six triangulations in figure 74 of \cite{GMN} (see figure
\ref{Nf=4graphs}(A) and (C)).\footnote{Our graphs are topologically
equivalent to those appearing in \cite{GMN}.} It can be seen that
the two Stokes graphs are related by a double flip, a simultaneous
flip about the $t_1-t_3$ and $t_2-t_4$ finite WKB lines in the
critical graph. We realize the double flip as two alternative sequences of two single
flips. We provide the corresponding intermediate graphs after the
single flips and perform the calculation we have familiarized
ourselves with by now.
Let us first consider the flip from Stokes graph \ref{Nf=4graphs}(A) to \ref{Nf=4graphs}(B$'$) via the $t_1-t_3$ flip. The relevant Stokes graphs and contours that generate the monodromy group are given in figure \ref{t1t3flipcontours}.
\begin{figure}[ht!]
\centering
\includegraphics[width=150mm]{diagram7a.pdf}
\caption{The two Stokes graphs related by the $t_1-t_3$ flip.}
\label{t1t3flipcontours}
\end{figure}
\subsubsection*{The Monodromy Group}
For Stokes graph \ref{Nf=4graphs}(A), using the contours as shown in figure \ref{t1t3flipcontours}(A),
the monodromy matrices are given by
\begin{align}
\begin{split}
M_{A, s_1} &=
\begin{pmatrix}
\nu_1& 0\\
0 &\frac{1}{\nu_1}
\end{pmatrix
\begin{pmatrix}
1& 0\\
\mathrm{i} u_4\nu_1^2& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{u_1u_{13}^2\nu_1^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1&-\frac{\mathrm{i}}{u_4}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_4& 1
\end{pmatrix} \, , \\
M_{A,s_2} &=
\begin{pmatrix}
\nu_2& 0\\
0& \frac{1}{\nu_2}
\end{pmatrix
\begin{pmatrix}
1& 0\\
\mathrm{i} u_4\nu_2^2& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{\mathrm{i}}{u_4\nu_2^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_1u_{13}^2\nu_1^2\nu_2^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_3u_{13}^2\nu_1^2\nu_2^2& 1
\end{pmatrix}
\\
&\hspace{3cm}
\times
\begin{pmatrix}
1& 0\\
-iu_2u_{24}^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_4& 1
\end{pmatrix
\begin{pmatrix}
1& -\frac{\mathrm{i}}{u_4}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_4& 1
\end{pmatrix} \, ,\\
M_{A,s_3} &=
\begin{pmatrix}
\nu_3& 0\\
0& \frac{1}{\nu_3}
\end{pmatrix
\begin{pmatrix}
1& 0\\
\mathrm{i} u_4\nu_3^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
\mathrm{i} u_1u_{13}^2\nu_3^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
\mathrm{i} u_3\nu_3^2& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{u_2\nu_3^2\nu_4^2}\\
0& 1
\end{pmatrix}
\\
&
\hspace{3cm}
\times
\begin{pmatrix}
1& -\frac{\mathrm{i}}{u_3}\\
0& 1
\end{pmatrix}
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_3& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_1u_{13}^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_4& 1
\end{pmatrix} \, ,\\
M_{A,s_4} &=
\begin{pmatrix}
\nu_4& 0\\
0& \frac{1}{\nu_4}
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_2\nu_4^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_3& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_1u_{13}^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_4& 1
\end{pmatrix} \, .
\end{split}
\end{align}
Next we consider the Stokes graph \ref{Nf=4graphs}(B$'$) and move along the contours $C_k$ around the singularities as shown in figure \ref{t1t3flipcontours}(B$'$).
We compute the monodromy matrices:
\begin{align}
\begin{split}
M_{B', s_1} &=
\begin{pmatrix}
\nu_1& 0\\
0 &\frac{1}{\nu_1}
\end{pmatrix
\begin{pmatrix}
1& 0\\
\mathrm{i}\tilde{u}_4\nu_1^2& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_3\nu_1^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_1\tilde{u}_{13}^2\nu_1^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1&-\frac{\mathrm{i}}{\tilde{u}_4}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i}\tilde{u}_4& 1
\end{pmatrix} \, ,\\
M_{B',s_2} &=
\begin{pmatrix}
\nu_2& 0\\
0& \frac{1}{\nu_2}
\end{pmatrix
\begin{pmatrix}
1& 0\\
\mathrm{i} \tilde{u}_4\nu_2^2& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{\mathrm{i}}{\tilde{u}_4\nu_2^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i}\tilde{u}_1\tilde{u}_{13}^2\nu_1^2\nu_2^2& 1
\end{pmatrix}\\
&\hspace{3cm}\times
\begin{pmatrix}
1& 0\\
-i\tilde{u}_2\tilde{u}_{24}^2& 1
\end{pmatrix}
\begin{pmatrix}
1& 0\\
-\mathrm{i}\tilde{u}_4& 1
\end{pmatrix
\begin{pmatrix}
1& -\frac{\mathrm{i}}{\tilde{u}_4}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i}\tilde{u}_4& 1
\end{pmatrix} \, ,\\
M_{B',s_3} &=
\begin{pmatrix}
\nu_3& 0\\
0& \frac{1}{\nu_3}
\end{pmatrix
\begin{pmatrix}
1& 0\\
\mathrm{i} \tilde{u}_4\nu_3^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
\mathrm{i} \tilde{u}_3\nu_3^2& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_2\nu_3^2\nu_4^2}\\
0& 1
\end{pmatrix}
\\
&\hspace{3cm}
\times
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_1}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& -\frac{\mathrm{i}}{\tilde{u}_3}\\
0& 1
\end{pmatrix}
\begin{pmatrix}
1& 0\\
-\mathrm{i} \tilde{u}_3& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} \tilde{u}_4& 1
\end{pmatrix} \, ,\\
M_{B',s_4} &=
\begin{pmatrix}
\nu_4& 0\\
0& \frac{1}{\nu_4}
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} \tilde{u}_2\nu_4^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i}\tilde{u}_3& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} \tilde{u}_4& 1
\end{pmatrix} \, .
\end{split}
\end{align}
We now implement the $t_2-t_4$ flip to go from Stokes graph \ref{Nf=4graphs}(B$'$) to graph \ref{Nf=4graphs}(C). The relevant Stokes graphs and
contours are given in figure \ref{t2t4flipcontours}. Notice that the
base point of the contours in figure \ref{t2t4flipcontours} is
different from that used in figure \ref{t1t3flipcontours}, however,
this is irrelevant because the
monodromy group is independent of the choice of a base point.
\begin{figure}[ht!]
\centering
\includegraphics[width=150mm]{diagram7b.pdf}
\caption{The $t_2-t_4$ flip and the contours around the singularities.}
\label{t2t4flipcontours}
\end{figure}
The monodromy matrices for the (B$'$) graph are given by
\begin{align}
\begin{split}
M_{B',s_1} &=
\begin{pmatrix}
\nu_1& 0\\
0& \frac{1}{\nu_1}
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{u_3\nu_1^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{u_1u_{13}^2\nu_1^2}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{u_4}\\
0& 1
\end{pmatrix} \, ,\\
M_{B',s_2} &=
\begin{pmatrix}
\nu_2& 0\\
0& \frac{1}{\nu_2}
\end{pmatrix
\begin{pmatrix}
1& \frac{\mathrm{i}}{u_4\nu_2^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_1u_{13}^2\nu_1^2\nu_2^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_2u_{24}^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_4& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{u_4}\\
0& 1
\end{pmatrix} \, ,\\
M_{B',s_3} &=
\begin{pmatrix}
\nu_3& 0\\
0& \frac{1}{\nu_3}
\end{pmatrix
\begin{pmatrix}
1& 0\\
\mathrm{i} u_3\nu_3^2& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{u_2\nu_3^2\nu_4^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{u_1}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{u_3}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_3& 1
\end{pmatrix} \, ,\\
M_{B',s_4} &=
\begin{pmatrix}
\nu_4& 0\\
0& \frac{1}{\nu_4}
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_4\nu_4^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_2\nu_4^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i} u_3& 1
\end{pmatrix} \, .
\end{split}
\end{align}
Finally, we consider Stokes graph \ref{Nf=4graphs}(C) and calculate the generators of the monodromy group
\begin{align}
\begin{split}
M_{C,s_1} &=
\begin{pmatrix}
\nu_1& 0\\
0& \frac{1}{\nu_1}
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_3\nu_1^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_1\tilde{u}_{13}^2\nu_1^2}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_2\tilde{u}_{24}^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1 &-\frac{\mathrm{i}}{\tilde{u}_4}\\
0& 1
\end{pmatrix} \, ,\\
M_{C,s_2} &=
\begin{pmatrix}
\nu_2& 0\\
0& \frac{1}{\nu_2}
\end{pmatrix
\begin{pmatrix}
1& \frac{\mathrm{i}}{\tilde{u}_4\nu_2^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{\mathrm{i}}{\tilde{u}_2\tilde{u}_{24}^2\nu_2^2}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i}\tilde{u}_1\tilde{u}_{13}^2\nu_1^2\nu_2^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i}\tilde{u}_2\tilde{u}_{24}^2& 1
\end{pmatrix}\\
&\hspace{3cm}\times
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_2\tilde{u}_{24}^2} \\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_4}\\
0& 1
\end{pmatrix} \, ,\\
M_{C,s_3} &=
\begin{pmatrix}
\nu_3& 0\\
0& \frac{1}{\nu_3}
\end{pmatrix
\begin{pmatrix}
1& 0\\
\mathrm{i}\tilde{u}_3\nu_3^2& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_4\nu_3^2\nu_4^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_2\nu_3^2\nu_4^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_1}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-\mathrm{i}}{\tilde{u}_3}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i}\tilde{u}_3& 1
\end{pmatrix} \, ,\\
M_{C,s_4} &=
\begin{pmatrix}
\nu_4& 0\\
0& \frac{1}{\nu_4}
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i}\tilde{u}_4\nu_4^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\mathrm{i}\tilde{u}_3& 1
\end{pmatrix} \, .
\end{split}
\end{align}
\subsubsection*{The Stokes Automorphism for the Double Flip}
The monodromy matrices
obtained above by encircling the singularities in both the pairs of
graphs in the sequence of flips have standard traces, since all the singularities are regular.
Hence, in order to compute the transformation of Voros symbols, we compute the traces
of products of monodromy matrices. We express the traces
in terms of the variables $u_{13}$ and $u_{34}$ in the $t_1-t_3$ flip and in terms
of the variables $u_{42}$ and $u_{21}$ in the $t_2-t_4$ flip.
Since the entries of the matrices are a bit cumbersome, we merely
present the results; here the variables in a given Stokes graph are denoted with the appropriate superscript:
This gives us the Stokes automorphism relations.
\begin{eqnarray}
\label{Stokes1u}
u_{13}^A=u_{13}^{B'},\ \ u_{34}^A=u_{34}^{B'}(1+u_{13}^{B'})
\end{eqnarray}
\begin{eqnarray}
\label{Stokes2u}
u_{42}^{B'}=u_{42}^C,\ \ u_{21}^{B'}=u_{21}^C(1+u_{42}^C)
\end{eqnarray}
Upon composing the Stokes relations from the two single flips, we
obtain the desired Stokes relation for the double flip from Stokes
graph (A) to (C).
If we consider a counter-clockwise loop encircling both the branch cuts
and the four singularities in graphs
(A), (B$'$) and (C), we get the relation,
\begin{eqnarray}
\label{conditionA}
u_{13}u_{42}\nu_1\nu_2\nu_3\nu_4=1
\end{eqnarray}
There is an analogous condition that is given by
\begin{eqnarray}
\label{conditionB}
u_{43}u_{12}\nu_1\nu_2\nu_3\nu_4=1 \,.
\end{eqnarray}
Using these and the Stokes automorphisms for the sequence of flips, we obtain the following relations for the independent variables of the Stokes graphs (A) and (C):
\begin{align}
u_{13}^A &=u_{13}^C \cr
u_{34}^A &=u_{34}^C\frac{1+u_{13}^C}{1+u_{42}^C} \,.
\end{align}
Because the double flip is composed of single flips, each taking place within their own arena,
the final result for the double flip is a composition of the
result for single flips \cite{NakanishiandIwaki}.
So far, we have implemented the double flip via a sequence of two
single flips, (A)$\rightarrow$ (B$'$)$\rightarrow$ (C) as in figure
\ref{Nf=4graphs}. The double flip can equivalently be implemented by
a different sequence of two single flips, (A)$\rightarrow$ (B)$\rightarrow$ (C) as shown in figure \ref{Nf=4graphs}. We have checked that this
results in the same Stokes automorphism relations as were arrived at
earlier. This is further confirmation of the rules for computing the monodromy
groups and of our resolution of the double flip into two
single flips.
\subsubsection*{Pops}
Finally, we consider two Stokes graphs that are
related by a pop rather than a flip, in the conformal gauge theory.
We concentrate on the situation depicted in figure \ref{pops} (see appendix \ref{sec:StokesGraphs} for more details).
This corresponds to
the degenerate triangulations in figure 74 of \cite{GMN}.
\begin{figure}[ht!]
\centering
\includegraphics[width=150mm]{diagram8}
\caption{The two Stokes graphs related by a pop. We refer the reader to \ref{BPop} for details.}
\label{pops}
\end{figure}
The pop is expected to give rise to a trivial Stokes automorphism for
our closed loop Voros symbols. We first
consider graph \ref{pops}(A) and calculate the generators of the monodromy
group. As the closed contour goes around $s_1$ counter-clockwise,
\begin{align}
M_{A,s_1} &=
\begin{pmatrix}
\nu_1& 0\\
0& \frac{1}{\nu_1}
\end{pmatrix
\begin{pmatrix}
1& -\frac{i}{u_4\nu_1^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{i\nu_2^2}{u_2u_{21}^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{-i}{u_1}\\
0& 1
\end{pmatrix
\underbrace{
\begin{pmatrix}
0& i\\
i& 0
\end{pmatrix
\begin{pmatrix}
1& 0\\
-\frac{i\nu_2^2}{u_1}& 1
\end{pmatrix
\begin{pmatrix}
0& -i\\
-i& 0
\end{pmatrix}
\begin{pmatrix}
1& -\frac{i}{u_2}\\
0& 1
\end{pmatrix} \, .
\end{align}
Next we compute the monodromy matrix as we traverse a closed contour that goes around the singularity $s_2$
\begin{align}
\begin{split}
M_{A,s_2} &=
\begin{pmatrix}
\nu_2& 0\\
0& \frac{1}{\nu_2}
\end{pmatrix
\begin{pmatrix}
1 &\frac{i}{u_2\nu_2^2}\\
0 &1
\end{pmatrix
\underbrace{
\begin{pmatrix}
0& i\\
i& 0
\end{pmatrix
\begin{pmatrix}
1& 0\\
\frac{i}{u_1}& 1
\end{pmatrix
\begin{pmatrix}
0& -i\\
-i& 0
\end{pmatrix}
\begin{pmatrix}
0& i\\
i& 0
\end{pmatrix
\begin{pmatrix}
1& iu_1\\
0& 1
\end{pmatrix
\begin{pmatrix}
0& -i\\
-i& 0
\end{pmatrix}\\&\time
\underbrace{
\begin{pmatrix}
0& i\\
i& 0
\end{pmatrix
\begin{pmatrix}
1& 0\\
\frac{-i\nu_2^2}{u_1}& 1
\end{pmatrix
\begin{pmatrix}
0& -i\\
-i& 0
\end{pmatrix}
\begin{pmatrix}
1& -\frac{i}{u_2}\\
0& 1
\end{pmatrix} \, .
\end{split}
\end{align}
Around the singularity $s_3$ and $s_4$, we find
\begin{align}
\begin{split}
M_{A,s_3}&=
\begin{pmatrix}
\nu_3& 0\\
0& \frac{1}{\nu_3}
\end{pmatrix
\begin{pmatrix}
1& 0\\
-iu_2\nu_3^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
\frac{iu_4u_{34}^2}{\nu_4^2}& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-iu_3 &1
\end{pmatrix}\\
&\hspace{2cm}\times
\underbrace{
\begin{pmatrix}
0& i\\
i& 0
\end{pmatrix
\begin{pmatrix}
1& \frac{-iu_3}{\nu_4^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
0& -i\\
-i& 0
\end{pmatrix}
\begin{pmatrix}
1& 0\\
-iu_4& 1
\end{pmatrix} \, ,\\
M_{A,s_4}&=
\begin{pmatrix}
\nu_4& 0\\
0& \frac{1}{\nu_4}
\end{pmatrix
\begin{pmatrix}
1& 0\\
iu_4\nu_4^2& 1
\end{pmatrix
\underbrace{
\begin{pmatrix}
0& i\\
i& 0
\end{pmatrix
\begin{pmatrix}
1& iu_3\\
0& 1
\end{pmatrix
\begin{pmatrix}
0& -i\\
-i& 0
\end{pmatrix}
\begin{pmatrix}
0& i\\
i& 0
\end{pmatrix
\begin{pmatrix}
1 &0\\
\frac{i}{u_3}& 1
\end{pmatrix
\begin{pmatrix}
0& -i\\
-i& 0
\end{pmatrix}\\&\time
\underbrace{
\begin{pmatrix}
0& i\\
i& 0
\end{pmatrix
\begin{pmatrix}
1& \frac{-iu_3}{\nu_4^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
0& -i\\
-i& 0
\end{pmatrix}
\begin{pmatrix}
1& 0\\
-iu_4& 1
\end{pmatrix} \, .
\end{split}
\end{align}
In the above, the set of matrices clubbed together by an underbrace gives the rule
for crossing a Stokes line that runs through a branch cut into the other sheet and approaches the turning point.
We repeat the above exercise for Stokes graph \ref{pops}(B) and the monodromy matrices
around the singularities in this case are given by
\begin{align}
\begin{split}
M_{B_{s_1}} &=
\begin{pmatrix}
\nu_1& 0\\
0& \frac{1}{\nu_1}
\end{pmatrix
\begin{pmatrix}
1& \frac{-i}{u_4\nu_1^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& \frac{i\nu_2^2}{u_2u_{21}^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
1 &\frac{-i\nu_2^2}{u_1}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &\frac{-i}{u_1}\\
0 &1
\end{pmatrix
\begin{pmatrix}
1 &\frac{-i}{u_2}\\
0 &1
\end{pmatrix} \, , \\
M_{B_{s_2}}&=
\begin{pmatrix}
\nu_2& 0\\
0& \frac{1}{\nu_2}
\end{pmatrix
\begin{pmatrix}
1& \frac{i}{u_2\nu_2^2}\\
0& 1
\end{pmatrix
\begin{pmatrix}
0& i\\
i& 0
\end{pmatrix
\begin{pmatrix}
1& -iu_1\nu_2^2\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
\frac{i}{u_1}& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
iu_1& 1
\end{pmatrix
\begin{pmatrix}
0& -i\\
-i& 0
\end{pmatrix
\begin{pmatrix}
1& \frac{-i}{u_2}\\
0& 1
\end{pmatrix} \, ,\\
M_{B_{s_3}}&=
\begin{pmatrix}
\nu_3& 0\\
0& \frac{1}{\nu_3}
\end{pmatrix
\begin{pmatrix}
1& 0\\
-iu_2\nu_3^2& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
\frac{iu_4u_{34}^2}{\nu_4^2}& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
\frac{-iu_3}{\nu_4^2}& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-iu_3& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
-iu_4& 1
\end{pmatrix} \, ,\\
M_{B_{s_4}}&=
\begin{pmatrix}
\nu_4& 0\\
0& \frac{1}{\nu_4}
\end{pmatrix
\begin{pmatrix}
1& 0\\
iu_4\nu_4^2& 1
\end{pmatrix
\begin{pmatrix}
0& -i\\
-i& 0\\
\end{pmatrix
\begin{pmatrix}
1& 0\\
\frac{-i}{u_3\nu_4^2}& 1
\end{pmatrix
\begin{pmatrix}
1& iu_3\\
0& 1
\end{pmatrix
\begin{pmatrix}
1& 0\\
\frac{i}{u_3}& 1
\end{pmatrix
\begin{pmatrix}
0& i\\
i& 0
\end{pmatrix
\begin{pmatrix}
1& 0\\
-iu_4& 1
\end{pmatrix} \, .
\end{split}
\end{align}
Again, it can be checked that the monodromy matrices have standard
traces in both the graphs.
Similarly it can be checked that the product of matrices from both the graphs have the same traces.
Thus, we have a consistency check on our calculation, which is the triviality of the Stokes automorphism acting on the
Voros symbols of graphs related by a pop \eqref{stokes pops}.
\section{The Gauge Theory Perspective}
\label{gaugetheory}
We have computed the monodromy groups associated to the differential
equations governing the instanton partition function with surface
operator insertion in terms of the exponents $\nu_i$ characterizing
the singularities, and the Borel resummed monodromies $u_{ij}$. The
characteristic exponents are readily calculated in terms of the masses
of the gauge theory using the explicit expression of the differential
equation. In this section, we further relate the exact
WKB parameters $u_{ij}$ with the Seiberg-Witten periods $a$ and $a_D$,
which are leading order approximations. As a result, we obtain the
monodromy groups in terms of the (deformed, resummed) gauge theory
data. Next, we present ideas on how to exploit this information
to obtain non-perturbative corrections to the prepotential.
\subsection{The Seiberg-Witten Variables}
We start by relating the monodromies $u_{ij}$ to more standard Seiberg-Witten
data. In the following figures, we mark only the turning points and the
singularities on the Riemann surface, and identify the $\hat{\alpha}$
and $\hat{\beta}$ cycles of the genus one Seiberg-Witten curve. For the conformal case
\cite{Ashok:2015gfa}, the cycles are identified as in figure \ref{Nf=4cycles}.
The cycles have a smooth limit when the masses are
set to zero, i.e. when the turning points coincide with the
singularities. Further, in \cite{Ashok:2015gfa}, it was explicitly checked that the prepotential of the conformal theory, obtained by calculating the period integrals with this choice of cycles, matches the results from equivariant localization methods. Once this identification is made, it is possible to go
down in the number of flavors sequentially, each time identifying the $\hat{\alpha}$ and
$\hat{\beta}$ cycles, until we finally reach the $N_f=0$ theory, where we obtain agreement with the results of \cite{Kashani-Poor:2015pca}.
\subsubsection*{The Conformal Theory}
From figure \ref{Nf=4cycles}, we read off the identification
\begin{align}
u_{13}\nu_1\nu_3=e^{a},\ \ u_{34}\nu_3\nu_4=e^{a_D} \, .
\end{align}
\begin{figure}[ht!]
\centering
\includegraphics[width=75mm]{cycles-A}
\caption{$N_f=4$ cycles}
\label{Nf=4cycles}
\end{figure}
\noindent
The Stokes automorphisms we have derived for the conformal theory then imply the following relations for the gauge theory variables between the Stokes graphs \ref{Nf=4graphs}(A) and \ref{Nf=4graphs}(C):
\begin{align}
(e^a)_{A} &=(e^{a})_C\,\, ,\nonumber\\
(e^{a_D})_A &=
(e^{a_D})_C
\Big[
\frac{1+ \nu_1^{-1}\nu_3^{-1}(e^{a})_C}
{1+\nu_2^{-1}\nu_4^{-1}(e^{-a})_C}
\Big] \, .
\end{align}
We will comment on the meaning of these relations after we list the cycles and appropriate gauge theory variables for the asymptotically free cases.
\subsubsection*{Three Flavours}
From figure \ref{Nf=3cycles}, for the three flavours case, we find the relation
\begin{align}
u_{12}\nu_1\nu_2=e^{a},\ \ u_{23}\nu_2\nu_3=e^{a_D} \, .
\end{align}
\begin{figure}[ht!]
\centering
\includegraphics[width=75mm]{cycles-B}
\caption{$N_f=3$ cycles}
\label{Nf=3cycles}
\end{figure}
\noindent
As before we can write the Stokes automorphisms in terms of the gauge
theory variables but we suppress these details and only give the
choice of cycles in our subsequent examples.
\subsubsection*{Two Flavours}
\noindent
For two flavours, from figure \ref{Nf=2cycles} we have,
\begin{align}
u_{23}\nu_1\nu_2=e^{a},\ \ u_{31}\nu_1=e^{a_D} \, .
\end{align}
\begin{figure}[ht!]
\centering
\includegraphics[width=75mm]{cycles-C}
\caption{$N_f=2$ cycles}
\label{Nf=2cycles}
\end{figure}
\subsubsection*{One Flavour}
\noindent
When we are left with a single flavour, we find from figure \ref{Nf=1cycles}
\begin{align}
u_{12}\nu_1=e^{a},\ \ u_{23}=e^{a_D} \, .
\end{align}
\begin{figure}[ht!]
\centering
\includegraphics[width=75mm]{cycles-D}
\caption{$N_f=1$ cycles}
\label{Nf=1cycles}
\end{figure}
\subsubsection*{Pure Super Yang-Mills}
\begin{figure}[ht!]
\centering
\includegraphics[width=75mm]{cycles-E}
\caption{$N_f=0$ cycles}
\label{Nf=0cycles}
\end{figure}
Finally, for pure super Yang-Mills, we have from figure \ref{Nf=0cycles}
\begin{align}
u_{21}=e^{a_D},\ \ x=e^{a} \, .
\end{align}
The Stokes automorphisms derived in Section \ref{pureSYM} can then be
reinterpreted in terms of the gauge theory variables as a relation
between $a$ and $a_D$,
\begin{align}
e^{a_D} &=e^{\tilde{a}_D}\, ,\nonumber\\
e^{a} &=e^{\tilde{a}}(1+e^{\tilde{a}_D})\, .
\end{align}
One can check that these precisely coincide with the Stokes automorphisms obtained for the pure super Yang-Mills case in \cite{Kashani-Poor:2015pca}.
\subsection{The Non-Perturbative Prepotential}
In this subsection, we restrict ourselves to some preliminary remarks on how to
exploit the results we have obtained on the monodromy group to find non-perturbative
corrections to the prepotential (parameterized in terms of the given exact
WKB monodromy data). We note that the expressions $e^{a(\epsilon_1)}$
and $e^{a_D(\epsilon_1)}$ used above (where we have now rendered
explicit the $\epsilon_1$ dependence) refer to the exact WKB, Borel
resummed $\epsilon_1$ perturbation series. At leading order in
$\epsilon_1$, the first term in the expansion matches the
Seiberg-Witten periods $a(0)$ and $a_D(0)$. In the $\epsilon_1$-deformed theory, these are corrected as a perturbation series in $\epsilon_1$. In
the exact WKB approach, these perturbation series are Borel resummed,
but the result depends on the region of Borel resummation, and the
resummed expressions $e^{a(\epsilon_1)}$ depend on the phase of Borel
resummation, or equivalently, on the phase of $\epsilon_1$. Thus, the
above identifications undergo the Stokes automorphisms of which we have
discussed many an example. Note that when we refer to such expressions, we
have in mind that they are valid with a given resummation angle.
The monodromy group itself, and in particular the gauge invariant
traces of products of monodromy matrices, are exact invariants of the
solution to the differential equations. In
\cite{Kashani-Poor:2015pca}, it was proposed in the context of the
theory with zero flavours to solve the equation for the Borel
resummed period $a(\epsilon_1)$ in terms of the invariant quantity
$a^{\text{exact}}$, determined by the exact periodicity (or Floquet exponent,
or monodromy) of the exact solution to the (Mathieu) differential
equation. This solution is valid at a particular resummation angle,
and $a$ is identified with $a^{\text{exact}}$, up to non-perturbative corrections.
Near a point in moduli space where a perturbative expansion is possible, like the
weak, magnetic or dyonic point, we can then solve the relation in terms of
a transseries \cite{Costin}.
On the other hand, since in this case $a_D(\epsilon_1)$ is independent
of the Borel resummation angle, we can posit
$a_D^{\text{exact}}=a_D(\epsilon_1)$ (independently of the specification of
the resummation angle).
The variable $u$ determining the derivative
of the prepotential is known as a function of $a$, and therefore as a function of $a^{\text{exact}}$. Inverting this relation and calculating the dual period integrals allows one to calculate non-perturbative corrections to the prepotential ${\cal F}$.
Thus, one application of the identifications above is to attempt to integrate up the
non-perturbative relation between the period $a$ and
the parameters coding the exact monodromy group, towards
non-perturbative corrections to the prepotential ${\cal F}$. The encompassing
case in which to execute this program is the conformal
$N_f=4$ theory.
\section{Conclusions}
\label{conclusions}
In our work, we filled in many details of the connection between
conformal field theory and four-dimensional gauge theory which was
made handily available starting from the matching of partition
functions and correlators in \cite{Alday:2009fs}. We applied the
technology developed in \cite{Awata:2010bz,AwataYamada,Gaiotto:2012sf,Gaiotto:2009ma} to study in greater detail
SU$(2)$ super Yang-Mills theories with a varying number of flavours. We thus provided a
complete list of $\epsilon_i$-deformed differential equations
satisfied by the five-point conformal block with surface operator
insertion, in the irregular conformal block limit. Subsequently, we
took a semi-classical limit and analyzed the resulting differential equations with
exact WKB methods. We used this technology to give
a detailed parameterization of the monodromy groups in terms of
Voros symbols and external conformal dimensions or, in the language of gauge theory, in terms of Borel resummed $\epsilon_1$-deformed Seiberg-Witten periods and masses. The Borel
resummed variables depend intrinsically on the resummation angle (and
the phase of $\epsilon_1$), and we illustrated how to bridge this
ambiguity using Stokes automorphisms, with at least one illustration for each
number of flavours.
The Stokes automorphisms we obtained are consistent with the general
theorems proved in \cite{NakanishiandIwaki}. The one subtlety arose
in the case of the conformal gauge theory: in this case, we analyzed a
pair of Stokes graphs that were related by a simultaneous flip along
two independent finite WKB lines. In such a case, we showed that the
resulting Stokes automorphisms could be obtained by treating the
double-flip as a sequence of single flips. We demonstrated consistency of
this approach by checking that the final Stokes automorphisms between
the Stokes graphs were independent of the order in which the single
flips were taken.
We believe it is instructive to develop these most elementary of
${\cal N}=2$ models in still further detail. There are many avenues to
explore. As was mentioned in the previous section, it should be
possible to calculate non-perturbative corrections in $\epsilon_1$ to
the prepotential, by suitably generalizing the analysis that was done for
the case of pure super Yang-Mills \cite{Kashani-Poor:2015pca}. Similarly, it
should be informative to match the cluster algebra
description of Stokes automorphisms of \cite{NakanishiandIwaki} to
the cluster algebra description of the spectrum of BPS states
\cite{KS}. (See also e.g. \cite{Gaiotto:2010be,Alim:2011kw,NakanishiandIwaki}.)
In the present work, we studied general properties of the Borel
resummed wave-functions of the null vector decoupling equations in the
semi-classical limit, without focusing on the specific form of the
wave-function. An interesting direction would be to study in greater
detail the dependence of the five-point conformal block on the
insertion point of the degenerate operator. This should yield
non-trivial information about the gauge theory in the presence of a
surface operator. Most interestingly,
perturbative and non-perturbative corrections in $\epsilon_2$ can now
be studied from the differential equation point of view.
All these projects have straightforward extensions, both to the
${\mathcal N} = 2^\star$ theory, where the Riemann surface is of genus
one \cite{Alday:2009aq}, as well as to higher rank conformal gauge
theories, the simplest of which is super Yang-Mills theory with
SU$(N)$ gauge group and $2N$ fundamental flavours. As explained in
Section \ref{gaugetheory}, one not only needs the description of the
monodromy group, but also requires a local expansion of the period
integrals in terms of the Coulomb moduli in order to carry out the
goal of calculating corrections to the prepotential that are
non-perturbative in $\epsilon_1$. For the ${\mathcal N} = 2^\star$
theory with gauge group SU$(2)$, the null vector decoupling equation
for the toroidal block has been well studied in the semi-classical
limit (see e.g.
\cite{Fateev:2009aw,Piatek:2013ifa, He:2014yka}) and the instanton series for the prepotential has been resummed
in terms of modular functions \cite{Kashani-Poor:2014mua}. One can therefore
attempt to study the monodromy group along the lines of the present
paper and hope to carry out the proposal put forward in Section
\ref{gaugetheory}. For higher rank ${\mathcal N}=2^\star$ theories,
there has been much progress on the gauge theory side in resumming the
instanton expansion for the prepotential using modular anomaly
equations in deformed gauge theories with arbitrary gauge group
\cite{Billo:2015ria, Billo:2015jta}. It remains an open problem to
reproduce these successes using conformal field theory methods and to
do the corresponding WKB analysis.
For the higher rank (and undeformed) SQCD theories, the instanton series has been resummed in a special locus with $\mathbb{Z}_N$ symmetry \cite{Ashok:2015cba, Ashok:2016oyh}. From the CFT approach to the problem, one has to work with Toda theory \cite{Wyllard:2009hg} and the corresponding null vector decoupling equations in Toda have been analyzed recently in \cite{PoghossianToda}. In the semi-classical limit, such differential equations can also be derived using a saddle point analysis of the Nekrasov integrand \cite{NPS13} and the resulting deformed Seiberg-Witten curve. It would be interesting to analyze these higher order differential equations using exact WKB methods and make contact with the general approach to these systems using spectral networks \cite{Gaiotto:2012rg, Gaiotto:2012db}.
\section*{Acknowledgments}
We would like to thank Anirban Basu, Soumyadeep Bhattacharya, Eleonora Dell'Aquila, Anshuman Maharana, Gautam Mandal, Shiraz Minwalla, and Ashoke Sen for useful discussions. We would especially like to thank Eleonora Dell'Aquila for help with the pictures.
S.A. and D.P.J. would like to thank TIFR, Mumbai for hospitality. D.P.J. would also like
to thank IMSc, Chennai for hospitality during the course of this work.
D.P.J. was partially supported by DAE XII-plan grant 12-R\& D-HRI-5.02-0303.
J.T. would like to acknowledge support from the grant ANR-13-BS05-0001.
\begin{appendix}
\section{The Null Vector Decoupling with Irregular Blocks}
\label{diffeqns}
In this section, we derive in detail the null vector decoupling equations that involve irregular conformal blocks.
These equations were summarized in a different form in \cite{Awata:2010bz}.
Our final results are listed in section \ref{CFT}.
\subsection{One Irregular Puncture}
The case of $N_f=4$ is standard and is described in section \ref{CFT} in sufficient detail.
We turn to the conformal block involving an irregular puncture. This involves
rendering one flavour very massive.
\subsubsection{$N_f=3$}
{From} the gauge theory point of view, the flavour decoupling is carried out by taking the limit:
\begin{equation}
m_2 \rightarrow \infty\qquad q\rightarrow 0\quad \text{with}\quad \Lambda_3 = q\, m_2 \quad \text{finite}\,.
\end{equation}
The parameter $\Lambda_3$ has mass-dimension one and is the strong coupling scale of the SU$(2)$ gauge theory with $N_f=3$.
As shown in \cite{Gaiotto:2012sf}, in the conformal field theory this involves a collision of the regular conformal block at $z=q$ and $z=0$ and leads to an irregular conformal block at $z=0$. We now consider the five point conformal block with the insertion of the degenerate field $\Phi_{2,1}(z)$:
\begin{equation}
\langle V_{\alpha_4}(\infty)\, V_{\alpha_3}(1) {\cal I}^{(4)}(0) \, \Phi_{2,1}(z) \rangle \, .
\end{equation}
The Ward identity for this chiral conformal block can be obtained by considering the five point block relevant for the conformal $N_f=4$ and scaling the wave-function in the following way \cite{Awata:2010bz,Gaiotto:2012sf}:
\begin{equation}
\Psi(z,q) = q^{-2\alpha_1\alpha_2}\, \psi_3(z, \Lambda_3) \, .
\end{equation}
On the rescaled wave-function $\psi_3(z)$, the $q$-derivative is traded for a $\Lambda_3$-derivative. Combining these, and taking the decoupling limit, we obtain the
null vector decoupling equation for the $N_f=3$ theory:
\begin{align}
\Bigg[&-\epsilon_1^2 \frac{\partial^2}{\partial z^2} +\frac{(m_3+m_4)^2}{4(z-1)^2}+ \frac{m_3 m_4}{z(1-z)} +\frac{m_1\Lambda_3}{z^3}+\frac{\Lambda_3^2}{4z^4}-\frac{\epsilon_1^2}{4(z-1)^2} +\epsilon_2^2\frac{(3-4z)}{4z(z-1)^2}\nonumber\\
&+\frac{1}{z^2-z^3}\Big(-\epsilon_1\epsilon_2\Lambda_3\frac{\partial}{\partial\Lambda_3}+ m_1^2+m_1(\epsilon_1+\epsilon_2)\Big)+\epsilon_1\epsilon_2\Big(\frac{1-2z}{z-z^2}\frac{\partial}{\partial z} + \frac{1-2z}{2z(z-1)^2}\Big)\Bigg]\psi_3(z,\Lambda_3) = 0 \, .
\end{align}
The quartic pole at $z=0$ in the OPE between the stress tensor and the irregular block explains our notation: we denote such an irregular block by ${\cal I}^{(4)}(0)$.
\subsubsection{$N_f=2$: asymmetric realization}
One can take a further limit in which we decouple the mass $m_1$:
\begin{equation}
m_1 \rightarrow \infty\qquad \Lambda_3\rightarrow 0\quad \text{with}\quad \Lambda_2^2 = m_1\Lambda_3 \quad \text{finite}\,.
\end{equation}
Simultaneously, we rescale the wave-function by
\begin{equation}
\psi_3(z,\Lambda_3) = \Lambda_3^{\frac{1}{\epsilon_1\epsilon_2}(m_1^2 + m_1(\epsilon_1+\epsilon_2))} \psi_2(z, \Lambda_2) \,.
\end{equation}
From the conformal field theory perspective, this amounts to tuning the coefficient of the quartic pole to zero, leaving behind only a cubic pole. This corresponds to the four point conformal block
\begin{equation}
\langle V_{\alpha_4}(\infty)\, V_{\alpha_3}(1) {\cal I}^{(3)}(0) \, \Phi_{2,1}(z) \rangle \, .
\end{equation}
The null vector decoupling equation takes the form:
\begin{multline}
\left[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} +\frac{(m_3+m_4)^2}{4(z-1)^2}+ \frac{m_3 m_4}{z(1-z)} +\frac{\Lambda_2^2}{z^3} -\frac{\epsilon_1\epsilon_2}{2(z^2-z^3)}\Lambda_2\frac{\partial}{\partial\Lambda_2}
\right.\cr
\left.+\epsilon_1\epsilon_2\left(\frac{1-2z}{z-z^2}\frac{\partial}{\partial z} + \frac{1-2z}{2z(z-1)^2} \right)-\frac{\epsilon_1^2}{4(z-1)^2} +\epsilon_2^2\frac{(3-4z)}{4z(z-1)^2} \right]\psi_2(z,\Lambda_2) = 0 \, .
\end{multline}
The equation exhibits a cubic pole at the irregular singularity $z=0$.
\subsection{Two Irregular Punctures}
Let us begin with the five point conformal block in which we have
regular conformal primaries at $z_i$, with $i \in \{1,2,3,4\}$ and the
degenerate field at $z$. In equation \eqref{tpsi}, we have obtained
the null vector decoupling equation for this case. We now set $z_1=0$,
$z_2=q$ and $z_4=1$ and consider the simultaneous collision of
punctures such that $q\rightarrow 0$ and $z_3 \rightarrow 1$. We
rescale the wave-function as in \cite{Gaiotto:2012sf}
\begin{equation}
\Psi(z_i) = q^{-2\alpha_1\alpha_2} (z_3-1)^{-2\alpha_3\alpha_4} \psi(z,\Lambda,\widetilde{\Lambda}) \, .
\end{equation}
In order to get finite results, we take the limit
\begin{align}
m_2 &\rightarrow \infty\quad,\quad q\rightarrow 0\quad \text{with}\quad \Lambda = q\, m_2 \quad \text{finite}\, , \cr
m_3 &\rightarrow \infty\quad,\quad (z_3-1)\rightarrow 0\quad \text{with}\quad \widetilde{\Lambda} = (z_3-1)\, m_3 \quad \text{finite}
\, .
\end{align}
Note that this introduces two independent scales in the problem. The conformal block we are considering is a three point function, with two irregular punctures of quartic order and a degenerate field $\Phi_{2,1}(z)$:
\begin{equation}
\langle {\cal I}^{(4)}(1) {\cal I}^{(4)}(0) \, \Phi_{2,1}(z) \rangle \, .
\end{equation}
The parameters $\Lambda$ and $\widetilde{\Lambda}$ respectively represent the quartic pole coefficient at $z=0$ and $z=1$. After the wave-function rescaling and the decoupling limits, we obtain the null vector decoupling equation for such a conformal block with two such irregular singularities and the degenerate field:
\begin{align}
\Bigg[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} +\frac{\Lambda^2}{4z^4}&+\frac{\Lambda m_1}{z^3(z-1)^2}+\frac{\widetilde{\Lambda} m_4}{z(z-1)^3} +\frac{\widetilde{\Lambda}^2}{4(z-1)^4}+\epsilon_1\epsilon_2 \frac{3z-1}{z(z-1)}\frac{\partial}{\partial z}+\frac{(2\epsilon_1\epsilon_2+3\epsilon_2^2)(2-3z)}{4z(z-1)^2}\nonumber\\
&+\frac{1}{z^2(z-1)^2}\Big(-\epsilon_1\epsilon_2\Lambda\frac{\partial}{\partial\Lambda} -2\Lambda m_1+ m_1^2+m_1(\epsilon_1+\epsilon_2)\Big)\Bigg]\psi_2(z, \Lambda, \widetilde{\Lambda}) = 0\,.
\end{align}
We can tune the two scales suitably in order to obtain the null vector decoupling equations for the remaining gauge theories
of our focus.
\subsubsection{$N_f=2$: symmetric realization}
We set $\Lambda = -\widetilde{\Lambda} = \Lambda_2$, in which case we obtain the null vector decoupling equation:
\begin{align}
\Bigg[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} +\frac{\Lambda_2^2}{4z^4} &+ \frac{\Lambda_2m_1}{z^3(z-1)^2}-\frac{\Lambda_2 m_4}{z(z-1)^3} +\frac{\Lambda_2^2}{4(z-1)^4}+\epsilon_1\epsilon_2 \frac{3z-1}{z(z-1)}\frac{\partial}{\partial z}+\frac{(2\epsilon_1\epsilon_2+3\epsilon_2^2)(2-3z)}{4z(z-1)^2} \nonumber\\
&+\frac{1}{z^2(z-1)^2}\Big(-\epsilon_1\epsilon_2\Lambda_2\frac{\partial}{\partial\Lambda_2} -2\Lambda_2m_1+ m_1^2+m_1(\epsilon_1+\epsilon_2)\Big)\Bigg]\psi_2(z, \Lambda_2) = 0\,.
\end{align}
\subsubsection{$N_f=1$}
We set $\Lambda = \Lambda_1$ and $\widetilde{\Lambda}\rightarrow 0$ such that $\widetilde{\Lambda}m_4 = -\frac{\Lambda_1^2}{4}$ is a finite combination. In other words, the relevant conformal block of interest is
\begin{equation}
\langle {\cal I}^{(3)}(1) {\cal I}^{(4)}(0) \, \Phi_{2,1}(z) \rangle\, .
\end{equation}
The coefficient of the cubic pole at $z=1$ is tuned and related to that of the quartic pole at $z=0$.
Such a conformal block satisfies the null vector decoupling equation:
\begin{align}
\Bigg[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} &+\frac{\Lambda_1^2}{4z^4} + \frac{\Lambda_1m_1}{z^3(z-1)^2}-\frac{\Lambda_1^2}{4z(z-1)^3}+\epsilon_1\epsilon_2 \frac{3z-1}{z(z-1)}\frac{\partial}{\partial z}+\frac{2-3z}{4z(z-1)^2}(2\epsilon_1\epsilon_2+3\epsilon_2^2)\nonumber\\
&+\frac{1}{z^2(z-1)^2}\Big(-\epsilon_1\epsilon_2\Lambda_1\frac{\partial}{\partial\Lambda_1} -2\Lambda_1m_1+ m_1^2+m_1(\epsilon_1+\epsilon_2)\Big)
\Bigg]\psi_1(z, \Lambda_1) = 0\,.
\end{align}
\subsubsection{$N_f=0$}
As for the earlier case with $N_f=2$ in the asymmetric realization, in order to obtain finite results, one has to rescale the wave-function
\begin{equation}
\psi(z, \Lambda, \widetilde{\Lambda}) = \Lambda^{\frac{1}{\epsilon_1\epsilon_2}(m_1^2 + m_1(\epsilon_1+\epsilon_2))} \psi_0(z, \Lambda_0) \,.
\end{equation}
We set $\Lambda\rightarrow 0$ and $\widetilde{\Lambda}\rightarrow 0$ such that the combinations $m_1\Lambda = \Lambda_0^2$ and $m_4\widetilde{\Lambda} = \Lambda_0^2$ are equal and finite. The relevant conformal block is given by
\begin{equation}
\langle {\cal I}^{(3)}(1) {\cal I}^{(3)}(0) \, \Phi_{2,1}(z) \rangle \,.
\end{equation}
This satisfies the null vector decoupling equation:
\begin{align}
\Bigg[-\epsilon_1^2 \frac{\partial^2}{\partial z^2} + \frac{\Lambda_0^2}{z^3(z-1)^2}+ \frac{\Lambda_0^2}{z(z-1)^3}&+\epsilon_1\epsilon_2 \frac{3z-1}{z(z-1)}\frac{\partial}{\partial z}+\frac{2-3z}{4z(z-1)^2}(2\epsilon_1\epsilon_2+3\epsilon_2^2)\nonumber\\
&+\frac{1}{z^2(z-1)^2}\Big(-\frac{1}{2}\epsilon_1\epsilon_2\Lambda_0\frac{\partial}{\partial\Lambda_0 }- 2\Lambda_0^2\Big)
\Bigg]\psi_0(z, \Lambda_0) = 0\,.
\end{align}
\section{Stokes Graphs}
\label{sec:StokesGraphs}
In this section, we plot actual machine-generated Stokes graphs for the $N_f=4$ theory as representative examples. The Stokes lines are defined by the condition
\begin{equation}
\text{Im} \left[ \int_{x_0}^{x} \text{d}x' \ \sqrt{Q_0(x')} \right] = 0 \, .
\end{equation}
The red dots indicate singularities, and the blue dots indicate
turning points. It is important to remember that what is relevant for
the monodromy calculations is the topology of the Stokes
graph. The cartoons in the body of the paper abstract away from graphs
given in this appendix.
\subsection{The Double Flip}
\label{BDoubleFlip}
The potential we use to plot these Stokes graphs has a convenient $\mathbb{Z}_4$ symmetric form \cite{GMN}
\begin{equation}
\label{eq:GMNZ4}
Q_0(z) = \frac{z^4 - u \left( z^4 - 1 \right)}{\left( z^4 - 1 \right)^4} \, ,
\end{equation}
where the singularities are at $z_1 = 1$, $z_2 = \text{i}$, $z_3 = -1$, and $z_4 = -\text{i}$. This potential is arrived at via an SL($2,\mathbb{C}$) transformation of \eqref{nullNf4Qfunctions} that maps all singularities to finite points for ease of plotting.
The masses of the fundamental hypermultiplets are $m_a =
\frac{1}{4} z_a$.
We have made the choice $u = \frac{1}{2}$ for plotting the double flip
Stokes graphs in figure \ref{fig:flips}.
\begin{figure}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{PreDFlip.pdf}
\caption{Before the double flip}
\label{fig:PreFlip}
\end{subfigure}
~
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\linewidth]{PostDFlip.pdf}
\caption{After the double flip.}\label{fig:PostFlip}
\end{subfigure}
~
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=1\linewidth]{CritFlip.pdf}
\caption{The critical graph.}\label{fig:CritFlip}
\end{subfigure}
\caption{Sequence of Stokes graphs related by a double flip about the critical graph}\label{fig:flips}
\end{figure}
\iffalse
\begin{figure}[!htb]
\minipage{0.5\textwidth}
\includegraphics[width=\linewidth]{PreDFlip.pdf}
\caption{Before the double flip.}\label{fig:PreFlip}
\endminipage\hfill
\minipage{0.5\textwidth}
\includegraphics[width=\linewidth]{PostDFlip.pdf}
\caption{After the double flip.}\label{fig:PostFlip}
\endminipage
\end{figure}
\begin{figure}[!htp]
\begin{center}
\includegraphics[width=0.5\linewidth]{CritFlip.pdf}
\caption{The critical graph.}\label{fig:CritFlip}
\end{center}
\end{figure}
\fi
\subsection{The Pop}
\label{BPop}
We continue to work with the potential \eqref{eq:GMNZ4} and, in order visualize the pops
in figure \ref{fig:pops}, we have made
the choice $u = \frac{1}{2} \text{exp} \left( \frac{3}{10} \text{i} \pi \right)$.
\begin{figure}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\textwidth]{PrePopInspiral.pdf}
\caption{Before the pop.}
\label{fig:PrePop}
\end{subfigure}
~
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=\linewidth]{PostPopInspiral2.pdf}
\caption{After the pop.}\label{fig:PostPop}
\end{subfigure}
~
\begin{subfigure}[b]{0.45\textwidth}
\includegraphics[width=1\linewidth]{CritPop.pdf}
\caption{The critical graph.}\label{fig:CritPop}
\end{subfigure}
\caption{Sequence of Stokes graphs related by a pop about the critical graph}\label{fig:pops}
\end{figure}
\iffalse
\begin{figure}[!htb]
\minipage{0.5\textwidth}
\includegraphics[width=\linewidth]{PrePopInspiral.pdf}
\caption{Before the pop.}\label{fig:PrePop}
\endminipage\hfill
\minipage{0.5\textwidth}
\includegraphics[width=\linewidth]{PostPopInspiral2.pdf}
\caption{After the pop.}\label{fig:PostPop}
\endminipage
\end{figure}
\begin{figure}[!htp]
\begin{center}
\includegraphics[width=0.5\linewidth]{CritPop.pdf}
\caption{The critical graph.}\label{fig:CritPop}
\end{center}
\end{figure}
\fi
\section{The Saddle Point Analysis}
We discussed regular (irregular) conformal blocks associated
to two-dimensional conformal field theories that, via the 2d/4d correspondence, are dual
to four-dimensional $\Omega$-deformed conformal (respectively,
asymptotically free) gauge theories. The null vector decoupling
equations together with the conformal Ward identities allow us to
arrive at Schr\"odinger equations that govern an integrable system
related to the $\Omega$-deformed gauge theory. While these analyses
are exact in $\epsilon_2$, it is enlightening to see how the
$\epsilon_2 \rightarrow 0$ limit of these differential equations are
derived from purely gauge-theoretic considerations. This will serve as
a consistency check of the 2d/4d correspondence, and our calculations.
In this section, we explain how to derive differential equations
starting from saddle-point difference equations, valid for any SU$(N)$
theory with $N_f=2N$ fundamental hypermultiplets. We specialize
to the case of the conformal SU$(2)$ theory ($N_f = 4$) and
consider various decoupling limits that give
asymptotically free theories with fewer fundamental hypermultiplets.
Our analysis begins with the Nekrasov partition function, and
considers its saddle-points in the limit $\epsilon_2 \rightarrow 0$,
with $\epsilon_1$ held constant and finite
\cite{Poghossian:2010,Fucito:2011pn,Fucito:2012xc,NPS13}. The
result of this analysis is the $\epsilon_1$-deformed Seiberg-Witten
equation:
\begin{equation}
y\left( x \right) + q \frac{M\left(x-\epsilon_1\right)}{y \left( x-\epsilon_1 \right)} = \left( 1+q \right) P\left(x\right) \, .
\end{equation}
Here, $q$ is the instanton counting parameter. The gauge polynomial $P(x)$ is of degree $N$ and encodes the Coulomb moduli of the gauge theory. The flavour polynomial $M(x)$ is of degree $2N$, and we choose to factorize it into two pieces:
\begin{equation}
\label{eq:Decomp1}
M(x) = A(x) D(x) \, ,
\end{equation}
where $A(x)$ and $D(x)$ are degree $N$ polynomials. As should be evident, this decomposition is far from being unique. Different decompositions can be mapped to the different ways in which the flavour symmetry is realized in the type II construction using two NS$5$ branes and a stack of D$4$ branes. One can associate the number of semi-infinite D$4$ branes on each side of the NS$5$ branes with the degree of the polynomials $A(x)$ and $D(x)$.
We now peel off a factor of $A(x)$ from the function $y(x)$ and express the remainder as the ratio of some rational function $\mathbf{Q}(x)$ as
\begin{equation}
\label{eq:Decomp2}
y(x) = A(x ) \frac{\mathbf{Q}(x )}{\mathbf{Q}(x -\epsilon_1)} .
\end{equation}
There are other ways to perform the split but this one will reduce, in the $\epsilon_1\rightarrow 0$ limit, to the correct M-theory curve. The result of these decompositions gives us the Baxter $T\mathbf{Q}$-relation
\begin{equation}
A(x) \mathbf{Q}(x) + q D(x-\epsilon_1) \mathbf{Q}(x-2\epsilon_1) - \left(1+q \right) P(x) \mathbf{Q}(x-\epsilon_1) = 0 .
\end{equation}
We now trade $\mathbf{Q}(x)$ for its Fourier transform \cite{NPS13}
\begin{equation}
\label{eq:FourierTransform}
\Psi(t) = \sum_{x \in \Gamma} \mathbf{Q}(x) \ e^{-x/\epsilon_1} ,
\end{equation}
and arrive at the differential equation of order $N$:
\begin{equation}
\left[ A(x) + q D(x-\epsilon_1) \ t^{-2} - (1+q) P(x) \ t^{-1} \right] \Psi(t) = 0 ,
\end{equation}
where in the above equation, the Fourier transform effectively sends $x \mapsto -\epsilon_1 t \partial_t$.
\subsection{$N_f=4$}
We specialize to the case of the conformal SU$(2)$ gauge theory, with $N_f=4$. The matter polynomials we start with are
\begin{align}
D(x) &= \left(x-m_1+\frac{\epsilon_1}{2}\right)\left(x-m_2+\frac{\epsilon_1}{2}\right) , \\
A(x) &= \left(x-m_3+\frac{\epsilon_1}{2}\right)\left(x-m_4+\frac{\epsilon_1}{2}\right) \,.
\end{align}
The gauge polynomial is
\begin{equation}
P(x) = x^2 - q \left( \frac{m_1 + m_2 + m_3 + m_4}{1+q} \right) x - \frac{u_S}{1+q} .
\end{equation}
Substituting this into the difference equation and taking the Fourier transform as described earlier leads to a second order differential equation.
We would like this differential equation to be of the Schr\"odinger form, i.e.\ with no linear derivative terms.
This may be achieved by peeling off an appropriate factor.\footnote{The terms we peel off are proportional to the products of square roots of eigenfunctions of the monodromy at $0, 1$ and $q$.}
The resulting differential equation is
\begin{equation}
\left[ -\epsilon_1^2 \frac{\partial^2}{\partial t^2} + Q(t, \epsilon_1) \right] \Psi(t) = 0 ,
\end{equation}
where the potential term $Q(t, \epsilon_1)$ has an expansion in powers
of $\epsilon_1$. In addition, we further shift of the Coulomb modulus $u_S$,
\begin{equation}
u_S = -\widetilde{u} + \frac{q-1}{2} \left( m_1^2 + m_2^2 \right) + q \left[ m_1 m_2 + m_3 m_4 + \frac{1}{2} \left(m_1 +m_2\right)\left( m_3 + m_4 \right) \right] + \frac{\epsilon_1^2}{4} \left( 1+q \right)
\end{equation}
Denoting the order $\epsilon_1^m$ coefficient in the potential by $Q_m(t)$, we obtain the non-zero potential terms:
\begin{align}
Q_0(t) &= -\frac{\widetilde{u}}{t(t-1)(t-q)} + \frac{\left( m_1 - m_2 \right)^2}{4t^2} + \frac{\left( m_1 + m_2 \right)^2}{4(t-q)^2} + \frac{\left( m_3 + m_4 \right)^2}{4(t-1)^2} + \frac{m_1^2 + m_2^2 + 2 m_3 m_4}{2t(1-t)} \, , \\
Q_2(t) &= -\frac{1}{4t^2}-\frac{1}{4(t-1)^2}-\frac{1}{4(t-q)^2} + \frac{1}{2(t-1)(t-q)} \, .
\end{align}
This matches the potentials in the text obtained from the null vector decoupling equations (\ref{nullNf4Qfunctions}).
\subsection{$N_f=3$}
The $N_f=3$ case is obtained by decoupling one of the masses; we
choose here to send $m_2\rightarrow\infty$, while simultaneously
sending $q\rightarrow 0$ such that the combination $\Lambda_3 = q\, m_2$ remains finite:
We identify $\Lambda_3$ to be the strong coupling scale in the SU$(2)$ gauge theory with $N_f=3$. The difference equation takes the form
\begin{equation}
A(x) \mathbf{Q}(x) -\Lambda_3 D(x-\epsilon_1) \mathbf{Q}(x-2\epsilon_1) - P(x) \mathbf{Q}(x-\epsilon_1) = 0 .
\end{equation}
The polynomial functions are given by
\begin{align}
D(x) &= \left(x-m_1+\frac{\epsilon_1}{2}\right), \\
A(x) &= \left(x-m_3+\frac{\epsilon_1}{2}\right)\left(x-m_4+\frac{\epsilon_1}{2}\right) \,,
\end{align}
while the gauge polynomial is
\begin{equation}
P(x) = x^2 - \Lambda_3 x - u_S\,.
\end{equation}
The differential equation can be obtained in a similar fashion as in the conformal case by taking the Fourier transform as in \eqref{eq:FourierTransform}. The analysis leading to the Schr\"odinger type equation can be repeated as before and we obtain the differential equation:
\begin{equation}\label{diffeqngeneral}
\left[ -\epsilon_1^2 \frac{\partial^2}{\partial t^2} + \sum_{m=0}^2 Q_m(t)\epsilon_1^m \right] \Psi(t) = 0 \, .
\end{equation}
The $u_S$ we use in the saddle-point analysis must be shifted in order
to make contact with the form of the potential in the text \eqref{Nf=3
semiclassical} and the shift is given by
\begin{equation}
u_S = \widetilde{u} -\frac{\Lambda_3}{2}\left( 2 m_1 + m_3 + m_4 \right) + \frac{\epsilon_1^2}{4} \, .
\end{equation}
The shift of the Coulomb modulus is accompanied by a global rescaling $m_1 \rightarrow \frac{m_1}{2}$ and $\Lambda_3 \rightarrow \frac{\Lambda_3}{2}$. After these shifts and rescalings, the non-zero potential functions $Q_m(t)$ are given by
\begin{align}
Q_0(t) &= \frac{(m_3+m_4)^2}{4(t-1)^2}+ \frac{m_3 m_4}{t(1-t)} +\frac{m_1\Lambda_3}{t^3}+\frac{\Lambda_3^2}{4t^4} +\frac{\widetilde{u}}{t^2(1-t)} , \\
Q_2(t) &= -\frac{1}{4(t-1)^2}\,.
\end{align}
which match the potential in equation (\ref{Nf=3 semiclassical}).
\subsection{$N_f=2$: asymmetric realization}
There are two distinct cases to be considered when $N_f=2$. These correspond to the fashion in which we decouple fundamental matter. We first consider the case when
\begin{equation}
m_1\rightarrow \infty \quad \text{and} \quad \Lambda_3\rightarrow0 \quad \text{with}\quad \Lambda_2^2 = m_1\Lambda_3\quad\text{finite}\,.
\end{equation}
The difference equation takes the form
\begin{equation}
A(x) \mathbf{Q}(x) +\Lambda_2^2 \mathbf{Q}(x-2\epsilon_1) - P(x) \mathbf{Q}(x-\epsilon_1) = 0 .
\end{equation}
The polynomials are given by
\begin{align}
A(x) &= \left(x-m_3+\frac{\epsilon_1}{2}\right)\left(x-m_4+\frac{\epsilon_1}{2}\right) \,,
\end{align}
while the gauge polynomial is
\begin{equation}
\label{eq:GaugePolyII}
P(x) = x^2 - u_S\,.
\end{equation}
The differential equation is once again given as in \eqref{diffeqngeneral}, and after the shift
\begin{equation}
u_S = \widetilde{u} - \Lambda^2 + \frac{\epsilon_1^2}{4} \, ,
\end{equation}
We find that the non-zero potential functions $Q_m(t)$ are
\begin{align}
Q_0(t) &= \frac{(m_3+m_4)^2}{4(t-1)^2}+ \frac{m_3 m_4}{t(1-t)} +\frac{\Lambda_2^2}{t^3} +\frac{\widetilde{u}}{t^2(1-t)} , \\
Q_2(t) &= -\frac{1}{4 (t-1)^2}\,.
\end{align}
which matches the potential in \eqref{Nf=2 asymmetric semiclassical}.
\subsection{$N_f=2$: symmetric realization}
An inequivalent way to realize the $N_f=2$ theory is to start from the $N_f=3$ differential equation and consider the limit
\begin{equation}
m_3\rightarrow \infty \quad \text{and} \quad \Lambda_3\rightarrow0 \quad \text{with}\quad \Lambda_2^2 = m_3\Lambda_3\quad\text{finite}\,.
\end{equation}
The difference equation takes the form
\begin{equation}
\Lambda_2A(x) \mathbf{Q}(x) +\Lambda_2 D(x-\epsilon_1)\mathbf{Q}(x-2\epsilon_1) - P(x) \mathbf{Q}(x-\epsilon_1) = 0 .
\end{equation}
The polynomial $A(x)$ is given by
\begin{align}
D(x) &= \left(x-m_1+\frac{\epsilon_1}{2}\right), \\
A(x) &= \left(x-m_4+\frac{\epsilon_1}{2}\right) \,,
\end{align}
while the gauge polynomial is given by \eqref{eq:GaugePolyII}.
The differential equation still takes the Schr\"odinger form \eqref{diffeqngeneral}.
Finally, in order to match with the form of the differential in the text \eqref{Nf=2 symmetric semiclassical}, we need to perform a conformal transformation $\left( t \rightarrow \frac{z-1}{z} \right)$. We choose to shift away the $O(\epsilon_1^2)$ term by redefining of the Coulomb modulus:
\begin{equation}
u_S = \widetilde{u} + m_1 \Lambda_2 - \frac{\Lambda^2}{2} + \frac{\epsilon_1^2}{4} \,.
\end{equation}
After this, the differential matches the form in \eqref{Nf=2 symmetric semiclassical}, with
\begin{equation}
Q_0(z) = \frac{\Lambda_2^2}{4z^4} + \frac{\Lambda_2m_1}{z^3(z-1)^2}-\frac{\Lambda_2 m_4}{z(z-1)^3} +\frac{\Lambda_2^2}{4(z-1)^4} +\frac{\widetilde{u}}{z^2(z-1)^2}
\, .
\end{equation}
\subsection{$N_f=1$}
We start with the symmetric realization of the $N_f=2$ case and take the limit
\begin{equation}\label{Nf=1limit}
m_4\rightarrow \infty \quad \text{and} \quad \Lambda_2\rightarrow 0 \quad\text{with}\quad \Lambda_1^3=m_4\Lambda_2^2\quad\text{finite}\,.
\end{equation}
The difference equation in this case takes the form
\begin{equation}
\Lambda_1^2 \mathbf{Q}(x) + \Lambda_1\, D(x-\epsilon_1)\mathbf{Q}(x-2\epsilon_1) - P(x) \mathbf{Q}(x-\epsilon_1) = 0 .
\end{equation}
The gauge polynomial is once again given by \eqref{eq:GaugePolyII}, and the flavour polynomial $D(x)$ is given by
\begin{equation}
D(x) = \left( x - m_1 + \frac{\epsilon_1}{2} \right) \, .
\end{equation}
While the differential equation in this case also takes the form of a Schr\"odinger equation, with the potential (after shifting $u_S$ so as to set $Q_2(t)$ to zero for convenience)
\begin{equation}
Q_0(t) = \frac{\Lambda_1^2}{4 t^4} - \frac{m_1 \Lambda_1}{t^3} + \frac{u_S}{t^2} + \frac{\Lambda_1^2}{t} \, ,
\end{equation}
its form requires a series of conformal transformations and rescalings before it can be easily compared
with the form in \eqref{Nf=1 semiclassical}; for convenience, we reproduce the transformations below:
\begin{align}
\begin{split}
t &\longrightarrow \frac{1}{z} \\
u_S &\longrightarrow 2^{4/3} \left( \widetilde{u}+m_1 \Lambda_1 \right) \, \\
m_1 &\longrightarrow - 2^{2/3} m_1 \, \\
z &\longrightarrow 2^{2/3} w \, \\
w &\longrightarrow \frac{1-z}{z} \, .
\end{split}
\end{align}
After this sequence of transformations, we get the form of the potential as in \eqref{Nf=1 semiclassical}:
\begin{equation}
Q_0(z) = \frac{\Lambda_1^2}{4z^4} + \frac{\Lambda_1m_1}{z^3(z-1)^2}-\frac{\Lambda_1^2}{4z(z-1)^3}
+\frac{\widetilde{u}}{z^2(z-1)^2} \, .
\end{equation}
\subsection{$N_f=0$}
The pure super Yang-Mills case is obtained by starting with the
$N_f=1$ case and taking the limit,
\begin{equation}
m_1\rightarrow \infty \quad \text{and} \quad \Lambda_1\rightarrow 0\quad\text{with}\quad \Lambda_0^4=m_1\Lambda_1^3\quad\text{finite}\,.
\end{equation}
The difference equation takes the form:
\begin{equation}
\Lambda_0^2 \mathbf{Q}(x) +\Lambda_0^2\mathbf{Q}(x-2\epsilon_1) - P(x) \mathbf{Q}(x-\epsilon_1) = 0 .
\end{equation}
While the differential equation in this case also takes the form of a Schr\"odinger equation, with the potential (after shifting $u_S$ so as to set $Q_2(t)$ to zero for convenience)
\begin{equation}
\phi_2 (t) = \frac{\Lambda_0^2}{t^3} + \frac{u_S}{t^2} + \frac{\Lambda_0^2}{t} \, ,
\end{equation}
its form requires a series of conformal transformations and rescalings before it can be literally compared with the form in \eqref{PureNVD}; for convenience, we reproduce these transformations below:
\begin{align}
\begin{split}
t &\longrightarrow \text{i} w \, \\
u_S &\longrightarrow - \left( \widetilde{u} + \Lambda_0^2 \right) \, \\
\Lambda_0 &\longrightarrow \text{e}^{\text{i}\frac{\pi}{4}} \Lambda_0 \, \\
w &\longrightarrow \frac{1-z}{z} \, .
\end{split}
\end{align}
After this sequence of transformations, we get the potential as in the bulk of the paper \eqref{PureNVD}:
\begin{equation}
Q_0(z) = \frac{\Lambda_0^2}{z^3(z-1)^2} +\frac{\widetilde{u}}{z^2(z-1)^2} + \frac{\Lambda_0^2}{z(z-1)^3} \, .
\end{equation}
Thus, we completed the derivation of the null vector decoupling equations (at $\epsilon_2=0$)
from the purely gauge theoretic instanton partition function.
\end{appendix}
|
\section{Introduction}
\label{Intro}
In this work we aim to fill a gap in the Bayesian literature by proposing two objective priors for the parameter of the Yule--Simon distribution. The distribution was firstly discussed in \cite{Yule25} and then re-proposed in \cite{Simon55}, and can be used in scenarios where the center of interest is some sort of frequency in the data. For example, \cite{Yule25} used it to model abundance of biological genera, while \cite{Simon55} exploited the distribution properties to model the addition of new words to a text. It goes without saying that other areas of applications can be considered where, for instance, frequencies represent the elementary unit of observation. For example, in this paper we show the employment of the Yule--Simon distribution in modelling daily increments of social network stock options, surnames and 'superstar' success in the music industry.
Despite the wide range of applications, the literature on the Yule--Simon distribution appears to be limited. And, more surprisingly, to the best of our knowledge it seems that no attention has been given to the problem by the Bayesian community. Given the challenges that classical inference faces in estimating the parameter of the distribution \citep{Garcia11}, the possibility of tackling the problem from a Bayesian perspective is, undoubtedly, appealing.
In addressing the estimation of the shape parameter of the Yule--Simon distribution by means of the Bayesian framework, we opted for an objective approach. We propose two priors: the first is the Jeffreys rule prior \citep{JE61}, while the second is obtained by applying the loss-based approach discussed in \cite{VillaWalker15}. Although we formally introduce the Yule--Simon distribution and its derivation in the next Section, it is important to give an anticipation of the general idea here, so to fully appreciate the gain in adopting an objective approach. As nicely illustrated in \cite{ChungCox94}, the shape parameter of the distribution is linked via a one-to-one transformation to the probability that the next observation will not take a value previously observed. For example, if we have observed $n$ words in a text, we wish to make inference on the probability that the $(n+1)$ observation is a word not yet encountered in the text, assuming this probability to be constant. It is then clear that the Yule--Simon distribution models extremely large events. As such, the information in the data about these events is limited and a ``wrongly'' elicited prior could end up dominating the data. On the other hand, a prior with minimal information content would allow the data ``to speak'', resulting in a more robust inferential procedure. We do not advocate that in every circumstance an objective approach is the only suitable. In fact, if reliable prior information is available, an elicited prior would represent, in general, the natural choice. Alas, in the presence of phenomena with extremely rare events, the above information is often insufficient or incomplete, and an objective choice would then represent the most sensible one.\\
The paper is organized as follows. In Section \ref{Prelim} we set the scene by introducing the Yule--Simon distribution and the notation that will be used throughout the paper. The proposed objective priors are derived and discussed in Section \ref{ObjectivePriors}. Section \ref{Simu} collects the analysis of the frequentist performances of the posterior distributions yielded by the proposed priors. Through a set of several simulation scenarios, we compare and analyse the inferential capacity of the objective priors here discussed. In Section \ref{RealData} we illustrate the application of the priors to three real-data applications. Finally, Section \ref{Concl} is reserved to concluding remarks and points of discussion.
\section{Preliminaries}
\label{Prelim}
The most known functional form of the Yule--Simon distribution, possibly, is the following:
\begin{equation}
f(k;\rho) = \rho\, \mbox{B}(k,\rho+1), \qquad k=1,2,\ldots \mbox{ and } \rho>0, \label{Yule_Org}
\end{equation}
where $\mbox{B}(\cdot,\cdot)$ is the beta function and $\rho$ is the shape parameter. The distribution in \eqref{Yule_Org} was firstly proposed by \cite{Yule25} in the field of biology; in particular, to represent the distribution of species among genera in some higher taxon of biotic organisms.
More recently, \cite{Simon55} noticed that the above distribution can be observed in other phenomena, which appear to have no connection among each others. These include, the distribution of word frequencies in texts, the distribution of authors by number of scientific articles published, the distribution of cities by population and the distribution of incomes by size. The derivation process followed by \cite{Yule25} was based on word frequencies, and it consisted of two assumptions:
\begin{itemize}
\item[(i)] The probability that the $(n+1)$-th word is a word observed $k$ times in the first $n$ words, is proportional to $k$; and
\item[(ii)] The probability that the $(n+1)$-th word is new (i.e. not being observed in the first $n$ words) is constant and equal to $\alpha\in(0,1)$.
\end{itemize}
\cite{Yule25} shows that, under the condition of stationarity, the process defined by the above two assumptions yields \eqref{Yule_Org} by setting $\rho = 1/(1-\alpha)$, obtaining:
\begin{equation}
f(k;\alpha) = \frac{1}{1-\alpha}\mbox{B}\left(k,\frac{1}{1-\alpha}+1\right). \label{Disc}
\end{equation}
An important consequence of the above assumption (ii) is that the shape parameter $\rho$ of the distribution takes values in $(1,+\infty)$. In other words, should we use the model as in \cite{Yule25}, which includes the possibility that $0<\rho\leq1$, we would loose the interpretation of the generating process described by the two assumptions above. In fact, for $\rho<1$, the probability of observing a new word would be negative; while for $\rho=1$ the probability would be zero, rendering the process trivial (i.e. all the observed words will be equal to the first one observed). Furthermore, the expectation of the Yule--Simon distribution is defined only for values of the shape parameter larger than one, and this property is something one would expect in most applications. For all the above reasons, in this work we focus on the parametrization of the Yule-Simon given in \eqref{Disc}, that is we will discuss prior distributions for $\alpha$.
In addition to the parametrization of the Yule--Simon distribution as in \eqref{Disc}, we will also consider the possibility of having the parameter $\alpha$ discrete. This is a common finding in literature, especially when implementations of the model are considered. See, for example, \cite{Simon55} and \cite{Garcia11}. The discretization of $\alpha$ will be discussed in detail in Section \ref{DiscrOb}.
\section{Objective Priors for the Yule-Simon distribution}
\label{ObjectivePriors}
This section is devoted to the derivation of two objective priors for the Yule-Simon distribution: the Jeffreys prior and loss-based prior. The former assumes that parameter space of $\alpha$ is continuous and it is based on the well-known invariance property proposed by \cite{JE61}; the latter assumes the parameter space discrete and is based on \cite{VillaWalker15}.
\subsubsection*{The Jeffreys Prior}
\label{DiscrJe}
The Jeffreys prior is defined in the following way \citep{JE61}:
\begin{equation}
\pi(\alpha) \propto \sqrt{\mathcal{I}(\alpha)} \notag
\end{equation}
where $\mathcal{I}(\alpha)=\mathbb{E}_{\alpha}\biggl[-\frac{\partial ^{2}\log(f(k;\alpha))}{\partial \alpha^{2}} \biggr]$ is the Fisher Information. In the next Theorem \label{Jef1} (which proof is in the Appendix) an explicit expression of the Jeffreys prior for the Yule-Simon distribution is provided.
\begin{theorem}\label{Jef1}
Let $f(k;\alpha)$ be the Yule-Simon distribution defined in equation (\ref{Disc}), with $0<\alpha<1$. The Jeffreys prior for $\alpha$ is
\begin{equation}\label{JefPrior}
\pi(\alpha) \propto q(\alpha) \\
\end{equation}
where
$$q(\alpha)=\frac{1}{1-\alpha}\sqrt{1-\frac{1}{(2-\alpha)^{2}} \,_{3}F_{2}\biggl(1,\frac{1}{1-\alpha}+1,1; \frac{1}{1-\alpha}+2,\frac{1}{1-\alpha}+2;1\biggr)}.$$
with $_{3}F_{2}$ being the hypergeometric distribution function.
\end{theorem}
\noindent The Jeffreys prior stated in Theorem \ref{Jef1} is a proper prior. In fact, let
\begin{equation}\label{JefPriorNorm}
\pi(\alpha) =\frac{q(\alpha)}{K}, \notag \\
\end{equation}
where
\begin{footnotesize}
\begin{equation}\label{NormConst}
K=\int_0^1\frac{1}{1-\alpha}\sqrt{1-\frac{1}{(2-\alpha)^{2}} \,_{3}F_{2}\biggl(1,\frac{1}{1-\alpha}+1,1; \frac{1}{1-\alpha}+2,\frac{1}{1-\alpha}+2;1\biggr)}d\alpha\notag
\end{equation}
\end{footnotesize}
is the normalizing constant of $\pi(\alpha)$.
\noindent It is not difficult to prove that $$K<\infty.$$
Indeed,
$$K\leq\int_0^1\sqrt{\frac{3-\alpha}{1-\alpha}}\frac{1}{2-\alpha}d\alpha=\frac{1}{3}\pi-\ln(2-\sqrt{3})<\infty. $$
The result above follows from the following inequality
$$_{3}F_{2}\biggl(1,\frac{1}{1-\alpha}+1,1; \frac{1}{1-\alpha}+2,\frac{1}{1-\alpha}+2;1\biggr)\geq 1.$$
The properness of the prior in (\ref{JefPrior}) ensures the properness of the yielded posterior distribution for $\alpha$, as such suitable for inference.
\subsubsection*{The Loss-based Prior}
\label{DiscrOb}
\cite{VillaWalker15} introduced a method for specifying and objective prior for discrete parameters. The idea is to assign a \emph{worth} to each parameter value by objectively measuring what is lost if the value is removed, and it is the true one. The loss is evaluated by applying the well known result in \cite{Berk66} stating that, if a model is misspecified, the posterior distribution asymptotically accumulates on the model which is the nearest to the true one, in terms of the Kullback--Leibler divergence.
Given that the parameter $\alpha\in(0,1)$ of the Yule--Simon is in principle continuous, the above method can not be applied. However, the boundedness of the interval allows for an easy discretization, directly we can consider the set
$$\mathbb{D}_{M}=\left\lbrace \alpha=\frac{i}{M}: i=1,\dots,M-1\right\rbrace.$$
Therefore, the \emph{worth} of the parameter value $\alpha$ is represented by the Kullback--Leibler divergence $D_{KL}(f(k|\alpha)\|f(k|\alpha^\prime))$, where $\alpha^\prime\neq\alpha$ is the parameter value that minimizes the divergence. To link the \emph{worth} of a parameter value to the prior mass, \cite{VillaWalker15} use the self-information loss function. This particular type of loss function measures the loss in information contained in a probability statement \citep{MerFed98}. As we now have, for each value of $\alpha$, the loss in information measured in two different ways, we simply equate them obtaining the loss-based prior of \cite{VillaWalker15}:
\begin{equation}
\pi(\alpha) \propto \exp{\biggl\{\min_{\alpha' \ne \alpha} D_{KL}(f(k|\alpha)\|f(k|\alpha'))\biggr\}} -1 \qquad\alpha,\alpha^\prime\in\mathbb{D}_{M},\label{OBB}
\end{equation}
where
\begin{align*}
D_{KL}(f(k|\alpha)\|f(k|\alpha'))
&=\log{\left(\frac{1-\alpha'}{1-\alpha}\right)}+\mathbb{E}_{\alpha}\left\{\log{\left[B\left(k;\frac{1}{1-\alpha}+1\right)\right]}\right\}\\
&\hspace{3cm}-\mathbb{E}_{\alpha}\left\{\log{\left[B\left(k;\frac{1}{1-\alpha'}+1\right)\right]}\right\}.
\end{align*}
As the discretized parameter space is finite, no matter what value of $M$ one chooses, the prior \eqref{OBB} is proper, hence, the yielded posterior will be proper as well.
An important aspect is that the value $\alpha^\prime$ minimizing the Kullback--Leibler divergence can not be analytically determined, and the prior has to be computationally derived. However, even for large values of $M$, the computational cost is trifling compared to the whole Monte Carlo procedure necessary to simulate from the posterior distribution. \\
\begin{figure}[h!]
\centering
{\includegraphics[width=7.5cm]{AllJeffNEW}}
\caption{Prior distribution for $\alpha$ obtained by applying Jeffreys rule (dashed line), the loss-based method with $M=10$ (continuous line), with $M=20$ (dotted line) and with $M=100$ (dash-dotted line).}
\label{JJvsKL}
\end{figure}
To have a feeling of the prior distributions derived above, we have plotted them in Figure \ref{JJvsKL}. The behaviour of the priors is similar, in the sense that they tend to increase as $\alpha$ increases and, for increasing values of $M$, the two distributions seem to converge. However, we note that the Jeffreys prior is flatter than the loss-based priors for large values of the parameter, i.e. for $\alpha$ approximately greater than $0.8$.
\section{Simulation Study}
\label{Simu}
The objective priors defined in Section \ref{ObjectivePriors} are automatically derived by taking into consideration properties intrinsic to the Yule--Simon distribution. In other words, they do not depend on experts knowledge or previous observations. It is therefore necessary, in order to validate them, to assess the goodness of the priors by making inference on simulated data. This section is dedicated in performing a simulation study on the parameter $\alpha$ using observations obtained from fully known distributions.
We have considered different sample sizes, $n=30$, $n=100$ and $n=500$, to analyse the behaviour of the prior distributions under different level of information coming from the data. Here we show the results for $n=100$ only, as the sole differences in using $n=30$ and $n=500$ sample sizes are limited to the precision of the inferential results: relatively low for $n=30$ and relatively high for $n=500$, as one would expect. Besides that, the differences in the performance of the two priors noted for $n=100$ remain for the other sample sizes. As the loss-based prior depends on the discretization of the parameter space, for illustration purposes, we have considered $M=10$ and $M=20$, that is $\alpha\in\{0.1,0.2,\ldots,0.9\}$ and $\alpha\in\{0.05,0.10,\ldots,0.95\}$, respectively.
Both the Jeffreys prior and the loss-based prior yield posterior distributions for $\alpha$ which are not analytically tractable, hence, it is necessary to use Monte Carlo methods. We have generated $100$ samples from a Yule--Simon distribution with the parameter $\alpha$ set to every value in the parameter space, 9 for $M=10$ and 19 for $M=20$. For each sample we have simulated from the posterior distribution of $\alpha$, under both priors, by running $10,000$ iterations, with a burn-in period of $2,000$ iterations.
To evaluate the priors we have considered two frequentist measures. The first is the frequentist coverage of the $95\%$ credible interval. That is, for each posterior, we compute the interval between the 0.025 and 0.975 quantiles and see if the true value of $\alpha$ is included in it. Over repeated samples, one would expect a proportion of about $95\%$ of the posterior intervals to contain the true parameter value. The second frequentist measure gives an idea of the precision of the inferential process, and it is represented by the square root of the mean squared error (MSE) from the mean, relative to the parameter value: $\sqrt{\mbox{MSE}(\alpha)}/\alpha$. We have considered the MSE from the median as well but, due to the approximate symmetry of the posterior, the results are very similar to the MSE from the mean.
\begin{figure}[h!]
\centering
\subfigure[
{\includegraphics[width=6.75cm]{PlotCoveragen100A01ConLinea.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{PlotsqrtMeanSe01.eps}}
\caption{Frequentist properties of the Jeffreys prior (dashed line) and the loss-based prior (continuous line) for $n=100$. The loss-prior is considered on the discretized parameter space with $M=10$. The left plot shows the posterior frequentist coverage of the $95\%$ credible interval, and the right plot represents the square root of the MSE from the mean of the posterior, relative to $\alpha$.}
\label{alpha01}
\end{figure}
Figure \ref{alpha01} details the results for the simulations with $n=100$ and a parameter space for $\alpha$ discretized with increments of $0.1$, that is $\alpha\in\{0.1,0.2,\ldots,0.9\}$. If we compare the coverage, we note that the loss-based prior tends to over-cover the credible interval, while the Jeffreys prior, although shows a better coverage for values of $\alpha<0.5$, deteriorates in performance as the parameter tends to the upper bound of its space. Looking at the MSE, both priors appear to have very similar performance, and the (relative) error tends to decrease and $\alpha$ increases.
\begin{figure}[h!]
\centering
\subfigure[
{\includegraphics[width=6.75cm]{PlotCoveragen100A005ConLinea.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{PlotsqrtMeanSe005.eps}}
\caption{Frequentist properties of the Jeffreys prior (dashed line) and the loss-based prior (continuous line) for $n=100$. The loss-prior is considered on the discretized parameter space with $M=20$. The left plot shows the posterior frequentist coverage of the $95\%$ credible interval, and the right plot represents the square root of the MSE from the mean of the posterior, relative to $\alpha$.}
\label{alpha005}
\end{figure}
In Figure \ref{alpha005} we have compared the frequentist performance of the Jeffreys prior with the loss-based prior defined over a more densely discretized parameter space, i.e. $\alpha=\{0.05,0.10,\ldots,0.95\}$. We note a smoother behaviour of the priors compared to Figure \ref{alpha01}, which is obviously due to the denser characterization considered. The coverage still reveals a tendency of the loss-based prior to over-cover, although less pronounced than the previous case. Jeffreys prior does not present any significant difference from the previous case, as one would expect. For what it concerns the MSE, the differences between the two priors are negligible, and the only aspect we note, as mentioned above, is a smoother decrease of the error as the parameter increases.
We look more into the details of the objective approach by analysing two i.i.d. samples. In particular, we consider a random sample of size $n=100$ from a Yule--Simon distribution with $\alpha=0.40$ and a sample, of the same size, from a Yule--Simon with $\alpha=0.68$.
In both cases, we have sampled from the posterior distribution via Monte Carlo methods with $10,000$ iterations and a burn-in period of $2,000$ iterations.
\begin{figure}[h!]
\centering
\subfigure[
{\includegraphics[width=6.75cm]{PosteriorSamplen100A040Int005.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{PosteriorHistogramJJn100A040.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{PosteriorSamplen100A040Int01.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{PosteriorHistogramn100A040Int001.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{PosteriorSamplen100A040Int005NEW.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{PosteriorHistogramn100A040Int005NEW.eps}}
\caption{Posterior samples (left) and histograms (right) of the analysis of an i.i.d. sample of size $n=100$ from a Yule--Simon distribution with $\alpha=0.40$. From top to bottom, we have Jeffreys prior, loss-based prior with $M=10$ and loss-based prior with $M=20$.}
\label{PosA040}
\end{figure}
Figure \ref{PosA040} shows the posterior samples and posterior histograms derived by applying the Jeffreys prior and the loss-based prior with two different discretizations, that is $M=10$ and $M=20$. The summary statistics of the three posteriors are reported in Table \ref{T40}, where we have the mean, the median, and the $95\%$ credible interval.
\begin{table}[h!]
\centering
\begin{tabular}{cccc}
\hline
Prior & Mean & Median & $95\%$ C.I. \\
\hline
Jeffreys & 0.40& 0.41 &(0.23,0.53) \\
Loss-based $(M=10)$ & 0.40 & 0.4 & (0.2,0.5) \\
Loss-based $(M=20)$ & 0.40 & 0.41 & (0.22,0.56) \\
\hline
\end{tabular}
\caption{Summary statistics of the posterior distributions for the parameter $\alpha$ of the simulated data from a Yule-Simon distribution with $\alpha=0.40$.}
\label{T40}
\end{table}
By comparing the mean of the posterior distributions, we see that they are all centered around the true parameter value. The credible interval yielded by the loss-based priors with the most dense discretization ($M=20$) is larger than the other two intervals. However, the difference is very small and we can conclude that the three prior distributions result in posteriors which carry the same uncertainty. In other words, the three objective priors perform in the same way.
\begin{figure}[h!]
\centering
\subfigure[
{\includegraphics[width=6.75cm]{PosteriorSampleJJn100A068.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{PosteriorHistogramJJn100A068.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{PosteriorSamplen100A068Int01.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{PosteriorHistogramn100A068Int01.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{PosteriorSamplen100A068Int005NEW.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{PosteriorHistogramn100A068Int005NEW.eps}}
\caption{Posterior samples (left) and histograms (right) of the analysis of an i.i.d. sample of size $n=100$ from a Yule--Simon distribution with $\alpha=0.68$. From top to bottom, we have Jeffreys prior, loss-based prior with $M=10$ and loss-based prior with $M=20$.}
\label{PosA068}
\end{figure}
\begin{table}
\centering
\begin{tabular}{cccc}
\hline
Prior & Mean & Median & $95\%$ C. I. \\
\hline
Jeffreys & 0.68 & 0.68 & (0.57,0.77) \\
Loss-based ($M=10)$ & 0.68 & 0.7 & (0.6,0.8) \\
Loss-based ($M=20)$ & 0.68 & 0.68 & (0.55,0.79) \\
\hline
\end{tabular}
\caption{Summary statistics of the posterior distributions for the parameter $\alpha$ of the simulated data from a Yule-Simon distribution with $\alpha=0.68$.}
\label{T68}
\end{table}
Similar considerations can be made for the case where we have sampled $n=100$ observations from a Yule--Simon distribution with $\alpha=0.68$. By inspecting Figure \ref{PosA068} and Table \ref{T68}, we note a very similar behaviour of the three priors, in the sense that the posterior distributions are still centered around the true value of $\alpha$ and that the credible intervals do not present important differences. Note that the choice of a true parameter value which would have not been included in any of the two discretized sample spaces, upon which the loss-prior is based, allows to show that the inferential process appears to be not affected by the discretization, hence motivating it.
To conclude, the simulation study shows no tangible differences in the performance of the prior distributions, in the spirit of objective Bayesian analysis.
\section{Real Data Application}
\label{RealData}
To illustrate the proposed priors, both the Jeffreys and the loss-based prior for the Yule-Simon distribution, we analyze three datasets. The first dataset concerns daily increments of four popular social networks stock indexes in the US market, the second contains the frequencies of surnames observed in the $1990$ US Census, and the last dataset consists of 'number one' hits in the US music industry.
\subsection{Social network stock indexes}
\label{Stock}
We analyze different data in the social media marketing, in particular we focus on Facebook, Twitter, Linkedin and Google. These four major companies are the most powerful social networks in the world and are listed in the Wall Street exchange market (\href{http://finance.yahoo.com}{http://finance.yahoo.com}). We analyze the daily increments for the stocks and, in particular, we consider the adjusted closing price from the $1^{st}$ of October 2014 to the $11^{th}$ of March 2016, for a total of $n=365$ observations. The daily increments are obtained by applying $z_{t}=\left|r_{t}/r_{t-1}-1\right|\cdot 100$, for $t=2,\dots, 365$, where $r_{t}$ is the adjusted closing price for the index at day $t$, and we built our frequency on it. These are shown in Figure \ref{DayliIncre}, while Figure \ref{DicretHists} shows the histogram of the frequencies of the discretized data. The discretization has been done by counting the number of times a daily return took a value truncated at the second decimal digit. For example, if two observed daily returns are $1.2494$ and $1.2573$, they were both considered as two occurrences of the same value.
\begin{figure}[h!]
\centering
\subfigure[Facebook]
{\includegraphics[width=6.75cm]{DailyIncrFacebook.eps}}
\subfigure[Google]
{\includegraphics[width=6.75cm]{DailyIncrGoogle.eps}}
\subfigure[Linkedin]
{\includegraphics[width=6.75cm]{DailyIncrLinkedin.eps}}
\subfigure[Twitter]
{\includegraphics[width=6.75cm]{DailyIncrTwitter.eps}}
\caption{Daily increments for Facebook, Google, Linkedin and Twitter from the $1^{st}$ of October 2014 to the $11^{th}$ of March 2016.}
\label{DayliIncre}
\end{figure}
By inspecting the histograms in Figure \ref{DicretHists} is seems that the (transformed) Yule--Simon distribution might be a suitable statistical model to represent the data. We apply the Bayesian framework and obtain the posterior distribution for the parameter of interest as
$$\pi(\alpha|\textbf{k}) \propto L(\textbf{k}|\alpha)\pi(\alpha),$$
where $\textbf{k}=(k_1,\dots,k_n)$ represents the set of observations, i.e. the frequencies of the discretized daily returns, $L(\textbf{k}|\alpha)$ the likelihood function and $\pi(\alpha)$ the prior distribution which, in turn, has the form of the Jeffreys prior in \eqref{JefPrior} or the loss-based prior \eqref{OBB}.
\begin{figure}[h!]
\centering
\subfigure[Facebook]
{\includegraphics[width=6.75cm]{HistFacebookDiscret.eps}}
\subfigure[Google]
{\includegraphics[width=6.75cm]{HistGoogleDiscret.eps}}
\subfigure[Linkedin]
{\includegraphics[width=6.75cm]{HistLinkedinDiscret.eps}}
\subfigure[Twitter]
{\includegraphics[width=6.75cm]{HistTwitterDiscret.eps}}
\caption{Histograms of the discretized daily returns for Facebook, Google, Linkedin and Twitter.}
\label{DicretHists}
\end{figure}
We have obtained the posterior distributions for the parameter $\alpha$ of the transformed Yule-Simon distribution by Monte Carlo methods. We run 25,000 iterations with a burn-in period of $5,000$ iterations. We have reported the chain and the histogram of the posterior distributions in Figure \ref{Facebook} and in Figure \ref{Google}, with the corresponding summary statistics in Table \ref{T2}. Note that, with the purpose of limiting the amount of space used, we have included the plots of the Facebook and Google daily returns only.
\begin{figure}[h!]
\centering
\subfigure[]
{\includegraphics[width=6.75cm]{JJFacebook.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{HistJJFacebook.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{KLFacebook.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{HistKLFacebook.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{KL05FacebookNEW.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{HistKL05FacebookNEW.eps}}
\caption{Posterior samples (left) and posterior histograms (right) for the Facebook daily returns obtained by applying the Jeffreys prior (top), the loss-based prior with $M=10$ (middle) and the loss-based prior with $M=20$ (bottom).}
\label{Facebook}
\end{figure}
\begin{figure}[h!]
\centering
\subfigure[]
{\includegraphics[width=6.75cm]{JJGoogle.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{HistJJGoogle.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{KLGoogle.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{HistKLGoogle.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{KL05GoogleNEW.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{HistKL05GoogleNEW.eps}}
\caption{Posterior samples (left) and posterior histograms (right) for the Google daily returns obtained by applying the Jeffreys prior (top), the loss-based prior with $M=10$ (middle) and the loss-based prior with $M=20$ (bottom).}
\label{Google}
\end{figure}
\begin{table}[h!]
\centering
\begin{tabular}{ccccc}
\hline
Company & Prior & Mean & Median & $95\%$ C.I. \\
\hline
Facebook & Jeffreys & 0.53 & 0.53 & (0.43, 0.61) \\
Facebook & Loss-based $(M=10)$ & 0.53 & 0.5 & (0.4, 0.6) \\
Facebook & Loss-based $(M=20)$ & 0.52 & 0.55 & (0.40, 0.60) \\
\hline
Google & Jeffreys & 0.47 & 0.47 & (0.37, 0.55) \\
Google & Loss-based $(M=10)$ & 0.47 & 0.5 & (0.4, 0.6) \\
Google & Loss-based $(M=20)$ & 0.47 & 0.46 & (0.35, 0.55) \\
\hline
Linkedin & Jeffreys & 0.56 & 0.57 & (0.47, 0.64) \\
Linkedin & Loss-based $(M=10)$ & 0.57 & 0.6 & (0.5, 0.6) \\
Linkedin & Loss-based $(M=20)$ & 0.56 & 0.55 & (0.45, 0.65) \\
\hline
Twitter & Jeffreys & 0.68 & 0.68 & (0.62, 0.73) \\
Twitter & Loss-based $(M=10)$ & 0.69 & 0.7 & (0.6, 0.7) \\
Twitter & Loss-based $(M=20)$ & 0.68 & 0.70 & (0.60, 0.75) \\
\hline
\end{tabular}
\caption{Summary statistics of the posterior distribution for the parameter $\alpha$ of the social network stock index data.}
\label{T2}
\end{table}
For all the four assets we notice that the results for $\alpha$ are very similar, as can be inferred by the minimal (or absence of) difference between the means and the medians. The credible intervals, as well, are very similar, with a slight larger size for the case where the loss-based prior with ($M=20$) is applied. One way of interpreting the results is as follows. The parameter $\alpha$ can be seen as the probability that the next observation is different from the ones observed so far, and therefore we note that Twitter has the highest chance to take a daily increment not yet observed, while Google has the smallest.
\subsection{Census Data - Surname analysis}
\label{Census}
\begin{table}
\begin{center}
\begin{tabular}{cccccc}
\hline
\#& Surname & Frequency & \#& Surname & Frequency \\
\hline
1& Smith & 2502021 & 6 & Davis & 1193807 \\
2& Johnson & 2014550 & 7 & Miller & 1054530 \\
3& Williams & 1738482 & 8 & Wilson & 843126 \\
4& Jones & 1544488 & 9 & Moore & 775975 \\
5& Brown & 1544488 & 10 & Taylor & 773488 \\
\hline
\end{tabular}
\caption{Ten most common Surname in United States from the Census 1990 analysis.}
\label{TCensus}
\end{center}
\end{table}
The second example we examine concerns with the frequency of surnames in the US (\href{http://www.census.gov/en.html}{http://www.census.gov/en.html}). From the population censuses \citep{Mar10}, we focus on the US Census completed in 1990 and consider the first 500 most common surnames. Refer to Table \ref{TCensus} for a list of the first 10 most frequent surnames.
Briefly, the process followed by \cite{Mar10} to obtain the data converts the surname with Senior (SR), Junior (JR) or a number in the last name field (f.e. Moore Sr or Moore Jr or Moore III are converted to Moore) and, in addition, the authors examined each name entry for the possibility of an inversion (e.g. a first name appearing in the last name fields or vice-versa). However, as there is the possibility of having many surnames that also inverted can sound absolutely right, the authors considered also the surname of the spouse, obtaining additional information to invert the name field of the entire family.
The analysis has been performed by running both the Markov Chain Monte Carlo for 25,000 iterations, with a burn-in of 5,000 iterations.
The posterior samples and the posterior histograms are shown in Figure \ref{Surname}, with the corresponding summary statistics of the posterior distributions reported in Table \ref{TSurn}. We again notice similarities to the simulation study and the analysis of daily increments, in the sense that means and medians are very similar for each prior, and the $95\%$ credible interval obtained by applying the loss-based prior with $M=20$ is slightly larger than the one obtained by using either the Jeffreys prior or the loss-based prior with $M=10$.
\begin{figure}[h!]
\centering
\subfigure[]
{\includegraphics[width=6.75cm]{JJSurname.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{HistJJSurname.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{KLSurname.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{HistKLSurname.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{KL05SurnameNEW.eps}}
\subfigure[]
{\includegraphics[width=6.75cm]{HistKL05SurnameNEW.eps}}
\caption{Posterior sample (left) and posterior histogram (right) for the surname data set obtained by applying the Jeffreys prior (top), the loss-based prior with $M=10$ (middle) and the loss-based prior with $M=20$ (bottom).}
\label{Surname}
\end{figure}
\begin{table}
\centering
\begin{tabular}{cccc}
\hline
Prior & Mean & Median & $95\%$ C. I. \\
\hline
Jeffreys & 0.53 & 0.54 & (0.47, 0.58) \\
Loss-based $(M=10)$ & 0.52 & 0.5 & (0.5, 0.6) \\
Loss-based $(M=20)$ & 0.53 & 0.55 & (0.45, 0.60) \\
\hline
\end{tabular}
\caption{Summary statistics of the posterior distributions for the parameter $\alpha$ of the Census surname analysis.}
\label{TSurn}
\end{table}
The estimated value of $\alpha$, on the basis of the 500 most common surnames in the US (and if we consider the mean) is, roughly, $1/2$. In other words, there are about $50\%$ chances that the next observed surname is not in the list of the 500. Obviously, a larger sample size would yield a smaller posterior mean, as the number of surnames is finite and the more we observe, the harder is to find a ``new'' one.
\subsection{`Superstardom' analysis}
The last example consists in modelling the number of `number one' hits a music artist had in the period 1955--2003 on the Billboard Hot 100 chart. The data, which is displayed in Table \ref{HITS-Data}, has been used by \cite{ChungCox94} and \cite{SpiVoo09} to show an apparent absence of correlation between talent and success in the music industry.
\begin{table}[h!]
\centering
\begin{tabular}{cc|cc}
\hline
Hits & Observations & Hits & Observations \\
\hline
1 & 119 & 9 & 4 \\
2 & 57 & 10 & 2 \\
3 & 30 & 11 & 1 \\
4 & 13 & 12 & 2 \\
5 & 10 & 13 & 1 \\
6 & 4 & 14 & 1 \\
7 & 1 & 15 & 1 \\
8 & 1 & 16 & 1 \\
\hline
\end{tabular}
\caption{Number of `number one' hits per artist from 1955 to 2003.}
\label{HITS-Data}
\end{table}
We have run the Monte Carlo simulation for 25,000 iterations, with a burn in period of 5,000, for each of the considered priors. The posterior samples and histograms are shown in Figure \ref{Hits}, with the correspondinf statistic summaries in Table \ref{THits}.
\begin{figure}[h!]
\centering
\subfigure[
{\includegraphics[width=6.75cm]{JJHits.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{HistJJHits.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{KLHits.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{HistKLHits.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{KL05HitsNEW.eps}}
\subfigure[
{\includegraphics[width=6.75cm]{HistKL05HitsNEW.eps}}
\caption{Posterior sample (left) and posterior histogram (right) for the music `number one' hits data set obtained by applying the Jeffreys prior (top), the loss-based prior with $M=10$ (middle) and the loss-based prior with $M=20$ (bottom).}
\label{Hits}
\end{figure}
\begin{table}[h!]
\centering
\begin{tabular}{cccc}
\hline
Prior & Mean & Median & $95\%$ C.I. \\
\hline
Jeffreys & 0.08 & 0.07 & (0.004, 0.24) \\
Loss-based $(M=10)$ & 0.13 & 0.1 & (0.1, 0.3) \\
Loss-based $(M=20)$ & 0.11 & 0.10 & (0.05, 0.25) \\
Loss-based $(M=100)$ & 0.10 & 0.08 & (0.01, 0.29) \\
\hline
\end{tabular}
\caption{Summary statistics of the posterior distribution for the parameter $\alpha$ of the analysis of the music `number one' hits.}
\label{THits}
\end{table}
This example of the music hits allows for some interesting points of dicussion. First, we note that the posterior distributions of for $\alpha$ are skewed; therefore, the posterior median represents a better centrality index than the posterior mean. Second, it is clear that the ``true'' value of $\alpha$ may be close to zero. As such, in order to explore better the parameter space when the loss-based prior is used, a denser discretization is more appropriate. We have then considered $M=100$, resulting the posterior summary statistics in Table \ref{THits}. We note now that the posterior median is similar to the one obtained using the Jeffreys prior. It is therefore recomendable that, when the inference on $\alpha$ indicates values near the parameter space boudaries, the level of discretization to be considered should be relatively dense.
\section{Discussions}
\label{Concl}
It is surprising how, from time to time, the Bayesian literature presents gaps even for problems which appear to be straightforward. The Yule--Simon distribution has undoubtedly many possibilities of application, as the discussed examples and the refereed papers show, and therefore demanded for a satisfactory discussion within the Bayesian framework.
Given the importance that objective Bayesian analysis can have in applications, and not only \citep{Berger06}, we have presented two priors which are suitable in scenarios with minimal prior information. The first prior is the Jeffreys prior which, as it is well known, has the appealing property of being invariant under monotone differentiable transformations of the parameter of interest. The second prior is derived considering the loss in information one would incur if the `wrong' model was selected. Although the latter requires a discretization of the parameter space, we have shown through simulation studies that the performance of the yielded posterior are very similar, both between the Jeffreys and the loss-based prior, and between different structures of the discretized parameter space. This is not surprising as both priors, i.e. the Jeffreys and the loss-based, have a similar behaviour, in the sense that they increase as the parameter $\alpha$ increases.
We have limited our analysis to the case where the shape parameter of the Yule--Simon distribution, $\rho$, is strictly larger than one. Doing so, we allow for a more convenient parametrization of the distribution where the new parameter $\alpha=(\rho-1)/\rho$ has the interpretation of being the probability that the next observation takes a value not observed before.
Besides through a simulation study, we have compared the objective priors by applying them on three data sets: the first related to financial data, the second to surnames in the US and the third one on the number of hits in the music industry. All comparisons allowed to show that the two proposed objective priors lead to similar results, in terms of posterior distributions. For obvious reasons, we have not considered if the choice of the Yule--Simon is the \emph{best} model to represent the data, but limited our analysis to make inference for the unknown parameter $\alpha$.
\section*{Acknowledgements}
Fabrizio Leisen was supported by the European Community's Seventh Framework Programme [FP7/2007-2013] under grant agreement no: 630677.
\vspace*{-8pt}
\bibliographystyle{biom}
|
\section{Introduction}
As usual, we denote by $a, b, c$ the sides of a given $\triangle ABC$ and by $S$ its area.
The positive orientation of the plane is determined by $\triangle ABC$.
Let $x, y, z$ be trilinear coordinates of a point with respect to $\triangle ABC$.
It is well known that the Lemoine point $K$ minimizes the sum $x^2+y^2+z^2$, i.e. \cite{L}
$$x^2+y^2+z^2 \geq \frac{4S^2}{a^2+b^2+c^2}\, .$$
Recently Kimberling \cite{K} obtained several inequalities for the power sums $x^q+y^q+z^q$.
In this note for an arbitrary point $X$ in the plane of $\triangle ABC$ we study a weighted sum $F(X)$
of the squares of the distances of $X$ to the sidelines of the triangle. We give a
geometric interpretation of the minimum (maximum) of the function $F(X)$.
\section{Preliminaries}
Let us recall some properties of isogonal conjugate points with respect to a given $\triangle ABC$.
Given the basic $\triangle ABC$ and its circumcircle $k(ABC)$. Denote by
$\imath$ the isogonal conjugation with respect to the triangle. The action of $\imath$ in the domains
(with respect to the vertex $A$) (Fig. 1) is as follows:
\begin{center}
\includegraphics[width=10cm]{f1.eps}
\vskip 2mm
Fig. 1
\end{center}
1) $\imath(\sigma)=\sigma$.
\noindent
If $M$ is a point on the side $BC$, then $\imath (M)=A$.
2) $\imath(\sigma_{12})=\sigma_{12}$.
\noindent
For any point $M$ on the ray $BD^{\rightarrow}$ we have $\imath (M)=C$; for any point $M \in CF^{\rightarrow}$ $\imath (M)=B$.
3) $\imath(\sigma_{13})=\sigma'_{13}$, \; $\imath(\sigma'_{13})=\sigma_{13}$.
\noindent
For any point $M \in AE^{\rightarrow}$ $\imath(M)=C$; \; for any point $M \in AG^{\rightarrow}$ $\imath(M)=B$.
The transformation $\imath$ can also be defined for points on the circumcircle $k$, different
from the vertices of the triangle. If $M$ is a point on the arc $\widearc {BC}$, then $\imath(M)$ is the point at infinity of the
line, which is symmetric to the line $AM$ with respect to the bisector of $\angle BAC$.
For an arbitrary point $X$ in the plane of
$\triangle ABC$ we denote by $x,\,y,\,z,$ the directed distances of the point $X$
to the lines $BC,\,CA,\,AB$, respectively. Then $(x,\, y,\, z)$ is the triple of trilinear coordinates of $X$
with respect to the basic triangle. The trilinear coordinates satisfy the equality $ax +by +cz = 2S.$
Further we denote by $S_1, \, S_2,\, S_3$ the oriented areas
of the triangles $BCX, CAX, ABX$, respectively. Then
$\ds{\lambda =\frac{S_1}{S}, \, \mu = \frac{S_2}{S}, \, \nu = \frac{S_3}{S}}; \; \lambda + \mu + \nu =1$
are the barycentric coordinates of $X$ with respect to the $\triangle ABC$. The relation between the trilinear coordinates
and the barycentric coordinates of the point $X$ is given by
$$\lambda = \frac{ax}{2S}\,, \quad \mu = \frac{by}{2S}\,, \quad \nu = \frac{cz}{2S}\,.$$
We also consider the function ${\ds J= \frac{a^2}{\lambda}\,+ \frac{b^2}{\mu}\,+
\frac{c^2}{\nu}}=\ds{\left(\frac{a}{x}+\frac{b}{y}+\frac{c}{z}\right)2S,}$ \,
which is defined for all points that do not lie on the sidelines of $\triangle ABC$.
The following characterization of the points in the $\angle BAC$ out of $\triangle ABC$ is useful.
\vskip 2mm
{\bf Lemma 1.} Let $M(x, y, z)$ be with trilinear coordinates satisfying the conditions
$x < 0, \\ y > 0, \, z > 0$, i.e. $M$ is in the $\angle BAC$ out of $\triangle ABC$. Then
1) $M \in \sigma_{12}$ iff $J > 0$;
2) $M$ lies on the arc $\widearc{BC}$ iff $J = 0$;
3) $M \in \sigma_{13}$ iff $J < 0$.
\vskip 2mm
We also need the following statement.
{\bf Lemma 2.} Let $M(x, y, z)$ be with trilinear coordinates satisfying the conditions
$x > 0, \\ y < 0, \, z < 0$, i.e. $M \in \sigma'_{13}$. Then $J<0$.
\vskip 2mm
\textbf{The isogonal conjugation with respect to barycentric coordinates }
Let $(\lambda, \mu, \nu), \, \lambda + \mu +\nu =1$ be the barycentric coordinates of a point $M$,
which does not lie on the lines $AB, BC, CA$ or on the circumcircle $k(ABC)$. If $(\lambda', \mu', \nu')$
are the barycentric coordinates of the point $N$, isogonal conjugate to $M$, then
$$\lambda'=\frac{a^2}{\lambda\, J}\,, \quad \mu'=\frac{b^2}{\mu\, J}\,, \quad \nu'=\frac{c^2}{\nu \,J}\,.\leqno(2.1)$$
If $(\lambda, \mu, \nu)$ are the barycentric coordinates of a point $M$, it is useful to consider the
\emph{homogeneous} barycentric coordinates of $M$:
$$(\rho \lambda,\, \rho \mu,\, \rho \nu), \quad \rho \neq 0.$$
Then the formulas
$$\lambda'=\frac{a^2}{\lambda}\,, \quad \mu'=\frac{b^2}{\mu}\,, \quad \nu'=\frac{c^2}{\nu}\leqno(2.2)$$
represent the isogonal conjugation even on the arcs $\widearc{BC}$, $\widearc{CA}$ or $\widearc{AB}$ of $k$.
If $M(\lambda, \mu, \nu)$ lies on\
the arc $\widearc{BC}$, then the point $\imath(M)= N(\lambda', \mu', \nu')$
satisfies the condition $\lambda'+\mu'+\nu'=0$ and lies on the line at infinity.
\vskip 2mm
\textbf{A general formulation of the problem}
\vskip 2mm
Now, let $(\lambda, \mu, \nu) \neq (0, 0, 0)$ be a triple of fixed real numbers. For any point
$X$ with trilinear coordinates $(x, y, z)$ consider the function
$$F(X)= \lambda x^2+\mu y^2 + \nu z^2, $$
which is a weighted sum of the squares of the directed distances $(x, y, z)$.
Our aim is to investigate the minima and maxima of the above function.
Further we consider three essential cases:
1. $\lambda \mu \nu \neq 0, \quad \lambda + \mu + \nu > 0 $;
2. $\lambda \mu \nu \neq 0, \quad \lambda + \mu + \nu < 0$;
3. $\lambda \mu \nu \neq 0, \quad \lambda + \mu + \nu = 0$.
\section{Weighted sum with $\lambda \mu \nu \neq 0, \quad \lambda + \mu + \nu > 0$}
Obviously both functions
$$\lambda x^2 + \mu y^2 + \nu z^2, \qquad \frac{\lambda x^2 + \mu y^2 + \nu z^2}{\lambda + \mu + \nu}$$
have minima and maxima at the same points. Without loss of generality
we can assume that $\lambda + \mu + \nu = 1$.
Thus the problem in this section is to find the minimum (maximum) of the function
$$F(X)=\lambda x^2 + \mu y^2 + \nu z^2, \quad \lambda + \mu + \nu = 1. \leqno(3.1)$$
First we consider the case
\vskip 2mm
\textbf{1.1.} $\lambda > 0, \; \mu > 0, \; \nu > 0$.
\vskip 2mm
\textbf{Problem 1.} {\it
$(i)$ Find the point $N$ that minimizes the function $F(X)$.
$(ii)$ If $M$ is the point with barycentric coordinates $(\lambda, \mu, \nu)$, prove that $M$ and $N$ are
isogonal conjugate.}
{\it Solution.} To solve (i), consider the system
$$\begin{array}{l}
{\ds F(X)=\lambda\,x^2+\mu\,y^2+\nu\,z^2,}\\
[2mm]
ax+by+cz=2S;
\end{array}
\qquad
x, y , z \in \R \leqno(3.2)
$$
and interpret $(x,y,z)$ as Cartesian coordinates in the three dimensional Euclidean space.
The level surfaces of the function $F(X)$ are the ellipsoids
$$\varepsilon(k) : \quad
\lambda\,x^2+\mu\,y^2+\nu\,z^2=k, \quad k={\rm const} \in (0, \infty).$$
Geometrically, to find the point that minimizes the function $F(X)$ in (2.2), means to find $k$ so that the ellipsoid $\varepsilon(k)$
is tangent to the plane $\pi: ax+by+cz=2S$ and then to determine the touch-point $N$ of $\varepsilon(k)$ to $\pi$.
The tangent plane to $\varepsilon(k)$ at a point $(x_0, y_0, z_0)$ is given by the
equality
$$\tau: \lambda\,x_0 \, x+\mu\,y_0 \, y+\nu\,z_0 \, z=k.$$
Then the condition $\tau \equiv \pi$ implies that
$$\begin{array}{l}
\ds{\frac{\lambda\,x_0}{a}=\frac{\mu\,y_0}{b}=\frac{\nu z_0}{c}=t,}\\
[4mm]
t(a x_0+b y_0+c z_0)=k,\\
[4mm]
ax_0+by_0+cz_0=2S.
\end{array} \leqno(3.3)
$$
Solving (3.3), we find:
$$t= \frac{2S}{J}\,, \quad k=\frac{4S^2}{J}\,,\; \left(J=\frac{a^2}{\lambda}+
\frac{b^2}{\mu}+\frac{c^2}{\nu}\right)$$
and
$$x_0=\frac{2S}{J}\,\frac{a}{\lambda}\,, \quad
y_0=\frac{2S}{J}\,\frac{b}{\mu}\,, \quad z_0=\frac{2S}{J}\,\frac{c}{\nu}\,.\leqno(3.4)$$
Now, taking into account (3.4), we conclude that the point $N$ minimizing the function $F(X)$
has barycentric coordinates $(\lambda', \mu', \nu')$ with respect to $\triangle ABC$
given by
$$\lambda'=\frac{ax_0}{2S} = \frac{a^2}{\lambda\,J}\,, \quad \mu'=\frac{b\,y_0}{2S} = \frac{b^2}{\mu\,J}\,,
\quad \nu'=\frac{cz_0}{2S} = \frac{c^2}{\nu \, J}\, \leqno(3.5)$$
and $\ds{F_{min} = k = \frac {4S^2}{J}}$, which solves (i).
To prove (ii), let us denote by $M$ the point with barycentric coordinates $(\lambda, \mu, \nu)$.
Comparing with (1.1) we conclude that formulas (3.5) are a representation of the isogonal conjugation in
barycentric coordinates. Hence, the point $N(\lambda', \mu', \nu')$ is the isogonal conjugate
one to the point $M(\lambda, \mu, \nu)$.
\qed
In this case $M$ and $N$ are in $\sigma$.
Next we consider the case
\vskip 2mm
\textbf{1.2.} $\lambda < 0, \; \mu > 0, \; \nu > 0$.
\vskip 2mm
In this case the problem states as follows:
\textbf{Problem 2.} {Prove that $F(X)$ has a minimum if and only if
$$\ds{J=\frac{a^2}{\lambda}+\frac{b^2}{\mu}+\frac{c^2}{\nu} < 0}.$$
$(i)$ Find the point $N$ that minimizes the function $F(X)$.
$(ii)$ If $M$ is the point with homogeneous barycentric coordinates $(\lambda, \mu, \nu)$, prove that $M$ and $N$ are
isogonal conjugate.}
{\it Solution.} Consider the system (3.2). In this case any level surface of the function $F(X)$
$$\varepsilon(k) : \quad
\lambda\,x^2+\mu\,y^2+\nu\,z^2=k, \quad k={\rm const} \in \R$$
is one of the following: one sheet hyperboloid if $k > 0$; cone if $k = 0$; two sheet hyperboloid if $k < 0$.
Geometrically, to find the point that minimizes the function $F(X)$ in (3.2), means to find $k<0$ so that the two sheet
hyperboloid $\varepsilon(k)$ is tangent to the plane $\pi: ax+by+cz=2S$ and then to determine the touch-point
$N$ of $\varepsilon(k)$ to $\pi$.
The plane $\pi$ can be tangent to $\varepsilon(k)$ only if $ k< 0 $.
Similarly to the solution of Problem 1 we obtain the system (3.4), which implies that
$$t J= 2S, \quad 2S t = k. $$
Therefore $\varepsilon(k)$ is tangent to the plane $\pi$ only in the case $J < 0$.
Further, we find the coordinates of the touch-point
$$x_0=\frac{2S}{J}\,\frac{a}{\lambda}\,, \quad
y_0=\frac{2S}{J}\,\frac{b}{\mu}\,, \quad z_0=\frac{2S}{J}\,\frac{c}{\nu}\,.\leqno(3.6)$$
Thus, the point $N$ minimizing the function $F(X)$ lies in the domain $\sigma'_{13}$.
Let $M$ be the point with barycentric coordinates $(\lambda, \, \mu, \, \nu)$. Then the formulas (3.6) show that
the point $N(\lambda', \, \mu', \, \nu')$ has barycentric coordinates
$$\lambda'=\frac{a^2}{\lambda\,J}\,, \quad \mu'=\frac{b^2}{\mu\,J}\,, \quad \nu'=\frac{c^2}{\nu\,J}$$
and it is isogonal conjugate to the point $M$.
Hence, the triple $(\lambda, \mu, \nu)$ determines a point $M$ in the domain $\sigma_{13}$ and
$$F_{min}= F(N) = \frac{4S^2}{J} < 0,$$
where $N \in \sigma'_{13}$ is the isogonal conjugate to the point $M$.
\qed
\vskip 2mm
\textbf{1.3.} $\lambda >0,\; \mu < 0,\; \nu < 0$.
In this case the problem states as follows:
\textbf{Problem 3.} {Prove that the function $F(X)$ has neither minimum, nor maximum.}
{\it Solution.} Let $M$ be the point with barycentric coordinates $(\lambda, \, \mu, \, \nu)$.
Then $M \in \sigma'_{13}$ and $J < 0$. Similarly to the case 1.2 we obtain that
in our case any level surface of the function $F(X)$
$$\varepsilon(k) : \quad
\lambda\,x^2+\mu\,y^2+\nu\,z^2=k, \quad k={\rm const} \in \R$$
is one of the following: one sheet hyperboloid if $k < 0$; cone if $k = 0$; two sheet hyperboloid if $k > 0$.
Since $J < 0$, then $\ds{k=\frac{4S^2}{J} < 0}$ and $\pi$ can not be a tangent plane to any two sheet hyperboloid.
Hence the function $F(X)$ has neither minimum nor maximum.
\qed
\vskip 2mm
\section{Weighted sum with $\lambda \mu \nu \neq 0, \; \lambda + \mu + \nu < 0$}
If $\lambda + \mu + \nu < 0$, then we put $\bar \lambda = - \lambda,\; \bar \mu = - \mu, \; \bar \nu = - \nu$
and consider the function
$$\bar F(X) = \bar \lambda x^2 + \bar \mu y^2 + \bar \nu z^2=- F(X), \quad \bar\lambda + \bar\mu + \bar\nu > 0.$$
We suppose again that $\bar\lambda + \bar\mu + \bar\nu = 1$ and consider the point $M$ with barycentric coordinates
$(\bar\lambda, \bar\mu, \bar\nu )$.
Comparing with Section 2 we have the following.
\vskip 2mm
\textbf{2.1. $\lambda<0, \; \mu <0, \; \nu < 0$}
Under these conditions $\ds{J=\frac{a^2}{\lambda}+\frac{b^2}{\mu}+\frac{c^2}{\nu} < 0}.$
Then
$$M(\bar \lambda, \bar \mu, \bar \nu) \in \sigma, \; N=\imath(M) \in \sigma, \; \bar J>0, \quad
F_{max}=F(N)= \frac{4S^2}{J} < 0.$$
\textbf{2.2. $\lambda>0, \, \mu <0, \; \nu < 0$}
\textbf{2.2.1. $J<0$}
The function $\bar F =-F$ has no minimum or maximum.
\textbf{2.2.2 $J>0$}
$$M(\bar \lambda, \bar \mu, \bar \nu) \in \sigma_{13}, \; N=\imath(M) \in \sigma'_{13}, \; \bar J<0, \quad
F_{max}=F(N)= \frac{4S^2}{J} > 0.$$
\textbf{2.3. $\lambda<0, \; \mu >0, \; \nu > 0$}
The function $\bar F =-F$ has no minimum or maximum.
\section{Weighted sum with $\lambda \mu \nu \neq 0, \; \lambda + \mu + \nu = 0$}
Let us consider the system of mass points $\{A(\lambda), B(\mu), C(\nu)\}$ and denote by $P(\mu+\nu)$ the center of
mass of the system $\{B(\mu), C(\nu)\}$ (Fig. 2). If $O$ is an arbitrary point, we have
$$\lambda \overrightarrow{OA} + \mu \overrightarrow{OB} + \nu \overrightarrow{OC}=-\lambda \overrightarrow{AP}
=\mu \overrightarrow{AB} + \nu \overrightarrow{AC} = \overrightarrow{v} = \overrightarrow{\rm const}.$$
Calculating
$$\overrightarrow{v}^2=-\lambda \mu \nu \left(\frac{a^2}{\lambda}+\frac{b^2}{\mu}+\frac{c^2}{\nu}\right)=
- \lambda \mu \nu J,$$
we obtain the geometric meaning of $J$ in the case $\lambda + \mu + \nu = 0$.
The condition $\overrightarrow{v}^2> 0$ implies that
$$\lambda \mu \nu J < 0. \leqno(5.1)$$
\begin{center}
\includegraphics[width=10cm]{f2.eps}
\vskip 2mm
Fig. 2
\end{center}
In this section we consider two cases: $(\lambda <0, \mu > 0, \nu > 0)$ and $(\lambda > 0, \mu < 0, \nu < 0)$.
\textbf{3.1. $\lambda \mu \nu < 0.$}
In this case $M(\lambda, \mu, \nu)$ can be interpreted as a point at infinity, i.e. the point at infinity of
the line $AP$ (Fig. 2). The inequality (5.1) implies that
$J>0$. Using similar arguments as in Section 2 we conclude that the function $F(X)$ has neither minimum
nor maximum.
\textbf{3.2. $\lambda \mu \nu > 0.$}
Under these conditions the inequality (5.1) implies that $J<0$.
We consider the function $\bar F=-F$. Since $\bar \lambda \bar \mu \bar \nu <0$, comparing with the previous case
we conclude that $\bar F$ i.e. $F$ has no minimum or maximum.
\section{Weighted sum with one or two zero weights}
\vskip 2mm
\textbf{4.1.} $\lambda = 0, \, \mu > 0, \, \nu > 0.$
\vskip 2mm
In this case the level surfaces of the function $F(X)$ are the elliptic cylinders
$$\mu\,y^2+\nu\,z^2=k, \quad k={\rm const} \in (0, \infty),$$
and the axis $Ox$, when $k=0$. The level surfaces intersect the plane $ax +by + cz = 2S$ into ellipses
by $k > 0$ and in the point $(2S, 0, 0)$ by $k=0$.
The minimum of the function $F(X)$ is $F_{min}=0$ and it occurs when $y=z=0$.
Thus $M(0, \mu, \nu)$ is a point on the sideline $BC$ and the minimum $F_{min}=0$
occurs when $N\equiv A$, i.e. $M$ and $N$ are again isogonal conjugate.
\vskip 2mm
\textbf{4.2. $\lambda=0, \; \mu <0, \; \nu < 0$.}
\vskip 2mm
$M(0, \bar \mu, \bar \nu) \in BC, N=\imath(M)=A,$ $F_{max}=F(N)=0$.
\vskip 2mm
\textbf{4.3. $\lambda=0, \; \mu \nu <0 $.}
\vskip 2mm
In this case the level surfaces of the function $F(X)$
$$\mu y^2 + \nu z^2 = k, \quad k\in (-\infty,+\infty)$$
are hyperbolic cylinders if $k \neq 0$ and two planes if $k=0$.
Hence the function $F(X)$ has no minimum or maximum.
\vskip 2mm
\textbf{5.1.} $\lambda = 1, \, \mu = \nu = 0.$
\vskip 2mm
It is clear that $F_{min}=0$ and it occurs when $x=0$.
Thus $M(1, 0, 0) \equiv A$ and $N $ is any point on $BC$, i.e. $M$ and $N$ are again isogonal conjugate.
\vskip 2mm
\textbf{5.2. $\lambda=-1, \; \mu =0, \; \nu = 0$.}
\vskip 2mm
$M(1,0,0)\equiv A, N \in BC$, $F_{max}=F(N)=0$.
\vskip 2mm
\textbf{Summarizing we get the following:}
\vskip 2mm
1. $\lambda + \mu + \nu > 0$.
If $M(\lambda, \mu, \nu)$ is an inner point for the circumcircle $k$, or coincides
with a vertex of $\triangle ABC$, then $F$ has a minimum and $F_{min}=F(N), \, N=\imath(M)$.
\vskip 2mm
2. $\lambda + \mu + \nu < 0$.
If $M(-\lambda, -\mu, -\nu)$ is an inner point for the circumcircle $k$, or coincides
with a vertex of $\triangle ABC$, then $F$ has a maximum and $F_{max}=F(N), \, N=\imath(M)$.
\vskip 2mm
3. $\lambda + \mu + \nu = 0$.
In this case the function $F$ has neither a minimum nor a maximum.
\section{Examples}
We choose as a typical example the following pair of conjugate points: the circumcenter $O$ and the orthocenter $H$.
1. Let $M\equiv O$. Then $\lambda =\sin 2\alpha, \, \mu = \sin 2\beta, \, \nu = \sin 2 \gamma$ and the
point $O$ generates the function
$$F(X)=\sin 2\alpha \, x^2+\sin 2\beta \, y^2+\sin 2 \gamma \, z^2.$$
1.1. $\triangle ABC$ is acute-angled and $O \in \sigma$. Simple calculations show that
$$J=\frac{4S}{\cos \alpha \, \cos \beta \, \cos \gamma}, \quad F_{min}=F(H)= 4S \cos \alpha \, \cos \beta \, \cos \gamma.$$
1.2. $\alpha = 90^0$ and $O$ is the midpoint of $BC$. Then
$$F_{min}=F(A)=0.$$
1.3. $\alpha > 90^0$ and $O \in \sigma_{13}$. Then
$$J=\frac{4S}{\cos \alpha \, \cos \beta \, \cos \gamma}\,<0, \quad
F_{min}=F(H)= 4S \cos \alpha \, \cos \beta \, \cos \gamma \, <0.$$
\vskip 2mm
{\bf Remark.} \emph{The function F(X) generates geometric inequalities.}
Let $M(\lambda, \, \mu, \, \nu)\in \sigma$ generate the function (3.1). If we choose a concrete triangle
center $X$ in $\sigma$ with trilinear coordinates $(x, y, z)$ and replace in (3.1), then we obtain
the geometric inequality
$$F(X)\geq F_{min}=\frac{4S^2}{J}\,.$$
The equality occurs if and only if $X\equiv \imath(M)$.
In a similar way the case $M \in \sigma_{13}$ generates even more interesting geometric inequalities.
|
\section{Introduction}
Blazars constitute an extreme class of radio-loud active galactic nuclei (AGN), which are characterized by a relativistic jet that
is pointed at small angles to the observer \citep[e.g.][]{begelman84}, and polarized and highly variable non-thermal continuum emission extending through the entire electromagnetic spectrum, from radio to X-rays and, or even, up to high and very high energy $\gamma$-rays
\citep[e.g.][]{Gupta, Wagner2009, Vercellone, Giommi, Abramowski2013, Abramowski2014}. The
rapid variability of blazars is visible at different wavelengths on different timescales down to hours or even shorter \citep[e.g.][]{wagner,2155flare,saito}.
This class of objects includes the flat-spectrum radio-loud quasars (FSRQs) as well as the BL Lacertae type (BL Lac) objects. For
FSRQs the presence of prominent broad and narrow emission lines are characteristic, while the featureless continuum emission in the optical band is attributed to the BL Lac type objects \citep[e.g.][]{urry}.
The spectral energy distribution (SED), in $\nu F_{\nu}$ representation, usually has a double-peaked structure that is generated by two emission components.
The first bump located between the optical and X-ray regimes is described by the synchrotron radiation from the relativistic electrons in the jet. The second bump can be explained either by leptonic or by the hadronic scenarios.
In the leptonic models the high-energy part of the SED, located in the hard X-ray-to-gamma-ray regime, is produced by the inverse-Compton (IC) emission of the same electron
population, involving either the jet synchrotron photons as a seed for the IC scattering
\citep[synchrotron self-Compton model, SSC; e.g. ][]{konigl, marscher, Band85}, or various photon
fields originating outside of the jet \citep[external-Compton models; e.g. ][]{dermer92,
sikora}.
Alternatively, the second bump can be explained in the framework of hadronic scenarios that are mostly initiated by the relativistic protons accelerated with the electrons \citep[see e.g.][]{Mannheim92, Mucke03, boe07}.
The location of the low-energy peak in the SED subdivides blazars into three subclasses of high-, intermediate-, and low-energy peaked sources (HBLs, IBLs, and LBLs, respectively), depending on the position of their synchrotron peak frequencies \citep[see e.g.][]{padovani95,fossati98, Abdo2010}.
\object{PKS\,0048$-$097}\ (RA$_{2000}$: $00^h50^m41.317^s$, DEC$_{2000}$: $-09\degree 29' 05.21''$) is a BL Lac-type blazar \citep{Plotkin08} located at redshift $z=0.635$ \citep{Landoni12}, which is well studied in the optical and radio regime \citep[e.g.][]{Ross70, Carswell73, Pica88, Wills92, Falomo96, Stickel93, Sefako01}.
The source was reported in the first \citep{1fgl} and second \citep{2fgl} Fermi-LAT Catalogs and Fermi Bright Gamma-ray Source List \citep{0fgl}, known as 1FGL, 2FGL and 0FGL, respectively. No detection in the very high energy gamma-rays range has been reported, meaning that there is only an upper limit on flux provided by the H.E.S.S. Collaboration \citep{hess_up}.
\cite{Abdo2010} found the frequency of lower peak to be $\log \nu_{syn}=14.3$, which allows classifying the object as an IBL-type source.
For the first time, we present here the results of nine years of multi-wavelength observations of \object{PKS\,0048$-$097}\ covering the radio, optical, UV, X-ray, and high-energy gamma-ray wavelengths.
The paper is organized as follows: Sect.~\ref{data} describes the multi-frequency observations of the source, Sect.~\ref{behaviour} shows the behaviour of \object{PKS\,0048$-$097}\ in both the quiescence and flaring states and the spectral properties in GeV range. Section~\ref{sec:correlations} discuss the multi-frequency correlations. The work is summarized in Sect.~\ref{summary}.
\section{Multi-wavelength data} \label{data}
\subsection{Gamma-ray monitoring with Fermi/LAT}
The LAT, on-board the \textit{Fermi} satellite, is a pair-conversion detector that is sensitive to photons in the energy band from $\sim 20$\,MeV to a few hundred GeV \citep{atwood}. This primary mission instrument covers the full sky every three\,hours, and the data are available publicly on the mission web page\footnote{\url{http://fermi.gsfc.nasa.gov/ssc/data/access/}}.
\object{PKS\,0048$-$097}\ is included in 2FGL (as well as in previous catalogues: 0FGL and 1FGL) as 2FGL\,0050.6$-$0926 with an average flux between 100\,MeV and 100\,GeV of $(3.85 \pm 0.25)\cdot10^{-9}$\,ph\,cm$^{-2}$\,s$^{-1}$ and the spectral index of $\Gamma_\textrm{2FGL} = 2.14 \pm 0.04$.
For comparison, the 1FGL reports an average flux in the same energy range of $(4.50 \pm 0.45) \cdot 10^{-9}$\,ph\,cm$^{-2}$\,s$^{-1}$ and a spectral index of $\Gamma_\textrm{1FGL} = 2.19 \pm 0.05$, while the reported flux level in 0FGL is $(7.2 \pm 1.0) \cdot 10^{-9}$\,ph\,cm$^{-2}$\,s$^{-1}$.
For this paper data collected between August 4,$^{}$ 2008 and July 1,$^{}$ 2014 were analysed using standard \textit{\textup{Fermi Science tools}} (version v9r33p0) with \verb|P7REP_SOURCE_V15_rev1| instrument response functions (IRFs). For the analysis the photons with a zenith angle $<105^\circ$ were selected in the energy range of 100\,MeV to 300\,GeV. Events were selected in a $15^\circ$ region of interest (ROI) centred on \object{PKS\,0048$-$097}\ .
The binned maximum-likelihood method \citep{Mattox96} was applied in the analysis.
The Galactic diffuse background was modelled using the \verb|gll_iem_v05| map
cube, and the extragalactic diffuse and residual instrument backgrounds were
modelled jointly using the \verb|isotropic_iem_v05| template. All the sources from the {\it Fermi}-LAT Second Source Catalog
\citep[2FGL,][]{2fgl} inside the ROI of \object{PKS\,0048$-$097}\ were modelled.
The long-term light curve binned in seven-day intervals is presented in Fig.~\ref{all_lc}.
\subsection{X-ray observations with \textit{Swift}}
The \textit{Swift} mission \citep{Gehrels04} is a multi-wavelength space observatory launched into orbit on November 20$^{}$, 2004. The instrument is equipped with the following detectors: the Burst Alert Telescope \citep[BAT,][]{Barthelmy05}, the X-ray Telescope \citep[XRT,][]{Burrows05}, and the Ultraviolet/Optical Telescope \citep[UVOT,][]{Roming05}.
\object{PKS\,0048$-$097}\ was monitored with \textit{Swift}/XRT\ in nine pointed observations (all in PC mode), resulting in total exposure of 32.1\,ks.
These data were analysed using version 6.15 of the HEASOFT package\footnote{\url{http://heasarc.gsfc.nasa.gov/docs/software/lheasoft}}. Data were recalibrated using the standard procedure \verb|xrtpipeline|. All the observations were checked for the pile-up effect, which was found to be negligible.
Spectral analysis was performed for data in the energy range of 0.3-10\,keV with the latest version of the \verb|XSPEC| package (version 12.8.2).
All data were binned to have at least 30 counts per bin. Spectra are well fitted with a power-law function with a Galactic absorption
value of $N_{H} = 3.22 \cdot 10^{20}$\,cm$^{-2}$ \citep{Kalberla05} set as a frozen parameter.
The power-law fit parameters are collected in Table~\ref{table_xrt}.
The long-term light curve of the integrated flux in the energy range of 2.0-10\,keV, presented in Fig.~\ref{all_lc}, shows the significant variability of the source. There are no significant changes in the spectral index for the period of observations (see Table~\ref{table_xrt}).
\subsection{UV observations with \textit{Swift}}
The UVOT instrument measures the UV and optical emission simultaneous to the X-ray telescope.
The observations are taken in the UV and optical bands with
central wavelengths of UVW2 (188 nm), UVM2 (217 nm), UVW1 (251 nm), U (345 nm), B
(439 nm), and V (544 nm).
The instrumental magnitudes were calculated using \verb|uvotsource| taking into account all
photons from a circular region with radius 5''. The background was determined from a circular region with radius 10'' near the source region.
The flux conversion factors used are provided by \cite{Poole08}.
All UVOT data were corrected for the dust absorption using the reddening $E(B-V)$ = 0.0274\,mag \citep{Schlafly} and the ratios of the extinction to reddening, $A_{\lambda} / E(B-V)$, for each filter \citep{Giommi06}.
All the measured magnitudes are collected in Table~\ref{table_uvot}, while the
\textit{Swift}/UVOT\ light curve is presented in Fig.~\ref{all_lc}.
\subsection{Optical monitoring with Catalina and with KAIT}
\object{PKS\,0048$-$097}\ is one of the frequently monitored targets by the two optical instruments Catalina and KAIT, which observe the blazar in V and R bands, respectively. Both light curves were used in this paper.
The Catalina Survey \cite[][]{Drake09} consists of the Catalina Sky Survey (CSS) and the Catalina Real-time Transient Survey (CRTS).
Here only CSS data in V band are used, which are publicly available on the instrument web page\footnote{\url{http://nesssi.cacr.caltech.edu/catalina/Blazars/Blazar.html}}.
The second optical monitoring we present was obtained with The Katzman Automatic Imaging Telescope Gamma-Ray Burst
KAIT \citep{Li00}. KAIT is the third robotic telescope in the Berkeley Automatic Imaging Telescope
(BAIT) program \citep{Richmond93,Treffers95}. The instrument monitors 163 AGNs with an average cadence of three days.
Data points in the light curves in R band \citep{Li03} are produced through a pipeline described by \cite{Cohen14} and available on the program web page\footnote{\url{http://brando.astro.berkeley.edu/kait/agn/}}.
The host galaxy of \object{PKS\,0048$-$097}\ remains unresolved \citep{Kotilainen98}, and because of this,
the data were not corrected for the influence of the host.
All optical magnitudes were corrected against the Galactic extinction based on the model by \cite{Schlegel98} with the most recent recalibration by \cite{Schlafly}, using the correction factors of $A_V=0.088$ and $A_R=0.069$, for V and R band, respectively.
\subsection{Radio observations with OVRO}
The radio observations of \object{PKS\,0048$-$097}\ were carried out at 15\,GHz with the Owens Valley Radio Observatory (OVRO), which is the 40\,m telescope dedicated to observe \emph{Fermi}-LAT targets \citep{Richards11}. Data used in this analysis were collected between January 6,$^{}$ 2008 to May 8,$^{}$ 2014 and were downloaded from the programme website\footnote{\url{http://www.astro.caltech.edu/ovroblazars}}.
\begin{figure*}[t]
\centering{\includegraphics[width=0.98\textwidth]{plots/only_lc.pdf}}
\caption[]{Multi-wavelength light curve of \object{PKS\,0048$-$097}. Panels from the top to the bottom show the radio observations from the OVRO telescope at 15\,GHz; optical (KAIT and Catalina) monitoring in R and V band, Swift/UVOT data in V, B, U, UVW1, UVM2, and UVW2 filters, X-ray observations with Swift/XRT, and Fermi-/LAT flux ($E>100$\,MeV). The horizontal dashed lines in all the panels represent the average flux for all observations presented here. In all the cases (excluding Fermi/LAT data) each point corresponds to one night of observations; for the case of Fermi/LAT monitoring data are binned in week-long intervals.
In the Fermi/LAT light curve, flux upper limits are represented with grey points.
The time periods corresponding to A and B flares are marked in grey, while the quiescence period Q is plotted in blue.}
\label{all_lc}
\end{figure*}
\section{Multi-wavelength behaviour and flaring activity of PKS\,0048$-$097} \label{behaviour}
Studies on multi-wavelength behaviour using simultaneous observations of the blazars, both in quiescence and flaring state, are crucial to understand the physics and nature of these objects.
\object{PKS\,0048$-$097}\ is a highly variable source, significant flux changes are observed in all wavelengths (see Fig.~\ref{all_lc}).
\subsection{Temporal and spectral variability of the blazar}
To quantify the temporal variability that is observed in different wavelengths, the fractional variability amplitude is calculated using formula provided by \cite{Vaughan03},
\begin{equation}
F_{var}= \sqrt{\frac{S^2-e^2}{F^2}},
\end{equation}
where $S^2$ is the variance, $e^2$ is the mean square error, and $F$ is the mean flux.
The uncertainty of $F_{var}$ is calculated following the formula by \cite{Poutanen08},
\begin{equation}
\delta F_{var}= \sqrt{F_{var}^2+(\sigma^2)} -F_{var},
\end{equation}
with the error in the normalised excess variance $\sigma$ given as \citep{Vaughan03}
\begin{equation}
\sigma = \sqrt{\left( \sqrt{\frac{2}{N} }\frac{e^2}{F^2} \right)^2 + \left( \sqrt{\frac{e^2}{N}}\frac{2 F_{var}}{F} \right)^2 },
\end{equation}
where $N$ is the number of data points in the light curve.
The calculated values of $F_{var}$ for all the energy bands we
analysed are collected in Table~\ref{table_fvar}.
Figure~\ref{fvar} shows changes of the $F_{var}$ values in different energy bands.
During the whole monitoring period of the blazar, significant variability was revealed in the $F_{var}$ values. The lowest $F_{var}$ value is for the GeV observation, the highest for X-ray range, while for the other ranges the $F_{var}$ values are similar.
We note that for \textit{Fermi}/LAT\ data $F_{var}$ was only calculated for
the flux points, and the upper limit points were omitted from the calculations.
It is important to mention that the values of $F_{var}$ are strongly dependent
on the size of the time bins in the light curves. Smaller bins allow showing
stronger flux variations, which in the case of larger time bins can be smoothed out
and lead to lower $F_{var}$ values. Obviously, the time binning is limited by
the characteristics of particular instruments.
The second factor that influences the values of $F_{var}$ are the flux
uncertainties. The uncertainties according to the definition are expected to be
constant (or at least very close to constant), which in practice is not always
true. For example, such a case is visible in the optical monitoring with
Catalina (see Fig.~\ref{all_lc}), in which some measurements have much larger
uncertainties than the others.
Figure~\ref{all_lc} reveals the variable nature of the source. Nine years of this this long-term monitoring allow pointing out significant variability patterns that are observed in different wavelengths.
\subsection*{Radio variability}
In the radio band the flux oscillates between $0.2\cdot10^{-13}$\,erg\,cm$^{-2}$\,s$^{-1}$ and $3.6\cdot10^{-13}$\,erg\,cm$^{-2}$\,s$^{-1}$.
The highest mentioned value corresponds to about 2.4\,Jy, the lowest to about 0.2\,Jy.
The highest flux level is observed during the period of MJD54762$-$MJD54998, while the lowest values are collected during the period of MJD55321$-$MJD55505.
As mentioned before, \object{PKS\,0048$-$097}\ was a target of several previous radio monitorings. Long-term radio observations provided by the Michigan group\footnote{\url{https://dept.astro.lsa.umich.edu/}} during the period of 1970-2010 showed, for instance, that the flux changes at 14.5\,GHz between 0.3-2.8\,Jy, with the highest value archived in 1993. The second prominent outburst reported by the scientists is observed at the end of 2009 with an observed flux level of about 2.4\,Jy. The second mentioned flare can correspond to the main outburst (MJD54900) visible in OVRO observations.
\subsection*{Optical variability}
The optical monitoring with Catalina includes 100\,months of observations, while the monitoring with KAIT includes about 54\,months. The significant optical variability of the source is revealed in several flares that were observed both in R and V band. The strongest outburst took place between MJD55822 and MJD55951, and it shows flux changes in the range of about $3\cdot10^{-11}$\,erg\,cm$^{-2}$\,s$^{-1}$ in V band (during this period \object{PKS\,0048$-$097}\ was not monitored with KAIT).
Observations of the blazar with KAIT reveal a prominent outburst with 3.5\,mag amplitude. We highlight here that this is the largest optical flare ever reported for this blazar. Previous studies that focused on the optical variability of \object{PKS\,0048$-$097}\ have also shown a significant outburst of 3\,mag \citep[][]{ Usher74}.
\subsection*{X-ray and UV variability}
Nine pointing observations of \object{PKS\,0048$-$097}\ with \textit{Swift}/UVOT\ and \textit{Swift}/XRT\ also show changes in the flux. In the X-ray range the largest outburst was observed in June 2009.
The highest flares observed with \textit{Swift}/UVOT\ occurred in June 2009 and January 2012. The first flare is simultaneous with that of \textit{Swift}/XRT, while the second optical-UV outburst is not correlated with the X-ray flux.
No significant variability is observed in the \textit{Swift}/XRT\ spectral index (see Fig.~\ref{spec_lc}).
\subsection*{GeV temporal and spectral variability}\label{gev_var}
The gamma-ray monitoring of the blazar presented here includes more than five years of observations beginning from the mission start date. The observed flux in this energy range shows significant temporal variability. The variability is also observed for the spectral index in this wavelength.
We note that in the GeV regime the observed emission is not very strong because the light curve (see Fig.~\ref{all_lc}) includes many of the flux upper limits points and the flux points have large error bars, even though the bin size is set to seven days.
To study variability patterns in the GeV domain in greater detail,
we investigated three intervals that are defined below.
All the intervals are marked in Fig.~\ref{all_lc}.
\begin{itemize}
\item \textbf{Flare A interval}.
The first flare (hereafter flare A) occurred between $\sim$MJD\,54683 (\textit{Fermi}-LAT mission start date) and MJD\, 54770.1 and is characterized by the elevated flux in GeV, optical, and radio range. There are no observations performed with \textit{Swift}/XRT\ and \textit{Swift}/UVOT\ during this period, therefore we cannot confirm or to exclude the hypothesis about a high state of the blazar in UV and soft X-rays.
\item \textbf{Flare B interval}.
The second flare (hereafter flare B) is observed between MJD\,55900 and MJD\,55980 in GeV, X-ray, and optical bands. Radio monitoring during this period only shows slight flux oscillation below the mean flux. Surprisingly, two observations with \textit{Swift}/XRT\ do not show an exceptionally high flux. Short time variations of the elevated flux can be noticed in the UV and optical data obtained with \textit{Swift}/UVOT, Catalina, and KAIT. The shape of the peak in the GeV range shows two separate components that cannot be distinguished in the optical observations in either KAIT or Catalina data.
\item \textbf{The quiescent-state interval}.
The quiescent state of the source has been chosen between MJD\,55450 and MJD\,55830. During this period the flux oscillates around the mean value. The choice of the quiescence state is dictated by two aspects: the flux during the chosen period should not exhibit any significant outbursts, and the length of the interval should be long enough to determine a good-quality spectrum. As mentioned before, large flux uncertainties can suggest variability, but within the error bars the variations are not significant
in the observation period.
\end{itemize}
To study the spectral properties of \object{PKS\,0048$-$097}, we used\ four time periods: two flares A and B, quiescent (low) state (as defined above), and the time period covering all observations.
For each of the defined intervals the photon spectrum was calculated using two spectral models following \cite[e.g.][]{Massaro04}, a single power-law:
\begin{equation}
F(E)=N_p \left( \frac{E}{E_0}\right)^{-{\Gamma}},
\end{equation}
and a log-parabolic one:
\begin{equation}
F(E)=N_l \left( \frac{E}{E_0}\right)^{-({\alpha+\beta \log (E/E_0)})}.
\end{equation}
For both models the break energy was set to $E_0=100$\,MeV and
was frozen in the fitting procedure.
The spectral points were calculated by dividing the data set into five logarithmically equal energy bins and a separate likelihood analysis was run for each bin. Unfortunately, \object{PKS\,0048$-$097}\ is too faint to obtain flux points in the highest energies in the GeV domain. In this case ($\textrm{TS}<9$), only the flux upper limits were derived. A one-sigma butterfly contour was calculated using the covariance matrix obtained with the \verb|gtlike| procedure \citep{Abdo09}.
The parameters of the spectral fit and the test statistic (TS) for each model and time interval are collected in Table~\ref{table_lat}. The TS for each interval favours the log-parabolic scenario for the \textit{Fermi}/LAT\ data. The log-parabolic fits and the spectral points are shown in the $\nu F_{\nu}$ representation in Fig.~\ref{sed}.
We recall that in the log-parabolic model the $\alpha$ parameter corresponds to the spectral index, while $\beta$ gives information about the curvature. The spectral parameters collected in Table~\ref{table_lat} do not show significant spectral variability; $\alpha$ is almost constant within the error bars. The curvature parameter changes, but again uncertainties are too large to confirm or exclude significant changes of this parameter. Furthermore, for the B flare $\beta$ = 0.04 $\pm$ 0.04, which means that here a scenario with zero curvature is also possible, which reduces to a single power-law description.
\subsection{Colour-magnitude relation}
Figure \ref{color-mag} shows the colour-magnitude diagram for the optical observations of \object{PKS\,0048$-$097}\ obtained with KAIT and Catalina. Neither a bluer-when-brighter nor a redder-when-brighter chromatism is found in data. The result is consistent with those reported by \cite{Ikejiri2011} and \cite{Wierzcholska2014} for this blazar. The Pearson correlation coefficient calculated for the colour-magnitude relation is $0.45\pm0.11$ (see Appendix \ref{appendix:error} for details on how the uncertainty was estimated).
We also note that the optical observations of \object{PKS\,0048$-$097}\ were not corrected for the contribution of the host galaxy or for the contamination from the emission of the accretion disc.
\section{Multi-frequency correlations} \label{sec:correlations}
The simultaneous long-term multi-frequency observations of \object{PKS\,0048$-$097}\ readily allow for correlation studies.
The standard way to quantify possible relations between two datasets is the discrete correlation function (DCF) following \cite{Edelson88}. Unfortunately, this method does not work for sparse and non-uniformly sampled light curves.
In this case, the cross-correlation function can be better estimated by the $z$-transformed discrete correlation function (ZDCF). The algorithm has been described in detail by \cite{tal97}.
The ZDCF was calculated for four cases: a comparison of optical data in R and V band, optical data in V band and radio data, optical data in V band and $\gamma$-ray data, and radio data and $\gamma$-ray. To find the peak location for the ZDCF, a maximum-likelihood was calculated for each case using the PLIKE algorithm \citep{tal13}. The peak location,
$\tau_\textrm{max}$, represents the most probable time-lag between the
two light curves.
For each combination of the two light curves, we calculated the following quantities: $\tau_\textrm{max}$, ZDCF$(\tau_\textrm{max})$, and the Pearson correlation coefficient for the given light curves shifted according to the $\tau_\textrm{max}$. The results are gathered in Table \ref{table_zdcf}.
In all the cases the flux-flux relation for offset $\Delta t = 0$ and for $\Delta t = \tau_\textrm{max}$ are presented in Fig.~\ref{ZDCF}. If necessary, the light curves were binned according to the bins of the \textit{Fermi}/LAT\ light curves. For the flux-flux relation the Pearson correlation coefficient was calculated, and its error was estimated using a Monte Carlo simulation (see Appendix \ref{appendix:error} for details).
Then the flux-flux relation was fitted with a linear function.
The calculated ZDCF values strongly suggest a correlation between the optical observations in V and R band with a time-lag of zero\,days.
For the other comparisons, no significant correlations for any time-lag were found.
Radio and optical data show a maximum of the ZDCF function for a time lag of about 200\,days
with a correlation coefficient of 0.5, which does not allow stating it as significant.
For $\gamma$-ray and optical data the maximum of ZDCF does not exceed 0.5.
Similar results were also found for the radio and $\gamma$-ray data set.
On the other hand, for the second and fourth case (comparison
of optical - radio and $\gamma$-ray - radio) the Pearson correlation coefficient calculated for the corresponding $\tau_\textrm{max}$ (see Table \ref{table_zdcf}) is about 0.6. This indicates a weak correlation between the emission at the two mentioned wavelengths. We note here that Pearson's correlation coefficient was calculated for binned data in seven-day intervals. This may influence the results because in this case the compared data are only quasi-simultaneous. Moreover, as mentioned before, $\gamma$-ray data have large uncertainties, which weakens the statistical importance of this relation.
\section{Summary and conclusions} \label{summary}
We have presented multi-wavelength monitoring of the blazar \object{PKS\,0048$-$097}\ for nine years, consisting of observations performed in the radio band with OVRO, in the optical and UV bands with KAIT, Catalina and \textit{Swift}/UVOT, in the X-ray band with \textit{Swift}/XRT, and in the high energy $\gamma$-ray wavelength with \textit{Fermi}/LAT.
It is the longest published monitoring of this source ever. It is also worth mentioning that except in the optical and radio ranges, \object{PKS\,0048$-$097}\ was not studied in detail at the different wavelengths.
During the period studied here, substantial variability is observed in all wavelengths, which in the case of the optical and radio band is consistent with observations reported for example by \cite{Ross70, Carswell73, Wills92, Falomo96, Stickel93, Sefako01}.
But for the case of other frequencies this is the first time that variability is reported.
We observed 3.5\,mag variations in the optical band in KAIT data during prominent outbursts, which is the strongest change in the flux for this source ever reported.
The variability of \object{PKS\,0048$-$097}\ that is observed in the other wavelengths seems to be more complex.
The first flare, called~A here, is characterized by higher flux in the GeV and optical range.
The interpolation of the radio light curve indicates that the observed emission does not show a flare. The second flare discussed in this paper, flare~B, is described by higher flux in the GeV, optical, and UV range. However, neither in radio range nor in X-ray are there any significant changes in the flux level.
Highly variable emission was quantified using $F_{var}$ values, and it indicates that the highest variability is found in the X-ray range and the lowest in the GeV domain.
The correlation studies of \object{PKS\,0048$-$097}\ show a strong correlation between the optical emission in V and R band and no significant correlations between other wavelengths. The analysis of the cross-correlation function did not show any time lag for which a linear relation could be resolved.
The variability observed in different time scales is a common feature for this class of objects.
But for the case of the eponymous blazar, the fact that variations are not correlated suggests that the standard, most popular leptonic SSC model might not be enough to describe the emission processes in this object.
The comparison of the optical colour and magnitude of \object{PKS\,0048$-$097}\ does not show either a bluer-when-brighter or a redder-when-brighter relation for the blazar. The lack of such a relation has also been reported by \cite{Ikejiri2011} and \cite{Wierzcholska2014}.
The bluer-when-brighter trend is a common feature of the BL Lac-type blazars, where the optical emission is mostly dominated by the synchrotron radiation from the jet. The lack of such behaviour can be caused by the fact that the correlation is hidden by a few overlapping branches, for which chromatism is present individually. For this case it is possible to distinguish separate states in the colour-magnitude diagram, for which bluer-when-brighter relation is evident. This scenario is possible for \object{PKS\,0048$-$097}\, where the colour-magnitude can disclose substructures with a clear bluer-when-brighter relation.
The optical data we presented were not corrected for either the contribution of the host galaxy or for the contamination from the accretion disc. \object{PKS\,0048$-$097}\ is a luminous LBL-type blazar (with an optical luminosity of about 10$^{46}$\,erg\,s$^{-1}$) with
an apparently high accretion rate \citep{Sbarrato12}. Not taking into account the accretion disc effects may lead to redder-when-brighter results in colour-magnitude diagrams, while the neglecting the host galaxy contribution causes a bluer-when-brighter effect. Hence, according to the optical observations of \object{PKS\,0048$-$097}, the lack of any chromatism in the colour-magnitude diagram does not definitely exclude such a relation for the blazar.
The spectral variability studies in GeV energy range show that in this domain the favourable model to describe spectra is the log-parabolic one.
Previously, such a model was successfully used in the spectral analysis for other blazars, as reported for instance by
\cite{Massaro04, Massaro04b, Massaro06, Massaro08, Tramacere07, Tramacere09, Tramacere11, Giommi05, Perri07, Donato05}.
The studies on the long-term emission in \object{PKS\,0048$-$097}\ confirm the importance of the simultaneous multi-frequency monitoring of blazars in
both the flaring and quiescence states. Our results strongly encourage further monitoring of \object{PKS\,0048$-$097}\ in the multi-frequency simultaneous campaigns.
\begin{figure}[t]
\centering{\includegraphics[width=0.45\textwidth]{plots/spec_lc.pdf}}
\caption[]{Temporal and spectral variability of \object{PKS\,0048$-$097}\ observed with Swift/XRT and Fermi-/LAT. The following panels present
the X-ray temporal and spectral variability and the $\gamma$-ray temporal and spectral variability.
For Swift/XRT observations one point correspond to one day of observations, while in the case of Fermi-/LAT data one point is for weekly integrated observations. We plot the flux upper
limits n the Fermi/LAT light curve with grey points.
The time periods corresponding to A and B flares are marked in grey, while the quiescence period Q is shown in blue.}
\label{spec_lc}
\end{figure}
\begin{table*}
\caption[]{Parameters of the spectral analysis of \textit{Swift}/XRT data. }
\centering
\begin{tabular}{c|c|c|c|c|c}
\hline
\hline
Observation ID & Observation date & Exposure & $F_{2-10\,\textrm{keV}}$ & $\Gamma$ & $\chi^2_{red}/n_{d.o.f.}$ \\
& &(ks)&(10$^{-12}$\,erg\,cm$^{-2}$\,s$^{-1}$) & & \\
(1) & (2) & (3) & (4) & (5) & (6) \\
\hline
36364001 & 04/06/2008 & 0.8 & $1.60\pm0.18$ & $2.33\pm0.14$ & 0.645/10 \\
38093001 & 24/05/2009 & 5.7 & $4.13\pm0.04$ & $2.27\pm0.04$ & 1.029/121 \\
36364002 & 25/05/2009 & 9.7 & $5.42\pm0.03$ & $2.18\pm0.03$ & 0.905/179 \\
36364003 & 20/06/2009 & 2.6 & $4.39\pm0.06$ & $2.12\pm0.06$ & 0.879/61 \\
36364004 & 23/12/2009 & 5.0 & $2.23\pm0.08$ & $2.12\pm0.07$ & 1.045/54 \\
41714001 & 18/11/2010 & 1.6 & $0.54\pm0.03$ & $1.74\pm0.67$ & 0.542/4 \\
38093002 & 10/07/2011 & 3.8 & $0.32\pm0.03$ & $2.15\pm0.18$ & 0.415/6 \\
38093003 & 30/12/2011 & 1.5 & $1.05\pm0.26$ & $1.98\pm0.21$ & 1.40/7 \\
38093004 & 03/01/2012 & 1.4 & $0.86\pm0.19$ & $2.49\pm0.17$ & 1.303/11 \\
All observations & & 32.1 & $3.20\pm0.02$ & $2.22\pm0.02$ & 1.091/252 \\
\hline
\hline
\end{tabular}
\tablefoot{The following columns report (1) the observation ID, (2) the observation time, (3) the exposure of analysed observations, (4) the integrated flux in the energy range from 2 to 10\,keV, (5) the photon index for the power law fit to the spectrum, (6) the reduced $\chi^2$ and the number of degrees of freedom for the power-law fit. The given $F_{2-10\,\textrm{keV}}$ fluxes are not corrected for Galactic absorption, since in energy band of 2-10\,keV this effect is found to be negligible.}
\label{table_xrt}
\end{table*}
\begin{table*}
\caption[]{Magnitudes for different epochs from \textit{Swift}/UVOT data for V, B, U, UVW1, UVM2, and UVW2 filters.}
\centering
\begin{tabular}{c|c|c|c|c|c |c}
\hline
\hline
Observation ID & V & B & U & UVW1 & UVM2 & UVW2 \\
\hline
36364001 & -- & -- & -- & -- & -- & $15.13\pm0.06$ \\
38093001 & $15.49\pm0.07$ & $15.91\pm0.06$ & $15.12\pm0.06$ & $15.07\pm0.06$ & $15.14\pm0.07$ & $15.25\pm0.06$ \\
36364002 & -- & -- & $15.09\pm0.05$ & -- & -- & -- \\
36364003 & -- & -- & -- & -- & $15.64\pm0.07$ & -- \\
36364004 & $16.39\pm0.10$ & $17.00\pm0.09$ & $16.03\pm0.08$ & $16.05\pm0.08$ & $16.04\pm0.09$ & $16.23\pm0.07$ \\
41714001 & $16.23\pm0.10$ & $16.67\pm0.09$ & $15.84\pm0.09$ & $15.98\pm0.09$ & $16.02\pm0.10$ & $16.21\pm0.08$ \\
38093002 & $15.67\pm0.07$ & $16.10\pm0.06$ & $15.32\pm0.05$ & $15.33\pm0.07$ & $15.33\pm0.07$ & $15.49\pm0.06$ \\
38093003 & $15.14\pm0.06$ & $15.54\pm0.05$ & $14.75\pm0.06$ & $14.81\pm0.06$ & $14.76\pm0.07$ & $15.04\pm0.06$ \\
38093004 & $14.69\pm0.05$ & $15.05\pm0.05$ & $14.21\pm0.05$ & $14.22\pm0.06$ & $14.24\pm0.06$ & $14.39\pm0.06$ \\
\hline
\hline
\end{tabular}
\tablefoot{The magnitudes are corrected for Galactic extinction.\\ (--) No observation taken in this filter for the given observation ID.}
\label{table_uvot}
\end{table*}
\begin{figure}[t]
\centering{\includegraphics[width=0.50\textwidth]{plots/fvar_vs_hz.pdf}}
\caption[]{Fractional variability vs. frequency for each observation regime. The numerical values can be found in Table~\ref{table_fvar}. The colours for the data points are the same as in Fig.~\ref{all_lc}.}
\label{fvar}
\end{figure}
\begin{table*}
\caption[]{Fractional variability in different energy bands. }
\centering
\begin{tabular}{c|c|c|c|c}
\hline
\hline
Instrument & Energy band/filter & $F_{var}$ &$\chi^2$/$n_{d.o.f.}$ & Bin size\\
(1) & (2) & (3) & (4) & (5) \\
\hline
\textit{Fermi}/LAT\ & 0.1 -- 300\,GeV & $0.340 \pm 0.073$ & 532/109 & 7\,days\\
\textit{Swift}/XRT\ & 2 -- 10\,keV & $0.850 \pm 0.004$ & 876/7 & --\\
\textit{Swift}/UVOT & UVW2 & $0.597 \pm 0.020$ & 1225/5 & -- \\
\textit{Swift}/UVOT & UVM2 & $0.654 \pm 0.017$ & 1449/5 & --\\
\textit{Swift}/UVOT & UVW1 & $0.651 \pm 0.029$ & 665/4 & -- \\
\textit{Swift}/UVOT & U & $0.590 \pm 0.019$ & 1022/5 & --\\
\textit{Swift}/UVOT & B & $0.640 \pm 0.017$ & 1394/4 & --\\
\textit{Swift}/UVOT & V & $0.590 \pm 0.022$ & 642/4 & --\\
Catalina & V & $0.500 \pm 0.039$ & 29830/316 & 1\,day \\
KAIT & R & $0.455 \pm 0.003$ & 349399/164 & 1\,day\\
OVRO & 15\,GHz & $0.495 \pm 0.002$ & 196843/377 & 1\,day \\
\hline
\hline
\end{tabular}
\tablefoot{The following columns present (1) the name of the instrument, (2) the energy band or filter, (3) the fractional variability, (4) the chi square value and the number of degrees of freedom for the fit with a constant, (5) the size of the data bins. In the case of \textit{Swift}/XRT\ and \textit{Swift}/UVOT\ data due to small number of pointing observations the bin sizes are not provided.}
\label{table_fvar}
\end{table*}
\begin{table*}
\caption[]{Summary of $z$-transformed discrete correlation function.}
\centering
\begin{tabular}{c|c|c|c|c|c}
\hline
\hline
energy bands & $\tau_\textrm{max}$ & time interval & ZDCF$(\tau_\textrm{max})$ & probability & $R_{bin}$ \\
& [days] & & & & \\
(1) & (2) & (3) & (4) & (5) & (6) \\
\hline
V -- R & 7.9 & (-6.0;+20.4) & $0.860\pm0.026$ & 77$\%$ & $0.95\pm0.11$ \\
V -- radio & 186.0 & (+7; +217) & $0.608\pm0.055$ & 52$\%$ & $0.57\pm0.06$ \\
V -- $\gamma$-ray & $-$23.0 & (-88; +21) & $0.493\pm0.012$ & 44$\%$ & $0.47\pm0.15$ \\
radio -- $\gamma$-ray & $-$166.0 & (-200;+148) & $0.536\pm0.074$ & 46$\%$ & $0.59\pm0.08$ \\
\hline
\hline
\end{tabular}
\tablefoot{The following columns present (1) the energy bands for which ZDCF is calculated, (2) the calculated time lag, (3) the time interval for which the time lag is calculated, (4) the ZDCF value for $\tau_\textrm{max}$, (5) the probability for ZDCF, (6) the Pearson correlation coefficient.}
\label{table_zdcf}
\end{table*}
\begin{figure}[t]
\centering{\includegraphics[width=0.48\textwidth]{plots/plot_color_mag.pdf}}
\caption[]{Colour-magnitude plot for \object{PKS\,0048$-$097} . Each point corresponds to seven days of binned observations. }
\label{color-mag}
\end{figure}
\begin{figure*}[]
\centering{\includegraphics[width=0.49\textwidth]{plots/zdcf_v_r.pdf}}
\centering{\includegraphics[width=0.49\textwidth]{plots/zdcf_v_ovro.pdf}}\\
\centering{\includegraphics[width=0.49\textwidth]{plots/zdcf_v_fermi.pdf}}
\centering{\includegraphics[width=0.49\textwidth]{plots/zdcf_ovro_fermi.pdf}}
\caption[]{Estimate of the cross-correlation function for different time lags for long-term light curve of \object{PKS\,0048$-$097}\ (see Fig.\ref{all_lc}). All data are binned in weekly averaged intervals. The following subplots show the CCF for (a) optical flux in V band, $F_V$ versus optical flux in R band, $F_R$ as a function of time delay; (b) optical flux in V band, $F_V$ versus radio flux $F_\textrm{RADIO}$ as a function of time delay; (c) optical flux in V band, $F_V$ versus $\gamma$-ray integrated flux $I_\gamma$ as a function of time delay; (d) radio flux, $F_\textrm{RADIO}$ versus $\gamma$-ray integrated flux, $I_\gamma$ as a function of time delay. }
\label{ZDCF}
\end{figure*}
\begin{figure*}[t]
\centering{\includegraphics[width=0.49\textwidth]{plots/cor_v_r.pdf}}
\centering{\includegraphics[width=0.49\textwidth]{plots/cor_v_ovro186.pdf}} \\
\centering{\includegraphics[width=0.49\textwidth]{plots/cor_v_fermi_dt-23.pdf}}
\centering{\includegraphics[width=0.49\textwidth]{plots/cor_ovro_fermi_dt-166.pdf}}
\caption[]{Subplot (a) shows the comparison of the optical flux in R ($F_R$) and V ($F_V$) band for \object{PKS\,0048$-$097}\ for simultaneous observations. The Pearson correlation coefficients for the two cases of a full set of data - solid line and without one data point (in right upper corner in the plot) - dashed line, are provided in the lower right corner. The following subplots show the relation for shifted light curves for (b) the radio flux, $F_\textrm{RADIO}$ vs $F_V$, (c) integrated gamma-ray flux, $I_\gamma$ vs $F_V$, and (d) $I_\gamma$ vs $F_\textrm{RADIO}$. The light curves are shifted by $\Delta t$ according to $\tau_\textrm{max}$ shown in Table. \ref{table_zdcf}. In the lower right corner the Pearson correlation coefficients are provided. }
\label{correlation_all}
\end{figure*}
\begin{table*}
\caption[]{Power-law and log-parabolic fit parameters to \textit{Fermi}/LAT\ data.}
\centering
\begin{tabular}{c|c|c|c|c|c}
\hline
\hline
& TS & N (10$^{-1}$ cm$^{-2}$ s$^{-1}$ MeV$^{-1}$) & $\gamma$ & $\alpha$ & $\beta$ \\
& (1) & (2) & (3) & (4) & (5) \\
\hline
All data -- power law & 2290 & $ 4.54 \pm 0.33$ & $2.09 \pm 0.03$ & -- & -- \\
All data -- log-parabola & 2293 & $ 3.55 \pm 0.39 $ & -- & $1.86 \pm 0.07$ & $0.04 \pm 0.01$ \\
Flare A -- power law & 555 & $ 13.5\pm 1.9 $ & $2.11 \pm 0.06$ & -- & -- \\
Flare A -- log-parabola & 559 & $ 9.0\pm 2.3 $ & -- & $1.67 \pm 0.22$ & $0.09 \pm 0.04$ \\
Flare B -- power law & 304 & $ 7.9\pm 1.8 $ & $2.02 \pm 0.08$ & -- & -- \\
Flare B -- log-parabola & 304 & $ 6.4\pm 2.2 $ & -- & $1.82 \pm 0.25$ & $0.04 \pm 0.04$ \\
Quiescent state -- power law & 379 & $ 3.51 \pm 0.70 $ & $2.01 \pm 0.07$ & -- & -- \\
Quiescent state -- log-parabola & 382 & $ 2.04 \pm 0.56 $ & -- & $1.55 \pm 0.17$ & $0.08 \pm 0.03$ \\
\hline
\hline
\end{tabular}
\tablefoot{The following columns present (1) the test statistic for a given fit, (2) the fit normalization: $N = N_p$ for power
law or $N = N_l$ for log-parabola, (3) the spectral index for the power-law fit, (4) the slope parameter for the log-parabolic fit, (5) the curvature parameter forthe log-parabolic fit. For all power-law and log-parabolic models the break energy was a frozen parameter with 100\,MeV.}
\label{table_lat}
\end{table*}
\begin{figure}[t]
\centering{\includegraphics[width=0.48\textwidth]{plots/fermi_sed.pdf}}
\caption[]{Spectral energy distribution for \textit{Fermi}/LAT\ data. Data are fitted using log-parabola models. The following colours corresponds to different time epochs: flare A (red), flare B (magenta), quiescent period (Q) defined in Sect.~\ref{gev_var} (green), and all data period (M, blue). }
\label{sed}
\end{figure}
\section*{Acknowledgements}
The author thanks the anonymous referee for constructive
comments that greatly improved the manuscript.
A.W. acknowledges support by Polish Ministry of Science and Higher Education in Mobility Plus Program.
This research has made use of data from the OVRO 40-m monitoring program
\citep{Richards11}, which is supported in part by NASA grants NNX08AW31G and NNX11A043G, and NSF grants AST-0808050 and AST-1109911.
The CSS survey is funded by the National Aeronautics and Space
Administration under Grant No. NNG05GF22G issued through the Science
Mission Directorate Near-Earth Objects Observations Program. The CRTS
survey is supported by the U.S.~National Science Foundation under
grants AST-0909182 and AST-1313422.
The author gratefully acknowledges the optical, X-ray, and $\gamma$-ray observations provided by the KAIT, \textit{Swift} and \textit{Fermi} teams, respectively.
This research was supported in part by PLGrid Infrastructure. The plots presented in this paper are rendered using Matplotlib \citep{matplotlib}.
\begin{appendix}
\section{Error estimation for the Pearson correlation coefficient}
\label{appendix:error}
To estimate the uncertainty of the Pearson correlation coefficient, one can use
the Monte Carlo approach.
Here we assumed that a
set of points $A = \{(x_i,y_i)\}$ is given and each of the point has its own
corresponding uncertainty values $(\Delta x_i,\Delta y_i)$. In the first step,
for each point new coordinates were drawn randomly according to the normal
distribution for which the mean was set to $x_i$ (or $y_i$) and the standard
deviation to $\Delta x_i$ (or $\Delta y_i$). This results in a new set of points
$A'$ , and its Pearson correlations coefficient is $C'$.
Repeating the procedure $N$ times gives a set of Pearson coefficients
$\{C'_n\}$. If $N$ is large enough, a histogram of the $\{C_n'\}$ should have
roughly a Gaussian shape. However, because the Pearson coefficient only has values in the range $[-1,1]$ it is good to apply a~Fisher transformation (see
below) on each of the $C'$ value before making the histogram. An example of
such a histogram is shown in Figure~\ref{fig:gauss}.
To find the uncertainty of the Pearson coefficient, a Gaussian function was
fitted to the histogram. The standard deviation of this fit can be used as an
estimate of the Pearson coefficient uncertainty of the original set of
points $A$. The value found by fitting should be transformed back by the
reverse Fisher transformation.
\subsection*{Fisher transformation}
The Fisher transformation allows representing values that span from $[-1,1]$
in a range of $(-\infty,\infty)$. It is defined as follows:
\begin{equation}
z = \frac{1}{2} \ln \left( \frac{1+r}{1-r} \right) = \atanh r,
\end{equation}
and the reverse:
\begin{equation}
r = \frac{\exp (2z) - 1}{\exp (2z) + 1}.
\end{equation}
\begin{figure}
\includegraphics[width=0.45\textwidth]{plots/error.pdf}
\caption{Histogram of the Monte Carlo-generated set of $10^5$ Pearson correlation coefficients $\{C'\}$. The red
line shows the Gaussian fit. The $C'$ values were transformed with
the Fisher transformation. The reverse Fisher transformation of the fit gives
a Pearson correlation coefficient equal to 0.87, and its error is estimated
to 0.03.}
\label{fig:gauss}
\end{figure}
\end{appendix}
\bibliographystyle{aa}
|
\section{Introduction}
\setcounter{equation}{0}
The Fourier transform on various locally
compact groups, and its properties with respect to different
operations on function spaces on these groups are well understood.
The interaction of the Fourier transform with the translations on
the groups, and with certain products on the functions defined on
these groups have been used to obtain characterisations of the
Fourier transform. For more details, refer to \cite{AAM1}-\cite{LT}, and the references therein.\\
We denote by $\mathcal{S}(\mathbb{R}^n),$ the Schwartz class of
rapidly decreasing functions on $\mathbb{R}^n,$ defined as
follows:\\
For a function $f: \mathbb{R}^n \rightarrow \mathbb{C},$ let
$$\|f\|_{\alpha,\beta} : = \sup \limits_{x\in \mathbb{R}^n} \ |x^\alpha \partial^\beta f(x)|,$$
where for multi-indices $\alpha=(\alpha_1,\cdots , \alpha_n),$ and
$\beta = (\beta_1,\cdots , \beta_n)\in ~\mathbb{N}^n,$ we denote
$x^\alpha = \prod\limits_{j=1}^n x_j^{\alpha_j},$ and $
\partial^\beta = \frac{\partial^{\beta_1}}{\partial
x_1^{\beta_1}}\frac{\partial^{\beta_2}}{\partial x_2^{\beta_2}}
\cdots \frac{\partial^{\beta_n}}{\partial x_n^{\beta_n}}.$ \\
The \textit{Schwartz class} of functions, denoted
$\mathcal{S}(\mathbb{R}^n),$ or simply $\mathcal{S},$ is defined
to be
$$\mathcal{S}(\mathbb{R}^n):=\{f:\mathbb{R}^n \rightarrow \mathbb{C} : f\in \mathcal{C}^\infty(\mathbb{R}^n),
\|f\|_{\alpha,\beta} <\infty \textrm{ \ for \ all \ } \alpha,
\beta \in \mathbb{N}^n\}.$$ Then the space
$\mathcal{C}_c^\infty(\mathbb{R}^n),$ also denoted
$\mathcal{C}_c^\infty,$ of compactly supported smooth functions
defined on $\mathbb{R}^n,$ is a
subspace of $\mathcal{S}(\mathbb{R}^n).$\\
\noindent The topology generated by the family of seminorms
$\{\|\cdot\|_{\alpha,\beta}: \alpha,\beta\in \mathbb{N}^n\}$ makes
$\mathcal{S}(\mathbb{R}^n)$ into a Fr\'{e}chet space over the
complex numbers. Also, $\mathcal{S}(\mathbb{R}^n)$ is closed under
the operations of pointwise and convolution product, where the
convolution of functions in $\mathcal{S}(\mathbb{R}^n)$ is defined
as
$$(f\ast g)(x) = \int\limits_{\mathbb{R}^n} f(x-y) \ g(y) \ dy, \ x\in \mathbb{R}^n.$$
For a function $f\in \mathcal{S}(\mathbb{R}^n),$ its Fourier
transform $\mathcal{F}f$ is defined as
$$\mathcal{F}f(\xi) = (2\pi)^{-n/2} \int\limits_{\mathbb{R}^n} f(x) \ e^{-ix\cdot \xi} \ dx, \ \xi\in \mathbb{R}^n.$$
The space of all continuous linear functionals on
$\mathcal{S}(\mathbb{R}^n)$ is called the \textit{space of
tempered distributions}, and is denoted by
$\mathcal{S}'(\mathbb{R}^n).$ We denote the action of $\varphi\in
\mathcal{S}'(\mathbb{R}^n)$ on a
function $f\in \mathcal{S}(\mathbb{R}^n)$ as $\langle \varphi,f \rangle .$\\
The operations of pointwise multiplication and convolution of
functions in $\mathcal{S}(\mathbb{R}^n)$ can be appropriately
extended to $\mathcal{S}'(\mathbb{R}^n)$ as follows:
\noindent For $f,g \in \mathcal{S}(\mathbb{R}^n)$ and $\varphi \in
\mathcal{S}'(\mathbb{R}^n)$,
\begin{eqnarray*}
\langle f\cdot \varphi,g\rangle &=& \langle \varphi,f\cdot g\rangle \\
\langle f\ast \varphi,g\rangle &=& \langle \varphi, \tilde{f} \ast g\rangle ,
\end{eqnarray*}
where $\tilde{f}(x)=f(-x)$ for $x \in \mathbb{R}^n.$
Then for $f,g\in \mathcal{S}(\mathbb{R}^n)$ and $\varphi \in
\mathcal{S}'(\mathbb{R}^n),$ we have $f\cdot \varphi \in \mathcal{S}'(\mathbb{R}^n)$ and $f\ast
\varphi \in \mathcal{S}'(\mathbb{R}^n).$\\
\noindent The Fourier transform, initially defined on
$\mathcal{S}(\mathbb{R}^n),$ can be extended to the space
$\mathcal{S}'(\mathbb{R}^n)$
via $$\langle \mathcal{F}\varphi,f \rangle = \langle \varphi, \mathcal{F}f \rangle , \ \textrm{for} \ f\in \mathcal{S}(\mathbb{R}^n), \ \varphi \in \mathcal{S}'(\mathbb{R}^n).$$
The Fourier transform is a topological isomorphism of
$\mathcal{S}'(\mathbb{R}^n)$ onto itself and satisfies
\begin{eqnarray*}
\mathcal{F} (f\cdot \varphi)&=& \mathcal{F}(f)\ast \mathcal{F}(\varphi) \\
\mathcal{F}(f\ast \varphi)&=& \mathcal{F}(f) \cdot
\mathcal{F}(\varphi).
\end{eqnarray*}
In \cite{AAM2} S. Alesker, S. Artstein-Avidan and V. Milman gave a
very interesting characterisation of the Fourier transform on the
Schwartz class of functions on $\mathbb{R}^n.$ The precise
statement of their result is as follows:
\begin{thm}
Assume that $T:\mathcal{S}(\mathbb{R}^n) \rightarrow
\mathcal{S}(\mathbb{R}^n)$ is a bijection which admits a bijective
extension $T': \mathcal{S}'(\mathbb{R}^n) \rightarrow
\mathcal{S}'(\mathbb{R}^n)$ such that for all $f\in
\mathcal{S}(\mathbb{R}^n)$ and $\varphi \in
\mathcal{S}'(\mathbb{R}^n),$ we have
$$T(f\ast \varphi) = T(f) T(\varphi) \textrm{ \ and \ } T(f\cdot \varphi) = T(f) \ast T(\varphi).$$
Then $T$ is essentially the Fourier transform: that is, for some
matrix $B\in GL(n,\mathbb{R})$ with $|det \ B|=1,$ we have either
$Tf= \mathcal{F}(f\circ B)$ or $Tf=\mathcal{F}(\overline{f\circ
B})$ for all functions $f\in \mathcal{S}(\mathbb{R}^n).$
\end{thm}
As the authors of the above result had remarked, the hypotheses of
this result involves only \textit{algebraic} properties of the map
on the class of tempered distributions, whereas the conclusion
states
that the map is essentially the Fourier transform.\\
Motivated by the above result, a characterisation of the Fourier
transform on the Schwartz space of the Heisenberg group was
obtained in \cite{LT}. This result did not involve any hypothesis
in terms of the tempered distributions. The anonymous referee of
\cite{LT} suggested if a characterisation of the Fourier transform
on $\mathbb{R}^n,$ without any assumptions on the tempered
distributions, could be obtained. This paper is an attempt towards
a positive answer to this question.\\
For a function $f:\mathbb{R}^n\rightarrow \mathbb{C},$ the support
of $f,$ denoted $Supp \ f,$ is defined as
$$Supp \ f := Closure({\{x\in \mathbb{R}^n : f(x) \neq 0\}}).$$
\section{A Characterisation of Fourier transform on $\mathcal{S}(\mathbb{R}^n)$}
We remark that our results are very much influenced by the those
of Alesker et al.\cite{AAM2} and their interesting proofs.\\
\noindent Our main result is the following:
\begin{thm}\label{LaT} Let $T:\mathcal{S}(\mathbb{R}^n) \rightarrow \mathcal{S}(\mathbb{R}^n)$ be a bijection satisfying the following conditions for all functions
$f,g \in \mathcal{S}(\mathbb{R}^n):$
\begin{description}
\item[(a)] $T(f+\overline{g}) = T(f)+[T(g)]^*,$ where $[Tg]^*(x)
= \overline{Tg(-x)}, \ x\in ~\mathbb{R}^n.$
\item[(b)] $T(f\cdot g)=
T(f) \ast T(g )$, \item[(c)] $T(f\ast g)= T(g)\cdot T(g).$
\end{description} Then there exists a matrix $B\in GL(n,\mathbb{R}),$ with $|det \ B|=1$ such that
either $Tf = \mathcal{F}(f\circ B)$ for all $f\in
\mathcal{S}(\mathbb{R}^n),$ or $Tf = \mathcal{F}(\overline{f\circ
B})$ for all $f\in \mathcal{S}(\mathbb{R}^n).$
\end{thm}
\begin{proof}
For $f\in \mathcal{S}(\mathbb{R}^n),$ we have $Tf \in
\mathcal{S}(\mathbb{R}^n).$ Since the Fourier transform
$\mathcal{F}$ is a bijection on $\mathcal{S}(\mathbb{R}^n),$ there
exists unique $g\in \mathcal{S}(\mathbb{R}^n)$ with
$Tf=\mathcal{F}g.$ Define a map $U:
\mathcal{S}(\mathbb{R}^n)\rightarrow \mathcal{S}(\mathbb{R}^n)$ as
$Uf:=g$ if $Tf=\mathcal{F}g.$ Then $Tf = \mathcal{F}(Uf)$ for all
$f\in \mathcal{S}.$ The map $U$ is a bijection of $\mathcal{S}$
onto itself and satisfies the following conditions for all
functions $f,g\in\mathcal{S}(\mathbb{R}^n):$
\begin{enumerate}
\item $U(f+\overline{g}) = U(f)+\overline{U(g)},$ \item $U(f\cdot
g)= U(f) \cdot U(g )$, \item $U(f\ast g)= U(f)\ast U(g).$
\end{enumerate}
The theorem is a then a consequence of the following result, which gives a precise description of the map $U.$\\
\end{proof}
\begin{thm}\label{LaU} Let $U:\mathcal{S}(\mathbb{R}^n) \rightarrow \mathcal{S}(\mathbb{R}^n)$ be a bijection satisfying the following conditions for all functions
$f,g \in \mathcal{S}(\mathbb{R}^n):$
\begin{enumerate} \item $U(f+\overline{g}) = U(f)+\overline{U(g)},$ \item $U(f\cdot g)=
U(f) \cdot U(g )$, \item $U(f\ast g)= U(f)\ast U(g).$
\end{enumerate} Then there exists a matrix $B\in GL(n,\mathbb{R})$ with $|det \ B| =1$ such that
either $Uf(x) = f(Bx)$ for all $f\in \mathcal{S}(\mathbb{R}^n),$
or $Uf(x) = \overline{f(Bx)}$ for all $f\in
\mathcal{S}(\mathbb{R}^n).$
\end{thm}
\begin{proof} We prove the result in 12 steps.
\noindent For $x_0\in \mathbb{R}^n,$ define $$C(x_0): = \{f \in
\mathcal{S}(\mathbb{R}^n): x_0\in Supp\ f\}.$$\\
\vspace{-0.8cm}
\noindent \textbf{Step 1.} Let $f,g \in \mathcal{S}.$ If $g=1$ on
$Supp \ f,$ then $Ug=1$ on
$Supp \ Uf.$\\
\noindent \textit{Proof of Step 1.} Since $g=1$ on $Supp \ f,$ we
have $f\cdot g =f.$ This gives $Uf= U(f\cdot g) = Uf \cdot Ug,$
and so $Ug=1$ on the set $\{x : Uf(x) \neq 0\}.$\\
\noindent Let $x\in Supp \ Uf$ with $Uf(x)=0.$ Then there is a
sequence $\{x_k\}_{k\in \mathbb{N}} \subseteq ~\mathbb{R}^n$ with
$Uf(x_k) \neq 0$ for all $k$ and $x_k \rightarrow x$ as $k
\rightarrow \infty.$ Since $Uf(x_k) \neq 0,$ we have $Ug(x_k) =1$
for all $k.$ Hence $Ug(x) = \lim \limits_{k \rightarrow \infty} \
Ug(x_k) = 1.$
Thus $Ug=1$ on $Supp \ Uf.$\\
\noindent \textbf{Step 2.} If $f \in \mathcal{C}_c^\infty,$ then $Uf \in \mathcal{C}_c^\infty.$\\
\noindent \textit{Proof of Step 2.} Choose $f\in
\mathcal{C}_c^\infty$ such that $f(x_0) \neq 0.$ Choose $g\in
\mathcal{S}$ such that $g=1$ on $Supp \ f.$ By Step 1, $Ug=1$ on
$Supp \ Uf.$
Since $Ug\in \mathcal{S},$ we get $Supp \ f$ is compact.\\
\noindent \textbf{Step 3.} For any $x_0 \in \mathbb{R}^n,$ there
exists $y_0 \in \mathbb{R}^n$ such
that $Uf\in C(y_0)$ whenever $f \in C(x_0).$\\
\noindent \textit{Proof of Step 3. }Let $E:=\{f\in \mathcal{S}:
f(x_0) \neq 0\}.$ Fix a function $g\in \mathcal{C}_c^\infty$ with
$g(x_0) \neq 0.$ By Step 2, we have $K: = Supp \ Ug$ is compact.\\
\noindent For $f\in E,$ define $K_f:= K \cap Supp \ Uf.$ For
functions $f_0:=g,f_1,\cdots ,f_k\in E,$ we have $\prod_{j=0}^k \
f_j\not \equiv 0,$ and so $\prod_{j=0}^k \ Uf_j= U(\prod_{j=0}^k \
f_j) \not \equiv 0,$ which gives $\cap _{j=0} ^k \ K_{f_j} \neq
\emptyset.$ This means, the collection $\{K_f: f\in E\}$ of closed
subsets of $K$ has finite intersection property. Since $K$ is
compact, this gives $\bigcap \limits_{f\in E} \ K_f \neq
\emptyset.$ Let $y_0 \in \bigcap
\limits_{f\in E} \ K_f.$\\
\noindent \textbf{Claim.} If $f\in C(x_0),$ then $Uf \in C(y_0).$\\
\noindent \textit{Proof of Claim.} We prove the claim in two
separate
cases.\\
\noindent \textbf{Case 1.} $f(x_0) \neq 0.$\\
Then
$f$ never vanishes on a neighborhood, say, $V$ of $x_0.$ Let $g\in
\mathcal{S}$ be such that $f\cdot g = 1$ on $V.$ Choose $h\in
\mathcal{S}$ such that $h=1$ on a neighborhood $W$ of $x_0,$ and
satisfies $W\subseteq Supp \ h \subseteq V.$ Since $f\cdot g=1$ on
$Supp \ h,$ by Step 1, we get $U(f\cdot g)=Uf\cdot Ug =1$ on
$Supp\
Uh.$ Since $h\in E,$ by definition, $y_0 \in Supp \ Uh .$ This implies $Uf(y_0) \neq 0,$ and hence $Uf\in C(y_0).$\\
\noindent Note that all our arguments till now can be applied to
the map $U^{-1}$ as well, and so we have proved that $f(x_0)\neq
0$ if and
only if $Uf(y_0) \neq 0.$\\
\noindent A function $f\in \mathcal{S}$ is said to satisfy the
condition $(\star)$ if the following holds:
$$(\star) \ \ \ \ \ \ \ \ \ \ f(x_0)=0 \textrm{ \ if \ and\ only \ if \ } Uf(x_0)=0.\hspace{4cm}$$
By the above discussion, we have that all functions in
$\mathcal{S}$ satisfy condition $(\star).$\\
\noindent \textbf{Case 2.} $f(x_0) =0.$\\
\noindent Suppose $Uf \not \in C(y_0).$ Then there is a
neighbourhood, say $W,$ of $y_0$ such that $Uf$ vanishes
identically on $W.$ Let $h\in \mathcal{S}$ with $Supp \ h
\subseteq W,$ and $h(y_0) \neq 0.$ There exists unique function $g
\in \mathcal{S}$ with $Ug=h.$ Then $U(f\cdot g) = Uf\cdot Ug = Uf
\cdot h \equiv 0.$ This gives $f\cdot g
\equiv 0.$ \\
\noindent On the other hand, since $Ug(y_0) = h(y_0) \neq 0,$ by
Condition $(\star),$ we have $g(x_0) \neq 0,$ and so g is never
zero near $x_0.$ Since $x_0\in Supp \ f,$
this implies $f\cdot g \not \equiv 0,$ a contradiction. Thus $Uf \in
C(y_0).$\\
\noindent \textbf{Step 4.} Define a map $A: \mathbb{R}^n
\rightarrow \mathbb{R}^n$ as follows: $Ax=y$ if $Uf\in C(y)$
whenever $f\in C(x).$ Then the map $A$ is well-defined.\\
\noindent \textit{Proof of Step 4.} Suppose for some $x_0 \in
\mathbb{R}^n,$ we have $Ax_0 = y_1$ and $Ax_0 = y_2$ with $y_1\neq
y_2.$ Let $V_1$ and $V_2$ be disjoint neighborhoods of $y_1$ and
$y_2,$ respectively. There exists functions $g_1$ and $g_2$ in
$\mathcal{S}$ which are supported
in $V_1$ and $V_2,$ respectively, such that $g_1(y_1) \neq 0$ and $g_2(y_2) \neq 0.$ Let $f_1,f_2 \in \mathcal{S}$ with $Uf_1=g_1$ and
$Uf_2=g_2.$ Then $0\equiv g_1\cdot g_2 =Uf_1 \cdot Uf_2 = U(f_1 \cdot f_2)$ and so $f_1\cdot f_2 \equiv
0.$\\
On the other hand, as $g(y_j) = Uf_j(y_j)\neq 0$ for $j=1,2,$ we
have by
Condition $(\star)$ that $f_j(x_0) \neq
0$
for $j=1,2,$ which is in contradiction to $f_1\cdot f_2\equiv 0.$\\
\noindent \textbf{Step 5.} $A:\mathbb{R}^n \rightarrow \mathbb{R}^n$ is a bijection.\\
\noindent \textit{Proof of Step 5.} The hypotheses of the theorem
hold good for the map $U^{-1}$ as well. Applying the preceding
steps to the map $U^{-1}$ gives rise to a well-defined function,
say,
$B:\mathbb{R}^n \rightarrow \mathbb{R}^n.$ Then $B = A^{-1}$, proving that $A$ is bijective.\\
\noindent Our observations can be summarised as:
$$A(Supp \ f) = Supp \ Uf,\textrm{ \ for \ all \ } f\in \mathcal{S}.$$
\noindent \textbf{Step 6.} The map $A$ is a homeomorphism of $\mathbb{R}^n$ onto itself.\\
\noindent \textit{Proof of Step 6.} Suppose not. Then there exist
$x\in \mathbb{R}^n,$ and sequence $\{x_k\}$ in $\mathbb{R}^n$ with
$x_k
\rightarrow x$ as $k \rightarrow \infty,$ but $Ax_k$ does not converge to $Ax.$\\
\noindent Let $V$ be a neighborhood of $Ax$ such that $Ax_k
\not\in V$ for any $k.$ Let $h\in \mathcal{S}$ with $Supp \
h\subseteq V,$ and $h(Ax) = 1.$ Let $g\in \mathcal{S}$ be such
that $Ug=h.$ Then $Ax_k \not \in Supp \ Ug$ for any $k,$ and so
$x_k \not\in Supp \ g$ for any $k.$ This gives
$g(x_k) = 0$ for all $k,$ implying $g(x)= 0,$ which is not
possible by Condition $(\star),$ since $Ug(Ax) =1.$\\
We observe that the above argument holds good when the maps $U$
and $A$ are replaced with $U^{-1}$ and $A^{-1}$ respectively,
yielding that
$A: \mathbb{R}^n \rightarrow \mathbb{R}^n$ is a homeomorphism.\\
\noindent \textbf{Step 7.} The map A satisfies $A(x+y) = Ax+Ay$ for all $x,y \in \mathbb{R}^n.$\\
\noindent \textit{Proof of Step 7.} Suppose $A(x+y) \neq Ax+Ay$ for some $x,y\in \mathbb{R}^n.$\\
\noindent Then there exist disjoint neighborhoods $V_{x+y},V_{xy}$
with $A(x+y) \in ~V_{x+y}$ and $Ax+Ay \in V_{xy}.$ By continuity
of the map $A,$ this gives rise to a neighborhood $W_{x+y,1}$ of
$(x+y)$ with $A(W_{x+y,1}) \subseteq V_{x+y}.$ By continuity of
addition in $\mathbb{R}^n,$ we get neighborhoods $W_{x,1},
W_{y,1}$ of $x$ and $y,$ respectively, such that $W_{x,1}+W_{y,1}
\subseteq W_{x+y,1}.$ Thus
\begin{eqnarray} \vspace{-1cm} A(W_{x,1}+W_{y,1}) &\subseteq&
A(W_{x+y,1}) \subseteq V_{x+y}.\label{eq1}
\end{eqnarray}
On the other hand, by continuity of addition in $\mathbb{R}^n,$
$Ax+Ay \in V_{xy}$ gives neighborhoods $V_{x,2},V_{y,2}$ such that
\begin{eqnarray} Ax\in V_{x,2}, \ Ay\in V_{y,2} \textrm{ \ and \ }
V_{x,2}+V_{y,2}\subseteq V_{xy}.\label{eq2} \end{eqnarray}
This
implies there exist neighborhoods $W_{x,2}, W_{y,2}$ of $x$ and
$y,$ respectively, with $A(W_{x,2})\subseteq V_{x,2}$ and
$A(W_{y,2}) \subseteq V_{y,2}.$\\
Define $W_x= W_{x,1}\cap W_{x,2}, \ W_y = W_{y,1}\cap W_{y,2}.$
Then
\begin{eqnarray}
A(W_x)& \subseteq & A(W_{x,2}) \subseteq V_{x,2} =V_x \ (say) \nonumber \\
A(W_y)& \subseteq & A(W_{y,2}) \subseteq V_{y,2} =V_y \ (say)\nonumber \\
A(W_x+W_y)& \subseteq & A(W_{x,1}+W_{y,1}) \subseteq A(W_{x+y,1}) \subseteq V_{x+y} \label{eq3}
\end{eqnarray}
Choose $f_x,f_y \in \mathcal{S}$ such that $Supp \ f_x \subseteq
W_x, Supp\ f_y \subseteq W_y$ and $f_x \ast ~f_y \not \equiv~0.$
Let $g_x = Uf_x$ and $g_y = Uf_y.$ Then $U(f_x\ast f_y) = g_x \ast
g_y \not\equiv 0.$\\
\noindent We have
\begin{eqnarray}
\nonumber Supp(g_x \ast g_y) &=&Supp \ U(f_x \ast f_y) \subseteq Supp \ Uf_x +Supp \ Uf_y \\
\nonumber &= & A(Supp \ f_x)+A(Supp \ f_y) \subseteq AW_x+AW_y
\\ & \subseteq & V_{x,2}+V_{y,2} \subseteq V_{xy} \textrm{ \ \ (by \
\ref{eq2})} \label{eq4}
\end{eqnarray}
But $Supp \ (f_x \ast f_y) \subseteq Supp \ f_x + Supp \ f_y
\subseteq W_x+W_y.$ By $(\ref{eq3}),$ this gives
\begin{eqnarray}
\nonumber Supp(g_x \ast g_y) &= &Supp (Uf_x \ast Uf_y) = Supp \ U(f_x \ast f_y)\\
& =& A(Supp (f_x \ast f_y)) \subseteq A(W_x+W_y)\subseteq
V_{x+y}. \label{eq5}
\end{eqnarray}
From $(\ref{eq4})$ and $(\ref{eq5})$, we get $$Supp(g_x \ast g_y)
\subseteq V_{xy} \cap V_{x+y} = \emptyset.$$ This gives $g_x\ast
g_y \equiv 0,$ a contradiction.
This proves the additivity of the map $A.$\\
\noindent \textbf{Step 8.} The map $A:\mathbb{R}^n \rightarrow
\mathbb{R}^n$ is a continuous additive bijection, and so also real
linear. Hence it is given by an invertible matrix,
which also we denote by $A.$\\
\noindent \textbf{Step 9.} 'Extension' of the map $U$ to scalars.\\
\noindent \textit{Illustration of Step 9.} For $f,g \in
\mathcal{S},$ and $c(\neq 0)\in \mathbb{C},$ we have
$$ U(cf)(x)\ Ug(x) = U(cfg)(x) = U(f)(x) \ U(cg)(x) , \ x\in \mathbb{R}^n.$$
Let $h \in \mathcal{S}$ be such that $Uh(x) \neq 0$ for any $x\in
\mathbb{R}^n.$ Then we have
\begin{eqnarray*}
U(cf)(x) &=& \frac{U(ch)(x)}{Uh(x)} \ {Uf}(x) \textrm{ \ for \ all \ } f\in \mathcal{S}\\
&=& m(c,x) \ Uf(x) \textrm{ \ (say)}.
\end{eqnarray*}
Thus $U(cf)(x) = m(c,x)\ Uf(x),$ for all $x\in \mathbb{R}^n.$ By
definition, the function $m(\cdot,\cdot)$ is continuous in the
second variable as a function
of $x\in \mathbb{R}^n.$\\
\noindent \textbf{Claim.} The function $m(\cdot,\cdot)$ is
independent of the second
variable.\\
\noindent \textit{Proof of Claim.} For $f,g \in \mathcal{S}, \ c
\in \mathbb{C},$ and $x\in \mathbb{R}^n,$ we have
\begin{eqnarray*}
U(cf\ast g) (x) &=& U(f\ast cg) (x) \\
(U(cf)\ast Ug)(x) &=& (Uf \ast U(cg))(x) \\
\int\limits_{\mathbb{R}^n} m(c,x-y)\ Uf(x-y) \ Ug(y) \ dy&=& \int\limits_{\mathbb{R}^n} Uf(x-y) \ m(c,y) \ Ug(y) \ dy \\
\end{eqnarray*}
As the above equation holds good for all functions $f,g \in
\mathcal{S},$ we have for all $F,G \in \mathcal{S},$
$$\int\limits_{\mathbb{R}^n} [m(c,x-y)-m(c,y)] \ F(x-y) \ G(y) \ dy =0, \ x\in \mathbb{R}^n.$$
Fix $x\in \mathbb{R}^n.$ Let $G\in \mathcal{S}$ with $G =1$ on $B(0,r),$ where $B(0,r)$ is
the open ball in $\mathbb{R}^n,$ centered at the origin and with radius
$r.$
Then for all functions $F \in \mathcal{C}_c^\infty$ with $Supp \ F
\subseteq B(x,r),$ we have
\begin{eqnarray*}
\int\limits_{B(0,r)} [m(c,x-y)-m(c,y)] \ F(x-y) \ dy &=&0. \\
\textrm{Thus \hspace{2.5cm} } m(c,x-y) -
m(c,y) &=& 0 \textrm{\ for all \ } y \in B(0,r).
\end{eqnarray*}
by the continuity of the map $m(\cdot,\cdot)$ in the second
variable. This gives in particular, $m(c,x) =m(c,0).$ As $x$ was
arbitrary, the above argument gives that the function $m(c,x)$ is
independent of the second variable $x\in \mathbb{R}^n.$ We define
$$m(c): = m(c,0).$$
\noindent \textbf{Step 10.} The map $m :\mathbb{C} \rightarrow
\mathbb{C}$ is an additive and multiplicative bijection, which
maps $\mathbb{R}$ onto $\mathbb{R},$ and hence we have either
$m(a) = a$ for all $a\in \mathbb{C},$ or $m(a) = \overline{a}$ for
all
$a\in \mathbb{C}.$\\
\noindent \textit{Proof of Step 10.} Let $g\in \mathcal{S}, \
a,b\in \mathbb{C}$ with $g(x) \neq 0$ for any $x\in \mathbb{R}^n.$
Then $Ug(y) \neq 0$
for any $y \in \mathbb{R}^n.$\\
Suppose $m(a) =m(b)$ for some $a,b\in \mathbb{C}.$ Then $$U(ag)
(x) = m(a) \ Ug(x)= m(b) \ Ug(x) =
U(bg)(x) , \ x\in \mathbb{R}^n.$$ Since $U$ is a bijection, this gives $a=b.$\\
\noindent By hypothesis(1), we have
\begin{eqnarray*}
m(a+\overline{b})\ Ug(x) &=& U((a+\overline{b})g)(x)
=U(ag+\overline{b}g)(x) \\
&=& U(ag)(x) + \overline{U(b\overline{g})}(x) = (m(a) +m(\overline{b} )) \ Ug(x).
\end{eqnarray*}
Since Ug is never zero, we get $m(a+\overline{b}) =
m(a)+\overline{m(b)}.$ In particular, $m(\overline{a})
=\overline{m(a)}$ for all $a\in \mathbb{C}.$\\
\noindent Now, hypothesis(2) gives
$$m(ab) Ug(x) = U(ab g)(x) = m(a)U(b g)(x) = m(a) m(b) Ug(x).$$
Again, since $Ug$ is nowhere vanishing, we get $m(ab)=m(a)m(b)$
for all $a,b\in \mathbb{C}.$\\
\noindent \textbf{Step 11.} For $f\in \mathcal{S},$ and $x_0 \in
\mathbb{R}^n,$ we have
$Uf(Ax_0) = m(f(x_0)).$\\
\noindent \textit{Proof of Step 11.} As before, choose $g \in
\mathcal{S}$ such that $g(x) \neq 0$
for any $x \in \mathbb{R}^n.$ Then $Ug(y) \neq 0$ for any $y \in \mathbb{R}^n.$ \\
\noindent Define $$h(x): = f(x_0) \ g(x) - f(x) \ g(x), \ x\in
\mathbb{R}^n.$$ Then $h\in \mathcal{S}$ and $h(x_0) = 0.$ By
Condition $(\star),$ we have $Uh(Ax_0) =0.$ This gives
\begin{eqnarray*}
0=Uh(Ax_0) &=& U(f(x_0) \cdot g - f \cdot g)(Ax_0)\\
&=& m(f(x_0)) \
Ug(Ax_0) - Uf(Ax_0) \ Ug(Ax_0). \end{eqnarray*}
Since $Ug$ is never zero, this gives
$Uf(Ax_0) = m(f(x_0)).$\\
\noindent Since $B=A^{-1},$ using Step 10, we get that either
$Uf(x_0) = f(Bx_0)$ or $Uf(x_0) = \overline{f(Bx_0)}.$\\
\noindent Thus we get that the map $U$ is as claimed by our
theorem. It remains to show that $|det \ B|=1.$\\
\noindent \textbf{Step 12.} The matrix $B$ satisfies $|det \ B| =1.$\\
\noindent \textit{Proof of Step 12.} We have
\begin{eqnarray*}
(f\ast g) (Bx) = U(f\ast g)(x) &=& (Uf \ast Ug)(x) \\
&=& \int\limits_{\mathbb{R}^n} Uf(x-y) \ Ug(y) \ dy \\
&=& \int\limits_{\mathbb{R}^n} f(B(x-y)) \ g(By) \ dy\\
&=& |det \ B|^{-n} \ \int\limits_{\mathbb{R}^n} f(Bx-y) \ g(y) \ dy
\end{eqnarray*}
Thus $|det \ B|=1,$ proving our result.
\end{proof}
|
\section{Introduction}
An integral domain is called IDF if every non-zero element of it has only a finite number of non-associate irreducible divisors. These domains were introduced by Grams and Warner in \cite{Grams1975}. They are also one of the generalizations of UFD's that were studied in the seminal paper \cite{anderson1990factorization}. Another important subclass of IDF domains are domains that contain no atoms at all; these domains were named antimatter and studied in \cite{Coykendall1999antimatter}. The class of IDF domains also includes FFD's. An integral domain is an FFD if every non-zero element of it has only a finite number of non-associate divisors. If $D$ is an FFD, then $D[X]$ is also an FFD (see \cite[Proposition~5.3]{anderson1990factorization}). A question first posed in \cite{anderson1990factorization} is whether the IDF property ascends in polynomial extensions. Malcolmson and Okoh in \cite{malcolmson2009polynomial}, answered this question in the negative. They actually proved that:
\begin{theorem}[{{\cite[Theorem~2.5]{malcolmson2009polynomial}}}]
\label{MOTheorem}
Every countable domain can be embedded in a countable antimatter domain $R_{\infty}$ such that $R_{\infty}[X]$ is not an IDF domain.
\end{theorem}
A natural question that follows is that under which additional conditions the IDF property does ascend in polynomial extensions. One such condition is when the domain, in addition to being IDF, is a valuation domain or more generally a GCD domain (see \cite[p.~14]{anderson1990factorization} and \cite[Theorem~1.9]{malcolmson2009polynomial}).
Another case is when the domain is atomic. In fact, atomic IDF domains are exactly FFD's (see \cite[Theorem~5.1]{anderson1990factorization}). In Theorem \ref{mainTheorem}, we see that the essential property of these domains sufficient for the IDF property to ascend is that any finite set of non-zero elements of such domains has only a finite (possibly zero) number of non-associate maximal common divisors (MCD for short); in this paper, we call any such domain MCD-finite. Actually, Theorem \ref{mainTheorem} shows that being MCD-finite is also a necessary condition for the ascent of IDF property. In Theorem \ref{MOstrong}, we use a modified version of a technique originally introduced by Roitman in \cite{roitman1993polynomial} to get a stronger version of Theorem \ref{MOTheorem}. In Section 3, we provide some more results and examples regarding MCD-finite domains.
In the remainder of this section, we state some terminology that are needed for the rest of the paper.
For a domain $D$, the set of its non-zero elements, units and non-zero non-units are denoted by $\nonz{D}$, $\units{D}$ and $\nunits{D}$, respectively. We call two elements $x, y \in D$ associates, and write $x \sim y$, if there exists $u \in \units{D}$ such that $a = ub$. For $x,y \in D$, we say that $x$ divides $y$, and write $x \mid y$, if there exists $z \in D$ such that $y = xz$. Also, we use the notations $a \assoc{D} b$ and $a \aadd{D} b$ to emphasize the underlying domain. Two elements $a, b \in \nonz{D}$ are called incomparable if $a \nmid b$ and $b \nmid a$ (i.e., the principal ideals $\langle a \rangle$ and $\langle b \rangle$ are incomparable). Also, following the notation and terminology of \cite{rand2015multiplicative}, the set of atoms of $D$ is denoted by $\atoms{D}$ and we call a subset $A$ of $D$ unit-closed if for every $u \in \units{D}$ and $a \in A$, we have $ua \in A$.
A domain $D$ is called atomic if every $x \in \nunits{D}$ can be written as a product of irreducible elements (atoms). If the set of principal ideals of $D$ satisfies the ascending chain condition, then $D$ is called an ACCP domain. We refer the reader to \cite{anderson1990factorization} for more on these and other domains with factorization properties.
Let $D$ be a domain and $S \subseteq \nonz{D}$. The set of all the common divisors of $S$ is denoted by
$\mathrm{CD}_D(S)$ (i.e., $\mathrm{CD}_D(S) = \{x \in \nonz{D} \colon\; x \mid s \; \text{for every} \; s \in S \}$). An element $c$ of $\mathrm{CD}_D(S)$ is called a maximal common divisor of $S$ whenever for every $d \in \mathrm{CD}_D(S)$, if $c \mid d$, then $c \sim d$. The set of all the maximal common divisors
of $S$ is denoted by $\mathrm{MCD}_D(S)$. As in \cite{roitman1993polynomial}, if for every finite subset $T$ of $\nonz{D}$, we have $\mathrm{MCD}_D(T) \neq \emptyset$, then $D$ is called an MCD domain. We say $D$ is MCD-finite if for every finite subset $T$ of $\nonz{D}$, we have $|\mathrm{MCD}_D(T)| < \infty$ (with the possibility $|\mathrm{MCD}_D(T)|=0$).
A commutative (additive) monoid $T$ is called cancellative if for all $a, b, c \in T$, if $a + b = a + c$, then $b = c$. Also, $T$ is torsion-free if for all $a, b \in T$ and $n \in \mathbb{N}$, if $na = nb$, then $a = b$. Finally, $T$ is reduced if it has no non-trivial units; i.e., if $a + b = 0$ for some $a, b \in S$, then $a = b = 0$. All the concepts defined in the two previous paragraphs can be extended to cancellative monoids in an obvious way. We also recall that the monoid ring $R[X;T]$ is a domain if and only if $T$ is cancellative and torsion-free and $D$ is a domain (see \cite[Theorem~8.1]{gilmer1984commutative}).
We say that an extension of domains $A \subseteq B$ is division-preserving if for all $x, y \in \nonz{A}$,
if $x \aadd{B} y$, then $x \aadd{A} y$. This can also be stated concisely by saying $B \cap K = A$ where $K$ is the field of fractions of $A$ (For some properties of these extensions that are relevant to this paper, see \cite[Remark~2.2]{roitman1993polynomial}).
Finally, we recall some definitions from \cite{herzog2011monomial}. Let $D$ be a domain, $R = D[X_1, \dotsc , X_n]$ and $f \in R$. Any product $X_1^{k_1} \dotsm X_n^{k_n}$ with $k_i \in \mathbb{N} \cup \{0\}$ is called a monomial. If we denote the set of all the monomials of $R$ by $\mathrm{Mon}(R)$, then $f = \sum_{u \in \mathrm{Mon}(R)}a_u u$ for some elements $a_u \in D$. The support of $f$ is defined as the set $\{u \in \mathrm{Mon}(R)\colon\; a_u \neq 0\}$. By a monomial order on $R$, we mean a total order $<$ on the set $\mathrm{Mon}(R)$ such that (1) for every $u \in \mathrm{Mon}(R)$, if $u \neq 1$, then $1 < u$, and (2) if $u, v \in \mathrm{Mon}(R)$ and $u < v$, then $uw < vw$ for all $w \in \mathrm{Mon}(R)$. For example, we can consider the pure lexicographic order (induced by the ordering $X_1 > \dotsb > X_n$), where $X_1^{a_1} \dotsm X_n^{a_n} < X_1^{b_1} \dotsm X_n^{b_n}$ if the leftmost nonzero component of $(a_1-b_1,\dotsc, a_n-b_n)$ is negative (see \cite[Example~2.1.2(c)]{herzog2011monomial}). For a fixed monomial order, the leading coefficient of $f$, denoted by $\lc{f}$, is defined as the coefficient of the largest monomial in the support of $f$.
\section{Main Results}
We begin with the following theorem, which states a necessary and sufficient condition for the ascent of IDF property in polynomial extensions.
\begin{theorem}
\label{mainTheorem}
The following are equivalent for a domain $D$:
\begin{enumerate}
\item
$D$ is IDF and MCD-finite.
\item
$D[X]$ is IDF.
\item
For any set $\mybar{X}$ of indeterminates, $D[\mybar{X}]$ is IDF.
\end{enumerate}
\end{theorem}
\begin{proof}
$(1 \Longrightarrow 3)$
If $f \in D[\mybar{X}]$, then there exist indeterminates $X_1, \dotsc, X_n \in \mybar{X}$ such that $f \in D[X_1, \dotsc, X_n]$. Now, if $g \in \atoms{D[\mybar{X}]}$ and $g \aadd{D[\mybar{X}]} f$, then it is easy to see that $g \in \atoms{D[X_1, \dotsc, X_n]}$ and $g \aadd{D[X_1, \dotsc, X_n]} f$. Moreover, for any $g_1, g_2 \in D[X_1, \dotsc, X_n]$, $g_1 \assoc{D[\mybar{X}]} g_2$ if and only if $g_1 \assoc{D[X_1, \dotsc, X_n]} g_2$. Therefore, we only need to prove this implication for the case where $\mybar{X}$ is finite. So let $R = D[X_1, \dotsc, X_n]$ and suppose that $D$ is an IDF domain such that $R$ is not IDF. We show that $D$ is not MCD-finite.
Let $f \in \nunits{R}$ be such that there exists an infinite set $\{f_i\}_{i \in I}$ of non-associate irreducible divisors of $f$. Let $K$ be the field of fractions of $D$ and $R' = K[X_1, \dotsc, X_n]$. Since $R'$ is a UFD, there exists an infinite subset $J$ of $I$ such that the elements of $\{f_i\}_{i \in J}$ are associate in $R'$. Hence, there exists a set $S$ of monomials such that for every $i \in J$ the support of $f_i$ is $S$. There are two types of irreducible elements in $R$ with a support of size 1. First type, up to associates, consists of the elements $X_1, \dotsc, X_n$. Obviously, $f$ cannot have an infinite number of non-associate irreducible divisors of this type. The second type consists of every $a \in \atoms{D}$. If an element $a \in \atoms{D}$ divides $f$, then $a$ divides every single coefficient of $f$ in $D$, and since $D$ is IDF, $f$ cannot have an infinite number of non-associate irreducible divisors of the second type. Therefore, $2 \leq |S|$, and so since for every $i \in J$, $f_i \in \atoms{R}$, the GCD of the coefficients of each $f_i$ is equal to $1$.
Now, we fix a monomial order. For all $i, j \in J$, there exists an element $u_{i, j} \in \nonz{K}$ such that $f_i u_{i, j} = f_j$, and so $u_{i,j} = \frac{\lc{f_j}}{\lc{f_i}}$. Now, fix an element $t$ of $J$. Then for every $i \in J$, $f_t = \left( \frac{\lc{f_t}}{\lc{f_i}} \right) f_i$, and so
\[ \lc{f} f_t = \lc{f} \frac{\lc{f_t}}{\lc{f_i}}f_i . \]
Using the definition of monomial order, it is not difficult to see that $\lc{f_i} \aadd{D} \lc{f}$. Hence, $\lc{f} \left( \frac{\lc{f_t}}{\lc{f_i}} \right) \in \nonz{D}$. Also, if $i \neq j$, then
\[ \lc{f}\frac{\lc{f_t}}{\lc{f_i}} \not{\assoc{D}} \lc{f}\frac{\lc{f_t}}{\lc{f_j}} ; \]
otherwise $\lc{f_i} \assoc{D} \lc{f_j}$, so $u_{i,j} \in \units{D}$, and so $f_i \assoc{R} f_j$, which is a contradiction. Since for every $i \in J$, the GCD of the coefficients of $f_i$ is equal to $1$, the set
\[ \left\{\lc{f}\frac{\lc{f_t}}{\lc{f_i}}\right\}_{i \in J} \]
is an infinite set of non-associate maximal common divisors of the coefficients of $\lc{f}f_t$, and hence $D$ is not MCD-finite.
$(3 \Longrightarrow 2)$ This is obvious.
$(2 \Longrightarrow 1)$
If $D$ is not IDF, then obviously neither is $D[X]$.
Now, let both of the domains $D$ and $D[X]$ be IDF domains, but there exist $n \in \mathbb{N}$ and elements
$a_0, a_1, \dots, a_n$ in $\nonz{D}$ with infinitely many non-associate maximal common divisors $\{c_i\}_{i \in I}$.
Let $f = a_0 + a_1X + \dotsb + a_nX^n$. For every $i \in I$, there exists an $f_i \in \nunits{D[X]}$
such that $f = c_if_i$. Since the GCD of the coefficient of each $f_i$ is 1, no element of $D$ can appear in a factorization of $f_i$ into non-units. Hence, there exists a factorization of $f_i$ with maximum number (bounded by $\deg(f_i)$) of non-unit factors; say $f_i = h_1 \dotsm h_k$. It follows that each $h_i$ is necessarily an atom. Therefore, each $f_i$ has an atomic factorization in $D[X]$. But every irreducible divisor of each of the elements $f_i$ is an irreducible divisor of $f$ too. Therefore, since $D[X]$ is an IDF domain, there exists a finite set $\{g_1, \dots, g_m \}$ of non-associate irreducible divisors of $f$ such that every $f_i$, up to associates, is equal to a product of the elements $g_i$. Hence, for every $i \in I$, there exist $t_{1,i}, \dotsc, t_{m,i} \in \mathbb{N} \cup \{0\}$ such that
\[f_i \assoc{D[X]} g_1^{t_{1,i}} \dotsm g_m^{t_{m,i}} .\]
Note that for $1 \leq j \leq m$, we have $1 \leq \deg(g_j)$. Hence, for $1 \leq j \leq m$, the set $\{t_{j,i}\}_{i \in I}$ is finite, and so there exist $k, \ell \in \mathbb{N}$ such that $k \neq \ell$ and
\[ g_1^{t_{1,k}} \dotsm g_m^{t_{m,k}} = g_1^{t_{1,\ell}} \dotsm g_m^{t_{m,\ell}} , \]
and so $f_k \assoc{D[X]} f_\ell$. Hence, $c_k \assoc{D[X]} c_\ell$, and so $c_k \assoc{D} c_\ell$, which is a contradiction.
\end{proof}
This theorem entails the previously known results \cite[Theorem~1.9]{malcolmson2009polynomial} and \cite[Proposition~5.3]{anderson1990factorization} regarding sufficient conditions for the ascent of IDF property. Moreover, in the introduction of \cite{malcolmson2009polynomial}, it is mentioned and attributed to Muhammad Zafrullah that the polynomial extension of a domain that is both IDF and pre-Schreier is IDF. Hence, by Theorem \ref{mainTheorem}, being pre-Schreier must imply being MCD-finite. Indeed, we recall that a domain $D$ is pre-Schreier if and only if the poset of its principal ideals satisfies Riesz interpolation property, i.e., for every pair of finite subsets $A$ and $B$ of principal ideals of $D$, if $A \leq B$ (i.e., for every $\langle a \rangle \in A$ and $\langle b \rangle \in B$, $\langle a \rangle \subseteq \langle b \rangle$), then there exists a principal ideal $\langle x \rangle$ such that $A \leq \langle x \rangle \leq B$ (see \cite[Theorem~1.1]{Zafrullah1087}). Hence, every finite set of non-zero elements of a pre-Schreier domain, up to associates, has at most one MCD. So the class of pre-Schreier domains is a subclass of MCD-finite domains as expected.
Before going on, it is worthwhile to compare the notions of IDF and MCD-finite domains ideal-theoretically. A domain $D$ is IDF if and only if for any non-zero principal ideal $I$, there only exists a finite number of ideals containing $I$ that are maximal with respect to being principal (see \cite[p.~12]{anderson1990factorization}). On the other hand, a domain $D$ is MCD-finite if and only if for any finitely generated ideal $I$, there only exists a finite number of ideals containing $I$ that are minimal with respect to being principal.
Now, we are going to find a way to embed an arbitrary domain $D$ in a domain $R$ such that $R$ is not MCD-finite. The task of finding monoids that are not MCD-finite is more straightforward than finding domains that are not MCD-finite. Next, we give two examples of such monoids.
\begin{example}
\label{monoiddomainex}
\begin{enumerate}
\item
The additive monoid $C = \{x \in \mathbb{Q}\colon\; 1 \leq x\} \cup \{0\}$ is not MCD-finite. In fact, let $A$ be pairwise incomparable subset of $C$ (i.e., there do not exist $x, y \in A$ such that $1 \leq |x- y|$), where $2 \leq |A|$. Then if there exists an element $y$ in $A$ such that $y < 2$, then $\mathrm{MCD}(A) = \{0\}$. Another possibility is when all the elements of $A$ are strictly greater than $2$. In this case, $A$ has an infinite number of MCD's. Finally, if $2$ is the minimum element of $A$, then $\mathrm{MCD}(A) = \{1\}$.
\item
Let $S$ be a cancellative, torsion-free and reduced monoid and suppose that there exist $s,t \in S$ such that $\mathrm{MCD}(s, t)$ contains at least two (non-associate) elements $b$ and $d$ (An elementary example of such monoid is the additive monoid $(\mathbb{N} \cup \{0\}) \setminus \{1\}$. Note that $2$ and $3$ are both MCD's of $5$ and $6$).
Set $T \coloneqq \prod_{i \in \mathbb{N}}S$. Let $s'$ and $t'$ be the elements of $T$ with all the components equal to $s$ and $t$, respectively. Let $c_i$ be the element of $T$ with $b$ in its $i$th place, and $d$ in all the other places.
Then $\{c_i\}_{i \in \mathbb{N}}$ is a set of non-associate MCD of $s'$ and $t'$ and hence, $T$ is not MCD-finite. Moreover, it is easy to see that $T$ is reduced, cancellative and torsion-free.
\end{enumerate}
\end{example}
Now, using monoid domains, we can easily embed any domain into a domain which is not MCD-finite (another way for doing this is given in Example \ref{ReesEx}). Explicitly:
\begin{lemma}
\label{monoiddomain}
Let $D$ be a domain and let $S$ be a torsion-free, cancellative monoid that is not MCD-finite. Let $R = D[X; S]$.
Then $R$ is not MCD-finite, the extension $D \subseteq R$ is division-preserving and $\atoms{D} \subseteq \atoms{R}$. Also, in the special case where $S$ is reduced, we additionally have $\units{D} = \units{R}$.
\end{lemma}
\begin{proof}
The essential observation is that if $S$ is torsion-free and cancellative, then any divisor of a monomial of $R = D[X; S]$ is itself a monomial (see \cite[Theorem~11.1]{gilmer1984commutative}). All the parts can be deduced from this (see also \cite[Lemma~3.1]{Coykendall2011embedding}).
\end{proof}
Finally, we are going to embed a domain which is not MCD-finite into a domain that is IDF and has the same unit elements in such a way that the MCD's of elements of the original domain is preserved in the new domain (thus ensuring that the new domain is not MCD-finite). For doing this, we use a modified version of a powerful construction originally used by Roitman in \cite{roitman1993polynomial} (this technique was used to construct an atomic domain such that its polynomial extension is not atomic). This technique has also been used in \cite{Coykendall2011embedding}, \cite{Coykendall2004ap} and \cite{rand2015multiplicative}.
Let $D$ be a domain and $S \subseteq \nonz{D}$. Set
\[ \mathcal{L}(D;S) \coloneqq D[ \{X_s, \frac{s}{X_s}\colon\; s \in S \} ] .\]
Rand in \cite{rand2015multiplicative} proved the following result which we restate by adding the construction used in the proof.
\begin{theorem}[{{\cite[Theorem~2.7]{rand2015multiplicative}}}]
\label{RandTheorem}
Let $D$ be a domain and let $S$ be a subset of $\atoms{D}$ that is unit-closed. For every $n \in \mathbb{N} \cup \{0\}$, we define a domain $\mathcal{T}_n(D)$ inductively as follows:
\begin{enumerate}
\item
Set $\mathcal{T}_0(D;S) \coloneqq D$.
\item
For every $n \in \mathbb{N}$, set
$\mathcal{T}_n(D;S) \coloneqq \mathcal{L}(\mathcal{T}_{n-1}(D); \atoms{\mathcal{T}_{n-1}(D)} \setminus S)$.
\end{enumerate}
Set $\mathcal{T}^{\infty}(D;S) \coloneqq \bigcup_{i \in \mathbb{N}} \mathcal{T}_i (D;S)$. Then $\units{\mathcal{T}^{\infty}(D;S)} = \units{D}$ and $\atoms{\mathcal{T}^{\infty}(D;S)} = S$.
\end{theorem}
Now, we are ready to prove the stronger version of Theorem \ref{MOTheorem}.
\begin{theorem}
\label{MOstrong}
Let $D$ be a domain and let $S$ be a unit-closed subset of $\atoms{D}$. Then there exists a domain $R$ containing $D$ such that $R$ is not MCD-finite, $\atoms{R} = S$ and $D \subseteq R$ is division-preserving.
In particular, if S does not contain an infinite set of non-associate elements, then R is an IDF domain.
\end{theorem}
\begin{proof}
Let $T$ be a cancellative, torsion-free and reduced monoid which is not MCD-finite (e.g., Example \ref{monoiddomainex}) and let $B = D[X, T]$. Then by Lemma \ref{monoiddomain}, $S \subseteq \atoms{B}$, $S$ is unit-closed in $B$, $D \subseteq B$ is division-preserving and $B$ is not MCD-finite.
Now, if we set $R \coloneqq \mathcal{T}^{\infty}(B;S)$, then by Theorem \ref{RandTheorem}, $\atoms{R} = S$. Also, by using induction and by \cite[Lemma~3.2(1)]{roitman1993polynomial}, for every $n \in \mathbb{N}$, the extension $B \subseteq \mathcal{T}_n(B;S)$ is division-preserving, and so the extensions $B \subseteq R$ and hence $D \subseteq R$ are also division-preserving.
Now, let $V$ be a finite subset of $\nonz{B}$ such that $\mathrm{MCD}_B(V)$, up to associates, is infinite. Then by \cite[Lemma~3.2(6)]{roitman1993polynomial}, for every $n \in \mathbb{N}$, we have $\mathrm{MCD}_B(V) = \mathrm{MCD}_{{\mathcal{T}}_n}(V)$. The family $\{\mathcal{T}_i(B;S)\}_{i \in \mathbb{N}}$ satisfies the conditions of \cite[Lemma~3.1]{roitman1993polynomial}, and so by part 4 of the same lemma, $\mathrm{MCD}_B(V) = \mathrm{MCD}_{R}(V)$. Finally, since $\units{B} = \units{R}$, we conclude that $R$ is not MCD-finite.
\end{proof}
\begin{remark}
In Theorem \ref{MOstrong}, if $S$ is empty, then $R$ is an antimatter domain which is not MCD-finite, and hence by Theorem \ref{mainTheorem}, $R[x]$ is not an IDF domain. Therefore, this theorem entails Theorem \ref{MOTheorem} (although, a completely different construction is used here). In particular, Theorem \ref{MOstrong} shows that, as conjectured in \cite[Problem~1]{malcolmson2009polynomial}, the countability of the domain $D$ is superfluous in Theorem \ref{MOTheorem}.
\end{remark}
\section{More on MCD-finite domains}
We begin this section by a result on the behavior of MCD-finite domains under the $D + M$ construction. First, we mention some general properties of the $D + M$ constructions.
\begin{remark}
Let $T$ be a domain that can be written in the form $K+M$, where $K$ is a field and $M$ is a non-zero maximal ideal. Let $D$ be a subfield of $K$ and $R = D + M$.
Any element of the form $m$ or $1+m$ (where $m \in M$) is an atom in $R$ if and only if it is an atom in $T$ (The case for $1 + m$ is proved in \cite[Lemma~1.5(i)]{Costa1986DM} and the case for $m$ can be proved similarly). Also, it is not difficult to see that for all $c_1, c_2 \in K$ and $0 \neq m \in M$, $c_1m \assoc{R} c_2m$ if and only if $c_1\nonz{D} = c_2\nonz{D}$.
Let $x = k+m \in T$ where $m \in M$ and $k \in K$. For convenience, we use the following notation:
\[
\widehat{x} \coloneqq
\begin{cases}
x(=m) & k= 0 \\
1 + k^{-1}m & k \neq 0
\end{cases}
\]
It can easily be proved that for all $x, y \in T$, where $x \not\in M$, if $x \aadd{T} y$, then $\widehat{x} \aadd{R} \widehat{y}$.
\end{remark}
\begin{theorem}
\label{kmnonmcdfinite}
Let $T$ be a domain that can be written in the form $K+M$, where $K$ is a field and $M$ is a non-zero maximal ideal. Let $D$ be a subfield of $K$ and $R = D + M$.
\begin{enumerate}
\item
Suppose that $M$ contains an atom. If $R$ is MCD-finite, then the group $\nonz{K}/\nonz{D}$ is finite and $T$ is MCD-finite.
\item
Suppose that every element of $M$, up to associates in $T$, has finitely many irreducible divisors in $M$. If $T$ is MCD-finite and the group $\nonz{K}/\nonz{D}$ is finite, then $R$ is MCD-finite.
\end{enumerate}
\end{theorem}
\begin{proof}
\begin{enumerate}
\item
Let $m \in M$ be an atom. Suppose that the group $\nonz{K}/\nonz{D}$ is infinite and let $\{c_i\}_{i \in I}$ be an infinite subset of $\nonz{K}$ such that $c_i\nonz{D} \neq c_j\nonz{D}$ for $i \neq j$.
Fix two elements $t$ and $v$ of $\nonz{K}$ such that $t\nonz{D} \neq v\nonz{D}$.
Then the set $\{c_im\}_{i \in I}$ is an infinite subset of non-associate divisors of $tm^2$ and $vm^2$ in $R$. On the other hand, all the elements $\{\frac{tm^2}{c_im}\}_{i \in I}$ and $\{\frac{vm^2}{c_im}\}_{i \in I}$ are atoms in $R$ and moreover, for every $i \in I$,
$ \frac{tm^2}{c_im} \not{\assoc{R}} \frac{vm^2}{c_im} ,$
and so $\{c_im\}_{i \in I}$ is an infinite set of non-associate MCD's of $tm^2$ and $vm^2$. Therefore, $R$ is not an MCD-finite domain.
Now suppose on the contrary that $T$ is not MCD-finite. So assume that there exist non-associate elements $x_1, \dots, x_n \in \nunits{T}$ with infinite number of non-associate MCD's; say $\{ y_i \}_{i \in I}$. If for some $i \in I$, $y_i \not\in M$, then $\widehat{y_i} \aadd{R} \widehat{x_j}$ for $1 \leq j \leq n$ and $\widehat{y_i} \in \mathrm{MCD}_R(\widehat{x_1}, \dotsc, \widehat{x_n})$. So we may assume that $y_i \in M$ for every $i \in I$. Suppose that
\[x_j = (k_{i,j} + m_{i,j}) y_i \qquad k_{i,j} \in K, m_{i,j} \in M .\]
Since $\nonz{K}/\nonz{D}$ is finite, we may assume that for every $1 \leq j \leq n$, the elements $k_{i,j}$ belong to the same coset.
Hence, there exists $z_j \in \nonz{K}$ such that $y_i \aadd{R} z_j x_j$ for every $i \in I$. Now
\[ \mathrm{GCD}_R \left( \frac{z_1 x_1}{y_i} , \dotsc, \frac{z_n x_n}{y_i} \right) = 1 ,\]
and so $\{ y_i \}_{i \in I}$ is also an infinite set of non-associate MCD's of $z_1 x_1 , \dotsc , z_n x_n$ in $R$, which is a contradiction.
\item
Suppose that $x_1, \dots, x_n \in \nunits{R}$ and, up to associates in $T$, $y_1, \dots, y_m$ are the MCD's of $x_1, \dotsc , x_n$ in $T$.
Suppose that $z \in \mathrm{MCD}_R(x_1, \dotsc, x_n)$.
If
\[ \mathrm{GCD}_T \left( \frac{x_1}{z}, \dotsc , \frac{x_n}{z} \right) = 1, \]
then $z \assoc{T} y_i$ for some $1 \leq i \leq m$. Now, assume that there exists $d \in \nunits{T}$ such that $d$ is a common divisor of $\frac{x_1}{z}, \dotsc , \frac{x_n}{z}$.
Note that $d$ must be in $M$ since otherwise $z\widehat{d}$ would be a common divisor of $x_1, \dotsc , x_n$ in $R$, which is a contradiction. For a similar reason, $d$ must be an atom and also an MCD of the elements $\frac{x_1}{z}, \dotsc , \frac{x_n}{z}$ in $T$. Hence, $ zd \assoc{T} y_i $ for some $1 \leq i \leq m$. Now, by the hypothesis, the set
\[ A= \left\{ \frac{y_i}{d} \in T\colon\; d \in M, d \in \atoms{T}, 1 \leq i \leq m \right\} \cup \{y_1, \dotsc, y_m \} ,\]
up to associates in $T$, is finite. We proved that for some $ [a] \in A / \!\!\!\assoc{T} $, $z \assoc{T} a$. Finally, if $c_1, \dotsc , c_k$ is a set of coset representatives of $\nonz{D}$ in $\nonz{K}$, then $z \assoc{R} c_i a$ for some $1 \leq i \leq k$.
\end{enumerate}
\end{proof}
As we have already mentioned in the paragraph after Theorem \ref{mainTheorem}, pre-Schreier domains are MCD-finite, and in fact, any finite set of a pre-Schreier domain, up to associates, has at most one MCD. Hence, every pre-Schreier domain that is not a GCD domain (see, e.g., the paragraph after \cite[Theorem~2.4]{Cohn1968}) is an example of an MCD-finite domain that is not an MCD domain. Conversely, the domain $\mathbb{Q} + X\mathbb{R}[X]$ is a non-Noetherian ACCP (and hence MCD) domain that is not MCD-finite (we recall that for any field extension $F \subseteq K$, the group $\nonz{K}/\nonz{F}$ is finite if and only if $K = F$ or $K$ is finite (see \cite[Theorem~7]{brandis1965multiplikative})). In fact, even Noetherian domains are not necessarily MCD-finite; consider $\mathbb{R} + X\mathbb{C}[X]$. Also, $\mathbb{Z} + X\mathbb{Q}[[X]]$ is an example of a GCD (and hence MCD-finite) domain that is not IDF (The behavior of IDF, ACCP, GCD and Noetherian domains under the $D + M$ construction is studied in \cite[Proposition~4.3]{anderson1990factorization}, \cite[Proposition~1.2]{anderson1990factorization}, \cite[Theorem~11]{Brewer} and \cite[Theorem~4]{Brewer}).
A trivial example of MCD-finite domains are FFD's. In particular, every Krull domain is an MCD-finite domain (see \cite[p.~14]{anderson1990factorization}). In fact, the stronger result \cite[Proposition~1]{Grams1975} also holds for MCD-finite domains. Actually, the proof of that theorem shows that any domain satisfying the hypothesis of the theorem has the following property:
\begin{center}
$(*)$ \text{The intersection of every infinite set of incomparable principal ideals is} 0 \end{center} and this property implies being both IDF and MCD-finite.
Now, we give an example of a domain with property $(*)$ that is neither pre-Schreier nor MCD domain.
\begin{example}
Let $D$ be an FFD which is not a UFD. For example, let $F \subsetneq K$ be an extension of finite fields and let $D = F + XK[X]$. Then by \cite[Proposition~5.2]{anderson1990factorization}, $D$ is an FFD. The domain $D$ satisfies ACCP and hence is an MCD domain. Since UFD's are exactly GCD domains which satisfy ACCP, we conclude that $D$ is not a pre-Schreier domain.
Now, let $R = D[Z, \{ \frac{X}{Z^n}, \frac{Y}{Z^n}\colon\; n \in \mathbb{N} \cup \{0\}] $. This domain is not an MCD domain and in fact the elements $X$ and $Y$ do not have any MCD's (see \cite[Example~5.1]{roitman1993polynomial}). But $R$ satisfies property $(*)$. To see this, suppose on the contrary that $\{ c_iR \}_{i \in I}$ is an infinite set of incomparable principal ideals of $R$ such that $\bigcap_{i \in I} c_iR \neq 0$. By \cite[Example~2]{Anderson1996FFD}, $D[Z, \frac{1}{Z}]$ is an FFD and hence there exist $i, j \in I$ such that $i \neq j$ and $c_i \assoc{T} c_j$, where $T = (D[Z, \frac{1}{Z}])[X, Y]$. But $c_i \not{\assoc{R}} c_j$ and hence there exists an $n \in \mathbb{N} \cup \{0\}$ such that $c_i = Z^nc_j$, which contradicts the incomparability assumption. Therefore, $R$ satisfies property $(*)$, hence is both IDF and MCD-finite. But it is neither pre-Schreier nor MCD domain.
\end{example}
Finally, we mention another way to embed a domain $D$ into a domain $R$ which is not MCD-finite. This time, for a finite non-singleton subset $A$ of $\nunits{D}$ that does not generate $D$, we construct the domain $R$ in such a way that the set $\mathrm{MCD}_R(A)$, up to associates, becomes infinite.
\begin{example}
\label{ReesEx}
Let $D$ be a domain and let $c_1, \dotsc, c_n$ be non-associate elements in $\nunits{D}$ such that $2 \leq n$ and $D \neq \langle c_1, \dotsc , c_n \rangle$. For every $i \in \mathbb{N} \cup \{0\}$, define the domain $R_i$ by induction as follows:
\begin{enumerate}
\item
Set $R_0 \coloneqq D$ and $I_0 \coloneqq \langle c_1, \dotsc, c_n \rangle_{R_0}$.
\item
For $i \in \mathbb{N}$, set $R_i \coloneqq R_{i-1}[X_i, \{ \frac{a}{X_i}\colon\; a \in I_{i-1} \}]$ and $I_i \coloneqq \langle c_1, \dotsc, c_n \rangle_{R_i}$.
\end{enumerate}
Note that for any $n \in \mathbb{N}$, $I_n \neq R_n$, and hence by \cite[Lemma~2.4]{roitman1993polynomial} and \cite[Lemma~2.10]{roitman1993polynomial}, for every $i \in \mathbb{N}$,
\[ X_1, \dotsc, X_i \in \mathrm{MCD}_{R_i}(c_1, \dotsc, c_n) .\]
Set $R \coloneqq \bigcup_{i \in \mathbb{N}} R_i$. By \cite[Lemma~2.3(1)]{roitman1993polynomial}, for all $i, j \in \mathbb{N} \cup \{0 \}$ with $i \leq j$, the extension $R_i \subseteq R_j$ is division-preserving, and hence by \cite[Lemma~3.1(3)]{roitman1993polynomial}, \[\{X_i\}_{i \in \mathbb{N}} \subseteq \mathrm{MCD}_R(c_1, \dotsc, c_n) ,\]
and so $R$ is not MCD-finite.
\end{example}
\noindent \textbf{Acknowledgement.} The authors would like to thank the referee whose careful reading and valuable comments improved the paper.
|
\section{Introduction}
Magnetic order in ultrathin films is largely driven by surface and interface effects that can lead to the appearance of new energy terms which are not present in the bulk. A well-known example is perpendicular magnetic anisotropy, which describes an easy anisotropy axis that appears perpendicular to the film plane in ultrathin films and multilayers. This phenomenon is induced by interface-driven changes to the orbitals in the ferromagnet and is important for current magnetic storage technologies because higher bit-densities can be achieved with perpendicular media. A more recent example of present interest concerns chiral interactions of the Dzyaloshinskii-Moriya form, which appear in similar ultrathin films in contact with a normal metal possessing large spin-orbit coupling. Despite its prediction over two decades ago~\cite{Fert:1980, fert_magnetic_1990}, compelling experimental evidence of its importance in ultrathin film systems has only been obtained during the past few years. For example, it has been shown that this interfacial Dzyaloshinskii-Moriya interaction (DMI) leads to spin spirals in Mn monolayers on W(110)~\cite{Bode:2007}, nanoscale skyrmion lattices in Fe monolayers on Ir(111)~\cite{heinze2011spontaneous}, and isolated skyrmions in Pd/Fe bilayers on Ir (111)~\cite{Romming:2013}.
While the DMI strength in such epitaxially-grown monolayer systems correlates well with predictions from \emph{ab initio} calculations~\cite{heinze2011spontaneous}, the situation for polycrystalline films grown by sputtering is not as clear. Recent experiments have shown that the interfacial DMI can be sufficiently large to promote chiral spin states in systems based on Pt/Co, where room temperature skyrmions have been reported in Pt/Co/Ir multilayers~\cite{Moreau_luchaireCNNano2016}, Pt/Co/MgO films~\cite{Boulle:2016} and in Pt/Co/Ta~\cite{Woo:2016} films, and homochiral N{\'e}el walls have been observed in Pt/Co/AlOx~\cite{Tetienne2015, Benitez:2015}. The results appear to be consistent with measurements of the frequency non-reciprocity of spin wave propagation with Brillouin light spectroscopy~\cite{Belmeguenai:2015} and predictions from electronic structure calculations~\cite{Freimuth:2014}. However, for ferromagnetic alloys such as CoFe or CoFeB, consensus is yet to be reached on the strength of possible induced chiral interactions at ferromagnet/heavy metal interfaces. Indeed, experiments have shown that chiral magnetic bubbles can be nucleated in Ta/CoFeB/TaOx~\cite{Jiang:2015}, yet other studies on the similar Ta/CoFeB/MgO system using single spin magnetometry have shown no evidence of any chiral interaction present~\cite{Tetienne2015}.
In this article, we seek to clarify the issue of the DMI strength in perpendicularly magnetized CoFeB/MgO films deposited on different heavy-metal underlayers, namely Hf, TaN, and W. In previous work, current-driven domain wall motion under applied magnetic fields was used to estimate the DMI strength for different thicknesses of these underlayers~\cite{Natcomtorrejon2014}. Here, we revisit this problem by employing techniques that do not rely on current-dependent spin torques, which possess different components (adiabatic, non-adiabatic, spin-Hall-like and Rashba-like fields) and whose collective effect on the motion of domain walls remains unclear. Instead, we employ two techniques to probe the Dzyaloshinskii-Moriya energy without strong assumptions on the dynamics. First, we use field-driven domain wall motion in the creep regime in which wall velocities are governed by a power law, where the dominant term arises from changes in the domain wall (elastic energy) due to the DMI~\cite{Je:2013, Hrabec:2014}. Second, we use Brillouin light spectroscopy to measure the non-reciprocal propagation of spin waves in the Damon-Eshbach geometry, where it has been shown that the frequency non-reciprocity is a direct measure of the DMI constant~\cite{Moon:2013, Di:2015, Belmeguenai:2015}. Both methods have been employed on the same multilayer films. While qualitative agreement is found for most cases, numerical estimates of the DMI strength for a given underlayer can differ considerably, which suggests inconsistencies remain in the underlying assumptions used to interpret these data.
This article is organized as follows. In Section II, we describe the sample structure and deposition methods. In Section III, we present results from field driven domain wall motion in the creep regime, which was characterized using Kerr effect microscopy. In Section IV, we present results from Brillouin light spectroscopy measurements in which frequency nonreciprocity is probed for spin waves in the Damon-Eshbach geometry. An analysis of the DMI strength obtained using the two methods is presented in Section V and some concluding remarks are given in Section VI.
\section{Samples}
Our multilayers were grown by sputtering and have the following nominal structure: Si/$\mathrm{SiO_2}$/X($t$)/$\mathrm{Co_{20}Fe_{60}B_{20}}$(1)/MgO(2)/Ta(1), where figures in parentheses denote the film thicknesses in nanometers. They were annealed in vacuum at 300 $^\circ$C for 1 hour. We consider in this study four different underlayers X: W (2 nm), W(3 nm), TaN (1 nm) and Hf (1 nm). The structural properties of the samples with metallic buffers were studied in \cite{LiuHayashiAPL2015}. High resolution transmission electron microscopy showed that W, Ta and Hf underlayers were amorphous in the thickness range investigated here. Larger thicknesses --i.e. 2 nm for Hf, and 3 to 5 for W-- were required to obtain crystalline transitions within the underlayer. The TaN buffers are also essentially amorphous~\cite{Natcomtorrejon2014}, while lattice fringes in MgO layers indicate a textured crystalline character that is partially replicated in the CoFeB layer \cite{LiuHayashiAPL2015}.
\section{Domain wall motion in the creep regime}
Domain wall motion in our ultrathin CoFeB layers was studied with a magneto-optical polar Kerr effect microscope with vector field capability. To determine the domain wall velocity, we first nucleated an approximately circular domain at the center of the field of view of the microscope. We then applied pulses of the perpendicular magnetic field, down to 15 $\mu$s in duration, and determined the distance travelled by the domain walls for both domain expansion and compression of the nucleated domain. The velocity was then estimated by dividing the total distance traveled by the domain wall during the duration of the field pulse~\cite{Metaxas2007}. These experiments were conducted in the presence of an additional static in-plane applied field, which was varied between -150 and 150 mT. An example of such measurements is given in Fig.~ \ref{Fig_Kerr}.
\begin{figure
\centering
\includegraphics[width=8.5cm]{FigKerr.pdf}
\caption{\label{Fig_Kerr}Differential Kerr images illustrating the expansion of a nucleated domain, where the black region indicates the area swept by the domain wall during the field pulse. The images compare motion under a pulsed perpendicular field, $H_z$, with an additional in-plane static field, $H_x$. The scale bar indicated in (a) represents 125~$\mu$m for (a)-(f) and 250~$\mu$m for (g)-(l).
(a)-(c) W (2 nm) underlayer under $\mu_0 H_z = 2.8$ mT with a pulse duration of 5~s in (a,c) and 50~s in (b).
(d)-(f) W (3 nm) underlayer under $\mu_0 H_z = 2.9$ mT with a pulse duration of 1.5~s in (d, f) and 15~s in (b).
(g)-(i) TaN (1 nm) underlayer under $\mu_0 H_z = 0.83$ mT with a pulse duration of 0.4~s in (g, i) and 1.5~s in (h).
(j)-(l) Hf (1 nm) underlayer under $\mu_0 H_z = 2.9$ mT with a pulse duration of 75~ms in (j, l) and 250~ms in (b).
The in-plane field is $\mu_0 H_x = -50$~mT in (a, d, g, j), $\mu_0 H_x = 0$~mT in (b, e, h, k), and $\mu_0 H_x = +50$~mT in (c, f, i, l).
}
\end{figure}
We paid particular attention to the placement of the sample with respect to the electromagnets in order to minimize artifacts due to crosstalk between in-plane and perpendicular field. This involved a specific procedure to precisely align the electromagnet in the sample plane which consisted in examining domain expansion for both polarities of the nucleated domain and applied in-plane fields.
For perpendicular fields well below the depinning threshold, the wall motion in ultrathin PMA films is described by the creep model. In this regime, the wall dynamics can be linked to the motion of a one-dimensional elastic string in a two dimensional disordered potential, where motion is driven by thermal activation and involves a series of avalanches. It is well established that the dependence of the wall velocity, $v$, on the applied perpendicular field, $H_z$, in this regime can be described by the following Arrhenius-type relation~\cite{Lemerle:1998},
\begin{equation}
v(H_z) = v_0 \exp\left[- \alpha\left(H_z\right)^{-\frac{1}{4}} \right],
\label{eq:creep}
\end{equation}
where $v_0$ is a velocity prefactor and $\alpha$ is a function of $H_z$ that depends on the wall (elastic energy), the pinning potential, and the thermal energy $k_B T$. A plot of the measured domain wall velocity for the different underlayers is given in Fig.~\ref{Fig_creep}.
\begin{figure}
\centering\includegraphics[width=8.5cm]{Figurecreepcurve.pdf}
\caption{(Color online) Domain wall velocity $v$ in CoFeB as a function of perpendicular field $\mu_0 H_z$ for different underlayers in zero in-plane field. The solid lines corresponds to fits based on the creep model [Eq.~(\ref{eq:creep})].
\label{Fig_creep}}
\end{figure}
By using a log-linear scale and by plotting the measured velocities as a function of $H_z^{-1/4}$, it is easy to identify the range of fields over which the wall motion remains in the creep regime. For the TaN underlayer, we observe a deviation from the creep behavior at higher applied fields where a change in the linear variation in Fig.~\ref{Fig_creep} can be seen, but for all other samples the motion remains in the creep regime for the range of applied fields considered. For the W underlayers, we note that the underlayer thickness plays an important role on the wall velocity (for the same nominal CoFeB film), where a clear difference in the slope of the velocity curves can be seen. This suggests that the domain wall pinning potential in the two multilayer stacks is different, which might arise from a difference in film morphology
To investigate the presence of a Dzyaloshinskii-Moriya interaction, we measured the wall motion in the presence of a finite in-plane magnetic field. As previous studies have shown~\cite{Kabanov:2010, Je:2013, Hrabec:2014, Lavrijsen:2015, Vanatka:2015}, an in-plane magnetic field can break the cylindrical symmetry of the domain wall energy, which is key to revealing the presence of a chiral interaction that prefers a given handedness for the domain wall. This asymmetry can be seen in the domain expansion in Fig.~\ref{Fig_Kerr}, particularly for the W underlayers [Fig.~\ref{Fig_Kerr}(a)-(c)], where the propagation is very different along the axis of the applied in-plane magnetic field. For the TaN [Fig.~\ref{Fig_Kerr}(d)-(f)] and Hf (not shown) underlayers, this asymmetry along the applied field direction is less pronounced but other features, such as an elliptical shaped bubble, are seen instead. In order to quantify this asymmetry, we plot in Fig.~\ref{Fig_Hip} the domain wall velocity as a function of the in-plane applied field $H_x$ at constant $H_z$, where the velocity along the field axis, $v_x$, and perpendicular to the field axis, $v_y$, for the different underlayers is shown.
\begin{figure}
\centering\includegraphics[width=8.5cm]{FigureTout.pdf}
\caption{(Color online) Domain wall velocity as a function of in-plane applied field $H_x$ for propagation along (squares) and perpendicular (circles) to the field direction. The propagation takes place under a static perpendicular field, $H_z$. (a) 2~nm thick W underlayer at $\mu_0 Hz = 2.8$ mT. (b) 3~nm thick W underlayer at $\mu_0 Hz = 2.9$ mT. (c) 1~nm thick TaN underlayer at $\mu_0 Hz = 0.83$ mT. (d) 1~nm thick Hf underlayer at $\mu_0 Hz = 2.9$ mT. The dashed vertical line indicates the offset field, $H_{\rm offset}$.
\label{Fig_Hip}}
\end{figure}
For each sample, a different perpendicular field $H_z$ was used to keep the wall velocities within a similar range to facilitate comparison. For the range of in-plane fields studied, the $V_y(H_x)$ curve is found to be symmetric with respect to $H_x = 0$ for all underlayers, although the curvature of the apex of the `V'-shaped curve is found to vary with the underlayer. For instance, the variation is found to be sharp for the TaN underlayer [Fig.~\ref{Fig_Hip}(c)], while the transition is smoother for both W thicknesses [Fig.~\ref{Fig_Hip} (a)-(b)]. The curves have been obtained from domain expansion for both up and down domains. We also note that the bubble growth has a strong asymmetric character in the direction parallel to the in-plane field axis [Fig.~\ref{Fig_Hip}(a)], which leads to difficulties in defining accurately the wall velocity in the direction perpendicular to the in-plane field axis ($\mathrm{v_y}$). Our convention is to define it from the vertical distance between the two points where the DW is parallel to the in-plane field [see arrow in Fig.~\ref{Fig_Kerr}(a)].
For wall motion along the field axis, on the other hand, the $v_x(H_x)$ curves exhibit a deformed `V'-shaped curve and are displaced along the field axis. The magnitude of this displacement differs for each underlayer. We determined the value of the applied in-plane field that results in a minimum in the DW velocity for each underlayer, which is denoted as $H_{\rm offset}$ in Table~\ref{Tab_Hoffest}. We find that the measured offset fields have the same sign for all underlayers (W, Hf and TaN), where the shift is towards negative values of the in-plane field. This asymmetry can be understood as follows. The in-plane field favors a certain orientation of the magnetization within the domain wall. If an ``up'' bubble domain is nucleated (i.e., $M_z > 0$ within the nucleated bubble), then a positive in-plane field $H_x >0$ favors a right-handed domain wall on the right side of the bubble, while a left-handed domain wall is favored on the left side of the bubble when viewed from above. In the absence of a chiral interaction, both handedness are degenerate and a circular expansion is expected. However, the presence of a DMI breaks this symmetry by favoring one handedness over the other, which means that the wall energies associated with motion parallel and antiparallel to the applied field are no longer equivalent. The minimum in the velocity curve coincides with the applied field that compensates the internal chiral DMI field, which therefore allows its magnitude to be estimated. Similarly, motion perpendicular to the applied field axis remains symmetric with respect to the applied field because the wall energies remain degenerate along this direction. We will revisit this analysis in more detail in Sec. V.
\section{Brillouin light spectroscopy}
In addition to domain wall motion, we have also characterized the DMI using Brillouin light scattering (BLS) measurements on the same multilayer films. In contrast to domain wall creep, where the domain wall energy is the quantity affected by the DMI, the BLS measurements probe propagating spin waves in the ultrathin CoFeB film where the DMI manifests itself as a propagation nonreciprocity for the spin waves, i.e., for a given wavelength, the two spin waves propagating in opposite directions have different frequencies when the static magnetization is in-plane and the wave vector is perpendicular to the static magnetization. For the interfacial form of the DMI present in ultrathin ferromagnets in contact with a heavy metal underlayer, spin wave propagation remains reciprocal in the absence of any in-plane applied fields in the uniform state, far from boundary edges~\cite{Garcia-Sanchez:2014}. However, when the magnetization is tilted away from its equilibrium orientation along the easy axis (perpendicular to the film plane) by an applied field, a nonreciprocity appears for spin wave propagating perpendicular to the magnetization direction. When the magnetization is saturated in the film plane by the applied field, the geometry corresponds to the Damon-Eshbach geometry for in-plane magnetized films in which magnetostatic surface spin waves exhibit nonreciprocal propagation in the direction perpendicular to the magnetization. For our CoFeB layers, we argue that the nonreciprocity is driven primarily by the presence of the DMI rather than dipolar effects~\cite{Moon:2013}.
In our BLS experiment, we applied in-plane magnetic fields that were sufficiently large to saturate the magnetization in the film plane. We employed the backscattering geometry and a 2$\times$3 pass Fabry-Perot interferometer to investigate propagating spin waves in the Damon-Eshbach geometry for different values of the incident wave vector, by measuring the frequency shifts of the inelastically scattered light with respect to the frequency of the incident laser beam with a wavelength of $\lambda$ = 532 nm. For each wave vector, the spectra were obtained after accumulating photons for a few hours in order to determine the scattered line position to an accuracy better than 0.1~GHz. The Stokes (S) and anti-Stokes (AS) frequencies were then determined from Lorentzian fits to the BLS spectra in order to obtain the desired frequency difference, $ \Delta f = f_S - f_{AS} $. In the following, as we refer to the properties of the SW, $f_\mathrm{S}$ denotes the absolute value of the Stokes frequency, and wave vectors along that of the photons are assumed to be positive.
In Fig.~\ref{Fig_BLS}, we present results from our BLS measurements for the four different underlayers studied.
\begin{figure}
\includegraphics[width=8.5cm]{TestfigBLS.pdf}
\caption{(a) BLS spectra measured for the 2 nm thick W underlayer at a light incidence corresponding to $k_{\rm SW}$ = 20.45 $\mu$m$^{-1}$ and an in-plane field of 1T to saturate the magnetization. Lorentzian fit are also superposed to show the nonreciprocal behavior of the spin-wave, in gray the Stokes and anti-Stokes are inverted. The inset shows a schematic illustration of the Damon-Eshbach configuration used. (b)-(e) Linear fit of the frequency difference $\Delta f$ as a function of wave vector for the different underlayers.
\label{Fig_BLS}}
\centering
\end{figure}
We observe a linear variation of the frequency nonreciprocity, $\Delta f$, with wave vector, for all the studied underlayers. The slopes obtained from a linear fit of this variation are summarized in Table~\ref{Tab_Hoffest}. The shifts are similar in magnitude for the two thicknesses of W and for TaN, while the nonreciprocity is approximately half as strong for the CoFeB film on Hf. We note that the nominal thickness for all the CoFeB layers studied is 1 nm, which allows us to exclude contributions from surface anisotropy~\cite{Stashkevich:2015} as the dominant mechanism for the nonreciprocity observed. Indeed, magnetostatic surface spin waves, which appear in thicker ferromagnetic films, can exhibit a frequency nonreciprocity even in the absence of a chiral interaction because they are localized to the film surfaces and probe different surfaces depending on their direction of propagation (e.g., left propagating waves are localized to the top surface while right propagation waves are localized to the bottom surface). In this way, a difference in surface anisotropies at the top and bottom film surfaces can induce a frequency nonreciprocity, but such effects should be negligible in the ultrathin films since $k_{\rm sw} t \ll 1$, where $t$ is the ferromagnetic film thickness. Moreover, the nonreciprocities observed are larger than what might be expected from a difference in surface anisotropies alone~\cite{Stashkevich:2015}.
\section{Discussion}
\begin{table*}
\begin{ruledtabular}
\begin{tabular}{cccccccccc}
Underlayer & $\mu_0 H_{\rm offset}$ (mT) & $D_{\rm creep}$ (mJ/m$^2$) & $\Delta f / k_{\rm SW}$ (MHz $\mu$m) & $D_{\rm BLS}$ (mJ/m$^2$) & $D_c$ (mJ/m$^2$) & $M_s$ (kA/m)~\cite{Natcomtorrejon2014} & $\alpha$\\
\hline
W (2 nm)& $35 \pm 5$&$0.23 \pm 0.03$&$40 \pm 6$&$0.25 \pm 0.04$&$0.10$&$729$&$0.039 \pm 0.003$\\
W (3 nm )& $15 \pm 5$&$0.12 \pm 0.03$&$41 \pm 6$&$0.27 \pm 0.04$&$0.11$&$788$&$0.033 \pm 0.006$\\
Ta$_{48}$N$_{52}$ (1 nm)& $5 \pm 2$&$0.05 \pm 0.02$&$30 \pm 6$&$0.31 \pm 0.06$&$0.26$&$1235$&$0.015 \pm 0.003$\\
Hf (1 nm) & $2 \pm 2$&$0.01 \pm 0.01$&$19 \pm 5$&$0.15 \pm 0.04$&$0.16$&$965$&$0.023 \pm 0.003$
\end{tabular}
\end{ruledtabular}
\caption{\label{Tab_Hoffest} Measured in-plane field value leading to a minimum for the DW velocity for different underlayer composition and the slope of the non reciprocity frequency versus the wave vector measured by BLS. Extracted DMI obtained by the creep method is compared to BLS measurement and the critical DMI ($D_c$) leading to a full rotation of the DW core toward N\'eel configuration. The measured magnetization and the damping parameter are also given.}
\end{table*}
In this section, we discuss and contrast the estimates of the DMI obtained using the two experimental methods employed. We first begin by discussing results from the BLS experiment. As we mentioned above, the interfacial DMI results in a frequency nonreciprocity in PMA materials when the equilibrium magnetization is tilted away from the film normal~\cite{Moon:2013}. Advantage has already been taken of this behavior in different ultrathin film systems, where the amplitude and sign of the DMI constant, $D$ has been deduced from the measured frequency nonreciprocity~\cite{Di:2015, Belmeguenai:2015, Cho:2015}. This nonreciprocity is characterized by the difference in frequency between the Stokes and anti-Stokes peaks, which depends on $D$ through the relationship~\cite{Di:2015}
\begin{equation}
\Delta f = \frac{2 \gamma}{\pi M_s} D k_{\rm sw},
\end{equation}
where $\gamma$ is the gyromagnetic constant and $M_{s}$ is the saturation magnetization. From the slope of the $\Delta f(k_{\rm SW})$ curve, we can therefore have a direct estimate of $D$ by assuming the bulk value of $\gamma$ and by using the values of $M_{s}$ obtained elsewhere~\cite{Natcomtorrejon2014}. The estimated values of $D$ from BLS measurements, $D_{\rm BLS}$, are reported in Table~\ref{Tab_Hoffest}. For all samples studied, we found that the Stokes frequencies are larger than the anti-Stokes frequencies, which indicates a positive DMI constant $D >0$ that favors a right-handed chirality~\cite{Belmeguenai:2015}.
Let us now turn our attention to our results of domain wall displacement in the creep regime. In PMA films, the wall motion at low fields in a disordered material is described by the creep model, in which the wall moves by series of thermally-activated correlated jump between successive pinning centers~\cite{Lemerle:1998, Chauve:2000}. The wall motion in this regime results from a competition between the disorder energy and the elastic energy associated with the domain wall, where the former favors a roughening of the wall by adapting to spatial profile of defects in the sample, while the latter favors a straight wall by minimizing the total length of the domain wall. In the present case, the wall energy is further modified by two additional competing energies: the Dzyaloshinskii-Moriya interaction, which favors a particular chirality of the domain wall, and the Zeeman energy associated with the in-plane applied-field, which can favor a different chirality depending on its orientation. To account for these additional terms, it has been proposed that the creep model can be modified by including a field-dependent wall energy $\sigma(H_{x})$~\cite{Je:2013},
\begin{equation}
v_x(H_z, H_x) = v_0 \exp \left[ -\left( \frac{\alpha \left[ \sigma(H_x) \right]}{H_z} \right)^{1/4} \right].
\end{equation}
Here, the domain wall energy plays the role of an elastic energy and it can be estimated from the one-dimensional domain wall model~\cite{Lemerle:1998, Chauve:2000}, which can be modified to include the Zeeman and DMI terms~\cite{Je:2013}. This model predicts that any applied in-plane field will only modify the velocity through changes in the wall energy, $\sigma(H_{x})$.
In the one-dimensional picture of domain wall motion, the DMI influences the domain wall motion as an effective magnetic field,
\begin{equation}
H_{\rm DMI}=\frac{D}{\mu_0 M_s \lambda},
\label{eq:DMI}
\end{equation}
where $\lambda$ is the domain wall width parameter $\sqrt{A/K_{\rm eff}}$, $A$ being the exchange constant, and $K_{\rm eff}$ the effective perpendicular anisotropy constant that takes into account the demagnetizing fields. This DMI field appears as an offset for the DW energy when we apply an in-plane field. With a non-zero DMI we expect a chiral magnetic N\'eel wall\cite{Thiaville:2012}. Since the domain wall velocity is directly linked to the domain wall energy $\sigma$ in the creep regime, one may expect that $H_{\rm DMI}$ also plays a role as an offset for the velocity curves shown in Fig.~\ref{Fig_Hip}. We have verified using micromagnetic simulations~\cite{Vansteenkiste:2014} that the maximum in the domain wall energy is indeed obtained for the DMI effective field (at least within a one-dimensional approximation for the domain wall). We have also used these simulations to determine the critical value $D = D_c$ at which the equilibrium domain wall profile is a homochiral N\'eel wall, assuming an exchange stiffness of $A = 22.5$ pJ/m \cite{Burrowes:2013}. This transition is governed by the cost in dipolar energy that must be overcome by the DMI to transform a Bloch-type wall, which possess no volume magnetic charges, to a N\'eel type. The $D_c$ values are given in Table~\ref{Tab_Hoffest} for the different systems as a point of comparison.
By equating the offset fields, $\mu_0 H_{\rm offset}$ (Table~\ref{Tab_Hoffest}), measured from the global minimum of the the $v_x(H_x)$ curves, to the DMI effective field defined in Eq.~\ref{eq:DMI}, we obtain an estimate of the DMI constant $D$ from the domain wall measurements, which are reported in Table~\ref{Tab_Hoffest} as $D_{\rm creep}$. These values are consistent with the behavior of the $v_y(H_x)$ curves (black curves in Fig.~\ref{Fig_Hip}). Since DMI interaction tends to align the magnetization of the DW core in the N\'eel configuration, when if an in-plane field is applied parallel to a DW, favoring a Bloch wall, the magnetization in the DW core will converge faster to a Bloch configuration if the DMI is small. In terms of velocity, the stronger the DMI is, the softer the rounding of the V-shaped curve.
From Table~\ref{Tab_Hoffest}, we can draw the striking conclusion that the agreement between the values of $D$ obtained by domain wall creep and BLS depends strongly on the underlayer. BLS measurements indicate that for all the samples $D_{\rm BLS} \geq D_c > 0$, therefore right-handed N{\'e}el walls are expected for all the different underlayers studied. This correlates well with the two thicknesses of the W underlayer, for which both $v_x(H_x)$ and $v_y(H_x)$ curves can be explained with the simple creep model presented above. For the Hf and TaN underlayers, on the other hand, the agreement between the two methods is very poor, where the minima in the $v_x(H_x)$ curves suggest that domain walls in these samples should be closer to Bloch-type walls. It should be noted that the velocity curves for these two cases exhibit features that are not accounted for in the creep model, such as the local minimum located at $\mu_0 H_x = -55 \pm 10$~mT for the Hf underlayer.
One possible explanation of this discrepancy relates to the domain wall dynamics in the creep regime. This is neglected in usual treatments where focus is on the domain wall and pinning energies, which appear in the exponential factor in Eq.~\ref{eq:creep}. The dynamics is partly captured in the velocity prefactor, $v_0$, which has been shown to exhibit a nontrivial variation as a function of applied in-plane fields~\cite{Lavrijsen:2015}. Another related aspect of the dynamics involves the Gilbert damping of the domain wall. In Table~\ref{Tab_Hoffest}, we have included estimates of the Gilbert damping constant $\alpha$ that have been determined from vector-network-analyzer ferromagnetic resonance measurements on the films studied. Indeed, one can observe a clear correlation between the level of quantitative agreement between the BLS and domain wall creep measurements for $D$, which is very good for the largest value of $\alpha$ (W, 2 nm) and very poor for the smallest value of $\alpha$ (TaN). This trend strongly suggests that dynamical processes in weakly damped systems can give rise to behavior that is not captured by the simple creep model in which the in-plane applied field and DMI only act on the wall energy. A manifestation of such dynamical processes can be seen directly in Fig.~\ref{Fig_Kerr}, where the weakly damped systems (TaN and Hf) exhibit dendritic domain growth that is absent in the more strongly damped systems (W, 2 and 3 nm). We hypothesize that this is related to roughening of domain walls during motion in PMA films with low damping, which has been observed previously in different studies~\cite{Yamada:2011, Burrowes:2013}. We note that the effective damping parameter experienced by the domain wall may differ from the value obtained by ferromagnetic resonance~\cite{Weindler2014}, but the overall trend should persist since the differences due to nonlocal damping are inversely proportional to the domain wall width~\cite{Kim:2015}, which is similar for the different underlayers studied.
Another possible explanation is that Brillouin light scattering and measurements of domain wall creep probe two very different processes in magnetic systems. In BLS, thermally-populated long wavelength spin waves (with wavelengths larger than 270~nm) are probed over areas of 100~$\mu$m, which corresponds to the typical laser spot size used. As such, estimates of the DMI obtained reflect its strength averaged out over these distances. On the other hand, domain wall creep involves probing the competition between the (elastic) wall energy and local pinning potentials, which concern length scales of the order of the domain wall width~\cite{Lemerle:1998}, i.e., about 10-20 nm in our films. Moreover, creep motion necessarily involves thermally-activated jumps or avalanches between pinning sites, and it is possible that the DMI is systematically reduced at these sites by virtue of their existence. In other words, if we suppose that domain wall pinning systematically involves some combination of interface roughness, grain boundaries, and non-magnetic substitutional disorder in the ferromagnetic film --- regions in which the DMI is reduced because the spin-orbit--mediated coupling through the heavy metal underlayer is diminished --- then it follows that creep motion only involves local regions of weak DMI. This idea appears to be corroborated by recent experiments involving steady-state domain wall motion over larger distances (as attained in the flow regime of propagation), which provide estimates of the DMI strength that show better agreement with values obtained by BLS~\cite{Natcomtorrejon2014, Vanatka:2015}. It is also consistent with recent studies using scanning-NV magnetometry, which have revealed that spatial variations in the domain wall structure (i.e., the degree to which a wall is Bloch- or N\'eel-like) can occur~\cite{Gross_NV}. This is likely due to a local variation in magnetic parameters that can lead to the discrepancy discussed here.
Finally, the DMI for the W samples estimated using the BLS and creep measurements quantitatively agrees with the that obtained from current driven DW velocity measurements~\cite{Natcomtorrejon2014}. All measurements return the same sign of DMI for the TaN samples but not for the Hf samples. It should be noted that the Hf thickness, and as a consequence, the structure of Hf\cite{LiuHayashiAPL2015}, is different from samples used in the current induced DW velocity measurements~\cite{Natcomtorrejon2014}. The Hf is thin and predominantly amorphous for the samples studied here whereas it is thicker and forms hcp structure for the samples used to evaluate DMI using the DW velocity measurements\cite{Natcomtorrejon2014}. As a consequence the structure of the Hf may play a role in defining the sign and amplitude of the DMI.
\section{Conclusion}
We have performed a detailed study, by boh DW dynamics under in-plane and out-of-plane magnetic field together with BLS measurement, of the influence of W, TaN or Hf on the induced interfacial DMI in CoFeB/MgO samples. BLS measurements indicate that for all the samples DMI is large enough to favor full N\'eel wall with a right handedness, in contrast to previous findings\cite{Natcomtorrejon2014}. These conclusions are not in quantitative agreement with those inferred from creep domain wall motion measurements. Qualitative agreement is obtained on the sign of the DMI constant, hence on the chirality of the magnetic textures in the samples. With regard to the amplitude of DMI, creep domain wall measurements show that the strength varies with the thickness of the underlayer while being larger than the critical value to have full N\'eel domain walls. In the case of Hf, and TaN samples, with a lower damping than the W samples, the DMI value is much smaller than the values obtained by BLS while velocity dependence on the in-plane magnetic field exhibit asymmetries and local minima that calls for a more robust description of the in plane magnetic fields and DMI in the creep regime. Discrepancies between the two methods show also that taking into account spatial inhomogeneities of the DMI might be a key to explain the experimental observations quantitatively.
\begin{acknowledgments}
The authors wish to thank S. Eimer for the design of the in-plane electromagnet and F. Garcia-Sanchez, A. Thiaville, J.-P. Tetienne, T. Hingant, I. Gross, L. J. Martinez and V. Jacques for fruitful discussions and revision of the manuscript. This work was partially supported by the Agence Nationale de la Recherche (France) under Contract No. ANR-14-CE26-0012 (ULTRASKY). RS acknowledges additional support from the LabEx NanoSaclay.
\end{acknowledgments}
|
\section{Introduction}
In his seminal work, \cite{turing1952} presented an elegant mathematical theory of reaction-diffusion type for pattern formation in developmental biology. He showed that, via a symmetry breaking, a homogeneous state which is linearly stable in the absence of diffusion may be driven unstable in the presence of diffusion to give rise to the emergence of a spatially inhomogeneous pattern. This process is now well known as {\it diffusion-driven instability} or {\it Turing instability}. Since then, reaction-diffusion systems have been proposed and applied to model many phenomena including cancer invasion and angiogenesis in cancer biology \citep{chap,chaplain,gatenby}, pattern formation in developmental biology \citep{hunding,maini}, wound healing in biomedicine \citep{dale,sherratt}, cell motility \citep{mogilner,moek,uduak} and material science \citep{bozzini,krinsky} among many others. Despite their numerous applications, Turing's theory of pattern formation has been widely criticised mainly due to the lack of robustness of the model system to changes in the parameters as well as the lack of experimental evidence of the existence of so-called morphogens with varying diffusivities. Only recently has the existence of chemical morphogens been experimentally validated in hair follicle pattern formation by \cite{sick2006}.
To-date mode selection and parameter identification for reaction-diffusion systems have been mainly carried out on regular planar domains and surfaces where the eigenvalue problem can be analytically solved to yield analytical forms of the wave numbers as well as their corresponding eigenfunctions \citep{ano,madzvamuse2003,uduak}. In this work, we will depart from this framework and extend computationally mode selection and parameter identification to include arbitrary domains and stationary surfaces. First, we will solve the eigenvalue problem numerically using finite elements on planar domains or surface finite elements on smooth surfaces, respectively, to obtain the eigenmodes and their corresponding eigenfunctions. Here, we employ the Krylov-Schur algorithm \citep{stewart} for solving the resulting algebraic system arising from the finite element discretisation. Second, we then pick an eigenmode to which we apply the necessary and sufficient conditions for Turing diffusion-driven instability in order to isolate reaction-kinetic model parameter values within a reaction-diffusion system. This process can be loosely thought of as an inverse problem for model parameter identification. Once the parameter values are isolated, the full reaction-diffusion system is then solved with these isolated parameter values to obtain an inhomogeneous spatially varying solution which is then compared to the numerically computed eigenfunction on the domain or surface. Alternatively, one could pose the following problem to which this methodology will provide insightful information which is otherwise out of reach with the current methodology: {\it Given a biological pattern on a domain or surface and a plausible reaction-diffusion system, what are the model parameter values within this reaction-diffusion system that will give rise to the observed pattern?} This article provides a theoretical and computational framework to answer such a question.
It must be observed that the eigenvalue problem and the reaction-diffusion system are both solved by a similar numerical method, the finite element method in multi-dimensions \citep{johnson}. The finite element method is well known for its capability to deal with complex irregular geometries \citep{barreira,elliot,venkataraman}. Alternative numerical methods such as finite differences \citep{beckett}, spectral methods \citep{chap,ruuth} and finite volume methods among others could be used but with considerable efforts in dealing with geometrical complexities. As mentioned above one interpretation of our approach is that it provides a means of estimating parameter values such that the pattern predicted by linear stability analysis is close to a desired pattern. It must be noted that in many cases the steady state pattern may not be an eigenfunction (or a linear combination of the eigenfunctions) of the Laplacian on the given domain. This is since the nonlinear terms play a role in the resultant steady state pattern \citep{murray2003}. In such a setting our approach may provide parameters which serve as a suitable initial guess for a more advanced parameter identification algorithm \citep{croft,garvie}
The remainder of this article is structured as follows. In Section \ref{sec:model} we introduce the mathematical model which we study in this work. We summarise the necessary and sufficient conditions for Turing {\it diffusion-driven instability} in Section \ref{sec:conditions}. We then detail how mode selection and parameter identification are carried out. In Sections \ref{sec:isolation} and \ref{sec:genisolation} we outline the new theoretical and computational framework for mode selection and parameter identification. The use of the finite element method is described in Section \ref{sec:ferds}. We then give specific examples in 2- and 3-dimensions for regular (by which we mean domains on which analytic expressions for the eigenfunctions are available) as well as general domains and surfaces. We discuss the implications of our framework in the context of current methodologies and conclude that given a biological pattern and a reaction-diffusion system, our approach provides a useful tool for estimating parameter values which may give rise to the observed pattern.
\section{Mathematical model framework}\label{sec:model}
In order to illustrate with clarity the novelty of our approach, we first introduce the standard theoretical framework for reaction-diffusion systems in multi-dimensions \citep{murray2003}. Let $\Omega \subset {\mathbb{R}}^m$ $(m=1,2,3)$ be a simply connected bounded stationary volume for all time $t \in I=[0,t_F]$, $t_F > 0$ and ${\partial \Omega}$ be the surface
boundary enclosing $\Omega$. Also let $ {\bfa u} = \left( u \left( {\bfa x},t \right), v \left( {\bfa x},t \right) \right)^T$ be a vector of two chemical concentrations at position ${\bfa x} \in \Omega \subset {\mathbb{R}}^m$ and time $t\in I$. The evolution equations for reaction-diffusion systems in the absence of cross-diffusion can be obtained from the application of the law of mass conservation and the extended Fick's first law \citep{murray2003,turing1952} to yield the dimensional system
\begin{equation} \label{eq:rden}
\begin{cases}
\begin{cases}
u_t = D_u \Delta u + f(u,v),\\
v_t = D_v\Delta v + g (u,v),
\end{cases}
\quad x \in \Omega, \; t> 0, \\
\\
{\bfa n} \cdot \nabla u= {\bfa n} \cdot \nabla v=0,\; x\,\text{on}\,\partial\Omega, \; t \ge 0, \\ \\
u(x,0)= u_0(x),\; \text{and} \; v(x,0)=v_0(x),\; x\,\text{on}\,\Omega, \; t=0,
\end{cases}
\end{equation}
where $\Delta$ denotes the usual cartesian Laplace operator, $D_u>0$ and $D_v>0$ are diffusion coefficients. Here, {\bfa n} is the unit outward normal to $\partial\Omega$. Initial conditions are prescribed through non-negative bounded functions $u_0 (x)$ and $v_0 (x)$. In the above, $f (u,v)$ and $g (u,v)$ represent nonlinear reactions.
In the case of surfaces, the Laplace operator is replaced by the Laplace-Beltrami operator $\Delta_\Gamma$, where $\Gamma$ is the (smooth) surface. This can be described as follows (For more details we refer the interested reader to see \cite{dziukelliott}). If $f:\Gamma\to\mathbb{R}$ is differentiable at $x\in\Gamma$ we can define the {\it tangential gradient} of $f$ at $x\in\Gamma$ by
\begin{equation}
\nabla_\Gamma f =\nabla\bar{f}-\nabla\bar{f}\cdot{\bfa n}\bn.
\end{equation}
Here $\bar{f}$ is a smooth extension of $f:\Gamma\to\mathbb{R}$ to an $(n+1)$-dimensional neighbourhood $U$ of the surface $\Gamma$, so that $\bar{f}|_\Gamma=f$. $\nabla$ is the gradient in $\mathbb{R}^{n+1}$ and ${\bfa n}$ is the unit normal. The {\it Laplace-Beltrami operator} applied to a twice differentiable function $f\in C^2(\Gamma)$ is given by
\begin{equation}
\Delta_\Gamma f = \nabla_\Gamma\cdot\nabla_\Gamma f.
\end{equation}
It must be observed that if the surface does not have a boundary, no boundary conditions are needed. If the surface has a boundary, we assume homogeneous Neumann boundary conditions.
Since the reaction terms are nonlinear, analytical solutions cannot normally be obtained. Therefore we investigate solution behaviour using linear stability theory and numerical methods. Linear stability analysis is one way of determining the behaviour of a nonlinear system near a given stationary point, normally a uniform steady state, of the given system. The idea is to find under what conditions on the nonlinear reaction kinetics is the uniform steady state linearly asymptotically stable in the absence of diffusion. When diffusion is introduced, the uniform steady state is driven unstable in what is now known as the process of {\it diffusion-driven instability} with the system converging to a spatially inhomogeneous steady state, thereby giving rise to patterning \citep{murray2003,turing1952}. The mathematical treatment of the derivation of the necessary conditions for {\it diffusion-driven} instability requires solving the well known eigenvalue problem, with $W$ a solution of
\begin{subequations}\label{eq:equn}
\begin{gather}
\Delta W+k^2W=0, \quad {\bfa x}\in \Omega,\\
({\bfa n}\cdot\nabla)W=0, \quad {\bfa x} \in \partial \Omega,
\end{gather}
\end{subequations}
where the solution pairs ($k$ (eigenvalues), $W_k({\bfa x})$ (eigenfunctions) obtained either analytically on certain spatial domains or numerically for the general case) of this vector equation can be compared to the spatially inhomogeneous steady state solutions of \eqref{eq:rden}, with good agreement expected near primary bifurcation points.
This approach is generally called mode isolation. The most famous exploration of this problem is the celebrated article "Can one hear the shape of the drum?" by Mark Kac \citeyearpar{kac}. The question being asked is if one knows all the eigenvalues of the eigenvalue problem is it possible to determine the domain? It was later proven by Gordon, Webb and Wolpert \citeyearpar{gww} that the answer is no and they gave examples of distinct regions with identical eigenvalues.
Other work concerned with mode isolation and linear stability theory for reaction-diffusion systems can be found in \cite{chap} and \cite{ano}, here the validation has been mainly restricted to special domains and volumes where the eigenvalue problem can be solved analytically. In this work we will depart from this framework, instead we will compute approximations of the eigenpairs on arbitrary, simply connected domains, volumes and surfaces. We then use these eigenvalues to calculate, by use of the Turing-parameter space restrictions, appropriate model parameter values. This approach can be thought to be analogous to an inverse parameter identification approach whereby, given the eigenvalues and eigenfunctions solving the eigenvalue problem (\ref{eq:equn}), find model parameter values that would give rise to an inhomogeneous spatially varying solution similar to that exhibited by the eigenfunction. To confirm numerical predictions, we use the computed model parameter values to solve the full nonlinear reaction-diffusion systems and compare approximated eigenfunctions on these arbitrary domains, volumes and surfaces to the spatially inhomogeneous solutions obtained numerically.
To proceed, next we show the two-component form which we will work with and state the conditions for {\it diffusion-driven} instability. These will help us to isolate particular modes.
\section{Conditions for diffusion driven instability for reaction-diffusion systems}\label{sec:conditions}
All two component reaction-diffusion systems of the form \eqref{eq:rden} can be non-dimensionalised and scaled to take the form
\begin{subequations}\label{eq:non}
\begin{align}
u_t = \gamma f(u,v)+\Delta u, \quad v_t = \gamma g(u,v)+d\Delta v, \quad {\bfa x}\in \Omega\subset \mathbb{R}^n, \; t\in [0,\infty], \label{eq:parta} \\
({\bfa n}\cdot \nabla)\begin{pmatrix}u \\ v\end{pmatrix}=0 \quad {\bfa x} \in \partial\Omega \; t\in [0,\infty],\\
u({\bfa x}, 0), v({\bfa x}, 0) \text{ given},
\end{align}
\end{subequations}
where $u=u({\bfa x},t), v=v({\bfa x},t)$, $d$ is the ratio of diffusion coefficients, $f(u,v)$ and $g(u,v)$ describe the reaction kinetics. For simplicity, we assume that $f$ and $g$ are continuously differentiable, $\gamma$ can be described as the relative strength of the reaction terms or alternatively as the domain size.
We have zero flux boundary conditions (homogeneous Neumann) because we want only internal sources of instability, ie. self-organisation of the system.
A uniform steady state $(u_s,v_s)$ is a fixed point where $(u,v)=(u_s,v_s)$, constant in time and space, satisfies \eqref{eq:non}, i.e. $(u_t,v_t)|_{u=u_s, v=v_s}={\bf 0}$.
We can find the steady state by solving $f(u_s,v_s)=g(u_s,v_s)=0$. \\
The conditions for instability due to diffusion are well known (see, for example \cite{murray2003}). Firstly, in the absence of diffusion, the steady state $(u_s,v_s)$ is linearly stable if and only if the partial derivatives of $f$ and $g$ at $(u_s,v_s)$ satisfy
\begin{equation}
f_u+g_v<0 \text{ and } f_ug_v-f_vg_u>0.
\end{equation}
Linear stability analysis considering small perturbations from the equilibrium ${\bf w}({\bfa x},t)=(\hat{u}({\bfa x},t),\hat{v}({\bfa x},t))$ leads us to the system
\begin{equation}
{\bf w}_t=\gamma\begin{pmatrix}f_u & f_v \\ g_u & g_v\end{pmatrix}{\bf w}+
\begin{pmatrix}1 & 0 \\ 0 & d \end{pmatrix} \Delta{\bf w},
\end{equation}
which can be solved by method of separation of variables to yield
\begin{equation}
{\bf w}({\bfa x},t)=\sum_kc_ke^{\lambda t}W_k({\bfa x}),
\end{equation}
where $W_k({\bfa x})$ solve the eigenvalue problem
\begin{subequations}
\begin{gather}
\Delta W+k^2W=0 \quad {\bfa x}\in \Omega \label{equn},\\
({\bfa n}\cdot\nabla)W=0 \quad {\bfa x}\in \partial \Omega \label{neumann}.
\end{gather}
\end{subequations}
These are modes that will decay with time unless the wavenumber $k^2$ satisfies
\begin{equation}\label{eq:disprel}
c(k^2)=d(k^2)^2-\gamma(df_u+g_v)k^2+\gamma^2(f_ug_v-f_vg_u)<0,
\end{equation}
this means that instability will occur if
\begin{equation}
df_u+g_v>0, \quad\quad (df_u+g_v)^2-4d(f_ug_v-f_vg_u)>0
\end{equation}
and $k^2$ lies in the range $k_-^2<k^2<k_+^2$ where
\begin{equation}\label{eq:kroots}
k^2_{\pm}=\gamma\frac{(df_u+g_v)\pm\sqrt{(df_u+g_v)^2-4d(f_ug_v-f_vg_u)}}{2d}.
\end{equation}
We exploit this range to isolate particular patterns/modes. The unstable modes will correspond to the eigenfunctions of the Laplacian (or Laplace-Beltrami) on the chosen domain or surface with the selected boundary conditions and $k^2$ the associated eigenvalues. The effect of varying $d$ and $\gamma$ on \eqref{eq:disprel} is shown in Figure \ref{fig:ck2}.\\
In summary the necessary conditions for diffusion driven instability are
\begin{subequations}
\begin{align}
f_u+g_v&<0, &f_ug_v-f_vg_u>0, \\
df_u+g_v&>0, &(df_u+g_v)^2-4d(f_ug_v-f_vg_u)>0.
\end{align}
\end{subequations}
Additionally, the sufficient conditions for patterning formation are that one must be able to isolate distinct real wave numbers and that the domain must be large enough \citep{madzvamuse2010,madzvamuse2015,murray2003}.
\subsection{Examples of reaction kinetics}
For illustrative purposes, we consider three classical reaction kinetics as summarised below. The work presented in this article holds true for other similar reaction kinetics capable of generating Turing patterns.
\subsubsection{Schnakenberg or activator-depleted substrate Kinetics}
The Schnakenberg kinetics \citep{schnak} are a condensed version of the well documented Brusselator model describing a series of autocatalytic reactions also known as activator-depleted models \citep{giermein,prig}, and can characterised by
\begin{equation}
A\rightleftharpoons X \quad\quad
B+X\to Y+D \quad\quad
2X+Y\to 3X.
\end{equation}
Using the Law of Mass Action and the non-dimensionalisation of $f$ and $g$, within system \eqref{eq:non}, we obtain that
\begin{equation}\label{eq:schnakkin}
f(u,v)=a-u+u^2v \quad \text{ and }\quad g(u,v)=b-u^2v,
\end{equation}
where $a$ and $b$ are positive parameters.
\subsubsection{Gierer-Meinhart Kinetics}
One of the models proposed by \cite{giermein} describes an system whereby an "activator" activates the production of an "inhibitor" which inhibits the production of the activator. Again the non-dimensionalised form can be obtained
\begin{equation}\label{eq:gm}
f(u,v)=a-bu+\frac{u^2}{v(1+ku^2)}, \quad \text{and} \quad g(u,v)=u^2-v,
\end{equation}
where $a$ and $b$ are positive parameters (representing constant production rate and linear degradation respectively) and $k$ can be thought of as the saturation concentration of $u$.
\subsubsection{Thomas Kinetics}
The Thomas model \citep{thomas} is an immobilized-enzyme substrate-inhibition mechanism which can be written in non-dimensional form as
\begin{equation}\label{eq:tho}
f(u,v)=a-u-\frac{\rho uv}{1+u+Ku^2}, \quad
g(u,v)=\alpha b - \alpha v-\frac{\rho uv}{1+u+Ku^2},
\end{equation}
where $a$, $\rho$, $K$, $\alpha$, $\beta$ are all non-negative parameters. This can be interpreted as in \cite{murray1982} by saying that $u$ and $v$
\begin{itemize}
\item are generated by constant production $a$ and $\alpha b$ respectively,
\item decay linearly proportional to $u$ and $\alpha v$ respectively and
\item are used up in a substrate inhibition manner $\frac{\rho uv}{1+u+Ku^2}$.
\end{itemize}
\section{Overview on mode isolation for reaction-diffusion systems}\label{sec:isolation}
The goal of mode isolation is to choose parameters, in our case ($d,\gamma$), so that a trajectory starting from a small random perturbation from the steady state will evolve into a spatial pattern generated by one that corresponds, or at least is close to, a chosen eigenfunction of the Laplacian on that domain.
Wavenumber isolation of reaction-diffusion systems is described by \cite{ano} in one dimension, squares and triangles. In \cite{uduak} wavenumbers of a visco-elastic model are isolated on the unit disk. We use similar ideas in the present work. The basic steps are as follows.
\begin{enumerate}
\item Determine a subset of eigenpairs of the Laplacian with suitable boundary conditions on the domain. For special domains this can be done analytically but in general must be done numerically.
\item Compute the dispersal relation \eqref{eq:disprel} for the chosen reaction kinetics (this is independent of the geometry) and the range of admissible wave numbers as a function of $d$ and $\gamma$.
\item Compute $d^*$ and $\gamma^*$ such that only one of the eigenvalues (wave numbers) computed in step 1 is in the range.
\item In order to compare with the patterned state, solve the reaction-diffusion system numerically with computed parameter values and compare with the numerically computed eigenfunctions.
\end{enumerate}
It is possible to implement the above procedure simply because if a domain is bounded and the boundary is sufficiently regular, the Neumann Laplacian has a discrete spectrum of infinitely many non-negative eigenvalues with no finite accumulation point
\begin{equation}
0<\lambda_1\leq\lambda_2\leq\cdots,\lambda_n\to\infty
\end{equation}
and this is due to the spectral theorem for compact self-adjoint operators (\citeauthor{benguria}, 2016; \citealp{kreyszig,taylor}).
The aim is to have an algorithm to find the parameter values $d$ and $\gamma$ for a given eigenpair $(k^2,W)$ such that only patterns analogous to $W$ will grow. For this, one needs that the corresponding $k$ is in the range defined in \eqref{eq:kroots}
\begin{equation}\label{eq:LR}
\gamma L=k^2_-<k<k^2_+=\gamma R
\end{equation}
where
\begin{subequations}
\begin{align}
L=\frac{(df_u+g_v)-\sqrt{(df_u+g_v)^2-4d(f_ug_v-f_vg_u)}}{2d},\\ R=\frac{(df_u+g_v)+\sqrt{(df_u+g_v)^2-4d(f_ug_v-f_vg_u)}}{2d},
\end{align}
\end{subequations}
and that no other $k$ is in this range.
\begin{figure}
\centering
\subfloat[$\gamma=15$]{\includegraphics[width=58mm]{Fig1a.jpg}}
\subfloat[d = 10]{\includegraphics[width=58mm]{Fig1b.jpg}}
\caption{Here the dispersal relation \eqref{eq:disprel} is plotted (for Schnakenberg kinetics). For a fixed value of $\gamma$, when $d$ is below the critical value $d_c$, $c(k^2)$ has no roots so no modes can be isolated. As $d$ increases as does the difference between the two roots so there is more chance the value of $k$ we seek will be between $k^2_-$ and $k^2_+$. Similarly, for a fixed value of $d$, increasing $\gamma$ causes both $k^2_-$ and $k^2_+$ to increase (Colour version online)}\label{fig:ck2}
\end{figure}
In other words, the sign of the polynomial $c(k^2)$ for a given $k$ determines if the mode will grow. Figure \ref{fig:ck2} illustrates how the graph of $c(k^2)$ changes as $d$ and $\gamma$ are varied. We define the critical diffusion ratio $d_c$ as the root of
\begin{equation}
d_c^2f^2_u + 2(2f_vg_u - f_ug_v)d_c + g^2_v = 0.
\end{equation}
We find $(k^2,W)$ either analytically or numerically. Then we propose the following algorithm described in pseudo-code:
{\bf Input:} $d=d_c+\epsilon$, $\epsilon\approx d_c/5$, $\gamma>0$, $f, g$ and the $k_{l,n}$ that we wish to be uniquely isolated.
\begin{enumerate}
\item Compute $k^2_-$ and $k^2_+$ from \eqref{eq:LR}.
\item If $k^2_{l,n}<k^2_-$ increase $\gamma$ by 1 (this number is arbitrary but should be small). This moves the curve to higher values of $k$.
\item If $k^2_{l,n}<k^2_+$ decrease $\gamma$ by 1. This moves the curve to lower values of $k$.
\item If there exists another $k_{l,n}^*\not= k_{l,n}$ such that $k^2_-<k_{l,n}^{*2}<k^2_+$ then decrease $\epsilon$ by $d_c/100$. This shifts the curve upwards so the difference between $k^2_-$ and $k^2_+$ is smaller.
\item If $k_{l,n}$ is uniquely isolated END. If not go to 3.
\end{enumerate}
{\bf Output:} The appropriate $d,\gamma$.\\
Note that we cannot have $d<d_c$ (because then $c(k^2)$ would have no roots) nor $\gamma<0$ (because $k^2>0$).
\section{Finite element method for reaction diffusion systems}\label{sec:ferds}
In order to validate that our mode isolation algorithm does indeed isolate the desired unstable mode, we will simulate the reaction-diffusion systems under consideration with the computed parameter values. To do this we employ a finite element method for the space discretisation and an implicit-explicit time-stepping scheme for the temporal approximation \citep{lakkis2013,madzvamuse2006,ruuth}.
In order to compute a finite element approximation, we write the weak formulation of \eqref{eq:non} as follows:
Find $u,v\in L^2(0,T;H^1(\Omega))$ such that for all $\phi\in H^1(\Omega)$ we have
\begin{align}\label{eqn:wf}
\begin{cases} \int_\Omega u_t\phi + \int_\Omega\nabla u \cdot \nabla\phi&=\gamma\int_\Omega f(u,v)\phi,\\
\int_\Omega v_t \phi+d\int_\Omega\nabla v \cdot \nabla\phi&=\gamma\int_\Omega g(u,v)\phi,
\end{cases}\quad {\bfa x}\in\Omega,\;t>0.
\end{align}
In this work we shall assume the well posedness of the weak formulation above. We note that for suitable parameter values existence and uniqueness of a classical solution, and hence a weak solution, to \eqref{eq:non} may be shown for example by the method of invariant regions proposed and analysed by Sm\"{o}ller \citeyearpar{smoller}.
\subsection{Spatial discretisation}
We define the computational domain $\Omega_h$ by requiring that $\Omega_h$ is a polyhedral approximation to $\Omega$. We define $T_h$ to be a triangulation of $\Omega_h$ made up of non-degenerate elements $\kappa_i$, i.e., $T_h=\bigcup_i\{\kappa_i\}$. We define the finite element space $V_h := \{ v_h\in C^0(\Omega):v_h|_\kappa \text{ is linear}\}$. The semidiscrete (space discrete) finite element approximation to (\ref{eqn:wf}) seeks a pair $(U,V)\in V_h^2$ such that
\begin{align}\label{eqn:fe_scheme}
\begin{cases} \int_{\Omega_h} U_t\phi+ \int_{\Omega_h}\nabla U \cdot \nabla\phi = \gamma\int_{\Omega_h} I_h\left[f(U,V)\right]\phi,\\
\int_{\Omega_h} V_t\phi+ d\int_{\Omega_h}\nabla V \cdot \nabla\phi = \gamma\int_{\Omega_h} I_h\left[g(U,V)\right]\phi,
\end{cases} \forall \phi \in V_h,
\end{align}
where we use the Lagrange interpolant of the initial data into $V_h$ as initial conditions for the scheme.
In order to illustrate a concrete example of the scheme, we focus on the reaction-diffusion system with Schnakenberg kinetics (\ref{eq:schnakkin}).
The finite element approximation (\ref{eqn:fe_scheme}) with the Schnakenberg kinetics can be written in matrix-vector form as follows
\begin{subequations}\label{eqn:mv_form_schnak}
\begin{align}
{\bf M} \boldsymbol{\alpha}_t+{\bf A} \boldsymbol{\alpha}=\gamma\left[a{\bf H}-{\bf M} \boldsymbol{\alpha}+ {\bf M} (\boldsymbol{\alpha})^2 \boldsymbol{\beta}\right],\\
{\bf M} \boldsymbol{\beta}_t+d{\bf A} \boldsymbol{\beta}=\gamma\left[b{\bf H}- {\bf M} (\boldsymbol{\alpha})^2 \boldsymbol{\beta}\right],
\end{align}
\end{subequations}
where $\boldsymbol{\alpha}$ and $\boldsymbol{\beta}$ are the coefficient vectors of the finite element functions $U$ and $V$ respectively and
\[
M_{i,j}=\int_{\Omega_h} \phi_i \phi_j, \quad A_{i,j}=\int_{\Omega_h} \nabla \phi_i\cdot\nabla\phi_j \quad \text{and} \quad H_j=\int_{\Omega_h}\phi_j.
\]
\subsection{Temporal discretisation}
For the temporal discretisation we employ an IMEX method \citep{lakkis2013,madzvamuse2006,ruuth} in which the diffusive term is treated implicitly and the reaction terms are treated explicitly, for simplicity we employ a uniform timestep $\tau$. Introducing the shorthand for a time discrete sequence of functions, $f^n=f(t_n)$,
the fully discrete scheme we employ reads, for $n=0,1,\dots$, given $(U^n, V^n)\in V_h^2$ find $(U^{n+1}, V^{n+1})\in V_h^2$
such that, $\forall \phi \in V_h$,
\begin{align}\label{eqn:discrete_fe_scheme}
\begin{cases} \int_{\Omega_h} \frac{1}{\tau}\left(U^{n+1}-U^n\right)\phi+ \int_{\Omega_h}\nabla U^{n+1} \cdot \nabla\phi = \gamma\int_{\Omega_h} I_h\left[f(U^n,V^n)\right]\phi,\\
\int_{\Omega_h} \frac{1}{\tau}\left(V^{n+1}-V^n\right)\phi+ d\int_{\Omega_h}\nabla V^{n+1} \cdot \nabla\phi = \gamma\int_{\Omega_h} I_h\left[g(U^n,V^n)\right]\phi,
\end{cases}
\end{align}
where we use Lagrange interpolant of the initial data into $V_h$ as initial conditions for the scheme.
This leads us to the following matrix vector form
\begin{subequations}\label{eq:matrixsystem}
\begin{align}
\left(\frac{1}{\tau}{\bf M}+{\bf A}\right) \boldsymbol{\alpha}^{m+1}=\gamma\left[a{\bf H}-{\bf M} \boldsymbol{\alpha}^m+{\bf M} (\boldsymbol{\alpha}^m)^2\boldsymbol{\beta}^m \right]+\frac{1}{\tau}{\bf M} \boldsymbol{\alpha}^m,\\
\left(\frac{1}{\tau}{\bf M}+d{\bf A}\right) \boldsymbol{\beta}^{m+1} =\gamma\left[b{\bf H}- {\bf M} (\boldsymbol{\alpha}^m)^2\boldsymbol{\beta}^m\right]+\frac{1}{\tau}{\bf M} \boldsymbol{\beta}^m.
\end{align}
\end{subequations}
Since we are interested in convergence to a spatially inhomogeneous steady state, for the stopping criteria we use the $L_2$ norm of the approximate time derivative of the discrete solution, stopping the computation if this decreases below some tolerance (see Figure \ref{l2si}).
\subsection{Numerical computations}\label{sec:parameters}
We take the parameter values as shown in Table \ref{table:parameters}, for the initial data we use small quasi-random perturbations around the uniform steady state values. The linear system \eqref{eq:matrixsystem} is solved using the conjugate gradient method \citep{dealii,golub,CG}.
\begin{table}[H]
\caption{Parameters for reaction kinetic models and the corresponding uniform steady states. The uniform states for Schnakenberg kinetics were obtained analytically while for the Gierer-Meinhardt and Thomas reaction kinetics these were calculated computationally using the Newton-Raphson method \citep{arfken,ano}.}
\label{table:parameters}
\begin{tabular}{lcccccccc
\hline\noalign{\smallskip}
Model & a & b & k & K & $\alpha$ & $\rho$ & $u_s$ & $v_s$ \\
\noalign{\smallskip}\hline\noalign{\smallskip}
Schnakenberg & 0.9 & 0.1 & & & & & 1 & 0.9 \\
Gierer-Meinhart & 0.1 & 1 & 0.5 & & & & 0.8395 & 0.7047 \\
Thomas & 150 & 100 & & 0.05 & 1.5 & 13 & 37.74 & 25.16 \\
\noalign{\smallskip}\hline
\end{tabular}
\end{table}
\begin{figure}
\centering
\includegraphics[width=115mm]{Fig2.jpg}
\caption{Plot of the $L_2$ norm of the discrete time-derivative over time for the example shown in Figure 8(b). There is an initial decay due to diffusion followed by a growth because of the exponentially growing modes which eventually decays, due to the dominant nonlinear terms (Colour version online)}
\label{l2si}
\end{figure}
\subsection{Convergence to a steady state}
Figure \ref{l2si} plots the $L_2$ norm of the discrete time derivative of $U$ and $V$ against the elapsed time. To begin with the difference is large. This quickly decays due to diffusion then there is a rapid growth, because of the exponentially growing modes. The time derivative eventually starts to decay due to the effects of the nonlinear terms that act to bound the exponentially growing solution thereby giving rise to a spatially inhomogeneous steady state.
\section{Isolating modes on general domains}\label{sec:genisolation}
On arbitrary domains, analytical solutions for the eigenvalue problem are not typically available but approximate eigenpairs can be computed numerically. Numerically approximating these pairs is a significant challenge. In general, as we are only typically interested in a small number of eigenpairs, it is not necessary to find all solution pairs, however for our approach to mode isolation to remain applicable, it is important that we obtain consecutive pairs.\\
As previously stated, the eigenvalue problem we wish to solve is as follows,
\begin{equation}\label{eq:geneig}
\begin{cases}
\Delta W+k^2W=0, \quad {\bfa x}\in\Omega, \\
({\bfa n}\cdot\nabla)W=0, \quad {\bfa x}\in \partial\Omega.
\end{cases}
\end{equation}
To approximate the solution we employ the finite element method for the spatial discretisation outlined in Section \ref{sec:ferds}. We work with the weak formulation of the eigenvalue problem and look for an approximate eigenpairs $(W_h,k_h^2)\in V_h\times\mathbb{R}_+$ (where $V_h$ contains all continuous piecewise linear functions on a given mesh) such that\begin{equation}
\int_\Omega \nabla W_h\cdot\nabla\phi=k^2\int_\Omega W_h\cdot\phi, \qquad \forall \phi \in V_h.
\end{equation}
As in \eqref{eqn:mv_form_schnak} this may be written in matrix-vector form, we want to find $(\boldsymbol{\alpha},k_h^2)\in\mathbb{R}^m\times\mathbb{R}_+$, where $m$ is the dimension of $V_h$ such that
\begin{equation}
{\bf A}\boldsymbol{\alpha} = k^2 {\bf M}\boldsymbol{\alpha}, \label{eq:ep}
\end{equation}
where ${\bf A}$ and ${\bf M}$ are stiffness and mass matrices defined respectively, by
\begin{equation}
A_{i,j}=\int_{\Omega_h} \nabla \phi_i\cdot\nabla\phi_j \quad \text{and} \quad M_{i,j}=\int_{\Omega_h} \phi_i \phi_j.
\end{equation}
This is a generalised eigenvalue problem.
We use the package {\bf deal.II} \citep{dealii} for its approximation using SLEPc and the Krylov-Schur algorithm. For completeness we give a description of the algorithm employed in Appendix \ref{krsch}.
\begin{figure}
\centering
\subfloat[Unit sphere]{\includegraphics[width=28mm]{Fig3a.jpg}\label{fig:sphmesh}}\;
\subfloat[Unit sphere cut to show inside]{\includegraphics[width=28mm]{Fig3b.jpg}\label{fig:sphcut}}\;
\subfloat[Surface of unit sphere]{\includegraphics[width=28mm]{Fig3c.jpg}\label{fig:ssmesh}}\;
\subfloat[Surface of unit sphere cut to show inside]{\includegraphics[width=28mm]{Fig3d.jpg}}\\
\subfloat[Ellipse]{\includegraphics[width=50mm]{Fig3e.jpg}}\;
\subfloat[Dumbell mesh]{\includegraphics[width=30mm]{Fig3f.jpg}\label{fig:dumbbellgrid}}\;
\subfloat[Inner structure of dumbell mesh]{\includegraphics[width=30mm]{Fig3g.jpg}}\\
\subfloat["fish" mesh]{\includegraphics[width=50mm]{Fig3h.jpg}\label{fig:smmesh}}\;
\subfloat["eel" meshes (with and without boundary)]{\includegraphics[trim = 20cm 4cm 20cm 10cm, clip,width=50mm]{Fig3i}}
\caption{Examples of mesh generation for different volumes and surfaces: (a-c) Mesh generation on the unit sphere. (d) The ellipse which is a deformation of a circle mesh. (e-f) The dumbbell is a deformation of the bulk of a sphere. (g) The "fish" shape is a deformation of the surface of a sphere. (h) An "eel" is modelled by a cylinder with an open boundary and additionally as the same cylinder with added rounded ends}\label{fig:meshes}
\end{figure}
\section{Mesh generation}
All the mesh generation is carried out using the {\bf deal.II} library. We use hexahedral meshes for the volumes and quadrilaterals for the ellipse and surfaces. In Figure \ref{fig:meshes} we exhibit different meshes generated by this package on which we will carry out computations. We also consider smooth surfaces; these meshes are generated by creating a triangulation $\Omega_h$ of the bulk of the domain $\Omega$ then the surface triangulation is defined by collecting the faces of the elements of the bulk triangulation that lie on the surface ($\Gamma_h=\Omega_h|_d\Omega$), i.e., the surface mesh is the trace of the volume mesh (in the example of the cylinder with open ends we use only the elements on the curved surface). For this reason the equations are not being approximated on the actual surface but on an approximation of it. For more details on surface mesh generation the reader is referred to \cite{dealii} and the references therein.
\section{Comparisons of eigenfunctions and spatially inhomogeneous steady states}
\subsection{Example 1: Sphere}
We start by considering the unit sphere, a domain for which the eigenvalue problem can be solved analytically.
\subsubsection{Eigenvalues and eigenfunctions of the Laplacian in the bulk of the unit sphere}
In order to solve \eqref{eq:equn} on the sphere, we convert the eigenvalue problem into spherical coordinates. The eigenvalue problem in spherical coordinates is as follows \citep{arfken,morimoto},
\begin{equation*}
\Delta w + k^2 w =\frac{1}{r^2}\pderi{r}\left(r^2\pder{w}{r}\right)+\frac{1}{r^2\sin\theta} \pderi{\theta}\left(\sin\theta\pder{w}{\theta}\right)+\frac{1}{r^2\sin\theta}\spdr{w}{\phi} + k^2 w=0,
\end{equation*}
with homogeneous Neumann boundary conditions.
The solutions of the above eigenvalue problem are well known and are obtained using separation of variables \citep{arfken,morimoto}. Following \cite{arfken} (p. 424-428) we find an infinite number of solutions of the form
\begin{subequations}\label{eq:efsph}
\begin{align*}
& \quad\quad\quad\quad w_{l,n}^m(r,\theta,\phi)=A_{l,n}^mJ_{l+\frac{1}{2}}(j'_{l+\frac{1}{2},n}r)e^{im\phi}P_l^m(\cos\theta),\\
\text{where} &
\begin{cases}
& l,m,n \text{ all integers such that } |m|\leq l \leq n, \\
& A_{l,n}^m \text{ are constants},\\
& J_\alpha(x)=\sum_{j=0}^\infty\frac{(-1)^j}{j!\Gamma(1+j+\alpha)}\left(\frac{x}{2}\right)^{2j+\alpha} \\ & \text{ with } \Gamma(n)=(n-1)! \text{ (i.e. a Bessel function of the first kind)},\\
& P_l^m(x) \text{ are associated Legendre polynomials},\\
& j'_{l+\frac{1}{2},n} \text{ are zeros of the differential of the spherical Bessel function.}
\end{cases}
\end{align*}
\end{subequations}
\begin{figure}
\centering
\subfloat[$w_{1,1}^1$]{\includegraphics[trim = 50mm 50mm 0mm 30mm, clip, width=55mm]{Fig4a.jpg}\label{fig:sphef1}}\;
\subfloat[$w_{2,1}^0$]{\includegraphics[trim = 0mm 50mm 50mm 30mm, clip, width=55mm]{Fig4b.jpg}\label{fig:sphef2}}\\
\subfloat[$w_{3,1}^0$]{\includegraphics[trim = 50mm 50mm 0mm 30mm, clip, width=55mm]{Fig4c.jpg}}\;
\subfloat[$w_{3,1}^{-2}$]{\includegraphics[trim = 0mm 45mm 50mm 30mm, clip, width=55mm]{Fig4d.jpg}}\\
\subfloat[$w_{4,1}^{-3}$]{\includegraphics[trim = 50mm 45mm 0mm 30mm, clip, width=55mm]{Fig4e.jpg}}
\caption{Analytical solutions to the eigenvalue problem on the unit sphere i.e. \eqref{eq:efsph} for selected values of $l$, $m$ and $n$. These are plotted using {\bf deal.II} (Colour version online)}\label{fig:sphereefs}
\end{figure}
We can find the eigenvalues $k_{l,n}^2=(j'_{l+\frac{1}{2},n})^2$ numerically (using the fact that $J'_{l+\frac{1}{2},n}=\frac{l}{k}J_{l+\frac{1}{2}}(k)-J_{l+\frac{3}{2}}(k)$). It follows that for each eigenvalue $\lambda_{l,n}=k_{l,n}^2$ there are $2l+1$ possible eigenfunctions. Figure \ref{fig:sphereefs} shows the eigenfunctions for some selected values of $l$, $m$ and $n$. For example $k_{1,1}=2.08158$ is the first zero of $J_{\frac{3}{2}}(x)$ and corresponds to the eigenfunctions
\[
w_{1,1}^m(r,\theta,\phi)=J_{\frac{3}{2}}(k_{1,1}r)e^{im\phi}P_1^m(\cos\theta), \text{ with } m=-1,0,1.
\]
The spherical Bessel function is given by $J_{\frac{3}{2}}(k_{1,1}r)=\frac{\sin(k_{1,1}r)}{(k_{1,1}r)^2}-\frac{\sin(k_{1,1}r)}{(k_{1,1}r)^2}$.
Meanwhile $Y_1^m=e^{im\phi}P_1^m(\cos\theta)$ are spherical harmonics whose real parts can be written in cartesian coordinates as $Y_1^{-1}=\sqrt{\frac{3}{4\pi}}\cdot\frac{y}{r}$, $Y_1^{0}=\sqrt{\frac{3}{4\pi}}\cdot\frac{z}{r}$ and $Y_1^{1}=\sqrt{\frac{3}{4\pi}}\cdot\frac{x}{r}$. Since the system we are solving is not sensitive to polarity we can consider these to be equivalent. Figure \ref{fig:sphereefs}\subref{fig:sphef1} shows a plot of the eigenfunction
\[
w_{1,1}^1=(\frac{\sin(k_{1,1}r)}{(k_{1,1}r)^2}-\frac{\sin(k_{1,1}r)}{(k_{1,1}r)^2})\cdot\frac{x}{r},
\]
where as usual $r^2=x^2+y^2+z^2$.
The second example, $k_{2,1}=3.34209$ corresponds to the eigenfunctions
\[
w_{2,1}^m(r,\theta,\phi)=J_{\frac{5}{2}}(k_{2,1}r)e^{im\phi}P_2^m(\cos\theta) \text{, with } -l\leq m\leq l.
\]
Choosing $m=0$, converting the above to cartesian coordinates and taking the real part gives
\begin{equation*}
\begin{split}
w_{2,1}^0 & (x,y,z)= \\ & \left(\left(\frac{3}{k_{2,1}^2r^2}-1\right)\frac{\sin(k_{2,1}r)}{k_{2,1}r}-\frac{3\cos(k_{2,1}r)}{k_{2,1}^2r^2}\right)
\left(\frac{1}{4}\sqrt{\frac{5}{\pi}}\cdot\frac{-x^2-y^2+2z^2}{r^2}\right).
\end{split}
\end{equation*}
The function $w_{2,1}^0$ is plotted in Figure \ref{fig:sphereefs}\subref{fig:sphef2}.
\subsubsection{Mode isolation on the sphere}
Using the method described in Section \ref{sec:isolation} and the values given in Table \ref{table:parameters} we can isolate the wavenumbers for the reaction-diffusion system with Schnakenberg kinetics and these are shown in Table \ref{table:schnak}. We can do the same for Thomas and Gierer-Meinhart (Table \ref{table:gmtom}). In these cases the interval $[k_-,k_+]$ is centered on $k_{l,n}$.
\begin{table}[H]
\caption{Given $d$ and $\gamma$ from the first two columns we obtain values for $k_-$ and $k_+$ and this means that particular given wavenumbers are isolated on the sphere for the reaction-diffusion system with Schnakenberg kinetics.}\label{table:schnak}
\begin{center}
\begin{tabular}{| c | c | c | c | c |}
\hline
$d$ & $\gamma$ & $k_-$ & $k_+$ & Wavenumbers excited \\ \hline \hline
10 & 15 & 1.7321 & 2.7386 & $k_{1,1}=2.08158$ \\ \hline
10 & 40 & 2.8284 & 4.4721 & $k_{2,1}=3.34209$ \\ \hline
9 & 60 & 3.9319 & 5.0866 & $k_{0,2}=4.49341$, $k_{3,1}=4.51410$ \\ \hline
8.81 & 85 & 4.8575 & 5.8955 & $k_{4,1}=5.64670$ \\ \hline
\end{tabular}
\end{center}
\end{table}
\begin{table}[H]
\caption{The values of $d$ and $\gamma$ which isolate the given wavenumbers on the sphere for the Gierer-Meinhart and Thomas reaction kinetics.}
\label{table:gmtom}
\begin{center}
\begin{tabular}{| r | r | c |}
\hline
Gierer-Meinhart & Thomas & Wavenumbers excited \\ \hline \hline
d=74 $\gamma$=30 & d=30 $\gamma$=15 & $k_{1,1}$ \\ \hline
d=74 $\gamma$=80 & d=30 $\gamma$=40 & $k_{2,1}$ \\ \hline
d=74 $\gamma$=160 & d=28, $\gamma$=60 & $k_{0,2}$, $k_{3,1}$ \\ \hline
d=72 $\gamma$=200 & d=27.5 $\gamma$=90 & $k_{4,1}$ \\ \hline
\end{tabular}
\end{center}
\end{table}
\subsubsection{Simulations of the reaction-diffusion systems on the unit sphere}\label{sec:sphere}
\begin{figure}
\centering
\subfloat[$\gamma=15$, d=10]{\includegraphics[trim = 10mm 30mm 10mm 30mm, clip,width=55mm]{Fig5a.jpg}}\;\;
\subfloat[$\gamma=40$, d=10]{\includegraphics[trim = 10mm 30mm 10mm 30mm, clip,width=55mm]{Fig5b.jpg}}\\
\subfloat[$\gamma=70$, d=9]{\includegraphics[trim = 10mm 30mm 10mm 30mm, clip,width=55mm]{Fig5c.jpg}}\;\;
\subfloat[$\gamma=85$, d=8.81]{\includegraphics[trim = 10mm 30mm 10mm 30mm, clip,width=55mm]{Fig5d.jpg}}
\caption{Converged solutions of system \eqref{eq:non} with Schnakenberg kinetics \eqref{eq:schnakkin}. These solutions represent the species $u$. The isolated modes are $w_{1,1}^1$, $w_{2,1}^0$, $w_{3,1}^0$ and $w_{4,1}^{-3}$ (Colour version online)}\label{fig:schnakspheres}
\subfloat[GM, $\gamma=80$, d=74]{\includegraphics[trim = 10mm 30mm 10mm 30mm, clip,width=55mm]{Fig6a.jpg}}\;\;
\subfloat[Thomas, $\gamma=40$, d=30]{\includegraphics[trim = 10mm 30mm 10mm 30mm, clip,width=55mm]{Fig6b.jpg}}\\
\subfloat[GM, $\gamma=160$, d=74]{\includegraphics[trim = 10mm 30mm 10mm 30mm, clip,width=55mm]{Fig6c.jpg}}\;\;
\subfloat[Thomas, $\gamma=70$, d=28]{\includegraphics[trim = 10mm 30mm 10mm 30mm, clip,width=55mm]{Fig6d.jpg}}\\
\subfloat[GM, $\gamma=200$, d=72]{\includegraphics[trim = 10mm 30mm 10mm 30mm, clip,width=55mm]{Fig6e.jpg}}\;\;
\subfloat[Thomas, $\gamma=90$, d=27.5]{\includegraphics[trim = 10mm 30mm 10mm 30mm, clip,width=55mm]{Fig6f.jpg}}
\caption{Converged solutions of system \eqref{eq:non} for the species $u$ with Gierer-Meinhart kinetics \eqref{eq:gm} on the left with isolated modes $w_{2,1}^0$, $w_{3,1}^3$ and $w_{4,1}^{-3}$ and Thomas \eqref{eq:tho} on the right with isolated modes $w_{2,1}^0$, $w_{3,1}^{-2}$ and $w_{4,1}^{-3}$ (Colour version online)}\label{fig:gmthomspheres}
\end{figure}
Solving using {\bf deal.II} we use the mesh shown in Figure \ref{fig:meshes}\subref{fig:sphmesh}. The timestep is taken to be $\tau=10^{-3}$. We take the initial conditions to be a small random perturbation from the previously computed homogeneous steady state. So for the reaction-diffusion system with Schnakenberg kinetics, at each point in the grid we set the initial conditions to be:
\begin{equation}
\alpha^0=0.995+0.01\epsilon, \quad \beta^0=0.895+0.01\epsilon,
\end{equation}
where $\epsilon$ is a uniformly distributed random variable between $0$ and $1$. \\
For each eigenvalue there are a number of different eigenfunctions. Computing using the values obtained with mode isolation, the solution converges to either one of the eigenfunctions or a linear combination. These converged solutions are shown in Figures \ref{fig:schnakspheres} and \ref{fig:gmthomspheres}. It is possible to force the solution to converge to an eigenfunction (which it does not appear to with random initial perturbation) by making a suitable choice of initial condition, for example a perturbation of the desired eigenfunction, suitably scaled. Hence, in the case where multiple wave numbers are excited, pattern selection is heavily influenced by the choice of initial conditions which act as the basin of attraction, one of the major criticisms of Turing's theory for pattern formation \citep{BardandLauder1974}.
\begin{figure}
\subfloat[$\lambda_1=3.52$]{\includegraphics[trim = 30mm 0 30mm 0, clip, width=32mm]{Fig7a.jpg}}\;
\subfloat[$\lambda_2=11.74$]{\includegraphics[trim = 30mm 0 30mm 0, clip, width=32mm]{Fig7b.jpg}}\;
\subfloat[$\lambda_3=12.52$]{\includegraphics[trim = 70mm 25mm 0 0, clip, width=39mm]{Fig7c.jpg}}\\
\subfloat[$\lambda_4=21.63$]{\includegraphics[trim = 30mm 0 30mm 0, clip, width=32mm]{Fig7d.jpg}}\;
\subfloat[$\lambda_5=24.51$]{\includegraphics[trim = 30mm 0 30mm 0, clip, width=32mm]{Fig7e.jpg}}\;
\subfloat[$\lambda_6=34.30$]{\includegraphics[trim = 30mm 0 30mm 0, clip, width=32mm]{Fig7f.jpg}}\\
\subfloat[$\lambda_7=41.75$]{\includegraphics[trim = 30mm 0 30mm 0, clip, width=32mm]{Fig7g.jpg}}\;
\subfloat[$\lambda_8=45.88$]{\includegraphics[trim = 30mm 0 30mm 0, clip, width=32mm]{Fig7h.jpg}}\;
\subfloat[$\lambda_9=50.97$]{\includegraphics[trim = 30mm 0 30mm 0, width=34mm]{Fig7i.jpg}}
\caption{Eigenfunctions corresponding to the labelled eigenvalues on an ellipse. These are solutions of \eqref{eq:geneig} approximated using {\bf deal.II} (Colour version online)}\label{fig:ellipseeigen}
\centering
\subfloat[$d$=10, $\gamma$=10]{\includegraphics[trim = 100mm 0 0 0, clip, width=36mm]{Fig8a.jpg}}\;
\subfloat[$d$=8.8, $\gamma$=30]{\includegraphics[trim = 100mm 0 0 0, clip, width=36mm]{Fig8b.jpg}\label{fig:ell30}}\;
\subfloat[$d$=8.8, $\gamma$=44]{\includegraphics[trim = 100mm 0 0 0, clip, width=36mm]{Fig8c.jpg}}\\
\subfloat[$d$=8.8, $\gamma$=57]{\includegraphics[trim = 100mm 0 0 0, clip, width=36mm]{Fig8d.jpg}}\;
\subfloat[$d$=8.7, $\gamma$=77]{\includegraphics[trim = 100mm 0 0 0, clip, width=36mm]{Fig8e.jpg}}\;
\subfloat[$d$=8.7, $\gamma$=95]{\includegraphics[trim = 100mm 0 0 0, clip, width=36mm]{Fig8f.jpg}}\\
\subfloat[$d$=8.63, $\gamma$=115]{\includegraphics[trim = 100mm 0 0 0, clip, width=36mm]{Fig8g.jpg}}\;
\subfloat[$d$=8.61, $\gamma$=135]{\includegraphics[trim = 100mm 0 0 0, clip, width=36mm]{Fig8h.jpg}}\;
\subfloat[$d$=8.61, $\gamma$=150]{\includegraphics[trim = 100mm 0 0 0, clip, width=36mm]{Fig8i.jpg}}
\caption{Converged solutions of system \eqref{eq:non}, with Schnakenberg kinetics \eqref{eq:schnakkin}, on an ellipse for the species $u$, they all match the associated eigenfunctions shown in Figure \ref{fig:ellipseeigen} (Colour version online)}\label{fig:ellipserd}
\end{figure}
\subsection{Example 2: Ellipse}
Eigenmodes on an ellipse have been investigated in various articles \citep{fox,greben,neves,wu}. Finding the solution involves numerically solving the Mathieu and modified Mathieu equations \citep{hbk}. In particular \cite{wu} analytically find the first eigenvalue of ellipses with Dirichlet boundary conditions, of various sizes of ellipse. Using the eigenvalue solver described in Section \ref{sec:genisolation}, with Dirichlet boundary conditions, we can reproduce their results (results not reported in the interests of brevity). In the following we consider Neumann conditions and choose the semimajor axis to be twice the semiminor axis. The eigenvalues and eigenfunctions are shown in Figure \ref{fig:ellipseeigen}. Figure \ref{fig:ellipserd} shows the converged solutions of the reaction diffusion system when the chosen values of $d$ and $\gamma$ isolate the corresponding wavenumbers $k^2_i=\lambda_i$.
\subsection{Example 3: Dumbbell}
As a third example we consider the dumbbell shaped domain shown in Figure \ref{fig:meshes}\subref{fig:dumbbellgrid}. The solver for the eigenvalue problem on this mesh gives the output of eigenvalues and eigenfunctions shown in Figure \ref{fig:dumeigen}. The corresponding steady state solution with the parameters obtained by mode isolation are shown in Figure \ref{fig:dumconverged}.
\begin{figure
\centering
\subfloat[$\lambda_1=1.49$]{\includegraphics[width=38mm]{Fig9a.jpg}}\;
\subfloat[$\lambda_2=12.68$]{\includegraphics[width=38mm]{Fig9b.jpg}}\;
\subfloat[$\lambda_3=22.86$]{\includegraphics[trim = 80mm 15mm 40mm 0, clip, width=38mm]{Fig9c.jpg}}\\
\subfloat[$\lambda_4=22.98$]{\includegraphics[width=38mm]{Fig9d.jpg}}\;
\subfloat[$\lambda_5=26.52$]{\includegraphics[width=38mm]{Fig9e.jpg}}\;
\subfloat[$\lambda_6=49.91$]{\includegraphics[width=38mm]{Fig9f.jpg}}\;
\caption{Eigenfunctions corresponding to the labelled eigenvalues on the dumbell. These are solutions of \eqref{eq:geneig} approximated using {\bf deal.II} (Colour version online)}\label{fig:dumeigen}
\centering
\subfloat[d=10, $\gamma=5$]{\includegraphics[trim = 90mm 15mm 20mm 0, clip, width=38mm]{Fig10a.jpg}}\;
\subfloat[d=9, $\gamma=40$]{\includegraphics[trim = 90mm 15mm 10mm 0, clip, width=38mm]{Fig10b.jpg}}\;
\subfloat[d=8.8, $\gamma=60$]{\includegraphics[trim = 90mm 15mm 20mm 0, clip, width=38mm]{Fig10c.jpg}}\;
\subfloat[d=8.8, $\gamma=88$]{\includegraphics[trim = 90mm 15mm 10mm 0, clip, width=38mm]{Fig10d.jpg}}\;
\subfloat[d=8.65, $\gamma=130$]{\includegraphics[trim = 90mm 15mm 10mm 0, clip, width=38mm]{Fig10e.jpg}}
\caption{Converged $u$ solutions of system \eqref{eq:non} with Schnakenberg kinetics \eqref{eq:schnakkin} on a dumbell. Eigenvalues $\lambda_1,\lambda_2,\lambda_5,\lambda_6$ have been isolated, however since $\lambda_3\approx\lambda_4$ in (c) we see a linear combination of their eigenfunctions (Colour version online)}\label{fig:dumconverged}
\end{figure}
\subsection{Example 4: Surface of a sphere}
\begin{figure}
\centering
\subfloat[The surface finite element solution with given parameters $d=9$ and $\gamma=35$]{\includegraphics[width=55mm]{Fig11a.jpg}}\;\;\;
\subfloat[Numerically computed eigenfunction corresponding to eigenvalue $\lambda_9=12.0186$]{\includegraphics[width=55mm]{Fig11b.jpg}}
\caption{Mode isolation for the reaction-diffusion system with Schnakenberg kinetics on the surface of the sphere (Colour version online)}\label{fig:ss}
\end{figure}
In all the previous examples we considered bulk, volumetric domains. In this example we have a curved surface as the domain. This means using the Laplace Beltrami operator $\Delta_\Gamma$ instead of the Laplacian $\Delta$ in \eqref{eq:geneig} and \eqref{eq:non}. To approximate solutions in this case, we employ the surface finite element method \citep{barreira,dziuk,dziukelliott,elliot2015,elliot,chung}. \\
The eigenpairs on the surface of the unit sphere can be found analytically and are well known and documented in \cite{chap} for example. The eigenfunctions are referred to as spherical harmonics. They are the restriction of the eigenfunctions \eqref{eq:efsph} to the surface. The eigenvalues are of the form $k^2=l(l+1)$, where $l$ is an integer, and the eigenfunctions are
\begin{equation}
w_l^m(\theta,\phi)=A_l^me^{im\phi}P_l^m(\cos\theta),
\end{equation}
where $m$ and $P_l^m$ are as in Section \ref{sec:sphere}. Therefore we can test the performance of the eigenvalue problem solver with this example. Using the eigenvalue solver on an approximated mesh of the surface of the sphere we obtain the following output of the first 30 eigenvalues computed to 4 decimal places
\begin{table}[H]
\begin{center}
\begin{tabular}{ c c c c c c c c}
$k_h^2$ = & 2.0014, & 2.0014, & 2.0014, & & \\ & 6.00664, & 6.00664, & 6.00671, & 6.0085, & 6.00857, & & \\ & 12.0186, & 12.0224, & 12.0224, & 12.023, & 12.0279, & 12.0284, & 12.0284,\\ & 20.0484, & 20.0484, & 20.0622, & 20.0622, & 20.0717, & 20.0717, & 20.0749, \\ & 30.1043, & 30.1043, & 30.1102, & 30.1102, & 30.1523, & 30.1523, & 30.1591.
\end{tabular}
\end{center}
\end{table}
As expected these are the first 5 values of the form $k^2=l(l+1)$ with $l=1,2,3,4,5$. The values are not exact because the mesh is an approximation of the actual surface of the sphere. The eigenfunctions are analogous to those detailed in Section \ref{sec:sphere} restricted to the boundary. This shows that the eigenvalue solver gives the required output. Since the results are shown in Section \ref{sec:sphere} we only show one example of mode isolation in Figure \ref{fig:ss}.
\subsection{Example 5: "fish" surface}
We now consider a smooth surface on which no analytical expression for the eigenpairs is available, the surface is taken to be diffeomorphic to the sphere and is shown in Figure \ref{fig:meshes}\subref{fig:smmesh}, it is meant to (very loosely) mimic the shape of a fish. We found the first 100 eigenpairs then chose several to isolate. These are shown in Figure \ref{fig:smsols}. Various patterns are observed including stripes, spots and concentric rings.
\begin{figure}
\centering
\subfloat[$d=8.9$, $\gamma=130$]{\includegraphics[trim = 0 35mm 0 50mm, clip,height=36mm]{Fig12a.jpg}}\;
\subfloat[$\lambda_5=40.18$]{\includegraphics[trim = 0 35mm 0 50mm, clip,height=36mm]{Fig12b.jpg}}\\
\subfloat[$d=8.58$, $\gamma=240$]{\includegraphics[trim = 0 35mm 0 50mm, clip,height=36mm]{Fig12c.jpg}}\;
\subfloat[$\lambda_{10}=79.56$]{\includegraphics[trim = 0 35mm 0 50mm, clip,height=36mm]{Fig12d.jpg}}\\
\subfloat[$d=8.58$, $\gamma=400$]{\includegraphics[trim = 0 35mm 0 50mm, clip,height=36mm]{Fig12e.jpg}}\;
\subfloat[$\lambda_{15}=134.73$]{\includegraphics[trim = 0 35mm 0 50mm, clip,height=36mm]{Fig12f.jpg}}\\
\subfloat[$d=8.58$, $\gamma=510$]{\includegraphics[trim = 0 35mm 0 50mm, clip,height=36mm]{Fig12g.jpg}}\;
\subfloat[$\lambda_{19}=175.98$]{\includegraphics[trim = 0 35mm 0 50mm, clip,height=36mm]{Fig12h.jpg}}
\caption{Surface finite element solutions corresponding to the $u$ species of the reaction-diffusion system with Schnakenberg kinetics with the given parameters on the left and numerically computed eigenfunctions corresponding to the given eigenvalue on the right (Colour version online)}\label{fig:smsols}
\end{figure}
\begin{figure}
\centering
\subfloat[$\lambda_{4(open)}=54.43$]{\includegraphics[height=24mm]{Fig13a.jpg}}\;
\subfloat[$\lambda_{4(closed)}=44.94$]{\includegraphics[height=24mm]{Fig13b.jpg}}\\
\subfloat[$\lambda_{23(open)}=253.69$]{\includegraphics[height=23mm]{Fig13c.jpg}}\;
\subfloat[$\lambda_{25(closed)}=257.54$]{\includegraphics[height=24mm]{Fig13d.jpg}}\\
\subfloat[$\lambda_{24(open)}=253.73$]{\includegraphics[height=23mm]{Fig13e.jpg}}\hspace{50mm}
\caption{Eigenfunctions of the Laplace-Beltrami operator on the "eel" shape with the corresponding eigenvalue. The left column shows the surface without a boundary and the right has a boundary. Note that, although the eigenfunctions are different, $\lambda_{23}\approx\lambda_{24}$ (Colour version online)}\label{fig:eelefs}
\subfloat[$d=8.8$, $\gamma=140$]{\includegraphics[height=24mm]{Fig14a.jpg}}\;
\subfloat[$d=8.8$, $\gamma=140$]{\includegraphics[height=24mm]{Fig14b.jpg}}\\
\subfloat[$d=8.6$, $\gamma=750$]{\includegraphics[height=24mm]{Fig14c.jpg}\label{fig:o750}}\;
\subfloat[$d=8.6$, $\gamma=750$]{\includegraphics[height=24mm]{Fig14d.jpg}}\\
\caption{Converged solutions corresponding to the $u$ species of the reaction-diffusion system with Schnakenberg kinetics on the surface of an eel. The surfaces on the right have a boundary whereas those on the left do not. We find that using the same parameter values on both surfaces gives very similar results (Colour version online)}\label{fig:eelsols}
\end{figure}
\subsection{Example 6 and 7 "eel" shapes}
When computing on surfaces, one has to consider whether or not the surface has a boundary. In papers modelling fish or eel patterns (see for example \cite{venkataraman}), a surface with a boundary is often used. To investigate whether having a boundary is significant in this example we consider a surface with and without boundary. We see that the eigenvalues and eigenfunctions are very similar and it is possible to isolate similar patterns using the same parameter values.
\section{Conclusion and further challenges}
In this paper we have considered reaction-diffusion systems and have presented a framework for isolating particular spatially inhomogeneous patterns. The method involves finding eigenpairs of the Laplacian and computing parameters such that when the reaction-diffusion system is solved numerically, only patterns analogous to a particular eigenfunction will grow. In previous works the eigenvalue problem is solved analytically whereas in this paper both the eigenvalue problem and the reaction-diffusion system are solved using the finite element method. Advances in numerical software mean that we can find 100 eigenpairs in a few minutes and we have demonstrated that these eigenpairs match analytical results. The approach is shown to work for 3 different examples of nonlinear reaction kinetics and on a variety of domains and surfaces.
In summary, the main observations are:
\begin{itemize}
\item Mode isolation is straightforward for low values of $k^2$ but can become slightly more difficult for higher values of $k^2$. This is due to the approximation of the nonlinear terms and clustering of the eigenvalues of a linear problem.
\item When two or more eigenvalues are clustered close to each other it becomes difficult to isolate them computationally. If two or more eigenvalues are in the permissible range then the inhomogeneous steady state could be a linear combination of the corresponding eigenfunctions.
\item We display an example of two surfaces where pattern formation appears to be robust despite the fact one has a boundary while the other does not. An interesting investigation would be to see if this can be true for other geometries. Note that this is only the case for zero-flux boundary conditions. Imposing Dirichlet or Robin-type boundary conditions would result in substantially different patterns.
\end{itemize}
In this paper we have only considered stationary domains/volumes and surfaces. However the domains of biological processes generally evolve with time \citep{barreira,elliot,lakkis2013,madzvamuse2006,venkataraman}. This adds more complexity to solving the reaction-diffusion systems. An interesting and natural extension of this work would be to introduce domain growth and surface evolution. For this extension, studies on the effects of initial conditions would also be worthwhile.
\section*{Data management}
All the computational data output is included in the present manuscript.
\section*{Acknowledgements}
This work (LM) was supported by an EPSRC Doctoral Training Centre Studentship through the University of Sussex. CV and AM acknowledge support from the Leverhulme Trust
Research Project Grant (RPG-2014-149) and the EPSRC grant (EP/J016780/1). This research was partly undertaken
whilst LM, CV and AM were participants in the Isaac Newton Institute Program,
Coupling Geometric PDEs with Physics for Cell Morphology, Motility and Pattern
Formation. This work (AM) has received funding from the European Union Horizon
2020 research and innovation programme under the Marie Sklodowska-Curie grant
agreement (No 642866). AM was partially supported by a grant from the Simons
Foundation. LM acknowledges the support from the University of Sussex ITS for
computational purposes.
\newpage
\bibliographystyle{plainnat}
|
\section{Introduction: direct products in algebra}
\label{secIntro}
The direct product is arguably the most elementary and ubiquitous construction in algebra.
Because of its uncomplicated relationship with the constituent factors, the `natural' statements of the form
\begin{quotation}
\it
$\A\times\B$ satisfies property $\pP$ if and only if $\A$ and $\B$ both satisfy $\pP$
\end{quotation}
abound, and are often taken as read. For example, in the class of groups, the above statement is true for
$\pP$ being any of the following: finitely generated, finitely presented, residually finite, having decidable word problem,
locally finite, torsion, solvable, nilpotent, \dots
Somewhat curiously, the statement does not hold for the property of being hopfian \cite{corner65}.
In fact, these statements hold in the wider class of monoids (excluding of course the group-specific properties of being nilpotent or solvable). However, if one moves on to the class of semigroups, the landscape begins to change and become more interesting. Specifically, the direct product $\S\times\T$ of two infinite semigroups is:
\begin{itemize}
\item
finitely generated if and only if $\S$ and $\T$ are finitely generated and have no indecomposable elements \cite[Theorem 2.1]{Robertson98};
\item
finitely presented if and only if both $\S$ and $\T$ are finitely presented, have no indecomposable elements, and satisfy an additional condition called stability \cite[Theorem 3.5]{Robertson98};
\item
residually finite if and only if both $\S$ and $\T$ are residually finite, but for apparently non-trivial reasons \cite[Theorem 1]{Gray09}.
\end{itemize}
To add further interest, it remains an open question whether it is possible to algorithmically determine if the direct product of two finitely presented semigroups (given by their finite presentations) is finitely presented.
By way of contrast, the question where one factor is known to be finite is decidable (while still non-trivial), see \cite[Theorem 1.2]{Araujo00}.
The purpose of this article is to raise this problematic to the level of general algebra, and initiate a comparative study across different classes of algebras. We will focus on three well known properties of being
\begin{itemize}
\item
finitely generated,
\item
finitely presented, and
\item
residually finite,
\end{itemize}
and ask under which conditions does the direct product of two algebras $\A$ and $\B$ from a certain class of algebras $\cC$ have one of these properties.
The classes $\cC$ we wish to cover are many, and we will resort to the language and methodology of universal algebra to enable us to state results for several classes at once. Here is the list of classes considered here:
\begin{itemize}
\item
groups;
\item
other `classical' structures: rings, (associative and non-associative) algebras, modules;
\item
group-like, non-associative algebras: loops and quasigroups;
\item
Mal'cev algebras (encompassing all of the above);
\item
semigroups and monoids;
\item
lattices;
\item
idempotent algebras (including lattices);
\item
algebras in congruence modular varieties (including Mal'cev algebras and lattices).
\end{itemize}
The main general preservation results we prove concern finite generation in
congruence permutable or idempotent varieties (Theorems~\ref{fg:Mal'cev} and~\ref{th:Gidempotent}), and residual
finiteness in congruence modular varieties (Theorem~\ref{th:RCM}). Contrasting these are some rather
surprising negative results, such as:
\begin{itemize}
\item
There exist expansions of groups (and expansions of lattices) $\A,\B$ that are finitely generated
without $\A\times\B$ being finitely generated (Remarks~\ref{re:Zx} and~\ref{re:Ns}).
\item
There exist algebras $\A,\B$ with $\B$ not finitely presented such that $\A\times\B$ is finitely presented
(Example~\ref{ex:Gset}).
\item
Direct products of two infinite loops (or two infinite lattices) are never finitely presented
(Theorems~\ref{th:Ploop} and~\ref{th:Lfp}).
\end{itemize}
Due to the great number of different settings that are considered here, we will not attempt a systematic definition of all the objects, concepts and properties. Instead, we will implicitly rely on the reader's prior, if not explicit or systematic, familiarity with many of the elementary notions used, and will confine ourselves to giving an outline of the basic definitions and facts before each individual result necessary to understand its statement and proof.
For a more systematic introduction to universal algebra we refer the reader to \cite{burris81,mckenzie87},
\cite{hungerford80} for classical algebraic structures, \cite{howie95} for semigroups and monoids, \cite{bruck58} for loops and quasigroups, \cite{Gr:LTF} for lattices.
Throughout the paper we will denote algebras (i.e. sets endowed by certain operations) by boldface capital letters $\A$, $\B$, $\C$,\dots; the underlying sets of elements will be denoted by the corresponding plain letters $A$, $B$, $C$,\dots.
A \emph{type} of algebras is a set $\fF$ of function symbols with a non-negative integer (the \emph{arity})
associated to each $f\in\fF$. An algebra $\A$ of type $\fF$ is a pair
$\algop{A}{\{ f^\A \setsuchthat f\in\fF \}}$ where $f^\A$ is an operation on $A$ whose arity equals
the arity of the symbol $f$.
We define \emph{terms} of type $\fF$ with variables $x_1,x_2,\dots$ and their length inductively:
every variable $x_1,x_2,\dots$ is a term of length $1$. For any $k$-ary $f\in\fF$ and terms $s_1,\dots,s_k$ we have
that $f(s_1,\dots,s_k)$ is a term of length $1+|s_1|+\dots+|s_k|$. For a term $t$ and an algebra $\A$ of type $\fF$
let $t^\A$ denote its induced \emph{term operation} on $\A$.
If there is no danger of confusion, we simply write $t$ instead of $t^\A$.
For two algebras $\A$, $\B$ of the same type $\fF$, their direct product $\C=\A\times\B$
has the carrier set $C=A\times B$ and componentwise operations
\[
f^{\A\times\B}((a_1,b_1),\dots,(a_k,b_k))=(f^\A(a_1,\dots,a_k),f^\B(b_1,\dots,b_k)),
\]
where $f\in\fF$ is a $k$-ary operation symbol, $a_1,\dots, a_k\in A$ and $b_1,\dots,b_k\in B$.
Occasionally we will resort to the vertical notation for pairs, writing $\vtt{a}{b}$ for $(a,b)$.
Associated with $\A\times \B$ are the \emph{natural projections} $\pi_A :A\times B\rightarrow A$, $(a,b)\mapsto a$ and $\pi_B :A\times B :\rightarrow B$, $(a,b)\mapsto b$, which are epimorphisms. In particular, $\A$ and $\B$ are homomorphic images of $\A\times\B$.
In addition, often, but not always, $\A$ and $\B$ naturally embed into $\A\times \B$; this is certainly the case with structures with a neutral element (such as groups, monoid, rings, etc.) and idempotent algebras.
We denote \emph{varieties} (i.e., classes of algebras of the same type that are defined by identities)
by calligraphic letters $\vV, \gG$,\dots For example, the variety of groups $\gG$ consists
of all algebras of type $\fF = \{\cdot,^{-1},1\}$ with arities $2,1,0$ defined by the identities
\[ (x\cdot y)\cdot z = x\cdot (y\cdot z),\ x\cdot 1 = x,\ x\cdot x^{-1} = 1. \]
For an algebra $\A$ we denote the set of all congruences on $\A$ by $\Con(\A)$, the trivial congruence (equality)
by $0_A$ and the total congruence by $1_A$.
The congruence on $\A$ generated by a set $R\subseteq A\times A$ is denoted by $\Cg_\A(R)$.
A variety $\vV$ is \emph{congruence permutable} or \emph{Mal'cev} if for every algebra $\A$ in $\vV$ all
congruences permute, i.e.,
\[ \forall\alpha,\beta\in\Con(\A)\colon \alpha\circ\beta = \beta\circ\alpha. \]
Equivalently $\vV$ is congruence permutable if and only if it has a ternary term $m$ such that the identities
\[
m(x,y,y)= x,\ m(y,y,x) = x
\]
hold in $\vV$. Mal'cev varieties contain in particular all varieties that have the operations of a group (multiplication, inversion, identity) or quasigroup (multiplication, left division, right division).
A variety $\vV$ is \emph{congruence modular} if every algebra in $\vV$ has a modular congruence lattice.
Congruence modular varieties contain Mal'cev varieties and all varieties with lattice operations.
\emph{Idempotent algebras} are another general class that will play a prominent role throughout.
We say that an algebra $\A$ of type $\fF$ is \emph{idempotent} if
\[
f^\A(x,x,\dots,x)=x
\]
for all $f\in \fF$ and all $x\in A$. Note that this implies that the above equality in fact holds for all the terms over $\fF$.
One distinguished example of idempotent algebras is provided by lattices, considered as algebras of type $(\meet,\join)$.
\section{Finite generation}
\label{secfg}
In this section we ask under which conditions is the direct product $\A\times \B$ of two algebras finitely generated. We immediately note that $\A$ and $\B$ are homomorphic images of $\A\times \B$, and so we have:
\begin{obs}
\label{obs1}
If $\A\times\B$ is finitely generated, then so are both $\A$ and $\B$.
\end{obs}
So the only question to consider is which conditions ensure that $\A$ and $\B$ finitely generated implies that $\A\times\B$ is finitely generated. The statement certainly holds for groups, and it actually turns out that this can be generalised to Mal'cev varieties.
\begin{thm}
\label{fg:Mal'cev}
Let $\vV$ be a Mal'cev variety of type $\fF$, and let $\A,\B\in\vV$.
Suppose that there exist $a_0\in A$ and $b_0\in B$ such that the sets
$\{ f^\A(a_0,\dots,a_0)\setsuchthat f\in\fF\}$ and
$\{ f^\B(b_0,\dots,b_0)\setsuchthat f\in\fF\}$ are finite.
Then $\A\times\B$ is finitely generated if and only if both $\A$ and $\B$ are finitely generated.
\end{thm}
\begin{proof}
This result was known to A. Geddes (unpublished).
Only the converse direction needs to be proved. Let
$\A,\B$ be respectively generated by finite sets $X,Y$, and
let
\[
X^\prime=X\cup\{ f^\A(a_0,\dots,a_0)\setsuchthat f\in\fF \},\
Y^\prime=Y\cup\{ f^\B(b_0,\dots,b_0)\setsuchthat f\in\fF \},
\]
which are finite sets by the assumptions in the theorem.
We claim that $\A\times\B$ is generated by
the finite set
\[
Z=(X^\prime\times \{b_0\})\cup (\{a_0\}\times Y^\prime)\cup \{(a_0,b_0)\}.
\]
First we prove
\begin{equation} \label{eq:svs}
A\times\{b_0\} \subseteq \langle Z\rangle.
\end{equation}
Let $a\in A$ be arbitrary. Since $\A$ is generated by $X$, supposing $X=\{x_1,\dots,x_m\}$,
we have an $m$-ary term $s$ over $\fF$
such that $s^{\A}(x_1,\dots,x_m) = a$.
Now we can use induction on the length of $s$.
If $s$ is a variable, then $a\in X$ and $(a,b_0)\in Z$ by definition.
So assume that $s = f(t_1,\dots,t_k)$ for some $k$-ary $f\in\fF$ and terms $t_1,\dots,t_k$.
Let $a_i = t_i^\A(x_1,\dots,x_m)$ for $1\leq i\leq k$.
By the induction hypothesis we have $(a_1,b_0),\dots,(a_k,b_0)\in\langle Z\rangle$.
From this and the definition of $Z$ it follows that the pairs
\[
\vtt{f^\A(a_1,\dots,a_k)}{f^\B(b_0,\dots,b_0)},
\vtt{a_0}{f^\B(b_0,\dots,b_0)} ,
\vtt{a_0}{b_0}
\]
all belong to $\langle Z\rangle$ (in fact, the second and third belong to $Z$).
By applying the Mal'cev term $m$ to the three pairs above, we obtain
$(f^\A(a_1,\dots,a_k),b_0)\in\langle Z\rangle$ which proves~\eqref{eq:svs}.
It follows similarly that
\[
\{a_0\}\times B \subseteq\langle Z\rangle.
\]
Now for an arbitrary $(a,b)\in A\times B$ we have that
\[
\vtt{a}{b_0} ,
\vtt{a_0}{b_0} ,
\vtt{a_0}{b}
\]
all belong to $\langle Z\rangle$.
Applying the Mal'cev term $m$ once more yields $(a,b)\in\langle Z\rangle$.
\end{proof}
\begin{rem} \label{re:Zx}
The extra assumption that the sets
$\{ f^\A(a_0,\dots,a_0)\setsuchthat f\in\fF\}$ and
$\{ f^\B(b_0,\dots,b_0)\setsuchthat f\in\fF\}$
be finite is essential.
Consider for example the algebra $\A$
of type $(+,-,c_a\ (a\in A))$, where $\algop{A}{+,-}$ is an abelian group that is not finitely generated, and $c_a$ is a constant symbol representing the element $a$ of $A$.
Then, clearly, $\A$ is finitely generated (by the empty set, say), and we claim that
the direct square $\A\times \A$ is not finitely generated.
Suppose that $\A\times \A$ is generated by a finite set $Z$.
This means that considered just as an abelian group (i.e. without the constants) it would be generated by the set $Z_1=Z\cup \{(a,a)\setsuchthat a\in A\}$.
It is straightforward to show that the subgroup generated by $Z_1$ to
$\{ (a,b)\in A\times A\setsuchthat a-b\in A^\prime\}$, where $A^\prime$ is the subgroup of $A$ generated by the set $\{ x-y\setsuchthat (x,y)\in Z\}$.
Since $Z$ is finite and $\algop{A}{+,-}$ is not finitely generated it follows that $A^\prime\neq A$, and hence $Z$ does not generate the entire $\A\times\A$, a contradiction.
\end{rem}
The extra assumption is sufficiently weak for Theorem \ref{fg:Mal'cev}
to hold in any Malcev varieties with finitely many basic operations, or those in which every member contains an idempotent. In particular:
\begin{cor}
Let $\vV$ be any of the following classes: groups, rings,
modules over a ring, Lie algebras,
loops, or quasigroups, and let $\A,\B\in\vV$.
Then $\A\times\B$ is finitely generated if and only if both $\A$ and $\B$ are finitely generated.
\end{cor}
Another large class for which the if and only if result holds is provided by idempotent algebras.
In fact we will prove a slightly more general result.
Note that the idempotency condition can be interpreted as the requirement for the unary clone of the algebra to consist only of the identity mapping. Recall that the $k$-ary clone $\Clo_k(\A)$ is obtained by taking the set of all
$k$-ary terms over the type of $\A$, and interpreting them as $k$-ary functions $A^k\rightarrow A$;
for details see \cite{mckenzie87}.
\begin{thm}
\label{th:Gidempotent}
Let $\A,\B$ be algebras of the same type.
Assume that all functions in $\Clo_1(\A)$ and $\Clo_1(\B)$ are surjective, and that,
moreover, one of $\Clo_1(\A)$ or $\Clo_1(\B)$ is finite (and hence a group of bijections).
Then $\A\times\B$ is finitely generated if and only if both $\A$ and $\B$ are finitely generated.
In particular, the direct product of two idempotent algebras is finitely generated if and only if both algebras are finitely generated.
\end{thm}
\begin{proof}
Only the converse direction needs to be proved.
Let $X=\{x_1,\dots,x_m\}$ and $Y=\{y_1,\dots,y_n\}$ be finite generating sets for $\A$ and $\B$ respectively.
Without loss of generality assume that $\Clo_1(\A)$ is a finite group, and that $X$ is closed under its action.
We will prove that $Z=X\times Y$ is a generating set for $\A\times\B$.
Let $a\in A$, $b\in B$ be arbitrary.
Since $X$ generates $\A$, there exists an $m$-ary term $p$ such that
\[
a=p^\A (x_1,\dots,x_m).
\]
Denote by $p_1$ the unary term $p(x,\dots,x)$.
By assumption the unary function $p_1^\B (x)$ is surjective, and so there exists $b^\prime\in B$ such that
\[
p_1^\B(b^\prime)=b.
\]
Since $Y$ generates $\B$ there exists an $n$-ary term $q$ such that
\[
b^\prime=q^\B (y_1,\dots,y_n).
\]
Let $q_1$ be the unary term $q(x,\dots,x)$, and let
\[
x_i^\prime = (q_1^\A)^{-1}(x_i)\in X\ \ (i=1,\dots,m).
\]
Now, for each $i=1,\dots,m$ we have
\[
\vtt{x_i^\prime}{y_1},\dots,\vtt{x_i^\prime}{y_n}\in Z,
\]
from which it follows that
\[
q^{\A\times\B} ( \vtt{x_i^\prime}{y_1},\dots,\vtt{x_i^\prime}{y_n} )
=\vtt{q_1^\A(x_i^\prime)}{q^\B(y_1,\dots,y_n)}=\vtt{x_i}{b^\prime}
\]
is in $\langle Z\rangle$.
Now applying $p$ to
\[
\vtt{x_1}{b^\prime},\dots,\vtt{x_m}{b^\prime}\in\langle Z\rangle
\]
we obtain
\[
p^{\A\times\B}(\vtt{x_1}{b^\prime},\dots,\vtt{x_m}{b^\prime})
=\vtt{p^\A(x_1,\dots,x_m)}{p_1^\B(b^\prime)}=\vtt{a}{b}
\in \langle Z\rangle
\]
as required.
\end{proof}
\begin{cor}
\label{corlatfg}
The direct product $\A\times \B$ of two lattices is finitely generated if and only if $\A$ and $\B$ are finitely generated.
\end{cor}
\begin{rem} \label{re:Ns}
The above corollary does not extend to arbitrary
expansions of lattices (algebras with near-unanimity term or
algebras in congruence distributive varieties)
in general. Indeed,
the algebra $\algop{\N}{\max,\min, s}$, where
$s:x\mapsto x+1$ is the successor function, is clearly generated by $1$.
We claim that its square is not finitely generated.
Suppose otherwise, and let $Z$ be a finite generating set.
Noting that if $(x,y)=m((x_1,y_1),(x_2,y_2))$, where $m$ is either of $\max$ or $\min$,
we have $|x-y|\leq \max(|x_1-y_1|,|x_2-y_2|)$,
while if $(x,y)=s(x_1,y_1)$ then $|x-y|=|x_1-y_1|$,
it follows that $\langle Z\rangle$ is contained in
$\{ (x,y)\setsuchthat |x-y|\leq M\}$, where $M=\max\{ |z-u|\setsuchthat (z,u)\in Z\}$,
which is clearly not the entire $\N\times\N$, a contradiction.
This example also shows that Theorem \ref{fg:Mal'cev} does not generalize to congruence modular varieties.
However, Theorem \ref{th:Gidempotent} does apply to lattices with involutions.
\end{rem}
\begin{rem}
The assumption that at least one of the two clones is finite in Theorem \ref{th:Gidempotent} is necessary, even if both consist entirely of bijections:
for instance, $\Z\times\Z$ is not finitely generated when considered as a $\Z$-set.
Likewise, finiteness of unary clones is not sufficient for finite generation:
Consider the free semigroup $\S$ in the variety defined by $x^2=0$ on $3$ generators.
It is well known that $\S$ is infinite; see \cite[Chapter 2]{lothaire97}.
Then $\S\times \S$ is not finitely generated by \cite[Theorem 2.1]{Robertson98}, even though $\Clo_1(\S)$ contains only two mappings, namely the identity and the constant $0$-mapping.
\end{rem}
\section{Finite presentability}
\label{secfp}
To be \emph{finitely presented} means to be isomorphic to a quotient of a finitely generated free algebra by a finitely generated congruence.
Compared with finite generation and residual finiteness, finite presentability presents a subtle issue:
to be finitely presented is \emph{not} an intrinsic property of the algebra, but rather it also makes reference to the class within which the free algebras are taken (typically a variety). For instance, a finitely generated free group is (obviously) finitely presented as a group, is also finitely presented as a monoid, but is not finitely presented as a magma (groupoid, a set with an arbitrary binary operation). For a variety $\vV$ of type $\fF$ we will denote by $ F_\vV(X)$ the \emph{free algebra} over $X$ in $\vV$. It consists of all terms of type $\fF$ over variables $X$ modulo the identities that hold in $\vV$.
\subsection{From products to factors}
Unlike for finite generation, the fact that a direct product is finitely presented
does not always imply that the factors are finitely presented as demonstrated by the following example
the main idea of which is due to Keith Kearnes and \'Agnes Szendrei.
\begin{exa} \label{ex:Gset}
Let $\G$ be the free group on two generators $x,y$.
Then $\G$ is a semidirect product of $A = \langle y\rangle$
and a normal subgroup $B = \langle x^a \setsuchthat a\in A\rangle$.
We let $\G$ act on itself by multiplication on the right to obtain a regular $\G$-set. Since as a group $\G$ is
generated by $x$ and $y$, this regular $\G$-set is term-equivalent to the algebra $\C = \algop{G}{f_x,g_x,f_y,g_y}$ where
for $z\in G$ we have
\[ f_x(z) = zx,\ g_x(z) = zx^{-1},\ f_y(z) = zy,\ g_y(z) = zy^{-1}. \]
The cosets of the subgroups $A$ and $B$ of $\G$ yield congruences
\begin{align*}
\alpha & = \{ (u,v) \in C^2 \setsuchthat Au= Av \}, \\
\beta & = \{ (u,v) \in C^2 \setsuchthat Bu = Bv \}
\end{align*}
on $\C$.
We claim:
\begin{enumerate}
\item \label{it:VC}
The algebra $\C$ is free in the variety $\vV$ of unary algebras of type $\{f_x,g_x,f_y,g_y\}$ defined by the identities
\[
g_xf_x(z)=z,\ f_xg_x(z)=z,\ g_yf_y(z)=z,\ f_yg_y(z)=z.
\]
\item \label{it:CaCb}
$\C$ is isomorphic to $\C/\alpha\times \C/\beta$.
\item \label{it:bnfg}
$\C/\beta$ is not finitely presented.
\end{enumerate}
For~\eqref{it:VC}, clearly $\C$ belongs to $\vV$ and is one-generated.
From the defining identities of $\vV$, it follows that $ F_\vV(z)$ is isomorphic to $\C$.
For~\eqref{it:CaCb} we note that $\alpha\wedge\beta = 0_C$
and $\alpha\circ\beta = \beta\circ\alpha = 1_C$ because $A\cap B=\{1\}$ and
$AB=BA=G$.
Finally for~\eqref{it:bnfg} it suffices to show that $\beta$ is not finitely generated as a congruence of
$\C$. Suppose otherwise that $\beta$ is generated by $(u_1,v_1),\dots,(u_k,v_k)$ for some $k\in\N$,
$u_1,\dots,u_k,v_1,\dots,v_k\in G$. Then it is straightforward that $B$ is generated as a group by $u_1v_1^{-1},\dots,u_kv_k^{-1}$.
But $B$ is a normal subgroup of a free group of infinite index, and so is not finitely generated by \cite[Proposition 1.3.12]{lyndon01}. Hence $\beta$
is not finitely generated, and the direct factor $\C/\beta$ is not finitely presented.
\end{exa}
To prove the implication from $\A\times\B$ being finitely presented to $\A$ and $\B$ being finitely presented it is sufficient to prove that the kernels of natural projections $\pi_A$, $\pi_B$ are finitely generated congruences on $\A\times\B$.
This is the case for a wide range of varieties. For instance:
\begin{thm} \label{th:fgVfp}
Let $\vV$ be a variety in which every direct product of two finitely generated algebras is finitely generated, and let $\A,\B\in\vV$.
If $\A\times \B$ is finitely presented, then both $\A$ and $\B$ are finitely presented.
\end{thm}
\begin{proof}
Assume $\A\times\B$ is finitely presented.
The kernel of the projection $\pi_A$ is $0_A\times 1_B$ which can be regarded as a subalgebra of $\A^2\times \B^2$.
As such it is isomorphic to $\A\times\B^2$. Since $\A$ and $\B$ are finitely generated,
by the assumption on $\vV$ also $\A\times\B^2$ will be finitely generated. Hence $0_A\times 1_B$ is finitely
generated as a congruence. Thus $\A$ and similarly $\B$ are finitely presented.
\end{proof}
In particular Theorem~\ref{th:fgVfp} applies to idempotent and Mal'cev varieties as we saw in Section~\ref{secfg}.
We also recall that finite generation is not preserved by direct products in all congruence modular varieties. However we will prove
the analog of Theorem~\ref{th:fgVfp} in these varieties as well.
For that
we need the following two auxiliary results.
\begin{lem} \label{le:rs}
Let $\A$ and $\B$ be algebras in a variety $\vV$, and let $\rho\in\Con(\A\times\B)$.
\begin{enumerate}
\item\label{it:ts}
The congruence $\rho\vee (0_A\times 1_B)$ is a product congruence, namely $\rho\vee (0_A\times 1_B) = \tau\times 1_B$ for
\[ \tau = \{ (u,v)\in A^2 \setsuchthat (u,b)\equiv_{\rho\vee (0_A\times 1_B)} (v,c) \text{ for all } b,c \in B \}. \]
\item\label{it:rs}
If $\vV$ is congruence modular, then $\rho\wedge (1_A\times 0_B) = \sigma\times 0_B$ for the congruence
\[ \sigma = \{ (u,v)\in A^2 \setsuchthat (u,b)\equiv_\rho (v,b) \text{ for some } b\in B \} \]
on $\A$.
\end{enumerate}
\end{lem}
\begin{proof}
For \eqref{it:ts} we first prove that the equality
\begin{equation} \label{eq:ts}
\rho\vee (0_A\times 1_B) = \tau\times 1_B
\end{equation}
holds. Let $u,v\in A, r,s\in B$. Assume $(u,r)\equiv_{\rho\vee (0_A\times 1_B)} (v,s)$.
Then $(u,b)\equiv_{\rho\vee (0_A\times 1_B)} (v,c)$ for all $b,c\in B$. Hence $(u,v)\in\tau$ and
$((u,r),(v,s))\in\tau\times 1_B$. Conversely, if $((u,r),(v,s))\in\tau\times 1_B$, then
$(u,r)\equiv_{\rho\vee (0_A\times 1_B)} (v,s)$ by the definition of $\tau$. Thus~\eqref{eq:ts} is proved.
It remains to check that $\tau$ is a congruence.
By~\eqref{eq:ts} we have that $\tau$ is the projection of $\rho\vee (0_A\times 1_B)$ onto $A$.
The projection of any congruence is preserved by the operations of $\A$, is reflexive and symmetric.
To see that $\tau$ is transitive, we use the particular form of $\rho\vee (0_A\times 1_B)$.
Let $(u,v)\in\tau$ and $(v,w)\in\tau$. Then $(u,b)\equiv_{\rho\vee (0_A\times 1_B)} (v,c)$ for all
$b,c\in B$ and $(v,d)\equiv_{\rho\vee (0_A\times 1_B)} (w,e)$ for all $d,e\in B$.
So $(u,b)\equiv_{\rho\vee (0_A\times 1_B)} (w,e)$ for all $b,e\in B$ and $(u,w)\in\tau$.
Thus $\tau\in\Con(\A)$ and item~\eqref{it:ts} is proved.
For~\eqref{it:rs} assume that $\vV$ is congruence modular. Let $\mu = \rho\wedge (1_A\times 0_B)$. Then
\begin{align*}
\mu & = \mu \vee [(1_A\times 0_B) \wedge (0_A\times 1_B)], \\
& = (1_A\times 0_B) \wedge [ \mu \vee (0_A\times 1_B)] \text{ by the modular law.}
\end{align*}
By~\eqref{it:ts} we have $\sigma\in\Con(\A)$ such that $\mu \vee (0_A\times 1_B) = \sigma\times 1_B$.
It follows that
\[ \mu = \sigma\times 0_B. \]
So $\sigma$ is the projection of $\mu$ onto $A$, i.e.,
\[ (u,v)\in\sigma \text{ iff } (u,b) \equiv_\mu (v,b) \text{ for some } b\in B. \]
Since $\mu = \rho\wedge (1_A\times 0_B)$, the latter condition is equivalent to $(u,b) \equiv_\rho (v,b)$
for some $b\in B$. Hence $\sigma$ satisfies~\eqref{it:rs}.
\end{proof}
\begin{lem} \label{le:CMfg}
Let $\vV$ be a congruence modular variety of type $\fF$, and let $\A,\B\in\vV$.
Suppose that there exist $a_0\in A$ and $b_0\in B$ such that the sets
$\{ f^\A(a_0,\dots,a_0)\setsuchthat f\in\fF\}$ and
$\{ f^\B(b_0,\dots,b_0)\setsuchthat f\in\fF\}$ are finite.
If $\A$ is finitely generated, then the kernel of the projection of $A\times B$ onto $B$
is a finitely generated congruence of $\A\times\B$.
\end{lem}
\begin{proof}
Let $X = \{x_1,\dots,x_k\}$ be a finite generating set for $\A$, and define
\begin{multline*}
\rho = \Cg_{\A\times\B}\bigl( \bigl\{ ((x,b_0),(a_0,b_0)) \setsuchthat x\in X \bigr\}
\\
\cup \bigl\{ ((f^\A(a_0,\dots,a_0),b_0),(a_0,b_0)) \setsuchthat f\in\fF \bigr\} \bigr).
\end{multline*}
Clearly $\rho\leq 1_A\times 0_B$. By Lemma~\ref{le:rs}~\eqref{it:rs} we have $\rho = \sigma \times 0_B$
for $\sigma\in\Con(\A)$ with
\begin{equation} \label{eq:scg}
\sigma = \Cg_\A\bigl( \bigl\{ (x,a_0) \setsuchthat x\in X \bigr\} \cup \bigl\{ (f^\A(a_0,\dots,a_0),a_0) \setsuchthat f\in\fF \bigr\} \bigr).
\end{equation}
Since $\rho$ is finitely generated, the result will follow once we have shown that $\sigma = 1_A$.
For that we claim
\begin{equation} \label{eq:usa}
a \equiv_\sigma a_0 \text{ for all } a\in A.
\end{equation}
Since $X$ generates $\A$, we have a $k$-ary term $t$ such that $t^\A(x_1,\dots,x_k) = a$.
We prove~\eqref{eq:usa} by induction on the length of $t$. If $t$ is a variable, the statement
is true by~\eqref{eq:scg}. So assume $t=f(s_1,\dots,s_\ell)$ for $f$ an $\ell$-ary operation
in $\fF$ and $k$-ary terms $s_1,\dots,s_\ell$. We obtain
\begin{align*}
a & = f^\A(s^\A_1(x_1,\dots,x_k),\dots,s^\A_\ell(x_1,\dots,x_k)) \span\omit\span\omit \\
& \equiv_\sigma f^\A(a_0,\dots,a_0) &&\text{(by the induction assumption)} \\
& \equiv_\sigma a_0 &&\text{(by~\eqref{eq:scg})}.
\end{align*}
This proves~\eqref{eq:usa}. Thus $\sigma = 1_A$ and $\rho = 1_A\times 0_B$ is finitely generated.
\end{proof}
\begin{thm} \label{th:CMfp}
Let $\vV$ be a congruence modular variety of type $\fF$, and let $\A,\B\in\vV$.
Suppose that there exist $a_0\in A$ and $b_0\in B$ such that the sets
$\{ f^\A(a_0,\dots,a_0)\setsuchthat f\in\fF\}$ and
$\{ f^\B(b_0,\dots,b_0)\setsuchthat f\in\fF\}$ are finite.
If $\A\times\B$ is finitely presented, then both $\A$ and $\B$ are finitely presented.
\end{thm}
\begin{proof}
Assume $\A\times\B$ is finitely presented in a congruence modular variety of finite type, i.e. it is a quotient of a finitely generated free algebra in $\vV$ by a finitely generated congruence.
In particular $\A$ and $\B$ are finitely generated, and hence, by Lemma~\ref{le:CMfg},
are quotients of $\A\times\B$ by finitely generated congruences.
Therefore both $\A$ and $\B$ are finitely presented.
\end{proof}
Of course, as in Section \ref{secfg}, Theorem \ref{th:CMfp} in particular applies to congruence modular varieties of finite type, or those whose members contain idempotents.
We also remark that the analog for Theorems~\ref{th:fgVfp} and~\ref{th:CMfp}
also holds for semigroups; see \cite{Robertson98}.
\subsection{From factors to products}
We first make a general observation.
\begin{prop} \label{prop:Pfree}
For any variety $\vV$ the following are equivalent:
\begin{enumerate
\item \label{it:free}
The direct product of any two finitely generated free algebras in $\vV$ is finitely presented.
\item \label{it:fp}
The direct product of any two finitely presented algebras in $\vV$ is finitely presented.
\end{enumerate}
\end{prop}
\begin{proof}
The implication \eqref{it:fp}$\Rightarrow$\eqref{it:free} is immediate.
For proving the converse, assume~\eqref{it:free}.
Note that this implies
in particular that the direct product of any two finitely generated algebras in $\vV$ is finitely generated.
Let $\A,\B\in\vV$ be finitely presented. Then we have a finite set $X = \{x_1,\dots,x_m\}$ and a finite
set $R$ of pairs of terms over $X$ such that $\A \cong F_\vV(X)/\Cg_{F_\vV(X)}(R)$.
Without loss of generality, we may assume that $R = R^{-1}$ and
$\{(x_i,x_i)\setsuchthat 1\leq i\leq m\} \subseteq R$.
Similarly $\B \cong F_\vV(Y)/\Cg_{F_\vV(Y)}(S)$ for finite $Y$ and $S$.
By assumption $F_\vV(X)\times F_\vV(Y)$ is finitely presented.
So $\A\times\B$ being finitely presented is equivalent to
$\Cg_{F_\vV(X)}(R)\times\Cg_{F_\vV(Y)}(S)$ being finitely generated as a congruence of $F_\vV(X)\times F_\vV(Y)$.
We claim that
\begin{equation} \label{eq:Rx0}
\Cg_{F_\vV(X)}(R)\times 0_{F_\vV(Y)} \text{ is finitely generated.}
\end{equation}
First observe that $\langle R\rangle\times 0_{F_\vV(Y)}\leq F_\vV(X)^2\times F_\vV(Y)^2$
is finitely generated as direct product of finitely generated algebras. Let $R'$ be a finite generating set.
Assume that $(u,v)\in\Cg_{F_\vV(X)}(R)$. By our assumptions $\langle R\rangle$ is reflexive and symmetric on $F_\vV(X)$.
So $\Cg_{F_\vV(X)}(R)$ is the transitive closure of $\langle R\rangle$. We have $w_0,\dots,w_n\in F_\vV(X)$
such that
\begin{equation*}
u = w_0, v = w_n, (w_i,w_{i+1}) \in\langle R\rangle
\end{equation*}
for all $i \in\{0,\dots,n-1\}$.
It then follows that
\[
(\vtt{w_i}{t},\vtt{w_{i+1}}{t})\in
\langle R\rangle\times 0_{F_\vV(Y)}
\]
for any $t\in F_\vV(Y)$, and hence, since $R'$ is a generating set for
$\langle R\rangle\times 0_{F_\vV(Y)}$, we have
\begin{equation*}
(\vtt{w_i}{t},\vtt{w_{i+1}}{t}) \in\langle R'\rangle
\end{equation*}
for all $i \in\{0,\dots,n-1\}$.
Hence $(\vtt{u}{t},\vtt{v}{t})$ is contained in the transitive closure of
$\langle R'\rangle$, that is, in $\Cg_{F_\vV(X)\times F_\vV(Y)}(R')$. Thus $\Cg_{F_\vV(X)}(R)\times 0_{F_\vV(Y)}$ is generated
by $R'$, and~\eqref{eq:Rx0} is proved.
Similarly $0_{F_\vV(X)}\times\Cg_{F_\vV(Y)}(S)$ is finitely generated. Since $\Cg_{F_\vV(X)}(R)\times\Cg_{F_\vV(Y)}(S)$
is the join of $\Cg_{F_\vV(X)}(R)\times 0_{F_\vV(Y)}$ and $0_{F_\vV(X)}\times\Cg_{F_\vV(Y)}(S)$, the result follows.
\end{proof}
As an immediate consequence, we obtain the if and only if statement for all the `classical' algebraic structures:
\begin{cor} \label{cor:class}
Let $\vV$ be any of the following varieties: groups, rings, modules over a ring $R$, algebras over a field,
Lie algebras, or monoids, and let $\A,\B\in\vV$. Then $\A\times\B$ is finitely presented if and only if $\A$ and $\B$ are finitely presented.
\end{cor}
\begin{proof}
Let $X,Y$ be disjoint finite sets. Then $F_\vV(X)\times F_\vV(Y)$ is isomorphic to $ F_\vV(X\cup Y)/\rho$ where
the congruence $\rho$ is generated by the following set $R$.
\begin{center}
\begin{tabular}{|c|c|} \hline
$\vV$ & $R$ \\ \hline
groups, monoids & $\{ (xy,yx) \setsuchthat x\in X, y\in Y \}$ \\
modules & $\emptyset$ \\
rings, algebras & $\{ (xy,0), (yx,0) \setsuchthat x\in X, y\in Y \}$ \\
Lie algebras & $\{ ([x,y],0), ([y,x],0) \setsuchthat x\in X, y\in Y \}$ \\ \hline
\end{tabular}
\end{center}
Since $\rho$ is finitely generated in each case, $F_\vV(X)\times F_\vV(Y)$ is finitely presented.
\end{proof}
In the light of Theorem \ref{fg:Mal'cev} one may wonder whether Corollary~\ref{cor:class}
generalizes to Mal'cev algebras.
This however is not the case, and it fails for loops in a very strong sense.
Recall that the variety $\lL$ of loops has type
$(\cdot,\backslash,/,1)$ (of arities $(2,2,2,0)$) and defining identities
\[
y\backslash (yx)=y(y\backslash x)=(xy)/y=(x/y)y=x1=1x=x;
\]
e.g. see \cite[Section II.1]{burris81}.
\begin{thm} \label{th:Ploop}
In the variety of loops, no infinite direct product $\L\times \M$ with $\L$ and $\M$ non-trivial is
finitely presented.
\end{thm}
\begin{proof}
Suppose that $\L\times\M$ is finitely presented, and, without loss of generality, that $\L$ is infinite.
Both $\L$ and $\M$ must be finitely generated. Choose finite generating sets $X$ and $Y$ for $\L$ and $\M$ respectively such that neither contains the identity element.
Then the set $Z=(X\times\{1\})\cup (\{1\}\times Y)$
is a finite generating set for $\L\times\M$.
We are going to make use of Evans' solution to the word problem for finitely presented loops \cite{Ev:WPAA,Ev:MSD}.
As a consequence of \cite[Theorem 4.1]{Ev:MSD} and its proof there exists
a finite presentation $F_\lL(\overline{Z})/\Cg_{F_\lL(\overline{Z})}(R)$
for $\L\times\M$ with the following properties:
\begin{itemize}
\item
The set $\overline{Z}$ is finite, say $\overline{Z}=\{z_1,\ldots,z_n\}$ and contains $Z$.
(The assumption that the generators in $X$ and $Y$ are not equal to $1$ plays a role in ensuring that this is true.)
\item
All pairs in $R$ have the form $(z_i\circ z_j,z_k)$ where $\circ$ is one of the basic loop operations
$\cdot,\backslash,/$ and $z_i,z_j,z_k\in \overline{Z}$.
\item
Given a term $t(z_1,\dots,z_n)$ in variables $\overline{Z}$, the unique normal form of the element
$t^{\L\times\M}(z_1,\dots,z_m)$ can be obtained by successively, and in arbitrary order,
applying the following transformations: (1) replace $z_i\circ z_j$ by $z_k$ for any $(z_i\circ z_j,z_k)\in R$;
(2) replace $u$ by $v$ where $u=v$ is a defining identity for the variety $\lL$ and $|u|>|v|$.
\end{itemize}
(In Evans' terminology, $R$ is a \emph{closed} set of defining relations.)
Let $\overline{X}\subseteq L$ be such that
\[
\overline{X}\times \{1\}=\overline{Z}\cap (L\times\{1\}),
\]
and without loss suppose
\[
\overline{X}\times \{1\}=\{z_1,\ldots,z_m\}
\]
where $m\leq n$.
From $X\times \{1\}\subseteq Z\subseteq \overline{Z}$ it follows that $X\subseteq \overline{X}$,
and hence $\overline{X}$ is a generating set for $\L$.
\begin{claim}
For every $l\in L$ the normal form of $(l,1)\in L\times M$ has the form
$p(z_1,\dots,z_m)$ for some $m$-ary term $p$, i.e. it involves solely variables from $\overline{X}\times\{1\}$.
\end{claim}
\begin{proof}
Since $\overline{X}$ is a generating set for $\L\cong\L\times\{1\}$, there exists an $m$-ary term $q$ such
that $(l,1)=q^{\L\times\M}(z_1,\dots,z_m)$.
Consider the process of reducing the term $q(z_1,\dots,z_m)$ to its normal form according to the process described above.
If a rule $(z_i\circ z_j,z_k)$ is applied, then necessarily $z_i,z_j\in\overline{X}\times\{1\}$; but then
$z_k\in \overline{X}\times\{1\}$, as otherwise the equality would not hold in $\L\times\M$.
If a loop identity is applied, replacing the longer term by the shorter, no new variables are introduced.
It follows that after the first step we obtain another term over variables $z_1,\dots,z_m$, and an inductive argument completes the proof of the claim.
\end{proof}
Returning to the proof of the theorem, since $\L$ is infinite and $\overline{X}$ is finite,
the set $(L\times{1})\setminus \{z_1,\dots,z_m\}$ is non-empty, and in particular there will exist a normal
form $z_i\circ z_j$ ($1\leq i,j\leq m$), which represents an element $(l,1)$ from this set.
Since $\M$ is non-trivial, the set $Y$ is non-empty; so let $y\in Y$ be arbitrary.
From $\{1\}\times Y\subseteq Z\subseteq\overline{Z}$ it follows that $(1,y)=z_k$ for some $k$ ($m<k\leq n$).
Now consider the terms $(z_i\circ z_j)\cdot z_k$ and $z_k\cdot (z_i\circ z_j)$.
Clearly, no rule from $R$ can be applied to either of those two terms, and neither can a loop identity.
Hence $(z_i\circ z_j)\cdot z_k$ and $z_k\cdot (z_i\circ z_j)$ are distinct normal forms with respect to the presentation $\langle \overline{X} | R\rangle$. On the other hand, these two terms clearly represent the same element $(l,y)$ of $\L\times \M$. This contradicts the fact that $\L\times\M\cong F_\lL(\overline{Z})/\Cg_{F_\lL(\overline{Z})}(R)$, and hence $\L\times\M$ is not finitely presented.
\end{proof}
Further classes in which this direction fails, even though the underlying result for finite generation holds, are
provided by idempotent magmas (or groupoids, where we have a single
binary operation symbol $\cdot$ and the defining identity $x\cdot x = x$) and lattices.
\begin{thm} \label{th:Pmagma}
In the variety $\iI$ of idempotent magmas (groupoids),
the direct product of two free algebras on two generators $F_\iI(x_1,x_2)^2$ is not finitely presented
(even though it is finitely generated).
\end{thm}
\begin{proof}
Consider the homomorphism
\[ h\colon F_\iI(x_{11},x_{12},x_{21},x_{22}) \to F_\iI(x_1,x_2)^2 \]
that is defined by
\[ h(x_{ij}) = (x_i,x_j) \text{ for } i,j\in\{1,2\}. \]
We note that $h$ is onto by the idempotence of $\cdot$ (cf. the proof of
Theorem~\ref{th:Gidempotent}). In particular $F_\iI(x_1,x_2)^2$ is finitely generated.
To prove that $F_\iI(x_1,x_2)^2$ is not finitely presented, it suffices to show that $\ker h$ is not generated
as a congruence by finitely many elements. So seeking a contradiction, we suppose that
there exist $n\in\N$ and $(s_1,t_1),\dots,(s_n,t_n)\in\ker h$ such that
\[ \Cg_{F_\iI(x_1,x_2)^2}( (s_1,t_1),\dots,(s_n,t_n) ) = \ker h. \]
Without loss of generality $s_i\neq t_i$ for all $i\leq n$.
Note that every term over the binary operation $\cdot$ has odd length.
Let $2k-1$ be the maximal length of the terms $s_1,\dots,s_n,t_1,\dots,t_n$.
Observe that for terms $s(x_{11},x_{12},x_{21},x_{22}),t(x_{11},x_{12},x_{21},x_{22})$ we have
\begin{equation} \label{eq:kerh}
(s,t)\in\ker h \text{ iff } s(x_1,x_1,x_2,x_2) = t(x_1,x_1,x_2,x_2), s(x_1,x_2,x_1,x_2) = t(x_1,x_2,x_1,x_2).
\end{equation}
In the absence of parentheses we will associate all products on the left in what follows. Let
\[ u = (x_{11}\underbrace{x_{21}\cdots x_{21}}_{k})(x_{12}\underbrace{x_{22}\cdots x_{22}}_{k}) \]
and consider its congruence class modulo $\ker h$.
We first note that the defining identity of $\iI$ yields a confluent rewriting system where
$tt$ can be replaced by $t$ for any term $t$. Hence we have normal forms for terms in the language
of $\iI$ which are obtained by repeatedly performing such replacements in any order.
Clearly $u$ is already in normal form.
Let $w$ be a term that reduces to $u$ modulo the identities in $\iI$. Then $w = w_1w_2$
for terms $w_1,w_2$ and one of the following holds:
\begin{enumerate}
\item either both $w_1$ and $w_2$ can be reduced to $u$ or
\item $w_1$ reduces to $x_{11}\underbrace{x_{21}\cdots x_{21}}_{k}$ and $w_2$ reduces to
$x_{12}\underbrace{x_{22}\cdots x_{22}}_{k}$.
\end{enumerate}
Consequently every subterm $t$ of $w$ of length at most $2k-1$ reduces to a subterm of $x_{11}x_{21}\cdots x_{21}$
or of $x_{12}x_{22}\cdots x_{22}$. So $t$ is a term either in $\{x_{11},x_{21}\}$ or in $\{x_{12},x_{22}\}$.
Hence $t$ is only congruent to itself modulo $\ker h$ by~\eqref{eq:kerh}. In particular none of the relations
$(s_1,t_1),\dots,(s_n,t_n)$ can be applied to $t$ or to $w$. It follows that $u$ is congruent only to
itself modulo $\ker h$. This contradicts the fact that by~\eqref{eq:kerh} and idempotence
\[ u \equiv (x_{11}x_{12})\underbrace{(x_{21}x_{22})\cdots(x_{21}x_{22})}_k \mod \ker h. \]
Thus $\ker h$ is not finitely generated, and $F_\iI(x_1,x_2)^2$
is not finitely presented.
\end{proof}
\begin{thm} \label{th:Lfp}
In the variety of lattices, no direct product of two infinite lattices is finitely presented.
\end{thm}
\begin{proof}
Seeking a contradiction let $\A,\B$ be infinite lattices such that $\A\times\B$ is finitely presented.
By~\cite[Theorem 1]{GHL:SFPL} (see also~\cite[Theorem 95]{Gr:LTF}) every finitely presented lattice
$\L$ has a congruence $\rho$ such that $\L/\rho$ is finite and every congruence class embeds into some
free lattice. Note that each congruence class is a sublattice because of the idempotence of the operations.
Since all congruences on $\A\times\B$
are product congruences (by \cite[Theorem 149]{Gr:LTF} and \cite[Lemma IV.11.10]{burris81}),
this implies we have $\alpha\in\Con(\A),\beta\in\Con(\B)$ such that $\A/\alpha$ and
$\B/\beta$ are finite and every $\alpha\times\beta$-class in $\A\times\B$ embeds into a free lattice.
Let $a\in A, b\in B$ such that $a/\alpha$ and $b/\beta$ are infinite.
By a result of J\'onsson~\cite[Theorem 2.7]{Jo:SFL} (see also~\cite[Theorem 5.4.5]{Gr:LTF})
every sublattice of finite length (i.e., without infinite chain) in a free lattice is finite.
Since $a/\alpha$ and $b/\beta$ are infinite and embed into a free lattice, each must contain either a copy of the
infinite chain $\Nl = \algop{\N}{\max,\min}$ or its dual $\Nl^\prime=\algop{\N}{\min,\max}$.
In particular, both $a/\alpha$ and $b/\beta$ contain a copy of the 4-element chain
$\mathbf{C}_4=\algop{\{0,1,2,3\}}{\max,\min}$, and
so $\mathbf{C}_4\times\mathbf{C}_4$ embeds into $(a,b)/(\alpha\times\beta)$ which in turn embeds
into some free lattice.
Now every free lattice $\F$ satisfies Whitman's condition~\cite[Theorem 5.44]{Gr:LTF},
namely that for all $x,y,u,v\in F$ we have
\[ x\wedge y \leq u\vee v \Longleftrightarrow x \leq u\vee v \text{ or } y \leq u\vee v \text{ or } x\wedge y \leq u \text{ or } x\wedge y \leq v. \]
But Whitman's condition does not hold in $\mathbf{C}_4\times\mathbf{C}_4$ as witnessed by $x=(1,3),y=(3,1),u=(0,2),v=(2,0)$.
So $(a,b)/(\alpha\times\beta)$ does not embed in any free lattice contradicting our assumption.
Thus $\A\times\B$ is not finitely presented.
\end{proof}
Finally we recall that the situation for semigroups is quite subtle, and is dealt in some detail in \cite{Robertson98,Araujo00}.
\section{Residual finiteness}
\label{secrf}
An algebra $\A$ is \emph{residually finite} if for all distinct $a,b\in A$ there exists a finite algebra $\C$
and a homomorphism $f$ from $\A$ to $\C$ such that $f(a) \neq f(b)$. Equivalently, there exists a congruence
$\rho$ of $\A$ such that $A/\rho$ is finite and $a\not\equiv_\rho b$.
Recall that it was an easy and completely general fact that if $\A\times \B$ is finitely generated, then so are
$\A$ and $\B$.
In the case of residual finiteness it is precisely the other direction that always holds:
\begin{prop}
If $\A$ and $\B$ are residually finite algebras, then so is $\A\times\B$.
\end{prop}
\begin{proof}
For distinct $(a,b),(c,d)\in A\times B$ we must have $a\neq c$ or $b\neq d$. Without loss assume $a\neq c$.
Since $\A$ is residually finite, there exists a homomorphism $f:\A\rightarrow \C$ with $\C$ finite and $f(a)\neq f(c)$. The composition of $f$ with the projection $\pi_\A : \A\times\B \rightarrow \A$ yields a homomorphism $\A\times\B\rightarrow \C$ that separates $(a,b)$ and $(c,d)$.
\end{proof}
In fact, the converse is often just as obvious. For instance, it is clear that residual finiteness is a hereditary
property, i.e. if $\A$ is residually finite and $\B$ is a subalgebra of $\A$, then $\B$ is residually finite as well.
Thus the if and only if statement holds for all classical algebraic structures, monoids, loops, idempotent algebras,
in particular lattices, and many other classes, because factors embed into the direct product.
In \cite{Gray09} the authors prove that the analogous statement holds for semigroups,
albeit for not altogether trivial reasons, and provide examples which show that it does sometimes fail (in particular for unary algebras).
Here we prove the result for congruence modular varieties.
\begin{thm} \label{th:RCM}
Let $\A$ and $\B$ be algebras in a congruence modular variety.
Then $\A\times\B$ is residually finite if and only if both $\A$ and $\B$ are residually finite.
\end{thm}
\begin{proof}
Our argument requires some commutator theory for congruence modular varieties as developed in~\cite{FM:CTC}.
Assume that $\A\times\B$ is residually finite. To show that $\A$ is residually finite, we consider distinct
elements $a_1,a_2\in A$ and fix some $b_1\in B$. Since $\A\times\B$ is residually finite,
we have $\rho_1\in\Con(\A\times\B)$ of finite index such that $(a_1,b_1) \not\equiv_{\rho_1} (a_2,b_1)$.
Note that the product congruences of $\A\times\B$ form a sublattice of $\Con(\A\times\B)$.
In particular, there exists a minimal product congruence $\alpha\times\beta$ above $\rho_1$
and a maximal product congruence $\gamma\times\delta$ below $\rho_1$.
Note that $\alpha\times\beta$ must have finite index, because it contains $\rho_1$ which has finite index; hence both $\alpha$ and $\beta$ are congruences of finite index.
If $\rho_1$ itself is a product of
congruences, then $\alpha=\gamma$ separates $a_1,a_2$ and the result is proved.
So we assume that $\rho_1 < \alpha\times\beta$ is a skew congruence. By factoring $\A\times\B$ by
$\gamma\times\delta$ we may assume that $0_A\times 0_B$ is the only product congruence below $\rho_1$.
We first claim that
\begin{equation} \label{eq:abc}
\beta \text{ is abelian, i.e. } [\beta,\beta] = 0_B.
\end{equation}
From the basic properties of the commutator~\cite[Propositon 3.4]{FM:CTC} we have that
$[\rho_1,0_A\times\beta] \leq\rho_1$ and
$[1_A\times 0_B,0_A\times\beta] = [1_A,0_A]\times [0_B,\beta] = 0_A\times 0_B \leq\rho_1$.
Since the commutator is join distributive by~\cite[Exercise 2.2]{FM:CTC}, this yields
\begin{equation} \label{eq:rv01}
[\rho_1\vee (1_A\times 0_B),0_A\times \beta] \leq\rho_1.
\end{equation}
By Lemma~\ref{le:rs} we have that $\rho_1\vee (1_A\times 0_B)$ is a product congruence that
contains $\rho_1$ and hence contains $\alpha\times\beta$ by minimality. Since the commutator
is monotone~\cite[Proposition 3.4]{FM:CTC},~\eqref{eq:rv01} yields
\[ [\alpha\times\beta,0_A\times\beta] \leq\rho_1. \]
So $[\alpha\times\beta,0_A\times\beta] = 0_A\times [\beta,\beta]$ is a product congruence
that is contained in $\rho_1$ and hence it is $0_A\times 0_B$. Thus~\eqref{eq:abc} is proved.
Recall that $\beta$ has finite index, say $|B/\beta| =: n$, and extend $b_1$ to a complete set of representatives $b_1,\dots,b_n$
for the $\beta$-classes in $B$. For every $i\in\{2,\dots,n\}$ we have $\rho_i\in\Con(\A\times\B)$
of finite index such that $(a_1,b_i) \not\equiv_{\rho_i} (a_2,b_i)$.
Then $\rho = \bigcap_{i=1}^n \rho_i$ has finite index in $\A\times\B$, $\rho < \alpha\times\beta$ and
\begin{equation} \label{eq:abi}
(a_1,b_i) \not\equiv_{\rho} (a_2,b_i) \text{ for all } i\in\{1,\dots,n\}.
\end{equation}
By Lemma~\ref{le:rs}~\eqref{it:rs} we have a congruence
\[ \sigma = \{ (u,v)\in A^2 \setsuchthat (u,b)\equiv_\rho (v,b) \text{ for some } b\in B \} \]
on $\A$.
Let $m = |(A\times B)/\rho|$. We claim that $|A/\sigma| \leq m$. Suppose otherwise that
$u_1,\dots,u_{m+1}\in A$ are pairwise distinct modulo $\sigma$.
Then for any $b\in B$ we have that $(u_1,b)\dots,(u_{m+1},b)$ are pairwise distinct modulo
$\rho$ which contradicts our assumption that $\rho$ has only $m$ classes.
Hence $|A/\sigma|\leq |A\times B/\rho|$.
Seeking a contradiction we now suppose that $a_1\equiv_\sigma a_2$. Then we have
$b\in B$ and some $i\in\{1,\dots,n\}$ such that $b\equiv_\beta b_i$ and
\begin{align*}
(a_1,b_i) & = (a_1,b_i) \\
(a_1,b) & = (a_1,b) \\
(a_1,b) & \equiv_\rho (a_2,b)
\end{align*}
By~\cite[Theorem 5.5]{FM:CTC} each congruence modular variety has a \emph{difference term},
which is a ternary term $d$ which satisfies the following for all algebras $\C$ in the variety:
\[ d^\C(x,x,y) = y \text{ and } d^\C(x,y,y) \equiv x \bmod [\Cg_\C(x,y),\Cg_\C(x,y)] \text{ for all } x,y\in C. \]
Together with~\eqref{eq:abc}, $b\equiv_\beta b_i$ implies $d^\B(b_i,b,b) = b_i$. So applying $d$ to the
three equations above yields
\[ (a_1,b_i) \equiv_\rho (a_2,b_i) \]
which contradicts~\eqref{eq:abi}. Hence $a_1\not\equiv_\sigma a_2$ and $\A$ is residually finite.
\end{proof}
As an immediate consequence, we obtain the if and only if statement for all the `classical' algebraic
structures:
\begin{cor}
Let $\vV$ be any of the following varieties: groups, rings, modules over a ring $R$,
associative algebras or Lie algebras over a field, lattices, quasigroups, or loops, and let $\A,\B\in\vV$.
Then $\A\times\B$ is residually finite if and only if $\A$ and $\B$ are residually finite.
\end{cor}
\section{Conclusion}
Throughout this paper we have seen various situations where finiteness properties are not preserved by
direct products and settings where they are by somehow non-trivial reasons.
We close the paper by giving some questions and suggestions for future research that -- to our mind -- would further enhance our
understanding of these phenomena.
In the variety of semigroups the direct product of free semigroups is not finitely presented in general.
Still certain direct products of infinite, finitely presented semigroups are finitely presented.
This is quite different from our result for loops (Theorem~\ref{th:Ploop}) where no direct product is ever finitely presented except for trivial reasons.
\begin{que}
Is there a congruence modular variety where direct products of free algebras are not finitely presented
but some product of non-trivial algebras is infinite and finitely presented?
\end{que}
For loops, and also idempotent magmas (Theorem~\ref{th:Pmagma}),
it is the non-associativity of operations that seems to be an obstacle for the preservation of finite
presentability in direct products.
\begin{que}
Is finite presentability preserved in some varieties of loops that satisfy some generalized associativity
law, like Bol loops or Moufang loops.
\end{que}
Theorem~\ref{th:Lfp} shows that it is difficult for an infinite direct product of lattices to be finitely presented
without ruling it out altogether.
\begin{que}
Is there an infinite, finitely presented lattice $\L$ and a non-trivial lattice $\M$ (necessarily finite)
such that $\L\times\M$ is finitely presented?
\end{que}
Throughout our paper unary algebras provide examples of unexpected behaviour.
\begin{que}
Classify classes of unary algebras where direct products preserve finite generation, finite presentability,
or residual finiteness.
\end{que}
Congruence modular varieties provided a broad context in which direct products were well-behaved with respect
to residual finiteness and finite presentability (from products to factors).
\begin{que}
Do our results on finite presentability and residual finiteness also hold for generalizations of
congruence modular varieties, e.g. varieties with Taylor term~\cite{HM:SFA}?
\end{que}
Another possible direction for future research is to consider preservation of finiteness conditions under
subdirect products, which play a prominent role in the theory of groups~\cite{BM:SFP,Bridson13} and
more generally Mal'cev algebras.
\section*{Acknowledgements}
The authors are grateful to Erhard Aichinger, Ralph Freese, Keith Kearnes, Michael Kinyon, \'{A}gnes Szendrei and Petr Vojt\v{e}chovsk\'{y} for helpful conversations on the material in this paper.
|
\section{Introduction}
\label{sec:intro}
The computational resources available to scientists and engineers have
never been greater. The ability to conduct simulations and analyses
on thousands of low-latency-connected computer processors has opened
up a world of computational research which was previously
inaccessible. Researchers using these resources rely on
scientific-computing and high-performance-computing techniques; a good
understanding of computational science is no longer optional for
researchers in a variety of fields, ranging from bioinformatics to
astrophysics.
Similarly, the advent of the internet has resulted in a paradigm where
information can be more easily captured, transmitted, stored, and
accessed than ever before. Researchers, both in academia and industry
\cite{Radermacher:2013:GIE:2445196.2445351}, have been actively
developing technologies and approaches for dealing with data of
previously-unimaginable scale. Researchers' ability to analyze data
has never been greater, and many branches of science are actively
using these newly-developed techniques.
Unfortunately, the skills needed to harness these computational and
data-empowered resources are not systematically taught in university
courses \cite{Richards:2011:TRN:2016741.2016801}. Some researchers, postdocs and students may find
non-academic programs to fill this void, but others either do not have
access to these courses or cannot commit the time to follow them.
These researchers typically end up learning by trial and error, or by
self-teaching, which is rarely optimal.
A number of academic programs that aim to address this issue have
emerged at universities across the world (a few examples are
\cite{Tel-Zur:2014:PEB:2690854.2690857,Burkhart:2014:THC:2690854.2690859}).
Some of these grew out of the training efforts of High Performance
Computing (HPC) centres and organizations ({\it
e.g}.\ \cite{Stewart:1995:HUC:224170.224209}). Recognizing the need
for additional skills in their users, computing centres such as those
in the \href{https://www.xsede.org/}{XSEDE} partnership
\cite{XSEDE2014} in the U.S., \href{http://www.prace-ri.eu/}{PRACE} in
Europe, and \href{https://www.computecanada.ca/}{Compute/Calcul
Canada} have been providing local and online HPC training as part of
their user support. Universities have also developed graduate
programs in both Scientific and High-Performance Computing, to train
scientists and engineers in the use of these computational resources.
A more-recent complement to these graduate programs is the development
of the degree in Data Science (DS), that is, degrees which focus on
the analysis of data, especially at scale. These degrees come in a
variety of forms, from multi-year academic graduate programs to
specialized private-sector training. These programs are in strong
demand at present, as large companies have discovered the value in
thoroughly analysing the vast quantities of customer data which they
collect. It is expected that this field will continue to grow, and
academic programs will continue to be introduced to meet this demand.
The SciNet HPC Consortium \cite{1742-6596-256-1-012026,
website:SciNet} is the high-performance-computing consortium of the
University of Toronto. SciNet provides both computational resources
and specialized user support for Canadian academic researchers, and as
members of its support team, we are responsible for training
researchers, postdocs and graduate students at the University of
Toronto in HPC techniques. In this paper we give a review of the
current state of graduate-level Scientific Computing, High-Performance
Computing and Data Science academic programs, and endeavour to design
an ideal HPC and DS graduate program. The paper is organized as
follows. In Sec.~\ref{sec:HPC-research-review} we discuss how
computation has become an essential ingredient in many academic
research endeavours; in Sec.~\ref{sec:HPC-review} we review the
current status of education in the areas of High-Performance and
Scientific Computing. In Sec.~\ref{sec:DS-review} we present the Data
Science education efforts at the academic and non-academic level.
Sec.~\ref{sec:futurePrograms} describes what HPC and DS Master's
programs could look like. We conclude with final remarks and
perspectives for the future in Sec.~\ref{sec:concl}.
\section{The role of HPC in current research}
\label{sec:HPC-research-review}
The breadth of science being as large as it is, it is essentially
impossible to give an overview of the uses of computational methods in
current scientific research. We will nonetheless attempt a review of
at least some computational scientific research, since the way
computers are used in research (and other realms of inquiry)
influences what should be taught to students.
Astrophysical computational research inherently involves large scale
computing, such as the simulation of gravitational systems with many
particles, magnetohydrodynamic systems, and bodies involving general
relativity. Atmospheric physics requires large weather and climate
models with many components to be simulated in a variety of scenarios.
High-energy particle physics projects, such as the ATLAS project at
CERN, require the analysis of many recorded events from large
experiments, while other high-energy physics projects have a need for
large scale simulations ({\it e.g.}\ lattice QCD investigations).
Condensed matter physics, quantum chemistry and material science
projects must often numerically solve quantum mechanical problems in
one approximation or another; the approximations make the calculations
feasible but still rely on large computing resources. Soft condensed
matter and chemical biophysics research often involve molecular
dynamics or Monte Carlo simulations, and frequently require sampling a
large parameter space. Engineering projects can involve optimizing or
analyzing complex airflow or combustion, leading to large fluid
dynamics calculations. Bioinformatics often involves vast quantities
of genomic input data to be compared or assembled, requiring many
small computations. Research in other data-driven sciences such as
social science, humanities, health care and biomedical science
\cite{Ihantola:2015:EDM:2858796.2858798}, is also starting to outgrow
the capacity of individual workstations and standard tools in their
respective fields.
Examining these cases in more detail, one can distinguish different
ways in which research relies on computational resources:
\begin{enumerate}
\item Research that is inherently computational, {\it i.e.}\ it
cannot reasonably be done without a computer, but which requires
relatively minor resources ({\it e.g.}\ a single workstation).
\item Research that investigates problems that do not fit on a
single computer, and therefore rely on multiple computing nodes
attached through a low-latency network.
\item Research that requires many relatively small computations.
\item Research that requires access to a large amount of storage,
but not necessarily a lot of other resources.
\item Research that requires access to a lot of storage, on which
many relatively small calculations are performed.
\end{enumerate}
The distinction between the various types of research determines the
appropriate systems and tools to use. Graduate students that are just
starting their research often do not have enough knowledge to make the
distinction (as nobody has taught them about this), let alone select
and ask for the resources that they will need \cite{Richards:2011:TRN:2016741.2016801}.
Note that all five categories fall under ``Advanced Research
Computing'' (ARC). The categories are not mutually exclusive, but
research of the second and third kind are usually associated with HPC,
while the fourth and fifth, and sometimes the first, are associated
with Data Science (DS). Although there is a lot of overlap between
HPC and DS, these fields require somewhat different techniques. For
that reason, we will consider separate programs for HPC and DS.
\section{Programs in High-Performance Computing}
\label{sec:HPC-review}
Much of the research presented in the previous section falls in the
category \textit{Scientific Computing} (SC). The growth in the
computational approach to research, both academic and industrial, has
prompted some institutions to develop graduate-level programs designed
to teach the skills needed to design, program, debug and run such
calculations. These programs, having been in development for more
than two decades, are now fairly wide spread and mature, and are known
by the names ``Scientific Computing'' or ``Computational Science and
Engineering''. Scientific Computing graduate degrees are offered
internationally in several graduate education hubs around the world
(U.S., England, Germany, Switzerland, {\it etc.}, --- lists of which
can be found at the
\href{https://www.siam.org/students/resources/cse_programs.php}{SIAM}
and \href{http://hpcuniversity.org/students/graduatePrograms}{HPC
University} \cite{Lathrop:2013:HUG:2484762.2484771} websites).
Canada is no exception here either, with at least eight universities
offering graduate-level programs in Computational Science.
These programs include one-year and two-year Master's programs, as
well as Ph.D.\ programs. Most of these programs ({\it e.g.}\ the ones
shown in Tables \ref{table:HPCProgram} and \ref{table:BioProgram})
require a final thesis. The projects and theses are faculty-guided
research projects and are usually one-term long, though, as with all
research, these projects sometimes take longer.
A typical curriculum for a two-year Master's program in Scientific
Computing (in this case from San Diego State University) is presented
in Table~\ref{table:HPCProgram}. It clearly shows that Scientific
Computing has its roots in research in the physical sciences; the
programs heavily emphasize numerical analysis and scientific
modelling. In some ways this is not surprising: computers are very
apt at solving such problems, and the formalism of the physical
sciences often lends itself easily to computer programming. Other
topics of study which are also often encountered in these programs
include finite element analysis, matrix computations, optimization,
stochastic methods, differential equations and stability.
\begin{table}[t]\centering
\begin{tabular}{lcc}\hline
Course Name & Type \\
\hline\hline
Introduction to Computational Science & required\\
Computational Methods for Scientists & required\\
Computational Modelling for Scientists& required\\
Computational Imaging & required\\
Scientific Computing & required\\
Applied Mathematics for Computational Scientists & required\\
Seminar Problems in Computational Science & required\\
Computational and Applied Statistics & elective\\
Computational Database Fundamentals & elective\\
Research & required\\
Thesis & required\\
\hline
\end{tabular}
\caption{The curriculum for the two-year
\href{http://www.csrc.sdsu.edu/masters.html}{Master's program} at
the Computational Science Research Center at San Diego State
University \cite{website:SDSUmasters}; this forms a good example of a
typical Scientific Computing graduate program.
}
\label{table:HPCProgram}
\end{table}
In contrast to Scientific Computing, HPC requires somewhat wider
knowledge; its practitioners need to understand more than just the
theoretical and numerical principles. They require skills such as
serial and parallel programming (often in several languages, and on
different platforms) and scripting, as well familiarity with numerics,
data handling, statistics, and supercomputers and their technical
bottlenecks. In addition, these practitioners are usually not
computer science students, so they must cope without that background.
This is somewhat unavoidable as they need to have sufficient domain
knowledge as well. Much of the same holds for Data Science.
\subsection{Academic HPC Programs}
\label{section:HPCprograms}
There are not many academic programs that focus on HPC. Part of the
reason may be that such programs require access to a
high-performance-computing machine so that students can develop their
skills on real hardware, in a real supercomputing environment. These
machines require multiple computing nodes which are connected by a
low-latency network. Fortunately, such systems do not need to be
local: as long as the machine is accessible through the internet the
machine could be used for teaching. Nonetheless, having the hardware
local to the students lends advantages, since most of the
administrators and analysts of the system are typically available to
assist students with optimizing their codes and developing good
computational strategies. Not surprisingly, the majority of the
currently offered HPC graduate programs seem to have been developed in
conjunction with or by supercomputer centres.
As examples of High Performance Computing programs, the University of
Edinburgh (UK) offers an
\href{https://www.epcc.ed.ac.uk/msc}{MSc in High Performance Computing},
the Universitat Polit\`ecnica de Catalunya/BarcelonaTech (Spain) offers a
\href{http://masters.fib.upc.edu/masters/master-high-performance-computing}{Master in High Performance Computing}
and a
\href{http://masters.fib.upc.edu/masters/master-data-mining-and-business-intelligence}{Master program in Data Mining and Business Intelligence},
SISSA/ICTP in Italy offers a
\href{http://www.mhpc.it/}{Master in High Performance Computing},
while a collaboration between the University ITMO
(Russia) and the University of Amsterdam (Netherlands) offers a
\href{http://en.hpc-magistr.escience.ifmo.ru/magistr/9/}{Double-Degree Master Programs in Applied Mathematics and Informatics (Computational Science)}.
Note that many of these programs emerged
from locations with a very strong tradition and consolidated
background in HPC.
\subsection{SciNet's HPC Programs}
\label{section:nonacadHPCprograms}
Many HPC centers provide training for their users to fill the
computational-skills gap for the wider scientific community,
such as,
\href{http://www.sdsc.edu/}{SDSC}, \href{http://www.psc.edu/}{PSC}, \href{https://www.tacc.utexas.edu}{TACC}, \href{http://www.ncsa.illinois.edu/}{NCSA},
\href{https://www.bsc.es/}{BSC}, \href{https://www.epcc.ed.ac.uk/}{EPCC}, \href{http://www.cscs.ch}{CSCS},
\href{https://www.sharcnet.ca/}{SHARCNET},
\href{http://www.ace-net.ca/}{AceNet},
\href{http://calculquebec.ca/en/}{Calcul Qu\'ebec},
among many others.
In its capacity as an HPC centre based at the University of Toronto,
SciNet has developed several education and training classes
\cite{website:SciNet-Edu} aimed at helping students and users obtain
the skills and knowledge required to get the most out of
advanced-research-computing resources. SciNet's training events and
courses are currently taken by researchers, postdocs, and graduate
students across many different departments and even from outside of
the University of Toronto (UofT). Some of these courses are
considered part of the official curricula and count as graduate level
courses within the Ph.D.\ programs at UofT.
Initially SciNet provided training specifically oriented toward
Scientific Computing, with the purpose of maximizing user
productivity. These early classes focused on parallel programming
(MPI and OpenMP), best coding practices, debugging, and other
scientific computing needs. Over the years the breadth of classes has
grown, with classes offered in Linux shell programming, parallel
input/output, advanced C++ and Fortran coding, accelerator
programming, and visualization. This is in addition to the annual HPC
Summer School which SciNet runs in collaboration with two
other HPC centres within Compute Ontario\cite{website:CO}, i.e.,
CAC\cite{website:cac} and SHARCNET\cite{website:sharcnet}. This
summer school is a week-long intensive workshop on HPC topics, and
more recently, also Data Science topics\footnote{Similar initiatives
and trends are being carried on by the International HPC Summer
School \cite{website:iHPCss} within the theme of HPC Challenges in
Computational Sciences.}.
Table~\ref{table:SciNet-courses} shows the training events and courses
that SciNet has already been teaching in the areas of HPC and Data
Science.
\begin{table*
\centering
\begin{tabular}{llc}
\hline
Course Name & Certificate & Credits\\
\hline\hline
%
\textit{Data Analysis with R}$^\ddag$ & DS/SC & 12 \\
%
Intro to Apache Spark & DS & 3 \\
Machine Learning Workshop & DS/SC & 6 \\
Hadoop Workshop & DS & 3 \\
Scalable Data Analysis Workshop & DS & 12 \\
Relational Database Basics & DS/SC & 6 \\
Storage and Input/Output in Large Scale Scientific Projects
& DS/SC & 6 \\
Workflow Optimization for Large Scale Bioinformatics & DS/HPC/CS & 6 \\
%
Python for High Performance Computing
& DS/HPC/SC & 12 \\
Parallel R & DS/HPC/SC & 3 \\
Python GUIs with Tkinter & DS/SC & 2 \\
Scientific Visualization & DS/SC & 6 \\
Visualizing Data with Paraview & DS/SC & 6 \\
\hline
\textit{Scientific Computing for Physicists}$^\ddag$\footnote{This includes 3 previously separate module courses: \emph{Scientific Software Development}, \emph{Numerical Tools for Physical Scientist}, and \emph{High Performance Scientific Computing}.}
& HPC/SC & 36 \\
\textit{Intro to Research Computing with Python}$^\ddag$ & HPC/SC & 12 \\
Intro to High Performance Computing & HPC/SC & 3 \\
Intro to Scientific C++ & HPC/SC & 6 \\
Intro to Scientific Programming with Modern FORTRAN & HPC/SC & 7 \\
Intro to Parallel Programming & HPC/SC & 7 \\
Programming Clusters with Message Passing Interface & HPC/SC & 12 \\
Programming Shared Memory Systems with OpenMP & HPC/SC & 6 \\
Practical Parallel Programming Intensive & HPC/SC & 32 \\
Intro to GPGPU with CUDA & HPC/SC & 9 \\
Programming GPUs with CUDA & HPC/SC & 12 \\
SciNet/CITA CUDA GPU Minicourse & HPC/SC & 12 \\
%
Coarray Fortran & HPC/SC & 2 \\
Parallel I/O & HPC/SC & 6 \\
Debugging, Optimization, Best Practices & HPC/SC & 6 \\
HPC Best Practices and Optimization & HPC/SC & 3 \\
HPC Debugging & HPC/SC & 3 \\
Intro to the Linux Shell & HPC/SC & 2 \\
Seminars in High Performance Computing & HPC/SC & 4 \\
Seminars in Scientific Computing & HPC/SC & 4 \\
\hline\hline
\end{tabular}
\caption{Courses taught by SciNet on
\emph{Data Science} (\textbf{DS}),
\emph{High-Performance Computing} (\textbf{HPC}),
and \emph{Scientific Computing} (\textbf{SC}).
$\ddag$ denotes courses already recognized by the university as graduate level credits.}
\label{table:SciNet-courses}
\end{table*}
The number and types of classes which SciNet teaches have grown
significantly~\cite{website:SciNet-Training}. This can be seen in
Figure~\ref{fig:classhours}, which presents the total student
class-hours taught by SciNet over the last four years and the
projection for the current year. This remarkable growth is a
testament to the latent need for this material to be taught.
\begin{figure}[b]\centering
\includegraphics[width=\columnwidth]{./teaching_stats-2012-2016}
\caption{Attendance hours at SciNet training and education events, per
year, for all SciNet classes and Data Science specific classes.}
\label{fig:classhours}
\end{figure}
The need for this training is supported by the enrolment statistics:
our students constitute 35\% of SciNet's total users, clearly showing
that even in a specialized audience this kind of training is still
needed.
For several years the four-week graduate-style classes offered by
SciNet have been accepted for \emph{graduate} class credit by the
departments of Physics, Chemistry and Astrophysics at UofT. This was
possible by accepting the classes as ``modular'' (or ``mini'')
courses, one-third semester long, and bundling three such classes into
a full-semester course. This arrangement has been so popular with
students and faculty that the Physics Department recently listed
SciNet's winter twelve-week HPC class in the course calendar
\cite{website:SciNet-PHY1610H}, allowing graduate students from other
departments in the university to take the class for university credit.
The skills that SciNet aims to transfer are rare and sought-after, and
complement and enhance the skills students learn in regular curricula.
That is why SciNet has developed a set of \emph{Certificate Programs}
\cite{website:SciNet-Certificates}, that users and students can pursue
in \textit{Scientific Computing}, \textit{High Performance Computing},
and/or \textit{Data Science}, once they have completed enough
credit-hours. As a document that proves the holder has highly
competitive skills, and in lieu of graduate credit for most
SciNet courses, the certificates are in high demand. In a
resounding endorsement of our teaching, thus far students have
completed a total of 78 certificates (52 in Scientific Computing, 19
in High-Performance Computing and 8 in Data Science). According to
the current registration and trends, we are projecting to have above a
100 of certificates completed by mid-2016. Moreover, the feedback
from some of our students was that their SciNet's certificates gave
them an advantage to get jobs in industry and the financial sector.
\section{Programs in Data Science}
\label{sec:DS-review}
The wide adoption of the internet in the professional and the personal
sphere ushered in the age of ``Big Data''. The ease of recording of
people's online behaviour, and the ability to rapidly move data, lead
to a large, diffuse, complex amount of data waiting to be mined for
useful information. Because of the typically large size of the data
special hardware and training are often needed. In contrast to
Scientific Computing and HPC, there are many applications of Data
Science in the private sector, in the medical science, banking,
retail, insurance, and internet industries.
Of these industries, Bioinformatics also has a large component in the
academic world. Though a more-recent addition to the HPC world, the
bioinformatics field is well-populated with graduate programs, a
testament to its rapid growth and latent demand. Its emergence as a
major user of HPC systems has resulted in the development of
``Master's of Bioinformatics'', and related degrees. A typical
Master's program is outlined in Table~\ref{table:BioProgram}, this one
from the Indiana-Purdue University at Indianapolis. While having many
features in common with a more-standard SC Master's program, such as
the study of programming and algorithms, it exhibits the particular
needs of the bioinformatics community, stressing the importance of
genetics and biological processes, and a lesser emphasis on
mathematics and programming theory.
\begin{table}[t]\centering
\begin{tabular}{lcc}\hline
Course Name & Type \\
\hline\hline
Introduction to Bioinformatics &required\\
Seminar in Bioinformatics & required\\
Biological Database Management &required\\
Programming for Life Science & required\\
High Throughput Data in Biology & required\\
Machine Learning in Bioinformatics &elective\\
Computational System Biology & elective\\
Structural Bioinformatics & elective\\
Transitional Bioinformatics Applications & elective\\
Algorithms in Bioinformatics & elective\\
Statistical Methods in Bioinformatics &elective\\
Computational Methods for Bioinformatics &elective\\
Next Generation Genomic Data Analytics & elective\\
Next Generation Sequencing & elective\\
Bioinformatics Project & required\\
\hline
\end{tabular}
\caption{The curriculum for the ``Project Track'' two-year
\href{https://soic.iupui.edu/biohealth/graduate/bioinformatics-masters}{Master's
of Science in Bioinformatics} at the Indiana University-Purdue
University in Indianapolis; this forms a good example of a typical
Bioinformatics graduate program.
}
\label{table:BioProgram}
\end{table}
Degrees in Data Science are relatively new, with the first Master's
program only being introduced in the U.S.\ (by North Carolina State
University) in 2007. A sample of some of the classes offered in one
such program is given in Table~\ref{table:DataProgram}. As can be
seen, these programs have a strong focus on data, with statistics,
machine learning, and databases being their standard focus. Analyzing
data that are too big to fit on a standard desktop computer requires
specialized equipment; such training is also part of these
graduate-level programs, as indicated by the presence of the ``Cloud
Computing'' and ``Distributed Systems'' classes. Like typical
graduate-level programs, these degrees usually require of the student
a final project or thesis.
\begin{table}[b]\centering
\begin{tabular}{lcc}\hline
Course Name & Type \\
\hline\hline
Analysis of Algorithms & required\\
Machine Learning & required\\
Advanced Database Concepts & required\\
Distributed Systems & elective\\
Advanced Database Concepts & elective\\
Cloud Computing & elective\\
Information Retrieval & elective\\
Data Mining & elective\\
Web Mining & elective\\
Applied Machine Learning & elective\\
Complex Networks and Their Applications & elective\\
Relational Probabilistic Models & elective\\
Internship in Data Science & elective\\
\hline
\end{tabular}
\caption{A selection of the courses available for the
\href{http://www.soic.indiana.edu/graduate/degrees/data-science/ms-data-science/index.html}{Master's
of Data Science} at the Indiana University.
}
\label{table:DataProgram}
\end{table}
One could argue that the novelty of methods in Data Science is due to
its roots in Business Analytics (BA), where the objective is to make a
decision. The field has certainly grown beyond that, and BA is now
considered a sub-field of Data Science. Another more-recently
developed sub-field is in the realm of health care (``Health
Informatics''). Because these sub-fields are directly applicable to
the private sector (and the associated revenue streams these present)
these have become the most-commonly implemented post-graduate
programs. The Business Analytics programs focus on using data to
refine business administration, as well as develop marketing
strategies. Health Informatics programs concentrate on using clinical
data to optimize health care processes.
The practical focus of Data Science is reflected in the presence of an
internship in the Data Science curriculum listed in
Table~\ref{table:DataProgram}. Internships in such programs are
similar to other co-op-type arrangements: the student works with an
employer for a semester, allowing the student to gain hands-on
experience applying the skills learnt during such period.
\subsection{Academic Data Science Programs}
Graduate level programs in Data Science are not difficult to find.
For instance, programs in bioinformatics (a data-driven field), can be
found on the
\href{https://www.iscb.org/iscb-degree-certificate-programs}{web site
of the International Society for Computational Biology}. It speaks
to the rapid rise of the field bioinformatics, that there are more
bioinformatics programs available than Scientific Computing programs.
Examples lists of other Data Science programs can be found
\href{http://analytics.ncsu.edu/?page_id=4184}{at the NCSU analytics
web site},
\href{http://www.predictiveanalyticstoday.com/online-business-analytics-programs}{the
online business analytics programs site of
{\tt predictiveanalyticstoday.com}} and
\href{http://www.onlinecoursereport.com/the-50-best-masters-in-data-science}{at
{\tt online.coursereport.com}}. They are not as common as programs in
Scientific Computing, due to the fact that Data Science is relatively
new field of study. Among those programs about half are offered in
the fields of Business Analytics and Health Informatics, with the
other half being Data Science programs proper.
\subsection{Non-academic Data Science training}
The demand for Data Science skills (or ``Data Analytics'' skills as
they are often called in the private sector) is so high
\cite{Radermacher:2013:GIE:2445196.2445351} that the private sector
has developed programs to meet the growing demand. A list of such
companies can be found
\href{http://www.skilledup.com/articles/list-data-science-bootcamps}{on
skilledup.com, which contains a list of data science boot-camps}. The format of these classes is varied, though they
are all oriented toward a ``boot-camp'' format: some are in person,
some online; some are one-week long, others twelve weeks. These
programs are very applied, often with one-on-one mentorship with a
seasoned Data Analytics expert. They also include direct contact with
possible future employers.
Moreover, a great number of these training programs are not focused on
developing analytical thinking or problem-solving skills,
\cite{bootcampsnogood} but rather are aimed at graduated Ph.D.s and
postdocs, whose problem-solving skills are assumed to have already
developed. This allows them to focus on the technical training
relevant to the job market. Some of these programs are free, some of
them offer fellowships, and many of them charge on the order of 10-30
thousand US-dollars for a training period of, typically, three months.
These programs have acquired such a level of popularity among young
and recent graduates that the companies offering these programs have
started to perform evaluation tests in order to assess which
candidates are more suitable to be accepted to their programs.
Perhaps the most appealing part for trainees is the networking
platform offered by these programs, as in most of the cases they
provide the opportunity to interact with actual companies looking for
new talent and avoid recruitment layers.
Institutions in the non-profit arena are also starting to offer
programs on Data Science. For instance the Fields Institute, a
traditional institution for mathematical research, has offered several
workshops and courses, and developed a thematic program on Big Data.
Other examples include the International Centre for Theoretical
Physics (ICTP) and the International School for Advanced Studies
(SISSA), prestigious institutions with a well known tradition in
theoretical physics, now offering training in ``Research Data
Science''.
HPC centers are also venturing into Data Science training, offering
workshops on R, Hadoop, machine learning, {\it etc}. SciNet started
offering classes with greater data-oriented content ({\it
cf}. Table~\ref{table:SciNet-courses}) in 2013, with a four-week
class in scientific analysis using Python. Having now finished its
third year, the class remains popular, with about twenty students
taking the class each year. The 2015 fall semester also inaugurated
SciNet's first ``Data Science with R'' class, a class focusing on data
analysis techniques using the R language. This class was very popular
with over twenty-five students finishing the class, and most students
requesting a second installment with more advanced material.
Continuing its growth in the Data Science area, in the last year
SciNet has held workshops in machine learning, scalable data analysis,
and Apache Spark.
Comparing the student- and taught-hours per year shown in
Fig.~\ref{fig:DSclasses}, one sees that the Data Science classes have
been growing consistently, both absolutely as well as relatively (Data
Science related courses roughly constituted less than $2\%$ (2012),
$4\%$ (2013), $12\%$ (2014), $15\%$ (2015), and we project around
$31\%$ (2016) of the total classes taught respectively in each year.
\begin{figure}
\begin{center}
\includegraphics[width=0.49\columnwidth]{./DS_courses-studentshours}
\includegraphics[width=0.49\columnwidth]{./DS_courses-taughthours}
\caption{Total student hours (left) and taught hours (right) per year,
for SciNet's Data Science related courses.}
\label{fig:DSclasses}
\end{center}
\end{figure}
\section{Designing Master's programs in HPC and Data Science}
\label{sec:futurePrograms}
As mentioned above, scientific computing is used by scientists and
engineers as never before, and graduate-level programs in Scientific
Computing are numerous in Canada and around the world. In contrast,
the development of HPC and Data Science programs is in its early
stages, both in academia and the private sector. These programs are
being developed to meet the continued shortfall in skill in these
areas, with the McKinsey Global Institute estimating
that the United States will be short 140,000 to 190,000 data analytics
professionals by 2018 \cite{McKinsey_BigData}.
One may wonder whether online learning could not satisfy this need. A
few examples of MOOCs (Massively Open Online Courses) in HPC and Data
Science do exist. However, seeing the growth in enrolment in SciNet's
in-person courses and the summer school over the years ({\it
cf}.~Figs.~\ref{fig:classhours} and \ref{fig:DSclasses}) shows that
many students still prefer the face-to-face format.
Similarly, one may wonder why certificate programs do not suffice
for HPC and DS education. As successful as these programs are, they
have a few disadvantages. Firstly, they are mostly
collections of fairly specific technical training: this leaves no
room for more fundamental material. Secondly, it is also hard to encorporate
an internship or thesis into such a certificate. Finally,
certificates tend to carry less weight than degrees, and, in line
with this, the demand for for-credit courses is larger than that for
not-for-credit courses, as our experience with SciNet's Scientific
Computing graduate course has shown.
A degree program in HPC or DS could offer more academic and
fundamental education, which would leave the student with the
analytical skills and high-level knowledge to stay on top of their
field regardless of changes in computational technology.
In the following sections we propose a curriculum for graduate-level
HPC and DS programs. One will notice a substantial overlap of topics
with the training courses currently taught by SciNet ({\it cf}. Table
\ref{table:SciNet-courses}). This is no coincidence: the training
program was developed on the basis of student feedback and requests,
and was a primary inspiration for the curricula proposed here. The
design was also influenced by the few existing examples of such
programs, as described above.
It should also be emphasized that the programs in HPC and DS are both
designed to allow students to a follow more
industrial/practically-oriented track or an academic/research-oriented
track, by selecting the appropriate set of elective courses and the
corresponding research project/internship. Additional \emph{advanced}
courses could be made available according to the interest and demand
of the students.
\subsection{Design of an HPC Master's Program}
\label{sec:SciNet-HPCprogram}
In this section we present a comprehensive and complete curriculum for
a two-year \emph{Master's Program in High-Performance Computing}.
Students would complete a total of twelve courses. The five
required courses, each one-term long, set the basis of HPC and ARC
knowledge (including topics such as modern and professional software
development, parallel techniques, performance and optimization, best
practices, distributed systems and resources). The seven elective
courses allow the student to specialize in a particular area. In
addition, a final internship or research project would be carried out.
As in any typical Master's program, a student entering the program
will be expected to possess a Bachelor's degree (B.Sc.\ or B.A.). It
is desirable that students have some background in sciences and the
basics of coding and programming ({\it e.g}.\ Fortran, C, C++);
otherwise it is strongly recommended that students take introductory
programming classes. Note that courses from the \emph{Data Science}
Master's Program are also eligible to be taken, with consent of the
graduate coordinator or adviser.
The course work for the High-Performance Computing program could
consist of the following courses.
\begin{description}
\item[Software Development (*):] The principles of creating
modern, maintainable code.
Special attention is given to designing modular code, and tackling
scientific computational projects. Languages: C++/C/Fortran.
\item[Best Practices (*):] Introduction and discussion of techniques
and methods to be considered when designing and implementing
computational research projects. Topics include: version control,
modularity, libraries.
\item[Performance \& Optimization (*):] Principles and tools for
measuring performance, finding bottlenecks, and optimizing existing
code.
\item[Basics of Parallel Programming (*):] A review of homogeneous and
heterogeneous architectures is presented, followed by parallel programming
paradigms such as OpenMP, MPI, and hybrid implementations.
\item[HPC Algorithms:] A review of commonly used algorithms in
computational science, such as Monte Carlo, implicit and explicit
methods to solve differential equations, timestepping techniques,
finite-element and finite-volume methods.
\item[Machine Learning:] Theory and practice of constructing
algorithms that create models from data. Topics: probabilistic
foundations, linear and logistic regression, neural networks,
Bayesian networks, tree models, support vector machines, density
estimation, accuracy estimation, normalization, model selection.
\item[Numerical Methods:] A review of the commonly used numerical
methods in scientific computing, such as linear solvers, fast
Fourier transforms. In contrast to the HPC Algorithms, it will
cover the basic principles and the theory behind the methods only
briefly, and focus mostly on the implementation and utilization via
libraries and specific examples.
\item[Scientific Computational Modeling:] This course will offer an
introduction to the basics of Scientific Computing and a review of
the most common algorithms and packages used in different fields of
research computing and computational sciences (astrophysics,
chemistry, genetics, {\it etc.} Introduction to computing modeling,
such as implementation of complex networks based on relational data
sets, and the evaluation of network properties using graph theory
elements, among many others.
\item[Programming Accelerators:] Some computational problems can be
computed much more efficiently on accelerators such as graphics
cards. This course will present an introduction to the use of
hardware accelerators (GPUs, FPGAs, many-core systems) in HPC.
Topics will include a review of the hardware, architectures, and
programming languages, such as, CUDA, OpenACC, and OpenCL.
\item[Research Data Management:] Design strategies, storage management
and I/O patterns, in order to prevent bottlenecks in massively
data-driven projects. Real use cases from various fields
(bioinformatics, molecular biophysics, medical physics,
biochemistry, quantum-chemistry, geophysics, {\it etc}.) have shown that,
quite often, approaches that work on a desktop do not perform on a
larger scale.
A review of the latest policies regarding scientific data
availability will also be discussed and presented.
\item[Visualization for Scientific Computing:] A review of basic
visualization concepts and methods with
applications to scientific data.
\item[Operating System Environment:] Scientific and high performance
computing is intimately linked to the hardware, OS, and application
framework on which they used. This course will help students become
comfortable working on *nix systems. Topics such as the command
line, shell scripting and advanced OS topics will be covered.
\item[HPC Hardware and System Administration:]
An integrated view on the
technology in HPC, the machines, hardware, network, file systems,
and what is involved in getting an HPC system up and running.
\item[Student HPC/ARC Seminar:] Weekly sessions running throughout the
year, with students presenting and discussing current papers and
research in the fields of HPC and ARC. Researchers and instructors
will providing guidance and supervision during the sessions.
\end{description}
In addition to this course work, the program would include an 4-month
internship or independent research project in the final year.
The program presented above is intended to be flexible. In the first
year of the HPC program, students might take \emph{HPC Algorithms},
\emph{Parallel Programming}, \emph{Software Development},
\emph{Performance and Optimization}, \emph{Best Practices}, and
\emph{Numerical Methods}, as well as attend the HPC/ARC seminar
series. In year two, they might take
\emph{Research Data Management}, \emph{Scientific Computational Modeling},
\emph{Graph Theory Applications},
\emph{Machine Learning},
\emph{Visualization for Scientific Computing}, and complete the degree
with a Research Project.
Eligible courses from the \emph{Data Science} Master's Program are
also possible to take with previous consent of the graduate
coordinator or adviser.
\subsection{Design of a Data Science Master's Program}
\label{sec:SciNet-DSprogram}
Here we present a comprehensive curriculum for a two-year
\emph{Master's Program in Data Science}. As with the HPC program,
students would complete a total of twelve courses. The five required
courses set the foundation of data analysis knowledge (including
topics such algorithms, databases, statistics, machine learning) and
seven elective courses allow the student to specialize in a particular
area, such as data mining, machine learning, complex systems. In
addition, a final internship or research project would be carried out,
in order to obtain real-world experience.
As in any typical Master's program, the entry level will be a
bachelor's degree (B.Sc.\ or B.A.). It is desirable that students
have some background in sciences and the basics of coding and
programming ({\it e.g}.\ Python, R); otherwise it is strongly
recommended that students take introductory programming classes.
Notice that courses from the \emph{High Performance Computing}
Master's Program may be taken with permission of the graduate
coordinator or adviser ({\it e.g}.\ \textit{Operating System
Environment}).
The fundamentals of data analysis should be at the core of a program
that will produce analysts capable of tackling real-life problems in
Data Science. Learning theoretical and practical approaches gives
students an advantage in the real world; this program proposes to
combine both in a unique fashion (similar to how most SciNet courses
are structured). Topics such as statistical analysis, algorithms and
large data sets are at the centre of the proposed program and
constitute the ``core'' (required) courses. Additionally, the
elective courses allow students to choose a specialization path by
gaining expertise in areas such as: social data mining, machine
learning, and representation of complex interactions. The course work
for the Data Science program could consist of the following courses:
\begin{description}
\item[Overview of Data Science:] An overview of the field of Data
Science, covering data-driven problems from several disciplines such
as, astronomy, bioinformatics, digital humanities, social sciences,
{\it etc}.
\item[Basics of Programming:] An introduction to programming, coding
structures, and basic algorithms. This course will focus on
languages with data-analytic capabilities, such as \emph{Python} and
\emph{R}.
\item[Data Analysis Algorithms (*):] An overview of the
major classes of algorithms, including comparison-based algorithms:
search, sorting, hashing; information extraction algorithms (graphs,
databases); graph algorithms: spanning trees, shortest paths, depth
and breath-first search.
\item[Machine Learning (*):] Theory and practice of constructing
algorithms that create models from data. Topics: probabilistic
foundations, linear and logistic regression, neural networks,
Bayesian networks, tree models, support vector machines, density
estimation, accuracy estimation, normalization, model selection.
Deep learning algorithms.
\item[Database Theory (*):] The mathematical
foundations of databases, search and query semantics, relational,
complex, object-oriented and semi-structured database models.
\item[Database Applications:] SQL, relational algebra, index,
views, constraints; query complexity; data models, including I/O
model, streaming model,
query optimization,
optimal join algorithm will be presented.
\item[High Dimensional Data Analysis (*):] Theory and methods for
exploring high-dimensional data will be presented, including linear
and non-linear dimension reduction, manifold learning; Euclidean
representation of proximity and network data; clustering,
statistical pattern recognition.
\item[Applied Statistics (*):] A review of statistical techniques,
with applications to data analysis problems. Topics will include
hypothesis testing, general linear models, generalized linear
methods, multivariate statistics.
\item[Best Practices:] Introduction and discussion of techniques and
methods to be considered when designing and implementing
computational research projects. Topics include: version control,
modularity, libraries.
\begin{comment}
\item[Advanced Machine Learning:] This course will cover advanced
topics, such as Memory-limited algorithms, no-regret algorithms,
bandit algorithms; large-scale kernel methods, Gaussian process
regression.
\item[Stochastic Machine Learning:]
Theory and application of machine
learning through stochastic optimization. Incremental estimation
algorithms, streaming data; online learning problems and
reinforcement learning; label propagation.
\end{comment}
\item[Data Mining:]
Algorithmic approaches to discovering patterns in
large data sets will be covered, including data exploration and
cleaning; association rules, clustering, anomaly detection, and
classification.
Other examples will include applications
to text and the web: crawling, indexing, ranking and
filtering algorithms; applications to search, classification and
recommendation; link analysis, significance tests.
Data mining techniques applied to social media. Sentiment analysis,
polarity classification; graph properties of social networks,
homophily, distance, influence, spectral methods, information
diffusion, probabilities models.
\item[Data Security and Integrity:] This course will cover technique
on how to safely protect the integrity and privacy of sensible data,
standards on data regulations ({\it e.g}.\ health, patient data,
denomic data), privacy, encryption algorithms.
\item[Search and Classification Algorithms:] Information retrieval
theory, with applications, will be explored. Examples of searching
algorithms, mathematical representations and matrix applications
will be covered.
\item[Distributed Systems:] Focus on distributed resources and data
sets across different architectures and systems. Students will learn
the skills and abilities of dealing with retrieving and screening
large data sets in \textit{cloud}-type based systems.
\item[Graph Theory Applications:] An introduction to graph theory and
complex systems and its applications will be offered. Material will
include the implementation of complex networks based on relational
data sets, and the evaluation of network properties.
\item[Visualization Techniques of Unstructured Data sets:] A review of
basic visualization concepts, with applications to unstructured
and/or large data sets and complex and dynamical systems, in order
to gain insights in the data and visually expose potential
correlations.
\item[Data Science Seminar Series:] Weekly sessions running throughout
the year, where students will present and discuss current papers and
research in the fields of data science and analytics. Researchers
and instructors will provide guidance and supervision during the
sessions.
\end{description}
In addition to this course work, the program would include an
4-month internship or research project.
The first year of classes could consist of
\emph{Overview of Data Science},
\emph{Data Analysis Algorithms},
\emph{Database Theory},
\emph{Applied Statistics} and
\emph{Visualization Techniques of Unstructured Data Sets},
while the second year could contain courses in
\emph{High Dimensional Data Analysis},
\emph{Graph Theory Applications},
\emph{Data Mining},
\emph{Search and Classification Algorithms}, and \emph{Distributed Systems}, and
would be completed by an Internship or a Research Project.
As with the HPC/ARC track, eligible courses from the \emph{HPC}
Master's Program are also possible to take with previous consent of
the graduate coordinator or adviser.
\section{Conclusion}
\label{sec:concl}
We have demonstrated the need for programs in higher education in
High-Performance Computing and Data Science. If the qualitative
evidence of this seems somewhat limited, it should be understood that
existing HPC and DS programs (academic and non-academic) are still
relatively new. While some such programs are already in existence, in
many cases students must use non-academic options, or teach the
material to themselves. Academic programs would offer the benefit of
not just teaching specific technical skills, but an education in the
fundamentals of HPC and DS and instilling the analytical skills needed
to adapt to an ever-changing technological landscape.
We have reviewed
existing academic and non-academic education programs, in both HPC and
DS. In light of this review, we presented a design for Master's
programs in HPC and DS, based on these examples and drawing from the
experience and enrollment statistics in not-for-credit training in HPC
and DS by the SciNet HPC Consortium at the University of Toronto.
To get well-founded graduate master's programs off the ground will not be
without challenges. It will likely involve partnerships and
discussions with other departments and institutes in order to offer a
stronger and multi-disciplinary program. Existing HPC Centers, which
already operate between different disciplines, can play a fundamental
role in bringing together such programs.
\bibliographystyle{unsrt}
|
\section{Introduction}
In the electromagnetic theory, Maxwell applied simultaneously the quaternion analysis and vector terminology to describe the electromagnetic properties. Hamilton invented the algebra of quaternions in 1843. During two decades from that time on, the quaternion was separated into the scalar and vector parts. In his studies, J. C. Maxwell mingled the quaternion analysis and vector terminology to depict electromagnetic properties in 1873. Presently some scholars apply the vector terminology to depict the physical properties of electromagnetic and gravitational fields, while some scholars begin to study the electromagnetic theory \cite{morita, edmonds} and gravitational theory \cite{majernik, rawat} with the algebra of quaternions.
Graves and Cayley invented independently the octonions, which can be considered as an ordered couple of quaternions. Some scholars applied the octonion analysis \cite{gogberashvili, demir} to describe the electromagnetic and gravitational theories \cite{mironov, negi}. The study reveals that the octonions (or the standard octonions, rather than the split octonions) can be divided into two parts, the quaternions and $S$-quaternions (short for the second quaternion), and their coordinate values are allowed to be complex numbers. And the $S$-quaternion in an octonion is similar to the imaginary number in the complex number. In the complex octonion flat space, the complex quaternion space is appropriate to describe the gravitational property, meanwhile the complex $S$-quaternion space is proper to depict the electromagnetic property (Appendix A). Further these properties can be extended from the complex octonion flat space \cite{weng1} to the complex octonion curved space. That is, the complex octonion curved space is possible to describe the gravitational and electromagnetic fields. The complex quaternion curved space fits for studying gravitational properties, and the complex $S$-quaternion curved space is apt for researching electromagnetic properties.
\subsection{Existing studies}
After the tensor analysis \cite{bishop, morandi} was evolved from the vector analysis and other theories, Einstein \emph{et al.} adopted the tensor analysis and pseudo-Riemannian space theory to study the gravitational properties in the curved four-space-time. Subsequently the quaternion and octonion curved spaces are applied to investigate physical properties of electromagnetic and gravitational fields in some situations. Marques-Bonham \cite{marques} developed the geometrical properties of a flat tangent space-time local to the manifold of the Einstein-Schr\"{o}dinger nonsymmetric theory on an octonionic curved space. Dundarer \cite{dundarer} defined a four-index antisymmetric and non-Abelian field, which satisfies a self-duality relation in eight-dimensional curved space. Tsagas \cite{tsagas} studied the evolution of electromagnetic fields in curved space-times, and calculated the wave equations. Castro \cite{castro} proposed a nonassociative octonionic ternary gauge field theory, based on a ternary bracket. Demir \cite{demir2} formulated the Maxwell-Proca type field equations of linear gravity in terms of hyperbolic octonions (split octonions). Chanyal \cite{chanyal} \emph{et al.} described the octonion formulation of Abelian/non-Abelian gauge theory in terms of the Zorn vector matrix realization. Kalauni \cite{kalauni} \emph{et al.} obtained the fully symmetric Dirac-Maxwell's equations as one single equation by using the matrix presentation, with the help of the algebraic properties of quaternions and octonions.
In the existing electromagnetic and gravitational theories described with the curved spaces, there are mainly three description methods as follows:
1) Curved four-space-time. Within the General Theory of Relativity in 1915, Einstein introduced the method of pseudo-Riemannian space to study gravitational properties in the curved four-space-time. Later some scholars attempt to apply the pseudo-Riemannian space to explore various properties of gravitational field, and even the electromagnetic field. But this method is not successful in depicting convincingly the physical property of electromagnetic field.
2) Real octonion curved space. Some scholars described the physical properties of the gravitational and electromagnetic fields in the flat space, by means of the real quaternion/octonion flat space. However this method was found to be inadequate finally \cite{weng2}. And it is not suitable to describe the gravitational and electromagnetic fields in the curved space either, making use of the real quaternion/octonion curved space. One reason is that the definition of the arc length for this space deviates from the space-time interval in the physics.
3) Complex octonion curved space. The complex quaternion/octonion flat space is able to depict the physical property of gravitational and electromagnetic fields in the flat space. Some scholars research gravitational and electromagnetic properties in the curved space, depending on the complex quaternion/octonion curved space. On the basis of the locally tangent frame and tensor analysis, the paper can explore the characteristics of complex quaternion/octonion curved space. There are a few contemporary results at present. The study reveals that the definition of the arc length for this space is identical with the space-time interval in the physics. The connection coefficient and curvature of the complex quaternion/octonion curved space, which are similar to the field strength and source, will exert an influence on the object's movement to a certain extent, no matter what kind of reason it is to result in the space curving evidently. On the contrary, measuring the field strength and source relevant to the gravitational and electromagnetic fields, one may be able to ascertain the deviation amplitude between the complex quaternion/octonion curved space and its flat space.
Making use of the comparison, the paper found that the four-dimensional Riemannian space corresponds to the four-dimensional Euclidean space. But both of them are not suitable to be applied to describe directly the physical properties of gravitational or electromagnetic field. The four-dimensional pseudo-Riemannian space corresponds to the four-dimensional pseudo-Euclidean space (or Minkowski space, four-space-time), and is appropriate for describing gravitational properties in the General Theory of Relativity. Similarly the real quaternion/octonion curved space corresponds to the four-dimensional Riemannian space, and is unsuitable to depict directly the physical property of gravitational or electromagnetic field either. In contrast, the complex quaternion/octonion curved space corresponds to the four-dimensional pseudo-Riemannian space, and is fit for researching the gravitational and electromagnetic properties. On the basis of this curved space, the paper will study various properties of gravitational and electromagnetic fields in the complex quaternion/octonion curved space, including the field potential, field strength, field source, linear momentum, angular momentum, torque, and force and so forth.
\subsection{Two topics}
In the curved four-space-time of the General Theory of Relativity, the scholars delved into the gravitational theory \cite{fischbach, adelberger}, attaining many scientific achievements. But there are still a few things to worry about.
1) Astrophysical jets. The scholars have being doubted the validity and relevant deductions of E\"{o}tv\"{o}s experiment for a long time \cite{baessler, turyshev}. They not only inspect the E\"{o}tv\"{o}s experiment in the laboratory again and again \cite{moffat}, but also measure repeatedly the gravitational constant in the astronomical observatory \cite{liu}. One recent theoretical result points out that the E\"{o}tv\"{o}s experiment has never been validated under the strong electromagnetic circumstance, suspecting that the gravitational mass may be varied with fluctuation of electromagnetic strength. Further the variable gravitational mass is able to explain the phenomenon of astrophysical jets \cite{weng3}, which is the defeat for the classical gravitational theory or the General Theory of Relativity. Up to the present, this suspicion is an ongoing process, and the answer has not been concluded yet. Obviously the query on the changeability of gravitational mass will impact further the theoretical basis of the General Theory of Relativity. The appeal of improving the General Theory of Relativity is gradually deepening and expanding. This open opinion provides one approach to develop the General Theory of Relativity.
2) Dark matter. An abundance of astronomical observational results reveal that there are a large amount of undetected matters in the galaxies and galaxy clusters. But the classical gravitational theory and the General Theory of Relativity cannot account for the existence of undetected matters. Obviously these two existing theories both are not perfect enough. Nowadays the scholars mainly develop two kinds of investigations for this problem. On the one hand, some scholars propose to modify the laws of gravity established by I. Newton and A. Einstein (such as MOND), attempting to explain the anomalous observations. But this method cannot account for the properties of galaxy clusters. On the other hand, some scholars introduce the concept of `dark matter', explaining the rotational velocity curves of galaxies and the gravitational lensing of galaxy clusters. All of these endeavors state that the classical gravitational theory and the General Theory of Relativity both may encounter the puzzling impediment, which may be very tough to overcome.
3) Curved space. The human being often breaks into exclamations of wonder of scientific laws, even if we are not able to comprehend why the scientific laws are so intriguing. Fortunately the human being possesses the ability to suppose, measure, and describe the particular contents of scientific laws to a certain extent. In other words, `the reason for the existence of scientific laws' and `the specific contents of scientific laws' are two disparate topics. Similarly the human being may not understand why the space is curved when we are in a curved space, although the scholars are able to measure and describe the bending degree of one space. If we find the reason why a space is flat, we may know why other spaces are curved. Therefore `the reason to curve the space' and `the bending degree of the curved space' are two distinct topics, they should be treated discriminatively. Einstein tried to uncover why the space is curved under certain circumstances, in the General Theory of Relativity. However this paper involves in one distinct topic, and focuses on describing the possible influence of the bending degree of curved space on some equations of the field theory.
The analysis in the above reveals that the concept, viewpoint, and experiment relevant to the curved space are evolving constantly and gradually, along with the development of science and technology. For the General Theory of Relativity, it may be necessary to have a fundamental rethink of existing concepts appropriately, and even supplement a few new results.
1) Distorting the result of the E\"{o}tv\"{o}s experiment. The General Theory of Relativity reckons subjectively that the gravitational mass is unchangeable, and does not contemplate the influence of the energy gradient (that is, one force term) on some other experiments \cite{reasenberg}. However some theories deduce that the gravitational mass may be varied with fluctuation of field strength and so forth. It results in the gravitational mass to depart from the inertial mass with a deviation. The energy gradient relevant to the gravitational or electromagnetic strength will act on the E\"{o}tv\"{o}s experiment. Therefore it is necessary to validate the E\"{o}tv\"{o}s experiment under the strong electromagnetic circumstance.
2) Surmising the reason for curving the space. As a pioneer, Einstein broke fresh ground in the field theory, that is, the curved space. This theory surmises subjectively that the curvature tensor is directly proportional to the energy-momentum tensor, deducing the field equations. However the field equations should be verified continually in the experiment. The field equations are doubtful, and call for more collateral evidences to support its claim. You can even guess that there may be more than one reason to curve the space.
3) Origin of mass. This is one topic that the General Theory of Relativity has never been touched on. At present a few scholars are engaging in some projects to figure out the origin and component of the mass. Undoubtedly this is one underlying challenge for some propositions in the General Theory of Relativity.
In a word, the paper will focus on studying the influence of curved space on the physical quantity, rather than try to find out the reason why the space is curved in the physics, by means of the affine frame of complex octonion curved space.
\section{Orthogonal affine frame}
In this paper, it is necessary to introduce the tangent-frame component, $C_\nu$ , of one point $P$ in the curved space, in order to define the component of a physics quantity in the curved coordinate system. And the tangent frame system $\{ C_\nu \}$ belongs to the affine frame. In general the tangent-frame components are neither equal-length, nor perpendicular to each other. For the pseudo-Riemannian space in the General Theory of Relativity, the tangent-frame component may be a vector or a scalar. In the quaternion or octonion curved space, the tangent-frame component, $C_\nu$ , may be the quaternion or octonion respectively. $\nu = 0, 1, 2, 3, 4, 5, 6, 7$.
In the physics, the most common coordinate systems are orthogonal and even orthonormal, to give prominence to the physical contents and lower down the degree of difficulty of the mathematics in the physics theories. That is, not all affine frames are suitable to be chosen as the curved coordinate system, for the gravitational and electromagnetic fields in the curved space. And it is essential to demarcate and filter out the appropriate affine frames further. When we study the physics quantity and relevant properties in the curved space, it is proper to stipulate to choose the orthogonal affine frame (or the orthogonal curved coordinate system), for the sake of reducing the correlative mathematical difficulty (such as, the metric tensor and invariants). On the other hand, the physics quantity in the orthogonal affine frame can be transformed into one in an orthogonal and equal-length affine frame, in order to draw a comparison between the physics quantity in the flat space (see Ref.[9]) with that in the tangent space of the curved space.
According to this arrangement, the gravitational and electromagnetic theories can be extended from the flat space into the curved space. Firstly, transform the gravitational and electromagnetic theories in the flat space into that under the orthogonal and equal-length affine frame of the tangent space in the curved space; Secondly, transform the gravitational and electromagnetic theories in the orthogonal and equal-length affine frame into that in the orthogonal affine frame; Thirdly, in the orthogonal affine frame, study the influence of the curved space on the force and so forth in the gravitational and electromagnetic fields.
In the curved space, the physics quantity will contain several parameters of the curved space. a) The integrating function, $\mathbb{X}$, of field potential can be chosen as the first-rank tensor, and then the field potential, $\mathbb{A}$, as the first covariant derivative of $\mathbb{X}$, will contain the connection coefficient. As the second covariant derivative, the field strength, $\mathbb{F}$, may contain the connection coefficient and curvature tensor. As the third covariant derivative, the field source, $\mathbb{S}$, will contain the connection coefficient, curvature tensor, and curvature derivative. b) If the field potential, $\mathbb{A}$, is chosen as the first-rank tensor, the field strength, $\mathbb{F}$, as the first covariant derivative of $\mathbb{A}$, will contain the connection coefficient. As the second covariant derivative, the field source, $\mathbb{S}$, may contain the connection coefficient and curvature tensor. c) When the angular momentum, $\mathbb{L}$, is chosen as the first-rank tensor, the torque, $\mathbb{W}$, as the first covariant derivative of $\mathbb{L}$, will contain the connection coefficient. As the second covariant derivative, the force, $\mathbb{N}$, may contain the connection coefficient and curvature tensor. It should be noted that three different first-rank tensors in the above may not be in the same tangent frame or curved space.
In a similar way, the acceleration, $\mathbb{D}$, can be chosen as the first-rank tensor, and then the jerk, $\mathbb{K}$, as the first covariant derivative of $\mathbb{D}$, will contain the connection coefficient. As the second covariant derivative, the jounce, $\mathbb{J}$, may contain the connection coefficient and curvature tensor. Of course, as one first-rank tensor, the curved space relevant to the acceleration, $\mathbb{D}$, may be different from that relevant to these three first-rank tensors in the above.
Obviously the covariant derivative of a physics quantity may include a few parameters of the curved space. Comparing the physics quantity in the curved space with that in the flat space, we are able to distinguish different bending degrees of the curved space. Contrasting several sets of measured values of physics quantities will conclude the bending degree of a curved space.
\section{Complex-quaternion curved space}
Under an orthogonal transformation of the coordinate system with the complex quaternions, the norm of the complex quaternions remains unchanged. The norm (or the arc length) is identical to the space-time interval in the physics, and is able to be applied to describe the gravitational field equations in the complex quaternion curved space.
Apparently, in the complex quaternion curved space, the underlying space is the quaternion space, and the tangent space is the quaternion space also. Making use of the complex quaternion orthogonality and affine frame, it is able to define the quaternion metric and covariant derivative, studying the gravitational properties in the complex quaternion curved space.
\subsection{Quaternion metric}
In the complex quaternion flat space, the complex quaternion radius vector is, $\mathbb{H}_g = i \textbf{\emph{i}}_0 h^0 + \textbf{\emph{i}}_r h^r $, with the basis vector being ${\emph{\textbf{i}}_j}$. According to the multiplication of quaternion, the norm $S$ is written as, $S^2 = \mathbb{H}_g \circ \mathbb{H}_g^* = - (h^0)^2 + (h^1)^2 + (h^2)^2 + (h^3)^2$. Obviously the differential, $d S$, of the norm is able to be chosen as the arc length of the complex quaternion (rather than the real-quaternion) flat space. And that this norm accords with the requirement of space-time interval in the physics. Herein $i$ is the imaginary unit. $h^j$ is real. $\circ$ is the quaternion multiplication. $*$ is conjugate of quaternion. $\emph{\textbf{i}}_0^2 = 1$, and $\emph{\textbf{i}}_r^2 = - 1$. $\emph{\textbf{i}}_j^2 = \emph{\textbf{i}}_j \circ \emph{\textbf{i}}_j$, for each subscript $j$. For the superscript and subscript, there are, $j, k, m, n = 0, 1, 2, 3$; $r, q = 1, 2, 3$.
In the tangent space (that is, the locally flat space) of the complex quaternion curved space, the radius vector is, $\mathbb{H}_g = i \textbf{\emph{e}}_0 c^0 + \textbf{\emph{e}}_r c^r $ , with the tangent-frame quaternion being ${\textbf{e}_j}$ . In the orthogonal and unequal-length affine frame, the metric of complex quaternion curved space is defined as,
\begin{equation}
d S^2 = d \mathbb{H}_g^* \circ d \mathbb{H}_g = g_{\overline{j}k} d \overline{u^j} d u^k ~,
\end{equation}
where the metric coefficient, $g_{\overline{j}k} = \textbf{e}_j^* \circ \textbf{e}_k$ , is quaternion-Hermitian. The orthogonal tangent-frame quaternion is, $\textbf{e}_j = \partial \mathbb{H}_g / \partial u^j$ . $\textbf{e}_0$ is the scalar part, while $\textbf{e}_r$ is the component of vector part. $\textbf{e}_k$ is unequal-length, and does not contain the imaginary unit $i$ . $ u^0 = i c^0 $ , and $ u^r = c^r $ . $ ( u^j )^* = \overline{u^j} $ , and it denotes that the correlated tangent-frame component, $\textbf{e}_j$ , is quaternion-conjugate. $c^0 = v_0 t$, $v_0$ is the speed of light, and $t$ is the time. $c^j$ is real, and $g_{\overline{j}k}$ is scalar.
\subsection{Quaternion parallel translation}
The normal parallel translation (Euclidean space) and the Levi-Civita parallel translation (Riemannian space) both are not suitable for the complex quaternion curved space. One reason is that the spaces associated with in the physics are not only the complex quaternion flat space (tangent space) but also the pseudo-Riemannian space (curved space). Therefore it is necessary to introduce one new concept of parallel translation. It is similar to the Levi-Civita parallel translation, and is able to satisfy the requirement of the complex quaternion space and of the space-time interval.
In a complex quaternion space, when the quaternion product of two complex quaternions, $\mathbb{G} ( g^0 , g^1 , g^2 , g^3 )$ and $\mathbb{Z} ( z^0 , z^1 , z^2 , z^3 )$, equals to zero, that is, $\mathbb{G}^* \circ \mathbb{Z} = 0$, the two quaternions, $\mathbb{G}$ and $\mathbb{Z}$ , are perpendicular to each other. This definition is called as the quaternion orthogonality. After that we can define a quaternion connection and parallel translation in the complex quaternion curved space, referring to the inference procedure of Levi-Civita parallel translation in the Riemannian space. Herein $g^j$ and $z^j$ are all scalar.
In the complex quaternion curved space, the complex quaternion physics quantity $\mathbb{A}_1$ , in the tangent space $\mathbb{T}_1$ at a point $M_1$ on the complex quaternion manifold, can be decomposed in the tangent space $\mathbb{T}_2$ at the point $M_2$ near $M_1$ . According to the definition of quaternion orthogonality, the physical quantity $\mathbb{A}_1$ can be separated into the projection component $\mathbb{A}_2$ in $\mathbb{T}_2$ , and the orthogonal component $\mathbb{G}_2$ perpendicular to $\mathbb{T}_2$ . In case the differential, $\mathbb{A}_2 - \mathbb{A}_1 = 0$, the physical quantity $\mathbb{A}_2$ is the parallel translation of $\mathbb{A}_1$ . And this definition is called as the quaternion parallel translation. Especially, when the scalar parts of $\mathbb{A}_1$ and of $\mathbb{A}_2$ are all null, $\mathbb{A}_1$ and $\mathbb{A}_2$ both will be degenerated into the vectors. Further the orthogonality of quaternion will be reduced to that of vector, and the quaternion parallel translation to the Levi-Civita parallel translation.
\subsection{Quaternion covariant derivative}
In the Riemannian space, substituting the Levi-Civita parallel translation and tangent frame vector for the normal parallel translation and basis vector in the flat space respectively, the approach to define the partial derivative via the limit can be directly extended into the definition of covariant derivative of the Riemannian space.
Similarly in the complex quaternion curved space, substituting the quaternion parallel translation, tangent-frame quaternion, and quaternion orthogonality for the Levi-Civita parallel translation, tangent-frame vector, and vector orthogonality respectively, the covariant derivative of the pseudo-Riemannian space can be extended to that of the complex quaternion curved space directly.
For the first-rank contravariant tensor $Y^j$ of a point $M_2$ in the complex quaternion curved space, the component of quaternion covariant derivative with respect to the coordinate $u^k$ is,
\begin{equation}
\triangledown_k Y^n = \partial ( \delta_j^n Y^j ) / \partial u^k + \Gamma^n_{jk} Y^j ~,
\end{equation}
where $\Gamma^n_{jk}$ is the connection coefficient. Making use of Eq.(1), one can deduce the connection coefficient from the above (Appendix B). $Y^n$ and $\Gamma^n_{jk}$ are all scalar.
\section{Gravitational field equations}
In the complex quaternion curved space, the gravitational potential had been transformed from the rectangular coordinate system (in the flat space) to the orthogonal and unequal-length affine frame (in the tangent space). The gravitational potential, $\mathbb{A}_g ( i a^0 , a^1 , a^2 , a^3 )$ , can be defined from the integrating function of gravitational potential, $\mathbb{X}_g$ ,
\begin{equation}
\mathbb{A}_g = i \lozenge^\star \circ \mathbb{X}_g ~,
\end{equation}
where $\mathbb{A}_g = i \lozenge^\star \circledcirc \mathbb{X}_g + i \lozenge^\star \circledast \mathbb{X}_g$ . $i \lozenge^\star \circledcirc \mathbb{X}_g$ and $i \lozenge^\star \circledast \mathbb{X}_g$ denote the scalar and vector parts of $\mathbb{A}_g$ respectively. $\lozenge a^j = i \textbf{e}_0 \triangledown_0 a^j + \delta^{rq} \textbf{e}_r \triangledown_q a^j$. $\nabla a^j = \delta^{rq} \textbf{e}_r \triangledown_q a^j$. $\textbf{e}^j = g^{jk} \textbf{e}_k$. $\mathbb{X}_g = x^j \textbf{e}_j$. The gauge equation is chosen as, $\nabla \times (x^r \textbf{e}_r) = 0$. $\mathbb{A}_g = i a + \textbf{a}$. $a = a^0 \textbf{e}_0$ , and $\textbf{a} = a^r \textbf{e}_r$ . Apparently, in the affine frame, the gravitational potential includes not only the physics quantity but also the spatial parameter of the complex quaternion curved space (Table 1). The quaternion operator $\lozenge$ is extended from the flat space (see Ref.[9]) to the curved space in this paper. And the quaternion operator is different from the Dirac operator. The former comes from the algebra of quaternions, while the latter derived from the mass-energy relation in the Special Theory of Relativity, which is applied to the Quantum Mechanics. $\star$ is the complex conjugate. $a^j$ is real.
The gravitational strength, $\mathbb{F}_g ( i f^0 , f^1 , f^2 , f^3 )$ , is defined as
\begin{equation}
\mathbb{F}_g = \lozenge \circ \mathbb{A}_g ~ ,
\end{equation}
where $\mathbb{F}_g = \lozenge \circledcirc \mathbb{A}_g + \lozenge \circledast \mathbb{A}_g $ . The scalar part of $\mathbb{F}_g$ is, $\lozenge \circledcirc \mathbb{A}_g = i f^0 \textbf{e}_0$, and the vector part of $\mathbb{F}_g$ is, $\lozenge \circledast \mathbb{A}_g = f^r \textbf{e}_r$. The gauge equation of gravitational potential is chosen as, $f^0 = 0$. The vector part of gravitational strength can be separated into two components, $f^r \textbf{e}_r = i \textbf{g} / v_0 + \textbf{b}$ . One component, $\textbf{g} / v_0 = (\textbf{e}_0 \triangledown_0) \circ \textbf{a} + \nabla \circ a$, is relevant to the acceleration, while the other, $\textbf{b} = \nabla \times \textbf{a}$ , is associated to the precession angular velocity. $f^0$ is real, and $f^r$ is a complex number.
The gravitational source, $\mathbb{S}_g ( i s^0 , s^1 , s^2 , s^3 )$ , is defined as,
\begin{equation}
- \mu \mathbb{S} = - ( \mu_g \mathbb{S}_g - i \mathbb{F}_g^* \circ \mathbb{F}_g / v_0) = ( \lozenge + i \mathbb{F}_g / v_0)^* \circ \mathbb{F}_g ~,
\end{equation}
or
\begin{equation}
- \mu_g \mathbb{S}_g = \lozenge^* \circ \mathbb{F}_g ~,
\end{equation}
where $- \mu_g \mathbb{S}_g = \lozenge^* \circledcirc \mathbb{F}_g + \lozenge^* \circledast \mathbb{F}_g $ . The scalar part of $\mathbb{S}_g$ is, $- \lozenge^* \circledcirc \mathbb{F}_g / \mu_g = i s^0 \textbf{e}_0$ , which is associated to the mass density. And the vector part of $\mathbb{S}_g$ is, $- \lozenge^* \circledast \mathbb{F}_g / \mu_g = s^r \textbf{e}_r$ , which is relevant to the density of linear momentum. $\mu$ and $\mu_g$ both are the coefficients. $s^j$ is real.
Considering the complex coordinate and definition of gravitational strength, it is possible to expand the above to achieve the four component equations of gravitational field in the complex quaternion curved space (Appendix C). Now the gravitational field equations deal with some spatial parameters of the complex quaternion curved space. For the first-rank contravariant tensor $Y^n$ in the complex quaternion curved space, the complex quaternion physics quantity $\lozenge Y^n$ is associated to the connection coefficient $\Gamma^n_{jk}$ . Meanwhile the complex quaternion physics quantities, $\lozenge^* \circ (\lozenge Y^n)$ and $\lozenge \circ (\lozenge^* Y^n)$, contain the connection coefficient $\Gamma^n_{jk}$ and the curvature $R_{j\overline{k}m}^{~~~~n}$ . As a result, choosing the gravitational potential as the first-rank contravariant tensor, the gravitational strength will be relevant to the connection coefficient $\Gamma^n_{jk}$ , while the gravitational source will be associated to the connection coefficient $\Gamma^n_{jk}$ and the curvature $R_{j\overline{k}m}^{~~~~n}$ .
In the tangent space, in order to facilitate comparison among the field equations in different coordinate systems, the gravitational field equations in the above may be firstly transformed from the orthogonal and unequal-length coordinate system into the orthogonal and equal-length coordinate system. Subsequently comparing the gravitational field equations in the orthogonal and equal-length coordinate system with that in the flat space, it is possible to appraise the bending degree of the complex quaternion curved space.
Certainly the deduction approach of the gravitational field equations in the complex quaternion curved space can be used as a reference to be extended to that of electromagnetic field equations.
\begin{table}[h]
\caption{The multiplication of the operator with the physics quantity of gravitational and electromagnetic fields, in the complex octonion curved space.}
\label{tab:table1}
\centering
\begin{tabular}{ll}
\hline\hline
definition & expression~meaning \\
\hline
$\nabla \cdot (\textbf{e}_r a^r)$ & $(\textbf{e}_1 \cdot \textbf{e}_1) \triangledown_1 a^1
+ (\textbf{e}_2 \cdot \textbf{e}_2) \triangledown_2 a^2
+ (\textbf{e}_3 \cdot \textbf{e}_3) \triangledown_3 a^3 $ \\
$\nabla \times (\textbf{e}_r a^r)$ & $(\textbf{e}_2 \times \textbf{e}_3) ( \triangledown_2 a^3 - \triangledown_3 a^2 )
+ (\textbf{e}_3 \times \textbf{e}_1) ( \triangledown_3 a^1 - \triangledown_1 a^3 )$ \\
$$ & ~~~ $ + (\textbf{e}_1 \times \textbf{e}_2) ( \triangledown_1 a^2 - \triangledown_2 a^1 )$ \\
$\nabla \circ (\textbf{e}_0 a^0) $ & $(\textbf{e}_1 \circ \textbf{e}_0) \triangledown_1 a^0
+ (\textbf{e}_2 \circ \textbf{e}_0) \triangledown_2 a^0
+ (\textbf{e}_3 \circ \textbf{e}_0) \triangledown_3 a^0 $ \\
$(\textbf{e}_0 \triangledown_0) \circ (\textbf{e}_r a^r)$ & $(\textbf{e}_0 \circ \textbf{e}_1) \triangledown_0 a^1
+ (\textbf{e}_0 \circ \textbf{e}_2) \triangledown_0 a^2
+ (\textbf{e}_0 \circ \textbf{e}_3) \triangledown_0 a^3 $ \\
$\nabla \cdot (\textbf{E}_r A^r)$ & $ (\textbf{e}_1 \cdot \textbf{E}_1) \triangledown_1 A^1
+ (\textbf{e}_2 \cdot \textbf{E}_2) \triangledown_2 A^2
+ (\textbf{e}_3 \cdot \textbf{E}_3) \triangledown_3 A^3 $ \\
$\nabla \times (\textbf{E}_r A^r)$ & $ (\textbf{e}_2 \times \textbf{E}_3) ( \triangledown_2 A^3 - \triangledown_3 A^2 )
+ (\textbf{e}_3 \times \textbf{E}_1) ( \triangledown_3 A^1 - \triangledown_1 A^3 )$ \\
$$ & ~~~ $ + (\textbf{e}_1 \times \textbf{E}_2) ( \triangledown_1 A^2 - \triangledown_2 A^1 )$ \\
$\nabla \circ (\textbf{E}_0 A^0)$ & $ (\textbf{e}_1 \circ \textbf{E}_0) \triangledown_1 A^0
+ (\textbf{e}_2 \circ \textbf{E}_0) \triangledown_2 A^0
+ (\textbf{e}_3 \circ \textbf{E}_0) \triangledown_3 A^0 $ \\
$(\textbf{e}_0 \triangledown_0) \circ (\textbf{E}_r A^r)$ & $ (\textbf{e}_0 \circ \textbf{E}_1) \triangledown_0 A^1
+ (\textbf{e}_0 \circ \textbf{E}_2) \triangledown_0 A^2
+ (\textbf{e}_0 \circ \textbf{E}_3) \triangledown_0 A^3 $ \\
\hline\hline
\end{tabular}
\end{table}
\section{Complex-octonion curved space}
Under an orthogonal transformation of the coordinate system with the complex octonions, the norms of the complex octonions remain unchanged. And the norm includes the space-time interval (or the arc length) in the physics, and is possible to be applied to describe simultaneously the gravitational field equations and the electromagnetic field equations in the complex octonion curved space. By means of the complex octonion orthogonality and affine frame, it is possible to define the octonion metric and covariant derivative, depicting the gravitational and electromagnetic properties in the complex octonion curved space.
What needs to be explained specially is, that the gravitational field involved in is the quaternion operator and physics quantity in the complex quaternion space. Meanwhile what the electromagnetic field concerned with is, the physics quantity in the complex $S$-quaternion space, and the quaternion operator in the complex quaternion space. Consequently the paper deals with one comparatively simple situation. That is, the paper will be only involved in the interval (or the arc length) of the complex quaternion space.
In the complex octonion curved space, the underlying space is the octonion space, and the tangent space is the octonion space also. A tangent frame system (complex quaternion) can discuss the gravitational field, while the other tangent frame system (complex $S$-quaternion) may research the electromagnetic field, in the curved space.
Some calculations in the context may be necessary to be dealt with the non-associativity of the octonions, for instance, some multiplications in the Appendices B, D, and Table 3. One can multiply the next argument by an existing octonion from the left-side, step by step, to ensure the existence and uniqueness of the product relevant to physical quantities.
\subsection{Octonion metric}
In the complex quaternion flat space, the complex quaternion radius vector is, $\mathbb{H}_g = i \textbf{\emph{i}}_0 h^0 + \textbf{\emph{i}}_r h^r $, with the basis vector being ${\emph{\textbf{i}}_j}$ . Meanwhile, in the complex $S$-quaternion flat space for the electromagnetic field, the complex $S$-quaternion radius vector is, $\mathbb{H}_e = i \textbf{\emph{I}}_0 H^0 + \textbf{\emph{I}}_r H^r $ , with the basis vector being ${\emph{\textbf{I}}_j}$ . In the complex octonion flat space, two radius vectors, $\mathbb{H}_g$ and $\mathbb{H}_e$ , can be combined together to become one complex octonion radius vector,
\begin{equation}
\mathbb{H} (h^\alpha) = \mathbb{H}_g + k_{eg} \mathbb{H}_e = i h^0 \emph{\textbf{i}}_0 + h^r \emph{\textbf{i}}_r + i h^4 \emph{\textbf{i}}_4 + h^{4+r} \emph{\textbf{i}}_{4+r} ~,
\end{equation}
where $h^{j+4} = k_{eg} H^j$ , $\emph{\textbf{i}}_{j+4} = \emph{\textbf{I}}_j$ . $\mathbb{H}_g = i h^0 \emph{\textbf{i}}_0 + h^r \emph{\textbf{i}}_r$ , $\mathbb{H}_e = i H^0 \emph{\textbf{I}}_0 + H^r \emph{\textbf{I}}_r$ . The coefficient $k_{eg}$ meets the demand for the dimensional homogeneity in the physics, according to the Appendix A. $h^j$ and $H^j$ are all real. $\emph{\textbf{i}}_0^2 = 1$, $\emph{\textbf{i}}_v^2 = - 1$. $\emph{\textbf{i}}_\alpha^2 = \emph{\textbf{i}}_\alpha \circ \emph{\textbf{i}}_\alpha$, for each subscript $\alpha$. $\alpha, \beta, \gamma, \lambda = 0, 1, 2, 3, 4, 5, 6, 7$. $v = 1, 2, 3, 4, 5, 6, 7$.
According to the multiplication of octonion, the norm $S$ is written as, $S^2 = \mathbb{H} \circ \mathbb{H}^*$. The differential, $d S$, of the norm is able to be chosen as the arc length of the complex octonion (rather than the real-octonion) curved space. Obviously, in case the contribution of $k_{eg} \mathbb{H}_e$ can be neglected, this norm will accord with the requirement of space-time interval in the physics. Herein $\circ$ and $*$ are upgraded to the octonion multiplication and conjugate respectively.
In the complex quaternion curved space for the gravitational field, the complex quaternion radius vector is, $\mathbb{H}_g = i \textbf{\emph{e}}_0 c^0 + \textbf{\emph{e}}_r c^r $, with the tangent-frame quaternion being ${\textbf{e}_j}$. In the complex $S$-quaternion curved space for the electromagnetic field, the complex $S$-quaternion radius vector is, $\mathbb{H}_e = i \textbf{\emph{E}}_0 C^0 + \textbf{\emph{E}}_r C^r $ , with the tangent-frame $S$-quaternion being ${\textbf{E}_j}$. Therefore, in the tangent space of complex octonion curved space, the complex octonion radius vector is, $\mathbb{H} = \mathbb{H}_g + k_{eg} \mathbb{H}_e$. Making use of the substitution, $c^{j+4} = k_{eg} C^j$ and $\textbf{e}_{j+4} = \textbf{E}_j$ , the complex octonion radius vector can be rewritten as, $\mathbb{H} (c^\alpha) = i c^0 \textbf{e}_0 + c^r \textbf{e}_r + i c^4 \textbf{e}_4 + c^{4+r} \textbf{e}_{4+r}$ . Herein $c^j$ and $C^j$ are all real.
In the orthogonal and unequal-length affine frame, the metric of complex octonion curved space is defined as,
\begin{equation}
d S^2 = d \mathbb{H}^* \circ d \mathbb{H} = g_{\overline{\alpha} \beta} d \overline{u^\alpha} d u^\beta ~,
\end{equation}
where the metric coefficient, $g_{\overline{\alpha} \beta} = \textbf{e}_\alpha^* \circ \textbf{e}_\beta$ , is octonion-Hermitian. The orthogonal tangent-frame octonion is, $\textbf{e}_\alpha = \partial \mathbb{H} / \partial u^\alpha$ . $\textbf{e}_0$ is the scalar part, while $\textbf{e}_v$ is the component of vector part. $\textbf{e}_\beta$ is unequal-length, and does not contain the imaginary unit $i$ . $u^0 = i c^0$ , $u^r = c^r$ , $u^4 = i c^4$ , and $u^{4+r} = c^{4+r}$ . $ ( u^\alpha )^* = \overline{u^\alpha} $ , and it denotes that the correlated tangent-frame component, $\textbf{e}_\alpha$ , is octonion-conjugate. $g_{\overline{\alpha} \beta}$ is scalar.
However, in the complex octonion curved space, what the paper will be involved in is, the mathematical manipulation between the quaternion operator (rather than the octonion operator) in the complex quaternion space, with the physics quantity of electromagnetic field in the complex $S$-quaternion space. So that the discussion of the space-time interval in the paper will be constrained to deal only with the component of the complex quaternion radius vector, $\mathbb{H}_g$ .
\begin{table}[h]
\caption{Comparison of major characteristics in some flat and curved spaces, including the pseudo-Riemannian space, quaternion space, and octonion space.}
\label{tab:table2}
\centering
\begin{tabular}{lllll}
\hline \hline
Terms & pseudo- & Quaternion & Octonion \\
& Riemannian space & space & space \\
\hline
tangent space & vector space & quaternion space & octonion space \\
orthogonality & vector & quaternion & octonion \\
parallel translation & Levi-Civita & quaternion & octonion \\
tangent frame & vector/scalar & quaternion & octonion \\
metric & scalar product & scalar product & scalar product \\
& & ~~ of quaternions & ~~ of octonions \\
connection coefficient & $\Gamma^n_{jk}$ & $\Gamma^n_{jk}$ & $\Gamma^\beta_{\alpha \gamma}$ \\
covariant derivative & $\nabla_k A^j$ & $\nabla_k A^j$ & $\nabla_\gamma A^\beta$ \\
\hline \hline
\end{tabular}
\end{table}
\subsection{Octonion parallel translation}
In the complex octonion curved space, they are not suitable enough that the normal parallel translation (Euclidean space), the Levi-Civita parallel translation (Riemannian space), and the quaternion parallel translation (quaternion curved space). One of reasons is that the spaces in the physics associate with not only the complex octonion flat space (tangent space) but also the pseudo-Riemannian space (curved space). Subsequently it is necessary to introduce one new concept of parallel translation. It is similar to the quaternion parallel translation, and is able to meet the demand for the complex octonion space and of the space-time interval.
In one complex octonion space, when the product of two complex octonions, $\mathbb{G} ( g^\alpha )$ and $\mathbb{Z} ( z^\beta )$ , is equal to zero, that is, $\mathbb{G}^* \circ \mathbb{Z} = 0$ , two octonions, $\mathbb{G} (g^\alpha)$ and $\mathbb{Z} (z^\beta)$ , are perpendicular to each other. This definition is called as the octonion orthogonality. Therefore it is able to define an octonion connection and parallel translation in the complex octonion curved space, referring to the inference procedure of quaternion parallel translation in the complex quaternion curved space. Herein $g^\alpha$ and $z^\beta$ are all scalar.
In the complex octonion curved space, the complex octonion physics quantity $\mathbb{A}_1$ , in the tangent space $\mathbb{T}_1$ of one point $M_1$ on the complex octonion manifold, can be decomposed in the tangent space $\mathbb{T}_2$ of the point $M_2$ near $M_1$. According to the definition of octonion orthogonality, $\mathbb{A}_1$ can be separated into the projection component $\mathbb{A}_2$ in $\mathbb{T}_2$ , and the orthogonal component $\mathbb{G}_2$ perpendicular to $\mathbb{T}_2$ . In case the differential, $\mathbb{A}_2 - \mathbb{A}_1 = 0$, the physical quantity $\mathbb{A}_2$ is the parallel translation of $\mathbb{A}_1$ . And this definition is called as the octonion parallel translation. Especially, when the scalar parts of $\mathbb{A}_1$ and of $\mathbb{A}_2$ are all null, $\mathbb{A}_1$ and $\mathbb{A}_2$ both will be degenerated into the vectors. Further the orthogonality of octonion is reduced to that of vector, and the octonion parallel translation to the Levi-Civita parallel translation. When the complex octonions are reduced to the complex quaternions, the orthogonality of octonion is degenerated into that of quaternion, and the octonion parallel translation into the quaternion parallel translation (Table 2).
\subsection{Octonion covariant derivative}
In the complex octonion curved space, substituting the octonion parallel translation, tangent-frame octonion, and octonion orthogonality for the quaternion parallel translation, tangent-frame quaternion, and quaternion orthogonality respectively, the covariant derivative of the complex quaternion curved space can be extended into that of the complex octonion curved space directly.
For the first-rank contravariant tensor $Y^\beta$ of a point $M_2$ in the complex octonion curved space, the component of octonion covariant derivative with respect to the coordinate $u^\gamma$ is,
\begin{equation}
\triangledown_\gamma Y^\beta = \partial ( \delta_\alpha^\beta Y^\alpha ) / \partial u^\gamma + \Gamma^\beta_{\alpha \gamma} Y^\alpha ~ ,
\end{equation}
where $\Gamma^\beta_{\alpha \gamma}$ is the connection coefficient. One can deduce the connection coefficient, from the Appendix B. $Y^\beta$ and $\Gamma^\beta_{\alpha \gamma}$ are all scalar.
\section{Electromagnetic field equations}
In the complex octonion curved space, the electromagnetic potential and gravitational potential had been transformed from the rectangular coordinate system (flat space) to the orthogonal and unequal-length affine frame (tangent space). From the octonion integrating function, $\mathbb{X}$, of field potential, the octonion field potential, $\mathbb{A} = \mathbb{A}_g + k_{eg} \mathbb{A}_e$ , is defined as,
\begin{equation}
\mathbb{A} = i \lozenge^\star \circ \mathbb{X} ~,
\end{equation}
where $\mathbb{A} = i \lozenge^\star \circledcirc \mathbb{X} + i \lozenge^\star \circledast \mathbb{X}$ . $\mathbb{X} = \mathbb{X}_g + k_{eg} \mathbb{X}_e$ . The integrating function, $\mathbb{X}_e$, of electromagnetic potential is one $S$-quaternion physics quantity, and $\mathbb{X}_e = X^j \textbf{E}_j$. The electromagnetic potential is $\mathbb{A}_e ( i A^0 , A^1 , A^2 , A^3 ) = i \lozenge^\star \circ \mathbb{X}_e$ , with $\mathbb{A}_e = i \lozenge^\star \circledcirc \mathbb{X}_e + i \lozenge^\star \circledast \mathbb{X}_e$ . $i \lozenge^\star \circledcirc \mathbb{X}_e$ and $i \lozenge^\star \circledast \mathbb{X}_e$ denote respectively the `scalar' and vector parts of $\mathbb{A}_e$ . The gauge equation is chosen as, $\nabla \times (X^r \textbf{E}_r) = 0$. $\mathbb{A}_e = i \textbf{A}_Q + \textbf{A}$. $\textbf{A}_Q = A^0 \textbf{E}_0$, and $\textbf{A} = A^r \textbf{E}_r$. Apparently, in the affine frame, the electromagnetic potential includes not only the physics quantity (in the complex $S$-quaternion curved space) but also the spatial parameter of curved space (in the complex quaternion curved space). $a^j$ and $A^j$ are all real.
The octonion field strength, $\mathbb{F} = \mathbb{F}_g + k_{eg} \mathbb{F}_e$ , is defined as,
\begin{equation}
\mathbb{F} = \lozenge \circ \mathbb{A} ~ ,
\end{equation}
where $\mathbb{F} = \lozenge \circledcirc \mathbb{A} + \lozenge \circledast \mathbb{A} $ . The electromagnetic strength is, $\mathbb{F}_e ( i F^0 , F^1 , F^2 , F^3 ) = \lozenge \circ \mathbb{A}_e$. $\mathbb{F}_e = \lozenge \circledcirc \mathbb{A}_e + \lozenge \circledast \mathbb{A}_e $. The `scalar' part of $\mathbb{F}_e$ is, $\lozenge \circledcirc \mathbb{A} = i F^0 \textbf{E}_0$ , and the vector part of $\mathbb{F}_e$ is, $\lozenge \circledast \mathbb{A}_e = F^r \textbf{E}_r$ . The gauge equation of electromagnetic potential is chosen as, $F^0 = 0$. The vector part of electromagnetic strength can be separated into two components, $F^r \textbf{E}_r = i \textbf{E} / v_0 + \textbf{B}$. One component, $\textbf{E} / v_0 = (\textbf{e}_0 \triangledown_0) \circ \textbf{A} + \nabla \circ \textbf{A}_Q$, is the electric field intensity, while the other, $\textbf{B} = \nabla \times \textbf{A}$, is the magnetic flux density. $F^0$ is real, and $F^r$ is one complex number.
The octonion field source, $\mu \mathbb{S} = \mu_g \mathbb{S}_g + k_{eg} \mu_e \mathbb{S}_e$ , is defined as,
\begin{equation}
- \mu \mathbb{S} = - ( \mu_g \mathbb{S}_g + k_{eg} \mu_e \mathbb{S}_e - i \mathbb{F}^* \circ \mathbb{F} / v_0) = ( \lozenge + i \mathbb{F} / v_0)^* \circ \mathbb{F} ~,
\end{equation}
or
\begin{equation}
- \mu_g \mathbb{S}_g = \lozenge^* \circ \mathbb{F}_g ~, ~~~ - \mu_e \mathbb{S}_e = \lozenge^* \circ \mathbb{F}_e ~,
\end{equation}
where $- \mu_g \mathbb{S}_g = \lozenge^* \circledcirc \mathbb{F}_g + \lozenge^* \circledast \mathbb{F}_g$ , while $- \mu_e \mathbb{S}_e = \lozenge^* \circledcirc \mathbb{F}_e + \lozenge^* \circledast \mathbb{F}_e$ . The electromagnetic source is $\mathbb{S}_e ( i S^0 , S^1 , S^2 , S^3 )$. The `scalar' part of $\mathbb{S}_e$ is, $- \lozenge^* \circledcirc \mathbb{F}_e / \mu_e = i S^0 \textbf{E}_0$, and is associated to the density of electric charge. And the vector part of $\mathbb{S}_e$ is, $- \lozenge^* \circledast \mathbb{F}_e / \mu_e = S^r \textbf{E}_r$ , and is relevant to the density of electric current. $\mu$ and $\mu_e$ are the coefficients. $S^j$ is real.
Considering the complex coordinate and definition of the electromagnetic strength, it is able to expand the above to achieve the four component equations of electromagnetic field in the complex $S$-quaternion curved space (Appendix D). Therefore the field equations deal with some spatial parameters of the complex octonion curved space. For the first-rank contravariant tensor $Y^\beta$ in the complex octonion curved space, the complex octonion physics quantity $\lozenge Y^\beta$ is associated to the connection coefficient $\Gamma^\beta_{\alpha \gamma}$ , while the complex octonion physics quantities, $\lozenge^* \circ (\lozenge Y^\beta)$ and $\lozenge \circ (\lozenge^* Y^\beta)$ , are relevant to the connection coefficient $\Gamma^\beta_{\alpha \gamma}$ and the curvature $R_{\alpha \overline{\gamma} \lambda}^{~~~~\beta}$ . As a result, choosing the electromagnetic (or gravitational) potential as the first-rank contravariant tensor, the electromagnetic (or gravitational) strength will be relevant to the connection coefficient $\Gamma^\beta_{\alpha \gamma}$, while the electromagnetic (or gravitational) source will be associated to the connection coefficient $\Gamma^\beta_{\alpha \gamma}$ and curvature $R_{\alpha \overline{\gamma} \lambda}^{~~~~\beta}$ .
In the $S$-quaternion tangent space, in order to facilitate comparison among the electromagnetic field equations in different coordinate systems, the electromagnetic field equations in the above may be firstly transformed from the orthogonal and unequal-length coordinate system into the orthogonal and equal-length coordinate system. Subsequently comparing the electromagnetic field equations in the orthogonal and equal-length coordinate system with that in the flat space, it is able to estimate the bending degree of the complex $S$-quaternion curved space. In the process of providing a contrast for the electromagnetic field equations, it is necessary to take into account the influence of the bending degree of complex quaternion curved space on the electromagnetic field equations. As a result, it is capable of measuring the bending degree of the complex octonion curved space, making use of contrasting the octonion field equations in the curved space with that in the flat space.
\begin{table}[h]
\caption{Some definitions of the physics quantity relevant to the gravitational and electromagnetic fields in the complex octonion curved space.}
\label{tab:table3}
\centering
\begin{tabular}{lll}
\hline\hline
physics~quantity & definition \\
\hline
radius~vector & $\mathbb{H} = \mathbb{H}_g + k_{eg} \mathbb{H}_e $ \\
integral~function & $\mathbb{X} = \mathbb{X}_g + k_{eg} \mathbb{X}_e $ \\
field~potential & $\mathbb{A} = i \lozenge^\star \circ \mathbb{X} $ \\
field~strength & $\mathbb{F} = \lozenge \circ \mathbb{A} $ \\
field~source & $\mu \mathbb{S} = - ( i \mathbb{F} / v_0 + \lozenge )^* \circ \mathbb{F} $ \\
linear~momentum & $\mathbb{P} = \mu \mathbb{S} / \mu_g $ \\
angular~momentum & $\mathbb{L} = ( \mathbb{H} + k_{rx} \mathbb{X} )^\star \circ \mathbb{P} $ \\
octonion~torque & $\mathbb{W} = - v_0 ( i \mathbb{F} / v_0 + \lozenge ) \circ \mathbb{L} $ \\
octonion~force & $\mathbb{N} = - ( i \mathbb{F} / v_0 + \lozenge ) \circ \mathbb{W} $ \\
\hline\hline
\end{tabular}
\end{table}
\section{Octonion angular momentum}
In the complex octonion curved space for the gravitational and electromagnetic fields, the octonion linear momentum, $\mathbb{P} ( p^j , P^j ) = \mathbb{P}_g + k_{eg} \mathbb{P}_e$, is defined from the octonion field source,
\begin{equation}
\mathbb{P} = \mu \mathbb{S} / \mu_g ~,
\end{equation}
where the component of the octonion linear momentum, $\mathbb{P}$ , in the complex quaternion space is, $\mathbb{P}_g = \{ \mu_g \mathbb{S}_g - ( i \mathbb{F} / v_0 )^* \circ \mathbb{F} \} / \mu_g = i p^0 \textbf{e}_0 + p^q \textbf{e}_q$ . And the component of the octonion linear momentum, $\mathbb{P}$ , in the complex $S$-quaternion space is, $\mathbb{P}_e = \mu_e \mathbb{S}_e / \mu_g = i P^0 \textbf{E}_0 + P^q \textbf{E}_q$. $p^j$ and $P^j$ are all real.
Subsequently, in the complex octonion curved space, the octonion angular momentum, $\mathbb{L} ( l^j , L^j ) = \mathbb{L}_g + k_{eg} \mathbb{L}_e$, can be defined from the octonion linear momentum and radius vector,
\begin{equation}
\mathbb{L} = ( \mathbb{H} + k_{rx} \mathbb{X} )^\star \circ \mathbb{P} ~,
\end{equation}
where $\mathbb{L} = ( \mathbb{H} + k_{rx} \mathbb{X} )^\star \circledcirc \mathbb{P} + ( \mathbb{H} + k_{rx} \mathbb{X} )^\star \circledast \mathbb{P}$ . The component of the octonion angular momentum, $\mathbb{L}$ , in the complex quaternion space is, $\mathbb{L}_g = ( \mathbb{H}_g + k_{rx} \mathbb{X}_g )^\star \circ \mathbb{P}_g + k_{eg}^2 ( \mathbb{H}_e + k_{rx} \mathbb{X}_e )^\star \circ \mathbb{P}_e = l^0 \textbf{e}_0 + l^q \textbf{e}_q$ . And the component of the octonion angular momentum, $\mathbb{L}$ , in the complex $S$-quaternion space is, $\mathbb{L}_e = ( \mathbb{H}_g + k_{rx} \mathbb{X}_g )^\star \circ \mathbb{P}_e + ( \mathbb{H}_e + k_{rx} \mathbb{X}_e )^\star \circ \mathbb{P}_g = L^0 \textbf{E}_0 + L^q \textbf{E}_q $. $l^j$ and $L^j$ are complex numbers.
In the above, the term, $( \mathbb{H} + k_{rx} \mathbb{X} )^\star \circledcirc \mathbb{P} = l^0 \textbf{e}_0 $, denotes the scalar part of the octonion angular momentum, $\mathbb{L}$ . While the term, $( \mathbb{H} + k_{rx} \mathbb{X} )^\star \circledast \mathbb{P} = l^q \textbf{e}_q + k_{eg} ( L^0 \textbf{E}_0 + L^q \textbf{E}_q )$, indicates the vector part of the octonion angular momentum, $\mathbb{L}$. The real part of the term, $l^q \textbf{e}_q$ , is relevant to the angular momentum. The term, $L^0 \textbf{E}_0$ , is associated with the `scalar' magnetic moment. For the term, $L^q \textbf{E}_q$ , its imaginary part is connected with the electric dipole moment, while its real part is connected with the magnetic dipole moment.
\section{Octonion torque}
In the complex octonion curved space, the octonion torque, $\mathbb{W} ( w^j , W^j ) = \mathbb{W}_g + k_{eg} \mathbb{W}_e$ , is defined from the octonion linear momentum,
\begin{equation}
\mathbb{W} = - v_0 ( \lozenge + i \mathbb{F} / v_0) \circ \mathbb{L} ~,
\end{equation}
where $\mathbb{W} = v_0 ( \lozenge + i \mathbb{F} / v_0) \circledcirc \mathbb{L} + v_0 ( \lozenge + i \mathbb{F} / v_0) \circledast \mathbb{L}$ . The component of the octonion torque, $\mathbb{W}$ , in the complex quaternion space is, $\mathbb{W}_g = - ( i \mathbb{F}_g \circ \mathbb{L}_g + i k_{eg}^2 \mathbb{F}_e \circ \mathbb{L}_e + v_0 \lozenge \circ \mathbb{L}_g ) = w^0 \textbf{e}_0 + w^q \textbf{e}_q$ . And the component of the octonion torque, $\mathbb{W}$, in the complex $S$-quaternion space is, $\mathbb{W}_e = - ( i \mathbb{F}_g \circ \mathbb{L}_e + i \mathbb{F}_e \circ \mathbb{L}_g + v_0 \lozenge \circ \mathbb{L}_e ) = W^0 \textbf{E}_0 + W^q \textbf{E}_q$. $w^j$ and $W^j$ are complex numbers.
In the above, the term, $v_0 ( \lozenge + i \mathbb{F} / v_0) \circledcirc \mathbb{L} = w^0 \textbf{e}_0$, denotes the scalar part of the octonion torque, $\mathbb{W}$ . And the term, $v_0 ( \lozenge + i \mathbb{F} / v_0) \circledast \mathbb{L} = w^q \textbf{e}_q + k_{eg} (W^0 \textbf{E}_0 + W^q \textbf{E}_q)$, indicates the vector part of the octonion torque, $\mathbb{W}$ . For the term, $w^0 \textbf{e}_0$ , its imaginary part is connected with the energy, while its real part is connected with the divergence of the angular momentum. Meanwhile for the term, $w^q \textbf{e}_q$, its imaginary part is associated with the force, and its real part is associated with the curl of the angular momentum. The real part of the term, $W^0 \textbf{E}_0$ , is relevant to the divergence of the magnetic dipole moment. And the part of the term, $W^q \textbf{E}_q$ , is dealt with the curl of the magnetic dipole moment and the derivative of the electric dipole moment and so on.
In the complex octonion curved space, when the octonion field potential is chosen as the first-rank tensor, the octonion field source, linear momentum, and angular momentum will be involved in the connection coefficient, curvature, and other spatial parameters of the complex octonion curved space. It means that the curved space will act on the octonion torque, including the energy, the torque, the divergence and curl of angular momentum, the divergence and curl of magnetic dipole moment, and the derivative of the electric dipole moment.
\section{Octonion force}
In the complex octonion curved space, the octonion force, $\mathbb{N} ( n^j , N^j ) = \mathbb{N}_g + k_{eg} \mathbb{N}_e$, is defined from the octonion torque,
\begin{equation}
\mathbb{N} = - ( \lozenge + i \mathbb{F} / v_0) \circ \mathbb{W} ~,
\end{equation}
where $\mathbb{N} = ( \lozenge + i \mathbb{F} / v_0) \circledcirc \mathbb{W} + ( \lozenge + i \mathbb{F} / v_0) \circledast \mathbb{W} $ .
For the octonion force, the component, $\mathbb{N}_g = n^0 \textbf{e}_0 + n^q \textbf{e}_q $ , in the complex quaternion space is,
\begin{equation}
\mathbb{N}_g = - ( i \mathbb{F}_g \circ \mathbb{W}_g / v_0 + \lozenge \circ \mathbb{W}_g + i k_{eg}^2 \mathbb{F}_e \circ \mathbb{W}_e / v_0 ) ~,
\end{equation}
and the component, $\mathbb{N}_e = N^0 \textbf{E}_0 + N^q \textbf{E}_q $ , in the complex $S$-quaternion space is,
\begin{equation}
\mathbb{N}_e = - ( i \mathbb{F}_g \circ \mathbb{W}_e / v_0 + \lozenge \circ \mathbb{W}_e + i \mathbb{F}_e \circ \mathbb{W}_g / v_0 ) ~,
\end{equation}
where $n^j$ and $N^j$ are complex numbers.
In the above, the term, $( \lozenge + i \mathbb{F} / v_0) \circledcirc \mathbb{W} = n^0 \textbf{e}_0$ , denotes the scalar part of the octonion force, $\mathbb{N}$. And the term, $( \lozenge + i \mathbb{F} / v_0) \circledast \mathbb{W} = n^q \textbf{e}_q + k_{eg} (N^0 \textbf{E}_0 + N^q \textbf{E}_q)$, indicates the vector part of the octonion force, $\mathbb{N}$ . The real part of the term, $n^0 \textbf{e}_0$ , is relevant to the mass continuity equation, while the real part of the term, $N^0 \textbf{E}_0$, is associated with the current continuity equation. For the term, $n^q \textbf{e}_q$ , its imaginary part is connected with the force, and corresponded with the linear acceleration. And that its real part is corresponded with the precession angular velocity.
In the complex octonion curved space, the imaginary part of $(n^r \textbf{e}_r) / 2$ is the force, in the gravitational and electromagnetic fields, that is,
\begin{equation}
\textbf{N} = Im \{ (n^r \textbf{e}_r) / 2 \} ~ ,
\end{equation}
where the force $\textbf{N}$ includes the inertial force, gravitational force, electromagnetic force, energy gradient force, and additional force term caused by the curved space. The additional force term is relevant to the connection coefficient and curvature and so forth of the complex octonion curved space.
A majority of force terms can be written approximately as,
\begin{eqnarray}
\textbf{N}_M \approx && - \nabla \circ (p^0 v_0 \textbf{e}_0) - ( \textbf{e}_0 \triangledown_0 ) \circ ( \textbf{p} v_0)
\nonumber \\
&& + ( \textbf{g} / v_0 ) \circ ( p^0 \textbf{e}_0 ) - \textbf{b} \times ( p^q \textbf{e}_q )
\nonumber \\
&& + ( \textbf{E} / v_0 ) \circ ( P^0 \textbf{E}_0 ) - \textbf{B} \times ( P^q \textbf{E}_q ) ~,
\end{eqnarray}
where $- \nabla \circ (p^0 v_0 \textbf{e}_0)$ is the energy gradient. $- ( \textbf{e}_0 \triangledown_0 ) \circ ( \textbf{p} v_0)$ is the inertial force. $ \{ ( \textbf{g} / v_0 ) \circ ( p^0 \textbf{e}_0 ) - \textbf{b} \times ( p^q \textbf{e}_q ) \}$ is the gravitational force. $ \{ ( \textbf{E} / v_0 ) \circ ( P^0 \textbf{E}_0 ) - \textbf{B} \times ( P^q \textbf{E}_q ) \}$ is the electromagnetic force. The product, $\textbf{E}_j \circ \textbf{E}_k$, belongs to the complex quaternion curved space, according to the multiplication of octonion.
When the octonion field potential is chosen as the first-rank tensor in the complex octonion curved space, the octonion linear momentum, angular momentum, and torque will be involved in the some spatial parameters of the complex octonion curved space (Table 3). It means that the curved space has an influence on the octonion force, including the force, the mass continuity equation, and the current continuity equation.
In general, the force $\textbf{N}$ consists of two components, in the complex octonion curved space. One is the force term $\textbf{N}_{flat}$ in the flat space, while the other is the additional force term $\textbf{N}_{curved}$ , caused by the curved space. Under most circumstances, the bending degree of curved space is quite tiny. As a result, the additional force term $\textbf{N}_{curved}$ is weak enough, while the force term $\textbf{N}_{flat}$ dominates the force $\textbf{N}$. Additionally, under some extreme conditions, the additional force term $\textbf{N}_{curved}$ may be strong enough, and even account for the main part.
\section{Conclusions}
In the complex quaternion curved space, from the definitions of quaternion metric, orthogonality, parallel translation, and covariant derivative, it is able to deduce the field potential, field strength, field source, linear momentum, angular momentum, torque, and force and so on in the gravitational field. The force includes the inertial force, gravitational force, and additional force term caused by the curved space. The connection coefficient and curvature of the curved space will act on the additional force term. When the gravitational potential is chosen as the first-rank tensor, the connection coefficient impacts the gravitational strength, while the curvature affects the gravitational strength and source.
In the complex octonion curved space, from the definitions of octonion metric, orthogonality, parallel translation, and covariant derivative, it is able to infer the octonion field potential, field strength, field source, linear momentum, angular momentum, torque, and force and so on in the gravitational and electromagnetic fields. The force consists of the inertial force, gravitational force, electromagnetic force, energy gradient, and additional force term caused by the complex octonion curved space. The connection coefficient and curvature of the complex octonion curved space may impact the additional force term. The study reveals that one may appraise the deviation amplitude of the complex octonion curved space departure from its flat space, by means of the measurements of the field potential, field strength, and force and so forth.
Specially, in the complex $S$-quaternion curved space for the electromagnetic field, it is capable of deducing the electromagnetic potential, strength, and source, and the additional force term caused by the curved space. The connection coefficient and curvature of the complex $S$-quaternion curved space have an influence on the additional force term. If the electromagnetic potential is the first-rank tensor, the connection coefficient of the complex $S$-quaternion curved space impacts the electromagnetic strength, while the curvature of the complex $S$-quaternion curved space affects the electromagnetic strength and source. Moreover the connection coefficient and curvature of the complex quaternion curved space will act on the electromagnetic strength and source as well.
It should be noted that the paper discussed only some simple cases about the influences of the complex octonion curved space on the field potential, field strength, and force and so forth. However it clearly states that the connection coefficient, curvature and other spatial parameters of the curved space exert an influence on the physical properties of gravitational and electromagnetic fields. In the following study, it is going to explore the impact of the force, in the strong gravitational and electromagnetic fields, on the movement statues of one charged objective in the complex octonion curved space. Moreover, it may intend to describe the gravitational and electromagnetic theories with any kind of frame, and then transform the field theories into that in the orthogonal affine frame by means of the appropriate transformation.
\section*{Acknowledgements}
The author is indebted to the anonymous referee for their valuable and constructive comments on the previous manuscript. This project was supported partially by the National Natural Science Foundation of China under grant number 60677039.
|
\section{Introduction}
State-space models are a powerful tool for handling nonlinear, non-Gaussian time series. This general class of models is widely used in many fields, including finance, ecology, biology and engineering. Over the last few decades, Sequential Monte Carlo (SMC) methods have become extremely popular for sequential state and parameter estimation in state-space models. These methods, however, have been largely ignored for Bayesian smoothing (i.e., retrospective analysis). Smoothing presents computational challenges because the target posterior distribution is often high-dimensional and intractable. In this paper, we propose two new SMC algorithms that overcome these challenges.
Markov chain Monte Carlo (MCMC) methods are the most common approach to Bayesian smoothing. \cite*{Carlin:1992} introduced the first MCMC approach for nonnormal and nonlinear models. \cite*{Carter:1994} and \cite*{Fruhwirth:1994} proposed the forward-filtering, backward-sampling (FFBS) algorithm and \cite*{Piet:1995} introduced the related simulation smoother for conditionally Gaussian models. The FFBS is an efficient block sampler that draws the states jointly given the parameters for linear, Gaussian state-space models. \cite{Shephard:1997} and \cite{Gamerman:1998} provided block sampling algorithms for non-Gaussian and exponential family measurement models, respectively. \cite*{Geweke:2001} proposed Metropolis-within-Gibbs algorithms for nonlinear and non-Gaussian state-space models, and \cite*{Stroud:2003} proposed a block sampling algorithm for nonlinear models with state-dependent variances. \cite*{Niemi2010} solved the nonlinear and non-normal case by sequential approximation of filtering and smoothing densities using normal mixtures.
Particle filtering is a sequential Monte Carlo method that has also been widely used for state estimation in state-space models and has been successful in many simulation studies and real data problems. The idea was first introduced by \cite*{Gordon:1993} with the name \quotes{bootstrap filter.} Then, \cite*{Pitt:1999} improved this by introducing the auxiliary particle filter. However, the problem of dealing with unknown parameters in Sequential Monte Carlo methods is not fully resolved. \cite*{Kitagawa:1998} proposed including the parameters into the state vector and proposed a particle filter on the augmented state vector. On the other hand, \cite*{West:2001} use a kernel smoothing density for the static parameters to avoid over-dispersion problems. This filter algorithm remains to be the most general method for sequential state and parameter estimation. Both \cite*{Storvik:2002} and \cite*{Fearnhead:2002} discussed generating samples of the parameters from the filtering distribution in situations where sufficient statistics for $\theta$ are available. In this case, the samples of parameters simulated at time $t$ do not depend on values simulated at previous times and the problem of impoverishment is mitigated. A comprehensive review of parameter estimation for state-space models was recently given by \cite*{kantas:2015}, in which both maximum likelihood methods and Bayesian methods were discussed.
In addition to the filtering problem, in which state estimation is conditional on the data available at time $t$, Sequential Monte Carlo methods can also be applied to state smoothing. In smoothing problems, we estimate the states conditional on all the observations. \cite*{Kitagawa:1996} introduced the idea of smoothing by storing the state vector, in which the smoothing process is realized by resampling the filtered particles within the smoothing window. But as time evolves and the smoothing window width increases in size, the smoothing samples at the start of the time series will degenerate to a single path. Other smoothing algorithms include: the forward-backward smoother of \cite*{Godsill:2004}, in which a backward recursion is included and the forward filter particles are reweighted; the two-filter smoother of \cite*{Kitagawa:1996}; the generalized two-filter smoother of \cite*{Doucet:2010}; and the new $O(N)$ and $O(N^2)$ smoothing algorithms of \cite*{Fearnhead:2010}.
All of the sequential Monte Carlo smoothing algorithms discussed above are based on the assumption that the fixed parameters are known. Research on particle smoothing with unknown parameters is limited. The particle learning and smoothing (PLS) algorithm of \cite*{Carvalho:2010} is one of the most well-known methods in this area. In their smoothing algorithm, however, the dependency between states and the parameters is ignored, which results in a failure of their smoothing algorithm at the beginning of the time series. In this paper, we take the dependency of state and parameters into consideration and adjust the resampling weights in the backward pass. In addition, we propose a new smoothing algorithm, in which we apply a forward-backward smoother on each parameter drawn from the last filter step and get a corresponding smoother sample. This provides smoothed samples of the states while accounting for parameter uncertainty.
There are three main advantages of our approach relative to existing methods. First, our refiltering algorithm is the only Sequential Monte Carlo method to provide an ``exact" solution to the Bayesian smoothing problem as the number of particles goes to infinity. Second, our smoothing algorithms can be easily parallelized, since communication between processors is minimal. Third, unlike MCMC approaches, marginal likelihood and Bayes' factors can be accurately computed at each time \citep{Carvalho:2010}, which is useful for sequential model comparison and model selection.
In addition, we find empirical evidence that the posterior dependence between states and parameters decreases as time $t$ goes to infinity. This suggests the possibility of new algorithms for on-line Bayesian state and parameter estimation that exploit this independence.
\newpage
This article is organized as follows. In Section 2, we give a brief review for particle filtering and smoothing algorithms. Two new smoothing algorithms are proposed in Section 3. In Section 4, the two new smoothing algorithms and PLS are tested on three models: an AR(1) plus noise model with three unknown parameters, a nonlinear growth model with five unknown parameters and a chaotic model with three unknown parameters. Finally, in Section 5, a real data smoothing problem is presented by modeling daily S\&P 500 index returns with a stochastic volatility model.
\section{Filtering and Smoothing with SMC}
Consider a general state-space model defined at discrete times $t=1,\ldots,T$:
\begin{eqnarray*}
\text{Initial} : ~ x_0 & \sim & p(x_0 |} %\;|\; \theta),\\
\text{Evolution} : ~ x_t & \sim & p(x_t |} %\;|\; x _{t-1},\theta),\\
\text{Observation} : ~ y_t & \sim &p(y_t |} %\;|\; x_t,\theta),
\end{eqnarray*}
where $y_t$ is the observation, $x_t$ is the hidden state, and $\theta$ are the model parameters. The Bayesian model is completed with a prior distribution, $\theta \sim p(\theta)$. The state-space model is characterized by two properties: (1) the states $x_t$ follow a first-order Markov process; and (2) the observations are conditionally independent given the states.
In a Bayesian framework, the objective is to compute the joint posterior distribution of the states and parameters, $p(x_t,\theta|y^s)$, where $y^s=(y_1,\ldots,y_s)$ denotes the observations up to time $s$. When $s=t$, this is called the filtering problem; and when $s=T$, this is called the smoothing problem. In most models, the joint posterior distribution is unavailable in closed form, and we rely on Monte Carlo methods to sample from the filtering and smoothing distributions. The goal of this paper is to draw samples from the joint smoothing distribution $p(x^T,\theta|y^T)$.
Traditionally, Sequential Monte Carlo methods assume that $\theta$ is known and are designed to approximate $p(x_t|} %\;|\; y^t,\theta)$ with a set of weighted samples or particles. In comparison to MCMC methods, SMC avoids convergence problems and allows for efficient calculation of marginal likelihoods, which is useful in parameter estimation or model selection problems.
The subsections below give a brief review of sampling importance resampling (SIR) particle filters, particle filters with unknown parameters, and the particle learning and smoothing algorithm. Our new smoothing algorithms are formulated based on this previous work.
\subsection{Particle Filtering}
The particle filter was first introduced by \cite*{Gordon:1993} to conduct state estimation in nonlinear/non-Gaussian state-space models. Based on importance sampling, we simply propagate the particles $x_{t-1}^{(i)}$ forward through the system equation and resample the new particles $\tilde{x}_t^{(i)}$ with weights $\omega_t^{(i)}$ proportional to the likelihood $p(y_t|} %\;|\; \tilde{x}_t^{(i)})$ to get filtered particles at time $t$: $x_t^{(i)}$. The filtering density $p(x_t |} %\;|\; y^t)$ can then be approximated by the empirical density of these particles.
\begin{equation}
\begin{split}
p(x_t|} %\;|\; y^t)
& \propto p(y_t|} %\;|\; x_t)\int p(x_t|} %\;|\; x_{t-1})p(x_{t-1}|} %\;|\; y^{t-1})dx_{t-1}\\
& \approx p(y_t|} %\;|\; x_t) \sum_{i=1}^{N} p(x_t|} %\;|\; x_{t-1}^{(i)}) \omega_t^{(i)}.
\end{split}
\end{equation}
\subsection{Particle Filtering with Unknown Parameters}
To deal with particle filtering with unknown parameters, \cite*{Kitagawa:1998} introduced the idea of augmenting the state by the parameters as $z_t=(x_t,\theta)'$, then applying a bootstrap filter to the augmented state vector $z_t$. Moreover, \cite*{Kitagawa:2001} proposed to add noise to the parameters in the transition density to avoid the collapse of samples as time progresses.
\cite*{West:2001} proposed an improvement to Kitagawa's method by drawing samples of the parameter from a smoothing kernel density of the form:
$$
p(\theta_{t+1} |} %\;|\; y^t) \approx \sum_{j=1}^{N} w_t^{(j)} \mathcal{N}(\theta_{t+1}|} %\;|\; m_t^{(j)},h^2V_t),
$$
at each filter step $t$, in which $m_t^{(j)} = a \theta_t^{(j)} + (1 - a)\bar{\theta_t}$, where $\bar{\theta}_t$ and $V_t$ are the sample mean and variance-covariance matrix of the posterior samples of $\theta_t$ at time $t$, and $a=\sqrt{1 - h^2}$ is a smoothing parameter between 0 and 1. Notice that $a=1$ implies the evolution equation $\theta_{t+1}=\theta_t$, which corresponds to state augmentation with no evolution noise. With this method, we have $V(\theta_{t+1}|} %\;|\; y^t) = V(\theta_{t}|} %\;|\; y^t)$ and thus, no information is lost over time.
In situations where the posterior distribution of $\theta$ depends on sufficient statistics that are easy to update recursively, the methods from \cite{Storvik:2002} can be applied to draw samples from its filtered distribution. We include sufficient statistics for $\theta$ into the state vector and draw samples of $\theta$ based on sufficient statistics at each time point $t$ in the filtering process. By doing this, the impoverishment problem is mitigated, and the true value of $\theta$ can be learned gradually through a filter process. The approach is based on the decomposition:
\begin{equation}
p(x^t,\theta|} %\;|\; y^t)
\propto p(x^{t-1}|} %\;|\; y^{t-1}) p(\theta |} %\;|\; s_{t-1}) p(x_t|} %\;|\; x_{t-1},\theta)p(y_t|} %\;|\; x_t,\theta)
\end{equation}
in which $s_t$ are the sufficient statistics for $\theta$. The details are listed below:
\vspace{.7cm}
\noindent \underline{\textsc{\Large Storvik's SIR Filter}}
\noindent For each time $t=1,\ldots, T$:
\begin{enumerate}
\item Sample $\theta^{(i)} \sim p(\theta |} %\;|\; s_{t-1}^{(i)})$ (for $i=1,\ldots,N$).
\item Propagate $x_t^{(i)} \sim p(x_t|} %\;|\; x_{t-1}^{(i)},\theta^{(i)})$ (for $i=1,\ldots,N$).
\item Compute weights $\omega_t^{(i)} \propto p(y_t|} %\;|\; x_t^{(i)},\theta^{(i)})$ (for $i=1,\ldots,N$).
\item Update sufficient statistics $s_t^{(i)} = S(x_t^{(i)},s_{t-1}^{(i)},y_t)$ (for $i=1,\ldots,N$).
\item Resample $N$ times from $\{(x_t^{(i)},s_t^{(i)})\}_{i=1}^N$ with weights $\omega_t^{(i)}$, to obtain a sample from $p(x_t,s_t|} %\;|\; y^t)$.
\end{enumerate}
\subsection{Particle Learning and Smoothing (PLS)}
In particle smoothing with unknown parameters, we are interested in estimating the states and parameters conditional on the whole data $y^T$ and drawing samples $(x^{T(i)},\theta^{(i)})$ from the joint posterior $p(x^T,\theta |} %\;|\; y^T)$, where $T$ denotes the number of time steps.
\cite{Carvalho:2010} showed that a backward pass can be run after the filtering and learning algorithm, and the filtered particles could be resampled to obtain draws from the smoothing distribution. The idea is based on Bayes' Rule and the decomposition of the joint posterior smoothing distribution as
\begin{equation}
p(x^T,\theta |} %\;|\; y^T) = p(x_T,\theta|} %\;|\; y^T) \prod_{t=1}^{T-1} p(x_t|} %\;|\; x_{t+1},\theta,y^t),
\end{equation}
where
\begin{equation}
p(x_t|} %\;|\; x_{t+1},\theta,y^t) \propto p(x_{t+1}|} %\;|\; x_t,\theta) p(x_t|} %\;|\; \theta,y^t).
\end{equation}
The steps of this algorithm are listed below.
\vspace{.5cm}
\noindent \underline{\textsc{\Large{PLS Algorithm}}}
\begin{enumerate}
\item {\bf (Forward Filter)} Run the particle learning algorithm to generate samples $\{(x_t^{(i)},\theta^{(i)})\}_{i=1}^N$ from $p(x_t,\theta|} %\;|\; y^t)$ at each time $t=1,\ldots,T$.
\item {\bf (Backward Smoother)} Select a pair $(x_T^{(i)},\theta^{(i)})$ from Step 1, and simulate backwards: For $t$ = $T-1,\ldots,1$, resample the particles $\{x_t^{(j)}\}_{j=1}^N$ from Step~1 with weights proportional to $\omega_t^{(j)} = p(x_{t+1}^{(i)}|} %\;|\; x_t^{(j)},\theta^{(i)})$ to generate $x_t^{(i)}$.
\end{enumerate}
According to these authors, this algorithm is an extension of \cite*{Godsill:2004} to state-space models with unknown parameters. However, this is not the case. Note that in the backward pass, we select a fixed $\theta^{(i)}$ first and evaluate the filter weights proportional to $p(x_{t+1}^{(i)}|} %\;|\; x_t^{(j)},\theta^{(i)})$. Thus, correspondingly, we should use samples drawn from $p(x_t|} %\;|\; \theta^{(i)},y^t)$, i.e. the filter samples with respect to this fixed $\theta^{(i)}$. But this is not the case for PLS.
Moreover, the particles in the filter process are in fact coming from the marginal density $p(x_t|} %\;|\; y^t)$, not from the conditional density $p(x_t|} %\;|\; \theta,y^t)$. Reweighting particles using the transition density ignores the dependence between states and parameters, which causes inaccurate smoothing estimates when the dependency is strong. Figure~\ref{fig:AR-corr} shows the dependency between the filtered samples of the states and parameters for the AR(1) plus noise model presented in Section 4. Correlations greater than 0.5 can be detected at the beginning of the time series. The simulation studies presented in Section 4 show that PLS gives poor smoothing estimates, particularly at early periods in the time series.
In the next section, we present two new smoothing algorithms. The first relies on a transformation of equation (4) and an adjustment of the weights in the backward pass to refine PLS. The second involves a separate forward-backward pass conditional on the sampled parameters.
\begin{figure}
\centering
\includegraphics[scale = 0.4]{Figure9.pdf}
\caption{\footnotesize AR(1) plus Noise Model. Top: means and 95\% intervals for the posterior absolute correlation between state and AR coefficient $\phi$ at every time step $t$. Results are based on 500 simulated datasets. Middle and Bottom: posterior samples of $x_t$ and $\phi$ at selected time steps for one dataset. The contours represent the fitted normal densities used in the PLS$_a$ algorithm.}
\label{fig:AR-corr}
\end{figure}
\section{Two New Smoothing Algorithms}
\subsection{PLS with Adjustment (PLS$_a$)}
As stated earlier, the PLS algorithm assumes we have samples from the conditional distribution, $p(x_t|\theta,y^t)$ in the filtering algorithm, when in fact we have samples from the joint, $p(x_t,\theta|y^t)$, and hence the marginal, $p(x_t|y^t)$. Thus, the reweighting scheme in PLS does not give us samples from the target smoothing distribution. To provide a remedy for this, we consider the following rearrangement of equation (4):
\begin{equation}
\underbrace{p(x_t|} %\;|\; x_{t+1},\theta,y^t)}_{\text{smoother}}
\propto \underbrace{p(x_{t+1}|} %\;|\; x_t,\theta) \; \dfrac{ p( x_t|} %\;|\; \theta,y^t)}{p(x_t|} %\;|\; y^t)}}_{\text{weights}} \underbrace{p(x_t|} %\;|\; y^t)}_{\text{filter}}.
\end{equation}
With samples from the filter, we use $\omega_t^{(j)} = p(x_{t+1}|} %\;|\; x_t^{(j)},\theta) p( x_t^{(j)}|} %\;|\; \theta,y^t)/p(x_t^{(j)}|} %\;|\; y^t)$ as our resampling weights in the backward pass. Only in this way can we use the filtered particles in the smoothing algorithm. Note that, in most cases, we cannot compute these resampling weights exactly, since the joint filtering distribution $p(x_t,\theta|y^t)$ is generally not available in closed form. To fix this problem, we propose to use a multivariate normal approximation to $p(x_t,\theta|y^t)$ based on the filtered particles $\{(x_t^{(i)},\theta^{(i)})\}_{i=1}^N$, using appropriate transformations if necessary.
The algorithm proceeds exactly as in PLS, but with modified weights in the backward pass. The details of the particle learning and smoothing algorithm with adjustment (PLS$_a$) are presented below. Based on the simulation results in Section 4, we find that the adjustment of the weights matters: the adjusted version outperformed the original one significantly, especially in the beginning of the series, where the PLS usually has problems.
\vspace{.5cm}
\noindent \underline{\textsc{\Large{PLS$_a$ Algorithm}}}
\begin{enumerate}
\item {\bf (Forward Filter)} Run a filtering and learning algorithm to obtain samples $\{(x_t^{(i)},\theta^{(i)})\}_{i=1}^N$ from $p(x_t,\theta |} %\;|\; y^t)$ for $t=1,\ldots,T$. Use the filtered samples to construct a multivariate normal approximation at each time $t$:
$$
p(x_t,\theta |} %\;|\; y^t)
\approx \mathcal{N}\left(\begin{pmatrix} \mu_t^{x} \\ \mu_t^{\theta} \end{pmatrix},
\begin{pmatrix} \Sigma_t^{x} & \Sigma_t^{x\theta} \\ \Sigma_t^{\theta x} & \Sigma_t^{\theta} \end{pmatrix}\right).
$$
This implies that the marginal and conditional distributions are also normal:
$p(x_t|} %\;|\; y^t) \approx \mathcal{N}(\mu_t^{x},\Sigma_t^{x})$, and $p(x_t|} %\;|\; \theta,y^t) \approx \mathcal{N}(\mu_t^{x|} %\;|\; \theta},\Sigma_t^{x |} %\;|\; \theta}),$
where the conditional mean and covariance are given by the well-known formulas for multivariate normal distributions.
\item {\bf (Backward Smoother)} Select a pair $(x_T^{(i)},\theta^{(i)})$ from Step 1, and simulate backwards: For $t$ = $T-1,\ldots,1$, resample the $\{x_t^{(j)}\}_{j=1}^N$ from Step~1 with weights proportional to
$$\omega_t^{(j)} = p(x_{t+1}^{(i)}|} %\;|\; x_t^{(j)},\theta^{(i)})\left( \frac{ p( x_t^{(j)}|} %\;|\; \theta^{(i)},y^t)}{p(x_t^{(j)}|} %\;|\; y^t)} \right)
\approx p(x_{t+1}^{(i)}|} %\;|\; x_t^{(j)},\theta^{(i)})\left( \frac{\mathcal{N}(x_t^{(j)}|} %\;|\; \mu_t^{x |} %\;|\; \theta^{(i)}},\Sigma_t^{x |} %\;|\; \theta} )}{\mathcal{N}(x_t^{(j)}|} %\;|\; \mu_t^{x},\Sigma_t^{x})}\right)$$
to generate $x_t^{(i)}$.
\end{enumerate}
\subsection{Refiltering Smoothing Algorithm}
In addition to the the PLS$_a$ modification, we propose the following new smoothing algorithm. The idea is simple but proved to be efficient and accurate in simulation studies. The algorithm is based on the decomposition:
\begin{equation}
p(x^T,\theta|} %\;|\; y^T) = p(x^T |} %\;|\; y^T,\theta) \; p(\theta |} %\;|\; y^T).
\end{equation}
We run Storvik's forward filter, or more generally a filter method as in \cite{West:2001}, to get samples of the parameter at the last time step, i.e. $\theta^{(i)} \sim p(\theta|} %\;|\; y^T)$. Then for each $\theta^{(i)}$, we apply a forward-backward smoothing algorithm as in \cite{Godsill:2004} to get one state trajectory $x^{T(i)}$ from $p(x^T|} %\;|\; y^T,\theta^{(i)})$. Repeating this process for each $i$, we obtain states from the marginal smoothing density $p(x^T|} %\;|\; y^T)$.
Since the run time for the forward filter is negligible compared to the backward smoother, ($O(N)$ vs $O(N^2)$, respectively), in simulation studies, we found that this algorithm almost has the same speed as PLS, but with significant improvement in accuracy. The algorithm is:
\vspace{.5cm}
\noindent \underline{\textsc{\Large Refiltering Algorithm}}
\begin{enumerate}
\item {\bf (Forward Filter)} Use Storvik, Particle Learning or Liu \& West to run a forward filter and learning algorithm and generate $\theta^{(i)} \sim p(\theta|} %\;|\; y^T)$;
\item {\bf (Backward Smoother)} For each $\theta^{(i)}, i=1,\ldots,N_0$, run a forward-backward smoothing algorithm to get a sample $x^{T(i)} \sim p(x^T|} %\;|\; y^T,\theta^{(i)})$.
\end{enumerate}
Note that this algorithm has a complexity of $O(TN^2)$, the same as PLS. But we can make it an $O(TN)$ algorithm in two ways. The first one is that we can choose a small number of states $n_0 \ll N$ for the forward-backward smoother in step 2. The second is to use a small number of parameter draws of size $N_0$ to use in step 2. The simulation study showed that both methods make the algorithm run much faster with only a minor loss of accuracy.
In the case where the model is linear and Gaussian: $x_t = G_t x_{t-1} + w_t, \enspace w_t \sim \mathcal{N}(0,W)$;
$y_t = F'_t x_t + v_t, \enspace v_t \sim \mathcal{N}(0,V)$;
we can incorporate a forward filtering, backward sampling algorithm as in \cite{Carter:1994} and \cite{Fruhwirth:1994} into step 2 : we run a Kalman filter forward pass then generate a sample backwards based on equation $(3)$. Note $p(x_t |} %\;|\; y^{t-1}) \sim \mathcal{N}(a_t,R_t)$ is the prior and $p(x_t |} %\;|\; y^t) \sim \mathcal{N}(m_t,C_t)$ is the posterior of the state at each time point $t$, which depends on the parameters $\theta=(F_t,G_t,V,W)$.
\vspace{.5cm}
\noindent \underline{\textsc{\Large Refiltering with FFBS}}
\begin{enumerate}
\item {\bf (Filter)} Use Storvik, Particle Learning or Liu \& West to run a forward filter and learning algorithm and generate $\theta^{(i)} \sim p(\theta|} %\;|\; y^T), i=1,..,N$;
\item {\bf (Smoother)} For each $\theta^{(i)}, i=1,..,N_0$, run a Kalman filter and store prior and posterior moments $a_t,R_t,m_t,C_t$. Sample $x_T^{(i)} \sim \mathcal{N}(m_T,C_T)$. For $t = T-1$ to 1, sample $x_t^{(i)} \sim p(x_t|} %\;|\; x_{t+1}^{(i)},\theta^{(i)},y^t) = \mathcal{N}(h_t,H_T)$, in which $h_t = m_t + B_t(x_{t+1}^{(i)} - a_{t+1})$, $H_t = C_t - B_t R_{t+1} B'_t$ and $B_t = C_t G'_{t+1} R^{-1}_{t+1}$. This provides a sample, $x^{T(i)} \sim p(x^T|} %\;|\; y^T,\theta^{(i)})$.
\end{enumerate}
\section{Examples}
\subsection{AR(1) Model with Three Unknown Parameters}
Assume that the states $x_t$ follow an AR(1) process where the observations $y_t$ equal $x_t$ plus Gaussian noise:
$$x_t = \phi x_{t-1} + w_t, \enspace w_t \sim \mathcal{N}(0,W),$$
$$y_t = x_t + v_t, \enspace v_t \sim \mathcal{N}(0,V).$$
This benchmark model has been widely used in SMC and MCMC simulation studies \cite*[see, for example,][]{Storvik:2002,Polson:2008}.
In this model, FFBS can be easily implemented and a long chain MCMC with 150,000 iterations was set as a standard to compare with other smoothing algorithms. We generate $T=100$ observations with parameter values $V = W = 1$, $\phi = 0.75$ and $x_0=0$.
For the analysis, we assume conjugate priors for the parameters: $(\phi,W) \sim \mathcal{NIG}(b_0,B_0,n_0,d_0)$, and $V \sim \mathcal{IG}(\nu_0,\delta_0)$ where $\mathcal{IG}(a,b)$ denotes the inverse-gamma distribution with scale and shape parameters $a$ and $b$, and $\mathcal{NIG}$ denotes the normal-inverse gamma distribution where $B_0$ represents the inverse of the scale factor in the normal variance. We assume $n_0 = \nu_0 = d_0 = \delta_0 = 2$ and $b_0=0.5, B_0=1$. The conjugate model for the parameters allows us to use Storvik's algorithm, with the sufficient statistics $s_t=(b_t,B_t,n_t,d_t,\nu_t,\delta_t)$, and the updating recursions:
\begin{align*}
B_t &= B_{t-1} + x_{t-1}^2, &
b_t &= B_t^{-1}(B_{t-1}b_{t-1}+ x_{t-1}x_t),\\
n_t &= n_{t-1} + 1/2, &
d_t &= d_{t-1} + (b_{t-1}^2B_{t-1} + x_t^2 - b_t^2B_t)/2,\\
\nu_t &= \nu_{t-1} + 1/2, &
\delta_t &= \delta_{t-1} + (y_t - x_t)^2/2.
\end{align*}
We first run Storvik's filtering algorithm. Figure \ref{fig:AR-learn} shows the parameter learning plots and the posterior distribution at the last time period $T = 100$. From the plot, we notice the true parameters values were learned properly and the samples of the parameters at the last time step are well concentrated around the true parameter values. Also the samples from the filter agree well with samples from a long MCMC. State smoothing by refiltering and the result of a long MCMC are also presented in Figure \ref{fig:AR-learn}. We notice that the mean, $2.5^{th}$ and $97.5^{th}$ quantiles of the smoothing samples almost coincide for the two methods at each time step $t$.
\begin{figure}
\centering
\includegraphics[scale = 0.4]{Figure1.pdf}
\caption{\footnotesize AR(1) Plus Noise Model. Top: posterior mean and the $2.5^{th}$, $97.5^{th}$ quantiles based on the refiltering smoother (blue) and full MCMC (red). Middle: parameter learning in Storvik forward filter. Bottom: histograms of parameter samples at last filter step, superimposed with density estimation from long MCMC (red line). The true parameter values are indicated by horizontal and vertical red lines.}
\label{fig:AR-learn}
\end{figure}
To show that PLS$_a$ outperforms PLS using the same computation time, we ran 500 simulations for each of these two methods and compared the standardized absolute errors over time, i.e., $\hat{e}^*_t = |\hat{x_t} - \hat{x}^{true}_t|/\sigma(x_t|} %\;|\; y^T)$ for $t = 1 \dots T$, where $\hat{x}_t^{true}$ and $\sigma(x_t|y^T)$ are the smoothed mean and standard deviation for $x_t$ computed from the long MCMC, and $\hat{x}_t$ is the smoothed mean from the other algorithms. The result is shown in Figure \ref{fig:AR-MAE}. From the plot, we can see the main difference between the two smoothing algorithms appears at the beginning of the series, in which the dependence of states and parameters is strong and therefore the adjustment matters. As time progresses, the dependency of states and parameters decreases, and the adjusted smoothing outcomes coincide with PLS. The results from the refiltering smoothing algorithm are also shown in the plot. For this model, refiltering substantially outperforms the other two methods, and its accuracy is consistent over time. Note that the number of particles for the three smoothing algorithms was adjusted to assure similar computation time.
\begin{figure}
\centering
\includegraphics[scale = 0.35]{Figure2.pdf}
\caption{\footnotesize AR(1) Plus Noise Model. Standardized absolute errors over time for three smoothing algorithms compared to a long MCMC. The results are based on 500 simulations. The long MCMC is treated as the truth. The solid lines are the means of standardized absolute errors at time $t$ among 500 simulations, while the dashed lines represent the $95^{th}$ quantile.}
\label{fig:AR-MAE}
\end{figure}
To compare the performance of all of the smoothing algorithms in this paper, we implement long and short MCMC runs, PLS, PLS$_a$, refiltering, $O(TN)$ refiltering, and refiltering with FFBS using 500 data simulations. All SMC based smoothing algorithms are run in parallel on 16 cores on a single node. Based on a similar run time, the mean standardized absolute errors over time (MAE* = $\sum_{t=1}^T|\hat{e}^*_t|/T$) are listed in Table~\ref{table:AR-MAE}. From the table, we see that the MAE* values for PLS$_a$ are about half as large as for PLS. For all of the refiltering algorithms, the MAE* magnitude is only about one fifth of of that for PLS. Hence, both of the new smoothing algorithms outperform the PLS smoothing algorithm of \cite{Carvalho:2010}. The column labeled MAEP* represents the mean standardized absolute error between the posterior mean of the parameters at the last time step for a long MCMC versus the other algorithms. From the table, we see that the learning of parameters using the particle filter is almost as good as the learning from a short MCMC.
\begin{table}
\centering
\caption{Comparison of smoothing algorithms in AR(1) plus noise model. MAE* and MAEP* denote the standardized mean absolute error for the states
and parameters, respectively. SMC smoothers are based on Storvik's algorithm with $N=50,000$ particles.}
\begin{tabular}{lcccc}
\hline\hline
\rule{0pt}{2.5ex} Algorithm & $N (N_0/n_0)$ & Time & MAE* & MAEP* \\ \hline
\rule{0pt}{2.5ex} MCMC & 5000 & 17s & 0.019 & 0.051 \\
\rule{0pt}{2.5ex} PLS & 2300 & 22s & 0.138 & 0.058 \\
\rule{0pt}{2.5ex} PLS$_a$ & 1050 & 22s & 0.060 & 0.058 \\
\rule{0pt}{2.5ex} Refiltering & 1,500/1500 & 22s & 0.026 & 0.058 \\
\rule{0pt}{2.5ex} Refiltering & 10,000/150 & 22s & 0.022 & 0.058 \\
\rule{0pt}{2.5ex} Refiltering & 1,000/2500 & 23s & 0.031 & 0.058 \\
\rule{0pt}{2.5ex} Refiltering/FFBS & 44,000 & 21s & 0.015 & 0.058 \\
\hline\hline & \\
\end{tabular}
\label{table:AR-MAE}
\end{table}
\subsection{Nonstationary Growth Model with Five Unknown Parameters}
Consider the nonstationary growth model:
\begin{align*}
x_t &= \alpha x_{t-1} + \beta \frac{x_{t-1}}{1+x_{t-1}^2} + \gamma \cos(1.2(t-1)) + w_t,\\
y_t & = x_t^2/20 + v_t,
\end{align*}
in which $w_t \sim \mathcal{N}(0,W)$ and $ v_t \sim \mathcal{N}(0,V)$. This benchmark nonlinear time series model has been used by \cite{Carlin:1992} to test MCMC smoothing, by \cite{Gordon:1993} to test the bootstrap filter, and by \cite{Doucet:2010} to test the Forward-Backward smoothing with known parameters. Here we test our smoothing methods on this model with unknown parameters.
We generate $T=100$ observations using parameter values $\alpha = 0.5$, $\beta = 25$ , $\gamma = 8$, $V =5$ and $W = 1$. We assume conjugate priors for the parameters similar to those given in \cite{Carlin:1992}, i.e. $((\alpha,\beta,\gamma)',W) \sim \mathcal{NIG}(b_0,B_0,n_0,d_0)$, and $V \sim \mathcal{IG}(\nu_0,\delta_0)$, where $b_0 = (0.5,25,8)'$, $B_0 = \text{diag}(1/0.25^2,1/10^2,1/4^2)$, and $n_0 = \nu_0 = d_0 = \delta_0 = 2$. The conjugate priors allow us to use Storvik's algorithm for filtering and parameter learning. The updating recursions for the sufficient statistics are given by
\begin{align*}
B_t &= B_{t-1} + F_t F_t', &
b_t &= B_t^{-1} (B_{t-1}b_{t-1} + F_t x_t),\\
n_t &= n_{t-1} + 1/2, &
d_t &= d_{t-1} + (b_{t-1}'B_{t-1}b_{t-1} + x_t^2 - b_t'B_tb_t)/2,\\
\nu_t &= \nu_{t-1} + 1/2, &
\delta_t &= \delta_{t-1} + (y_t - x_t^2/20)^2/2,
\end{align*}
where $F_t=(x_{t-1},x_{t-1}/(1+x_{t-1}^2),\cos(1.2(t-1))$. The parameter learning process and the posterior histograms of the parameters at time $T=100$ are plotted in Figure \ref{fig:NL-learn}. In the figure, the 95\% confidence bands narrow down quickly as time increases. In the histograms, the samples concentrate around the true parameter values. A total of $N = 50,000$ particles were used for the forward pass.
Furthermore, we compare the refiltering smoothing algorithm using $N_0=5000$ and $n_0 = 1000$, with a long MCMC using $N = 150,000$ iterations. The smoothing plot is also presented in Figure \ref{fig:NL-learn}. The results from the two smoothing algorithms closely agree with each other.
\begin{figure}
\centering
\includegraphics[scale = 0.4]{Figure4.pdf}
\caption{\footnotesize Nonlinear Model.
Top: refiltering smoothing compared with MCMC.
Middle: parameter learning in Storvik forward filter.
Bottom: histogram of parameter samples at last filter step, superimposed with density estimation from long MCMC (red line).}
\label{fig:NL-learn}
\end{figure}
Table~\ref{table:NL-MAE} gives a summary of the overall performance of the three smoothing algorithms compared to a long MCMC, using 500 simulated datasets. A decrease in the mean absolute error for the new methods relative to PLS is obvious. The plot of the standardized absolute errors over time of the three smoothing algorithms (not shown) illustrates the same patterns as for the AR(1) model: the main improvement of the two new smoothing algorithms over PLS is evident at the beginning of the time series.
Note that for this model, it is difficult to distinguish between the positive and negative sign of the states based on the data, thus it is difficult to assign initial values for the states for the MCMC algorithm based on observations. With a bad starting values for the states, the MCMC chain takes much longer to converge. In contrast, smoothing based on SMC does not suffer from the initialization problem.
\begin{table}
\centering
\caption{Comparison of smoothing algorithms for the nonlinear growth model.
SMC smoothers are based on Storvik's algorithm with $N=50,000$ (run time 96s).}
\begin{tabular}{lcccc
\hline\hline
\rule{0pt}{2.5ex} Algorithm & $N (N_0/n_0)$ & Time & MAE* & MAEP* \\ \hline
\rule{0pt}{2.5ex} MCMC & 20000 & 228s & 0.075 & 0.246 \\
\rule{0pt}{2.5ex} PLS & 10000 & 200s & 0.373 & 0.213 \\
\rule{0pt}{2.5ex} PLS$_a$ & 5000 & 208s & 0.189 & 0.213 \\
\rule{0pt}{2.5ex} Refiltering & 5000/1000 & 231s & 0.097 & 0.213 \\
\hline\hline
\end{tabular}
\label{table:NL-MAE}
\end{table}
\subsection{Chaotic Model with Three Unknown Parameters}
Now let us consider data generated from the model:
\begin{align*}
N_{t} &= r N_{t-1} e^{-N_{t-1} + z_t}, \enspace z_t \sim \mathcal{N}(0,\sigma^2),\\
y_t &\sim Pois(\phi N_t).
\end{align*}
This model is widely used in the field of ecology \citep*{Fasiolo:2016}, where $N_t$ stands for the density of the population at generation $t$, and $r$ is the growth rate of the population. This model is characterized by its sensitivity to parameter variations: small increments in $r$ will lead to significant oscillations in the likelihood function. As a result, parameter estimation via maximum likelihood methods is challenging. \cite*{Fasiolo:2016} described the pathological likelihood function for this model and compared the performance of information reduction approaches and state-space methods for this model. A time series of 100 observations is generated from this model with true parameter values $r = e^{3.8}$, $\sigma^2=0.3$ and $\phi = 10$.
To estimate this model using our framework, we first make the transformations, $x_t = \log(N_t)$ and $\mu=\log(r)$. Then the system and observation equations become
\begin{align*}
x_t &= \mu + x_{t-1} - e^{x_{t-1}} + z_t,\\
y_t & \sim Pois(\phi e^{x_t}).
\end{align*}
We assume diffuse conjugate priors for the parameters of the form, $\phi \sim \mathcal{G}(a_0,b_0);$ and $(\mu, \sigma^2) \sim \mathcal{NIG}(m_0,c_0,n_0,d_0)$, where $\mathcal{G}$ denotes the gamma distribution, with $a_0=15, b_0=1, m_0=5, c_0=.1, n_0=2, d_0=2$. The sufficient statistics are $s_t=(a_t,b_t,m_t,c_t,n_t,d_t)$, and the updating recursions are
\begin{align*}
a_t & = a_{t-1} + y_t, &
b_t & = b_{t-1} + e^{x_t}, \\
c_t & = c_{t-1} + 1, &
m_t & = c_t^{-1} (c_{t-1}m_{t-1} + x_t - x_{t-1} + e^{x_{t-1}}),\\
n_t & = n_{t-1} + 1/2,&
d_t & = d_{t-1} + \{c_{t-1}m_{t-1}^2 + (x_t - x_{t-1} + e^{x_{t-1}})^2 - c_tm_t^2 \}/2.
\end{align*}
The parameter learning process is summarized in Figure \ref{fig:Chaos-learn}. A total of $N = 50,000$ particles were used for the filtering. The true parameter values were learned quickly and the posterior samples of the parameters settle around the true values. Figure \ref{fig:Chaos-learn} also provides a comparison of refiltering with $N_0=5000$ and $n_0=1000$ to a long MCMC with 150,000 iterations for smoothing, which shows similar results for both methods.
\begin{figure}
\centering
\includegraphics[scale = 0.4]{Figure7.pdf}
\caption{\footnotesize Chaotic Model.
Top: refilter smoothing compared with MCMC.
Middle: parameter learning in Storvik forward filter process.
Bottom: histogram of parameter samples at last filter step, superimposed with density estimation from long MCMC (red line). }
\label{fig:Chaos-learn}
\end{figure}
To allow a comparison of the three smoothing methods, 100 simulations were performed. We examined plots of the MAE* values over time (not shown), and Table~\ref{table:Chaos-MAE} presents numerical summaries based on the simulations. From the plots, we find similar patterns for this example as in the previous two: PLS$_a$ and refiltering dominate PLS early in the time series (up to about time $t=80$), and the three methods coincide afterwards. From Table~\ref{table:Chaos-MAE}, we notice a decrease in MAE* for the two new methods compared to PLS. We also find that smoothing method based on refiltering performs better compared to a short MCMC for this model.
\begin{table}
\centering
\caption{Comparison of smoothing algorithms for the chaotic model.
SMC smoothers are based on Storvik's algorithm with $N=50,000$ (run time 11s).}
\begin{tabular}{lcccc}
\hline\hline
\rule{0pt}{2.5ex} Algorithm & $N (N_0/n_0)$ & Time & MAE* & MAEP* \\ \hline
\rule{0pt}{2.5ex} MCMC & 30000 & 184s & 0.097 & 0.166 \\
\rule{0pt}{2.5ex} PLS & 10000 & 219s & 0.190 & 0.206 \\
\rule{0pt}{2.5ex} PLS$_a$ & 5000 & 180s & 0.108 & 0.206 \\
\rule{0pt}{2.5ex} Refiltering & 5000/1000 & 40s & 0.089 & 0.206 \\
\hline\hline
\end{tabular}
\label{table:Chaos-MAE}
\end{table}
\subsection{Analysis of S\&P 500 Returns}
In this section, we analyze daily returns on the S\&P 500 index from January 2008 to March 2009, during the financial crisis, and compare the PLS, PLS$_a$ and refiltering smoothers with MCMC where daily returns $y_t$ follow a stochastic volatility model:
\begin{align*}
x_{t} &= \alpha + \beta x_{t-1} + \omega_t, \enspace \omega_t \sim \mathcal{N}(0,W),\\
y_t &= \mu + \exp(x_t/2)v_t, \enspace v_t \sim \mathcal{N}(0,1).
\end{align*}
Here $y_t=\log(P_t/P_{t-1})$ are the daily returns, $P_t$ are the prices, $\mu$ is the expected return, and $x_t$ is the unobserved log-variance at time $t$, which is assumed to follow an AR(1) model with drift $\alpha$. The AR coefficient $\beta$ measures the autocorrelation present in the logged squared data. This model has been widely used to analyze financial time series with volatility clustering \cite*[see, for example,][]{Jacquier:1994,Kim:1998}.
We assume conjugate priors for the parameters $\theta=(\mu,\alpha,\beta,W)$. For the expected returns, $\mu \sim \mathcal{N}(a_0,b_0)$, and for the volatility parameters, we assume $((\alpha,\beta)',W) \sim \mathcal{NIG}(m_0,C_0,n_0,d_0)$, where $a_0=0, b_0=1, m_0=(0,.9)', C_0=\mbox{diag}(1,1), n_0=2, d_0=2$. The refiltering algorithm is implemented with $N=10,000$ and $n_0 = 1000$. The parameter learning and state smoothing estimates are compared to an MCMC with 15,000 iterations, using the single-state updating scheme of \cite{Jacquier:1994}.
From the sequential learning plots in Figure 6, a significant change in the parameters is observed in September 2008, especially for $\alpha$ and $\beta$. The change corresponds to the collapse of Lehman Brothers, and an increase in the volatility of the index. Figure 7 shows the filtered and smoothed volatilities for each algorithm. These plots show clear evidence that PLS and MCMC do not match, especially from September-November 2008, when the volatility changes abruptly. PLS$_a$ reduces this discrepancy somewhat, and among the three new smoothing algorithms, refiltering is by far the most accurate. Given that the run times for refiltering and MCMC are roughly the same, and the close match between the corresponding posterior distributions, we conclude that these two algorithms are comparable.
\begin{figure}
\centering
\includegraphics[scale = 0.8]{sv-param-bands-learn-3x2.pdf}
\caption{\footnotesize Stochastic Volatility Model.
Top row: Daily prices and returns on the S\&P 500 index from January 2008 to March 2009.
Middle and bottom rows: filtered medians and 95\% intervals for the parameters $\mu$, $\alpha$,
$\beta$ and $W$. Based on Storvik's algorithm with 50,000 particles.}
\label{fig:sv-params}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale = 0.65]{sv-state-bands-learn.pdf}
\caption{\footnotesize Stochastic Volatility Model.
Top left: filtered median and 95\% posterior intervals for $x_t$ from Storvik's algorithm.
The other three panels show the smoothed median and 95\% intervals for $x_t$
using MCMC and the PLS, PLS$_a$ and refiltering algorithms. The gray bands in
each plot are the 100\% filtering intervals (i.e., range) from Storvik's algorithm.}
\label{fig:sv-states}
\end{figure}
\section{Conclusions}
In this paper, we proposed two new SMC-based smoothing algorithms that simultaneously deal with parameter learning. The first is a modification of the PLS algorithm of \cite{Carvalho:2010}, that adds a correction term in the backward resampling weights. The second is a two-step algorithm, called {\em refiltering}, that includes a parameter learning step followed by a forward-backward algorithm for smoothing. Refiltering is well suited for parallel implementation, since the smoothing step requires essentially no communication between processors.
We tested the new methods on four models: a benchmark AR(1) plus noise model, a nonlinear growth model, a chaotic model from ecology, and a stochastic volatility model from finance, and compared the estimates with the widely-used smoothing method known as PLS. For all examples, both new methods showed significant improvement over PLS, and refiltering was competitive with MCMC. Overall, our proposed methods are quite general, and may be applied to a wide class of state-space models for parameter and state estimation. In future work, we plan to apply the methods to other real data applications in finance and ecology.
\newpage
\section*{Appendix A: Marginal Likelihood of the Model}
The marginal likelihood is important in Bayesian model selection. As noted by \cite{Carvalho:2010}, the marginal likelihood can be computed trivially from the output of SMC-based Bayesian filtering and learning algorithms \cite[e.g.,][]{Storvik:2002,Carvalho:2010}.
Define $\omega_t^{j} = p(y_t|x_t^{j},\theta^{j},\mathcal{M})$, where $(x_t^j,\theta^j) \sim p(x_t,\theta|y^{t-1},\mathcal{M})$ for given model $\mathcal{M}$. Then, the log marginal likelihood for model $\mathcal{M}$ is estimated by
\begin{equation*}
\log(f(y |} %\;|\; \mathcal{M})) \approx \sum_{t=1}^{T} \log\left(\sum_{j=1}^{N}\omega_t^{j}\right) - T \log(N).
\end{equation*}
By comparison, many different MCMC-based estimates of the marginal likelihood have been proposed. One of the most commonly used and straightforward methods is the harmonic mean estimator \citep{Newton:1994}, which can be computed based on the joint distribution of the data:
\begin{equation*}
\log(f(y |} %\;|\; \mathcal{M})) \approx \log\left(\dfrac{1}{\frac{1}{N}\sum_{j=1}^{N}\frac{1}{p(y|} %\;|\; \psi^j,\mathcal{M})}}\right),
\end{equation*}
where $y=(y_1,\ldots,y_T)$ are the observations, $N$ is the total number of MCMC iterations, and $\psi^j=(x^j,\theta^j), j=1,\ldots,N$ are the posterior draws of the states and parameters.
An important advantage of SMC over MCMC is that the estimation of marginal likelihood from SMC output is stable. As shown in Table 4, in our simulation studies, we find that the SMC-based marginal likelihood estimator converges quickly as $N$ increases, while for MCMC, the harmonic mean estimator fails to converge even for $N$ larger than 500K in all three models.
\begin{table}[h]
\centering
\caption{Comparison of Log Likelihood}
\begin{tabular}{l l l l l l l l l l}
\hline
\rule{0pt}{2.5ex} & & \multicolumn{2}{l}{ AR(1) + Noise } & & \multicolumn{2}{l}{ Nonlinear Growth } & & \multicolumn{2}{l}{ Chaotic Model } \\\cline{3-4}\cline{6-7}\cline{9-10}
\rule{0pt}{2.5ex}$N$ & & Storvik & MCMC & & Storvik & MCMC & & Storvik & MCMC \\ \hline\hline
\rule{0pt}{2.5ex}1K & & -176.44 & -164.40 & & -212.91 & -253.27 & & -286.34 & -193.37 \\
\rule{0pt}{2.5ex}5K & & -176.71 & -171.74 & & -212.51 & -169.92 & & -284.88 & -200.35 \\
\rule{0pt}{2.5ex}10K & & -176.71 & -169.88 & & -211.66 & -170.42 & & -284.96 & -195.64 \\
\rule{0pt}{2.5ex}50K & & -176.87 & -169.86 & & -212.45 & -170.91 & & -285.16 & -199.54 \\
\rule{0pt}{2.5ex}100K & & -176.91 & -170.84 & & -212.10 & -162.56 & & -284.98 & -196.61 \\
\rule{0pt}{2.5ex}500K & & -176.88 & -173.56 & & -212.30 & -164.06 & & -285.10 & -197.99 \\ \hline
\end{tabular}
\end{table}
\label{table:loglike}
\clearpage
\bibliographystyle{rss}
|
\section{Introduction}
\label{sec:intro}
Throughout the past few decades, problems of computer science have
become subjects of intense interest to theoretical physicists as
paradigms of complex systems that could benefit from theoretical
approaches and insights inspired by statistical physics. These include
neural networks, Boltzmann machines and deep learning, compressed
sensing, satisfiability problems, and a host of other approaches to
data mining and machine
learning~\cite{montanari_mezard,MPZ,ganguli,pankaj}. The interest in
the constraints on computation and information processing placed by
physical laws is even older and dates to work by Landauer and
Bennett~\cite{landauer1,landauer2,bennett}. One of the holy grails at
the interface between physics and computer science is the physical
realization of a large-scale quantum computer in which the processing
of information makes use of quantum-mechanical concepts such as
superposition and entanglement~\cite{nielsen-chuang,review}. However,
building a quantum computer remains a challenging task because of the
practical difficulty associated with maintaining coherence over the
duration of the computation.
This paper aims at bringing a new class of problems to the
physics-computer science interface by introducing a two-dimensional
(2D) representation of a generic reversible classical computation, the
result of which is encoded in the ground state of a statistical
mechanics vertex model with appropriate boundary conditions. The
vertex model is defined in terms of Boolean variables (or spins
degrees of freedom) placed on the bonds or links of an anisotropic 2D
lattice with vertices representing logic gates. The corresponding gate
constraints are implemented through short-ranged one- and two-body
interactions involving the spins of the vertex (as we show, this
construction can be realized in physical programmable machines, such
as the D-Wave machine.) One direction of the lattice represents
``computational (rather than real) time'', as introduced by Feynman in
the history representation of quantum computation~\cite{feynman}, but
here used for classical reversible circuits. The two boundaries of the
lattice transverse to the ``time'' direction contain the input and
output bits of the computation. It is important to stress that we are
not limiting ourselves to forward computations with fixed inputs. More
interesting are problems in which only partial information about both
inputs and outputs is known. In that case, reaching the ground state
requires flow of information both forwards and backwards across the
lattice, processes that are naturally built into our approach.
The idea of encoding classical computation in the ground state of a
many-body spin model was introduced earlier for irreversible
computation in Ref.~\cite{Biamonte2008,crosson2010,Biamonte2012}. Here
we focus on reversible rather than irreversible computation in order
to address problems with both fixed-input and mixed-boundary
conditions on inputs and outputs, as explained above. Mapping onto a
regular 2D lattice as opposed to an arbitrary graph allows us to use
intuitive ideas from equilibrium and non-equilibrium statistical
mechanics, especially of classical and quantum phase
transitions. Also, while in Ref.~\cite{crosson2010} an error
correction scheme was required to implement fault tolerant
computation, in our approach accurate computation without error
correction is possible below moderate temperatures that scale only as
the inverse of the logarithm of the system size, a consequence of the
exponential scaling of the static correlation length with inverse
temperature (see below).
Most importantly, the mapping proposed here defines statistical
mechanics vertex models that, irrespective of the computation they
represent, display no bulk thermodynamic transition down to zero
temperature. Thus our work emphasizes that the
dynamics of relaxation to the ground state rather than the
thermodynamics of the model is essential for understanding the
complexity of ground state computation.
The absence of a thermodynamical phase transition removes an obvious
impediment to reaching the ground state of the vertex model. For
instance, a suboptimal mapping from a computational problem into a
physical system may place the solution within a glassy phase, even in
the case of easy computational problems. The mapping of XORSAT (a
problem in P) into a diluted $p$-spin model is such an
example~\cite{Ricci-Tersenghi}. The fact that our vertex model is free
of thermodynamic transitions does not mean that the ground state can
be reached easily. This remains true even for problems with unique
solutions which are encoded by vertex models with unique ground
states. Such problems are in the complexity class UNIQUE-SAT, which
under randomized reduction is as hard as
SAT~\cite{Valiant-Vazirani}. Hence, even in the absence of a
thermodynamic transition finding the unique ground state of vertex
models encoding problems with a single solution is a problem in
NP-complete~\cite{gavey-johnson,papadimitriou,arora-barak}. Of course,
this does not mean that one cannot benefit from speed-ups allowed by
either physics inspired heuristics or by special-purpose physical
hardware, such as quantum annealers.
This paper focuses on the study of vertex-model representations of
random circuits for which the complexity of the computation is
reflected in the concentration of TOFFOLI gates, the length of the
input and output boundaries $L$, and the depth of the circuit $W$. We
concentrate on computational problems with a single solution -- or
problems for which one can discern among an ${\cal O}(1)$ number of
solutions with a small overhead -- a class of problems that encompass
factoring of semi-primes, an important and nontrivial example that we
shall explore in a future publication.
In our discussion of dynamics we deploy thermal annealing as well as
introduce a more efficient ``annealing with learning'' protocol. The
latter translates into an algorithm for solving classical problems for
which, as expected, forward computation from a fixed input boundary
reaches solution in a time linear in the depth of the computational
circuit. Finally, we note that reaching the ground state of the
vertex model could be accelerated by replacing classical annealing
with quantum
annealing~\cite{apolloni1989,finnila1990,kadowaki1998,farhi2001}. While
approaching computational problems through quantum annealing is left
for future investigations, the current paper includes the formal
derivation of the quantum version of the statistical mechanics model
of reversible classical computation. This provides the background for
an explicit mapping of our lattice model onto the Chimera architecture
of the D-Wave machine, a development that points to the potential
usefulness of the vertex model as a programming platform for special
purpose quantum annealers.
\section{Results}
\label{sec:Results}
\subsection{The vertex model for reversible classical computation}
\label{sec:tile-model}
Our starting point is the fact that any Boolean function can be
implemented in terms of TOFFOLI gates, which are reversible logic
gates with three inputs and three outputs. Starting from a circuit of
TOFFOLI gates, our construction proceeds by first using SWAP gates to
repeatedly swap distant bits in the input that are acted upon by
particular gates of the circuit, until the operation of every gate is
reduced to adjacent bits. The second step is to associate tiles with
each of the gates, as shown in Fig. ~\ref{fig:gate_tile_coupling},
where one should imagine placing input and output bits at the
intersections of the tile surfaces with the horizontal lines, as
described in detail in the Methods Section.
\begin{figure}[t]
\centering
\includegraphics[angle=0,origin=c,width=9cm]{Fig1.pdf}
\caption{{\bf Tile representation of reversible computational
gates.} (a)\&(b) Elementary tiles representing the three
computational gates for reversible circuits: ID (identity), SWAP and
TOFFOLI. (c) construction of the Hamiltonians that encode the
gate-satisfying states in the ground state manifolds. Spins are
placed on the boundary of the tiles. For the TOFFOLI gate, an
ancilla spin is placed in the center of the rectangular
tile. Couplings needed in the Hamiltonians for the three different
gates (tiles) are indicated by purple lines connecting two
spins. The dashed line denotes the boundary of the tile.}
\label{fig:gate_tile_coupling}
\end{figure}
The tiles representing the gates can then be laid down side-by-side on
a plane to implement the computational circuit, as shown in
Fig.~\ref{fig:ripple-carry} for the example of the ``ripple-carry
adder'', which computes the carry bit that is ``rippled'' to the next
bit when adding two numbers \cite{vedral}. (The ``ripple-carry adder''
is the building block for more complicated circuits such as addition
and multiplication.) As can be seen from this example, one may also
need to include the Identity (ID) gate in addition to the TOFFOLI and
SWAP gates in order to represent particular logic circuits via
tiling. Implied in the figure is that common boundaries of adjacent
tiles contain a pair of ``twin'' bits (one on each tile) whose values
must coincide. The derivations of spin Hamiltonians implementing
the truth tables of individual tiles, the short range inter-tile
Hamiltonian enforcing the consistency between bits of neighboring
tiles, and the boundary conditions specifying inputs and outputs
are presented in the Methods Section.
\begin{figure}[hbt]
\centering
\includegraphics[angle=0,origin=c,width=8.5cm]{Fig2.pdf}
\caption{{\bf Tile and vertex model representation of the
ripple-carry adder.} (a) The ripple-carry adder which computes the
carry bit that is ``rippled'' to the next bit. We add one additional
control line $s_n$ and set it to 1 to implement the original CNOT
gate with a TOFFOLI gate. (b) The ripple-carry adder implemented on
the tile lattice, with different gates depicted in different colors:
blue tile: ID; green tile: SWAP; gold tile: TOFFOLI. Spins between
adjacent tiles are forced to be equal by the ferromagnetic `grout'
coupling $K$. (c) The ripple-carry adder mapped to a vertex model
with periodic boundary condition in the transverse direction. After
each column of gate (vertex) operation, bit states are labeled at
each bond. Light yellow and grey stripes represent the $P$ and $T$
matrices used in the transfer matrix calculation of the partition
function.}
\label{fig:ripple-carry}
\end{figure}
The final step of our mapping, also detailed in the Methods Section,
is to construct a vertex model on a tilted square lattice, with each
vertex representing either a TOFFOLI gate or four possible rectangular
tiles obtained by combining square ID and SWAP tiles (ID-ID, ID-SWAP,
SWAP-ID, SWAP-SWAP), as shown in Fig.~\ref{fig:ripple-carry}. This
construction can always be done by an appropriate retiling of the
circuit so that each rectangular tile has four neighbors (hence the
square lattice). There are six Boolean (or spin) variables associated
to each vertex: two on each of the two double bonds and one on each of
the two single bonds tied to a vertex. In deriving the vertex model we
work in the limit in which the spin coupling defining the gate
Hamiltonians, $J\to \infty$ (see the Methods Section), in which case
all gate truth tables are satisfied exactly. Consequently, each vertex
can be in one of $r=2^3=8$ states. Three of the spins are inputs, and
we use the state $q$ of the vertex, where $q=0,1,\dots,7$, to read-off
the inputs in binary (which are uniquely related to the spin): $x^{\rm
IN}_a={\rm bit}[a,q]$, $a=1,2,3$ for the three bits of the number
$q$. The output bits are the bits of the 3-bit number $G(q)$, where
$G$ is the gate function: $x^{\rm OUT}_a={\rm bit}[a,G(q)]$,
$a=1,2,3$. The energy cost for two adjacent gates that are
incompatible with each other is determined by the ferromagnetic
coupling $K$.
The resulting vertex model Hamiltonian can be written as
\begin{eqnarray}
\label{eq:H-vertex}
\hat H&=& \sum_{\langle s s'\rangle} \sum_{q_s,q_{s'}}\;K^{g_s
g_{s'}}_{q_s,q_{s'}} \;|q_sq_{s'}\rangle\langle q_sq_{s'}|
\nonumber\\ &&+ \sum_{s\in {\rm boundary}}
\sum_{q_s}\;h_{q_s}\;|q_s\rangle\langle q_s| \nonumber\\ &&+
\sum_{s} \sum_{q_s,q'_{s}}\;\Delta_{q_s,q'_{s}}\;|q_s\rangle\langle
q'_s| \;,
\end{eqnarray}
where $K^{g_s g_{s'}}_{q_s,q_{s'}}$ encodes the energy cost for
mismatched nearest-neighbor vertices (the energies, with scale set by
$K$, depend on the state of the vertices $q_s$ and $q_{s'}$, as well
as on the types of gates $g_s$ and $g_{s'}$ present at neighboring
vertices $s,s'$ -- an explicit example is given in the Supplementary
note 2); $h_{q_s}$ encodes the boundary conditions, which we
associate directly with the vertex rather than with the input or
output bits of a gate (since the relationship is one-to-one); and
finally, the transition matrix elements $\Delta_{q_s,q'_s}$ between
the states within a vertex $s$. All these couplings can be determined
given a computational circuit and the boundary conditions. The quantum
term $\Delta_{q_s,q'_{s}}$ can be designed from the internal couplings
within the tiles; For simplicity, one should consider the case,
$\Delta_{q_s,q'_{s}}=\Delta$ for all $q_s,q_{s'}$, which then
represents the 8-state counterpart of a transverse field.
The vertex model defined by Eq. (\ref{eq:H-vertex}) is the starting
point for all the subsequent discussions of this paper. For example, a
quantum annealing protocol for solving a factoring problem would start
with $K\ll \Delta$, where the ground state is a superposition of all
locally satisfied gates independent of one another, and end with $K\gg
\Delta$, with the ground state in which each tile satisfies the gate
constraint and also passes and receives the right information to and
from its neighbors.
\subsection{The quantum vertex model phase diagram}
\label{sec:annealing_results}
Figure \ref{fig:phase-diagram} shows our conjectured equilibrium phase
diagram of the vertex model described by the Hamiltonian in
Eq.~(\ref{eq:H-vertex}). For $T,\Delta \ll J$, the local gate
constraints are satisfied, which we indicate by ``local SAT''. The
solution of the computational problem resides at the origin
($T/K=\Delta/K=0$), where all the gates are locally satisfied and
globally consistent, which we indicate by ``global SAT''. In the
Methods Section we show explicitly that along the classical axis,
$\delta=\Delta/K= 0$, the vertex model displays no finite temperature
bulk thermodynamic transition irrespective of the computational
circuit it represents. In particular, the resulting bulk thermodynamic
behavior is always that of a paramagnet:
\begin{equation}
\beta F =-\left[3L(W-1)\right]\; \ln (2\cosh \beta K) \;.
\end{equation}
\begin{figure}[h!]
\centering
\includegraphics[angle=0,origin=c,width=8cm]{Fig3.pdf}
\caption{{\bf Phase diagram of the vertex model.} Our exact calculation of the partition
function shows that there is no phase transition along the classical
path ($\delta = 0$). We argue that there should be a quantum phase
transition for some critical $\delta_{\rm c}$.}
\label{fig:phase-diagram}
\end{figure}
Moreover, along the ``quantum'' axis $T=0$ the vertex model must
encounter a zero-temperature quantum phase transition at a finite
value of $\delta$. This follows from considering trivial classical
circuits with no TOFFOLI gates in which case an $L\times W$ vertex
model is equivalent to $3L$ decoupled Ising chains of size $W$ in a
transverse magnetic field. Just as in the one-dimensional Ising model
in a transverse field, in the limit of no TOFFOLI gates one expects a
zero-temperature second-order quantum phase transition at $\delta_{\rm
c} = 1$. The addition of TOFFOLI gates complicates the analysis, but
on physical grounds we expect that the phase transition cannot simply
disappear but rather change character instead, possibly from second
order to first order. This could be the case if the no-TOFFOLI
critical point happens to be an endpoint of a phase boundary in the
$\delta$-$x_{\rm T}$ plane, where $x_{\rm T}$ is the concentration of
TOFFOLI gates. Determining the order of the transition for the vertex
model describing a generic computation is a difficult problem, which
we expect to address via quantum Monte Carlo simulations in a future
publication.
\subsection{Thermal annealing of the classical vertex model}
\label{sec:numerics}
Here we study the dynamics of relaxation to the ground state as a
function of the size and depth of the computation via thermal
annealing \cite{SA}. This proceeds by cooling the system from a high
temperature of order $K$ down to zero temperature over a total time
duration, $\tau$, according to the ramp protocol, $T(t)=K (1-t/\tau)$.
The dynamics is extracted by following an order parameter $m$
that measures the overlap of the final state $\{q^{\rm final}\}$
reached at $t=\tau$ with the reference (solution) state
$\{q^{\text{sol}}\}$:
\begin{equation}
m = \frac{8}{7} \left[\frac{1}{LW} \sum_s \delta_{q^{\rm
final}_s,q^{\text{sol}}_s} - \frac{1}{8}\right] \;.
\label{eqn:order}
\end{equation}
(Below we explain in detail how a unique solution state
$\{q^{\text{sol}}\}$ is obtained.) Notice that the order parameter
reaches $m=1$ when the final state agrees with the solution, and $m=0$
if the state is random, in which case it agrees with the solution by
chance in 1/8th of the sites. We remark that the ``solution overlap''
is a much better indicator of the evolution towards solution than the
total energy. This is because a single vertex flip into an incorrect
state in the middle of the circuit may cost little energy but it
throws other vertices into a completely different state from the
correct one.
\begin{figure*}[ht]
\centering
\includegraphics[angle=0,origin=c,width=18cm]{Fig4.pdf}
\caption{{\bf Scaling of the dynamical correlation length $\ell(\tau)$}. For all sizes and cases, 2000 realizations of the
boundary states were used. The data-point code used in panel (a)
applies to all other panels. When not visible, the error bars are
smaller than the size of the data points. (a)-(c) fixed input and
output; (d)-(f) mixed boundary conditions; (a)\&(d) 20\% TOFFOLI;
(b)\&(e) 40\% TOFFOLI; (c)\&(f) 100\% TOFFOLI. For systems with the
smallest depth $W$ studied, $16\times 42$, and for circuits with few
TOFFOLI gates (20\%) and fixed input and output boundary conditions,
$\ell(\tau)$ tends to saturate, indicating that complete solutions
have been reached. Notice that the functional form of the scaling
does not depend on the boundary conditions, and depends solely on
the concentration of TOFFOLI gates.}
\label{fig:MC_scaling_plot1}
\end{figure*}
\begin{figure}[h!]
\centering
\includegraphics[angle=0,origin=c,width=8.5cm]{Fig5a.pdf}
\includegraphics[angle=0,origin=c,width=8.5cm]{Fig5b.pdf}
\caption{{\bf Effects of boundary conditions and TOFFOLI concentration.} (a) The saturation of the scaling curve for fixed input and
output boundary conditions with 20\% TOFFOLI gates when
$\ell(\tau)\sim W/2$. This indicates that the solutions have been
reached, and is consistent with the domain growth picture. Notice
that for mixed input and output boundary conditions, the curve does
not saturate when $\ell(\tau)\sim W/2$. (b) The functional form of
the scaling curves as a function of different TOFFOLI
concentrations. The correlation grows more slowly as the
concentration of TOFFOLI gates increases. The dashed line
corresponds to the fitting $\ell(\tau)=\ell_0[(\tau/\tau_0)/{\rm
ln}(\tau/\tau_0)]^{1/2}$, with $\ell_0=1.42$ and $\tau_0=8.33$.}
\label{fig:saturation}
\end{figure}
The details of the numerical Metropolis simulations are presented in
the Methods Section. Our results are represented in the form inspired
by the dynamic scaling theory of Ref.~\cite{Liu-Polkovnikov-Sandvik}
that builds on the Kibble-Zurek mechanism~\cite{kibble,zurek}, namely:
\begin{equation}
\ell(\tau) = \langle m \rangle (\tau)\; WL/L_{\partial} \;,
\label{eqn:scale}
\end{equation}
which defines a dynamical correlation length,
$\ell(\tau)$. $L_{\partial}$ is the number of pinned vertices on both
boundaries (see the Methods Section). To motivate
Eq.~(\ref{eqn:scale}) we note that the domain of satisfied gates that
contribute to $\langle m \rangle (\tau)$, the fraction of gates that
reach their correct states at time $\tau$, grows from the pinned
states at the boundaries, and covers an area $L_{\partial}\times
\ell(\tau)$. Thus $\ell(\tau)$ describes the growth of correlated
regions of satisfied gates that eventually connect the two boundaries
of the circuit. (We note that recently, the Kibble-Zurek mechanism has
been extended to include systems with zero-temperature
order~\cite{Rubin}, the case relevant to the current discussion).
We note that at any temperature $T$ along the annealing path, the
correlation length is $\ell_T (\tau) \leq \ell _T (\tau \to \infty) =
\xi_T$, where $\xi_T\sim e^{K/2T}$ is the thermal correlation length
in the paramagnetic state, and $K$ is a characteristic ferromagnetic
interaction strength in our model. In thermal equilibrium all gate
constraints defining the computational circuit would be satisfied once
$\xi_T$ reaches the depth of the computation, $W$. Notice that the
exponential dependence of $\xi_T$ on temperature implies that
achieving the correct assignment of gates does not require very low
temperatures on the scale of $K$ since $\xi_T \sim W$ already for
temperatures below $T\sim K/\ln W$. However, reaching the solution to
the computational problem is a dynamical process that cannot proceed
to completion until the {\it dynamic correlation length} at the end of
the annealing protocol, $\ell (\tau) = \ell_{T=0} (\tau)$, reaches
$W$, allowing the input and output boundaries of the system that
specify the computation to communicate.
In Fig.~\ref{fig:MC_scaling_plot1} we present the numerical results
for fixed input and output, and mixed boundary conditions, with
different concentrations of TOFFOLI gates (see the Methods Section for
details). Remarkably, we find that the curves for different system
sizes $L$ and $W$ collapse very well when scaled as in
Eq.~(\ref{eqn:scale}). In addition, notice that for shorter circuit
with fixed input and output and low concentration of TOFFOLI gates
(20\%), $\ell(\tau)$ begins to saturate for large enough $\tau$
(Fig.~\ref{fig:MC_scaling_plot1}a). As shown more clearly in
Fig.~\ref{fig:saturation}a, this saturation occurs when the dynamical
correlation length $\ell(\tau)$ reaches $W/2$, where the growing
domains of satisfied gates meet. Since in this case $L_{\partial} =
2L$, $\ell(\tau_s) \sim W/2$ corresponds to $\langle m \rangle
(\tau_s) = 1$ establishing $\tau_s$ as the time-to-solution. For
mixed boundary conditions, however, $\ell(\tau)\sim W/2$ initiates the
communication between the two boundaries and establishes the system's
capacity to ``learn'' (see below) but is not sufficient for
negotiating solution. Indeed, Fig.~\ref{fig:saturation}a shows that
$\langle m \rangle (\tau)$ does not yet saturate when $\ell(\tau)\sim
W/2$. As can be seen from Fig.~\ref{fig:correlation_growth} for
computations with mixed boundary conditions, correlations must develop
along the transverse direction ({\it i.e.}, parallel to the
boundaries) before solution can be reached. In those cases it is this
slower process that determines the time-to-solution and dominates the
complexity of computations.
Finally, all non-trivial operations between input and output bits
involve TOFFOLI gates, and it is thus expected that the increasing the
concentration of these gates slows down the growth of
correlations. This expectation is confirmed in
Fig.~\ref{fig:saturation}b, where we show curves for the same system
size with different concentrations of TOFFOLI gates. The case of no
TOFFOLI gates is equivalent to $3L$ decoupled ferromagnetic Ising
chains. In this case the dynamic correlation length behaves as
$\ell(\tau)=\ell_0[(\tau/\tau_0)/{\rm ln}(\tau/\tau_0)]^{1/2}$ (with
$\ell_0=1.42$ and $\tau_0=8.33$) as illustrated by the dashed line in
Fig.~\ref{fig:saturation}b. This behavior is in agreement with the
exact result for the Kibble-Zurek dynamical scaling of the density of
domain walls in a ferromagnetic Ising chain \cite{Krapivsky}.
\subsection{Annealing with learning}
\label{sec:learning}
\begin{figure*}[hbt]
\centering
\includegraphics[angle=0,origin=c,width=18cm]{Fig6.pdf}
\caption{{\bf Growing correlation length without learning.} The average local overlaps $\langle\delta_{q^{\rm sol}_s,
q^{\rm final}_s}\rangle$ of 2000 replicas with $\{q^{\rm sol}\}$
for a given circuit and boundary state without learning. The system
size is $16\times 42$, with 20\% TOFFOLI gates. (a) Fixed
input and output; (b) mixed boundary conditions.}
\label{fig:correlation_growth}
\end{figure*}
Simple thermal annealing is not necessarily an optimal way to reach
the ground state. For example, in the case of forward computation, the
time scale for the dynamic correlation length to grow to $\ell(\tau)
\sim W/2$ (so as to reach solution) is slower than ballistic (or
linear in $\tau$), as expected for deterministic forward
computation. This can be already seen from the exactly solvable case
with no TOFFOLI gates. Moreover, for single-solution problems with
mixed boundary conditions the growth of correlations establishing
communication between boundaries scales with the same form as in
direct computation (see Fig.~\ref{fig:MC_scaling_plot1}). However, in
that case negotiating solution requires the establishment of much
slower correlations {\it along} the boundaries, a process for which a
``vanilla'' thermal annealing approach is extremely inefficient and
would require unreasonably large computational resources.
These shortcomings are addressed by using a heuristic ``learning''
protocol in which annealing proceeds
through the following steps: (1) one starts by annealing $N_R$
identical replicas of a circuit over some time $\tau_a$, during which
the correlation lengths grow beyond a few columns of gates such that
the probability for assigning correct gates within that region, $p\sim
\exp -|x|/\xi > 1/2$, within each replica; (2) one then assigns a
specific identity to each gate (with $p >1/2$) provided that a
fraction of the $N_R$ replicas, greater than or equal to $\alpha$,
agree on this assignment; (3) with the agreed upon gates frozen, the
annealing process is independently applied again to each of the
replicas allowing only gates not yet fixed to participate in the
Metropolis algorithm; finally, (4) the procedure is iterated until all
gates are fixed, thus establishing the solution to the problem.
This protocol raises the question of how many replicas $N_{R\epsilon}$
are needed to ensure that the learning algorithm reaches the correct
result with a probability greater than $1-\epsilon$. In particular,
how does $N_{R\epsilon}$ depend on the system size $L\times W$ and the
threshold $\alpha$? As we show in the Supplementary note 3, the
number of replicas needed to ensure an error rate smaller than
$\epsilon$ is given by $N_{R\epsilon} = \frac{{\rm ln} \left[
\frac{2p-1}{p} \frac{\epsilon}{LW} \right]} {{\rm ln} \left[
2p^{1-\alpha}(1-p)^{\alpha}\right]}$, where $p>\frac{1}{2}$ is the
probability of a correct gate assignment for one replica. Note that,
for fixed $\alpha$ and error rate $\epsilon$, the number of replicas
grows only \textit{logarithmically} with the system size, and thus in
practice the learning algorithm works with reasonable resources.
Before describing the results of applying ``annealing with learning''
to computations with both fixed and mixed boundary conditions,
in Fig.~\ref{fig:correlation_growth} we plot the average
local overlaps of 2000 replicas with the solution for a fixed circuit
and boundary condition before applying the learning algorithm. The
agreement with the data presented in Fig.~\ref{fig:MC_scaling_plot1}
substantiates the fact that the local majority rule implemented
through the independent annealing of the replicas recapitulate the
behavior of the correct solution to the computational problem.
\begin{figure*}[ht]
\centering
\includegraphics[angle=0,origin=c,width=18.5cm]{Fig7.pdf}
\caption{{\bf Growing correlation length with learning for fixed input boundary condition.} Annealing with learning for the fixed input boundary case for
a system of size $16\times 42$ with 20\% TOFFOLI gates. The
annealing time within each iteration is $\tau_a=2^{13}$ and the gate
state probability threshold $\alpha=0.7$.}
\label{fig:learning_forward}
\end{figure*}
We start from fixed input boundary condition. Using the algorithm described above,
we choose $\tau_a=2^{13}$ for each iteration, and set the majority
rule threshold at $\alpha=0.7$. In Fig.~\ref{fig:learning_forward} we
show the local order parameter of the final states averaged over 2000
replicas after each iteration. We emphasize that even though we are
plotting the average overlap with the actual solution as a benchmark,
in the learning algorithm no reference to $\{q^{\rm sol}\}$ is made. The
weight of each possible state of each gate in the circuit is computed
solely from the replicas. After each iteration, with $\tau_a=2^{13}$
the correlation length grows to $\ell(\tau_a)\sim 10$, and by pinning
gates with high percentage of agreement on certain states we are
pushing the ``boundary'' forward until all gates are fixed. Since the
total number of iterations $n_a$ scales linearly with the circuit
depth $W$, $n_a \propto W$, the total time to solution
$\tau=n_a\tau_a$ also scales linearly with $W$, $\tau \propto W
\tau_a$, consistent with the expectations for the time-to-solution for
forward computation. For the computation shown in
Fig.~\ref{fig:learning_forward}, it is clear that the ``annealing with
learning'' process proceeds ballistically and reaches solution with
$n_a=9$ steps.
Now we look at mixed boundary conditions. The results presented in
Fig.~\ref{fig:learning_mixed} are obtained by applying the learning
algorithm with $\alpha=0.7$. However, in the case of mixed boundary
conditions the process of ``learning'' proceeds through two series of
annealing steps with different time scales: an initial set of
iterations with $\tau_a = 2^{13}$ which build longitudinal
correlations required for learning, followed by a set of longer
annealing steps with $\tau_b = 2^{18}$ that allow the slower
correlations along the transverse direction to
develop. Figure~\ref{fig:learning_mixed} shows the progression to
solution, which could not be reached for the same computation using
the ``vanilla'' thermal annealing for our longest accessible times
($\tau \sim 2^{25}$).
We note that this protocol can also be used to solve problems with a
``few'', ${\cal O}$(1), solutions. This is best illustrated for the
case of two solutions, which can be addressed by carrying out $2n$
computations with mixed boundary conditions, where $n$ is the number
of unknown bits in the input. The idea is to define $2n$ problems by
fixing each bit at a time to be 0 or 1, while leaving the other $n-1$
bits floating. Since the two solutions must differ in at least one of
the $n$ bits, after at most $2n$ steps, this scheme transforms the
problem into two separate problems, each of which can be solved by the
techniques discussed in this paper. An important problem that falls
precisely within this case is factorization of semi-prime numbers
$s=p\times q$, where there are exactly two solutions, corresponding to
the two ordered pairs $(p,q)$ and $(q,p)$ of primes $p,q$ (assumed to
be different).
\begin{figure*}[ht]
\centering
\includegraphics[angle=0,origin=c,width=18cm]{Fig8.pdf}
\caption{{\bf Growing correlation length with learning for mixed boundary conditions.} Annealing with learning for mixed boundary conditions and
systems of size $16\times 42$ with 20\% TOFFOLI gates. The annealing
time within each iteration is $\tau_a=2^{13}$, and the probability
threshold $\alpha=0.7$. After iteration 6 (not shown), the correlations fully build up along the
longitudinal direction, $\tau_a$ is then increased to $\tau_b=2^{18}$.}
\label{fig:learning_mixed}
\end{figure*}
Finally, we turn to the analysis of cases with multiple solutions and
no solution. In both of these cases it is not sensible to compute the
local overlap with a solution, as we did for circuit problems with
only one solution. Instead, we plot the largest weight of each gate
state in the circuit obtained from 2000 replicas. This is shown in
Fig.~\ref{fig:multi_no_sol} for an instance with 8 solutions obtained
by fixing fewer gates (than in the single solution case) on each
boundary; and an instance with no solutions, obtained by fixing a few
gates on one boundary to the wrong states. Fig.~\ref{fig:multi_no_sol}
shows that the learning algorithm eventually gets stuck when the
replicas cease to agree on gate assignments above the threshold
$\alpha$. We note that the learning algorithm cannot differentiate
between these two cases. We interpret the freezing of the system as an
effect of frustration in satisfying the local gate constraints in the
bulk induced by incompatible boundaries in the case of no solution or
compatible but competing boundaries in the case of multiple solutions.
\begin{figure}[ht]
\centering
\includegraphics[angle=0,origin=c,width=9cm]{Fig9.pdf}
\caption{{\bf Cases with multiple solutions or no solution.} Color plot of the largest weight of each gate state in the
circuit for a system of size $16\times42$ and 20\% TOFFFOLI gates
after a relaxation time $\tau\approx 2^{20}$. (a) Case with
8 solutions; (b) case with no solution. The learning
algorithm eventually gets stuck at the point where no more gates
have majority weight above the threshold $\alpha$.}
\label{fig:multi_no_sol}
\end{figure}
\subsection{Mapping onto the D-Wave Chimera graph for quantum annealing}
\label{sec:chimera}
We close this paper by describing a scheme for ``programming'' our
vertex model into a quantum annealer. In particular, we present an
explicit embedding of the tile model of universal classical computing
circuits into the Chimera graph architecture of the D-Wave
machine. The idea is to use one unit cell to represent one square tile
of our construction presented previously. Rectangular tiles
(\textit{i.e.}, TOFFOLI gates) can be viewed as consisting of two
square tiles, thus requiring two unit cells to be embedded in the
Chimera graph. We then implement the Hamiltonians of
Eqs. (\ref{H-ID})-(\ref{H-Toffoli}) using the programmable couplers
available in the D-Wave machine, as illustrated in
Fig.~\ref{fig:chimera} and described in more detail in the Methods
Section.
\begin{figure*}[!ht]
\centering
\includegraphics[angle=0,origin=c,width=0.8\textwidth]{Fig10.pdf}
\caption{{\bf Mapping onto the Chimera graph.} Procedure for embedding a $4\times 4$ tile lattice into the
Chimera graph. (a) Left: a generic tile lattice rotated by
$45^\circ$. Spins are put on the boundary of each tile. The lattice
can be further divided into two sublattices, depicted by dark and
light grey respectively; right: embedding of the tile lattice into
the Chimera graph. The ``grout couplings'' are indicated by red
links. (b) Embedding of each gate into the unit cells of the Chimera
graph. (i) Left: a $K_{4,4}$ unit cell of the Chimera graph; middle:
in order to couple qubits in the same column, we slave the qubits to
their neighbors in the other column using additional ferromagnetic
couplings indicated by red links; right: effectively we are left
with four qubits that are fully connected. For simplicity, we
hereafter denote the effective couplings between spins in the same
column by a single green link. However, one should keep in mind that
they are obtained by slaving the spins to the opposite column via
large ferromagnetic couplings. (ii) The four qubits in the rotated
square tile are labeled by their locations on the tile: N (North), S
(South), W (West) and E (East). Tiles corresponding to different
sublattices must be embedded differently due to the special
connectivity of the Chimera graph. (iii) Embedding of the TOFFOLI
gate consisting of two square tiles into two unit cells. $(a, b, c,
d)$ corresponds to the input and output bits of the gate, and $S$ is
the ancilla bit. In the unit cell, ferromagnetic couplings that copy
spins are indicated by purple links, and couplings required in
Hamiltonian (\ref{H-Toffoli}) are indicate by black links.}
\label{fig:chimera}
\end{figure*}
\section{Discussion}
\label{sec:conclusions}
The results of this paper were motivated by an attempt to use our
statistical mechanics intuition about lattice models of spin systems
to uncover some of the salient features of universal classical
reversible computation. There are questions posed and open problems
raised by these studies. Here we list four that we find most
important.
First, one should understand the scaling of time-to-solution of the
various schemes discussed here, including those that utilize learning,
as a function of input size and depth for specific computational
problems. Under a trivial reduction scheme, one can solve problems
with two solutions using similar annealing with learning techniques
that we deployed for problems with a unique solution. As an important
application we are already investigating the problem of the
factorization of semi-primes. The scaling properties of the
time-to-solution in the context of this concrete and relevant problem
should be contrasted to that obtained in the random circuit with the
same concentration of TOFFOLI gates.
A second question raised by our work is the nature of the
zero-temperature quantum phase transition encountered in the quantum
vertex model, as depicted in Fig.~\ref{fig:phase-diagram}. We
demonstrated that, in the limit of the trivial computational circuit
with no TOFFOLI gates, this transition is second order, in direct
analogy to the case of the one-dimensional Ising model in a transverse
magnetic field. Whether the transition remains second order or becomes
first order for realistic computations (corresponding to a finite
concentration of TOFFOLIs) has very important consequences for
solutions of computational problems via quantum annealing.
Third, the computational problems discussed here should also be
studied directly in a {\it bona fide} quantum annealer. An important
result of this paper is the programming of generic reversible
computational circuits into the Chimera architecture of the D-Wave
machine. This paves the way for using this type of hardware to study
annealing protocols along the $\delta$ axis, as well as arbitrary
directions in the $\delta - T$ plane. Our approach should also be used
as a guide to the development of alternative machine architectures
optimized for direct implementations of the vertex model.
Finally, we close with a brief discussion of the broader implications
of the mapping of reversible classical computation onto the vertex
model on the individual disciplines of computer science and
physics. As already mentioned earlier, the line of argumentation in
this paper follows a physics perspective, namely, it concentrates on
``typical behavior'' based on heuristic approach to explicit
instantiations of the vertex model. Computer science could benefit
from further work on more sophisticated theoretical and computational
heuristic approaches, special purpose hardware ({\it i.e.}, quantum
annealers), and new formal proofs that rely on statistical mechanics
representations of computational problems. At the same time there are
lessons to be learned from computer science that we believe may have
interesting implications for physics. For example, if NP$\neq$P, the
vertex model representing the hardest problems in UNIQUE-SAT can be
also viewed as describing a physical glassy system that displays slow
dynamics even though the model involves no frustrating interactions,
has a unique non-degenerate ground state, and displays no bulk
thermodynamic transitions down to zero temperature! There are known
examples of systems with glassy dynamics in the absence of a
thermodynamic phase transition, such as the kinetically constrained
models discussed in
\cite{Newman-Moore1999,Garrahan-Newman2000,Ritort-Solich2003}.
However, the non-Arrhenius relaxation characteristic of these models
only translate into a quasi-polynomial time-to-solution of a
computational problem. Thus, within the vertex model approach, the
existence of hard UNIQUE-SAT problems with exponential or
sub-exponential behavior of the time-to-solution would suggest the
existence of a novel family of glassy physical systems without a
thermodynamic transition but with exponentially large barriers and
corresponding astronomically-long relaxation times. This example
underscores the richness of the possibilities opened by explorations
of the vertex model of classical computation and more generally, of
problems at the interface between physics and computer science.
\section{Methods}
\subsection{Implementing gates with one- and two-body spin interactions}
\label{sec:model-definition}
We start by representing Boolean variables $x_i= (1+\sigma_i)/2$ in
terms of spins $\sigma_i=\pm 1$ placed on the boundary of each tile,
as depicted in Fig~\ref{fig:gate_tile_coupling}. Operations of logic
gates are then implemented in a similar way as in
Ref.~\cite{Biamonte2008}, by designing a Hamiltonian acting on the
spins associated with individual tiles such that (a) the interactions
are short ranged and involve at most two bodies; and (b) spin (i.e.,
bit) states that satisfy the gate constraint are ground states of the
tile Hamiltonian and all other ``unsatisfying'' spin-states are pushed
to high energies.
\noindent
{\textbf{Identity (ID) Gate:}} The ID gate takes two bits $(a,b)$ into
$(a,b)$. This is easily enforced by adding ferromagnetic interactions
($J>0$) that align input bits $a$ and $b$ to output bits $c$ and $d$,
respectively, leading to an energy
\begin{equation}
E_{{\rm ID}}(\sigma_a, \sigma_b; \sigma_c, \sigma_d) = - J(\sigma_a
\sigma_c+\sigma_b \sigma_d).
\label{H-ID}
\end{equation}
\noindent
{\textbf{SWAP Gate:}} The SWAP gate takes $(a,b)$ into $(b,a)$, and
can be implemented in the same manner as the ID gate through a
ferromagnetic interaction ($J>0$),
\begin{equation}
E_{{\rm SWAP}}(\sigma_a, \sigma_b; \sigma_c, \sigma_d) = - J(\sigma_a
\sigma_d +\sigma_b\sigma_c).
\label{H-SWAP}
\end{equation}
\noindent
{\textbf{TOFFOLI Gate:}} The TOFFOLI gate is represented by a
rectangular tile with the three input bits $(a, b, c)$ and three
output bits $({a'},{b'},d)$ placed on the boundary, as shown in
Fig.~\ref{fig:gate_tile_coupling}. Notice that in this case we also
place an additional ancilla bit in the center of the rectangular tile,
which is essential in order to satisfy the gate constraint with no
more than two-body interactions. The TOFFOLI gate takes the three-bit
input state $(a,b,c)$ into $(a,b,ab\oplus c)$. The copying of the
first two input bits from the input into the output is accomplished as
before through a ferromagnetic coupling: $-J(\sigma_a
\sigma_{a'}+\sigma_b \sigma_{b'})$. Enforcing the third output bit
$d=ab\oplus c$ requires a more involved interaction. We present the
result below, and leave the detailed justification for the
Supplementary note 1. The complete energy cost associated to the
TOFFOLI gate reads
\begin{widetext}
\begin{eqnarray}
E_{{\rm TOFFOLI}}(\sigma_a, \sigma_b, \sigma_c; \sigma_{a'},
\sigma_{b'}, \sigma_{d}; \sigma_S) & = & -
J(\sigma_a\sigma_{a'}+\sigma_b\sigma_{b'}) + J(\sigma_a -3\sigma_b -
2\sigma_c +2\sigma_d +4\sigma_S) \nonumber \\ & &
+\ J(-3\sigma_a\sigma_b - 2\sigma_a\sigma_c + 4\sigma_b\sigma_c +
2\sigma_a\sigma_d-4\sigma_b\sigma_d - 4 \sigma_c\sigma_d \nonumber
\\ & & \;\;\;\;\;\;
+\ 4\sigma_a\sigma_S - 8\sigma_b\sigma_S - 6\sigma_c\sigma_S +
6\sigma_d\sigma_S).
\label{H-Toffoli}
\end{eqnarray}
\end{widetext}
\subsection{The global constraint and coupling of adjacent tiles}
\label{sec:tile-coupling}
In addition to satisfying each gate separately, spins shared by
neighboring tiles must be matched across the entire system in order
for the tile model to accurately represent the desired computational
circuit. To be precise one can imagine splitting each boundary spin
into two ``twin'' spins and identifying input/output spins with each
tile. Within this picture, adjacent spins at the boundary between
tiles must be locked together, a constraint we implement by
introducing a ferromagnetic ``grout'' coupling $K>0$ between spins on
adjacent tiles. The corresponding term in the energy is then written
as
\begin{equation}
E_{\rm grout}(\{\sigma\}) = - K \sum_{\langle i,j\rangle}
\sigma_i\;\sigma_j,
\end{equation}
where $\langle i,j\rangle$ labels pairs of ``twin'' spins $i$ and $j$
on the boundary between two adjacent tiles and the sum ranges over all
such pairs of the system.
\subsection{Boundary conditions}
\label{sec:boundary_conditions}
Completing the description of the two-dimensional model of universal
classical computation requires a discussion of boundary conditions,
which determine the type of computational problem one is addressing.
For example, if the $N$-bit input is fully specified and one is
interested in the output, all that is needed is to transfer the
information encoded into the input left to right by applying
sequentially the gates one column of tiles at a time. In this case, if
the depth (i.e., the number of steps) of the computation is a
polynomial in $N$, this column-column computation reaches the output
boundary, and thus solves the problem, in polynomial time.
As mentioned earlier, by using reversible gates one can also represent
computational problems with mixed input-output boundary conditions for
which only a fraction of the bits on the left (input) edge and a
fraction of the bits on the right (output) edge are fixed. A concrete
example is the integer factorization problem implemented in terms of a
reversible integer multiplication circuit. A reversible circuit for
multiplying two $N$-bit numbers $p$ and $q$ can be constructed using
$5N+1$ bits in each column. One needs two $N$-bit registers for the
two numbers $p$ and $q$ to be multiplied, one $N$-bit carry register
$c$ for the ripple-sums, a $2N$-bit register $s$ for storing the
answer $p\times q=s$, and one ancilla bit $b$. For multiplication, one
only fixes the boundary conditions on the input: $p$ and $q$ are the
two numbers to be multiplied, and $c$, $s$ and $b$ are all 0's. For
factorization we must impose mixed boundary conditions: On the input
side the $c$, $s$ and $b$ registers are fixed to be all 0's; on the
output side the $s$ register is now fixed to the number to be
factorized, and $c$ and $b$ are again all set to 0. Thus, $3N+1$ bits
in the input and output are fixed, while $2N$ bits are floating on
both boundaries.
Boundary conditions on inputs, outputs, or both are imposed by
inserting longitudinal fields at the appropriate bit sites, namely,
\begin{equation}
E_{\rm boundary}(\{\sigma\}) = -\!\!\!\sum_{i\in{\rm boundary}} h_i\,
\sigma_i,
\end{equation}
with $|h_i| = h \gg J$. The sign of an individual $h_i$ field
determines the value of the spin $\sigma_i$ and thus of the binary
variable $x_i$: For $h_i>0$, $x_i=1$, while for $h_i<0$, $x_i=0$. If
no constraint is imposed on a binary variable $x_i$, then $h_i=0$.
\subsection{Construction of the vertex model}
\label{sec:strong_coupling}
Combining the contributions above leads us to a classical Hamiltonian
that includes the energy functions internal to each tile, the coupling
between the spins at the boundary between adjacent tiles, and the
magnetic fields associated with the input and output bits defining the
boundary conditions of the computation, namely,
\begin{equation}
\label{eq:E_C}
H_C = \sum_{g} E^J_g (\{\sigma\}_g) - K \sum_{\langle i,j\rangle}
\sigma_i\, \sigma_j
- \!\!\!\sum_{i\in \text{boundary}} h_i\, \sigma_i ,
\end{equation}
where $\{\sigma\}_g$ labels all the spins and $E^J_g (\{\sigma\}_g)$
represents the energy function of tile (\textit{i.e.}, gate) $g$.
This Hamiltonian is the starting point for our mapping of universal
classical computation into the Chimera architecture of the D-Wave
machine, one of the important results of the paper, which we discuss
in detail below. In order to anticipate the fact that quantum rather
than classical thermal annealing may be a more effective way of
reaching the ground state and therefore the solution of these
computational problems, we add a transverse magnetic field $\Gamma$ to
Eq. (\ref{eq:E_C}) to obtain the quantum Hamiltonian
\begin{eqnarray}
\label{eq:E_QA2}
\hat H &=& \sum_{g} E^J_g (\{\hat\sigma^z\}_g) - K \sum_{\langle
i,j\rangle} {{\hat\sigma^z}_i}\, {{\hat\sigma^z}_j} -
\!\!\!\sum_{i\in \text{boundary}} h_i\, \hat\sigma^z_i \nonumber \\ &
& +\ \Gamma \sum_i \hat{\sigma}_i^x .
\end{eqnarray}
However, we find it more expedient and intuitive to work directly with
tiles which satisfy the logic gate constraint exactly. We thus proceed
by projecting the system onto the manifold of states where all local
gate constraints are satisfied by working in the limit in which both
$h$ and $J$ are very large; and we imagine varying $K$ and $\Gamma$,
with $K,\Gamma \ll J,h$ \cite{castelnovo2006}. This limit is best
understood if we switch off the coupling between tiles, $K$. Within a
given tile, the configurations that satisfy the logic gate constraints
span the degenerate ground state manifold, while the unsatisfying
configurations have energies of order $J$ and higher. Let
$\{|q_a\rangle\}$, $a=1,\dots,r$ be all the $r$ states spanned by the
spin configurations $|\sigma_1,\dots,\sigma_n\rangle$ that define the
ground state manifold. For two-bit (four-spin) gates we have $r=4$,
while for three-bit (six-spin) gates $r=8$.
As long as $\Gamma\ll J$ we can understand the effect of a transverse
field $\Gamma$ on the $r$ degenerate states by degenerate perturbation
theory. Since for reversible gates maintaining the gate constraints
requires at least two spin flips, the transverse field $\Gamma$
induces an effective, second-order or higher spin-spin interaction on
the ground state manifold of a given tile of order $\Delta =
\Gamma^2/J$ or lower. This discussion leads naturally to the quantum
vertex model presented in Eq.~(\ref{eq:H-vertex}).
Switching on the $K$ coupling penalizes configurations in which the
states of adjacent tiles are incompatible. Thus, in order to satisfy
both intra- and inter-tile constrains that define the computational
process we must reach the limit of $\Delta \ll K\ll J,h$.
\subsection{Thermodynamics of the classical vertex model}
\label{sec:transfermatrix}
We start by considering the partition function of the classical limit
of the Hamiltonian in Eq.~(\ref{eq:H-vertex}), i.e., $\Delta=0$, which
we obtain via a transfer matrix calculation. Consider first a system
with free boundary conditions at both ends. The partition function for
the vertex model can be more easily written using the spin variables
on the links of the lattice. Let $\{\sigma\}_j$ denote the spin states
on a vertical line, which cuts across $3L$ spins (with $L$ the number
of vertices or 3-bit gates in a column). For convenience we shall
utilize the notation $|\{\sigma\}_j\rangle$ for the vectors in this
transfer matrix calculation. Within this notation, one can write the
partition function as
\begin{widetext}
\begin{eqnarray}
Z= \sum_{\{\sigma_{1}\},\dots,\{\sigma_{2W}\}}
\langle\{\sigma_{1}\}|P_1|\{\sigma_{2}\}\rangle
\langle\{\sigma_{2}\}\}|T|\{\sigma_{3}\}\rangle
&\times\cdots\times&
\langle\{\sigma_{2j-1}\}|P_j|\{\sigma_{2j}\}\rangle
\langle\{\sigma_{2j}\}|T|\{\sigma_{2j+1}\}\rangle \times\cdots
\nonumber\\ &\cdots\times&
\langle\{\sigma_{2W-1}\}|P_W|\{\sigma_{2W}\}\rangle \;,
\end{eqnarray}
\end{widetext}
where the matrix $T$ encodes the energy costs for matching spins
across the links, and the matrices $P_j$ encode the computations
performed by one column of gates. The two types of slices are depicted
in Fig.~\ref{fig:ripple-carry}. Notice that the $T$ is the same for
all slices, and its matrix elements are given by
\begin{equation}
\langle\{\sigma_{2j}\}|T|\{\sigma_{2j+1}\}\rangle =
\exp\left({\sum_{a=1}^{3L}
\beta\,K\,\sigma_{2j,a}\,\sigma_{2j+1,a}}\right),
\end{equation}
whereas $\langle\{\sigma_{2j-1}\}|P_j|\{\sigma_{2j}\}\rangle$
represents the matrix element of $P_j$ at the $j$th column and thus
depends on the particular set of gates within that column. However,
all $P_j$ are permutation matrices since all gates are
reversible. This fact is essential because it allows us to compute the
partition function exactly, irrespective of the circuit.
For the next step notice that the vector $|\Sigma\rangle = \sum_{\{\sigma\}}
|{\{\sigma\}}\rangle$ is an eigenvector of $P_j$ for any operation
$P_j$:
\begin{eqnarray}
P_j\;|\Sigma\rangle & = & P_j\;\sum_{\{\sigma\}}|{\{\sigma\}}\rangle
= \sum_{\{\sigma\}} P_j\;|{\{\sigma\}}\rangle =
\sum_{\{\sigma'\}}|{\{\sigma'\}}\rangle \nonumber \\ & = &
|\Sigma\rangle \;,
\end{eqnarray}
where we used that we can relabel the states after the
permutation. The vector $|\Sigma\rangle = \sum_{\{\sigma\}}
|{\{\sigma\}}\rangle$ is also an eigenvector of $T$:
\begin{eqnarray}
T\;|\Sigma\rangle & = & T\;\sum_{\{\sigma\}}|{\{\sigma\}}\rangle=
\sum_{\{\sigma\},\{\sigma'\}} |\{\sigma'\}\rangle
\langle\{\sigma'\}|T|\{\sigma\}\rangle \nonumber \\ &=&
\sum_{\{\sigma'\}} |\{\sigma'\}\rangle \sum_{\sigma}
\exp\left({\sum_{a=1}^{3L} \beta\,K\,\sigma'_{a}\,\sigma_{a}}\right)
\nonumber\\ &=& \sum_{\{\sigma'\}} |\{\sigma'\}\rangle
\;\;\prod_{a=1}^{3L}\;\; \sum_{\sigma_{a}=\pm 1} e^{\;\beta
K\,\sigma'_{a}\,\sigma_{a}} \nonumber\\ &=& \sum_{\{\sigma'\}}
|\{\sigma'\}\rangle \; (2\cosh \beta K)^{3L} \nonumber\\ &=& (2\cosh
\beta K)^{3L}\; |\Sigma\rangle \; \;.
\end{eqnarray}
By collecting all the factors we arrive at the partition function
\begin{eqnarray}
Z &=& \langle\Sigma|\;P_1\;T\;P_2\;T\dots\;T\;P_{W}\;|\Sigma\rangle
\nonumber\\ &=& (2\cosh \beta
K)^{3L(W-1)}\;\langle\Sigma|\Sigma\rangle \;.
\end{eqnarray}
The overlap $\langle\Sigma|\Sigma\rangle=2^{3L}$ reflects the $2^{3L}$
degenerate ground states corresponding to open boundary conditions on
both boundaries. Had we fixed one of the boundaries to a particular
state $|\{\sigma\}_{\rm fixed}\rangle$ we would have instead obtained
an overlap $\langle \{\sigma\}_{\rm fixed}|\Sigma\rangle=1$. More
generally, in the thermodynamic limit boundaries contribute an
entropic term that counts the number of ground states, but does not
affect the bulk thermodynamics. In particular, the bulk free energy is
that of a paramagnet:
\begin{equation}
\beta F =-\left[3L(W-1)\right]\; \ln (2\cosh \beta K) \;.
\end{equation}
This also implies that thermodynamics alone, which is independent of
the specific form of the circuit, cannot reveal the complexity of a
ground-state computation, which is reflected in the dynamics of the
system's relaxation into its ground state.
\subsection{Metropolis algorithm for thermal annealing}
The Metropolis simulations are carried out as follows. We work on a
lattice of $L\times W$ vertices, using the Hamiltonian of
Eq.~(\ref{eq:H-vertex}) with $\Delta_{q_s,q_s'}=0$. Periodic boundary
conditions are used in the transverse direction, i.e., the circuit is
laid down on the surface of a tube of length $W$ and circumference
$L$. We consider four types of circuits corresponding to different
concentrations of TOFFOLI gates (the other four types of gates are
assigned equal concentrations): a circuit with only TOFFOLI gates
(100\% concentration), and random circuits with 40\%, 20\%, and 0\%
concentration of TOFFOLI gates.
The first step of the simulation is
to construct a reference state $\{q^{\text{sol}}\}$ that solves the
circuit, by fixing the states $q^{\text{sol}}_s$ for the vertices $s$
at the left boundary, and determining and storing all other states
$q^{\text{sol}}_s$ for vertices $s$ in the rest of the circuit. Next,
we construct three explicit boundary conditions consistent with the
reference state, $\{q^{\text{sol}}\}$, that will serve as the input
states for our simulations: 1. fixed input, for which we apply pinning
fields at the left boundary that fix the states to match
$q^{\text{sol}}_s$ for the vertices $s$ at the left boundary, and
leave the other boundary free (no pinning field); 2. fixed input and
output, for which we pin all vertices $s$ on the left and right
boundaries to those defined by $q^{\text{sol}}_s$; and 3. mixed
boundaries, for which we pin $L_{\partial}=L/2+3$ vertices on both the
left and right boundaries to the solution values $q^{\text{sol}}_s$,
but leave all the remaining boundary vertices free. Our computations
proceed in each of these three cases by averaging over 2000
independent random instances of input states for a given circuit with
a fixed concentration of TOFFOLI gates corresponding to different
$\{q^{\text{sol}}\}$ and computing the average order parameter
$\langle m \rangle$ as a function of the relaxation time $\tau$. Here
it is important to stress that the partial specification of boundaries
in the case of mixed boundary conditions generically leads to multiple
solutions which compete in establishing the local configurations of
gates consistent with the global constraints defining the
computational circuit. While we also discuss cases with multiple
solutions and no solution in Results, the focus of
this paper is on problems with a single solution. To ensure a single
solution in the case of mixed boundary conditions we always check that
each of the random instances of the input state for a given circuit
allows for one and only one solution.
\subsection{Mapping onto the Chimera architecture of the D-Wave machine}
Figure \ref{fig:chimera} shows the `flow chart' of embedding a
$4\times 4$ tile lattice into the Chimera graph. The entire tile
lattice is rotated by $45^\circ$ for convenience, and spins living on
the boundary of each tile are shown explicitly. The lattice of tiles
can be further divided into two sublattices labeled by dark and light
grey, for reasons that should become clear shortly. Now let us first
consider how to encode the ID and SWAP gates represented by a single
square tile into a unit cell. The embedding involves internal
couplings $J$ that enforce the gate constraints, and the ``grout''
couplings $K$ that match adjacent tiles. The Chimera unit cell forms a
complete bipartite graph $K_{4,4}$, as depicted in
Fig.~\ref{fig:chimera}b-(i), with each spin in one column coupled to
all spins in the other, but not to those in their own column
\cite{Choi}. In order to obtain the generic spin couplings to
represent the gates on a single tile, which inevitably involves
couplings between qubits in the same column as well, we use an
additional ferromagnetic coupling to slave the spins in one column to
their nearest neighbors in the other column. Thus, effectively we are
left with four spins that are fully connected. Details are shown in
Fig. \ref{fig:chimera}b-(i).
In order to use the connectivity of the Chimera graph architecture and
couple adjacent tiles properly, it is convenient to explore the
bipartiteness of the square lattice. Let us take one tile from the
rotated tile lattice, and label the four qubits by their locations on
the tile: N (North), S (South), W (West) and E (East), as shown in
Fig. \ref{fig:chimera}b-(ii). The spins in the adjacent tiles are
matched by the ``grout'' coupling $K$. Upon a careful inspection of
the resulting tile lattice, we notice that the qubits labeled by N and
W in one tile are always connected respectively to qubits S and E in
its neighbor. Therefore, once we fix the embedding of one sublattice
in the unit cell, the embedding of the other sublattice must be
different, because qubits in one unit cell are only coupled to those
at the same place in the neighboring unit cell. We map the two
sublattices of the tile lattice in the unit cells as illustrated in
Fig. \ref{fig:chimera}b-(ii).
Finally, let us show how to embed the TOFFOLI gate, which corresponds
to a rectangular tile, into the unit cells. The TOFFOLI gate can be
viewed as consisting of two square tiles, thus requiring two unit
cells, and the spins coupled between these two tiles exactly provide
the ancilla bit needed in Hamiltonian of
Eq. (\ref{H-Toffoli}). Similar to the square tiles considered above,
within a unit cell we use additional ferromagnetic couplings to slave
qubits from one column to the other when necessary. The explicit
mapping is shown in Fig. \ref{fig:chimera}b-(iii). As we have already
seen, in order to come up with a proper embedding, one has to
carefully take into account how qubits are coupled to adjacent unit
cells. Putting all of the above ingredients together, we arrive at the
embedding of the entire $4\times4$ tile lattice including TOFFOLI into
the Chimera graph, as shown in Fig. \ref{fig:chimera}a.
\vspace{0.8cm}
{\bf Data availability:}
The data that support the main findings of this study are available from the corresponding author upon request.
{\bf Author contributions:}
All authors contributed to all aspects of this work.
{\bf Competing financial interests:}
The authors declare no competing
financial interests.
\newpage
|
\section{Introduction}
The $\Lambda CDM$ cosmology predicts that the halo of Milky Way like galaxies should contain hundreds of sub-haloes, in which dwarf galaxies may be embedded. Many dwarf galaxies and tidal substructures in the Galactic halo have been discovered in the last decade~\citep[][etc.]{nyetal02,rochapinto2004,Bel06,2009ApJ...700L..61N,bonaca2012}. The most prominent one is the Sagittarius (Sgr) stream, which is believed to be a tidal debris of the disrupting Sgr dwarf galaxy. After it was firstly discovered by \citet{Mateo1996}, the stream has been mapped over $2\pi$ radians on the sky by the 2MASS \citep{Majewski2003} and SDSS \citep{Bel06,Koposov2012} surveys. The stream, which is composed of the leading and the trailing tails, wraps at least once around the Galaxy. Some modeling works even claimed that the stream should wrap more than once \citep{Penarrubia2010,2010ApJ...714..229L}. One prominent feature of the Sgr stream is the bifurcations in both north and south hemisphere.
Other than the large scale and prominent substructures, there are also some local and less prominent substructures in the region of the Sgr stream. Some of them may be related to the Sgr stream, some may be new tidal debris with different origins. \citet[][hereafter C07]{Chou2007} found two substructures in the region of the Sgr leading stream using M giant stars. One is likely the part of the Sgr leading tail with similar metallicity and velocity. The authors further separated the tracer stars as the \emph{best} and the \emph{less certain} subsamples. The former are located between 10 and 20\,kpc, while the latter is located within 5\,kpc in distance. They thought most of the best and less subsample should belong to the Sgr stream because this region is far way from the Galactic disk and no evidence for the existence of other tidal substructure is found from the star count of the M giant stars. The other substructure, which is denoted as the north Galactic cap (\emph{NGC}) group by the authors, shows opposite velocity to the Sgr leading tail and is proposed to be the Sgr trailing tail overlapped with the leading tail in the north.
\citet{Pila2014} photometrically observed three fields around the Sgr leading tail and found that the distance of the Sgr leading tail well matches the nearer branch in the model from \citet{Penarrubia2010}. They claimed that this belongs to a new wrap of the Sgr tidal stream together with another branch in the trailing region claimed in the same paper. However, the distance from \citet{Pila2014} is not consistent with from the north trailing tail in the simulation from~\citet[][hereafter LM10]{2010ApJ...714..229L}, requiring more kinematical data to confirm their conclusion.
It is noted that \citet{heidi07} explored an overdensity at ($\rm\Lambda$, $g_0$) $=$ ($240^\circ$, $16.7$\,mag), whose origin is unclear, from the blue horizontal branch stars in the same region of the Sgr leading tail. $g_0$ is the g-band apparent magnitude with reddening corrected. Addition, they also found a moving group near S297+63-20.5 \footnote{The "S" signifies stream (or substructure), followed by the Galactic longitude, then Galactic latitude, and the apparent magnitude of the reddening corrected $g_0$ turnoff.\cite{heidi07} } with $Vgsr=-76 \pm 10$\,kms$^{-1}$\ and distance of $15.8$ kpc to the Sun. Since the location is quite close to the substructures discussed in \cite{Chou2007} and \citet{Pila2014}, it is curious whether they are relevant with each other.
In this work, we use the M giant stars observed from the LAMOST survey to revisit the less prominent substructures in the region of the north Sgr leading tail. As of June 2014, the LAMOST survey has already obtained more than 4 million stellar spectra with sufficient signal-to-noise ratio (S/N). \citet{ZJ15} developed a template-matching technique to reliably identify the M giant stars from the low resolution spectra.
Applying their method to the LAMOST DR2 data, we obtain 17,589 M giant stars, which is so far the largest sample of the M giant spectra. On the other hand, \citet[hearafter ~Paper I]{2016arXiv160300262L}, employed a photometric-based method to estimate [Fe/H] and distance for the M giant stars.
The paper is organized as following. In section 2, we describe how to select the data and improve the estimation of the metallicity. In section 3, the substructures in the region of the Sgr leading tail is unveiled and featured in spatial-kinematic-metallicity space. In section 4, we discuss the possible origins of the new unveiled substructures. Finally, a brief conclusion is drawn in the last section.
\section{Data}
\subsection{The updated LAMOST M giant star catalog}
The Large Sky Area Multi-Object Fiber Spectroscopic Telescope (LAMOST; also called Guo Shou Jing Telescope) is a National Key Scientific facility built by the Chinese Academy of Sciences \citep{cui2012,zhao2012}. The LAMOST spectroscopic survey, started since 2011, mainly aimed for the understanding of the structure of the Milky Way \citep{2012RAA....12..735D}. Although the standard data processing pipeline provides quite accurate estimation of the stellar parameters for the FGK type stars, it does not reliably identify the M type stars \citep{Luo2015}. In order to identify the M type stars and classify the M giant and dwarf stars from the LAMOST dataset,
\citet{ZJ15} constructed the M-giant templates and identified around 9,000 M-giant stars from LAMOST DR1 dataset. We apply the same method to the LAMOST DR2 dataset and expand the M giant sample to 21,696. Then we purify the sample by excluding a few contaminated K giant stars and M dwarf stars using the criteria of the WISE color index $(W_1-W_2)_0$\ according to~Paper I\ and \citet{ZJ15}, finally we get 17,589 M giant stars. The interstellar reddening is corrected using the same spatial model of the extinction mentioned in~Paper I. We adopt the $E(B-V)$ maps of \citet{Schlegel1998}, in combination with $A_r/E(B-V) = 2.285$ from \citet{Schlafly2011} and $A_\lambda/A(r)$ from \citet{Davenport2014} but has a latitude dependence.
The contamination of 166 carbon stars have also been excluded by cross-matching with the latest LAMOST carbon star catalog (Ji et al. in preparation).
\subsection{Distance and metallicity determination}
Paper I\ developed a photometric-based method to estimate the distance for the M giant stars. Specifically, the absolute magnitude in $J$ band is determined from
\begin{equation}
\rm M_{J}=3.12[(J-K)_{0}^{-2.6}-1]-4.61.
\label{eq:distance}
\end{equation}
For the LAMOST M giant sample used in this work, we adopt the equation (\ref{eq:distance}) to estimate the distance. Based on the high-resolution spectroscopic data, C07\ presented metallicities for 59 M giant stars, which are the member candidates for the Sgr stream according to their positions and velocities. We use this sample to improve the photometric metallicity derived by~Paper I. We apply the photometric cut mentioned in section 3.1 of~Paper I\ to the C07\ sample and obtain 41 stars with high quality in photometry. The mean error of $W_1-W_2$ for these stars is only 0.031 mag. Figure~\ref{fig:feh} shows the metallicities of the 41 M giant stars derived by C07\ as a function of $(W_1-W_2)_0$. Then, the correlation between [Fe/H] and $(W_1-W_2)_0$\ is fitted with the following linear relation:
\begin{equation}
{\rm [Fe/H]}_{\rm WISE}=-2.477\times(W_1-W_2)_{0} -1.083,
\label{eq:metallicity}
\end{equation}
and shown with the black line in the figure. The residual scatter of the linear relation, as shown in the inset of Figure~\ref{fig:feh}, is 0.24\,dex. As a comparison, the red line shows the linear relationship obtained from the APOGEE M giant stars in~Paper I. It is obvious that the linear relationship between $(W_1-W_2)_0$\ and [Fe/H] in~Paper I\ does not well fit the data from C07.
In~Paper I, the selection of the M giant stars is purely dependent on the photometry (see their section 3.3), which may be polluted by the K giant contaminations. This difference also could be caused by population effect. Considering the large fraction of K giant stars in the APOGEE data, this contamination may not be ignored and hence, to some extent, induce systematic bias in the [Fe/H]--$(W_1-W_2)_0$\ correlation. For the M giant sample from C07, the M giant identification has been confirmed from the high-resolution spectra and thus should not be affected by the K giant stars. Therefore, the updated linear relationship based on the C07\ sample (the black line) should be more reliable. In the rest of this paper, we use this improved metallicity estimates for the LAMOST M giant stars.
\section{Results}\label{sect:result}
In order to better compare with the Sgr stream, we adopt the Sgr coordinates originally introduced by \citet{Majewski2003}. The equator of the coordinate system is defined by the mid-plane of the Sgr stream. The latitude $\rm B$\ of the Sgr coordinates is parallel with the mid-plane of the stream as viewed from the Sun. The zero-point of the longitude $\rm\Lambda$\ is at the core of the Sgr dwarf galaxy and $\rm\Lambda$\ increases toward the direction of the trailing tidal tail. The coordinates with $\rm\Lambda$$<180^\circ$ are mostly located in the south of the Galactic disk mid-plane and the coordinates with $\rm\Lambda$$>180^\circ$ are mostly in the north.
The line-of-sight velocity measured from the M giant spectra is with respect to the Sun. We convert it to $V_{gsr}$, the line-of-sight velocity with respect to the Galactic standard of rest, using the following equation:
\begin{equation}
V_{gsr}=V_{los}+10{\rm cos}l{\rm cos}b+225.2{\rm sin}l{\rm cos}b+7.2{\rm sin}b,
\end{equation}
in according to~\citep{2012ApJ...757..151L,1998MNRAS.298..387D}.
Figure~\ref{fig:Mgiant_full_lambda} shows the line-of-sight velocities (top panel) and distances (bottom panel) along $\rm\Lambda$\ for the LAMOST DR2 M giant stars. The green circles show all the M giant stars located with $-15^\circ<$$\rm B$$<15^\circ$ and heliocentric distance larger than $10$\,kpc. The red lines indicate the Sgr stream from \citet[][hereafter B14]{Belokurov2014}, who marked them with sub-giant, red giant branch, and blue horizontal branch stars. And the grey dots are the simulation data containing the last wrap of the stream from LM10. The black diamonds are from the substructures of C07, with the distance estimated using the same way as in~Paper I. In the region of $90^\circ<$$\rm\Lambda$$<150^\circ$, which targets the trailing tail, the M giant stars fit quite well in both velocity and distance with either the simulation data from LM10\ or the observed data from B14. The clumpy data located at $130^\circ<$$\rm\Lambda$$<200^\circ$ and $V_{gsr}$$\sim0$\,kms$^{-1}$ are mostly the contamination from the disk population since this region crosses through the Galactic disk in the anti-center direction. However, in the top panel, it is noted that a narrow tail located within $130^\circ<$$\rm\Lambda$$<150^\circ$ with quite small velocity dispersion is relatively isolated out of the clumpy disk contaminations. These stars are likely members of the leading tail.
In the region of $200^\circ<$$\rm\Lambda$$<320^\circ$, the stars are separated into two groups in distance. The one with distance larger than 20\,kpc is consistent with both LM10\ and B14\ in distance and $V_{gsr}$\ and hence belongs to the Sgr leading tail. The other stars located between 10 and 20\,kpc in distance is not identified by B14\ and neither the distance nor the velocity is in agreement with the simulation data from LM10. Some of the stars in this group are well overlapped with the Sgr leading tail in velocity as shown in the top panel of figure~\ref{fig:Mgiant_full_lambda}. Some stars are well overlapped with the Sgr leading tail in velocity are from substructure of C07. A few stars shift from the Sgr leading tail by about 100kms$^{-1}$ toward the positive velocity are NGC group of C07. Therefore, it seems that the stars with distance at 10--20\,kpc may be further separated into two groups, one with similar $\rm\Lambda$--$V_{gsr}$\ trend as the Sgr leading tail, but located in much nearer distance, the other with larger $V_{gsr}$\ than the Sgr leading tail by $\sim100$\,kms$^{-1}$\ at the same $\rm\Lambda$\ and overlapped with the \emph{NGC}\ group from C07.
The LAMOST M giant stars well indicate the Sgr tidal stream in both north and south tails. This can therefore better constrain the dynamics of the tidal stream. However, before addressing the orbital properties of the Sgr tidal stream, it is very important to clarify whether those not prominent but clearly displayed substructures, i.e. the two groups of stars located within 10--20\,kpc, are related with the tidal stream.
\subsection{Substructures in $210^\circ<$$\rm\Lambda$$<290^\circ$}
In order to further investigate the confusing velocity substructures shown in figure~\ref{fig:Mgiant_full_lambda}, we select the stars within $210^\circ<$$\rm\Lambda$$<290^\circ$ to avoid possible contaminations from the thick disk and the some over-density. Indeed, the stars located at $290^\circ<$$\rm\Lambda$$<320^\circ$ well overlap with the RR Lyrae substructure discovered by~\cite{Duffau2014}. At the other end, the region of $200^\circ<$$\rm\Lambda$$<210^\circ$ is very close to the Galactic disk in the Galactic anti-center region. The rest of the region well overlaps the Sgr leading tail in the longitude. The stars in interest are located between 10 and 20\,kpc, while the Sgr leading tail is well identified at the distance between 20 and 60 kpc.
As seen from the top panel of figure~\ref{fig:Mgiant_full_lambda}, both the stars in interest and the distance-identified Sgr leading tail contribute to the group with velocity from $\sim0$\,kms$^{-1}$\ at $\rm\Lambda$$\sim290^\circ$ to $\sim-150$\,kms$^{-1}$\ at $\rm\Lambda$$\sim210^\circ$. However, only the stars located between 10 and 20 kpc contribute to the group with the velocity from $\sim+200$\,kms$^{-1}$\ at $\rm\Lambda$$\sim290^\circ$ to $0$\,kms$^{-1}$\ at $\rm\Lambda$$\sim210^\circ$. A apparent concern is that the group of stars with larger velocity are possibly contaminated by the thick disk. Therefore, we firstly investigate the distribution of the metallicity for these stars and remove those with high possibility to be the thick disk stars.
Figure~\ref{fig:feh_compare} shows the metallicity distribution for the M giant stars in three ranges of distances, 0--10\,kpc (top panel), 10--20\,kpc (middle panel), and beyond $20$\,kpc (bottom panel). In principle, the thick disk stars should dominate the metallicity distribution function for the nearest stars. Indeed, the peak of the metallicity distribution for the stars within 10\,kpc in distance is at around -0.7\,dex, well consistent with the thick disk populations. In the distance larger than 20\,kpc (the bottom panel), a metal-poor population dominates the metallicity distribution with almost no stars fell in [Fe/H]$>-0.75$\,dex. In the distance between 10 and 20\,kpc, the metallicity distribution function seems a mixture of the thick disk stars with typical metallicity distribution as shown in the top panel and the metal-poor population represented by the bottom panel. If we select the M giant stars with [Fe/H]$<-0.75$\,dex in the distance between 10 and 20\,kpc, we can significantly reduce the contaminations from the thick disk. Assume that the stars within 10 kpc are all from the thick disk. This is obviously a quite crucial assumption. Then we can read from the top panel that about 20\% stars are within [Fe/H]$<-0.75$\,dex. Given that in the middle panel, all stars with [Fe/H]$>-0.75$\,dex are from the thick disk, the possible contamination from the thick disk in the region of [Fe/H]$<-0.75$\,dex is only about 15\%, according to the star counts in the panel. Therefore, cutting out the stars with [Fe/H]$>-0.75$\,dex can effectively avoid the effect of the thick disk in velocity.
We then quantify the distribution of $V_{gsr}$\ for the M giant stars with $210^\circ<$$\rm\Lambda$$<290^\circ$ and [Fe/H]$<-0.75$\,dex. Because the velocity $V_{gsr}$\ is correlated with $\rm\Lambda$, it is not easy to directly derive the velocity distribution. To solve this issue, we select the velocity of the Sgr leading tail as the baseline, and derive the distribution of the velocity offset from the baseline at a given $\rm\Lambda$\ for the selected stars. We adopt the velocity trend of the Sgr leading tail from B14\ (the red line in Figure~\ref{fig:Mgiant_full_lambda}) as the baseline. Figure~\ref{fig:structure_vgsr_hist} shows the distribution of the velocity offset (black line) with the bin size of $50$\,kms$^{-1}$. The distribution shows clear bimodality, one peak is located at around 0\,kms$^{-1}$, while another narrower peak is located at about 150\,kms$^{-1}$. In order to investigate whether the bimodality is real or just statistical fluctuation, we calculate the Akaiki?s information criterion (AIC) and Bayesian information criterion (BIC) for the mixed Gaussian models with various components. We find the BIC (AIC) for 1-, 2-, and 3-Gaussian models are 16.5 (15.6), 11.7 (10.2), and 18.6 (16.2), respectively. According to the definition of BIC and AIC, the model with minimum value is the best model for the data. This suggests that a two-Gaussian model is the best one to fit the data.
A two-Gaussian model is then fitted to the distribution (yellow line) and indicates that the best fit peaks are $-3$\,kms$^{-1}$\ with velocity dispersion $83$\,kms$^{-1}$ and $+162$\,kms$^{-1}$\ with velocity dispersion of $26$\,kms$^{-1}$, respectively. According to the bimodal fitting, we separate the stars into two groups at $V_{gsr}-V{gsr,Sgr}=100$\,kms$^{-1}$\ and display their $\rm\Lambda$--distance--$V_{gsr}$\ distribution in Figure~\ref{fig:vgsr_compare1}. The stars with velocity offset smaller than $100$\,kms$^{-1}$\ are denoted as \emph{Vel-3+83}\ and marked as blue filled circles and those with offset larger than $100$\,kms$^{-1}$\ are denoted as the \emph{Vel+162+26}\ and marked as red filled circles. For better comparison, Figure~\ref{fig:vgsr_compare1} also indicates the NGC group from C07\ (red diamonds), the possible thick disk M giant stars with [Fe/H]$>-0.75$\,dex in 10--20\,kpc(yellow dots), and the M giant member candidates of the Sgr leading tail located beyond 30\,kpc (green crosses).
First, the velocity dispersion for the substructure~\emph{Vel-3+83}\ is $83$\,kms$^{-1}$, which is consistent with the large velocity dispersion shown with the red points in figure 2 of C07. The correlation between $V_{gsr}$\ and $\rm\Lambda$\ is also quite similar to the sample from C07. Second, although the trend of $V_{gsr}$\ along with $\rm\Lambda$\ is also comparable with the Sgr leading tail beyond 30\,kpc in distance, the velocity dispersion measured from our sample is significantly larger than the Sgr leading tail, whose dispersion is only $36$\,kms$^{-1}$\ (but see further discussion at section~\ref{sect:disc_structA}). The M giant stars belonging to this group are listed in table~\ref{tab:nearbybranch}.
Second, the \emph{NGC}\ group stars from C07\ are located at exactly the same velocity as the substructure~\emph{Vel+162+26}. Considering that the two groups of stars are also at similar distance as shown in the bottom panel of Figure~\ref{fig:Mgiant_full_lambda}, they are likely from the same substructure. The purple straight line shows the linear fitting of $V_{gsr}$\ along $\rm\Lambda$\ for the substructure \emph{Vel+162+26}\ using the data both from this work and C07. Because the \emph{NGC}\ group stars in C07\ are actually start from 5\,kpc, two more stars located slightly smaller than 10\,kpc in the same sky area and velocity regime are added to \emph{Vel+162+26}. The candidates of the~\emph{Vel+162+26}\ are listed in table~\ref{tab:NGC}.
Finally, we present that the metal-rich M giant stars (yellow dots) do have similar kinematics to the thick disk. For this purpose, we build a oversimplified thick disk kinematical model with azimuthal velocity of $170$\,kms$^{-1}$\ and naively assume that all the stars are located at 15 kpc from the Sun. We calculate the line-of-sight velocity at various $\rm\Lambda$\ for the thick disk in the following way \citep{2012ApJ...757..151L}:
\begin{eqnarray}
V_{td}=-{{\sin l}\over{|\sin l |}}V_{rot}\sin a\cos b,\nonumber\\
r= \sqrt{d^2+8^2-2 \times8 \times d \times \cos l},\\
a=\arccos({{d^2+r^2-8^2}\over{2dr}}).\nonumber
\label{eq:disk_velocity}
\end{eqnarray}
where $8$\,kpc is the adopted distance from the Sun to the Galactic center, $d$ is the distance to the observed star (in this case $15$\,kpc), $r$ is the distance from the Galactic center, $a$ is the azimuth angle of the observed star with respect to the Sun--Galactic center line, $V_{rot}$ is the adopted azimuthal velocity of the thick disk ($170$\,kms$^{-1}$), and $V_{td}$ is the predicted line-of-sight velocity with respect to the Galactic standard of rest. The yellow line in Figure~\ref{fig:vgsr_compare1} shows the predicted line-of-sight velocity for the thick disk. The metal-rich stars within 10--20\,kpc (yellow dots) are well consistent with it, implying that most of the stars are exactly the thick disk stars, except 2 of them located beyond $200$\,kms$^{-1}$. Moreover, neither the substructure~\emph{Vel-3+83}\ nor \emph{Vel+162+26}\ show any similarity to the model velocity of the thick disk. This again confirms that both the substructures are not likely contributed by the thick disk.
\section{Discussion}\label{sect:discussion}
In this section we clarify whether the two substructures located in $210^\circ<$$\rm\Lambda$$<290^\circ$ and distance between 10 and 20\,kpc are related to the Sgr tidal stream.
\subsection{\emph{Vel-3+83}}\label{sect:disc_structA}
The substructure \emph{Vel-3+83}\ shows very similar velocity to the Sgr leading tail but the velocity dispersion is quite larger than the Sgr leading tail. However, when the member stars of \emph{Vel-3+83}\ are mapped to the $V_{gsr}$\ vs. distance plane (see figure~\ref{fig:dist_vgsr1}), we find that the large velocity dispersion of the \emph{Vel-3+83}, shown as the blue filled circles, is due to the rapid variation of $V_{gsr}$\ when distance changes. The actual velocity dispersion can be derived from the velocity scattering at each given distance bin. We then obtain that the velocity dispersion is only 30.4\,kms$^{-1}$, quite comparable with the stars in the Sgr leading tail (green crosses). Moreover, the substructure \emph{Vel-3+83}\ seems continuously connected with the Sgr leading tail at the distance of 20\,kpc. However, putting \emph{Vel-3+83}\ and the leading tail in 3-dimensional space, the connection at 20\,kpc produces a folding 3D structure, which should not be a realistic tidal stream.
According to the $\rm\Lambda$--distance--$V_{gsr}$\ relation, we are able to constrain the direction of motion of the substructure~\emph{Vel-3+83}. \emph{Vel-3+83}\ member stars (color dots) and the Sgr leading tails (gray dots) from LM10\ are plotted in $X$--$Z$ plane in the left panel of figure~\ref{fig:branch_cartoon}. The colors of the \emph{Vel-3+83}\ stars indicate $V_{gsr}$. It shows that at around ($X$, $Z$)=(20, 10)\,kpc, $V_{gsr}$\ of the \emph{Vel-3+83}\ is toward the Sun, implying that the substructure \emph{Vel-3+83}\ is moving to negative $X$ direction. At around ($X$, $Z$)=(5, 10)\,kpc, $V_{gsr}$\ is roughly zero, meaning that at this point the substructure is roughly moving along the tangental direction with respect to the line of sight. Combine these information together, we can infer that \emph{Vel-3+83}\ is likely moving from right to left in the $X$--$Z$ plane (toward the inner Galaxy), showing as a red arrow in the figure.
It is obvious that the direction of motion for \emph{Vel-3+83}\ is opposite to any of the Sgr tidal tails. Hence, it is very difficult to attribute the substructure to a part or branch of the Sgr tidal tail. Alternatively, the substructure \emph{Vel-3+83}\ could be a disrupting satellite of the Sgr dwarf galaxy. However, the distance from the \emph{Vel-3+83}\ to the leading tail is as large as 40\,kpc, which seems too far to be a satellite of the Sgr dwarf galaxy. Therefore, we infer that it is very likely a new tidal debris not related to the Sgr dwarf galaxy.
It is noted that~\citet{heidi07} discovered an unknown over-density from the BHB stars at ($\rm\Lambda$, $g_0$)=(240$^\circ$, 16.7). \emph{Vel-3+83}\ is exactly located at the same location of the over-density (see the comparison in the bottom panel of figure~\ref{fig:vgsr_compare1}). Table~\ref{tab:nearbybranch} also shows that most of the member M giant stars of \emph{Vel-3+83}\ are located above the mid-plane of the Sgr tidal stream ($\rm B$$>0^\circ$), in agreement with the figure 3 in \citet{heidi07}.
It is also worthy to note that the ambiguous arm found by \citet{Pila2014}, who claimed that it is a new wrap of the Sgr stream without velocity information, covers the same sky area as the substructure~\emph{Vel-3+83}. Therefore, it seems that what the authors found is not part of the Sgr tidal debris, but the same tidal debris unveiled in this work.
\subsection{\emph{Vel+162+26}}
Figure~\ref{fig:vgsr_compare1} also displays the member stars of the substructure~\emph{Vel+162+26}\ with red filled circles combined with the NGC group members from C07\ with red diamonds. C07\ argued that these stars are likely the dynamically old Sgr members from the wrapped trailing tail. Similar to \emph{Vel-3+83}\, we can also estimate the approximate direction of motion for \emph{Vel+162+26}\ using the relationship of $\rm\Lambda$--distance--$V_{gsr}$\ and find that \emph{Vel+162+26}\ is coarsely moving toward smaller $X$ and larger $Z$, as shown in the right panel of figure~\ref{fig:branch_cartoon}. This direction of motion is opposite to all the Sgr tidal streams predicted by the simulation data from LM10\ and thus seems against the statement by C07. Therefore, we propose that this substructure is also likely a new tidal debris not relating to the Sgr tidal stream.
\section{CONCLUSION}
Using 17,000 M giant stars from the LAMOST DR2, we are able to map the Sgr tidal stream in the whole northern sky. Both the leading tail and the trailing tail are well sampled. These data will be very important to constrain the orbit of the Sgr tidal stream. Before addressing the orbital properties of the Sgr tidal stream, we investigate whether there is any more weaker and less prominent substructures nearby the Sgr tidal stream and if there is any, whether it belongs to the Sgr.
We find two substructures in the north Galactic cap region around the Sgr leading tail. The substructure~\emph{Vel-3+83}\ shows very similar $\rm\Lambda$--$V_{gsr}$\ relation to the leading tail with nearer distance between 10 and 20\,kpc. However, based on the analysis of the spatial position and the possible direction of motion, it is very likely a new tidal debris not belonging to the Sgr leading tail, nor the trailing tail in north. We point out that the spatial position of \emph{Vel-3+83}\ is consistent with the unknown over-density discovered by \citet{heidi07}.
The other substructure~\emph{Vel+162+26}\ shows positive $V_{gsr}$\ in the similar range of distance. It well overlaps with the \emph{NGC}\ group found by C07, who attribute it to the earlier wrapped Sgr trailing tail. However, when we map the stars from both \emph{Vel+162+26}\ and the NGC group of C07\ to the $X$--$Z$ plane, we find that the direction of motion of this combined substructure is also not in consistent with any branch of the Sgr tidal stream. Therefore, we propose that this is another weak tidal debris not relating to the Sgr tidal stream.
Our proposal for the substructure \emph{Vel-3+83}\ and \emph{Vel+162+26}\ is based on that the simulation of the Sgr tidal process from LM10\ is relatively correct in the distance and velocity. However, if the simulation has bias from the real Sgr tidal stream, which is still far from completely observed, then our conclusion has to be revisited.
In this work, we also revise the relationship between metallicity and WISE color index $(W_1-W_2)_0$, using 41 M giant stars with reliable metallicity estimates from the high resolution spectra (C07).
In the next two years, LAMOST will finish its survey and we expect that it will expand the M giant sample by a factor of 2. Then, we will have much more sample to better address the two substructure and give tighter constraints on their origins.
\section{ACKNOWLEDGEMNTS}
We thank Martin Smith for his kind support on this project.
This work is supported by the Strategic Priority Research Program ``The Emergence of Cosmological Structures" of the Chinese Academy of Sciences grant No. XDB09000000 and the National Key Basic Research Program of China 2014CB845700. C.L. acknowledges the National Natural Science Foundation of China (NSFC) under grants 11373032, 11333003, and U1231119.
J.L, J.L.H and J.Z. thank the NSFC under grants 11173044, 11503066 and the Shanghai Natural Science Foundation 14ZR1446900.Guoshoujing Telescope (the Large Sky Area Multi- Object Fiber Spectroscopic Telescope LAMOST) is a National Major Scientific Project built by the Chinese Academy of Sciences. Funding for the project has been provided by the National Development and Reform Commission. LAMOST is operated and managed by the National Astronomical Observatories, Chinese Academy of Sciences.
|
\section{Introduction}
Given an input sequence, {\it segmentation} is the problem of identifying and assigning tags to its subsequences.
Many natural language processing (NLP) tasks can be cast into the segmentation problem, like named entity recognition \cite{okanohara-EtAl:2006:COLACL}, opinion extraction \cite{yang-cardie:2012:EMNLP-CoNLL}, and Chinese word segmentation \cite{andrew:2006:EMNLP}.
Properly representing {\it segment} is critical for good segmentation performance.
Widely used sequence labeling models like conditional random fields \cite{Lafferty:2001:CRF:645530.655813} represent the contextual information of the segment boundary as a proxy to entire segment and achieve segmentation by labeling input units (e.g. words or characters) with boundary tags.
Compared with sequence labeling model, models that directly represent segment are attractive because they are not bounded by local tag dependencies and can effectively adopt segment-level information.
Semi-Markov CRF (or semi-CRF) \cite{NIPS2005_427} is one of the models that directly represent the entire segment.
In semi-CRF, the conditional probability of a semi-Markov chain on the input sequence is explicitly modeled, whose each state corresponds to a subsequence of input units, which makes semi-CRF a natural choice for segmentation problem.
However, to achieve good segmentation performance, conventional semi-CRF models require carefully hand-crafted features to represent the segment.
Recent years witness a trend of applying neural network models to NLP tasks.
The key strengths of neural approaches in NLP are their ability for modeling the compositionality of language and learning distributed representation from large-scale unlabeled data.
Representing a segment with neural network is appealing in semi-CRF because various neural network structures \cite{Hochreiter:1997:LSM:1246443.1246450} have been proposed to compose sequential inputs of a segment and the well-studied word embedding methods \cite{DBLP:journals/corr/MikolovSCCD13} make it possible to learn entire segment representation from unlabeled data.
In this paper, we combine neural network with semi-CRF and make a thorough study on the problem of representing a segment in neural semi-CRF.
\textcite{DBLP:journals/corr/KongDS15} proposed a segmental recurrent neural network (SRNN) which represents a segment by composing input units with RNN.
We study alternative network structures besides the SRNN.
We also study segment-level representation using {\it segment embedding} which encodes the entire segment explicitly.
We conduct extensive experiments on two typical NLP segmentation tasks: named entity recognition (NER) and Chinese word segmentation (CWS).
Experimental results show that our concatenation alternative achieves comparable performance with the original SRNN but runs 1.7 times faster and our neural semi-CRF greatly benefits from the segment embeddings.
In the NER experiments, our neural semi-CRF model with segment embeddings achieves an improvement of 0.7 F-score over the baseline and the result is competitive with state-of-the-art systems.
In the CWS experiments, our model achieves more than 2.0 F-score improvements on average.
On the PKU and MSR datasets, state-of-the-art F-scores of 95.67\% and 97.58\% are achieved respectively.
We release our code at \url{https://github.com/ExpResults/segrep-for-nn-semicrf}.
\section{Problem Definition}
Figure \ref{fig:ne-and-cws} shows examples of named entity recognition and Chinese word segmentation.
For the input word sequence in the NER example, its segments ({\it ``Michael Jordan'':PER, ``is'':NONE, ``a'':NONE, ``professor'':NONE, ``at'':NONE, ``Berkeley'':ORG}) reveal that ``Michaels Jordan'' is a person name and ``Berkeley'' is an organization.
In the CWS example, the subsequences (\begin{CJK*}{UTF8}{gkai}``浦东/Pudong'', ``开发/development'', ``与/and'', ``建设/construction''\end{CJK*}) of the input character sequence are recognized as words.
Both NER and CWS take an input sequence and partition it into disjoint subsequences.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\columnwidth,trim={0cm 0.4cm 7.8cm 12cm},clip]{necws}
\caption{Examples for named entity recognition (above) and Chinese word segmentation (below).}
\label{fig:ne-and-cws}
\end{figure}
Formally, for an input sequence $\mathbf{x}=( x_1, .., x_{|\mathbf{x}|} )$ of length $|\mathbf{x}|$, let $x_{a:b}$ denote its subsequence $(x_a,...,x_b)$.
A {\it segment} of $\mathbf{x}$ is defined as $(u, v, y)$ which means the subsequence $x_{u:v}$ is associated with label $y$.
A {\it segmentation} of $\mathbf{x}$ is a {\it segment} sequence $\mathbf{s} = (s_1,..,s_p)$, where $s_j=(u_j,v_j,y_j)$ and $u_{j+1}=v_j+1$.
Given an input sequence $\mathbf{x}$, the {\it segmentation} problem can be defined as the problem of finding $\mathbf{x}$'s most probable {\it segment} sequence $\mathbf{s}$.
\begin{figure*}[t]
\begin{subfigure}[b]{0.241\textwidth}
\centering
\includegraphics[width=\columnwidth,trim={3cm 2.8cm 5cm 5cm},clip]{std}
\caption{semi-CRF}\label{fig:std}
\end{subfigure}
~
\begin{subfigure}[b]{0.241\textwidth}
\centering
\includegraphics[width=\columnwidth,trim={3cm 2.8cm 5cm 3.9cm},clip]{rnn}
\caption{SRNN}\label{fig:rnn}
\end{subfigure}
~
\begin{subfigure}[b]{0.241\textwidth}
\centering
\includegraphics[width=\columnwidth,trim={3cm 2.8cm 5cm 3.9cm},clip]{cnn}
\caption{SCNN}\label{fig:cnn}
\end{subfigure}
~
\begin{subfigure}[b]{0.241\textwidth}
\centering
\includegraphics[width=\columnwidth,trim={3cm 2.8cm 5cm 3.9cm},clip]{concate}
\caption{SCONCATE}\label{fig:concate}
\end{subfigure}
\caption{An illustration for the semi-CRF, SRNN, SCNN and SCONCATE.
In these figures, circles represent the inputs, blue rectangles represent {\it factors} in graphic model and yellow rectangles represent generic nodes in the neural network model.}
\label{fig:semi-crf-vs-srnn}
\end{figure*}
\section{Neural Semi-Markov CRF}
Semi-Markov CRF (or semi-CRF, Figure \ref{fig:std}) \cite{NIPS2005_427} models the conditional probability of $\mathbf{s}$ on $\mathbf{x}$ as
\[
p(\mathbf{s}|\mathbf{x})=\frac{1}{Z(\mathbf{x})}\exp\{W \cdot G(\mathbf{x},\mathbf{s})\}
\]
where $G(\mathbf{x},\mathbf{s})$ is the feature function, $W$ is the weight vector and $Z(\mathbf{x})=\sum_{\mathbf{s'}\in \mathbf{S}} \exp\{W \cdot G(\mathbf{x}, \mathbf{s'})\}$ is the normalize factor of all possible {\it segmentations} $\mathbf{S}$ over $\mathbf{x}$.
By restricting the scope of feature function within a segment and ignoring label transition between segments (0-order semi-CRF), $G(\mathbf{x},\mathbf{s})$ can be decomposed as $G(\mathbf{x}, \mathbf{s})=\sum_{j=1}^p g(\mathbf{x},s_j)$ where $g(\mathbf{x},s_j)$ maps segment $s_j$ into its representation.
Such decomposition allows using efficient dynamic programming algorithm for inference.
To find the best segmentation in semi-CRF, let $\alpha_j$ denote the best segmentation ends with $j$\textsuperscript{th} input and $\alpha_j$ is recursively calculated as
\[\mathbf{\alpha}_j=\max_{l=1..L,y} \Psi(j-l,j,y) + \mathbf{\alpha}_{j-l-1}\]
where $L$ is the maximum length manually defined and $\Psi(j-l,j,y)$ is the transition weight for $s=(j-l,j,y)$ in which $\Psi(j-l,j,y)=W\cdot g(\mathbf{x}, s)$.
Previous semi-CRF works \cite{NIPS2005_427,okanohara-EtAl:2006:COLACL,andrew:2006:EMNLP,yang-cardie:2012:EMNLP-CoNLL} parameterize $g(\mathbf{x},s)$ as a sparse vector, each dimension of which represents the value of corresponding feature function.
Generally, these feature functions fall into two types: 1) the {\it CRF style features} which represent input unit-level information such as ``the specific words at location $i$'' 2) the {\it semi-CRF style features} which represent segment-level information such as ``the length of the segment''.
\textcite{DBLP:journals/corr/KongDS15} proposed the segmental recurrent neural network model (SRNN, see Figure \ref{fig:rnn}) which combines the semi-CRF and the neural network model.
In SRNN, $g(\mathbf{x}, s)$ is parameterized as a bidirectional LSTM (bi-LSTM).
For a segment $s_j=(u_j,v_j,y_j)$, each input unit $x$ in subsequence $(x_{u_j},..,x_{v_j})$ is encoded as {\it embedding} and fed into the bi-LSTM.
The rectified linear combination of the final hidden layers from bi-LSTM is used as $g(\mathbf{x}, s)$.
\textcite{DBLP:journals/corr/KongDS15} pioneers in representing a segment in neural semi-CRF.
Bi-LSTM can be regarded as ``neuralized'' {\it CRF style features} which model the input unit-level compositionality.
However, in the SRNN work, only the bi-LSTM was employed without considering other input unit-level composition functions.
What is more, the {\it semi-CRF styled} segment-level information as an important representation was not studied.
In the following sections, we first study alternative input unit-level composition functions (\ref{sec:alt-inp-rep}).
Then, we study the problem of representing a segment at segment-level (\ref{sec:seg-rep}).
\subsection{Alternative Seg-Rep. via Input Composition}\label{sec:alt-inp-rep}
\subsubsection{Segmental CNN}
Besides recurrent neural network (RNN) and its variants, another widely used neural network architecture for composing and representing variable-length input is the convolutional neural network (CNN) \cite{Collobert:2011:NLP:1953048.2078186}.
In CNN, one or more filter functions are employed to convert a fix-width segment in sequence into one vector.
With filter function ``sliding'' over the input sequence, contextual information is encoded.
Finally, a pooling function is used to merge the vectors into one.
In this paper, we use a filter function of width 2 and max-pooling function to compose input units of a segment.
Following SRNN, we name our CNN segment representation as SCNN (see Figure \ref{fig:cnn}).
However, one problem of using CNN to compose input units into segment representation lies in the fact that the max-pooling function is insensitive to input position.
Two different segments sharing the same vocabulary can be treated without difference.
In a CWS example, \begin{CJK*}{UTF8}{gkai}``球拍卖'' (racket for sell) and ``拍卖球'' (ball audition)\end{CJK*} will be encoded into the same vector in SCNN if the vector of \begin{CJK*}{UTF8}{gkai}``拍卖''\end{CJK*} that produced by filter function is always preserved by max-pooling.
\subsubsection{Segmental Concatenation}
Concatenation is also widely used in neural network models to represent fixed-length input.
Although not designed to handle variable-length input, we see that in the inference of semi-CRF, a maximum length $L$ is adopted, which make it possible to use padding technique to transform the variable-length representation problem into fixed-length of $L$.
Meanwhile, concatenation preserves the positions of inputs because they are directly mapped into the certain positions in the resulting vector.
In this paper, we study an alternative concatenation function to compose input units into segment representation, namely the SCONCATE model (see Figure \ref{fig:concate}).
Compared with SRNN, SCONCATE requires less computation when representing one segment, thus can speed up the inference.
\subsection{Seg-Rep. via Segment Embeddings}\label{sec:seg-rep}
For segmentation problems, a segment is generally considered more informative and less ambiguous than an individual input.
Incorporating segment-level features usually lead performance improvement in previous semi-CRF work.
Segment representations in Section \ref{sec:alt-inp-rep} only model the composition of input units.
It can be expected that the segment embedding which encodes an entire subsequence as a vector can be an effective way for representing a segment.
In this paper, we treat the segment embedding as a lookup-based representation, which retrieves the embedding table with the surface string of entire segment.
With the entire segment properly embed, it is straightforward to combine the segment embedding with the composed vector from the input so that multi-level information of a segment is used in our model (see Figure \ref{fig:with-seg}).
However, how to obtain such embeddings is not a trivial problem.
\begin{figure}[t]
\centering
\includegraphics[width=0.7\columnwidth,trim={2cm 2.8cm 2cm 3.9cm},clip]{with_seg}
\caption{Our neural semi-CRF model with segment representation from input composition and segment embeddings.}\label{fig:with-seg}
\end{figure}
A natural solution for obtaining the segment embeddings can be collecting all the ``correct'' segments from training data into a lexicon and learning their embeddings as model parameters.
However, the in-lexicon segment is a strong clue for a subsequence being a correct segment, which makes our model vulnerable to overfitting.
Unsupervised pre-training has been proved an effective technique for improving the robustness of neural network model \cite{Erhan:2010:WUP:1756006.1756025}.
To mitigate the overfitting problem, we initialize our segment embeddings with the pre-trained one.
Word embedding gains a lot of research interest in recent years \cite{DBLP:journals/corr/MikolovSCCD13} and is mainly carried on English texts which are naturally segmented.
Different from the word embedding works, our segment embedding requires large-scale segmented data, which cannot be directly obtained.
Following \textcite{wang-EtAl:2011:IJCNLP-2011} which utilize automatically segmented data to enhance their model, we obtain the auto-segmented data with our neural semi-CRF baselines (SRNN, SCNN, and SCONCATE) and use the auto-segmented data to learn our segment embeddings.
Another line of research shows that machine learning algorithms can be boosted by ensembling {\it heterogeneous} models.
Our neural semi-CRF model can take knowledge from heterogeneous models by using the segment embeddings learned on the data segmented by the heterogeneous models.
In this paper, we also obtain the auto-segmented data from a conventional CRF model which utilizes hand-crafted sparse features.
Once obtaining the auto-segmented data, we learn the segment embeddings in the same with word embeddings.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth,trim={0cm 0cm 3.5cm 11.2cm},clip]{tran}
\caption{An example for fine-tuning decreases the generalization power of pre-trained segment embedding.
``1994\_World\_Cup'' does not occur in the training data and its similarity with ``1998\_World\_Cup'' is broken because ``1998\_World\_Cup'' is tuned.}\label{fig:wo-ft}
\end{figure}
A problem that arises is the fine-tuning of segment embeddings.
Fine-tuning can learn a task-specific segment embeddings for the segments that occur in the training data, but it breaks their relations with the un-tuned out-of-vocabulary segments.
Figure \ref{fig:wo-ft} illustrates this problem.
Since OOV segments can affect the testing performance, we also try learning our model without fine-tuning the segment embeddings.
\subsection{Model details}
In this section, we describe the detailed architecture for our neural semi-CRF model.
\subsubsection{Input Unit Representation}
Following \textcite{DBLP:journals/corr/KongDS15}, we use a bi-LSTM to represent the input sequence.
To obtain the input unit representation, we use the technique in \textcite{dyer-EtAl:2015:ACL-IJCNLP} and separately use two parts of input unit embeddings: the pre-trained embeddings $E^p$ without fine-tuning and fine-tuned embeddings $E^t$.
For the $i$th input, $E^p_i$ and $E^t_i$ are merged together through linear combination and form the input unit representation
\[I_i=\textsc{ReLU}(W^\mathcal{I} [E_i^p ; E_i^t] + b^\mathcal{I})\]
where the notation of $W[X_1;..;X_n]$ equals to $X_1,..,X_n$'s linear combination $W_1 X_1 + .. + W_n X_n$ and $b^I$ is the bias.
After obtaining the representation for each input unit, a sequence $(I_1,...,I_{|\mathbf{x}|})$ is fed to a bi-LSTM.
The hidden layer of forward LSTM $\overrightarrow{H_i}$ and backward LSTM $\overleftarrow{H_i}$ are combined as \[H_i=\textsc{ReLU}(W^\mathcal{H}[\overrightarrow{H_i};\overleftarrow{H_i}]+b^\mathcal{H})\]
and used as the $i$\textsuperscript{th} input unit's final representation.
\subsubsection{Segment Representation}
Given a segment $s_j=(u_j,v_j,y_j)$, a generic function \textsc{SComp}$(H_{u_j},...,H_{v_j})$ stands for the segment representation that composes the input unit representations $(H_{u_j},..,H_{v_j})$.
In this work, \textsc{SComp} is instantiated with three different functions: SRNN, SCNN and SCONCATE.
Besides composing input units, we also employ the segment embeddings as segment-level representation.
Embedding of the segment $s_j=(u_j,v_j,y_j)$ is denoted as a generic function \textsc{SEmb}$(x_{u_j}...x_{v_j})$ which converts the subsequence $(x_{u_j},...,x_{v_j})$ into its embedding through a lookup table.
At last, the representation of segment $s_j$ is calculated as
\[S_j=\textsc{ReLU}(W^\mathcal{S}[\textsc{SComp}_j;\textsc{SEmb}_j;E^Y_{y_j}]+b^\mathcal{S})\]
where $E^Y$ is the embedding for the label of a segment.
\begin{table}[t]
\small
\centering
\begin{tabular}{r|l}
\hline
fixed input unit embedding $E^p_i$ size & 100 \\
fine tuned input unit embedding $E^t_i$ size & 32 \\
input unit representation $I_i$ size & 100 \\
LSTM hidden layer $H_i$ size & 100 \\
seg-rep via input composition {\sc SComp} & 64 \\
seg-rep via segment embedding {\sc SEmb} & 50 \\
label embedding $E^Y_{y_i}$ size & 20 \\
final segment representation $S_i$ size & 100 \\
\hline
\end{tabular}
\caption{Hyper-parameter settings}
\label{tbl:hyper-parameter}
\end{table}
Throughout this paper, we use the same hyper-parameters for different experiments as listed in Table \ref{tbl:hyper-parameter}.
\subsubsection{Training Procedure}
In this paper, negative log-likelihood is used as learning objective.
We follow \textcite{dyer-EtAl:2015:ACL-IJCNLP} and use stochastic gradient descent to optimize model parameters.
Initial learning rate is set as $\eta_0=0.1$ and updated as $\eta_t=\eta_0/(1+0.1t)$ on each epoch $t$.
Best training iteration is determined by the evaluation score on development data.
\section{Experiment}
We conduct our experiments on two NLP segmentation tasks: named entity recognition and Chinese word segmentation.
\subsection{Dataset and Word Embedding}
For NER, we use the CoNLL03 dataset which is widely adopted for evaluating NER models' performance.
F-score is used as evaluation metric.\footnote{{\tt conlleval} script in CoNLL03 shared task is used.}
For CWS, we follow previous study and use three Simplified Chinese datasets: PKU and MSR from 2\textsuperscript{nd} SIGHAN bakeoff and Chinese Treebank 6.0 (CTB6).
For the PKU and MSR datasets, last 10\% of the training data are used as development data as \cite{pei-ge-chang:2014:P14-1} does.
For CTB6 data, recommended data split is used.
We convert all the double byte digits and letters in the PKU data into single byte.
Like NER, CWS performance is evaluated by F-score.\footnote{{\tt score} script in 2\textsuperscript{nd} SIGHAN bakeoff is used.}
Unlabeled data are used to learn both the input unit embeddings (word embedding for NER, character embedding for CWS) and segment embeddings.
For NER, we use RCV1 data as our unlabeled English data.
For CWS, Chinese gigawords is used as unlabeled Chinese data.
Throughout this paper, we use the word embedding toolkit released by \textcite{ling-EtAl:2015:NAACL-HLT} to obtain both the input unit embeddings and segment embeddings.\footnote{\tt https://github.com/wlin12/wang2vec}
\subsection{Baseline}
We compare our models with three baselines:
\begin{enumerate}
\item {\sc Sparse-CRF}: The CRF model using sparse hand-crafted features.
\item {\sc NN-Labeler}: The neural network sequence labeling model making classification on each input unit.
\item {\sc NN-CRF}: The neural network CRF which models the conditional probability of a label sequence over the input sequence.
\end{enumerate}
BIESO-tag schema is used in all the CRF and sequence labeling models.\footnote{O tag which means OUTSIDE is not adopted in CWS experiments since CWS doesn't involve assigning tags to segments.}
For {\sc Sparse-CRF}, we use the baseline feature templates in \textcite{guo-EtAl:2014:EMNLP2014} for NER and
\textcite{jiang-EtAl:2013:ACL2013}'s feature templates for CWS.
Both {\sc NN-Labeler} and {\sc NN-CRF} take the same input unit representation as our neural semi-CRF models but vary on the output structure and do not explicitly model segment-level information.
\subsection{Comparing Different Input Composition Functions}
\begin{table*}[t]
\small
\centering
\begin{tabular}{r|r||cc | cc cc cc| c}
\hline
\multicolumn{2}{r||}{} & \multicolumn{2}{c|}{NER} & \multicolumn{6}{c|}{CWS} & \\
\multicolumn{2}{r||}{} & \multicolumn{2}{c|}{CoNLL03} & \multicolumn{2}{c}{CTB6} & \multicolumn{2}{c}{PKU} & \multicolumn{2}{c|}{MSR} & \\
\multicolumn{2}{r||}{\it model} & dev & test & dev & test & dev & test & dev & test & spd \\
\hline
\multirow{3}{*}{\it baseline} & \sc NN-Labeler & 93.03 & 88.62 & 93.70 & 93.06 & 93.57 & 92.99 & 93.22 & 93.79 & \bf 3.30 \\
& \sc NN-CRF &\bf 93.06 &\bf 89.08 & 94.33 & 93.65 & 94.09 & 93.28 & 93.81 & 94.17 & 2.72 \\
\cline{2-11}
& \sc Sparse-CRF & 88.87 & 83.43 &\bf 95.68 &\bf 95.08 &\bf 95.85 &\bf 95.06 &\bf 96.09 &\bf 96.54 & \\
\hline
\multirow{3}{*}{\it neural semi-CRF}& \sc SRNN & 92.97 & 88.63 & 94.56 & 94.06 & 94.86 & 93.91 & 94.38 & 95.21 & 0.62 \\
& \sc SCONCATE & 92.96 & 89.07 & 94.34 & 93.96 & 94.41 & 93.57 & 94.05 & 94.53 & 1.08 \\
& \sc SCNN & 91.53 & 87.68 & 87.82 & 87.51 & 79.64 & 80.75 & 85.04 & 85.79 & 1.46 \\
\hline
\end{tabular}
\caption{The NER and CWS results of the baseline models and our neural semi-CRF models with different input composition functions.
{\it spd} represents the inference speed and is evaluated by the number of tokens processed per millisecond.}
\label{tbl:close-result}
\end{table*}
We first consider the problem of representing segments via composing input units and compare different input composition functions.
Results on NER and CWS data are shown in Table \ref{tbl:close-result}.
From this table, the SRNN and SCONCATE achieve comparable results and perform better than the SCNN.
Although CNN can model input sequence at any length, its invariance to the exact position can be a flaw in representing segments.
The experimental results confirm that and show the importance of properly handling the input position.
Considering SCNN's relatively poor performance, we only study SRNN and SCONCATE in the following experiments.
Comparing with {\sc NN-Labeler}, structure prediction models (NN-CRF and neural semi-CRF) generally achieve better performance.
The best structure prediction model outperforms {\sc NN-Labeler} by 0.4\% on NER and 1.11\% averagely on CWS according to Table \ref{tbl:close-result}.
But the difference between the neural structure prediction models is not significant.
NN-CRF performs better than the best neural semi-CRF model (SCONCATE) on NER while the both SRNN and SCONCATE outperform NN-CRF on three CWS datasets.
We address this to the fact either the NN-CRF or the neural semi-CRF merely takes input-level information and not sufficiently adopts segment-level information into the models.
A further comparison on inference speed shows that SCONCATE runs 1.7 times faster than SRNN, but slower than the {\sc NN-Labeler} and NN-CRF, which is resulted from the intrinsic difference in time complexity.
\subsection{Comparing Different Segment Embeddings}
Next we study the effect of different segment embeddings.
Using a segmentation model, we can obtain auto-segmented unlabeled data, then learn the segment embeddings.
In this paper, we tried two segmentation models.
One is the neural semi-CRF baseline which represents segment by composing input and another one is the CRF model using sparse hand-crafted features.
For convenience, we use {\sc SEmb-Homo} and {\sc SEmb-Hetero} to note the segment embeddings learned from their auto-segmented data respectively.
\subsubsection{Effect of Pre-trained Segment Embeddings}
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth,trim={0cm 0.4cm 0cm 0.5cm},clip]{learn_curve}
\caption{Negative log-likelihood (blue lines) and development F-score (red lines) by iterations.
Solid lines show the model with randomly initialized segment embeddings.
Dashed lines show that initialized with pre-trained.}
\label{fig:learning-curve}
\end{figure}
We first incorporate randomly initialized segment embeddings into our model and tune the embeddings along with other parameters.
However our preliminary experiments of adding these embeddings into SRNN witness a severe drop of F-score on the CoNLL03 development set (from 92.97\% to 77.5\%).
A further investigation shows that the randomly initialized segment embeddings lead to severe overfitting.
Figure \ref{fig:learning-curve} shows the learning curve in training the NER model.
From this figure, the model with randomly initialized segment embeddings converge to the training data at about 5\textsuperscript{th} iteration and the development performance stops increasing at the same time.
However, by initializing with {\sc SEmb-Homo}, the development set performance increase to 93\%, which shows the necessity of pre-trained segment embeddings.
\subsubsection{Effect of Fine-tuning Segment Embeddings}
\begin{table}[t]
\small
\centering
\setlength{\tabcolsep}{5.5pt}
\begin{tabular}{r||c|ccc}
\hline
\it model & CoNLL03 & CTB6 & PKU & MSR \\
\hline
SRNN & 92.97 & 94.56 & 94.86 & 94.80 \\
\hdashline[1pt/3pt]
\sc +SEmb-Homo w/FT & 92.97 & 95.83 &\bf 96.70 &\bf 97.32 \\
\sc +SEmb-Homo wo/FT &\bf 93.14 &\bf 95.91 & 96.64 & 96.59 \\
\hline
SCONCATE & 92.96 & 94.34 & 94.41 & 94.05 \\
\hdashline[1pt/3pt]
\sc +SEmb-Homo w/FT & 93.07 & 95.79 &\bf 96.75 &\bf 97.29 \\
\sc +SEmb-Homo wo/FT &\bf 93.36 &\bf 95.88 & 96.50 & 96.44 \\
\hline
OOV & 46.02 & 5.45 & 5.80 & 2.60 \\
\hline
\end{tabular}
\caption{Effect of fine-tuning (FT) segment embedding on development data.
For CoNLL03 data, a named entity is ``out-of-vocabulary'' when it is not included in the training data as a named entity.}
\label{tbl:tuning-effect}
\end{table}
We study the effect of fine-tuning the segment embeddings by imposing {\sc SEmb-Homo} into our model.
Table \ref{tbl:tuning-effect} shows the experimental results on development data.
We find that our models benefit from fixing the segment embeddings on CoNLL03.
While on MSR, fine-tuning the embeddings helps.
Further study on the out-of-vocabulary rate shows that the OOV rate of MSR is very low, thus fine-tuning on segment embeddings help to learn a better task-specified segment representation.
However, on CoNLL03 data whose OOV rate is high, fine-tuning the segment embedding harms the generalization power of pre-trained segment embeddings.
\subsubsection{Effect of Heterogeneous Segment Embeddings}
\begin{figure}[t]
\includegraphics[width=\linewidth,trim={0cm 0.8cm 0cm 0.5cm},clip]{segrep}
\caption{Comparison between models with {\sc SEmb-Homo} and {\sc SEmb-Hetero} on development data.
The rows show different baseline neural semi-CRF models and the columns show whether fine-tuning (FT) the segment embeddings.}
\label{fig:bl-vs-crf}
\end{figure}
In previous sections, our experiments are mainly carried on the segment embeddings obtained from homogeneous models.
In this section, we use our {\sc Sparse-CRF} as the heterogeneous model to obtain {\sc SEmb-Hetero}.
We compare the models with {\sc SEmb-Hetero} and {\sc SEmb-Homo} on the development data in Figure \ref{fig:bl-vs-crf}.
These results show that {\sc SEmb-Hetero} generally achieve better performance than the {\sc SEmb-Homo}.
On the CoNLL03 and MSR dataset, the differences are significant.
Meanwhile, we see that fine-tuning the segment embedding can narrow the gap between {\sc SEmb-Hetero} and {\sc SEmb-Homo}.
\subsubsection{Final Result}
\begin{table}[t]
\small
\centering
\begin{tabular}{r||c |c c c}
\hline
\it model & CoNLL03 & CTB6 & PKU & MSR \\
\hline
\sc NN-labeler & 88.62 & 93.06 & 92.99 & 93.79 \\
\sc NN-CRF & 89.08 & 93.65 & 93.28 & 94.17 \\
\hline
\sc Sparse-CRF & 83.43 & 95.08 & 95.06 & 96.54 \\
\hline
SRNN &88.63 & 94.06 & 93.91 & 95.21 \\
\sc +SEmb-Hetero & 89.59 &\bf 95.48 & 95.60 & 97.39 \\
&\it +0.96 &\it +1.42 &\it +1.69 &\it +2.18 \\
\hdashline[1pt/3pt]
SCONCATE & 89.07 & 93.96 & 93.57 & 94.53 \\
\sc +SEmb-Hetero & \bf 89.77 & 95.42 &\bf 95.67 &\bf 97.58 \\
&\it +0.70 &\it +1.43 &\it +2.10 &\it +3.05 \\
\hline
\end{tabular}
\caption{Comparison between baselines and our neural semi-CRF model with segment embeddings.}
\label{tbl:seg-result}
\end{table}
At last, we compare our neural semi-CRF model leveraging additional segment embeddings with those only represent segment by composing input.
Table \ref{tbl:seg-result} shows the result on the NER and CWS test data.
Style of segment embeddings ({\sc Homo} or {\sc Hetero}) and whether fine-tune it is decided by the development data.
From this result, we see that segment-level representation greatly boost up model's performance.
On NER, an improvement of 0.7\% is observed and that improvement on CWS is more than 2.0\% on average.
We compare our neural semi-CRF model leveraging multi-levels segment representation with other state-of-the-art NER and CWS systems.
Table \ref{tbl:ne-stoa} shows the NER comparison results.
\begin{table}[t]
\small
\centering
\begin{tabular}{r|r||c}
\hline
\it genre & \it model & CoNLL03 \\
\hline
\multirow{2}{*}{\it NN} & \cite{Collobert:2011:NLP:1953048.2078186} & 89.59 \\
& \cite{DBLP:journals/corr/HuangXY15} & 90.10 \\
\hdashline[1pt/3pt]
\multirow{3}{*}{\it non-NN} & \cite{Ando:2005:FLP:1046920.1194905} & 89.31 \\
& \cite{guo-EtAl:2014:EMNLP2014} & 88.58 \\
& \cite{passos-kumar-mccallum:2014:W14-16} & \bf 90.90 \\
\hline
\multicolumn{2}{r||}{our best} & 89.77 \\
\hline
\end{tabular}
\caption{Comparison with the state-of-the-art NER systems.}\label{tbl:ne-stoa}
\end{table}
The first block shows the results of neural NER models and the second one shows the non-neural models.
All these work employed hand-crafted features like capitalization.
\textcite{Collobert:2011:NLP:1953048.2078186}, \textcite{guo-EtAl:2014:EMNLP2014}, and \textcite{passos-kumar-mccallum:2014:W14-16} also utilize lexicon as an additional knowledge resource.
Without any hand-crafted features, our model can achieve comparable performance with the models utilizing domain-specific features.
\begin{table}[t]
\small
\centering
\setlength{\tabcolsep}{5.4pt}
\begin{tabular}{r|r||c c c}
\hline
\it genre & \it model & CTB6 & PKU & MSR \\
\hline
\multirow{4}{*}{\it NN} & \cite{zheng-chen-xu:2013:EMNLP} & - & 92.4 & 93.3 \\
& \cite{pei-ge-chang:2014:P14-1} & & 94.0 & 94.9 \\
& \cite{pei-ge-chang:2014:P14-1} w/bigram & - & 95.2 & 97.2 \\
& \cite{DBLP:journals/corr/KongDS15} & & 90.6 & 90.7 \\
\hdashline[1pt/3pt]
\multirow{4}{*}{\it non-NN} & \cite{Tseng05aconditional} & - & 95.0 & 96.4 \\
& \cite{zhang-clark:2007:ACLMain} & - & 95.1 & 97.2 \\
& \cite{sun-EtAl:2009:NAACLHLT09} & - & 95.2 & 97.3 \\
& \cite{wang-EtAl:2011:IJCNLP-2011} &\bf 95.7 & - & - \\
\hline
\multicolumn{2}{r||}{our best} & 95.48 &\bf 95.67 &\bf 97.58 \\
\hline
\end{tabular}
\caption{Comparison with the state-of-the-art CWS systems.}\label{tbl:cws-stoa}
\end{table}
Table \ref{tbl:cws-stoa} shows the comparison with the state-of-the-art CWS systems.
The first block of Table \ref{tbl:cws-stoa} shows the neural CWS models and second block shows the non-neural models.
Our neural semi-CRF model with multi-level segment representation achieves the state-of-the-art performance on PKU and MSR data.
On CTB6 data, our model's performance is also close to \textcite{wang-EtAl:2011:IJCNLP-2011} which uses semi-supervised features extracted auto-segmented unlabeled data.
According to \textcite{pei-ge-chang:2014:P14-1}, significant improvements can be achieved by replacing character embeddings with character-bigram embeddings.
However we didn't employ this trick considering the unification of our model.
\section{Related Work}
Semi-CRF has been successfully used in many NLP tasks like information extraction \cite{NIPS2005_427}, opinion extraction \cite{yang-cardie:2012:EMNLP-CoNLL} and Chinese word segmentation \cite{andrew:2006:EMNLP,sun-EtAl:2009:NAACLHLT09}.
Its combination with neural network is relatively less studied.
To the best of our knowledge, our work is the first one that achieves state-of-the-art performance with neural semi-CRF model.
Domain specific knowledge like capitalization has been proved effective in named entity recognition \cite{ratinov-roth:2009:CoNLL}.
Segment-level abstraction like whether the segment matches a lexicon entry also leads performance improvement \cite{Collobert:2011:NLP:1953048.2078186}.
To keep the simplicity of our model, we didn't employ such features in our NER experiments.
But our model can easily take these features and it is hopeful the NER performance can be further improved.
Utilizing auto-segmented data to enhance Chinese word segmentation has been studied in \textcite{wang-EtAl:2011:IJCNLP-2011}.
However, only statistics features counted on the auto-segmented data was introduced to help to determine segment boundary and the entire segment was not considered in their work.
Our model explicitly uses the entire segment.
\section{Conclusion}
In this paper, we systematically study the problem of representing a segment in neural semi-CRF model.
We propose a concatenation alternative for representing segment by composing input units which is equally accurate but runs faster than SRNN.
We also propose an effective way of incorporating segment embeddings as segment-level representation and it significantly improves the performance.
Experiments on named entity recognition and Chinese word segmentation show that the neural semi-CRF benefits from rich segment representation and achieves state-of-the-art performance.
\section*{Acknowledgments}
This work was supported by the National Key Basic Research Program of China via grant 2014CB340503 and the National Natural Science Foundation of China (NSFC) via grant 61133012 and 61370164.
\bibliographystyle{named}
|
\section{Introduction}
\label{intro}
A complex manifold is said \textbf{irreducible symplectic} if it is simply connected and the vector space of holomorphic $2$-forms is spanned by a nowhere degenerate form. Irreducible symplectic manifolds form, together with Calabi-Yau manifolds and complex tori, one of the three fundamental classes of K\"ahler manifolds with trivial canonical bundle. We are going to denote by $X$ an irreducible symplectic manifold and by $f\colon X\dashrightarrow X$ a bimeromorphic transformation of $X$.\\
On the second cohomology of $X$ we can define a quadratic form, the Beauville-Bogomolov form, whose restriction to $H^{1,1}(X,\mb R)$ is hyperbolic (i.e. has signature $(1, h^{1,1}(X)-1)$) and which is preserved by the linear pull-back action $f^*$ induced by $f$ on cohomology; the setting is therefore similar to that of a compact complex surface, where the intersection form makes the second cohomology group into a hyperbolic lattice. In the surface case, the action of an automorphism $f\colon S\to S$ on cohomology translates into dynamical properties of $f$ (see Paragraphs \ref{isometries hyperbolic} and \ref{parabolic case} for details), and we can hope to have similar results in the irreducible symplectic case.
If $g\colon M\dashrightarrow M$ is a meromorphic transformation of a compact K\"ahler manifold $M$, for $p=0,1,\ldots ,\dim (M)$ the $p$-th \textbf{dynamical degree} of $g$ is
$$\lambda_p(g):= \limsup_{n\to +\infty}||(g^n)_p^*||^{\frac 1n},$$
where $(g^n)_p^*\colon H^{p,p}(M)\to H^{p,p}(M)$ is the linear morphism induced by $g^n$ and $||\cdot ||$ is any norm on the space $End (H^{p,p}(M))$. Note that in the case of an automorphism, $\lambda_p(f)$ is just the maximal modulus of eigenvalues of $f^*_p$.
Let $g\colon M\dashrightarrow M$ be a bimeromorphic transformation of a compact K\"ahler manifold. A meromorphic fibration $\pi\colon M\dashrightarrow B$ onto a compact K\"ahler manifold $B$ such that $\dim B\neq 0, \dim X$ is called \textbf{$g$-invariant} if there exists a bimeromorphic transformation $h\colon B\dashrightarrow B$ such that $\pi\circ g=h\circ \pi$.\\
\centerline{\xymatrix{
M \ar@{-->}[d]^\pi \ar@{-->}[r]^g & M \ar@{-->}[d]^\pi\\
B \ar@{-->}[r]^h & B
}}
The transformation $g$ is said to be \textbf{primitive} (see \cite{MR3431659}) if it admits no invariant fibration.
In the surface case, an automorphism whose action on cohomology has infinite order admits an invariant fibration onto a curve if and only if all the dynamical degrees are equal to $1$ (Theorem \ref{fibrations surfaces}). Our main result establish an analogue of the "only if" direction.
\begin{THM}
\label{thm A}
Let $X$ be an irreducible symplectic manifold, $f\colon X\dashrightarrow X$ a bimeromorphic transformation with at least one dynamical degree $>1$. Then
\begin{enumerate}
\item $f$ is primitive;
\item $f$ admits at most $\dim(X)+b_2(X)-2$ periodic hypersurfaces;
\item the generic orbit of $f$ is Zariski-dense.
\end{enumerate}
\end{THM}
Here a hypersurface $H\subset X$ is said to be $f$-periodic if its strict transform $(f^n)^*H$ by some iterate of $f$ is equal to $H$.
\begin{rem}
Point $(2)$ follows from point $(1)$ and \cite[Theorem B]{MR2727612}; point $(3)$ follows from point $(1)$ and \cite[Theorem 4.1]{MR2400885}, but is proven here as a lemma (Lemma \ref{dense orbits}).
\end{rem}
In order to prove the Main Theorem, we establish a result on the dynamics of birational transformations of projective manifolds that has its own interest.
\begin{PROP}
\label{thm B}
Let $X,B$ be projective manifolds, $f\colon X\dashrightarrow X, g\colon B\dashrightarrow B$ birational transformations and $\pi\colon X\to B$ a non-trivial fibration such that $\pi\circ f=g\circ \pi$. If the generic orbit of $g$ is Zariski-dense and the generic fibre of $\pi$ is of general type, then
\begin{enumerate}
\item $\pi$ is isotrivial over an open dense subset $U\subset B$;
\item there exists an étale cover $U'\to U$ such that the induced fibration $X'=U'\times _U \pi^{-1}(U)$ is trivial: $X'\cong U'\times F$ for a fibre $F$;
\item the images by the natural morphism $X'\to \pi^{-1}(U)$ of the varieties $U'\times \{ pt\}$ are $f$-periodic; in particular the generic orbit of $f$ is not Zariski-dense.
\end{enumerate}
\end{PROP}
\begin{rem}
Point $(1)$ is equivalent to point $(2)$ by \cite[Proposition 2.6.10]{MR2247603}.
\end{rem}
In Section \ref{dyn deg} we recall the definition and main results about dynamical degrees, in the absolute and relative context; Section \ref{sect HK} is consecrated to irreducible symplectic manifolds, with a focus on the invariance of the Beauville-Bogomolov form under the action of a birational transformation; in Section \ref{proof main results} and \ref{sec: invariant subvarieties} we prove the Main Theorem and the Proposition above; Section \ref{alternative approach} presents a different approach to the proof of the Main Theorem, which allows to prove a slightly weaker version of it.
\subsection{Acknowledgements}
I would like to thank Serge Cantat for proposing the topic and strategy of this paper, and for his precious help in all the stages of its redaction; I am also grateful to Mathieu Romagny for the fruitful conversations about Hilbert schemes.
\section{Dynamical degrees}
\label{dyn deg}
Throughout this section $M$ will be a compact K\"ahler manifold of dimension $d$.
\subsection{Definition and entropy}
Let $f\colon X\dashrightarrow Y$ be a dominant meromorphic map between compact K\"ahler manifolds; the map $f$ is then holomorphic outside its indeterminacy locus $\mathcal I\subset X$, which has codimension at least $2$. The closure $\Gamma $ of its graph over $X\setminus \mathcal I$ is an irreducible analytic subset of dimension $d$ in $X\times Y$. Let $\pi_X,\pi_Y$ denote the restrictions to $\Gamma$ of the projections from $X\times Y$ to $X$ and to $Y$ respectively; then $\pi_X$ induces a biholomorphism $\pi_X^{-1}(X\setminus \mathcal I)\cong X\setminus \mathcal I$ and we can identify $f$ with $\pi_Y\circ \pi_X^{-1}$.\\
Let $\alpha$ be a smooth $(p,q)$-form on $Y$; we define the pull-back of $\alpha$ by $f$ as the $(p,q)$-current (see \cite{demailly1997complex} for the basic theory of currents) on $X$
$$f^*\alpha := (\pi_X)_*(\pi_Y^*\alpha).$$
It is not difficult to see that if $\alpha$ is closed (resp. positive), then so is $f^* \alpha$, so that $f$ induces a linear morphism between the Hodge cohomology groups. This definition of pull-back coincides with the usual one when $f$ is holomorphic.
Remember that the $p$-th dynamical degree of a dominant meromorphic map $f\colon M\dashrightarrow M$ are defined as
$$\lambda_p(f)= \limsup_{n\to +\infty}||(f^n)_p^*||^{\frac 1n}.$$
Thanks to the above definition of pull-back, one can prove that
$$\lambda_p(f)=\lim_{n\to +\infty} \left( \int_M (f^n)^*\omega^p \wedge \omega^{d-p}\right)^{\frac 1n}$$
for any K\"ahler form $\omega$. See \cite{MR2119243}, \cite{MR2920277} for details.
The $p$-th dynamical degree measures the exponential growth of the volume of $f^n(V)$ for subvarieties $V\subset M$ of dimension $p$ \cite{MR2139697}.
\begin{rem}
By definition $\lambda_0(f)=1$; $\lambda_d(f)$ coincides with the topological degree of $f$: it is equal to the number of points in a generic fibre of $f$.
\end{rem}
\begin{rem}
Let $f$ be an automorphism. Then we have $(f^n)^*=(f^*)^n$, so that $\lambda_p(f)$ is the maximal modulus of eigenvalues of the linear automorphism $f^*_p\colon H^{p,p}(M,\mb R)\to H^{p,p}(M,\mb R)$; since $f^*$ also preserves the positive cone $\mathcal K_p\subset H^{p,p}(M,\mb R)$, a theorem of Birkhoff \cite{MR0214605} implies that $\lambda_p(f)$ is a positive real eigenvalue of $f^*_p$.\\
It should be noted however that in the bimeromorphic setting we have in general $(f^n)^*\neq (f^*)^n$.
\end{rem}
\begin{rem}
\label{inverse degrees}
If $f$ is bimeromorphic we have
$$\lambda_p(f)=\lambda_{d-p}(f^{-1}).$$
Indeed, for $f$ biregular we have
$$\int_M (f^n)^*\omega^p \wedge \omega^{d-p}=\int_M (f^{-n})^*(f^n)^*\omega^p \wedge (f^{-n})^*\omega^{d-p}=\int_M \omega^p\wedge (f^{-n})^*\omega^{d-p},$$
which proves the equality by taking the limit.\\
If $f$ is only bimeromorphic, for all $n$ we can find two dense open subsets $U_n,V_n\subset M$ such that $f^n$ induces an isomorphism $U_n\cong V_n$; by the definition of pull-back the measures $(f^n)^*\omega^p \wedge \omega^{d-p}$ and $\omega^p\wedge (f^{-n})^*\omega^{d-p}$ have no mass on any proper closed analytic subset, so that
$$\int_M (f^n)^*\omega^p \wedge \omega^{d-p}=\int_{U_n} (f^n)^*\omega^p \wedge \omega^{d-p}=\int_{V_n} \omega^p\wedge (f^{-n})^*\omega^{d-p}=\int_M \omega^p\wedge (f^{-n})^*\omega^{d-p},$$
which proves the equality in the bimeromorphic case as well.
\end{rem}
The main interest in the definition of dynamical degrees lies in the following theorem by Yomdin and Gromov \cite{MR1095529}.
\begin{thm}
If $f\colon M\to M$ is an automorphism, then the topological entropy of $f$ is given by
$$h_{top}(f)=\max_{p=0,\ldots, d} \log\lambda_p(f).$$
\end{thm}
The topological entropy is a positive real number which measures the disorder created by iterations of $f$.\\
It is also possible to give a definition of topological entropy in the bimeromorphic context (see \cite{MR2391122}), but in this situation we only have
$$h_{top}(f)\leq \max_{p=0,\ldots, d} \log\lambda_p(f).$$
\subsection{Relative setting}
Dinh, Nguy\^en and Truong have studied the behaviour of dynamical degrees in the relative setting (\cite{MR2851870} and \cite{MR2989646}). Throughout this paragraph we denote by $f\colon M\dashrightarrow M$ a meromorphic transformation of a compact K\"ahler manifold $M$ of dimension $d$, by $\pi\colon M\dashrightarrow B$ a meromorphic fibration onto a compact K\"ahler manifold $B$ of dimension $k$ and by $g\colon B\dashrightarrow B$ a meromorphic transformation such that
$$g\circ \pi=\pi\circ f.$$
The $p$-th \textbf{relative dynamical degree} of $f$ is defined as
$$\lambda_p(f|\pi)=\limsup_{n\to +\infty}\left( \int_M (f^n)^*\omega_M^p\wedge \pi^*\omega_B^k \wedge \omega_M^{d-p-k} \right)^{\frac 1n},$$
where $\omega_M$ and $\omega_B$ are arbitrary K\"ahler forms on $M$ and $B$ respectively. In particular $\lambda_p(f|\pi)=0$ for $p>d-k$.
Roughly speaking, $\lambda_p(f|\pi)$ gives the exponential growth of $(f^n)^*$ acting on the subspace of classes in $H^{p+k,p+k}(M,\mb R)$ that can be supported on a generic fibre of $\pi$; if $M$ is projective, it also represents the growth of the volume of $f^n(V)$ for $p$-dimensional subvarieties $V\subset \pi^{-1}(b)$ of a generic fibre of $\pi$.
\begin{rem}
\label{bimeromorphic invariant}
Dynamical degrees and relative dynamical degrees are bimeromorphic invariants \cite{MR2851870}. In other words, if there exist bimeromorphic maps $\phi\colon M\dashrightarrow M'$, $\psi\colon B\dashrightarrow B'$ and a meromorphic fibration $\pi'\colon M'\dashrightarrow B'$ such that $\pi'\circ \phi=\psi\circ \pi$, then
$$\lambda_p(f)=\lambda_p(\phi\circ f\circ \phi^{-1}),\qquad \lambda_q(f|\pi)=\lambda_q(\phi\circ f\circ \phi^{-1}|\pi').$$
\end{rem}
\begin{rem}
\label{relative=fibre}
If $F=g^{-1}(b)$ is a regular, $f$-invariant, non-multiple fibre, then $\lambda_p(f|\pi)=\lambda_p(f_{|F})$ for all $p$ (see \cite{MR2851870}).
\end{rem}
The following theorem is due to Dinh, Nguy\^en and Truong \cite{MR2851870}.
\begin{thm}
\label{dinh nguyen}
Let $M$ be a compact K\"ahler manifold, $f\colon M\dashrightarrow M$ a meromorphic transformation, $\pi\colon M\dashrightarrow B$ a meromorphic fibration and $g\colon B\dashrightarrow B$ a meromorphic transformation such that $\pi\circ f=g\circ \pi$. Then for all $p=0,\ldots \dim(M)$
$$\lambda_p(f)=\max_{q+r=p}\lambda_q(f|\pi)\lambda_r(g).$$
\end{thm}
\subsection{Log-concavity}
Dynamical degrees and their relative counterparts enjoy a log-concavity property (see \cite{khovanskii1979geometry},\cite{MR524795}, \cite{MR1095529}, \cite{MR2920277} for the original result, \cite{MR2851870} for the relative setting).
\begin{prop}
\label{log-concavity}
If $f\colon M\dashrightarrow M$ is a meromorphic dominant map, the sequence $p\mapsto \log \lambda_p(f)$ is concave on the set $\{0,1,\ldots , d\}$; in other words
$$\lambda_p(f)^2\geq \lambda_{p-1}(f)\lambda_{p+1}(f) \qquad \text{for }p=1,\ldots, d-1.$$
Analogously, if $\pi\colon M\dashrightarrow B$ is an $f$-invariant meromorphic fibration, then the sequence $p\mapsto \log \lambda_p(f|\pi)$ is concave on the set $\{0,1,\ldots , \dim(M)-\dim(B)\}$.
\end{prop}
As a consequence we have $\lambda_p\geq 1$ for all $p=0,\ldots, d$; furthermore, there exist $0\leq p\leq p+q\leq d$ such that
\begin{equation}
1=\lambda_0(f)<\cdots < \lambda_p(f)=\lambda_{p+1}(f)=\cdots=\lambda_{p+q}(f)>\cdots > \lambda_d(f).
\end{equation}
\section{Irreducible symplectic manifolds}
\label{sect HK}
We give here the basic notions and properties of irreducible symplectic manifolds (see \cite{MR1963559}, \cite{MR2964480} for details).
\begin{rem}
Because of the non-degeneracy of $\sigma$, one can easily prove that an irreducible symplectic manifold has even complex dimension.
\end{rem}
Throughout this section $X$ denotes an irreducible symplectic manifold of dimension $2n$ and $\sigma$ a non-degenerate holomorphic two-form on $X$.
Here is a list of the known examples of such manifolds that are not deformation equivalent.
\begin{enumerate}
\item Let $S$ be a $K3$ surface, i.e. a simply connected K\"ahler surface with trivial canonical bundle. Then the Hilbert scheme $S^{[n]}=Hilb^n(S)$, parametrizing $0$-dimensional subschemes of $S$ of length $n$, is a $2n$-dimensional irreducible symplectic manifold.
\item Let $T$ be a complex torus of dimension $2$, let $\phi\colon Hilb^n(T)\to \Sym^n(T)$ be the natural morphism and let $s\colon \Sym^n(T)\to T$ be the sum morphism. Then the kernel $K_{n-1}(T)$ of the composition $s\circ \phi$ is an irreducible symplectic manifold of dimension $2n-2$, which is called a \emph{generalized Kummer variety}.
\item O'Grady has found two sporadic examples of irreducible symplectic manifolds of dimension $6$ and $10$.
\end{enumerate}
An irreducible symplectic manifold is said of type $K3^{[n]}$ (respectively of type generalized Kummer) if it is deformation equivalent to $Hilb^n(S)$ for some $K3$ surface $S$ (respectively to $K_{n-1}(T)$ for some two-dimensional complex torus $T$).
\subsection{The Beauville-Bogomolov form}
We can define a natural quadratic form on the second cohomology $H^2(X,\mb R)$ which enjoys similar properties to the intersection form on compact surfaces; for details and proofs see \cite{MR1963559}.
\begin{dfn}
Let $\sigma$ be a holomorphic two-form such that $\int(\sigma\bar \sigma)^n=1$. The \emph{Beauville-Bogomolov quadratic form $q_{BB}$} on $H^2(X,\mb R)$ is defined by
$$q_{BB}(\alpha)=\frac n2 \int_X \alpha^2(\sigma\bar \sigma)^{n-1}+(1-n)\left(\int_X \alpha\sigma^n\bar\sigma^{n-1}\right)\left(\int_X \alpha\sigma^{n-1}\bar \sigma^n\right).$$
\end{dfn}
The Beauville-Bogomolov form satisfies two important properties: first the Beauville-Fujiki relation, saying that there exists a constant $c>0$ such that
$$q_{BB}(\alpha)^n=c\int_X \alpha^{2n} \qquad \text{for all }\alpha\in H^2(X,\mb R).$$
In particular, some multiple of $q_{BB}$ is defined over $\mb Z$.\\
Second, the next Proposition describes completely the signature of the form.
\begin{prop}
\label{signature BB}
The Beauville-Bogomolov form has signature $(3,b_2(X)-3)$ on $H^2(X,\mb R)$.\\
More precisely, the decomposition $H^2(X,\mb R)=H^{1,1}(X,\mb R)\oplus \left(H^{2,0}(X)\oplus H^{0,2}(X)\right)_\mb R$ is orthogonal with respect to $q_{BB}$, and $q_{BB}$ has signature $(1,h^{1,1}(X)-1)$ on $H^{1,1}(X,\mb R)$ and is positive definite on $\left(H^{2,0}(X)\oplus H^{0,2}(X)\right)_\mb R$.
\end{prop}
\begin{rem}
\label{non-fixed q non-negative}
For a divisor $D\in Div(X)$, we define $q_{BB}(D):=q_{BB}(c_1(\mathcal O_X(D)))$.\\
If $D$ is effective and without fixed components, then $q_{BB}(D)\geq 0$. Indeed, let $D'$ be an effective divisor linearly equivalent to $D$ and with no components in common with $D$. We have
$$q_{BB}(D)=\frac n2 \int_{D\cap D'}(\sigma \bar\sigma)^{n-1},$$
where each irreducible component of the intersection $D\cap D'$ is counted with its multiplicity. The integral on the right hand side is non-negative because $\sigma$ is a holomorphic form.\\
If furthermore $D$ is ample, then by Beauville-Fujiki relation $q_{BB}(D)>0$.
\end{rem}
\subsection{Bimeromorphic maps between irreducible symplectic manifolds}
A bimeromorphic map $f\colon M\dashrightarrow M'$ between compact complex manifolds is an isomorphism in codimension $1$ if there exist dense open subsets $U\subset M$ and $U'\subset M'$ such that
\begin{enumerate}
\item $\codim(X\setminus U)\geq 2, \codim (X'\setminus U')\geq 2$;
\item $f$ induces an isomorphism $U\cong U'$.
\end{enumerate}
A \textbf{pseudo-automorphism} of a complex manifold $X$ is a bimeromorphic transformation which is an isomorphism in codimension $1$.
\begin{prop}[Proposition 21.6 and 25.14 in \cite{MR1963559}]
\label{pseudo-aut}
Let $f\colon X\dashrightarrow X'$ be a bimeromorphic map between irreducible symplectic manifolds. Then $f$ is an isomorphism in codimension $1$ and induces a linear isomorphism $f^*\colon H^2(X',\mb Z)\xrightarrow{\sim} H^2(X,\mb Z)$ which preserves the Beauville-Bogomolov form.\\
In particular, the group of birational transformation of an irreducible symplectic manifold $X$ coincides with its group of pseudo-automorphisms and acts by isometries on $H^2(X,\mb Z)$.
\end{prop}
\subsection{Isometries of hyperbolic spaces}
\label{isometries hyperbolic}
Proposition \ref{signature BB} establishes a parallel between the dynamics of automorphisms of compact K\"ahler surfaces and that of bimeromorphic transformations of irreducible symplectic manifolds: in both cases the map on the manifold induces an isomorphism at the level of the integral cohomology group $H^2(X,\mb Z)$ preserving a non-degenerate quadratic form (the intersection form in the surface case and the Beauville-Bogomolov form in the irreducible symplectic one). By Hodge's index theorem, the intersection form on the Picard group of a surface $S$ has signature $(1, \rho(S)-1)$, which leads to a classification of automorphisms of surfaces as loxodromic, parabolic or elliptic depending on their action on the hyperbolic lattice $NS_\mb Z(S)$ (see \cite{MR3289919}).\\
Analogously if $X$ is an irreducible symplectic manifold, the restriction of the Beauville-Bogomolov form to $H^{1,1}(X,\mb R)$ has signature $(1,h^{1,1}(X)-1)$. Since $H^{1,1}(X,\mb R)$ is invariant by the action of a bimeromorphic transformation $f\colon X\dashrightarrow X$, and since the two lines $\mb C\sigma$ and $\mb C\bar\sigma$ are also invariant (the action of $f^*$ being given by multiplication by a complex number of modulus $1$), we can also classify bimeromorphic transformations of irreducible symplectic manifolds depending on their action on $H^2(X,\mb Z)$ as follows.
\begin{dfn}
Let $f\colon X\dashrightarrow X$ a bimeromorphic transformation of an irreducible symplectic manifold (respectively, an automorphism of a compact K\"ahler surface) and denote by $f_1^*$ the linear automorphism of $H^{1,1}(X,\mb R)$ induced by $f$. We say that $f$ is
\begin{itemize}
\item \emph{loxodromic} if $f_1^*$ admits an eigenvalue of modulus strictly greater than $1$ (or, equivalently, if $\lambda_1(f)>1$);
\item \emph{parabolic} if all the eigenvalues of $f_1^*$ have modulus $1$ and $||(f^n)_1^*||$ is not bounded as $n\to +\infty$;
\item \emph{elliptic} if $||(f^n)_1^*||$ is bounded as $n\to +\infty$.
\end{itemize}
\end{dfn}
In each of the cases above, simple linear algebra arguments allow to further describe the situation.\\
Denote by $\mathcal C_{\geq 0}\subset H^{1,1}(X,\mb R)$ (respectively $\mathcal C_0\subset H^{1,1}(X,\mb R)$) the positive (resp. null) cone for the Beauville-Bogomolov (repsectively, intersection) form $q$:
$$\mathcal C_{\geq 0}= \{\alpha\in H^{1,1}(X,\mb R) | q(\alpha)\geq 0\},$$
$$\mathcal C_{0}= \{\alpha\in H^{1,1}(X,\mb R) | q(\alpha)=0\}.$$
$\mathcal C_0$ is called the \emph{isotropic cone} for the Beauville-Bogomolov form.
For a proof of the following result, see \cite{MR2533769} (for irreducible symplectic manifolds) and \cite{MR3289919} (for surfaces).
\begin{thm}
\label{characterization birational}
Let $f\colon X\dashrightarrow X$ a bimeromorphic transformation of an irreducible symplectic manifold (respectively, an automorphism of a compact K\"ahler surface).
\begin{itemize}
\item If $f$ is loxodromic, then $f_1^*$ has exactly one eigenvalue with modulus $>1$ and exactly one eigenvalue with modulus $<1$; these eigenvalues are real, simple and they are the inverse of each another; their eigenspaces are contained in $\mathcal C_0$, they are the only $f_1^*$-invariant lines in $\mathcal C_{\geq 0}$ and they are not defined over $\mb Z$.
\item If $f$ is parabolic, then all eigenvalues of $f^*_1$ are roots of unity; the Jordan form of $f^*$ has exactly one non-trivial Jordan block, which is of dimension $3$ (in other words $||(f^n)_1^*||$ has quadratic growth); for every $\alpha\in H^{1,1}(X,\mb R)$, $(f^n)_1^*\alpha / n^2$ converges to a class contained in $\mathcal C_0$, which (for every $\alpha$ outside a proper subspace of $H^{1,1}(X,\mb R)$) spans the only $f_1^*$-invariant line of $\mathcal C_{\geq 0}$.
\item If $f$ is elliptic, then some iterate of $f_1^*$ is equal to the identity.
\end{itemize}
\end{thm}
\subsection{The parabolic case}
\label{parabolic case}
In the case of surfaces, an automorphism being of parabolic type has a clear geometric interpretation (see \cite{gizatullin1980rational}, \cite{grivaux2013parabolic}, or \cite{MR1867314} for the birational case).
\begin{thm}
\label{fibrations surfaces}
Let $S$ be a compact K\"ahler
surface; an automorphism $f\colon S\to S$ is of parabolic type if and only if there exists an $f$-invariant fibration $\pi\colon S\to C$ onto a nonsingular compact curve $C$.
\end{thm}
We could expect the situation to be similar in the irreducible symplectic context; indeed, Hu, Keum and Zhang have proved a partial analogue to Theorem \ref{fibrations surfaces}, see \cite{MR3431659}:
\begin{thm}
\label{hu zhang}
Let $X$ be a $2n$-dimensional projective irreducible symplectic manifold of type $K3^{[n]}$ or of type generalized Kummer and let $f\in Bir(X)$ be a bimeromorphic transformation which is not elliptic; $f$ is parabolic if and only if it admits a rational Lagrangian invariant fibration $\pi\colon X\dashrightarrow \mb P^n$ such that the induced transformation on $\mb P^n$ is biregular, i.e. there exists $g\in Aut(\mb P^n)$ such that $\pi\circ f=g\circ \pi$.
\end{thm}
The hard direction is to exhibit an invariant fibration for a parabolic transformation. The Main Theorem generalizes the converse, proving that the dynamics of a loxodromic transformation is too complicated to expect an invariant fibration.
\section{Proof of the main results}
\label{proof main results}
Throughout this section, $f\colon X\dashrightarrow X$ denotes a loxodromic bimeromorphic transformation of an irreducible symplectic manifold $X$, $\pi\colon X\dashrightarrow B$ a meromorphic invariant fibration onto a K\"ahler manifold $B$ such that $0<\dim B<\dim X$ and $g\colon B\dashrightarrow B$ the induced transformation of the base.
\centerline{\xymatrix{
X \ar@{-->}[d]^\pi \ar@{-->}[r]^f & X \ar@{-->}[d]^\pi\\
B \ar@{-->}[r]^g & B
}}
The results in this Section are largely inspired by \cite{MR2400885}.
\subsection{Meromorphic fibrations on irreducible symplectic manifolds}
\label{fibrations HK}
We collect here some useful facts about the fibration $\pi$.
\begin{rem}
\label{projective base}
If $B$ is K\"ahler, then it is projective. Indeed, if $B$ wasn't projective, by Kodaira's projectivity criterion and Hodge decomposition
$$H^2(B,\mb C)=H^{2,0}(B)\oplus H^{1,1}(B)\oplus H^{0,2}(B),$$
we would have $H^{2,0}(B)\neq \{0\}$, meaning that $B$ carries a non-trivial holomorphic $2$-form $\sigma_B$. Since the indeterminacy locus of $\pi$ has codimension at least $2$, the pull-back $\pi^*\sigma_B$ can be extended to a global non-trivial $2$-form on $X$ which is not a multiple of $\sigma$, contradicting the hypothesis on $X$.
\end{rem}
Here we use the same conventions as in \cite{MR2400885}: let $\eta\colon \tilde X\to X$ be a resolution of the indeterminacy locus of $\pi$ (see \cite{MR0506253}), and let $\nu\colon \tilde X\to B$ be the induced holomorphic fibration, whose generic fibre is bimeromorphic to that of $\pi$.
\centerline{\xymatrix{
\tilde X \ar[d]^\eta \ar[rd]^\nu&\\
X \ar@{-->}[r]^\pi & B
}}
The pull-back $\pi^*D$ of an effective divisor $D\in Div(B)$ is defined as
$$\pi^*D=\eta_*\nu^*D,$$
where $\eta_*$ is the pushforward as cycles. The pull-back induces linear morphisms $Pic(B)\to Pic(X)$ and $NS(B)\to NS(X)$, and is compatible with the pull-back of smooth forms defined in Section \ref{dyn deg}.
Now let $H\in Pic(B)$ be an ample class, and let $L=\pi^*H$. The pull-back of the complete linear system $|H|$ is a linear system $U\subset |L|$, whose associated meromorphic fibration is exactly $\pi$. In particular, $L$ has no fixed component, and by Remark \ref{non-fixed q non-negative} we have $q_{BB}(L)\geq 0$.
Let $NS(B)\subset H^{1,1}(B,\mb R)$ denote the Neron-Severi group with real coefficients of $B$. The following Lemma is essentially proven in \cite{MR2400885}.
\begin{lemma}
\label{pullback isotropic}
The restriction of the Beauville-Bogomolov form to the pull-back $\pi^*NS(B)$ is not identically zero if and only if the generic fibre of $\pi$ is of general type. If this is the case, then $X$ is projective.
\end{lemma}
\begin{proof}
Remark first that, since the generic fibre of $\nu$ is bimeromorphic to the generic fibre of $\pi$ and the Kodaira dimension is a bimeromorphic invariant, the generic fibre of $\pi$ is of general type if and only if the generic fibre of $\nu$ is.\\
As a second remark, by \cite{moishezon1967criterion} if there exists a big line bundle on a compact K\"ahler manifold $X$, then $X$ is projective.
Suppose that the generic fibre of $\pi$ is of general type. Let $H$ be an ample divisor on $B$ and let $L=\pi^*H$. By \cite{MR2400885}[Theorem 2.3] we have $\kappa(X,L)=\dim(B)+\kappa(F)$, where $F$ is the generic fibre of $\nu$; we conclude that $L$ is big (and in particular $X$ is projective). We can thus write $L=A+E$ for an ample divisor $A$ and an effective divisor $E$ on $X$. Now, if $q$ denotes the Beauville-Bogomolov form, we have
$$q(L)=q(L,A)+q(L,E)\geq q(L,A)= q(A,A)+q(A,E)\geq q(A,A)>0,$$
where the first and second inequalities are consequences of $L$ and $A$ being without fixed components and the last one follows directly from Remark \ref{non-fixed q non-negative}
. This proves the "if" direction.
Now assume that the restriction of $q_{BB}$ to $\pi^*NS(B)$ is not identically zero. Since ample classes generate $NS_\mb R(B)$, there exists an ample line bundle $H\in Pic(B)$ such that, denoting $L=\pi^*H$, $q(L)\neq 0$; furthermore, $L$ is without fixed components, so that $q(L)>0$ by Remark \ref{non-fixed q non-negative}. It follows by\cite{MR2050205}[Theorem 4.3.i] that $L$ is big (thus $X$ is projective), and so is $\eta^*L$ since $\eta$ is a birational morphism. Therefore, the restriction $\eta^*L|_F$ to a generic fibre of $\nu$ is also big (see \cite{MR2095471}[Corollary 2.2.11]).
Now we have
$$\eta^*L=\nu^*H+\sum a_iE_i \qquad \text{for some }a_i\geq 0,$$
where the sum runs over all the irreducible components of the exceptional divisor of $\eta$. The adjunction formula leads to
$$K_F=K_{\tilde X}|_F+\det N^*_{F/\tilde X}=K_{\tilde X}|_F=\sum e_iE_i|_F \qquad \text{for some }e_i> 0,$$
since the conormal bundle $N^*_{F/\tilde X}$ is trivial.\\
This implies that, for some $m>0$, the divisor $mK_F-\eta^*L|_F$ is effective because $\nu^*H|_F$ is trivial. Thus
$$\kappa(F)\geq \kappa(F,\eta^*L|_F)=\dim(F),$$
meaning that $F$ is of general type. This proves the "only if" direction.
\end{proof}
\begin{cor}
\label{pullback line}
If the generic fibre of $\nu$ is not of general type, then $\pi^*NS(B)\subset H^{1,1}(X,\mb R)$ is a line contained in the isotropic cone $\mathcal C_0$.
\end{cor}
\begin{proof}
By Lemma \ref{pullback isotropic}, $\pi^*NS(B)$ is contained in the isotropic cone. The pull-back $L$ of an ample line bundle on $B$ is effective and non-trivial, so that its numerical class is also non-trivial; thus $\pi^*NS(B)$ cannot be trivial. To conclude it suffices to remark that $\pi^*NS(B)$ is a linear subspace of $H^{1,1}(X,\mb R)$, and the only non-trivial subspaces contained in the isotropic cone are lines.
\end{proof}
\subsection{Density of orbits}
The following theorem was proven in \cite{MR2400885}.
\begin{thm}
\label{orbit fibration}
Let $X$ be a compact K\"ahler manifold and let $f\colon X\dashrightarrow X$ be a dominant meromorphic endomorphism. Then there exists a dominant meromorphic map $\pi\colon X\dashrightarrow B$ onto a compact K\"ahler manifold $B$ such that
\begin{enumerate}
\item $\pi \circ f=\pi$;
\item the general fibre $X_b$ of $\pi$ is the Zariski closure of the orbit by $f$ of a generic point of $X_b$.
\end{enumerate}
\end{thm}
\begin{lemma}
\label{pullback pseudo-aut}
Let $\phi\colon X\dashrightarrow Y$, $\psi\colon Y\dashrightarrow Z$ be meromorphic maps between compact complex manifolds. If $\phi$ is an isomorphism in codimension $1$, then for all $D\in Div(Z)$
$$(\psi\circ \phi)^*D=\phi^*\psi^*D.$$
\end{lemma}
\begin{proof}
Let $U\subset X$, $V\subset Y$ two open sets such that $\phi$ induces an isomorphism $U\cong V$ and such that $\codim (X\setminus U)\geq 2,\codim(Y\setminus V)\geq 2$. It is easy to see that, for every effective divisor $D_Y\in Div(Y)$, we have an equality $\phi^*D_Y=\overline{\phi|_U^*(D_Y\cap V)}$; therefore the equality is true for every divisor in $Div(Y)$.\\
Up to shrinking $V$ to some other open subset whose complement has codimension at least $2$, we can suppose that $\psi$ is regular on $V$; therefore the composition $\psi\circ \phi$ is regular on $U$, and since the complement of $U$ has codimension $\geq 2$ in $X$, for all $D\in Div(Z)$ we have
$$(\psi\circ \phi)^*D=\overline{(\psi\circ \phi)|_U^*D}=\overline{\phi|_U^*(\psi|_V^*D)}=\overline{\phi|_U^*(\psi^*D\cap V)}=\phi^*\psi^*D,$$
where the third equality follows again from the fact that the complement of $V$ has codimension at least $2$ in $Y$. This proves the claim.
\end{proof}
Let us prove point $(3)$ of the Main Theorem.
\begin{lemma}
\label{dense orbits}
Let $f\colon X\dashrightarrow X$ be a bimeromorphic loxodromic transformation of an irreducible symplectic manifold. Then the generic orbit of $f$ is Zariski-dense.
\end{lemma}
\begin{proof}
If the claim were false, then by Theorem \ref{orbit fibration} we could construct a commutative diagram\\
\centerline{\xymatrix{
X \ar@{-->}[d]^\pi \ar@{-->}[r]^f & X \ar@{-->}[d]^\pi\\
B \ar[r]^{id} & B
}}
where $\pi$ is a meromorphic map whose general fibre $X_b$ coincides with the Zariski-closure of the $f$-orbit of a generic point of $X_b$. Remark \ref{projective base} applies in the case where the fibres are not connected; therefore the base $B$ is projective.\\
Now, remark that $f^*$ acts as the identity on the space $\pi^*NS(B)\subset NS(X)$, which is defined on $\mb Q$: indeed, for $v\in NS(B)$, we have
$$f^*\pi^*v=(\pi\circ f)^*v=(id_B \circ \pi)^* v =\pi^*v,$$
where the first equality follows from Lemma \ref{pullback pseudo-aut}; by Theorem \ref{characterization birational}, the Beauville-Bogomolov form is negative definite on $\pi^*NS(B)$.\\
Now, let $H\in Pic(B)$ be an ample line bundle and let $L=\pi^*H$. We have seen in \ref{fibrations HK} (again the hypothesis on fibres being connected was irrelevant) that $L$ is a numerically non-trivial line bundle such that $q_{BB}(L)\geq 0$, contradiction. This proves the claim.
\begin{comment}
Up to replacing $X$ by a birational model, we can suppose that $\pi$ is a morphism.\\
Let
$$X \xrightarrow{\pi'} B' \xrightarrow{\rho} B$$
be the Stein factorization of $\pi$, where $\pi'$ has connected fibres and $\rho$ is a finite morphism.
Now, the general fibre $F$ of $\pi$ being of general type, the action of $f|_F$ is finite by Theorem \ref{birational group general type}; this shows that the generic orbit of $f$ is finite. Now apply Theorem \ref{orbit fibration}: there exists a meromorphic fibration $\pi\colon X\dashrightarrow B$ such that the generic fibre is the Zariski closure of the orbit of a generic point in the fibre. The orbit being finite, the generic fibre of $\pi$ is also finite; up to replacing $X$ by a birational model, we may assume that $\pi$ is regular, and the above discussion shows that the orbits over a dense open subset $U\subset B$ have finite cardinality. Then, after shrinking $U$, we can assume that the cardinality of all the fibres is the same (equal to the degree of $\pi|_{\pi^{-1}(U)}$; in particular, some iterate of $f$ is equal to the identity, which contradicts the hypothesis of $f$ being loxodromic.
\begin{comment}
let $U\subset X$ be the open set of points whose $f$-orbit is finite. The map
\begin{align*}
U& \to \mb N\\
p&\mapsto \# \{f^n(p), n\in \mb N\}
\end{align*}
is lower
semi-continuous. This shows that $f$ has finite order, and in particular we have $\lambda_1(f)=1$, a contradiction.
\begin{comment}
Now clearly $\lambda_q(id_B)=1$ for all $q=0,1,\ldots, \dim(B)$. \\
Moreover, if $F=\pi^{-1}(b)$ is a smooth non-multiple fibre, then $\lambda_r(f|\pi)=\lambda_r(f|_F)$ (Remark \ref{relative=fibre}); now by Lemma \ref{key lemma} $F$ is of general type, therefore its group of birational transformations is finite (Theorem \ref{birational group general type}), and thus
$$\lambda_r(f|\pi)=1 \qquad r=0,\ldots , \dim(F).$$
By Theorem \ref{dinh nguyen}, we then have $\lambda_p(f)=1$ for all $p=0,\ldots ,\dim(X)$, contradiction.
\end{comment}
\end{proof}
\subsection{The key lemma}
The following key lemma, together with the Proposition in Section \ref{intro}, implies the Main Theorem.
\begin{lemma}[Key lemma]
\label{key lemma}
Let $X$ be an irreducible symplectic manifold, $f\colon X\dashrightarrow X$ a loxodromic bimeromorphic transformation and $\pi\colon X \dashrightarrow B$ a meromorphic $f$-invariant fibration onto a compact K\"ahler manifold. Then $X$ is projective and the generic fibre of $\pi$ is of general type.
\end{lemma}
\begin{proof}
Let $g\colon B\dashrightarrow B$ be a bimeromorphic transformation such that $g\circ \pi=\pi\circ f$.
Let us define
$$V:=Span\left\{(h \circ \pi)^*NS_\mb R(B)| h \colon B\dashrightarrow B \text{ birational transformation}\right\}\subset NS_\mb R(X).$$
The linear subspace $V$ is clearly defined over $\mb Q$. Since the pull-back by $\pi$ of an ample class is numerically non-trivial, we also have $V\neq\{0\}$.\\
Furthermore, $V$ is $f^*$-invariant: if $v=(h\circ \pi)^*w$ for some $w\in NS(B)$ and for some birational transformation $h\colon B\dashrightarrow B$, then
$$f^*v=f^*(h\circ \pi)^* w=(h\circ \pi\circ f)^*w=(h\circ g\circ \pi)^*w=(\tilde h \circ \pi)^*w,$$
where $\tilde h =h\circ g\colon B\dashrightarrow B$ is a birational transformation and the second equality follows from Lemma \ref{pullback pseudo-aut}.
Now suppose that the generic fibre of $\pi$ is not of general type; we are first going to show that $V$ is contained in the isotropic cone $\mathcal C_0=\{v\in H^{1,1}(X,\mb R)| q_{BB}(v)=0\}$. The generic fibre of the meromorphic fibration $h\circ \pi$ is bimeromorphic to that of $\pi$. By Lemma \ref{pullback isotropic} we know that $(h\circ \pi)^*NS_\mb R(B)$ is contained in the isotropic cone for all birational transformations $h\colon B\dashrightarrow B$. We just need to show that for all birational transformations of $B$ onto itself $h_i,h_j$ and for all $w_i, w_j\in NS_\mb R(B)$ we have
$$q_{BB}((h_i\circ \pi)^*w_i, (h_j\circ \pi)^*w_j)=0.$$
Let $h=h_j\circ h_i^{-1}$, and let $\rho \colon \tilde B\to B$ be a resolution of the indeterminacy locus of $h$; denote by $\tilde h \colon \tilde B \to B$ the induced holomorphic transformation, and let $\tilde \pi=\rho^{-1} \circ h_i \circ \pi\colon X\dashrightarrow \tilde B$; $\tilde \pi$ is a meromorphic fibration onto the birational model $\tilde B$, whose generic fibre is bimeromorphic to that of $g$. Finally, let $\eta\colon \tilde X\to X$ be a resolution of singularities of $\tilde \pi$ and let $\nu\colon \tilde X\to B$ be the induced holomorphic map.
\centerline{
\xymatrix{
\tilde X \ar[dd]^\eta \ar[rd]^\nu&&\\
&\tilde B \ar[d]^\rho \ar[dr]^{\tilde h}&\\
X \ar@{-->}[r]^{h_i\circ \pi} \ar@{-->}[ru]^{\tilde\pi}& B \ar@{-->}[r]^h &B
}}
Now it is clear that $\eta\colon \tilde X\to X$ is a resolution of singularities of both $h_i\circ \pi$ and $h_j\circ \pi=h\circ h_i\circ \pi$. Therefore
$$(h_i\circ \pi)^*w_i=\eta_*\nu^*\rho^*w_i=\tilde \pi^*\rho^*w_i\in \tilde \pi^*NS(\tilde B)$$
and
$$(h_j\circ \pi)^*w_j=\eta_*\nu^*\tilde h^*w_j=\tilde\pi^* \tilde h^*w_j\in \tilde \pi^* NS(\tilde B).$$
Since the fibres of $\tilde \pi$ are not of general type, it suffices to apply Lemma \ref{pullback isotropic} to the fibration $\tilde \pi\colon X\dashrightarrow \tilde B$ to conclude that $q_{BB}((h_i\circ \pi)^*w_i, (h_j\circ \pi)^*w_j)=0$. This proves that $V$ is contained in the isotropic cone.
Now the only non trivial vector subspaces of $NS_\mb R(X)$ contained in the isotropic cone are lines; by Theorem \ref{characterization birational}, $V$ is then an $f^*$-invariant line contained in the isotropic cone and not defined over $\mb Q$. But this contradicts the definition of $V$. We have thus proved that the generic fibre of $\pi$ is of general type.
In order to prove that $X$ is projective it suffices to apply the last part of Lemma \ref{pullback isotropic}.
\end{proof}
By \cite[Corollary 14.3]{MR0506253}) we know that the group of birational transformations of a variety of general type is finite. Therefore, we expect the dynamics of $f$ on the fibres to be simple.
\subsection{Relative Iitaka fibration}
\label{relative fibration}
Before giving the proof of the Main Theorem, we are going to recall the basic results about the relative Iitaka fibration. We will follow the approach of \cite{tsuji2010global} with some elements from \cite{MR0506253}. See also \cite{MR3075000}, \cite{MR0217084}.
Let $X$ be a smooth projective variety, and suppose that some multiple of $K_X$ has some non trivial section. Recall that, for $m>0$ divisible enough, the rational map
\begin{align*}
\phi_{|mK_X|} \colon X & \dashrightarrow \mb P H^0(X,mK_X)^\vee\\
p & \mapsto \{s\in \mb P H^0(X,mK_X) | s(p)=0\}
\end{align*}
has connected fibres. Moreover the rational map $\phi_{|mK_X|}$ eventually stabilize to a rational fibration that we call \emph{canonical fibration} of $X$.
\begin{rem}
\label{induced linear}
If $f\colon X\dashrightarrow X$ is a bimeromorphic transformation of $X$, the pull-back of forms induces a linear automorphism $f^*\colon H^0(X,mK_X)\to H^0(X,mK_X)$. For example, for $m=1$ a section $\sigma \in H^0(X,K_X)$ is a holomorphic $d$-form ($d=\dim X$); $f$ is defined on an open set $U\subset X$ such that $X\setminus U$ has codimension at least $2$. Therefore by Hartogs theorem the pull-back $f|_U^*\sigma$ can be extended to $X$. It is easy to see that the construction is invertible and induces a linear automorphism of $\mb P H^0(X,mK_X)^\vee$ which commutes with the Iitaka fibration:
\centerline{
\xymatrix{
X \ar@{-->}[r]^f \ar@{-->}[d]^{\phi_{|mK_X|}}& X \ar@{-->}[d]^{\phi_{|mK_X|}}\\
\mb P H^0(X,mK_X)^\vee \ar[r]^{\tilde f}& \mb P H^0(X,mK_X)^\vee
}}
\end{rem}
The above construction can be generalized to the relative setting: let $\pi\colon X\to B$ be a regular fibration onto a smooth projective variety $B$, and let $K_{X/B}=K_X\otimes \pi^*K_B^{-1}$ be the relative canonical bundle.\\
For some fixed positive integer $m>0$ (divisible enough), let $\mathcal S=\pi_*(mK_{X/B})^{\vee}$. $\mathcal S$ is a coherent sheaf over $B$; therefore one can construct (generalizing the construction of the projective bundle associated to a vector bundle, see \cite{MR0506253} for details) the algebraic projective fibre space
$$\eta \colon \Proj(\mathcal S)\to B$$
associated to $\mathcal S$, which is a projective scheme (a priori neither reduced nor irreducible) $Y$ over $B$. Its generic geometric fibre $Y_b$ over a generic point $b\in B$ is canonically isomorphic to $\mb P H^0(X_b, mK_{X_b})^\vee $. The Iitaka morphisms $\phi_b\colon X_b\dashrightarrow \mb P H^0(X_b, mK_{X_b})^\vee$ induce a rational map $\phi\colon X\dashrightarrow Y$ over $B$.\\
The \emph{relative canonical fibration} of $X$ with respect to $\pi$ is
\begin{align*}
\phi\colon X & \dashrightarrow Y\\
x\in X_b &\mapsto \left[\{s\in H^0(X_b; mK_{X_b})| s(x)=0\}\right]\in Y_b.
\end{align*}
It can be shown that, for $m$ divisible enough:
\begin{itemize}
\item $\phi$ stabilizes to a certain rational fibration;
\item the image by $\phi$ of the generic fibre $X_b=\pi^{-1}(b)$ of $\pi$ is contained inside the fibre $\eta^{-1}(b)$ of the natural projection $\eta\colon Y\to B$;
\item the restriction of $\phi$ to a generic fibre $X_b$ is birationally equivalent to the canonical fibration of $X_b$.
\end{itemize}
\begin{rem}
\label{induced linear relative}
The construction in Remark \ref{induced linear} can also be generalized to the relative setting: let $f\colon X\dashrightarrow X$ and $g\colon B\dashrightarrow B$ be birational transformations such that $\pi\circ f=g\circ \pi$. For a generic $b\in B$ define
\begin{align*}
\tilde f|_{Y_b}\colon \mb P H^0(X_{b},mK_{X_{b}})^\vee &\dashrightarrow \mb P H^0(X_{g(b)},mK_{X_{g(b)}})^\vee\\
[s^*]&\mapsto \left\{ [s]\in \mb P H^0(X_{g(b)},mK_{X_{g(b)}})| s^*(f^*s)=0 \right\}.
\end{align*}
These are well defined linear automorphisms because, for a fibre $X_b$ of $\pi$ not contained in the indeterminacy locus of $f$, the restriction $f\colon X_b \dashrightarrow X_{g(b)}$ is a birational map, and thus induces a linear isomorphism
$$f^*\colon H^0(X_{g(b)},mK_{X_{g(b)}})\to H^0(X_b,mK_{X_b}).$$
Furthermore the $\tilde f_{X_b}$ can be glued to a birational transformation $\tilde f\colon Y\dashrightarrow Y$ such that $\eta\circ \tilde f=g\circ \eta$.
\end{rem}
Now suppose the generic fibre of $\pi$ is of general type. Since the restriction of $\phi$ to a generic fibre of $g$ is birational onto its image and the images of fibres are disjoint, $\phi$ itself must be birational onto its image; denote by $Z$ the closure of the image of $\phi$ and let $f_Z=\phi\circ f\circ \phi^{-1} \colon Z\dashrightarrow Z$ be the birational transformation induced by $f$. \\
By the above Remark, $f_Z$ is the restriction of the birational transformation $\tilde f\colon Y \dashrightarrow Y$. In particular $f_Z$ induces an isomorphism between generic fibres of $\eta|_Z$.
\centerline{\xymatrix{
X \ar@{-->}[d]^\pi \ar@{-->}[r]^\phi \ar@(ul,ur)[]|{f} & Z \ar@(ul,ur)[]|{f_Z} \ar@{^{(}->}[r] \ar@{-->}[ld]& Y \ar@(ul,ur)[]|{\tilde f} \ar@{-->}[lld]^{\eta}\\
B \ar@(dl,ul)[]|{g}
}}
\vspace{0.3cm}
\subsection{Proof of the Main Theorem}
\begin{lemma}
\label{dense orbit isotrivial}
Let $X,B$ be projective manifolds, $f\colon X\dashrightarrow X$ and $g\colon B\dashrightarrow B$ birational transformations and $\pi\colon X\to B$ a fibration such that $\pi\circ f=g\circ \pi$.\\
\centerline{\xymatrix{
X \ar[d]^\pi \ar@{-->}[r]^f & X \ar[d]^\pi\\
B \ar@{-->}[r]^g & B
}}\\
Assume that the generic fibre of $\pi$ is of general type and that
the generic orbit of $g$ is Zariski-dense. Then all the fibres over a non-empty Zariski open subset of $B$ are isomorphic.
\end{lemma}
\begin{proof}
Denote as before
$$\phi\colon X \dashrightarrow Y$$
the relative Iitaka fibration. We are going to identify $X$ with its birational model $\overline{\phi(X)}$.\\
Let $F=\pi^{-1}(b_0)$ be the fibre of $\pi$ over a point $b_0$ whose orbit is Zariski-dense in $B$, and let
$$\mathfrak{I}:=\mathfrak{Isom}_B(X,F\times B)$$
be the $B$-scheme of isomorphisms over $B$ between $X$ and $F\times B$; the fibre $\mathfrak I_b$
parametrizes the isomorphisms $X_b\cong F$. We can realize $\mathfrak I$ as an open subset of the Hilbert scheme $\mathfrak{Hilb}_B(X\times_B (B\times F))$ by identifying a morphism $X_b\to F$ with its graph in $X_b\times F$. Therefore,
$$\mathfrak I=\coprod _{P\in \mb Q[\lambda]} \mathfrak I^P,$$
where the fibre $\mathfrak I^P_b$ is the (a priori non irreducible and non reduced) quasi-projective scheme of (graphs of) isomorphisms $X_b\xrightarrow{\sim} F$
having fixed Hilbert polynomial $P(\lambda)$; such polynomials are calculated with respect to the restriction to the fibre $X_b\times F$ of a fixed line bundle $L$ on $X\times_B (B\times F)$ relatively very ample over $B$. We shall fix
$$L=H_Y|_{X} \boxtimes_B H_F,$$
where $H_Y$ is a very ample line bundle on $Y$ and $H_F$ is a very ample line bundle on $F$.\\
Now, the pull-back of forms by $f$ induces a linear isomorphism
$$\tilde f_b\colon \mb P H^0(X_b, mK_{X_b})^\vee \xrightarrow{\sim} \mb P H^0(X_{g(b)}, mK_{X_{g(b)}})^\vee $$
between fibres of $\eta\colon Y\xrightarrow{\sim} B$, which restricts to an isomorphism $X_b\to X_{g(b)}$; under the canonical identification of fibres of $\eta$ with $\mb P^N$, $H_Y|_{Y_b}\cong \mathcal O_{\mb P^N}(d)$ (meaning that the section $H_Y|_{Y_b}$ has degree $d$)
for some $d>0$ independent of the fibre. Under the identification, the action of $\tilde f_b$ is linear, so that $\tilde f_b^*(H_Y|_{Y_{g(b)}})$ also has degree $d$ on $\mb P^N$. In particular we have
$$\tilde f_b^*(H_Y|_{X_{g(b)}})=H_Y|_{X_b}.$$
Now take any isomorphism $X_{b_0}\xrightarrow{\sim} F$, which we can identify with its graph $\Gamma\subset X_{b_0}\times F$; the image of $\Gamma$ by the isomorphism $\tilde f_{b_0}\times id_F\colon X_{b_0}\times F\xrightarrow{\sim} X_{g(b_0)}\times F$ is the graph $\Gamma'$ of an isomorphism $X_{g(b_0)}\xrightarrow{\sim} F$. Furthermore, since $(\tilde f_{b_0}\times id_F)^*(L|_{X_{g(b_0)}\times F})=L|_{X_{b_0}\times F}$, $\Gamma'$ has the same Hilbert polynomial as $\Gamma$. Iterating this reasoning we find that for some $P\in \mb Q[\lambda]$ the image of the natural morphism $\psi\colon \mathfrak I^P\to B$ is Zariski-dense.\\
By Chevalley's theorem (\cite[Theorem 3.16]{MR1416564}) we also know that $\psi(\mathfrak I^P)$ is constructible; since every constructible Zariski-dense subset of an irreducible scheme contains a dense open set \cite[Proof of Theorem 3.16]{MR1416564}, we have $X_b\cong F$ for all $b$ in an open dense subset of $B$. This concludes the proof.
\end{proof}
\begin{proof}[Proof of the Proposition in Section \ref{intro}]
\begin{comment}
By Lemma \ref{dense orbit isotrivial}, all the fibres over a dense open subset $U\subset B$ are isomorphic to one another; suppose that $U$ is the maximal open subset with this property. By \cite[Proposition 2.6.10]{MR2247603}, there exists an étale cover $\epsilon \colon U'\to U$ such that the fibration $X':=U'\times X_U$ induced by $\pi_U\colon X_U\to U$ is trivial:
$$X'\cong U'\times F.$$
Suppose first that $g_U\colon U\to U$ is biregular. Then $g_U$ can be lifted to a continuous application $g'\colon U'\to U'$ if and only if $(g\circ \epsilon)_*\pi_1(U')\subset \epsilon_*\pi_1(U')$.\\
\centerline{\xymatrix{
U' \ar[d]^\epsilon \ar@{-->}[r]^{?} \ar[rd]^{g\circ \epsilon} & U' \ar[d]^\epsilon\\
U \ar[r]^{g_U} & U
}}\\
Now $H:=\epsilon_*\pi_1(U')$ is a finite index subgroup of the finitely generated group $G:=\pi_1(U)$; therefore, by Lemma \ref{finitely generated group}, there exist a finite index subgroup $H'\leq H$ such that $g_* (H')=H'$. In other words, up to replacing $U'$ by some étale cover, $g$ can be lifted to $g'\colon U'\to U'$.
Now, if $g_U$ is only birational, let $V\subset U$ be the maximal open set where $g_U$ is defined. We will show first that the codimension of $U\setminus V$ is at least $2$. Suppose the contrary, and let $D\subset U\setminus V$ be a hypersurface where $g_U$ is not defined. Since $B$ is projective, the indeterminacy locus of $g$ has codimension at least $2$; therefore, $g$ is defined on a dense open subset do $D$, and then $g(D)\subset B\setminus U$. This means that
\end{comment}
By Lemma \ref{dense orbit isotrivial}, all the fibres over a dense open subset $U\subset B$ are isomorphic, which shows $(1)$. By \cite[Proposition 2.6.10]{MR2247603}, there exists an étale cover $\epsilon \colon U'\to U$ such that the induced fibration $X'_{U'}:=X\times_U U'$ is trivial:
$$X'_{U'}\cong U'\times F.$$
This shows $(2)$.
Now suppose that the generic fibre $F$ is of general type. This implies that the group $G:=\Aut(F)$ is finite; therefore, for any $x\in F$, we can define the subvariety
$$W^x:=U'\times G\cdot x\subset X'_{U'}.$$
We are going to show that the image of $W^x$ by the cover $\epsilon_X\colon X'_{U'}\to X_U$ is $f$-invariant.
Remark that the fibration $\pi_U\colon X_U\to U$ is locally trivial in the euclidean topology. Let $\{U_i\}_{i\in I}$ be a covering of $U$ by euclidean open subsets such that the restriction of the fibration to each $X_{U_i}$ is trivial: there exist biholomorphisms $X_{U_i}\cong U_i\times F$. Then the subvarieties
$$V_i^x:=U_i\times G\cdot x \subset X_{U_i}$$
patch together to algebraic subvarieties $V^x$ of $X_U$ which are exactly the images of the $W^x$.\\
Now we will prove that the varieties $V^x$ are $f$-invariant. Let $p\in X_U$ be a point where $f$ is defined and such that $g$ is defined on $\pi(p)$, and let $i\in I$ be such that $p\in X_{U_i}$; up to shrinking $U_i$, we can suppose that $g(U_i)\subset U_j$. By an identification $X_{U_i}\cong U_i\times F, X_{U_j}\cong U_j\times F$, we can write $f(x,y)=(g(x),h(x,y))$; here, for all $x$ in on open dense subset of $U_i$, the continuous map $b\mapsto h(x,\bullet)\in \Bir(F)$ is well defined. Since $\Bir(F)$ is a finite (hence discrete) group, $h$ doesn't depend on $x$, which shows that all the varieties $V^x$ are $f$-invariant.
Now remark that the varieties $\epsilon_X(U'\times G\cdot \{x\})$ are the disjoint union of varieties of type $\epsilon_X(U'\times \{y\})$; since the first are $f$-invariant, the latter must be $f$-periodic, which concludes the proof.
\begin{comment}
Let $\{U_i\}_{i\in I}$ be a covering of $U$ by euclidean open subsets such that the restriction of the fibration to each $X_{U_i}$ is trivial: there exist biholomorphisms $X_{U_i}\cong U_i\times F$. We are going to show that, for every $x\in F$, the subvarieties
$$V_i^x:=U_i\times G\cdot x \subset X_{U_i}$$
(where $G\cdot x$ denotes the orbit of $x$ by the finite group $G:=\Aut(F)$) patch together to algebraic subvarieties $V^x$ of $X_U$ which are $f$-invariant. This implies that the generic orbit of $f$ is not Zariski-dense, a contradiction with the hypothesis.\\
Let us first show that the local varieties $V_i^x\subset U_i$ can be patched together to a global (holomorphic) subvariety of $X_U$ for every $x\in F$ such that all the elements of $G$ are defined on $x$. Let $X_{U_i}\cong U_i\times F, X_{U_j}\cong U_j\times F$ be two local trivializations of $\pi$; over every point $b\in U_i\cap U_j$, the change of trivialization is given by an automorphism $\varphi_b$ of $F$. The map $b\mapsto \varphi_b$ being continuous and $\Aut(F)$ being discrete, we have $\varphi_b$ constant for every $b\in U_i\cap U_j$. This implies that the varieties $U_i\times G\cdot x$ are locally well defined regardless of the chosen trivialization. To check that they are globally well defined, remark that the obstruction lies in the holonomy representation $\pi_1(U)\to \Aut(F)$; however the orbit $G\cdot x$ is clearly $\Aut(F)$-invariant, so that the claim is proven.\\
Now let us show that the subvarieties $V^x$ are algebraic. By \cite[Proposition 2.6.10]{MR2247603}, there exists an étale cover $U'\to U$ such that the induced fibration $X'_{U'}:=X\times_U U'$ is trivial:
$$X'_{U'}\times F.$$
Then each $V^x$ is the images of a subvariety $U'\times G\cdot x$ by the natural projection $X'_{U'}\to X$.\\
In order to conclude, we only need to prove that the varieties $V^x$ are $f$-invariant. Let $p\in X_U$ be a point where $f$ is defined, and let $i\in I$ be such that $p\in X_{U_i}$; up to shrinking $U_i$, we can suppose that $g(U_i)\subset U_j$. By an identification $X_{U_i}\cong U_i\times F, X_{U_j}\cong U_j\times F$, we can write $f(x,y)=(g(x),h(x,y))$; here, for all $x$ in on open dense subset of $U_i$, the continuous map $b\mapsto h(x,\bullet)\in \Bir(F)$ is well defined. Since $\Bir(F)$ is a finite (hence discrete) group, $h$ doesn't depend on $x$, which shows that all the varieties $V^x$ are $f$-invariant. This concludes the proof.
\end{comment}
\end{proof}
\begin{comment}
\begin{lemma}
\label{finitely generated group}
Let $G$ be a finitely generated group, $H\leq G$ a finite index subgroup and $\phi\colon G\to G$ an automorphism. Then there exists a finite index subgroup $H'\leq H$ such that $\phi(H')=H'$.
\end{lemma}
\begin{proof}
Let $K$ be the normal core of $H$. Then $K$ is a normal, finite index subgroup, and we are going to show that $\phi^N(K)=K$ for some $N>0$.\\
The subgroups $K,\phi(K), \phi^2(K),\ldots \leq G$ are all normal subgroups of same index $i$; therefore, each of them is the kernel of a surjective group morphism $\psi_j\colon G\to G_j$, where the $G_j$-s are groups of order $i$. The number of possible $G_j$-s is finite, and since $G$ is finitely generated, for a fixed $G_j$ the number of morphisms $G\to G_j$ is also finite. This implies that $\phi^m(K)=\phi^n(K)$ for some $0\leq n<m$. Now let
$$H'=\bigcap _{j=0}^{n-1}\phi^j(K).$$
The subgroup $H'$ has finite index; furthermore, by the above discussion it is clear that $\phi(H')\subset H'$. Since $H'$ and $\phi(H')$ have the same index in $G$, we must have equality.
\end{proof}
\end{comment}
\begin{proof}[Proof of the Main Theorem, point (1)]
Let $X$ be an irreducible symplectic manifold, $f\colon X\dashrightarrow X$ a birational loxodromic transformation, and suppose by contradiction that $f$ is imprimitive: there exist thus a meromorphic fibration $\pi\colon X\dashrightarrow B$ and a bimeromorphic transformation $g\colon B\dashrightarrow B$ such that $\pi\circ f=g\circ \pi$.\\
By Lemma \ref{key lemma}, $X$ is projective and the generic fibre of $\pi$ is of general type. However, we also know by Lemma \ref{dense orbits} that the generic orbit of $f$ is Zariski-dense; therefore, by the Proposition in Section \ref{intro}, the generic fibre of $\pi$ cannot be of general type, a contradiction.
\end{proof}
\begin{comment}
\begin{thm}
\label{dense orbit general type fibre}
Let $X,B$ be projective manifolds, $f\colon X\dashrightarrow X$ and $g\colon B\dashrightarrow B$ birational transformations and $\pi\colon X\to B$ a fibration such that $\pi\circ f=g\circ \pi$.\\
\centerline{\xymatrix{
X \ar[d]^\pi \ar@{-->}[r]^f & X \ar[d]^\pi\\
B \ar@{-->}[r]^g & B
}}\\
If the generic orbit of $f$ is Zariski-dense, then the generic geometric fibre of $\pi$ is not of general type.
\end{thm}
\end{comment}
\section{Invariant subvarieties}
\label{sec: invariant subvarieties}
Let $X$ be a compact complex manifold. If $f\colon X\to X$ is an automorphism, we say that a subvariety $W\subset X$ is \emph{invariant} if $f(W)=W$, or, equivalently, if $f^{-1}(W)=W$. We say that $W\subset X$ is \emph{periodic} if it is invariant for some positive iterate $f^n$ of $f$.\\
Now let $f\colon X\dashrightarrow X$ be a pseudo-automorphism of $X$ (i.e. a bimeromorphic transformation which is an isomorphism in codimension $1$). We say that a hypersurface $W\subset X$ is invariant if the strict transform $f^*W$ of $W$ is equal to $W$ (as a set); since $f$ and $f^{-1}$ don't contract any hypersurface, this is equivalent to $f(W)=W$ (here $f(W)$ denotes the analytic closure of $f|_U(W\cap U)$, where $U\subset X$ is the maximal open set where $f$ is well defined). We say that a hypersurface is periodic if it is invariant for some positive iterate of $f$.
The following Theorem is a special case of \cite{MR2727612}[Theorem B].
\begin{thm}
\label{invariant hypersurfaces fibration}
Let $f\colon X\dashrightarrow X$ be a pseudo-automorphism of a compact complex manifold $X$. If $f$ admits at least $\dim(X)+b_2(X)-1$ invariant hypersurfaces, then it preserves a non-constant meromorphic function.
\end{thm}
\begin{proof}[Proof of the Main Theorem, point (2)]
Let $f\colon X\dashrightarrow X$ be a loxodromic bimeromorphic transformation of an irreducible symplectic manifold $X$ (which is a pseudo-automorphism by \ref{pseudo-aut}).\\
Suppose that $f$ admits more than $\dim(X)+b_2(X)-2$ periodic hypersurfaces; then some iterate of $f$ satisfies the hypothesis of Theorem \ref{invariant hypersurfaces fibration}. Therefore $f^n$ preserves a non-constant meromorphic function $\pi\colon X\to \mb P^1$, and, up to considering the Stein factorization of $\pi$, we can assume that $\pi$ is an $f^n$-invariant fibration onto a curve. This contradicts point $(1)$ of the Main Theorem.
\end{proof}
The following example shows that we cannot hope to obtain an analogue of point $(2)$ of the Main Theorem for higher codimensional subvarieties.
\begin{ex}
Let $f\colon S\to S$ be a loxodromic automorphism of a $K3$ surface $S$, and let $X=Hilb^n(S)$. Then $X$ is an irreducible symplectic manifold and $f$ induces a loxodromic automorphism $f_n$ of $X$. By point $(2)$ of the Main Theorem, $f_n$ admits only a finite number of invariant hypersurfaces. However $f$ admits infinitely many periodic points (\cite{MR1864630},\cite{MR3289919}); if $x$ is a periodic point in $S$, then (the image in $X$ of) $\{x\}^p\times S^{n-p}$ is a periodic subvariety of codimension $2p$.
\end{ex}
Thus we have showed the following Proposition.
\begin{prop}
For all integers $0<p\leq n$, there exist a $2n$-dimensional projective irreducible symplectic manifold $X$ and a loxodromic automorphism $f\colon X\to X$ admitting infinitely many periodic subvarieties of codimension $2p$.
\end{prop}
\section{Appendix: An alternative approach to the Main Theorem}
\label{alternative approach}
In this section we describe a different approach to the proof of the Main Theorem which doesn't require the Proposition in Section \ref{intro}. The result we obtain is actually slightly weaker than the Main Theorem; however this approach allows to prove point $(2)$ and $(3)$, as well as point $(1)$ for automorphisms.
We have already seen in Proposition \ref{characterization birational} that the first dynamical degree of a bimeromorphic transformation $f\colon X\dashrightarrow X$ is either $1$ or an algebraic integer $\lambda$ whose conjugates over $\mb Q$ are $\lambda^{-1}$ and some complex numbers of modulus $1$ (so that $\lambda$ is a quadratic or a Salem number). In the case of automorphisms, the following Proposition from Verbitsky \cite{MR1406664} allows to completely describe all the other dynamical degrees as well.
\begin{prop}
\label{verbitsky}
Let $X$ be an irreducible symplectic manifold of dimension $2n$ and let $SH2(X,\mb C)\subset H^*(X,\mb C)$ be the subalgebra generated by $H^2(X,\mb C)$. Then we have an isomorphism
$$SH^2(X,\mb C)=\Sym ^*H^2(X,\mb C)/\langle \alpha^{n+1}|q_{BB}(\alpha)=0 \rangle$$.
\end{prop}
The following Corollary is due to Oguiso \cite{MR2533769}.
\begin{cor}
\label{remark oguiso}
Let $f\colon X\to X$ be an automorphism of an irreducible symplectic manifold of dimension $2n$. Then for $p=0,1,\ldots, n$
$$\lambda_p(f)=\lambda_{2n-p}(f)=\lambda_1(f)^p.$$
\end{cor}
\begin{proof}
By Proposition \ref{verbitsky} the cup-product induces an injection
$$\Sym^p H^2(X, \mb C)\hookrightarrow H^{2p}(X,\mb C)$$
for $p=1,\ldots, n$.\\
Let $v_1\in H^2(X,\mb C)$ be an eigenvector for the eigenvalue $\lambda=\lambda_1(f)$. Then $v_p:=v_1^p\in H^{2p}(X,\mb C)$ is a non-zero class for $p=1,\ldots, n$ and $f^*v_p=(f^*v_1)^p=\lambda^p v_p$. This implies that $\lambda_p(f)\geq \lambda_1(f)^p$, and we must have equality by log-concavity (Proposition \ref{log-concavity}). This proves the result for $p=0,1,\ldots, n$.\\
Now by Remark \ref{inverse degrees} we have $\lambda_{2n-p}(f)=\lambda_p(f^{-1})$. Applying what we have just proved to $f^{-1}$ we obtain
$$\lambda_{2n-1}(f)=\lambda_1(f^{-1})=\lambda_n(f^{-1})^{1/n}=\lambda_n(f^{-1})^{1/n}=\lambda_1(f)$$
and thus, for $p=0, \ldots ,n$,
$$\lambda_{2n-p}(f)=\lambda_{p}(f^{-1})=\lambda_1(f^{-1})^p=\lambda_1(f)^p,$$
which concludes the proof.
\end{proof}
\begin{lemma}
\label{trivial relative dyn}
Let $X$ be a smooth projective variety, $f\colon X \dashrightarrow X$ a birational transformation of $X$, $\pi\colon X\dashrightarrow B$ a rational $f$-invariant fibration onto a smooth projective variety $B$. If the generic fibre of $\pi$ is of general type, then all the relative dynamical degrees $\lambda_p(f|\pi)$ are equal to $1$ (for $p=0, \ldots , \dim(X)-\dim(B)$).
\end{lemma}
\begin{proof}
Since the Kodaira dimension and the relative dynamical degrees are bimeromorphic invariants (Remark \ref{bimeromorphic invariant}), up to considering a resolution of the indeterminacy locus of $\pi$, we can suppose that $\pi$ is regular.\\
Let
\centerline{\xymatrix{
\phi\colon X \ar[d]^\pi \ar[r] & Y:=\Proj(\pi_*K_{X/Y}^{\otimes m}) \ar[ld]^\eta\\
B }}
be the Iitaka fibration. Since $\phi$ is birational onto its image, denoting $Z\subset Y$ the closure of $\phi(X)$, the claim is equivalent to $\lambda_p(f_Z|\eta_Z)=1$, where $\eta_Z$ denotes the restriction of $\eta$ to $Z$ and $f_Z=\phi\circ f\circ \phi^{-1} \colon Z\dashrightarrow Z$.\\
The construction of Remark \ref{induced linear relative} provides a birational transformation $\tilde f\colon Y\dashrightarrow Y$ extending $f_Z$.
\centerline{\xymatrix{
X \ar[d]^\pi \ar@{-->}[r]^\phi \ar@(ul,ur)[]|{f} & Z \ar@(ul,ur)[]|{f_Z} \ar@{^{(}->}[r] \ar[ld]& Y \ar@(ul,ur)[]|{\tilde f} \ar[lld]^{\eta}\\
B \ar@(dl,ul)[]|{g}
}}
\vspace{0.3cm}
Now we will prove that if $\lambda_p(\tilde f|\eta)=1$ then $\lambda_p(f_Z|\eta_Z)=1$. Let $H_Y\in Pic(Y)$ and $H_B\in Pic(B)$ be ample classes; therefore $H_Y|_Z$ is an ample class on $Z$. The map
\begin{align*}
H^{2n,2n}(Y,\mb R)&\to \mb R\\
\alpha &\mapsto \int_Y \alpha \wedge c_1(H_Y)^{\dim(Y)-\dim(X)}=\alpha \cdot H_Y^{\dim(Y)-\dim(X)}
\end{align*}
is linear and strictly positive (except on $0$) on the closed positive cone $\mathcal K_{2n}\subset H^{2n,2n}(Y,\mb R)$. Since $\alpha\mapsto \alpha\cdot [Z]$ is linear too, we can define
$$M=\max_{\alpha\in \mathcal K_{2n}\setminus \{0\}}\frac {\alpha\cdot [Z]}{\alpha \cdot H_Y^{\dim(Y)-\dim(X)} }\geq 0.$$
Now
$$\lambda_p(f_Z|\eta_Z)=\lim_{n\to + \infty} \left( (\tilde f^n)^*H_Y^p\cdot \eta^* H_B^{\dim(B)}\cdot H_Y^{2n-p-\dim(B)}\cdot [Z] \right)^{\frac 1n}\leq$$
$$\lim_{n\to + \infty} \left( M(\tilde f^n)^*H_Y^p\cdot \eta^* H_B^{\dim(B)}\cdot H_Y^{\dim(Y)-p-\dim(B)} \right)^{\frac 1n}=$$
$$\lim_{n\to + \infty} \left( (\tilde f^n)^*H_Y^p\cdot \eta^* H_B^{\dim(B)}\cdot H_Y^{\dim(Y)-p-\dim(B)} \right)^{\frac 1n}=\lambda_p(\tilde f|\eta)=1,$$
and since all relative dynamical degrees are $\geq 1$ (Proposition \ref{log-concavity}) we have $\lambda_p(f_Z|\eta_Z)=1$.
Now all is left to prove is that $\lambda_p(\tilde f|\eta)=1$. There exists $k>0$ such that $\eta^*H_B^{\dim(B)}\equiv_{num}k[F]$, where $[F]$ is the numerical class of a fibre $F$ of $\eta$. We have
$$\lambda_p(\tilde f|\eta)=\lim_{n\to + \infty} \left( (\tilde f^n)^*H_Y^p\cdot \eta^* H_B^{\dim(B)}\cdot H_Y^{\dim(Y)-p-\dim(B)} \right)^{\frac 1n}=$$
$$\lim_{n\to + \infty} \left( (\tilde f^n)^*H_Y^p\cdot k[F]\cdot H_Y^{\dim(Y)-p-\dim(B)} \right)^{\frac 1n}=$$
$$\lim_{n\to + \infty} \left( \left((\tilde f^n)^*H_Y\right)|_F^p\cdot H_Y|_F^{\dim(Y)-p-\dim(B)} \right)^{\frac 1n}.$$
For each fibre we have a canonical identification $F\cong \mb P^N$, and by this identification $H_Y|_F\cong \mathcal O_{\mb P^N}(d)$, meaning that the hyperplane section $H_Y|_F$ is defined by an equation of degree $d$. Under the identification, the action of $\tilde f$ from one fibre to another is linear, so that $\left((\tilde f^n)^*H_Y\right)|_F$ is also defined by an equation of degree $d$ on $\mb P^N$. This means that
$$\lambda_p(\tilde f|\eta)=\lim_{n\to +\infty}(d^{\dim(F)})^{\frac 1n}=1$$
as we wanted to show. This concludes the proof.
\end{proof}
The following Proposition is a weaker version of point $(1)$ of the Main Theorem.
\begin{prop}
\label{inequality dim B}
Let $f\colon X\dashrightarrow X$ be a loxodromic transformation of an irreducible symplectic manifold $X$ of dimension $2n$, and let
$$1=\lambda_0(f)<\cdots < \lambda_{p_0}(f)=\cdots =\lambda_{p_0+k}(f)>\cdots>\lambda_{2n}(f)=1$$
be its dynamical degrees.\\
If $\pi\colon B\dashrightarrow B$ is an $f$-invariant meromorphic fibration, then $\dim(B)\geq 2n-k$. In particular, if $f$ is an automorphism (or, more generally, if all the consecutive dynamical degrees of $f$ are distinct), then it is primitive.
\end{prop}
\begin{proof}
Let $g\colon B\dashrightarrow B$ be a birational transformation such that $g\circ \pi=\pi\circ f$.\\
\centerline{\xymatrix{
X \ar@{-->}[d]^\pi \ar@{-->}[r]^f & X \ar@{-->}[d]^\pi\\
B \ar@{-->}[r]^g & B
}}
We know by Lemma \ref{key lemma} that the generic fibre of $\pi$ is of general type; by Lemma \ref{trivial relative dyn} this implies that all the relative dynamical degrees $\lambda_p(f|\pi)$ are equal to $1$. By Theorem \ref{dinh nguyen} we then have
$$\lambda_p(f)=\max_{p-\dim(F)\leq q\leq p}\lambda_q(g),$$
where $\dim(F)=\dim(X)-\dim(B)$ is the dimension of a generic fibre.\\
Let $q\in \{0, 1, \ldots, \dim(B)\}$ be such that $\lambda_q(g)$ is maximal. Then
$$\lambda_q(f)=\lambda_{q+1}(f)=\cdots= \lambda_{q+\dim(F)}(f)=\lambda_q(g),$$
meaning that $k\geq \dim(F)=2n-\dim(B)$. This concludes the proof.
\end{proof}
\begin{rem}
Since in the Theorem we have $k\leq 2n-1$, the base of an invariant fibration cannot be a curve. Therefore Proposition \ref{inequality dim B} implies point $(2)$ of the Main Theorem.
\end{rem}
|
\section{Introduction}
In this paper we define new families of numbers and polynomials. By using
these new numbers and polynomials, we give computation algorithm for the
Euler type numbers of negative orders and combinatorial sums involving
binomial coefficients. We give combinatorial interpretations of these
numbers and polynomials.
Let $a$ and $b$ be real numbers and let $\lambda $ be real or complex
numbers.
For $n$ and $k$ nonegative integers, define $y_{3}(n,k;\lambda ;a,b)$ and
W_{n}(\lambda )$ by means of the following generating functions,
respectively
\begin{equation}
F_{y_{3}}(t,k;\lambda ;a,b)=\frac{e^{bkt}}{k!}\left( \lambda
e^{(a-b)t}+1\right) ^{k}=\sum_{n=0}^{\infty }y_{3}(n,k;\lambda ;a,b)\frac
t^{n}}{n!}. \label{St2a1}
\end{equation}
an
\begin{equation}
F_{w}(t;\lambda )=\frac{1}{\lambda e^{t}+\lambda ^{-1}e^{-t}+2
=\sum_{n=0}^{\infty }W_{n}(\lambda )\frac{t^{n}}{n!}. \label{w1}
\end{equation}
Our purpose of this paper is to derive and investigate some properties of
the numbers $y_{3}(n,k;\lambda ;a,b)$, the numbers $W_{n}(\lambda )$\ and
their generating functions. We can show how the numbers $y_{3}(n,k;\lambda
;a,b)$ and $W_{n}(\lambda )$ are related to certain sequences of numbers and
polynomials such as the Stirling numbers, the central factorial numbers, the
array polynomials, the Bernstein basis functions, the first and the second
kind Euler numbers and others. We compute some values of these numbers and
polynomials, which are given by the tables. By using these numbers and their
numerical values, we also compute the first and the second kind Euler
numbers of negative order.
A summary by sections follows:
In Section 2 is a background section containing basic definitions,
identities, relations and terminology we needed. In this section, we
breiefly give certain sequences of numbers and polynomials such as the Euler
numbers, the Stirling numbers, the central factorial numbers and the array
polynomials, the Bernstein basis functions. In Section 3, we some properties
of the numbers $y_{3}(n,k;\lambda ;a,b)$ and their generating function. By
using these functions and their functional equations, we derive various
identities and relations. We also give relations between these numbers and
the well-known special numbers and polynomials. By using partial derivative
equation of the generating function, we give recurrence relations and
derivative formula of these numbers. We compute a few values of the numbers
y_{3}(n,k;\lambda ;a,b)$. These values are given by the tables. In Section
4, by using the numbers $y_{3}(n,k;\lambda ;a,b)$, we compute the first and
the second kind Euler numbers of negative order. We give a few values of
these numbers, which are given by the tables. We also give a computational
algorithm in order to compute the second kind Euler polynomials and numbers
of negative order. In Section 5, we give some properties of the numbers
W_{n}(\lambda )$. We define a new sequence of polynomials related to the
numbers $W_{n}(\lambda )$. These numbers and polynomials are related to the
first and the second kind Euler type numbers. We compute some numerical
values of these numbers. In Section 6, we give further remarks and
observations. We also give not only some combinatorial applications,
including the rook numbers and polynomials, but also combinatorial
interpretation for special values $\lambda $, $a$ and $b$ for the numbers
y_{3}(n,k;\lambda ;a,b)$.
\section{Background}
Throughout this paper, we use the following standard notations:
\begin{eqnarray*}
\mathbb{N} &=&\{1,2,3,\ldots \}, \\
\mathbb{N}_{0} &=&\{0,1,2,3,\ldots \}=\mathbb{N}\cup \{0\},
\end{eqnarray*
$\mathbb{Z}$ denotes the set of integers, $\mathbb{R}$ denotes the set of
real numbers and $\mathbb{C}$ denotes the set of complex numbers.
The principal value $\ln z$ is the logarithm whose imaginary part lies in
the interval $(-\pi ,\pi ]$. Furthermor
\begin{equation*}
0^{n}=\left\{
\begin{array}{cc}
1, & (n=0) \\
0, & (n\in \mathbb{N)
\end{array
\right.
\end{equation*
and also if $\lambda $ is a complex number, we use the following notation
\begin{equation*}
\left(
\begin{array}{c}
\lambda \\
\end{array
\right) =1\text{ and }\left(
\begin{array}{c}
\lambda \\
\end{array
\right) =\frac{\lambda (\lambda -1)\cdots (\lambda -v+1)}{v!}=\frac{\left(
\lambda \right) _{v}}{v!}\text{ }(n\in \mathbb{N}\text{, }\lambda \in
\mathbb{C)}
\end{equation*
(\textit{cf}. \cite{Bayad}, \cite{Comtet}, \cite{SimsekNEW}, \cit
{SrivastavaChoi2012}). For combinatorial example, we also use the notations
of Bona \cite{Bona}. Let $\{1,2,\ldots ,n\}$ be a distinctly $n$-element
set. In work of Bona \cite[PP. 11-13.]{Bona}, the shorter notation $[n]$
stands for $(n)_{k}$: the number $n(n-1)(n-2)\cdots (n-k+1)$ of all $k
-element lists from $[n]$ without repetition occurs in combinatorics.
The first kind Apostol-Euler polynomials of order $k$ are defined by means
of the following generating functions
\begin{equation}
F_{P1}(t,x;k,\lambda )=\left( \frac{2}{\lambda e^{t}+1}\right)
^{k}e^{tx}=\sum_{n=0}^{\infty }E_{n}^{(k)}(x;\lambda )\frac{t^{n}}{n!}.
\label{Cad3}
\end{equation}
We observe tha
\begin{equation*}
E_{n}^{(k)}(\lambda )=E_{n}^{(k)}(0;\lambda )
\end{equation*
which denotes the first kind Apostol-Euler numbers of order $k$ (\textit{cf
. \cite{Grademir}-\cite{Qi}; see also the references cited in each of these
earlier works).
Substituting $k=\lambda =1$ into (\ref{Cad3}), we have the first kind Euler
numbers $E_{n}=E_{n}^{(1)}(1)$, which are defined by means of the following
generating function
\begin{equation*}
\frac{2}{e^{t}+1}=\sum_{n=0}^{\infty }E_{n}\frac{t^{n}}{n!},
\end{equation*
where $\left\vert t\right\vert <\pi $ (\textit{cf}. \cite{Boyadzhiev}-\cit
{Qi}; see also the references cited in each of these earlier works).
The second kind Euler numbers $E_{n}^{\ast }$ are defined by means of the
following generating function
\begin{equation*}
\frac{2}{e^{t}+e^{-t}}=\sum_{n=0}^{\infty }E_{n}^{\ast }\frac{t^{n}}{n!},
\end{equation*
where $\left\vert t\right\vert <\frac{\pi }{2}$ (\textit{cf}. \cite{Grademir
-\cite{Qi}; see also the references cited in each of these earlier works).
\begin{remark}
By using generating functions of the numbers $E_{n}$ and $E_{n}^{\ast }$, we
can easily give a relationship between both of these numbers as follows
\begin{equation*}
E_{n}^{\ast }=2^{n}E_{n}\left( \frac{1}{2}\right)
\end{equation*
(\textit{cf}. \cite{DSKim.JIA}, \cite{MS.KIMjnt}, \cite{KimJNT}, \cit
{RJMP2010}, \cite{jnt2003}; see also the references cited in each of these
earlier works).
\end{remark}
The second kind $\lambda $-Stirling numbers $S(n,v;\lambda )$, generalized
of the classical Stirling number of the second kind, are defined by means of
the following generating function
\begin{equation}
F_{S}(t,v;\lambda )=\frac{\left( \lambda e^{t}-1\right) ^{v}}{v!
=\sum_{n=0}^{\infty }S(n,v;\lambda )\frac{t^{n}}{n!}, \label{SN-1}
\end{equation
For the numbers $S(n,v;\lambda )$, the reader should consult \cite{Luo} and
\cite{SimsekFPTA} see also (\textit{cf}. \cite{SimsekMANISA}, \cit
{Srivastava2011}).
From (\ref{SN-1}), we easily see tha
\begin{equation*}
S(n,v)=S(n,v;1),
\end{equation*
which denotes the Stirling numbers of the second kind. These numbers are
computing by the following formula
\begin{equation*}
S(n,v)=\frac{1}{v!}\sum_{j=0}^{v}\left(
\begin{array}{c}
v \\
\end{array
\right) (-1)^{j}\left( v-j\right) ^{n}
\end{equation*
(\textit{cf}. \cite{Grademir}-\cite{Qi}; see also the references cited in
each of these earlier works). A recurrence relation for these numbers is
given b
\begin{equation*}
S(n,k)=S(n-1,k-1)+kS(n-1,k),
\end{equation*
wit
\begin{equation*}
S(n,0)=0\text{ (}n\in \mathbb{N}\text{); }S(n,n)=1\text{ (}n\in \mathbb{N
\text{); }S(n,1)=1\text{ (}n\in \mathbb{N}\text{)}
\end{equation*
and $S(n,k)=0$ ($n<k$ or $k<0$) (\textit{cf}. \cite{Grademir}-\cite{Qi}; see
also the references cited in each of these earlier works).
Let $\lambda \in \mathbb{C}$ and $k\in \mathbb{N}_{0}$. The $\lambda $-array
polynomials $S_{v}^{n}(x)$ are defined by means of the following generating
function
\begin{equation}
F_{A}(t,x,k;\lambda )=\frac{1}{k!}e^{tx}\left( \lambda e^{t}-1\right)
^{k}=\sum_{n=0}^{\infty }S_{k}^{n}(x;\lambda )\frac{t^{n}}{n!} \label{Ary}
\end{equation
(\textit{cf}. \cite{SimsekFPTA}, \cite{Bayad}). Substituting $\lambda =1$
into the above equation, we hav
\begin{equation*}
S_{k}^{n}(x)=\frac{1}{k!}\sum_{j=0}^{k}(-1)^{k-j}\left(
\begin{array}{c}
k \\
\end{array
\right) \left( x+j\right) ^{n}
\end{equation*
wit
\begin{equation*}
S_{0}^{0}(x)=S_{n}^{n}(x)=1,S_{0}^{n}(x)=x^{n}.
\end{equation*
If $k>n$, tha
\begin{equation*}
S_{k}^{n}(x)=0
\end{equation*
(\textit{cf}. \cite{Bayad}, \cite{Chan}, \cite{SimsekFPTA}, \cite{AM2014};
see also the references cited in each of these earlier works).
The second kind central factorial numbers $T(n,k)$ are defined by means of
the following generating function
\begin{equation}
F_{T}(t,k)=\frac{1}{(2k)!}\left( e^{t}+e^{-t}-2\right)
^{k}=\sum_{n=0}^{\infty }T(n,k)\frac{t^{2n}}{(2n)!} \label{CT-1}
\end{equation
(\textit{cf}. \cite{Bona}, \cite{Cigler}, \cite{Comtet}, \cite{SrivastavaLiu
, \cite{AM2014}; see also the references cited in each of these earlier
works).
\begin{remark}
In the work of Alayont et al. \cite{Aloyat-1}, we observe that the central
factorial numbers are related to the rook polynomials, which count the
number of ways of placing non-attacking rooks on a chess board. In the work
of Alayont and Krzywonos \cite{Alayont}, we've noticed the following elegant
result which are related to the he central factorial numbers and the rook
numbers:The number of ways to place $k$ rooks on a size $m$ triangle board
in three dimensions is equal to
\begin{equation*}
T(m+1,m+1-k),
\end{equation*}
where $0\leq k\leq m$.
\end{remark}
In \cite{SimsekNEW}, we defined the numbers $y_{1}(n,k;\lambda )$ by means
of the following generating functions
\begin{equation}
F_{y_{1}}(t,k;\lambda )=\frac{1}{k!}\left( \lambda e^{t}+1\right)
^{k}=\sum_{n=0}^{\infty }y_{1}(n,k;\lambda )\frac{t^{n}}{n!}. \label{ay1}
\end{equation}
\begin{theorem}
Let $n$ be a positive integer. Then we have
\begin{equation}
y_{1}(n,k;\lambda )=\frac{1}{k!}\sum_{j=0}^{k}\left(
\begin{array}{c}
k \\
\end{array}
\right) j^{n}\lambda ^{j}. \label{ay2}
\end{equation}
\end{theorem}
By substituting $\lambda =1$ into (\ref{ay2}), then we se
\begin{equation}
B(n,k)=k!y_{1}(n,k;1). \label{CC2}
\end{equation
In \cite{golombek}, Golombek gave the following formula for (\ref{ay2})
\begin{equation*}
B(n,k)=\frac{d^{n}}{dt^{n}}\left( e^{t}+1\right) ^{k}\left\vert
_{t=0}\right. .
\end{equation*}
If we substitute $\lambda =-1$ into (\ref{ay2}), then we get the following
well-known numbers, which are so-called Stirling numbers of the second kind
\begin{equation*}
S_{2}(n,k)=(-1)^{k}B(n,k).
\end{equation*
We \cite{SimsekNEW} gave the following \textbf{conjecture}
\begin{equation*}
B(d;k)=(k^{d}+x_{1}k^{d-1}+x_{2}k^{d-2}+\cdots +x_{d-2}k^{2}+x_{d-1}k)2^{k-d}
\end{equation*
where $x_{1},x_{2},\ldots ,x_{d-1},d$ are positive integers.
In \cite{SimsekNEW}, we defined the numbers $y_{2}(n,k;\lambda )$ by
means of the following generating functions
\begin{equation}
F_{y_{2}}(t,k;\lambda )=\frac{1}{(2k)!}\left( \lambda e^{t}+\lambda
^{-1}e^{-t}+2\right) ^{k}=\sum_{n=0}^{\infty }y_{2}(n,k;\lambda )\frac{t^{n
}{n!}. \label{C1}
\end{equation}
Note that there is one generating function for each value of $k$.
\begin{theorem}
\begin{equation}
y_{2}(n,k;\lambda )=\frac{1}{\left( 2k\right) !}\sum_{j=0}^{k}\left(
\begin{array}{c}
k \\
\end{array}
\right) 2^{k-j}\sum_{l=0}^{j}\left(
\begin{array}{c}
j \\
\end{array}
\right) \left( 2l-j\right) ^{n}\lambda ^{2l-j} \label{CCC3}
\end{equation}
\end{theorem}
We will see in the following section that the above formulas provide
important insight when we are trying to compute the first and the second
kind Euler numbers of negative order.
\section{A family of new numbers $y_{3}(n,k;\protect\lambda ;a,b)$}
In this section, we give some properties of numbers $y_{3}(n,k;\lambda ;a,b)
. We compute a few values of these numbers. These numbers are related to the
combinatorial sums such as the stirling numbers, the array polynomials.
By using (\ref{St2a1}), we give the following explicit formula for the
numbers $y_{3}(n,k;\lambda ;a,b)$:
\begin{theorem}
\begin{equation}
y_{3}(n,k;\lambda ;a,b)=\frac{1}{k!}\sum_{j=0}^{k}\left(
\begin{array}{c}
k \\
\end{array}
\right) \lambda ^{j}(bk+j(a-b))^{n}, \label{St2a}
\end{equation}
\end{theorem}
\subsubsection{Setting $a=b$ in (\protect\ref{St2a}), we hav
\protect\begin{equation*}
y_{3}(n,k;\protect\lambda ;a,a)=\frac{\left( ak\right) ^{n}}{k!}\left( 1
\protect\lambda \right) ^{k}.
\protect\end{equation*
For $k=0,1,2,3,4,5$ compute a few values of the numbers $y_{3}(n,k;\protec
\lambda ;a,b)$ given by Equation (\protect\ref{St2a}) as follows
\protect\begin{equation*}
y_{3}(n,0;\protect\lambda ;a,b)=0,
\protect\end{equation*
\protect\begin{equation*}
y_{3}(n,1;\protect\lambda ;a,b)=a^{n}\protect\lambda +b^{n},
\protect\end{equation*
\protect\begin{equation*}
y_{3}(n,2;\protect\lambda ;a,b)=\frac{\left( 2b+2(a-b)\right) ^{n}}{2
\protect\lambda ^{2}+(b+a)^{n}\protect\lambda +2^{n-1}b^{n},
\protect\end{equation*
\protect\begin{equation*}
y_{3}(n,3;\protect\lambda ;a,b)=\frac{(3b+3(a-b))^{n}}{6}\protect\lambda
^{3}+\frac{(3b+2(a-b))^{n}}{2}\protect\lambda ^{2}+\frac{(2b+a)^{n}}{2
\protect\lambda +\frac{3^{n-1}b^{n}}{2},
\protect\end{equation*
\protect\begin{eqnarray*}
y_{3}(n,4;\protect\lambda ;a,b) &=&\frac{(4b+4(a-b))^{n}}{24}\protect\lambda
^{4}+\frac{(4b+3(a-b))^{n}}{6}\protect\lambda ^{3}+\frac{(4b+2(a-b))^{n}}{4
\protect\lambda ^{2} \\
&&+\frac{(3b+a)^{n}}{6}\protect\lambda +\frac{4^{n-1}b^{n}}{6},
\protect\end{eqnarray*
\protect\begin{eqnarray*}
y_{3}(n,5;\protect\lambda ;a,b) &=&\frac{(5b+5(a-b))^{n}}{120}\protec
\lambda ^{5}+\frac{(5b+4(a-b))^{n}}{24}\protect\lambda ^{4}+\frac
(5b+3(a-b))^{n}}{12}\protect\lambda ^{3} \\
&&+\frac{(5b+2(a-b))^{n}}{12}\protect\lambda ^{2}+\frac{(4b+a)^{n}}{24
\protect\lambda +\frac{5^{n-1}b^{n}}{24}.
\protect\end{eqnarray*
Identities and Relations}
Here, we give some identities and relations for special values of the
numbers $y_{3}(n,k;\lambda ;a,b)$, which are given below
\begin{equation*}
y_{3}(n,k;\lambda ;1,1)=\frac{k^{n}}{k!}\left( 1+\lambda \right) ^{k},
\end{equation*
\begin{equation*}
y_{3}(0,k;\lambda ;1,1)=\frac{1}{k!}\left( 1+\lambda \right) ^{k},
\end{equation*
\begin{equation*}
y_{3}(k,k;\lambda ;1,1)=\frac{k^{k}}{k!}\left( 1+\lambda \right) ^{k},
\end{equation*
\begin{equation*}
y_{3}(k,k;1;1,1)=\frac{2^{k}k^{k}}{k!}.
\end{equation*
If substitute $a=1$ and $b=0$ into (\ref{St2a}), than we ge
\begin{equation*}
y_{2}(n,k;\lambda )=y_{3}(n,k;\lambda ;1,0).
\end{equation*
If substitute $a=1$ and $b=-1$ into (\ref{St2a}), than we obtai
\begin{equation*}
y_{3}(n,k;\lambda ;1,-1)=\sum_{m=0}^{n}\left(
\begin{array}{c}
n \\
\end{array
\right) (-1)^{n-m}k^{n-m}2^{m}y_{1}(m,k;\lambda ).
\end{equation*
If substitute $\lambda =-1$, $a=-1$ and $b=0$\ into (\ref{St2a}), than we
hav
\begin{equation*}
S_{2}(n,k)=(-1)^{k+n}y_{3}(n,k;-1;-1,0).
\end{equation*
Now by using functional equation for the generating functions, we derive
some identities related to the numbers $y_{3}(n,k;\lambda ;a,b)$,
y_{1}(n,k;\lambda )$, $S_{2}(n,k)$, the Bernstein basis functions and the
array polynomials.
\begin{theorem}
\begin{equation*}
y_{3}(n,k;\lambda ;a,b)=\sum_{j=0}^{k}\sum_{m=0}^{n}\left(
\begin{array}{c}
n \\
\end{array}
\right) a^{m}b^{n-m}y_{1}(m,j;\lambda )S_{2}(n-m,k-j).
\end{equation*}
\end{theorem}
\begin{proof}
By combining (\ref{SN-1}) and (\ref{ay1}) with (\ref{St2a1}), we get the
following functional equation
\begin{equation*}
F_{y_{3}}(t,k;\lambda ;a,b)=\sum_{j=0}^{k}F_{y_{1}}(at,j;\lambda
)F_{S}(bt,k-j;1).
\end{equation*
From this equation, we ge
\begin{equation*}
\sum_{n=0}^{\infty }y_{3}(n,k;\lambda ;a,b)\frac{t^{n}}{n!
=\sum_{j=0}^{k}\sum_{n=0}^{\infty }a^{n}y_{1}(n,j;\lambda )\frac{t^{n}}{n!
\sum_{n=0}^{\infty }b^{n}S_{2}(n,k-j)\frac{t^{n}}{n!}.
\end{equation*
Therefor
\begin{equation*}
\sum_{n=0}^{\infty }y_{3}(n,k;\lambda ;a,b)\frac{t^{n}}{n!
=\sum_{n=0}^{\infty }\sum_{j=0}^{k}\sum_{m=0}^{n}\left(
\begin{array}{c}
n \\
\end{array
\right) a^{m}b^{n-m}y_{1}(m,j;\lambda )S_{2}(n-m,k-j)\frac{t^{n}}{n!}.
\end{equation*
Comparing the coefficients of $\frac{t^{n}}{n!}$ on both sides of the above
equation, we arrive at the desired result.
\end{proof}
A relation between the numbers $y_{1}(n,k;\lambda )$ and $y_{3}(n,k;\lambda
;a,b)$ is given by the following theorem:
\begin{theorem}
\begin{equation*}
y_{3}(n,k;\lambda ;a,b)=\sum_{j=0}^{n}\left(
\begin{array}{c}
n \\
\end{array}
\right) (a-b)^{j}(bk)^{n-j}y_{1}(j,k;\lambda ).
\end{equation*}
\end{theorem}
\begin{proof}
By using (\ref{ay1}) and (\ref{St2a1}), we obtain the following functional
equation
\begin{equation}
F_{y_{3}}(t,k;\lambda ;a,b)=e^{bkt}F_{y_{1}}((a-b)t,k;\lambda ). \label{BER}
\end{equation
From this equation, we ge
\begin{equation*}
\sum_{n=0}^{\infty }y_{3}(n,k;\lambda ;a,b)\frac{t^{n}}{n!
=\sum_{n=0}^{\infty }(bk)^{n}\frac{t^{n}}{n!}\sum_{n=0}^{\infty
}(a-b)^{n}y_{1}(n,k;\lambda )\frac{t^{n}}{n!}.
\end{equation*
Therefor
\begin{equation*}
\sum_{n=0}^{\infty }y_{3}(n,k;\lambda ;a,b)\frac{t^{n}}{n!
=\sum_{n=0}^{\infty }\sum_{j=0}^{n}\left(
\begin{array}{c}
n \\
\end{array
\right) (a-b)^{j}(bk)^{n-j}y_{1}(j,k;\lambda )\frac{t^{n}}{n!}.
\end{equation*
Comparing the coefficients of $\frac{t^{n}}{n!}$ on both sides of the above
equation, we arrive at the desired result.
\end{proof}
Combining (\ref{Ary}) and (\ref{St2a1}), we obtain the following functional
equation
\begin{equation*}
F_{y_{3}}(t,k;-1;a,b)=(-1)^{k}F_{A}\left( (a-b)t,\frac{b}{a-b},k\right) .
\end{equation*}
By using this equation, a relation between the array polynomials and the
numbers $y_{3}(n,k;-1;a,b)$ by the following theorem:
\begin{theorem}
\begin{equation*}
y_{3}(n,k;-1;a,b)=(-1)^{k}(a-b)^{n}S_{v}^{n}\left( \frac{b}{a-b}\right) .
\end{equation*}
\end{theorem}
By replacing $\lambda $ by $-\lambda ^{2}$\ in (\ref{Ary}), using (\ref{ay1
) and (\ref{St2a1}), we also have the following functional equation
\begin{equation*}
F_{y_{3}}(2t,k;-\lambda ^{2};a,b)=(-1)^{k}k!F_{y_{1}}((a-b)t,k;\lambda
)F_{A}\left( (a-b)t,\frac{2b}{a-b},k;\lambda \right) .
\end{equation*
By using this equation with (\ref{Ary}), (\ref{ay1}) and (\ref{St2a1}), we
get the following theorem:
\begin{theorem}
\begin{equation*}
y_{3}(n,k;-\lambda ^{2};a,b)=(-1)^{k}k!\left( \frac{a-b}{2}\right)
^{n}\sum_{j=0}^{n}\left(
\begin{array}{c}
n \\
\end{array}
\right) y_{1}(n,j;\lambda )S_{k}^{n-j}\left( \frac{b}{a-b}\right) .
\end{equation*}
\end{theorem}
By using (\ref{CT-1}) and (\ref{St2a1}), we give the following functional
equation
\begin{equation*}
F_{y_{3}}(t,k;1;1,-1)=\frac{1}{k!}\sum_{j=0}^{k}\left(
\begin{array}{c}
k \\
\end{array
\right) (-2)^{k-j}\left( 2j\right) !F_{T}(t,j).
\end{equation*
By this equation, we get a relation between the central factorial numbers
and the numbers $y_{3}(n,k;1;1,-1)$ by the following theorem:
\begin{theorem}
If $n$ is an even integer, we have
\begin{equation*}
y_{3}(n,k;1;1,-1)=\frac{1}{k!}\sum_{j=0}^{k}\left(
\begin{array}{c}
k \\
\end{array}
\right) (-2)^{k-j}\left( 2j\right) !T(n,j).
\end{equation*}
If $n$ is an odd integer, we have
\begin{equation*}
y_{3}(n,k;1;1,-1)=0.
\end{equation*}
\end{theorem}
Substituting $b=x$ and $a=1$ into (\ref{St2a}), we give relationships
between the numbers $y_{3}(n,k;\lambda ;a,b)$, the numbers
y_{1}(n-m,k;\lambda )$\ and the Bernstein basis functions $B_{k}^{n}(x)$ by
the following corollary:
\begin{corollary}
We have
\begin{equation*}
y_{3}(n,k;\lambda ;1,x)=\frac{1}{k!}\sum_{j=0}^{k}\left(
\begin{array}{c}
k \\
\end{array
\right) \lambda ^{j}\sum_{m=0}^{n}k^{m}j^{n-m}B_{m}^{n}\left( x\right)
\end{equation*
and
\begin{equation}
y_{3}(n,k;\lambda ;1,x)=\sum_{m=0}^{n}k^{m}B_{m}^{n}\left( x\right)
y_{1}(n-m,k;\lambda ), \label{BE-1}
\end{equation
where $B_{m}^{n}\left( x\right) $\ denotes the Bernstein basis functions
\begin{equation*}
B_{m}^{n}\left( x\right) =\left(
\begin{array}{c}
n \\
\end{array
\right) x^{k}\left( 1-x\right) ^{n-k}
\end{equation*
(\textit{cf}. \cite{Lorenz}, \cite{mmas2015}; see also the references cited
in each of these earlier works).
\end{corollary}
Integrating both sides of Equation (\ref{BE-1}) from $0$ to $1$, and using
\begin{equation*}
\int_{0}^{1}B_{m}^{n}\left( x\right) dx=\frac{1}{n+1}
\end{equation*
(\textit{cf}. \cite{Lorenz}, \cite{mmas2015}; see also the references cited
in each of these earlier works), we get the following theorem:
\begin{theorem}
\begin{equation*}
\int_{0}^{1}y_{3}(n,k;\lambda ;1,x)dx=\frac{1}{n+1}
\sum_{m=0}^{n}k^{m}y_{1}(n-m,k;\lambda )
\end{equation*}
or
\begin{equation*}
\int_{0}^{1}y_{3}(n,k;\lambda ;1,x)dx=\frac{1}{\left( n+1\right) k!}
\sum_{j=0}^{k}\left(
\begin{array}{c}
k \\
\end{array}
\right) j^{n}\lambda ^{j}\sum_{m=0}^{n}k^{m}j^{-m}.
\end{equation*}
\end{theorem}
\subsection{Recurrence relation for the numbers $y_{3}(n,k;\protect\lambda
;a,b)$}
Here using derivative operator to generating function for the numbers
y_{3}(n,k;\lambda ;a,b)$, we give a recurrence relation for these numbers.
Taking derivative of (\ref{St2a1}), with respect to $t$, we obtain the
following partial differential equation:
\begin{equation}
\frac{\partial }{\partial t}F_{y_{3}}(t,k;\lambda
;a,b)=bkF_{y_{3}}(t,k;\lambda ;a,b)+(a-b)\lambda F_{y_{3}}(t,k-1;\lambda
;a,b). \label{St2a2}
\end{equation
Combining this equation with (\ref{St2a1}), we derive a recurrence relation
for the numbers $y_{3}(n,k;\lambda ;a,b)$ by the following theorem:
\begin{theorem}
Let $k$ be a positive integer. Then we have
\begin{equation*}
y_{3}(n+1,k;\lambda ;a,b)=bky_{3}(n,k;\lambda ;a,b)+(a-b)y_{3}(n,k-1;\lambda
;a,b).
\end{equation*}
\end{theorem}
\begin{proof}
By using (\ref{St2a2}) and (\ref{St2a1}), we ge
\begin{eqnarray*}
&&\sum_{n=1}^{\infty }y_{3}(n,k;\lambda ;a,b)\frac{t^{n-1}}{\left(
n-1\right) !} \\
&=&bk\sum_{n=0}^{\infty }y_{3}(n,k;\lambda ;a,b)\frac{t^{n}}{n!
+(a-b)\sum_{n=0}^{\infty }y_{3}(n,k-1;\lambda ;a,b)\frac{t^{n}}{n!}.
\end{eqnarray*
Therefor
\begin{eqnarray*}
&&\sum_{n=0}^{\infty }y_{3}(n+1,k;\lambda ;a,b)\frac{t^{n}}{n!} \\
&=&bk\sum_{n=0}^{\infty }y_{3}(n,k;\lambda ;a,b)\frac{t^{n}}{n!
+(a-b)\sum_{n=0}^{\infty }y_{3}(n,k-1;\lambda ;a,b)\frac{t^{n}}{n!}.
\end{eqnarray*
Comparing the coefficients of $\frac{t^{n}}{n!}$ on both sides of the above
equation, we arrive at the desired result.
\end{proof}
Taking derivative of (\ref{St2a1}), with respect to $\lambda $, we obtain
the following partial differential equation
\begin{equation*}
\frac{\partial }{\partial \lambda }F_{y_{3}}(t,k;\lambda ;a,b)=e^{\left(
a-b\right) t}F_{y_{3}}(t,k-1;\lambda ;a,b).
\end{equation*
By using the same processes in the above theorem, we obtain the following
theorem
\begin{equation*}
\frac{\partial }{\partial \lambda }y_{3}(n,k;\lambda
;a,b)=\sum_{m=0}^{n}\left(
\begin{array}{c}
n \\
\end{array
\right) \left( a-b\right) ^{n-m}y_{3}(m,k-1;\lambda ;a,b).
\end{equation*}
\section{Computation of the Euler numbers of negative order}
In this section, by using the numbers $y_{3}(n,k;\lambda ;a,b)$, we compute
values of the second kind Apostol type Euler polynomials of negative order.
We also give a computation algorithm for computing the values of these
polynomials.
We \cite{SimsekNEW} defined the second kind Apostol type Euler polynomials
of order $k$, $E_{n}^{\ast (k)}(x;\lambda )$ by means of the following
generating functions
\begin{equation}
F_{P}(t,x;k,\lambda )=\left( \frac{2}{\lambda e^{t}+\lambda ^{-1}e^{-t}
\right) ^{k}e^{tx}=\sum_{n=0}^{\infty }E_{n}^{\ast (k)}(x;\lambda )\frac
t^{n}}{n!}. \label{Eul.2}
\end{equation
We observe tha
\begin{equation*}
E_{n}^{\ast (k)}(\lambda )=E_{n}^{\ast (k)}(0;\lambda )
\end{equation*
denotes the second kind Apostol type Euler numbers of order $k$.
We can give a motivation on (\ref{Eul.2}) as follows
\begin{equation*}
F_{P}(t,x;k,\lambda )=F_{H}\left( t,\frac{x+k}{2};k,-\lambda ^{-2}\right) ,
\end{equation*
wher
\begin{equation*}
F_{H}\left( t,x;k,u\right) =\frac{1-u}{e^{t}-u}e^{tx}=\sum_{n=0}^{\infty
}H_{n}^{(k)}(x;u)\frac{t^{n}}{n!}
\end{equation*
$u\neq 1$ and $H_{n}^{(k)}(x;u)$ denotes the Frobenius-Euler polynomials of
higher order. From the above functional equation, we ge
\begin{equation*}
E_{n}^{\ast (k)}(x;\lambda )=\frac{2^{n}}{\lambda ^{k-2}(\lambda ^{2}+1)
H_{n}^{(k)}\left( \frac{x+k}{2};-\lambda ^{-2}\right) .
\end{equation*
These numbers are also related to the twisted Euler numbers and polynomials
\textit{cf}. \cite{DSKim.JIA}, \cite{KimJNT}, \cite{jnt2003}).
The first kind Apostol-Euler numbers of order $-k$ are defined by means of
the following generating functions
\begin{equation}
\left( \frac{\lambda e^{t}+1}{2}\right) ^{k}=\sum_{n=0}^{\infty
}E_{n}^{(-k)}(\lambda )\frac{t^{n}}{n!} \label{ae-1}
\end{equation
(\textit{cf}. \cite{SimsekNEW}, \cite{SrivastavaChoi2012}; see also the
references cited in each of these earlier works). The second kind Apostol
type Euler numbers of order $-k$ are defined by means of the following
generating functions
\begin{equation}
F_{N}(t;-k,\lambda )=\left( \frac{\lambda e^{t}+\lambda ^{-1}e^{-t}}{2
\right) ^{k}=\sum_{n=0}^{\infty }E_{n}^{\ast (-k)}(\lambda )\frac{t^{n}}{n!}
\label{Cac3}
\end{equation
(\textit{cf}. \cite{SimsekNEW}; see also the references cited in each of
these earlier works)
\begin{theorem}
(\cite{SimsekNEW}) Let $k$ be nonnegative integer. Then we have
\begin{equation}
E_{n}^{(-k)}(\lambda )=k!2^{-k}y_{1}(n,k;\lambda ). \label{Cab3}
\end{equation}
\end{theorem}
\begin{remark}
Byrd \cite{Byrd} and Liu \cite{Liu1} also gave a formula for the numbers
E_{n}^{(-k)}$. In (\cite{SimsekNEW}), we computed a few values of the first
kind Euler numbers of order $-k$ by the following formula
\begin{equation}
E_{n}^{(-k)}=2^{-k}\sum_{j=0}^{k}\left(
\begin{array}{c}
k \\
\end{array
\right) j^{n} \label{Caa3}
\end{equation
as follows: for $n=1,2,\ldots ,9$ and $k=0,-1,-2,\ldots ,-9$, we compute a
few values of the numbers $E_{n}^{(-k)}$, given by the above relations, as
follows:
\begin{equation*}
\begin{tabular}{lllllllllll}
$n\backslash k$ & $0$ & $-1$ & $-2$ & $-3$ & $-4$ & $-5$ & $-6$ & $-7$ & $-8$
& $-9\cdots $ \\
$0$ & $0$ & $\frac{1}{2}$ & $\frac{3}{4}$ & $\frac{7}{8}$ & $\frac{15}{16}$
& $\frac{33}{32}$ & $\frac{33}{64}$ & $\frac{81}{64}$ & $\cdots $ & $\cdots $
\\
$1$ & $0$ & $\frac{1}{2}$ & $1$ & $\frac{3}{2}$ & $2$ & $\frac{5}{2}$ & $3$
& $\frac{7}{2}$ & $4$ & $\frac{9}{2}\cdots $ \\
$2$ & $0$ & $\frac{1}{2}$ & $\frac{3}{2}$ & $3$ & $5$ & $\frac{15}{2}$ &
\frac{21}{2}$ & $14$ & $18$ & $\frac{45}{2}\cdots $ \\
$3$ & $0$ & $\frac{1}{2}$ & $\frac{5}{2}$ & $\frac{27}{4}$ & $14$ & $25$ &
\frac{81}{2}$ & $\frac{245}{4}$ & $88$ & $\frac{243}{2}\cdots $ \\
$4$ & $0$ & $\frac{1}{2}$ & $\frac{9}{2}$ & $\frac{33}{2}$ & $\frac{85}{2}$
& $90$ & $168$ & $287$ & $459$ & $\frac{1395}{2}\cdots $ \\
$5$ & $0$ & $\frac{1}{2}$ & $\frac{17}{2}$ & $\frac{171}{4}$ & $137$ &
\frac{1375}{4}$ & $738$ & $1421$ & $2524$ & $4212\cdots $ \\
$6$ & $0$ & $\frac{1}{2}$ & $\frac{33}{2}$ & $\frac{231}{2}$ & $\frac{925}{2}
$ & $\frac{5505}{4}$ & $\frac{13587}{4}$ & $7364$ & $14508$ & $26550\cdots $
\\
$7$ & $0$ & $\frac{1}{2}$ & $\frac{65}{2}$ & $\frac{1287}{4}$ & $1619$ &
5725$ & $\frac{65007}{4}$ & $\frac{317275}{8}$ & $86608$ & $173664\cdots $
\\
$8$ & $0$ & $\frac{1}{2}$ & $\frac{129}{2}$ & $\frac{1833}{2}$ & $\frac{1166
}{2}$ & $\frac{49155}{2}$ & $\frac{160671}{2}$ & $\frac{441469}{2}$ & $\frac
1068453}{2}$ & $1173240\cdots $ \\
$\underset{\vdots }{9}$ & $0$ & $\frac{1}{2}$ & $\frac{513}{2}$ & $\frac
15531}{2}$ & $\frac{161365}{2}$ & $\frac{1951155}{4}$ & $\frac{8499057}{4}$
& $7418789$ & $22071123$ & $\frac{232549335}{4}\cdots
\end{tabular
\end{equation*}
\end{remark}
By using the numbers $y_{3}\left( n,k;\lambda ;a,b\right) $, we can compute
the second kind Apostol type Euler polynomials and numbers of order $-k$.
Substituting $t=2z$, $a=\frac{x+k}{2k}$, $b=\frac{x-k}{2k}$ into (\ref{St2a1
), and replacing $k$ by $-k$ in (\ref{Eul.2}), we ge
\begin{equation*}
F_{P}(z,x;-k,\lambda )=\frac{k!}{2^{k}\lambda ^{k}}F_{y_{3}}\left(
2z,k;\lambda ^{2};\frac{x+k}{2k},\frac{x-k}{2k}\right) .
\end{equation*
From this equation we ge
\begin{equation*}
\sum_{n=0}^{\infty }E_{n}^{\ast (-k)}(x;\lambda )\frac{t^{n}}{n!}=\frac{k!}{
2^{k}\lambda ^{k}}\sum_{n=0}^{\infty }2^{n}y_{3}\left( n,k;\lambda ^{2}
\frac{x+k}{2k},\frac{x-k}{2k}\right) \frac{t^{n}}{n!}.
\end{equation*
Comparing the coefficients of $\frac{t^{n}}{n!}$ on both sides of the above
equation, we arrive at the following theorem.
\begin{theorem}
Let $n$ be nonnegative integers. Then we have
\begin{equation}
E_{n}^{\ast (-k)}(x;\lambda )=\frac{k!2^{n-k}}{\lambda ^{k}}y_{3}\left(
n,k;\lambda ^{2};\frac{x+k}{2k},\frac{x-k}{2k}\right) . \label{Eul.3c}
\end{equation}
\end{theorem}
The next assertion confirms and extend a formula in Equation (\ref{Caa3}).
Substituting $x=0$ into (\ref{Eul.3c}), we get the following corollary:
\begin{corollary}
Let $n$ be nonnegative integers. Then we have
\begin{equation}
E_{n}^{\ast (-k)}(\lambda )=\frac{k!2^{n-k}}{\lambda ^{k}}y_{3}\left(
n,k;\lambda ^{2};\frac{1}{2},-\frac{1}{2}\right) . \label{Eul.3d}
\end{equation}
\end{corollary}
\begin{remark}
Substituting $\lambda =1$ into (\ref{Eul.3d}), we hav
\begin{equation*}
E_{n}^{\ast (-k)}=k!2^{n-k}y_{3}\left( n,k;1;\frac{1}{2},-\frac{1}{2}\right)
\end{equation*
(\textit{cf}. \cite{Liu1}, \cite{SimsekNEW}, \cite{SrivastavaChoi2012}; see
also the references cited in each of these earlier works).
\end{remark}
We are finally ready to compute the some values of the second kind Apostol
type Euler polynomials of negative order. By using (\ref{St2a}) and (\re
{Eul.3c}), we compute a few values of the polynomials $E_{n}^{\ast
(-k)}(x;\lambda )$ as follows
\begin{eqnarray*}
E_{n}^{\ast (0)}(x;\lambda ) &=&{x}^{n}, \\
E_{n}^{\ast (-1)}(x;\lambda ) &=&\frac{{\left( x+1\right) }^{n}\lambda +
\frac{{\left( x-1\right) }^{n}}{\lambda }}{2}, \\
E_{n}^{\ast (-2)}(x;\lambda ) &=&\frac{{\left( x+2\right) }^{n}\lambda ^{2}+
\frac{{\left( x-2\right) }^{n}}{\lambda ^{2}}+2{x}^{n}}{4}, \\
E_{n}^{\ast (-3)}(x;\lambda ) &=&\frac{{\left( x+3\right) }^{n}\lambda
^{3}+3 {\left( x+1\right) }^{n}\lambda +\frac{3{\left( x-1\right) }^{n}}
\lambda }+ \frac{{\left( x-3\right) }^{n}}{\lambda ^{3}}}{8}, \\
E_{n}^{\ast (-4)}(x;\lambda ) &=&\frac{{\left( x+4\right) }^{n}\lambda
^{4}+4 {\left( x+2\right) }^{n}\lambda ^{2}+\frac{4{\left( x-2\right) }^{n}}{
\lambda ^{2}}+\frac{{\left( x-4\right) }^{n}}{\lambda ^{4}}+6{x}^{n}}{16}, \\
E_{n}^{\ast (-5)}(x;\lambda ) &=&\frac{{\left( x+5\right) }^{n}\lambda
^{5}+5 {\left( x+3\right) }^{n}\lambda ^{3}+10{\left( x+1\right)
^{n}\lambda + \frac{10{\left( x-1\right) }^{n}}{\lambda }+\frac{5{\left(
x-3\right) }^{n}}{ \lambda ^{3}}+\frac{{\left( x-5\right) }^{n}}{\lambda ^{5
}}{32}
\end{eqnarray*
Substituting $\lambda =1$ into the above table, we hav
\begin{eqnarray*}
E_{n}^{\ast (0)}(x) &=&{x}^{n}, \\
E_{n}^{\ast (-1)}(x) &=&\frac{{\left( x+1\right) }^{n}+{\left( x-1\right) }
^{n}}{2}, \\
E_{n}^{\ast (-2)}(x) &=&\frac{{\left( x+2\right) }^{n}+2{x}^{n}+{\left(
x-2\right) }^{n}}{4}, \\
E_{n}^{\ast (-3)}(x) &=&\frac{{\left( x+3\right) }^{n}+3{\left( x+1\right) }
^{n}+3{\left( x-1\right) }^{n}+{\left( x-3\right) }^{n}}{8}, \\
E_{n}^{\ast (-4)}(x) &=&\frac{{\left( x+4\right) }^{n}+4{\left( x+2\right) }
^{n}+6{x}^{n}+4{\left( x-2\right) }^{n}+{\left( x-4\right) }^{n}}{16}, \\
E_{n}^{\ast (-5)}(x) &=&\frac{{\left( x+5\right) }^{n}+5\,{\left( x+3\right)
}^{n}+10{\left( x+1\right) }^{n}+10{\left( x-1\right) }^{n}+5{\left(
x-3\right) }^{n}+{\left( x-5\right) }^{n}}{32}
\end{eqnarray*
By using (\ref{St2a}) and (\ref{Eul.3d}), we compute a few values of the
polynomials $E_{n}^{\ast (-k)}(\lambda )$ as follows
\begin{eqnarray*}
E_{n}^{\ast (0)}(\lambda ) &=&1, \\
E_{n}^{\ast (-1)}(\lambda ) &=&\frac{\lambda +\frac{{\left( -1\right) }^{n}}{
\lambda }}{2}, \\
E_{n}^{\ast (-2)}(\lambda ) &=&\frac{{2}^{n}\lambda ^{2}+\frac{{\left(
-2\right) }^{n}}{\lambda ^{2}}}{4}, \\
E_{n}^{\ast (-3)}(\lambda ) &=&\frac{{3}^{n}\lambda ^{3}+3\lambda +\frac{3\,
\ \left( -1\right) }^{n}}{\lambda }+\frac{{\left( -3\right) }^{n}}{\lambda
^{3} }}{8}, \\
E_{n}^{\ast (-4)}(\lambda ) &=&\frac{{4}^{n}\lambda ^{4}+{2}^{n+2}\lambda
^{2}+\frac{{\left( -2\right) }^{n+2}}{\lambda ^{2}}+\frac{{\left( -4\right) }
^{n}}{\lambda ^{4}}}{16}, \\
E_{n}^{\ast (-5)}(\lambda ) &=&\frac{{5}^{n}\lambda ^{5}+5.{3}^{n}\lambda
^{3}+10\lambda +\frac{10{\left( -1\right) }^{n}}{\lambda }+\frac{5{\left(
-3\right) }^{n}}{\lambda ^{3}}+\frac{{\left( -5\right) }^{n}}{\lambda ^{5}}}{
32}
\end{eqnarray*}
Substituting $\lambda =1$ into the above table, we have a few values of the
second kind Apostol Euler numbers of order $-k$ as follows:
\begin{eqnarray*}
E_{n}^{\ast (0)} &=&1, \\
E_{n}^{\ast (-1)} &=&\frac{{\left( -1\right) }^{n}+1}{2}, \\
E_{n}^{\ast (-2)} &=&\frac{{2}^{n}+{\left( -2\right) }^{n}}{4}, \\
E_{n}^{\ast (-3)} &=&\frac{{3}^{n}+3{\left( -1\right) }^{n}+{\left(
-3\right) }^{n}+3}{8}, \\
E_{n}^{\ast (-4)} &=&\frac{{2}^{n+2}+{\left( -2\right) }^{n+2}+{4}^{n}+{\
\left( -4\right) }^{n}}{16}, \\
E_{n}^{\ast (-5)} &=&\frac{{5}^{n}+5.{3}^{n}+10{\left( -1\right) }^{n}+5{\
\left( -3\right) }^{n}+{\left( -5\right) }^{n}+10}{32}
\end{eqnarray*}
When we look carefully the above table and the generating functions, we can
easily get the following result:
\begin{equation*}
E_{2n+1}^{\ast (-k)}=0,
\end{equation*}
where $n\geq 0$.
That is for $n=1,2,\ldots ,9$ and $k=0,-1,-2,\ldots ,-9$, we compute a few
values of the numbers $E_{n}^{\ast (-k)}$, given by the above relations, as
follows:
\begin{equation*}
\begin{tabular}{lllllllllll}
$n\backslash k$ & $0$ & $-1$ & $-2$ & $-3$ & $-4$ & $-5$ & $-6$ & $-7$ & $-8$
& $-9\cdots $ \\
$0$ & $1$ & $0$ & $1$ & $0$ & $1$ & $0$ & $1$ & $0$ & $1$ & $0$ \\
$1$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ \\
$2$ & $0$ & $1$ & $2$ & $3$ & $4$ & $5$ & $6$ & $7$ & $8$ & $9$ \\
$3$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ \\
$4$ & $0$ & $1$ & $8$ & $21$ & $40$ & $65$ & $96$ & $133$ & $176$ & $225$ \\
$5$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ \\
$6$ & $0$ & $1$ & $32$ & $183$ & $544$ & $1205$ & $2256$ & $3787$ & $5888$ &
$8649$ \\
$7$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ \\
$8$ & $0$ & $1$ & $128$ & $1641$ & $8320$ & $26465$ & $64896$ & $134953$ &
250496$ & $427905$ \\
$\underset{\vdots }{9}$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$
& $0\cdots
\end{tabular
\end{equation*}
\subsection{Algorithm for our computations}
The theory of the algorithms has been very important in Mathematics and in
Computer Science and also in Communications Systems. We know that there are
many ways to compute the second kind Euler polynomials of the negative
order. In this section we give a computation algorithm for computing the
values of these polynomials, which are given by Equations (\ref{St2a}) and (
\ref{Eul.3c}).
\iffalse
\begin{algorithm}[H]
\caption{Let $n$ be nonnegative integers and $\lambda$ real or complex numbers. This algorithm will return the values of the second kind Euler polynomials of the negative
order $E_{n}^{\ast (-k)}(x;\lambda )$ given by Equation (\ref{Eul.3c}).}
\label{Algorithm-1}
\begin{algorithmic}
\Procedure{$Second$\_$Kind$\_$Euler$\_$Poly$}{$n, -k, x, \lambda$}
\State {$\textbf{Begin}$}
\State {$\textbf{Inputs:}$}
\State {${E_{n}^{\ast (-k)}} \leftarrow {k!2^{n-k}}/{\lambda ^{k}}$}
\State {${Y_{3} \leftarrow 0}$}
\State {${\lambda \leftarrow \lambda^{2}}$}
\State {${a \leftarrow \left({x+k}\right)/{2k}}$}
\State {${b \leftarrow \left({x-k}\right)/{2k}}$}
\State {$\textbf{Outputs:}$}
\State {$E_{n}^{\ast (-k)}(x;\lambda ) \leftarrow {E_{n}^{\ast (-k)}}$}
\ForAll{$j$ in $\{ 0,1,2, \ldots ,k\}$}
\State {$Y_{3} \leftarrow Y_3 + {\rm{Binomial\_Coef}}\left( {k,j} \right)*{\rm{Power}}\left( {\lambda,j} \right)*{\rm{Power}} \left( {bk+j\left( a-b\right),n} \right)$}
\EndFor
\State {$y_{3}(n,k;\lambda ;a,b) \leftarrow \left(1/{k!}\right)*{Y_3}$}
\State {${E_{n}^{\ast (-k)}} \leftarrow {E_{n}^{\ast (-k)}}*y_{3}(n,k;\lambda ;a,b)$}
\State \textbf{return} {${E_{n}^{\ast (-k)}}$}
\EndProcedure
\end{algorithmic}
\end{algorithm}
\fi
\section{New families of numbers and polynomials}
In this section, we investigate some properties of the .numbers
W_{n}(\lambda )$, which are related to the the second kind Apostol type
Euler polynomials of order $2$, $E_{n}^{\ast (2)}(1;\lambda )$.
For $n$ and $k$ nonnegative integers, $W_{n}^{(k)}(\lambda )$ define by
means of the following genearting functio
\begin{equation}
F_{w}(t;\lambda ;k)=\frac{1}{\left( \lambda e^{t}+\lambda
^{-1}e^{-t}+2\right) ^{k}}=\sum_{n=0}^{\infty }W_{n}^{(k)}(\lambda )\frac
t^{n}}{n!}. \label{w1a}
\end{equation}
By using the Umbral calculus convention in (\ref{w1}), we get a recurrence
relation for the numbers $W_{n}(\lambda )$. Therefore, we set the following
functional equatio
\begin{equation*}
1=\left( \lambda e^{t}+\lambda ^{-1}e^{-t}+2\right) \sum_{n=0}^{\infty
}W_{n}(\lambda )\frac{t^{n}}{n!}.
\end{equation*
We make some elementary calculations in the above equation, we hav
\begin{equation*}
1=\lambda \sum_{n=0}^{\infty }\frac{t^{n}}{n!}\sum_{n=0}^{\infty
}W_{n}(\lambda )\frac{t^{n}}{n!}+\lambda ^{-1}\lambda \sum_{n=0}^{\infty
\frac{\left( -t\right) ^{n}}{n!}\sum_{n=0}^{\infty }W_{n}(\lambda )\frac
t^{n}}{n!}+2\sum_{n=0}^{\infty }W_{n}(\lambda )\frac{t^{n}}{n!}.
\end{equation*
By using the Cauchy product and $W^{n}(\lambda )$ is replaced by
W_{n}(\lambda )$ in the above equation, we obtai
\begin{equation*}
1=\lambda \sum_{n=0}^{\infty }\sum_{m=0}^{n}\left(
\begin{array}{c}
n \\
\end{array
\right) W_{m}(\lambda )\frac{t^{n}}{n!}+\lambda ^{-1}\sum_{n=0}^{\infty
}\sum_{m=0}^{n}(-1)^{n-m}\left(
\begin{array}{c}
n \\
\end{array
\right) W_{m}(\lambda )\frac{t^{n}}{n!}+2\sum_{n=0}^{\infty }W_{n}(\lambda
\frac{t^{n}}{n!}.
\end{equation*}
By comparing the coefficients of $\frac{t^{n}}{n!}$ on both sides of the
above equation, we arrive at the following theorem:
\begin{theorem}
Let $n$ be a positive integer. and le
\begin{equation*}
W_{0}(\lambda )=\frac{\lambda }{\left( \lambda +1\right) ^{2}}.
\end{equation*
The following recurrence relation holds true
\begin{equation}
2W_{n}(\lambda )+\lambda \sum_{m=0}^{n}\left(
\begin{array}{c}
n \\
\end{array
\right) W_{m}(\lambda )+\lambda ^{-1}\sum_{m=0}^{n}(-1)^{n-m}\left(
\begin{array}{c}
n \\
\end{array
\right) W_{m}(\lambda )=0. \label{w2}
\end{equation}
\end{theorem}
By using (\ref{w2}), we compute a few values of the numbers $W_{n}(\lambda )$
as follows
\begin{equation*}
W_{1}(\lambda )=-\frac{\lambda \left( \lambda -1\right) }{(\lambda +1)^{3}
,W_{2}(\lambda )=-\frac{2\lambda ^{2}}{(\lambda +1)^{4}},W_{3}(\lambda )
\frac{4\lambda (1-\lambda )(\lambda ^{2}-\lambda +1)}{(\lambda +1)^{5}
,\cdots
\end{equation*}
We can show that the numbers $W_{n}(\lambda )$ are associated with the the
second kind Apostol type Euler polynomials of order $2$, $E_{n}^{\ast
(2)}(1;\lambda )$. By combining (\ref{Eul.2}) with (\ref{w1}), we ge
\begin{equation*}
\sum_{n=0}^{\infty }W_{n}(\lambda )\frac{t^{n}}{n!}=\sum_{n=0}^{\infty
\frac{\lambda }{4}E_{n}^{\ast (2)}(1;\lambda )\frac{t^{n}}{n!}.
\end{equation*
Comparing the coefficients of $\frac{t^{n}}{n!}$ on both sides of the above
equation, we get the following relation
\begin{equation*}
W_{n}(\lambda )=\frac{\lambda }{4}E_{n}^{\ast (2)}(1;\lambda ).
\end{equation*
Substituting $k=2$ into (\ref{Eul.2}) and combining with the above equation,
we get the following theorem:
\begin{theorem}
\begin{equation*}
W_{n}(\lambda )=\frac{\lambda }{4}\sum_{m=0}^{n}\left(
\begin{array}{c}
n \\
\end{array
\right) E_{m}^{\ast }(1;\lambda )E_{n-m}^{\ast }(1;\lambda ).
\end{equation*}
\end{theorem}
By using (\ref{C1}) and (\ref{w1a}), we get the following functional
equation
\begin{equation*}
F_{y_{2}}(t,k;\lambda )F_{w}(t;\lambda ;k)=1.
\end{equation*
From this equation, we ge
\begin{equation*}
\sum_{n=0}^{\infty }W_{n}^{(k)}(\lambda )\frac{t^{n}}{n!}\sum_{n=0}^{\infty
}y_{2}(n,k;\lambda )\frac{t^{n}}{n!}=1.
\end{equation*
Therefor
\begin{equation*}
\sum_{n=0}^{\infty }\sum_{m=0}^{n}\left(
\begin{array}{c}
n \\
\end{array
\right) W_{n-m}^{(k)}(\lambda )y_{2}(m,k;\lambda )\frac{t^{n}}{n!}=1.
\end{equation*
Comparing the coefficients of $\frac{t^{n}}{n!}$ on both sides of the above
equation, we give a relation between the numbers $y_{2}(n,k;\lambda )$ and
W_{n}^{(k)}(\lambda )$ by the following theorem:
\begin{theorem}
Let $n$ be a positive integer. Than we hav
\begin{equation*}
\sum_{m=0}^{n}\left(
\begin{array}{c}
n \\
\end{array
\right) W_{n-m}^{(k)}(\lambda )y_{2}(m,k;\lambda )=0.
\end{equation*}
\end{theorem}
For $x$ real numbers, define $W_{n}^{(k)}(x;\lambda )$ by means of the
following generating functio
\begin{equation}
G_{w}(t,x,k;\lambda )=e^{tx}F_{w}(t,k;\lambda )=\sum_{n=0}^{\infty
}W_{n}^{(k)}(x;\lambda )\frac{t^{n}}{n!}. \label{w1b}
\end{equation}
Combining (\ref{w1a}) with (\ref{w1b}), we ge
\begin{equation*}
\sum_{n=0}^{\infty }W_{n}^{(k)}(x;\lambda )\frac{t^{n}}{n!
=\sum_{n=0}^{\infty }\frac{t^{n}}{n!}\sum_{n=0}^{\infty }W_{n}^{(k)}(\lambda
)\frac{t^{n}}{n!}.
\end{equation*
Therefor
\begin{equation*}
\sum_{n=0}^{\infty }W_{n}^{(k)}(x;\lambda )\frac{t^{n}}{n!
=\sum_{n=0}^{\infty }\sum_{m=0}^{n}\left(
\begin{array}{c}
n \\
\end{array
\right) x^{n-m}W_{m}^{(k)}(\lambda )\frac{t^{n}}{n!}.
\end{equation*
Comparing the coefficients of $\frac{t^{n}}{n!}$ on both sides of the above
equation, we get the following theorem:
\begin{theorem}
\begin{equation*}
W_{n}^{(k)}(x;\lambda )=\sum_{m=0}^{n}\left(
\begin{array}{c}
n \\
\end{array
\right) x^{n-m}W_{m}^{(k)}(\lambda ).
\end{equation*}
\end{theorem}
By using (\ref{w1b}) and (\ref{Cad3}), we get the following functional
equation
\begin{equation*}
G(t,x,k;\lambda )=\frac{1}{4^{k}}F_{P1}\left( t,\frac{x}{2};k,\lambda
\right) F_{P1}\left( -t,\frac{x}{2};k,\lambda ^{-1}\right) .
\end{equation*
By combining this equation with (\ref{w1b}) and (\ref{Cad3}), we ge
\begin{equation*}
\sum_{n=0}^{\infty }W_{n}^{(k)}(x;\lambda )\frac{t^{n}}{n!}=\frac{1}{4^{k}
\sum_{n=0}^{\infty }E_{n}^{(k)}\left( \frac{x}{2};\lambda \right) \frac{t^{n
}{n!}\sum_{n=0}^{\infty }(-1)^{n}E_{n}^{(k)}\left( \frac{x}{2};\lambda
^{-1}\right) \frac{t^{n}}{n!}.
\end{equation*
Therefor
\begin{equation*}
\sum_{n=0}^{\infty }W_{n}^{(k)}(x;\lambda )\frac{t^{n}}{n!}=\frac{1}{4^{k}
\sum_{n=0}^{\infty }\sum_{m=0}^{n}(-1)^{n-m}\left(
\begin{array}{c}
n \\
\end{array
\right) E_{m}^{(k)}\left( \frac{x}{2};\lambda \right) E_{n-m}^{(k)}\left(
\frac{x}{2};\lambda ^{-1}\right) \frac{t^{n}}{n!}.
\end{equation*
Comparing the coefficients of $\frac{t^{n}}{n!}$ on both sides of the above
equation, we get the following theorem:
\begin{theorem}
\begin{equation}
W_{n}^{(k)}(x;\lambda )=\frac{1}{4^{k}}\sum_{m=0}^{n}(-1)^{n-m}\left(
\begin{array}{c}
n \\
\end{array
\right) E_{m}^{(k)}\left( \frac{x}{2};\lambda \right) E_{n-m}^{(k)}\left(
\frac{x}{2};\lambda ^{-1}\right) . \label{w1c}
\end{equation}
\end{theorem}
Substituting $x=0$ into (\ref{w1c}), we ge
\begin{equation*}
W_{n}^{(k)}(x;\lambda )=\frac{1}{4^{k}}\sum_{m=0}^{n}(-1)^{n-m}\left(
\begin{array}{c}
n \\
\end{array
\right) E_{m}^{(k)}\left( \lambda \right) E_{n-m}^{(k)}\left( \lambda
^{-1}\right) .
\end{equation*
Setting $k=1$ in the above equation, we hav
\begin{equation*}
W_{n}(x;\lambda )=\frac{1}{4}\sum_{m=0}^{n}(-1)^{n-m}\left(
\begin{array}{c}
n \\
\end{array
\right) E_{m}\left( \lambda \right) E_{n-m}\left( \lambda ^{-1}\right) .
\end{equation*
By using (\ref{C1}), (\ref{Cac3}) and (\ref{w1a}), we get the following
functional equations, respectively
\begin{equation}
F_{w}(t;\lambda ;-k)=(2k)!F_{y_{2}}(t,k;\lambda ) \label{1F}
\end{equation
an
\begin{equation}
F_{N}(t;-k,\lambda )=\sum_{m=0}^{k}\left(
\begin{array}{c}
k \\
\end{array
\right) \frac{1}{2^{m}}F_{w}(t;\lambda ;-m). \label{2F}
\end{equation
By using (\ref{1F}), the numbers $W_{n}(\lambda )$ of order $-k$, denoted by
$W_{n}^{(-k)}(\lambda )$, can be computed by the following theorem:
\begin{theorem}
\begin{equation}
W_{n}^{(-k)}(\lambda )=(2k)!y_{2}(n,k;\lambda ). \label{w2A}
\end{equation}
\end{theorem}
Thanks to Equation (\ref{2F}), a relation between the numbers $E_{n}^{\ast
(-k)}(\lambda )$ and $W_{n}^{(-k)}(\lambda )$ are given by the following
theorem:
\begin{theorem}
\begin{equation*}
E_{n}^{\ast (-k)}(\lambda )=\sum_{m=0}^{k}\left(
\begin{array}{c}
k \\
\end{array
\right) \frac{1}{2^{m}}W_{n}^{(-m)}(\lambda ).
\end{equation*}
\end{theorem}
By using (\ref{CCC3}), for $k=0;1;2;3$ and we $n=0;1;2;3;4$, we can compute
a few values of the numbers $W_{n}^{(-k)}(\lambda )$ given by Equation (\re
{w2A}) as follows:
\begin{array}{ccccc}
n\backslash k & 0 & 1 & 2 & 3 \\
0 & 1 & \lambda +\frac{1}{\lambda } & \lambda ^{2}+4\lambda +\frac{4\lambda
+1}{\lambda ^{2}} & \lambda ^{3}+6\lambda ^{2}+15\lambda +\frac{\lambda }{15
+\frac{6\lambda +1}{\lambda ^{3}} \\
1 & 0 & \lambda -\frac{1}{\lambda } & 2\lambda ^{2}+4\lambda -\frac{4\lambda
+2}{\lambda ^{2}} & 3\lambda ^{3}+12\lambda ^{2}+15\lambda -\frac{\lambda }
15}-\frac{12\lambda +3}{\lambda ^{3}} \\
2 & 0 & \lambda +\frac{1}{\lambda } & 4\lambda ^{2}+4\lambda +\frac{4\lambda
+4}{\lambda ^{2}} & 9\lambda ^{3}+24\lambda ^{2}+15\lambda +\frac{\lambda }
15}+\frac{24}{\lambda ^{2}}+\frac{9}{\lambda ^{3}} \\
3 & 0 & \lambda -\frac{1}{\lambda } & 4\lambda ^{2}+2\lambda -\frac{2\lambda
+4}{\lambda ^{2}} & 27\lambda ^{3}+48\lambda ^{2}+15\lambda -\frac{15}
\lambda }-\frac{48}{\lambda ^{2}}-\frac{27}{\lambda ^{3}} \\
4 & 0 & \lambda +\frac{1}{\lambda } & 16\lambda ^{2}+8\lambda +\frac
4\lambda +16}{\lambda ^{2}} & 81\lambda ^{3}+96\lambda ^{2}+15\lambda +\frac
\lambda }{15}+\frac{96}{\lambda ^{2}}+\frac{81}{\lambda ^{3}
\end{array
$
\section{Further remarks and observations}
In this section we outline some application of our numbers to combinatorial
analysis. In \cite{SimsekNEW}, we gave combinatorial interpretations of the
numbers, $y_{1}(n,k)$, $y_{2}(n,k)$ and the central factorial numbers. We
pursued that the numbers $B(n,k)$ related to the enumerative combinatorics
with the following Bona \cite[P. 46, Exercise 3-4]{Bona} exercises:
\textbf{Exercise 3}. Find the number of ways to place n rooks on an $n\times
n$ chess board so that no two of them attack each other.
\textbf{Exercise 4}. How many ways are there to place some rooks on an
n\times n$ chess board so that no two of them attack each other?
As a result, our new special numbers with their generating functions have
many applications in combinatorial analysis and in analytic number theory.
These numbers are related to the many well-known numbers and polynomials:
the Bernstein basis functions, the array polynomials, the Stirling numbers
of the second kind, the central factorial numbers and also the Golombek's
problem \cite{golombek} \textquotedblleft \textbf{Aufgabe 1088
\textquotedblright .
\begin{acknowledgement}
The paper was supported by the \textit{Scientific Research Project
Administration of Akdeniz University.}
\end{acknowledgement}
|
\section{Introduction}
Competition for resources is one of the most primitive activities of human.
The problems of distributing resources among competing agents are found wherever human exists.
We have developed various systems and tools, e.g., economic markets, auctions/mechanisms, voting systems and money,
to facilitate the distributions and exchanges of resources.
The study of these problems in various settings is now a popular topic concerning both computer scientists and economists,
in the field coined as \emph{Multiagent Resource Allocation};
see \cite{Chev2006} for a fairly recent survey.
While money is provably reducing the complexity of many such problems, in many scenarios monetary transfer is inapplicable;
canonical examples include hospital/facility location determination, peer-to-peer sharing network,
and distribution of computing resources among staffs in the same company.
Mechanism design problems without monetary transfer are thus naturally motivated.
Briefly speaking, a mechanism is a combination of a communication protocol and an algorithm,
for agents to reveal their preferences to an auctioneer, and for the auctioneer to determine a \emph{good allocation} of resources.
There are different ways to interpret the meaning of ``good allocation'', including
social welfare, various measures of fairness (often coined as ``cake cutting''),
or revenue of the auctioneer (when monetary transfer is feasible).
Independent of the interpretation, a favourable feature of a mechanism is strategyproofness
--- the mechanism \emph{motivates} the agents to reveal their preferences \emph{truthfully}.
Strategyproof (SP) mechanism design without monetary transfer has been studied under various contexts, e.g.,\cite{SchummerV2002,ProcacciaT2009,DokowFMN2012,DughmiG2010}.
We focus on the problem formulated by Guo and Conitzer~\cite{GuoC2010}:
design SP mechanisms for allocating divisible items among two agents who have linear preferences over the items.
The target is to attain social welfares which are competitive against those in the first-best mechanism (which is not SP).
In contrast to most of the prior work, we take on an analytic approach for the problem.
While all known SP mechanisms (which we will discuss next) are somewhat naturally motivated,
they do not shed any insight on how an optimal mechanism should look like.
We will present results obtained using analytic methods,
which will suggest that analytical insights are necessary for seeking the optimal mechanisms.
\subsection{Related Work}
Guo and Conitzer~\cite{GuoC2010} considered a sub-class of SP mechanisms called Swap-Dictatorial mechanisms,
in which each agent is a \emph{dictator} with probability $\frac 12$,
who chooses her favourite allocation from a predefined set of allowable allocations,
and the other agent is allocated the remaining items.\footnote{Swap-Dictatorial mechanisms can be generalized to more than two agents,
by first generating a random order of the agents, and then each agent takes turn to choose her favourite allocation.}
They studied two sub-classes of Swap-Dictatorial mechanisms, Increasing-Price (IP) mechanisms and Linear-Increasing-Price (LIP) mechanisms.
For the case with two items, they showed that there is a LIP mechanism which is $0.828$-competitive against the first-best mechanism;
they used a linear program to show that no SP mechanism can be better than $0.841$-competitive.
They also showed that as the number of items goes to infinity, IP and LIP mechanisms have maximal competitiveness of $\frac 12$.
Han et al.~\cite{HanSTZ2011} showed a number of upper bound results on the competitiveness of SP mechanisms, when the numbers of agents and/or items increase.
In particular, they showed that no swap-dictatorial mechanism can be better than $\left(\frac 12 + o_m(1)\right)$-competitive for $m$ items.
In addition, they proved the following characterization result:
in the case with two items, if a mechanism $A$ is symmetric and second order continuously differentiable,
then $A$ is SP if and only if $A$ is swap-dictatorial.
Cole, Gkatzelis and Goel \cite{ColeGG2013-EC} proposed another sub-class of SP mechanisms (for any number of agents)
called \emph{Partial Allocation} (PA) mechanisms, which are not swap-dictatorial.
They showed in another work \cite{ColeGG2013-AAMAS} that a variant of PA mechanism is $\frac 23$-competitive for two agents and any number of items.
Recently, non-SP resource allocation mechanisms are also under study.
The mechanisms are considered as games and the bids of agents are strategies.
The canonical measure of efficiency is the Price of Anarchy.
See~\cite{FLZ2009,BranzeiCDFF2014} and the references therein for more details.
We note that in every swap-dictatorial mechanism, all items are always completely allocated among the agents
--- the auctioneer never holds some of the items from being allocated. However, this is not true in PA mechanism.
We say a mechanism is \emph{full} if all items are always completely allocated among the agents,
and say it is \emph{partial} otherwise.
\subsection{Our Contribution}
Our main contribution is to use analytic methods to \emph{derive} SP mechanisms with competitiveness better than those previously known.
We also improve a linear-program (LP) based proof to show new upper bounds on the competitive ratios for SP full and partial mechanisms.
In addition, we provide the first \emph{compact} upper bound proof.
In Sections \ref{sect:suff-condition}---\ref{sect:human-upper-bound}, we focus on the case with two agents and two items.
In Section \ref{sect:suff-condition}, we first prove a characterization of symmetric SP mechanisms,
which is essentially the same as the Rochet's characterization~\cite{Rochet1985}.
Then we provide a set of sufficient conditions for symmetric SP mechanisms.
We note that while our set of sufficient conditions and the characterization in~\cite{HanSTZ2011} are both of analytical flavor,
the two results are not comparable:
their result focuses on conditions that yield equivalence between SP mechanisms and swap-dictatorial mechanisms (which must be full mechanisms),
while our result is applicable for a broad sub-class of partial mechanisms,
and also mechanisms which are not second order continuously differentiable.
In Section \ref{sect:fully-two-items}, we look into the solution to a LP of Guo and Contizer~\cite{GuoC2010}
for making a few observations and heuristic assumptions, which allow us to \emph{derive} a $\frac 56$-competitive full mechanism;
we believe it is an optimal full mechanism.\footnote{There may be more than one optimal full mechanisms.}
In Section \ref{sect:partial-two-items}, by using our set of sufficient conditions, we consider a sub-family of SP partial mechanisms,
and show that one of such mechanisms is strictly better than $\frac 56$-competitive.
This may be surprising to some practitioners, since it suggests that in general,
the competitiveness can be improved by suitably holding a fraction of items from being allocated.
Guo and Conitzer used the LP to show an upper bound of $0.841$ on the competitiveness of SP full mechanisms.
We will discuss how to prune out a lot of \emph{unnecessary} constraints from their LP.
This allows us to solve the LP with much refined resolution,
and to improve the upper bound to $\frac 56 + \epsilon$, where $\epsilon < 10^{-9}$; we believe the final answer is $\frac 56$.
With a minor modification to their LP, we show an upper bound of $0.8644$ on the competitiveness of SP partial mechanisms.
While the LP-based upper bound proofs are \emph{legitimate},
they may look unsatisfactory to some researchers, due to two reasons.
First, such proofs are hardly verifiable by researchers without the use of a computer.
Second, such LPs are extremely huge for three or more items and thus not solvable in practice,
so they shed no insight for providing a better upper bound when the number of items increases.
Therefore, a \emph{compact} upper bound proof, i.e., a proof which can be easily verifiable by researchers, is preferred.
In Section \ref{sect:human-upper-bound}, we use the SP characterization of Rochet~\cite{Rochet1985}
to provide the first compact proof; the upper bound is $0.9523$.
While this upper bound is worse than those yielded by LP-based proofs,
the compact proof is worth an attention since it might shed insight for generalizations.
In Section \ref{sect:avg-PA}, we consider the cases with two agents and any number of items.
By taking a suitable \emph{average} of some PA mechanisms of Cole et al.~\cite{ColeGG2013-EC,ColeGG2013-AAMAS},
we design a SP mechanism which is at least $0.67776$-competitive.
In Section \ref{sect:DIP}, we propose a new sub-class of SP mechanisms for any number of agents and items,
called \emph{Dynamic-Increasing-Price} (DIP) mechanisms.
DIP mechanism is similar to IP mechanism in the sense that both introduce \emph{virtual money} and \emph{virtual prices}.
However, there is no dictator-swapping process in DIP, and DIP is not swap-dictatorial in general.
Also, the prices in an IP mechanism is independent of agents' preferences, but in a DIP mechanism,
for each agent, the prices of the items depend on other agents' preferences.
In other words, a DIP mechanism enforces all agents to be \emph{complete price takers}.
We note that DIP is well motivated by the classical context of markets: when the scale of the market is large,
each agent in the market has tiny effect on the prices,
so the prices she face are \emph{almost} completely depending on other agents preferences.
We show that the $\frac 56$-competitive mechanism is a DIP mechanism.
\section{Preliminaries}
\paragraph{Problem Setting.}
We study the problem of allocating $m\geq 2$ divisible items, each of one unit, among two agents, referred to as agents 1 and 2.
A vector $(c_1,\cdots,c_m)$ is \emph{normalized} if each $c_j\geq 0$ and $\sum_{j=1}^m c_j = 1$.
Each agent $i$ has a normalized\footnote{See~\cite[Section 2]{GuoC2010} for an explanation on why the utility functions are normalized.}
linear utility function $u_i(\vec{x}_i) = \sum_{j=1}^m u_{ij} x_{ij}$, where $\sum_{j=1}^m u_{ij} = 1$;
her utility function is identified to the normalized utility vector $\vec{u}_i :=(u_{i1},\cdots,u_{im})$.
Each agent $i$ reports to mechanism $A$ a normalized bid vector $\vec{b}_i = (b_{i1},\cdots,b_{im})$.
The mechanism $A$, based on the bids, allocates $A_{ij} (\vec{b}_1,\vec{b}_2)$ unit of item $j$ to agent $i$.
The allocation must be \emph{feasible}, i.e., for any $i,j$, $A_{ij} (\vec{b}_1,\vec{b}_2)\geq 0$ and
for any $j$, $A_{1j} (\vec{b}_1,\vec{b}_2) + A_{2j} (\vec{b}_1,\vec{b}_2) \leq 1$.
Let
$$u_i^A(\vec{b}_1,\vec{b}_2) ~:=~ \sum_{j=1}^m u_{ij}\cdot A_{ij}(\vec{b}_1,\vec{b}_2),$$
which is the utility attained by agent $i$ when agent 1 bids $\vec{b}_1$ and agent 2 bids $\vec{b}_2$.
Mechanism $A$ is \emph{strategyproof} (SP) if for any agent $i$ and for any normalized vectors $\vec{b}_i,\vec{b}_{3-i}$,
$$u_i^A(\vec{u}_i,\vec{b}_{3-i}) ~\geq~ u_i^A(\vec{b}_i,\vec{b}_{3-i}),$$
i.e., agent $i$ is always better off to bid her true utility vector.
Strategyproofness is generally accepted as a favourable feature of a mechanism,
since it discourages agents from having strategic consideration for reporting bids.
Let
$$\text{\textsf{SW}}^A(\vec{u}_1,\vec{u}_2) ~:=~ u_1^A(\vec{u}_1,\vec{u}_2) + u_2^A(\vec{u}_1,\vec{u}_2),$$
which is the social welfare when both agents bid truthfully to mechanism $A$.
Let
$$\text{\textsf{SW}}_{\text{\textsf{OPT}}}(\vecu_1,\vecu_2) ~:=~ \sum_{j=1}^m ~\max\{u_{1j},u_{2j}\},$$
which is the maximum possible social welfare among all feasible allocations;
the feasible allocation that attains the maximum possible social welfare is called the \emph{first-best allocation}.
A SP mechanism $A$ is $\alpha$-competitive if for any $\vec{u}_1,\vec{u}_2$,
$\text{\textsf{SW}}^A(\vec{u}_1,\vec{u}_2) \geq \alpha\cdot \text{\textsf{SW}}_{\text{\textsf{OPT}}}(\vecu_1,\vecu_2)$.
Our target is to design SP mechanisms with high competitive ratios.
\paragraph{Useful Definitions, Facts and Tool.}
It is known that every $\alpha$-competitive SP mechanism has a corresponding
symmetric-over-agents and symmetric-over-items $\alpha$-competitive SP mechanism \cite[Claim 1]{GuoC2010}.
Thus, from now on we focus only on such symmetric SP mechanisms.
When the agents' utility functions are linear, any \emph{weighted average} over SP mechanisms is also SP:
if $A^1,\cdots,A^k$ are SP mechanisms, then $\bar{A}$, defined by the allocation rule
$$\bar{A}_{ij}(\vec{b}_1,\vec{b}_2) ~:=~ \sum_{\ell=1}^k \beta_\ell \cdot A^\ell_{ij}(\vec{b}_1,\vec{b}_2),$$
where the $\beta_\ell$'s are positive and $\sum_{\ell=1}^k \beta_\ell = 1$, is also SP.
We will write $\bar{A}$ as $\sum_{\ell=1}^k \beta_\ell \cdot A^\ell$.
For any utility functions of the agents $u_1,u_2$, let their \emph{attainable utility region} (AUR) be
$$\text{\textsf{AUR}}(u_1,u_2) ~:=~\left\{(r_1,r_2)\left|~\exists\text{ a feasible allocation }(\vec{x}_1,\vec{x}_2)\text{~s.t.~}
u_1(\vec{x}_1) = r_1\text{ and }u_2(\vec{x}_2) = r_2\right.\right\}.$$
\begin{prop}\label{prop:aur-convex}
If $u_1,u_2$ are increasing, concave and continuous functions, then \emph{$\text{\textsf{AUR}}(u_1,u_2)$} is a convex subset of $\mathbb{R}^2$.
\end{prop}
\begin{proof}
Suppose $(r_1,r_2),(r'_1,r'_2)\in\text{\textsf{AUR}}(u_1,u_2)$ and $\beta$ is any real number in $[0,1]$.
It suffices to prove that $(\beta r_1 + (1-\beta) r'_1, \beta r_2 + (1-\beta) r'_2)\in \text{\textsf{AUR}}(u_1,u_2)$.
Let $(\vec{x}_1,\vec{x}_2)$ be a feasible allocation such that for $i=1,2$, $u_i(\vec{x}_i) = r_i$.
Let $(\vec{x}_1',\vec{x}_2')$ be a feasible allocation such that for $i=1,2$, $u_i(\vec{x}_i') = r_i'$.
Note that $(\beta\cdot\vec{x}_1 + (1-\beta)\cdot\vec{x}_1',\beta\cdot\vec{x}_2 + (1-\beta)\cdot\vec{x}_2')$ is also a feasible allocation.
Since $u_i$ is a concave function,
$$u_i(\beta\cdot\vec{x}_i + (1-\beta)\cdot\vec{x}_i') ~\geq~ \beta \cdot u_i(\vec{x}_i) + (1-\beta)\cdot u_i(\vec{x}_i') ~=~ \beta r_i + (1-\beta) r_i'.$$
Then, since $u_i$ is increasing and continuous, there exists $\gamma_i\in [0,1]$ such that
$$u_i(\beta\gamma_i\cdot\vec{x}_i + (1-\beta)\gamma_i\cdot\vec{x}_i') ~=~ \beta r_i + (1-\beta) r_i',$$
i.e., $(\beta\gamma_1\cdot\vec{x}_1 + (1-\beta)\gamma_1\cdot\vec{x}_1',\beta\gamma_2\cdot\vec{x}_2 + (1-\beta)\gamma_2\cdot\vec{x}_2')$ is a feasible allocation
that verifies $(\beta r_1 + (1-\beta) r'_1, \beta r_2 + (1-\beta) r'_2)\in \text{\textsf{AUR}}(u_1,u_2)$.
\end{proof}
\section{A Set of Sufficient Conditions for Strategyproofness}\label{sect:suff-condition}
In Sections \ref{sect:suff-condition}---\ref{sect:human-upper-bound}, we focus on the case with two items.
In this case, each normalized utility vector has the form $(t,1-t)$, which is essentially single-parameter.
We assume that each agent $i$ bids a number $b_i\in [0,1]$,
which is supposed to be the first entry of her normalized utility vector.
A symmetric mechanism $A$ can be described by a single function $A:[0,1]^2\rightarrow\mathbb{R}^+$, such that
\begin{align*}
A_{11}(b_1,b_2) \equiv A(b_1,b_2) ~~&~~A_{12}(b_1,b_2) \equiv A(1-b_1,1-b_2)\\
A_{21}(b_1,b_2) \equiv A(b_2,b_1) ~~&~~A_{22}(b_1,b_2) \equiv A(1-b_2,1-b_1)
\end{align*}
\medskip
In this section, we first prove a characterization of SP symmetric mechanisms (Theorem \ref{thm:Rochet}),
which follows almost directly from a characterization result of Rochet~\cite[Theorem 1]{Rochet1985}; we will provide a self-contained proof.
Then we use the characterization to provide a set of sufficient conditions for strategyproofness (Theorem \ref{thm:SP-char}),
which will be used in the next three sections.
Let
$$\hat{u}^A(b_1,b_2) := b_1 \cdot A(b_1,b_2) + (1-b_1) \cdot A(1-b_1,1-b_2),$$
which is the utility attained by agent 1 if her true utility vector is $(b_1,1-b_1)$, she bids truthfully, and agent 2 bids $b_2$.
\begin{theorem}[{\cite[Theorem 1]{Rochet1985}}] \label{thm:Rochet}
Let $A$ be a symmetric mechanism for two items. $A$ is SP if and only if
\begin{enumerate}
\item[(a)] for any fixed $b_2\in [0,1]$, $\hat{u}^A(b_1,b_2)$ is a convex function of $b_1$, and
\item[(b)] for any fixed $b_2\in [0,1]$, $z := A(t_1,b_2) - A(1-t_1,1-b_2)$ is a sub-gradient of $\hat{u}^A(b_1,b_2)$ at $b_1 = t_1$,
i.e., for any $b_1\in [0,1]$,
$$\hat{u}^A(b_1,b_2) ~\geq~ \hat{u}^A(t_1,b_2) + z \cdot (b_1 - t_1).$$
\end{enumerate}
\end{theorem}
We note that since we are considering symmetric mechanisms, in the above theorem,
stating the conditions (a) and (b) w.r.t.~agent $1$ only is without loss of generality.
\begin{proof}
If $A$ is SP, then $\hat{u}^A(b_1,b_2) = \sup_{b_1'} b_1 \cdot A(b_1',b_2) + (1-b_1) \cdot A(1-b_1',1-b_2)$,
which is a supremum of linear functions of $b_1$, thus $\hat{u}^A(b_1,b_2)$ is convex w.r.t.~$b_1$, i.e., condition (a) holds.
Next, we show that condition (b) is equivalent to strategyproofness as below; the second, third and the fourth statements below are equivalent
since the R.H.S.~of the inequalities in them are indeed identical.
\begin{align*}
~&~A\text{ is SP}\\
\Leftrightarrow~&~\forall b_1,b_2,t_1,~\hat{u}^A(b_1,b_2) ~\geq~ b_1\cdot A(t_1,b_2) + (1-b_1)\cdot A(1-t_1,1-b_2)\\
\Leftrightarrow~&~\forall b_1,b_2,t_1,~\hat{u}^A(b_1,b_2) ~\geq~ t_1\cdot A(t_1,b_2) + (1-t_1)\cdot A(1-t_1,1-b_2) + z\cdot (b_1 - t_1)\\
\Leftrightarrow~&~\forall b_1,b_2,t_1,~\hat{u}^A(b_1,b_2) ~\geq~ \hat{u}^A(t_1,b_2) + z\cdot (b_1 - t_1)\\
\Leftrightarrow~&~\text{condition (b) holds.}
\end{align*}
\end{proof}
\begin{theorem}\label{thm:SP-char}
Let $A$ be a symmetric mechanism, described by function $A(b_1,b_2)$.
If for any fixed $b_2$, $A(b_1,b_2)$ is increasing, continuous and piecewise continuously differentiable w.r.t.~$b_1$,
and if for any $t_1,t_2\in [0,1]$, the equality\footnote{A clarification on the perhaps misleading notation:
at any specific point $(y_1,y_2)$, $\frac{\partial A}{\partial b_1} (y_1,y_2)$ is
the value of the partial derivative of $A$ w.r.t.~its first parameter at that point.
To be crystal clear, $\frac{\partial A}{\partial b_1} (y_1,y_2) = \lim_{\delta\rightarrow 0} \frac{A(y_1+\delta,y_2) - A(y_1,y_2)}{\delta}$.}
\begin{equation}\label{eq:SP-char}
t_1 \cdot \frac{\partial A}{\partial b_1} (t_1,t_2) = (1-t_1) \cdot \frac{\partial A}{\partial b_1} (1-t_1,1-t_2)
\end{equation}
holds within each piecewise interval, then $A$ is SP.
\end{theorem}
\begin{proof}
Recall that
$$\hat{u}^A(b_1,b_2) ~=~ b_1\cdot A(b_1,b_2) + (1-b_1)\cdot A(1-b_1,1-b_2).$$
By the assumptions on $A$, $\frac{\partial \hat{u}^A}{\partial b_1}(t_1,t_2)$ exists everywhere (except perhaps at the endpoints of the piecewise intervals),
and its value is
$$A(t_1,t_2) ~-~ A(1-t_1,1-t_2) ~+~ t_1 \cdot \frac{\partial A}{\partial b_1} (t_1,t_2) ~-~ (1-t_1)\cdot \frac{\partial A}{\partial b_1} (1-t_1,1-t_2).$$
The final two terms cancel out due to \eqref{eq:SP-char}. Hence,
\begin{equation}\label{eq:SP-char-Rochet}
\frac{\partial \hat{u}^A}{\partial b_1}(t_1,t_2) ~=~ A(t_1,t_2) ~-~ A(1-t_1,1-t_2).
\end{equation}
Since $A$ is continuous and piecewise continuously differentiable w.r.t.~its first parameter,
at any endpoint $(t_1,t_2)$ of a piecewise interval,
the left and right partial derivatives of $\hat{u}^A$ w.r.t.~its first parameter are equal,
i.e., $\frac{\partial \hat{u}^A}{\partial b_1}(t_1,t_2)$ exists at the endpoint too.
When $t_1$ increases, $A(t_1,t_2)$ increases but $A(1-t_1,1-t_2)$ decreases.
By \eqref{eq:SP-char-Rochet}, $\frac{\partial \hat{u}^A}{\partial b_1}(t_1,t_2)$ increases with $t_1$ within each piecewise interval.
Thus, $\hat{u}^A(b_1,b_2)$ is convex w.r.t.~$b_1$, i.e., condition (a) in Theorem \ref{thm:Rochet} holds.
\eqref{eq:SP-char-Rochet} and condition (a) imply that $A(t_1,t_2) - A(1-t_1,1-t_2)$ is a subgradient of $\hat{u}^A(b_1,t_2)$ at $b_1 = t_1$,
i.e., condition (b) in Theorem \ref{thm:Rochet} holds. Then by Theorem \ref{thm:Rochet}, $A$ is SP.
\end{proof}
Note that the allocation functions of Partial Allocation mechanisms in~\cite{ColeGG2013-EC,ColeGG2013-AAMAS} are discontinuous,
so Theorem \ref{thm:SP-char} is not applicable.
\section{A $\frac 56$-Competitive Full Mechanism for Two Items}\label{sect:fully-two-items}
Guo and Conitzer~\cite{GuoC2010} introduced the linear program (LP) below,
which represents the optimal full mechanism when the bids are restricted to be multiples of $1/N$ for some integer $N$.
Let $[N]$ denote the set of all multiples of $1/N$ which are between zero and one.
\begin{align}
&~~~~~~\max \lambda\nonumber\\
&\forall t_1,t_1',t_2\in [N],&&\hat{u}^A(t_1,t_2) \geq t_1 \cdot A(t_1',t_2) + (1-t_1)\cdot A(1-t_1',1-t_2); & \text{(stragyproofness)}\nonumber\\
&\forall t_1,t_2\in [N],&& \text{\textsf{SW}}^A(t_1,t_2)\geq (1+|t_1-t_2|)\lambda; & \text{(competitiveness)}\nonumber\\
&\forall t_1,t_2\in [N],&& A(t_1,t_2) + A(t_2,t_1) = 1; & \label{eq:fully}\\
&\forall t_1,t_2\in [N],&& A(t_1,t_2)\geq 0. & \nonumber
\end{align}
They solved the LP with $N=50$. The optimal $\lambda$ value, which is $0.841$,
is an upper bound on the optimal competitiveness of SP full mechanisms for two items.
The LP has $\Theta(N^2)$ variables and $\Theta(N^3)$ constraints, which is efficiently solvable only for small $N$.
However, one would expect that the strategyproofness constraints with large $|t_1-t_1'|$ are unnecessary.
So we keep only those constraints with $|t_1-t_1'| = 1/N$.
This reduces the number of constraints to $\Theta(N^2)$,
allowing us to solve the LP with a much refined resolution of $N=400$.
We obtain an improved upper bound of $\frac 56 + \epsilon$, where $\epsilon < 10^{-9}$.
We believe that $\frac 56$ is the final answer.
We make two observations from the solution to the LP, and make two heuristic assumptions.
We then use the observations and assumptions to \emph{derive} a full mechanism which is $\frac 56$-competitive.
\medskip
\noindent\textbf{Observation 1. }There exists a function $f:[0,1]\rightarrow\mathbb{R}$ such that $A(t_1,t_2) = f(t_1) - f(t_2) + \frac 12$.
Furthermore, $f$ is increasing, continuous and piecewise differentiable.
\medskip
\noindent\textbf{Observation 2. }For all $t\in \left[0,\frac 15\right]$, $f(t) = 0$. For all $t\in\left[\frac 45,1\right]$, $f(t)$ is a constant.
\medskip
\noindent\textbf{Assumption 3. }The function $A(t_1,t_2)$ satisfies the equality \eqref{eq:SP-char},
except at points where $t_1\in \left\{1/5,4/5\right\}$.
\medskip
With Observation 1, Observation 2 and Assumption 3, \eqref{eq:SP-char} yields
\begin{equation}\label{eq:SP-char-f}
\forall t\in [0,1]\setminus\left\{1/5,4/5\right\},~~t f'(t) = (1-t) f'(1-t).
\end{equation}
With Observation 2, the social welfare attained when $t_2=0$ is $1+t_1 \left(f(t_1) - f(1-t_1) + \frac 12\right)$,
while $\text{\textsf{SW}}_{\text{\textsf{OPT}}}(t_1,t_2) = 1+t_1$.
For the mechanism to be $\frac 56$-competitive, the following inequality must hold:
$1+t_1\left(f(t_1) - f(1-t_1) + \frac 12\right) \geq \frac 56 (1+t_1)$, or equivalently
\begin{equation}\label{eq:f-diff}
f(t) - f(1-t) \geq \frac 13 - \frac{1}{6t}.
\end{equation}
Observe that by \eqref{eq:SP-char-f}, once the values of $f(t)$ for $0\leq t\leq \frac 12$ are known,
the values of $f(t)$ for $\frac 12 \leq t\leq 1$ can be determined.
We now state the final heuristic assumption: in \eqref{eq:f-diff}, the equality holds for $t\in \left[\frac 15,\frac 12\right]$.
\medskip
\noindent\textbf{Assumption 4. }$\forall t\in\left[\frac 15, \frac 12\right]$, $f(t) - f(1-t) = \frac 13 - \frac{1}{6t}$.
\medskip
With \eqref{eq:f-diff} and Assumption 4, we can solve $f$ using calculus, which is:
\begin{equation}\label{eq:f-explicit}
f(t) = \begin{cases}
0, & t\in\left[0,\frac 15\right];\\
\frac 56 - \frac{1}{6t} - \frac 16 \ln(5t), & t\in\left[\frac 15,\frac 12\right];\\
\frac 12 - \frac 16 \ln(5-5t), & t\in\left[\frac 12,\frac 45\right];\\
\frac 12, & t\in \left[\frac 45,1\right].
\end{cases}
\end{equation}
\begin{theorem}
The full mechanism $A$ as described in Assumption 1 and \eqref{eq:f-explicit} is feasible, SP and $\frac 56$-competitive.
\end{theorem}
\begin{proof}
Feasibility trivially holds. Strategyproofness follows from Observation 1, Assumption 3 and Theorem \ref{thm:SP-char}.
For competitiveness, note that
$$\text{\textsf{SW}}^A(t_1,t_2) ~=~ 1 + (t_1 - t_2) \left[ f(t_1) - f(t_2) - f(1-t_1) + f(1-t_2) \right].$$
Showing that $A$ is $\frac 56$-competitive is equivalent to showing that $\text{\textsf{SW}}^A(t_1,t_2) \geq \frac 56 (1 + |t_1-t_2|)$,
which can be done with an appropriate case analysis (which is needed due to the piecewise definition of $f$) and simple calculus;
we skip the details.
\end{proof}
\section{A Partial Mechanism for Two Items -- Strictly Better than $\frac 56$-Competitive}\label{sect:partial-two-items}
By changing the equality sign in \eqref{eq:fully} to a $\leq$ sign, Guo and Conitzer's LP covers partial mechanisms also.
The modified LP provides an upper bound of $0.8644$.
We look into its solution, as we did in Section \ref{sect:fully-two-items}, but we do not recognize a nice pattern.
Since we solve the modified LP with high resolution, we \emph{believe} that an optimal partial mechanism attains competitive ratio close to $0.8644$,
which beats the $\frac 56+\epsilon$ upper bound for full mechanism.
Yet, to formally \emph{prove} that an optimal partial mechanism is strictly better than an optimal full mechanism,
we ought to provide a concrete, and preferably compact, SP partial mechanism which is strictly better than $(\frac 56+\epsilon)$-competitive.
This is the purpose of the current section.
Let $f_1 : \left[0,1/2\right]\rightarrow \mathbb{R}$ and $f_2: \left[1/2, 1\right] \rightarrow \mathbb{R}$ be two increasing and continuously differentiable functions
such that for all $t\in \left[0,1/2\right]$, $t f_1'(t) = (1-t) f_2'(1-t)$, and $f_1(0)=f_2\left(1/2\right) = 0$.
Also, let $Q,R: [0,1]\rightarrow\mathbb{R}^+$ be two functions.
Then define the function
$$A(t_1,t_2) ~:=~
\begin{cases}
Q(t_2)\cdot f_1(t_1) + R(t_2),&t_1\in\left[0,1/2\right];\\
A\left(\frac 12,t_2\right) + Q(1-t_2)\cdot f_2(t_1),&t_1\in\left(1/2,1\right].
\end{cases}$$
It is easy to verify that the above function $A$ satisfies all conditions required in Theorem \ref{thm:SP-char},
and thus it yields a SP mechanism, modulo feasibility constraint.
Our strategy is to pick some choice of $f_1,f_2$, and then use an LP to find out $Q,R$,
such that $A$ is feasible and attains a good competitiveness.
As before, we formulate the LP with bids restricted to be multiples of $1/N$ for some integer $N$.
The LP is stated below; note that we impose a slightly stricter feasibility constraint,
in which only $(1-\delta)$ fraction, for some $\delta>0$, of each item can be allocated.
The reason will be clear later.
\begin{align*}
&~~~~~~\max \lambda\\
&\forall t_1,t_2\in [N],&&~ A(t_1,t_2) + A(t_2,t_1)\leq 1-\delta;\\
&\forall t_1,t_2\in [N],&&~ \text{\textsf{SW}}^A(t_1,t_2)\geq (1+|t_1-t_2|)\lambda.\\
&\forall t\in [N],&&~ Q(t),R(t)\geq 0.
\end{align*}
It is easy to verify that the above program is an LP with variables $Q(t),R(t)$ for $t\in [N]$, plus an extra variable $\lambda$.
Lacking further insight on how a good choice of $f_1,f_2$ should be, we try the natural candidate $f_1(t) := t$,
and hence $f_2(t) := \ln (2t) - t + 1/2$. Then we solve the above LP with resolution $N=1000$ and $\delta = 2.92/2000$.
The optimal $\lambda$ is larger than $0.835524$. Let the optimal solution be $Q^*,R^*$.
We note that the maximum entry in $Q^*$ is less than $1.46$.
Now, we are ready to describe the desired symmetric SP mechanism $\tilde{A}$.
$\tilde{A}$ takes $t_1,t_2$ as bids from the two agents.
Let $\tt_1,\tt_2$ be the values by rounding $t_1,t_2$ to its nearest multiple of $1/N$.
$$\tilde{A}(t_1,t_2) ~:=~
\begin{cases}
Q^*(\tt_2)\cdot f_1(t_1) + R^*(\tt_2),&t_1\in\left[0,1/2\right];\\
\tilde{A}\left(\frac 12,\tt_2\right) + Q^*(1-\tt_2)\cdot f_2(t_1),&t_1\in\left(1/2,1\right].
\end{cases}$$
The rounding is needed because the domains of $Q^*,R^*$ are $[N]$.
The rounding does not destroy strategyproofness; one can verify that $\tilde{A}$ is SP using Theorem \ref{thm:SP-char}.
\footnote{In general, if an allocation function for agent $i$, denoted by $A_i'(\vec{u}_i,\vec{u}_{-i})$, yields a SP mechanism,
then any other allocation function $A_i''(\vec{u}_i,\vec{u}_{-i}) \equiv A_i'(\vec{u}_i,T(\vec{u}_{-i}))$,
where $T$ is an \emph{arbitrary} function with range compatible with the domain of the second parameter of $A_i'$, yields a SP mechanism too.}
Since the maximum entry of $Q^*$ is less than $1.46$, and since the derivatives of $f_1,f_2$ are bounded by $1$,
$\frac{\partial \tilde{A}}{\partial b_1} (b_1,b_2) < 1.46$ for all $b_1,b_2$.
Also, note that $|t_i - \tt_i| \leq 1/2000$. Due to the first constraint of the LP above, $\tilde{A}(\tt_1,\tt_2) + \tilde{A}(\tt_2,\tt_1) \leq 1-\delta$.
Thus,
$$\tilde{A}(t_1,t_2) + \tilde{A}(t_2,t_1) ~<~ \tilde{A}(\tt_1,\tt_2) + \frac{1.46}{2000} + \tilde{A}(\tt_2,\tt_1) + \frac{1.46}{2000} ~\leq~ 1.$$
This verifies the feasibility of $\tilde{A}$.
To bound the competitiveness of $\tilde{A}$, first note that by Theorem \ref{thm:Rochet},
$$\hat{u}^{\tilde{A}}(t_1,\tt_2) \geq \hat{u}^{\tilde{A}}(\tt_1,\tt_2) - \frac{1}{2000}~~~~\text{and}~~~~\hat{u}^{\tilde{A}}(t_2,\tt_1) \geq \hat{u}^{\tilde{A}}(\tt_2,\tt_1) - \frac{1}{2000}.$$
From the LP, we have
$$\frac{\hat{u}^{\tilde{A}}(\tt_1,\tt_2) + \hat{u}^{\tilde{A}}(\tt_2,\tt_1)}{1+|\tt_1-\tt_2|} ~\geq~ 0.835524.$$
With the above two sets of inequalities, we proceed a simple error analysis to show that the competitiveness of $\tilde{A}$ with bids $t_1,t_2$ is
$$\frac{\hat{u}^{\tilde{A}}(t_1,t_2) + \hat{u}^{\tilde{A}}(t_2,t_1)}{1+|t_1-t_2|} ~=~ \frac{\hat{u}^{\tilde{A}}(t_1,\tt_2) + \hat{u}^{\tilde{A}}(t_2,\tt_1)}{1+|t_1-t_2|} ~\geq~ 0.833689,$$
which is strictly larger than $\frac 56+\epsilon$.
\medskip
To program mechanism $\tilde{A}$, we need to store the values of $Q^*,R^*$ at $1001$ discrete values,
and to compute $f_1,f_2$ at arbitrary real values in their domain.
While this may not look compact to some people, $\tilde{A}$ is a concrete SP partial mechanism that breaks the $\frac 56$ barrier.
\section{A Compact Upper Bound Proof}\label{sect:human-upper-bound}
In this section, we provide a compact upper bound proof on the competitiveness of SP mechanisms.
First of all, we make the following qualitative observation.
Suppose $\vec{u}_i = (t_i,1-t_i)$.
For some $t_1,t_2$, if agent $1$ earns a \emph{too} high utility from the mechanism,
then the utility earned by agent $2$ has to be very low, forcing a low competitive ratio;
conversely, if agent $1$ earns \emph{too} low a utility, this forces a low competitive ratio too.
Thus, to attain a high competitive ratio $h$, there is a restricted range of utility values that each agent can earn.
Geometrically, the utilities earned by the agents must lie in the intersection of
$\text{\textsf{AUR}}(t_1,t_2)$ and $R_h(t_1,t_2) := \{(r_1,r_2)\,|\,r_1+r_2 \geq h \cdot \text{\textsf{SW}}_{\text{\textsf{OPT}}}(t_1,t_2)\}$.
Briefly, our proof strategy is: for some $h,t_1,t_2$,
since the utilities earned by both agents are restricted to certain range, the allocations are restricted too.
Then due to Theorem \ref{thm:Rochet}, the subgradients of $\hat{u}^A$ at certain points are also restricted.
We then show that if $h$ is too high, the above restrictions add up to forbid the existence of $\hat{u}^A$.
Let
\begin{align*}
U_h(t_1,t_2) &:= \max\{r_1\,|\,\exists (r_1,r_2)\in\text{\textsf{AUR}}(t_1,t_2)\cap R_h(t_1,t_2)\}\\
L_h(t_1,t_2) &:= \min\{r_1\,|\,\exists (r_1,r_2)\in\text{\textsf{AUR}}(t_1,t_2)\cap R_h(t_1,t_2)\}.
\end{align*}
\begin{theorem}
For the case with two items, no SP mechanism is better than $0.9523$-competitive.
\end{theorem}
\begin{proof}
By~\cite[Claim 1]{GuoC2010}, it suffices to prove that no symmetric SP mechanism is better than $0.9523$-competitive.
Suppose the contrary that there exists an $h$-competitive symmetric SP mechanism $A$, where $h=0.9523$.
We compute the following functions explicitly:
$$
U_h(t_1,0.1) = \begin{cases}
1 - \frac{(\frac{11}{10}-t_1)h-1}{\frac{1}{10}-t_1} t_1, & t_1\in \left[0,\frac{11}{10}-\frac{1}{h}\right];\\
1, & t_1 \in \left[\frac{11}{10}-\frac{1}{h},\frac{1}{h}-\frac{9}{10}\right];\\
1 + \frac{(t_1+\frac{9}{10})h-1}{t_1-\frac{1}{10}} (t_1-1), & t_1 \in \left[\frac{1}{h}-\frac{9}{10},1\right].
\end{cases}
$$
$$L_h(t_1,0.1) = \begin{cases}
\frac{(\frac{11}{10}-t_1)h-1}{\frac{1}{10}-t_1} (1-t_1), & t_1\in \left[0,\frac{11}{10}-\frac{1}{h}\right];\\
0, & t_1 \in \left[\frac{11}{10}-\frac{1}{h},\frac{1}{h}-\frac{9}{10}\right];\\
\frac{(t_1+\frac{9}{10})h-1}{t_1-\frac{1}{10}} t_1, & t_1 \in \left[\frac{1}{h}-\frac{9}{10},1\right].
\end{cases}
$$
$$U_h(t_1,0) = \min\left\{1,1 + \frac{(t_1+1)h-1}{t_1} (t_1-1)\right\}.$$
In particular, we have
$$0.4753 = L_h(0,0.1) \leq \hat{u}^A(0,0.1) \leq U_h(0,0.1) = 1.$$
Let $q := \hat{u}^A(0,0.1)$. Then $A(1,0.9) = q$. By feasibility, $A(0,0.1)\geq 0$.
Thus, $A(0,0.1) - A(1,0.9) \geq -q$.
By Theorem \ref{thm:Rochet}, for any $t_1\in [0,1]$,
\begin{equation}\label{eq:support1}
\hat{u}^A(t_1,0.1) \geq q - q t_1.
\end{equation}
Since $A$ is $h$-competitive, $\hat{u}^A(0.1,0) \geq \frac{11}{10}h-q$.
Since $A(1,0.9) = q$, $A(0.9,1) \leq 1-q$.
Then we have
$$A(0.1,0) ~=~ \frac{\hat{u}^A(0.1,0)-0.9\cdot A(0.9,1)}{0.1} ~\geq~ 11h-9-q,$$
and hence $A(0.1,0) - A(0.9,1) \geq 11h-10$.
By Theorem \ref{thm:Rochet}, for any $t_1\in [0.1,1]$,
\begin{equation}\label{eq:support2}
\hat{u}^A(t_1,0) \geq \frac{11}{10}h-q + (11h-10) (t_1-0.1).
\end{equation}
Next, we show that for any possible value of $q$, there exists $t_1$ such that
either condition \eqref{eq:support1} or condition \eqref{eq:support2} is violated.
Observe that when $q$ increases, the lower bound in \eqref{eq:support1} increases,
while the lower bound in \eqref{eq:support2} decreases.
Thus, once we find a $q^*$ such that
\begin{enumerate}
\item[(a)] $\exists t_1'\in [0,1]$ with $q^*-q^* t_1' > U_h(t_1',0.1)$, and
\item[(b)] $\exists t_1'' \in [0.1,1]$ with $\frac{11}{10}h-q^* + (11h-10) (t_1''-0.1) > U_h(t_1'',0)$,
\end{enumerate}
then for any $q\geq q^*$, by \eqref{eq:support1} and (a), $\hat{u}^A(t_1',0.1) > U_h(t_1',0.1)$, a contradiction;
for any $q < q^*$, by \eqref{eq:support2} and (b), $\hat{u}^A(t_1'',0) > U_h(t_1'',0)$, again a contradiction.
We find that $q^* = 0.6979$ satisfies (a) and (b), with $t_1' = 0.26$ and $t_1'' = 0.32$.
\end{proof}
\section{A $0.67776$-Competitive Mechanism for Multiple Items}\label{sect:avg-PA}
Cole et al.~\cite{ColeGG2013-EC} introduced a family of SP mechanisms called \emph{Partial Allocation} (PA) mechanisms,
which work for multiple agents and multiple items. We describe the two-agent version $\mathbf{PA}_c$ below.\footnote{\label{fn:pa}Since
the utility functions are normalized, $0<u_2(\vec{a}_2)^c,u_1(\vec{a}_1)^{1/c}\leq 1$. So Step 2 of $\mathbf{PA}_c$ is legitimate.\\
Also, we note that the eventual utility attained by agent 1 is $u_1(\vec{a}_1)\cdot u_2(\vec{a}_2)^c = W(u_1,u_2)$,
and the eventual utility attained by agent 2 is $u_1(\vec{a}_1)^{1/c}\cdot u_2(\vec{a}_2) = W(u_1,u_2)^{1/c}$.}
In~\cite{ColeGG2013-AAMAS}, they showed that a variant of a PA mechanism, which we denote by \textbf{PA}$_{\max}$,
is SP and it is $\frac 23$-competitive for two agents and multiple items.
We will show that by taking a suitable weighted average of two PA mechanisms and \textbf{PA}$_{\max}$, we break the $\frac 23$ barrier.
\smallskip
\noindent
\begin{tabular}{|l|l|}
\hline
$\mathbf{PA}_c(\vec{u}_1,\vec{u}_2)$ for $0 < c < \infty$: & \textbf{PA}$_{\max}$$(\vec{u}_1,\vec{u}_2)$:\\
1. Compute the feasible allocation $(\vec{a}_1,\vec{a}_2)$ & 1. Compute the allocation of $\mathbf{PA}_1(\vec{u}_1,\vec{u}_2)$.\\
~~~~that maximizes $u_1(\vec{a}_1) \cdot u_2(\vec{a}_2)^c$. & 2. Compute the allocation in which each\\
~~~~Let $W(u_1,u_2)$ denote the maximal value. & ~~~~item is split evenly among the two agents.\\
2. Agent $1$ is allocated a $u_2(\vec{a}_2)^c$ fraction of $\vec{a}_1$; & 3. Output the allocation that yields higher\\
~~~~agent $2$ is allocated a $u_1(\vec{a}_1)^{1/c}$ fraction of $\vec{a}_2$. & ~~~~social welfare.\\
~~~~(See footnote \ref{fn:pa}). & \\
\hline
\end{tabular}
\smallskip
To build up an intuition on why this might work, we look at an almost worst case scenario for \textbf{PA}$_{\max}$,
where there are two items, $\vec{u}_1=(0.99,0.01)$ and $\vec{u}_2=(0.5,0.5)$.
In $\mathbf{PA}_1$, $\vec{a}_1=(1,0)$ and $\vec{a}_2=(0,1)$,
$u_1(\vec{a}_1) = 0.99$ and $u_2(\vec{a}_2) = 0.5$.
The eventual allocation to agent $1$ is $\frac 12 \vec{a}_1$, i.e., reducing $\vec{a}_1$ by half.
This reduction harms the eventual social welfare hugely.
If we consider $\mathbf{PA}_c$ for some $c$ less than $1$, say $c=0.5$,
the eventual allocation to agent $1$ is much better, which is $0.5^{0.5} \vec{a}_1 \approx 0.707 a_1$.
The eventual allocation to agent $2$ in $\mathbf{PA}_{0.5}$ is $0.99^2 \vec{a}_2$, which is slightly worse than $0.99 \vec{a}_2$,
the eventual allocation to agent $2$ in $\mathbf{PA}_1$.
But overall, the social welfare in $\mathbf{PA}_{0.5}$ is much better.
However, there are bad scenarios for $\mathbf{PA}_{0.5}$, e.g., when the utility functions of the two agents in the last paragraph are swapped.
To attain an overall good competitiveness, we consider some weighted average of $\left(\frac 12\cdot \mathbf{PA}_c + \frac 12\cdot \mathbf{PA}_{1/c}\right)$ and \textbf{PA}$_{\max}$.
For each choice of $c$ and the weights, by using the tool of AUR,
we can compute the competitive ratio with math software.
We find that for a suitable choice of $c$ and the weights, a competitive ratio of $0.67776$ is attained.
\begin{theorem}
The mechanism
$$\left(\frac{1029}{4000} \cdot \mathbf{PA}_{0.421} + \frac{1029}{4000} \cdot \mathbf{PA}_{1/0.421} + \frac{971}{2000}\cdot \text{\emph{\textbf{PA}$_{\max}$}}\right)$$
is SP, and it is at least $0.67776$-competitive.
\end{theorem}
\begin{proof}
The mechanism is SP since it is a weighted average of three SP mechanisms.
For any $u_1,u_2$, let $(u^*_1,u^*_2)\in \text{\textsf{AUR}}(u_1,u_2)$ be a point that attains the optimal social welfare $\text{\textsf{SW}}_{\text{\textsf{OPT}}}(\vecu_1,\vecu_2)$.
Observe that $(1,0),(0,1)\in \text{\textsf{AUR}}(u_1,u_2)$.
Thus by Proposition \ref{prop:aur-convex}, $\text{\textsf{AUR}}(u_1,u_2)$ contains the line segment $\ell_1(u^*_1,u^*_2)$ that joins $(1,0)$ and $(u^*_1,u^*_2)$,
and also the line segment $\ell_2(u^*_1,u^*_2)$ that joins $(0,1)$ and $(u^*_1,u^*_2)$.
Let $\ell(u^*_1,u^*_2) := \ell_1(u^*_1,u^*_2) \cup \ell_2(u^*_1,u^*_2)$.
Recall the notion $W(u_1,u_2)$ defined in $\mathbf{PA}_c$.
Note that $\text{\textsf{SW}}^{\mathbf{PA}_c}(u_1,u_2) = W(u_1,u_2) + W(u_1,u_2)^{1/c}$, and
$$W(u_1,u_2) \geq \max_{(r_1,r_2)\in \ell(u^*_1,u^*_2)} r_1 \cdot (r_2)^c.$$
These allow us to turn to the easier single-variate optimization problems along the two line segments.
For our choices of $c$ and the weights, we run through all possible values of $(u^*_1,u^*_2)$,
which is the set $\{(u^*_1,u^*_2)\,|\,0\leq u^*_1,u^*_2\leq 1\text{ and }1\leq u^*_1+u^*_2\leq 2\}$, to compute the competitive ratio.
We perform this with math software, but first only considering discrete points where $u^*_1,u^*_2$ are multiples of $\frac{1}{2000}$.
The competitive ratio over these discrete points is at least $0.67844$.
For general $(u^*_1,u^*_2)$, we first round them down to the nearest multiple of $\frac{1}{2000}$ to $(\tilde{u}^*_1,\tilde{u}^*_2)$.
Observe that
$$\max_{(r_1,r_2)\in \ell(u^*_1,u^*_2)} r_1 \cdot (r_2)^c \geq \max_{(r_1,r_2)\in \ell(\tilde{u}^*_1,\tilde{u}^*_2)} r_1\cdot (r_2)^c,$$
and $1\leq u^*_1 + u^*_2 \leq \tilde{u}^*_1 + \tilde{u}^*_2 + \frac{1}{1000}$.
Thus, the competitive ratio over general points is at least $0.67844\times \left(1-\frac{1}{1000}\right) > 0.67776$.
\end{proof}
\section{Dynamic-Increasing-Price Mechanisms}\label{sect:DIP}
In this section, we propose a sub-class of SP mechanisms for any number of agents and items.
For simplicity, we describe the general form of DIP mechanisms for the case with two agents.
It is easy to see how to generalize to the cases with any number of agents.
Recall that the number of items is $m$.
For each agent $i$, she has one unit of \emph{virtual money}.
For each item $j$, agent $i$ will be given a price function $P_j^{u_{3-i}}: [0,1]\rightarrow \mathbb{R}^+ \cup \{0,+\infty\}$,
which is an increasing function that depends on the utility function of the agent $3-i$.
We will write $P_j^{-i}$ as a shorthand for $P_j^{u_{3-i}}$.
The value $P_j^{-i}(y)$ describes the marginal price when agent $i$ has already purchased $y$ units of item $j$.
In other words, if agent $i$ purchases $x_j$ units of item $j$,
she needs to pay $T_j^{-i}(x_j) := \int_0^{x_j} P_j^{-i}(y)\,dy$ unit of her virtual money.
Agent $i$ will purchase an allocation $\vec{x} = (x_1,x_2,\cdots,x_m)$ which is in
$$\argmax_{\substack{\forall j,~0\leq x_j\leq 1 \\ \sum_{j=1}^m T_j^{-i}(x_j) \leq 1}}\,\,\,\,\sum_{j=1}^m u_{ij} x_j.$$
A DIP mechanism is obviously SP, since each agent is accepting prices which she herself cannot influence, and uses them to decide an optimal purchase.
Yet, feasibility is a delicate issue, particularly when $m$ is large.
While there are some similarities between DIP and IP mechanisms,
we note that there is no process of dictator-swapping in DIP, and DIP is not swap-dictatorial and not full in general.
Another difference between DIP and IP is that in DIP the price functions are \emph{price versus quantity of item already purchased},
while in IP they are \emph{price versus virtual money already spent}.
DIP is well motivated, as explained below.
In the first-best allocation, if $u_{1j} > u_{2j}$, then agent $1$ gets all of item $j$.
When $u_{1j}$ is high but $u_{2j}$ is low, ideally we want to construct the price function $P_j^{-1},P_j^{-2}$ such that
$P_j^{-2}$ is lower to encourage agent $1$ to purchase more item $j$,
but $P_j^{-1}$ is higher to discourage agent $2$ from purchasing more item $j$.
Adjusting the prices dynamically in this manner might help pushing the allocation in DIP towards the first best allocation,
and therefore might be hopeful to attain better competitiveness.
The way that DIP works is similar to \emph{price taking} in the context of Fisher market:
suppose an agent enters a Fisher market where items are sold, the prices she takes surely depend on other agents' preferences,
while she herself can also have influence on the prices --- as a mechanism, DIP deliberately removes such influence by herself.
In a non-rigorous sense, we may say that DIP is a \emph{rescue} for Fisher markets from being non-strategyproof.
We note that Cole et al.~\cite[Theorem 1]{ColeGG2013-AAMAS} proved that when there are two agents and multiple items,
Fisher market equilibrium allocation (which is also the \emph{proportional fair} allocation) is highly competitive against the first-best allocation.
As a simple example, the SP mechanism that allocates half of each item to both agents is DIP,
by setting $P_j^{-i}(y) = 0$ for $y\in\left[0,\frac 12\right]$ and $P_j^{-i}(y) = +\infty$ for $y\in\left[\frac 12,1\right]$.
\begin{prop}
The $\frac 56$-competitive mechanism given in Section \ref{sect:fully-two-items} is a DIP mechanism.
\end{prop}
We give the price functions for agent $1$ on item $1$ below; other price functions are defined symmetrically.
Recall that $\vec{u}_2 = (t_2,1-t_2)$, and also the definition of $f$ in \eqref{eq:f-explicit}.
Let $\tau=\frac 12 - f(t_2)$. Then
$$P_1^{-1} (y) :=
\begin{cases}
0, & y\in \left[0,\tau\right]\\
C, & y\in \left(\tau,f(\frac 12) + \tau\right]\\
\frac{C}{g(y)} - C, & y\in \left(f(\frac 12) + \tau, \frac 12 + \tau\right]\\
+\infty,& y\in \left(\frac 12 + \tau, 1\right],
\end{cases}$$
where $g(y)$ denotes the unique value of $z\in \left[\frac 15,\frac 12\right]$ such that $f(1-z) - f(t_2) + \frac 12 = y$,
and $C$ is the positive constant such that $\int_0^{1-f(t_2)}P_1^{-1} (y)\,dy = 1$.
\section{Discussion and Open Problems}
The most important problem for future research is to seek optimal competitive mechanisms.
As we have already seen, even for the case with two items, where the setting is essentially single-parameter,
the use of analytical tools seems unavoidable --- for instance,
it looks unlikely to have a \emph{natural} interpretation of the $\frac 56$-competitive mechanism,
which we believe to be an optimal full mechanism.
In the study of revenue-optimal mechanisms with prior, more advanced analytical tools,
including duality theory and variational calculus, have played key roles. See, e.g., \cite{PP2011,GK2014,GK2015}.
We believe that such tools will be useful for our problem too;
for instance, duality theory is likely to be useful for showing that the upper bound for full mechanism is \emph{exactly} $\frac 56$.
In Section \ref{sect:human-upper-bound}, we use Rochet's characterization to prove a non-trivial upper bound.
The proof only considers the restrictions to two cross sections of $\hat{u}^A$,
so clearly it has not yet fully exploited the power of the characterization.
An interesting research agenda is to seek a more sophisticated use of the characterization for proving better upper bounds,
for either the case with two items, or for those with more items.
\section*{Acknowledgments}
The author thanks the anonymous reviewers for their thoughtful suggestions which help improving the structure and presentation of this paper.
\bibliographystyle{alpha}
|
\section{Introduction}\label{Intro_fR}
The energy scales in particle physics are arranged in a hierarchical manner. While the scale of Weak interaction corresponds to $E\sim 10^{3}~\textrm{GeV}$, the Strong interaction at a scale of $E\sim 10^{16}~\textrm{GeV}$ exceeds the Weak scale by a factor of $10^{13}$. This large difference leads to a fine tuning problem in the scheme of renormalization --- known as the \textit{gauge hierarchy problem}. This fine tuning is absolutely necessary to renormalize the mass of Higg's Boson, which recently have been detected with a mass of $127~\textrm{GeV}$. At face value, this fine tuning is what nature prefers and the question is \textit{why?} Hence it is natural to ask, is there a more fundamental principle from which this fine tuning would appear naturally.
There have been a large amount of work to address the hierarchy problem, a few candidates have emerged out of it --- supersymmetry, technicolor and extra dimensions. In this work we will be concerned solely about the third alternative, i.e., we will assume the actual spacetime has more than three spatial dimensions (commonly referred to as the \emph{bulk}), while the spacetime we live in is a four-dimensional hypersurface in the bulk (commonly referred to as the \emph{brane}). The two immediate observable consequences are --- change of $1/r^{2}$ gravitational force law at the length scale of extra dimensions, existence of massive graviton through Kaluza-Klein tower \cite{ArkaniHamed:1998rs,Antoniadis:1998ig,Antoniadis:1990ew,Rubakov:1983bb,Horava:1996ma,Kaloper:1999sm,
Cvetic:1996vr,Cohen:1999ia,Maia:2001gq}.
To probe these extra dimensions one needs to have high enough energy or high enough curvature, such that the relevant energy scale of the problem becomes close to the Planck scale. General Relativity, described by the Einstein-Hilbert action is considered to be an effective theory of gravity, valid much below the fundamental Planck scale \cite{Buchbinder1992}. Once energies approach the Planck scale, one not only expects to observe deviations from the Einstein-Hilbert action but also signatures of the extra dimensions. This is particularly relevant, since future colliders will probe higher and higher energies such that aspects beyond general relativity should become apparent. Since the ultraviolet behaviour of the true gravity theory is yet unknown one hopes that in these high energy/high curvature regimes deviations from standard model or deviations from Einstein gravity may appear through existence of extra dimensions. To capture some of the aspects of ``quantum gravity'' one is tempted to consider how the presence
of higher curvature (and higher derivative) invariants in the higher dimensional gravitational action modifies the well-known results \cite{Rizzo:2006wf,Konoplya:2008ix,Brown:2007dd,Rizzo:2006sb,Cai:2006pq}.
The higher derivative terms that one can add to the Einstein-Hilbert action are not unique. However many of these terms can lead to a linear instability, called Ostrogradski instability, leading to appearance of ghost fields and hence will not be considered in this work. Among the higher curvature theories, the Lanczos-Lovelock gravity and the $f(R)$ gravity are of much importance. The Lanczos-Lovelock gravity is special in the sense that the field equations derived from the Lanczos-Lovelock action contain only second order derivatives of the metric and have natural thermodynamic interpretation \cite{Chakraborty:2015wma,Chakraborty:2015hna,Chakraborty:2015kva,Chakraborty:2014rga,
Chakraborty:2014joa}. On the other hand, $f(R)$ gravity was first introduced to explain both early and late time exponential expansion of the universe without invoking additional matter components, e.g., dark energy \cite{Nojiri:2001ae,Nojiri:2010wj,Sotiriou:2008rp,DeFelice:2010aj,Paliathanasis:2015aos,Basilakos:2011rx,Zhong:2010ae}. But only addressing cosmological observations do \emph{not} lead to a viable model, for that the $f(R)$ theories should pass the local gravity tests --- perihelion precession of Mercury and bending angle of light as well. It turns out that Solar System experiments do not exclude the viability of $f(R)$ theories at scales shorter than the cosmological ones, but provide constraints on $f''(R)$ and hence constraints on the parameters of the model. Thus it is can be affirmed that extended gravity theories cannot be ruled out, definitively, using Solar System experiments \cite{Capozziello:2005bu,Capozziello:2007ms,Sotiriou:2005xe,Capozziello:2006jj}.
It is also well known that $f(R)$ gravity theories can be related to scalar-tensor theories by a conformal transformation at the action level \cite{Barrow:1988xh,Capozziello:1996xg,Nojiri:2010wj,Sotiriou:2008rp,DeFelice:2010aj,Anand:2014vqa,Bahamonde:2016wmz,
Parry:2005eb,Catena:2006bd,Chiba:2013mha}. Thus it is important to consider the following situation --- obtaining field equations from the scalar-tensor representation and from $f(R)$ gravity representation. Since the two actions are related by a conformal transformation, the field equations should also be equivalent. However the situation is \emph{not} trivial, since the metric in scalar-tensor representation depends on the conformal factor, its variation can potentially lead to various additional terms, which must cancel other terms \emph{exactly} in order to arrive at the equivalence. If the equivalence exist, we can use it to solve field equations for scalar-tensor theory and obtain the solution corresponding to $f(R)$ action and vice-versa. This would be advantageous, since in \emph{general} solving the field equations for $f(R)$ gravity, where $R$ is not a constant, is difficult\footnote{Note that in cosmological context one can solve the field equations for $f(R)$ gravity by a trick,
known as the \emph{reconstruction method} \cite{Nojiri:2006gh,Capozziello:2006dj}.
We will have occasion to comment on this method when we compare our technique introduced in this
work with the \emph{reconstruction scheme}.} \cite{Nojiri:2010wj,Chakraborty:2014zya,Bhattacharya:2016lup,Capozziello:2011et,Barausse:2007pn,
Capozziello:2009nq,Aghamohammadi:2010yq,Capozziello:2011wg}. While the corresponding scalar-tensor solution could in principle be much simpler. The same should work on the brane hypersurface as well. The effective field equations on the brane derived through Gauss-Codazzi formalism in the $f(R)$ representation \cite{Chakraborty:2014xla,Dadhich:2000am,Shiromizu:1999wj,Harko:2004ui,Chakraborty:2015bja,Chakraborty:2015taq} should be equivalent to the same but derived from the scalar-tensor representation. The non-triviality of this result originates from the quadratic combination of energy momentum tensor and extrinsic curvature appearing in effective field equations.
As an aside, we should mention that the conformal transformation is well motivated \emph{only} when the spacetime does not have singularities. Such singularity free spacetimes have been obtained earlier in the context of cosmology with moduli dependent loop corrections of the gravitational part of superstring effective action with orbifold compactifications \cite{Antoniadis:1993jc}. However to obtain a singularity free description it was necessary that the stress energy tensor associated with the modulus should violate the strong energy condition. In circumstances, where the energy conditions are obeyed, one obtains singular solutions in general. For singular spacetimes, viz., cosmological spacetimes near the Big Rip or Big Crunch the transformation can break down. In those contexts it can exhibit peculiar behavior, e.g., the Big Rip singularity, which may appear in some versions of $f(R)$ gravity can either map itself to infinite past or future, or can be replaced by a Big Crunch singularity
\cite{Briscese:2006xu,Bahamonde:2016wmz}. Another point requires clarification at this stage, this has to do with physical non-equivalence of the two frames. All the comments phrased above has to do with mathematical equivalence, but the physical solutions can be very different \cite{Briscese:2006xu}. This is evident, since the conformal factor can change the complete structure of the spacetime. This fact was pointed out earlier in \cite{Dabrowski:2008kx} by showing that through conformal transformation one can create matter and as a result, one frame is empty while another has matter, clearly they are physically non-equivalent. This should not come as a surprise, since Schwarzschild metric under conformal transformation no longer satisfies Einstein's equations. Further in the cosmological context for $f(R)$ gravity model it was explicitly demonstrated \cite{Briscese:2006xu,Capozziello:2010sc} that neither the Hubble parameter nor the deceleration parameter matches in Jordan and Einstein frame, showing the
physical non-equivalence. In view of the above, the phrase ``equivalence'' in the following sections should be understood in a mathematical sense, \emph{not} in a physical sense. Further, we will contend ourselves, with \emph{only those spacetimes (or regions of spacetimes) which are regular}, such that the conformal transformation between the two frames is well defined throughout the region of interest.
The paper is organized as follows: In \ref{Sec_01} we present a brief review of the equivalence between the $f(R)$ gravity and scalar-tensor theory in five dimensions and hence the equivalence between the bulk field equations as well. \ref{Sec_02} is devoted to show the equivalence between the effective field equations on the brane. The application of the bulk equivalence is presented in \ref{Sec_03}. There we have started from scalar-tensor theory and have solved the bulk equations, from which the solution in $f(R)$ representation is obtained. In \ref{Sec_04} we consider brane spacetime where the solution in scalar-tensor representation starting from $f(R)$ representation is derived as another explicit example. We conclude with a brief discussion on this technique. All the relevant calculations are presented in \ref{GBAPP_01}.
We have set the fundamental constants, $c$ and $\hbar$ to unity and shall work with mostly positive signature of the metric. The Latin indices, $a,b,\ldots$ runs over the full spacetime indices, while Greek indices, $\mu,\nu,\ldots$ stands for four dimensional spacetime.
\section{Equivalence of gravitational field equations in the bulk}\label{Sec_01}
The starting point for any fundamental theory corresponds to correct identification of the dynamical variable and the associated field equations. An useful trick to obtain the field equations is to introduce an action principle, extremizing which one can obtain the field equations. Along similar lines, in gravitational theories as well one considers the metric $g_{ab}$ as dynamical. Given an appropriate action, when arbitrary variations of $g_{ab}$ are considered, the gravitational action reaches extremum value only if $g_{ab}$ satisfies the gravitational field equations. For example, Einstein's equations follow from variation of the Einstein-Hilbert action, which is the Ricci scalar. At high enough curvature (or energy) the Einstein-Hilbert action is most likely to be supplemented by higher curvature corrections. Among many such viable modifications, $f(R)$ theories are of particular interest. The action for $f(R)$ gravity model (also known as the Jordan frame) in five spacetime dimensions reads,
\begin{align}\label{J_Action}
S_{J}\equiv \int d^{5}x\sqrt{-g}\left[\frac{f(R)}{2\kappa _{5}^{2}}+\mathcal{L}_{m}\right]=\int d^{5}x\sqrt{-g}\left[\left(\frac{f'R}{2\kappa _{5}^{2}}-\frac{f'R-f}{2\kappa _{5}^{2}}\right)+\mathcal{L}_{m}\right]~,
\end{align}
where $\kappa _{5}$ is the five dimensional gravitational constant, $\mathcal{L}_{m}$ is the matter Lagrangian and $f'$ stands for $df/dR$. Conformal transformation of the Jordan frame metric $g_{ab}$ results in: $\mathbf{g}_{ab}=\Omega ^{2}g_{ab}$, where $\Omega$ is the conformal factor. The resulting action, written in terms of $\mathbf{g}_{ab}$ can be obtained using the transformation properties of curvature tensor yielding (as presented in \ref{GBAPP_01}),
\begin{align}\label{E_Action_01}
S&=\int d^{5}x\sqrt{-\mathbf{g}}\left[\Omega ^{-5}\left\lbrace \frac{f'}{2\kappa _{5}^{2}}\left(\Omega ^{2}\mathbf{R}+8\Omega ^{2}\boldsymbol{\square}\ln \Omega -12\mathbf{g}^{ab}\nabla _{a}\Omega \nabla _{b}\Omega\right)-U(f)\right\rbrace +\boldsymbol{\mathcal{L}_{m}}\right]~.
\end{align}
Here $U(f)$ stands for $(f'R-f)/2\kappa _{5}^{2}$ and $\boldsymbol{\mathcal{L}_{m}}$ is the matter Lagrangian in the conformally transformed action. Note that, in our convention, all the metric dependent quantities originating from conformal transformation of the Jordan frame are boldfaced. We will follow this convention throughout this work. Under the following identifications
\begin{align}\label{conditions_01}
\Omega ^{3}=f';\qquad \kappa _{5}\phi =\frac{2}{\sqrt{3}}\ln f'=2\sqrt{3}\ln \Omega~,
\end{align}
the action presented in \ref{E_Action_01} reduces to the following form (known as the Einstein frame action),
\begin{align}\label{E_Action_02}
S_{E}=\int d^{5}x\sqrt{-\mathbf{g}}\left(\frac{\mathbf{R}}{2\kappa _{5}^{2}}-\frac{1}{2}\mathbf{g}^{ab}\nabla _{a}\phi \nabla _{b}\phi-V(\phi)+\boldsymbol{\mathcal{L}_{m}}\right)+\int d^{4}x \sqrt{h}\frac{4}{\kappa _{5}^{2}}n^{c}\nabla _{c}\ln \Omega~.
\end{align}
The Einstein frame action can be divided into a bulk term and a surface term, as evident from \ref{E_Action_02}. If we are only interested in variation of the action and derivation of the field equations, the boundary term can be safely ignored. However, while concentrating on brane dynamics, which is a boundary effect, the surface term will be important. Further, the potential term $V(\phi)$ appearing in \ref{E_Action_02} corresponds to,
\begin{align}\label{conditions_02}
V(\phi)=\frac{f'(\phi)R(\phi)-f(\phi)}{2\kappa _{5}^{2}f'(\phi)^{5/3}}~,
\end{align}
which can be obtained by solving \ref{conditions_01}. Extremizing the Einstein frame action presented in \ref{E_Action_02} with respect to arbitrary variations of the metric $\mathbf{g}^{ab}$ we readily obtain the field equations in the Einstein frame,
\begin{align}\label{E_bulk}
\mathbf{G}_{ab}\equiv \mathbf{R}_{ab}-\frac{1}{2}\mathbf{R}\mathbf{g}_{ab}=\kappa _{5}^{2}\left[\nabla _{a}\phi \nabla _{b}\phi-\mathbf{g}_{ab}\left(\frac{1}{2}\mathbf{g}^{cd}\nabla _{c}\phi \nabla _{d}\phi+V(\phi)\right)\right]+\kappa _{5}^{2}\mathbf{T}_{ab}~.
\end{align}
The energy momentum tensor $\mathbf{T}_{ab}$ is obtained from the matter action $\boldsymbol{\mathcal{L}_{m}}$ by variation of the Einstein frame metric $\mathbf{g}_{ab}$. Using the conformal transformation, the energy momentum tensor $\mathbf{T}_{ab}$ in the Einstein frame can be related to the energy momentum tensor $T_{ab}$ in the Jordan frame as,
\begin{align}
\mathbf{T}_{ab}&\equiv-\frac{2}{\sqrt{-\mathbf{g}}}\frac{\delta \left(\sqrt{-\mathbf{g}}\boldsymbol{\mathcal{L}_{m}}\right)}{\delta \mathbf{g}^{ab}}
\nonumber
\\
&=-\frac{2}{\sqrt{-g}\Omega ^{5}}\frac{\delta \left(\sqrt{-g}\mathcal{L}_{m}\right)}{\Omega ^{-2}\delta g^{ab}}
=\frac{1}{f'}T_{ab}~.
\end{align}
Since the Jordan frame action $S_{J}$ is derivable from Einstein frame action $S_{E}$ through conformal transformation, the bulk equations derived from them should coincide. The above statement, though physically well motivated is by no means trivial. This has to do with the fact that while deriving the Einstein frame equations one should vary the Einstein frame metric $\mathbf{g}_{ab}$. This in turn leads to arbitrary variation of the Jordan frame metric $g_{ab}$ and the conformal factor $\Omega$. Since the conformal factor can be written in terms of the curvature tensor due to the identification in \ref{conditions_01}, it can lead to various additional correction terms. It is not clear a priori, how these terms combine and yield correct field equations in the Jordan frame. Since there exist no explicit derivation of the same, it is worthwhile to explicitly demonstrate the equivalence.
In order to prove the same we will start from \ref{E_bulk} and shall try to write every curvature tensor components in terms of the Jordan frame metric $g_{ab}$. This can be done using transformation properties of curvature tensors between the two frames related by conformal transformation, leading to,
\begin{align}
\mathbf{R}_{ab}=R_{ab}-\frac{\nabla _{a}\nabla _{b}f'}{f'}+\frac{4}{3}\frac{\nabla _{a}f'\nabla _{b}f'}{f'^{2}}-g_{ab}\frac{\square f'}{3f'}~,
\end{align}
and
\begin{align}
\mathbf{g}_{ab}\mathbf{R}=g_{ab}R-\frac{8}{3}g_{ab}\frac{\square f'}{f'}+\frac{4}{3}g_{ab}\frac{g^{cd}\nabla _{c}f'\nabla _{d}f'}{f'^{2}}~.
\end{align}
Having expressed both the Ricci tensor and Ricci scalar in the Einstein frame in terms of the conformal factor and corresponding curvature components in the Jordan frame, the Einstein tensor in the Einstein frame can be expressed as,
\begin{align}
\mathbf{G}_{ab}=G_{ab}-\frac{\nabla _{a}\nabla _{b}f'}{f'}+\frac{4}{3}\frac{\nabla _{a}f'\nabla _{b}f'}{f'^{2}}-\frac{2}{3}g_{ab}\frac{g^{cd}\nabla _{c}f'\nabla _{d}f'}{f'^{2}}+g_{ab}\frac{\square f'}{f'}~.
\end{align}
Further the contribution from scalar field present in the right hand side of \ref{E_bulk} can be written in terms of $f(R)$ and its various derivatives as,
\begin{align}\label{Tab_phi}
\nabla _{a}\phi \nabla _{b}\phi &-\mathbf{g}_{ab}\left(\frac{1}{2}\mathbf{g}^{cd}\nabla _{c}\phi \nabla _{d}\phi+V(\phi)\right)
=\frac{1}{\kappa _{5}^{2}}\left[\frac{4}{3}\frac{\nabla _{a}f'\nabla _{b}f'}{f'^{2}}-\frac{2}{3}g_{ab}\frac{g^{cd}\nabla _{c}f'\nabla _{d}f'}{f'^{2}}-g_{ab}\frac{f'R-f}{2f'}\right]~.
\end{align}
Using these relations between Einstein frame and Jordan frame, the field equations in Einstein frame presented in \ref{E_bulk} can be written as,
\begin{align}
f'G_{ab}+\frac{f'R-f}{2}g_{ab}-\nabla _{a}\nabla _{b}f'+g_{ab}\square f'=\kappa _{5}^{2}T_{ab}~,
\end{align}
which is precisely the field equations one would have obtained by extremizing \ref{J_Action} for arbitrary variation of the Jordan frame metric $g_{ab}$. Hence follows the equivalence. As a consequence if one can solve for $\mathbf{g}_{ab}$ starting from the field equations in the Einstein frame, the solution in Jordan frame can be obtained through a conformal transformation and vice versa. We should emphasize that the above statement though mathematically correct, practically might require suitable approximations for inverting various functional relations connecting the two frames. We will provide detailed comments on this aspect later on, while providing concrete examples.
Even though we have used metric formalism to arrive at the equivalence, one can also use another method known as \emph{Palatini method}. For discussions on the same we refer our reader to \cite{Olmo:2005hc,Iglesias:2007nv,Barausse:2007ys,Sotiriou:2008rp,Olmo:2011uz}.
\section{Equivalence of effective field equations on the brane}\label{Sec_02}
In the previous section, we have shown the equivalence between the bulk field equations derived from the Einstein and the Jordan frame. However from the perspective of brane world, governed by effective field equations derived from the bulk action, the equivalence of the effective field equations are more important. The effective field equations involve various quadratic combinations of the extrinsic curvature and the matter energy momentum tensor. Thus all the additional terms with their appropriate factors present in the Einstein frame must cancel each other such that effective field equations in the Jordan frame is obtained. In this connection we would like to highlight that in most of the works related to $f(R)$ gravity the surface term in the Einstein frame is ignored, however in order to prove the equivalence on the brane this term is absolutely necessary. Hence the equivalence of effective field equations too, is a non-trivial statement. In this section we will explicitly demonstrate the same.
The bulk field equations in the Einstein frame involves energy momentum tensor of the scalar field along with any other matter fields which may be present in the bulk. The bulk energy momentum tensor $T_{ab}^{\rm bulk}$ (the trace is denoted by $T^{\rm bulk}$) will induce an effective brane energy momentum tensor $T_{ab}^{\rm brane}$ as,
\begin{align}\label{bulk_brane}
\kappa _{4}^{2}T_{\mu \nu}^{\rm brane}=\frac{2}{3}\kappa _{5}^{2}\left[T_{ab}^{\rm bulk}e^{a}_{\mu}e^{b}_{\nu}+h_{\mu \nu}\left(T_{ab}^{\rm bulk}n^{a}n^{b}-\frac{1}{4}T^{\rm bulk}\right)\right]~.
\end{align}
Here $\kappa _{5}$ is the five-dimensional (i.e., bulk) gravitational constant while $\kappa _{4}$ is the four-dimensional (i.e., brane) gravitational constant. Moreover, the object, $e^{a}_{\mu}$ stands for $\partial x^{a}/\partial y^{\mu}$, where $y^{\mu}$ corresponds to the brane coordinates and $x^{a}$ are the bulk coordinates. The normal to the brane hypersurface being $n_{a}$ such that the induced metric on the brane hypersurface becomes $h_{\mu \nu}=e^{a}_{\mu}e^{b}_{\nu}(g_{ab}-n_{a}n_{b})$ \cite{Poisson}.
To obtain the effective field equations in the Jordan frame, we need to express the scalar field in terms of $f(R)$ and its derivatives. Thus, the bulk energy momentum tensor $\mathbf{T}_{ab}^{(\phi)}$ for the scalar field reads,
\begin{align}
\kappa _{5}^{2}\mathbf{T}_{ab}^{(\phi)}=\frac{4}{3}\nabla _{a}\ln f'\nabla _{b}\ln f'-\frac{2}{3}g_{ab}\left(g^{cd}\nabla _{c}\ln f'\nabla _{d}\ln f'\right)-g_{ab}\frac{f'R-f}{2f'}~.
\end{align}
Using which the following results can be obtained
\begin{align}
\kappa _{5}^{2}\mathbf{T}_{ab}^{(\phi)}\mathbf{n}^{a}\mathbf{n}^{b}&=\frac{1}{\Omega ^{2}}\Big[\frac{4}{3}\left(n^{a}\nabla _{a}\ln f'\right)^{2}-\frac{2}{3}\nabla _{a}\ln f'\nabla ^{a}\ln f'-\frac{f'R-f}{2f'}\Big]~,
\\
\kappa _{5}^{2}\mathbf{T}^{(\phi)}&=\frac{1}{\Omega ^{2}}\Big[\frac{4}{3}\nabla _{a}\ln f'\nabla ^{a}\ln f'-\frac{10}{3}\nabla _{a}\ln f'\nabla ^{a}\ln f'-5\frac{f'R-f}{2f'}\Big]~.
\end{align}
From these expressions of the bulk energy momentum tensor of the scalar field, the brane energy momentum tensor, after some simplifications (using \ref{bulk_brane} in particular), leads to,
\begin{align}
\kappa _{4}^{2}\mathbf{T}^{(\phi)\textrm{brane}}_{\mu \nu}&=\frac{8}{9}\nabla _{\mu}\ln f'\nabla _{\nu}\ln f'-\frac{5}{9}h_{\mu \nu}\nabla _{a}\ln f'\nabla ^{a}\ln f'
\nonumber\
\\
&-\frac{3}{4}h_{\mu \nu}\frac{f'R-f}{3f'}+\frac{8}{9}h_{\mu \nu}\left(n^{a}\nabla _{a}\ln f'\right)^{2}~.
\end{align}
Let us now work through the last bit of this analysis regarding Einstein tensor. Using the transformation properties of Riemann tensor and Ricci scalar (see for example, \ref{GBAPP_01}) we immediately obtain the following result for the induced Einstein tensor,
\begin{align}
\mathbf{G}_{\mu \nu}=G_{\mu \nu}-2\frac{\nabla _{\mu}\nabla _{\nu}\Omega}{\Omega}+2h_{\mu \nu}\frac{h^{\alpha \beta}\nabla _{\alpha}\nabla _{\beta}\Omega}{\Omega}+4\frac{\nabla _{\mu}\Omega \nabla _{\nu}\Omega}{\Omega ^{2}}-h_{\mu \nu}\frac{h^{\alpha \beta}\nabla _{\alpha}\Omega \nabla _{\beta}\Omega}{\Omega ^{2}}~.
\end{align}
In this particular case, the conformal factor $\Omega$ is related to $df/dR$ through the relation $\Omega=f'^{1/3}$. Using this relation, after some straightforward manipulation and simplification we arrive at,
\begin{align}
\mathbf{G}_{\mu \nu}=G_{\mu \nu}-\frac{2}{3}\frac{\nabla _{\mu}\nabla _{\nu}f'}{f'}+\frac{8}{9}\frac{\nabla _{\mu}f'\nabla _{\nu}f'}{f'^{2}}
-\frac{5}{9}h_{\mu \nu}h^{\alpha \beta}\frac{\nabla _{\alpha}f'\nabla _{\beta}f'}{f'^{2}}+h_{\mu \nu}h^{\alpha \beta}\frac{2\nabla _{\alpha}\nabla _{\beta}f'}{3f'}~.
\end{align}
The only remaining part corresponds to the electric part $\mathbf{E}_{\mu \nu}$ of the bulk Weyl tensor $C_{abcd}$. From the transformation property of the Weyl tensor it immediately follows: $\mathbf{E}_{\mu \nu}=E_{\mu \nu}$. Combining all these, in the Einstein frame the effective gravitational field equations on the brane takes the form,
\begin{align}
\mathbf{T}_{\mu \nu}^{\rm brane}&=\mathbf{G}_{\mu \nu}-\left\lbrace \mathbf{K}\mathbf{K}_{\mu \nu}-\mathbf{K}^{\alpha}_{\mu}\mathbf{K}_{\nu \alpha}-\frac{1}{2}\mathbf{h}_{\mu \nu}\left(\mathbf{K}^{2}-\mathbf{K}_{\mu \nu}\mathbf{K}^{\mu \nu}\right)\right\rbrace +\mathbf{E}_{\mu \nu}-\kappa _{4}^{2}\mathbf{T}_{\mu \nu}^{(\phi)}
\nonumber
\\
&=G_{\mu \nu}-\left\lbrace KK_{\mu \nu}-K^{\alpha}_{\mu}K_{\nu \alpha}-\frac{1}{2}h_{\mu \nu}\left(K^{2}-K_{\mu \nu}K^{\mu \nu}\right)\right\rbrace +E_{\mu \nu}
\nonumber
\\
&-\frac{2}{3}\frac{\nabla _{\mu}\nabla _{\nu}f'}{f'}+h_{\mu \nu}\frac{f'R-f}{4f'}+h_{\mu \nu}h^{\alpha \beta}\frac{2\nabla _{\alpha}\nabla _{\beta}f'}{3f'}~,
\end{align}
which is precisely the effective gravitational field equations in the Jordan frame with the identification $T^{\rm brane}_{\mu \nu}=(1/f')\mathbf{T}_{\mu \nu}^{\rm brane}$. Hence the equivalence works at the level of effective field equations as well. However, the practical implementation of the above result again requires inversion of complicated functional forms and hence invites approximations.
\section{A Comparison with reconstruction methods in f(R) gravity}
In the above two sections we have shown the equivalence of gravitational field equations both in the bulk and in the brane respectively. In this section we will present a comparison of our method with an existing well known method in $f(R)$ gravity, the \emph{reconstruction method}. As already emphasized, due to presence of higher derivatives in the field equations for $f(R)$ gravity, obtaining a straightforward solution in a general case is very difficult. Even for systems with large number of symmetries, e.g., in cosmology which has a single unknown function $a(t)$, solving field equations \emph{directly} in the Jordan frame is very complicated. This lends its way to reconstruction method which we will briefly summarize \cite{Nojiri:2009kx,Nojiri:2006be,Nojiri:2009xh,Carloni:2010ph}.
In the reconstruction method one assumes that the expansion history of the universe is known exactly and by inverting the field equations one can determine what class of $f(R)$ theories can give rise to the observed universe. For example, power law solutions for the scale factor singles out $R^{n}$ to be the gravitational action. Since the scale factor $a(t)$ is known, the Ricci scalar is also known as $R(t)$. This can be inverted to get $t=g(R)$ and hence the Hubble parameter is known to be a function of Ricci scalar. This when used in the field equations, leads to a differential equation for $f(R)$, which can be solved to know the gravity model \cite{Nojiri:2009kx,Nojiri:2006be,Nojiri:2009xh}. There have been other variants of this model, e.g, assuming every physical quantity to be function of scale factor or function of Hubble parameter, which ultimately leads to a differential equation for the gravity model \cite{Carloni:2010ph}. The essential ingredients remain the same but one particular case may be
convenient in comparison to the other in a particular situation. Let us now explicitly point out the advantages
and disadvantages of reconstruction scheme as well as our approach.
\begin{itemize}
\item An important limitation of the reconstruction method is, only very simple cosmic histories, e.g. simple power law behaviours can be connected to $f(R)$ theory in an exact way. Our method, has similar disadvantages. Even though the field equations can be exactly solved in the Einstein frame for a few cases of interest, the inversion of the potential to $f(R)$ theory can be performed only in simple situations.
\item The reconstruction method is adapted to cosmological spacetimes only, since the cosmic history is known through experiments. However the situation we are interested in corresponds to, higher dimensional physics in presence of higher curvature gravity, possible behaviour of the warp factor and the brane separation. Since there is no experimental backdrop for extra dimensions it is not possible to come up with a physical ansatz. Thus one needs to solve the field equations at face value, which can be efficiently done using our method as we have illustrated in the next sections.
\item The utility of reconstruction scheme lies in its quick and straightforward analysis. Given a phenomenological scale factor $a(t)$ one needs to solve a single differential equation to get $f(R)$, given the inversion $t=t(R)$. While in our method one first need to solve the gravity plus scalar field system to get the solution in Einstein frame, which itself is a formidable task. Then one needs to invert the potential $V(\phi)$ to get $f(R)$ and finally the conformal transformation will yield the solution in Jordan frame.
\end{itemize}
Thus both the methods have their own advantages and disadvantages. The reconstruction method is very simple, applicable even in presence of singularity and useful in cosmological context, while not so useful when applied to other scenarios e.g. extra dimensions. On the other hand, the method introduced in this work even though requires regular spacetime region, involves more steps to arrive at the solution, is very robust. It will work for any regular spacetime region, from cosmological scenarios to extra dimensions as explained in later examples.
\section{Einstein to Jordan frame in the bulk: Explicit examples}\label{Sec_03}
We will now illustrate through simple examples how one might obtain solutions to bulk field equations in $f(R)$ gravity, which involve higher derivative terms, by exploiting the equivalence with scalar tensor theory depicted in the previous sections. As emphasized before, due to occurrence of higher derivative terms it is difficult to solve for the bulk equations of $f(R)$ gravity. On the other hand, solving a set of coupled equations of gravity plus scalar field system is much simpler. Hence through the equivalence shown earlier, if we can obtain a solution for the bulk metric $g_{ab}$ in the Einstein frame, corresponding solution in the Jordan frame will differ only by a conformal factor.
Before we jump into detailed calculations it is worthwhile to sketch the flowchart we are going to follow --- (a) We will start with the bulk action in the Einstein frame. (b) For some suitable potential, we will find out the metric describing bulk spacetime, by solving the bulk field equations. (c) We will match the potential in the Einstein frame with the corresponding $f(R)$ theory in the Jordan frame and finally (d) The conformal transformation will yield the corresponding bulk metric in the Jordan frame. In the examples to follow we will explicitly illustrate all the four steps mentioned above.
\paragraph*{Bulk field equations in Einstein frame} We start by solving the field equations of gravity and the scalar field in the Einstein frame. We assume that the branes are flat, viz., $\eta _{\mu \nu}$ is the spacetime metric on the brane. Further, the two branes are assumed to be separated by the stabilized value of the radion field $r_{c}$ \cite{Goldberger:1999uk,Chakraborty:2013ipa} such that the metric ansatz turns out to be (this ansatz is useful, particularly in the context of gauge hierarchy problem)
\begin{align}
d\mathbf{s}^{2}=e^{-2A(y)}\eta _{\mu \nu}dx^{\mu}dx^{\nu}+r_{c}^{2}dy^{2}~,
\end{align}
where $A(y)$ is the warp factor and is dependent on the extra spacetime coordinate alone. From the above metric ansatz the non-zero components of the Ricci tensor are immediate,
\begin{align}
\mathbf{R}_{\mu \nu}=\frac{e^{-2A}}{r_{c}^{2}}\left(A''-4A'^{2}\right)\eta _{\mu \nu};\qquad \mathbf{R}_{yy}=4A''-4A'^{2}~.
\end{align}
Note that we are following the previously mentioned convention: all the metric dependent quantities in the Einstein frame are boldfaced. From the components of the Ricci tensor, straightforward computation of the Ricci scalar leads to,
\begin{align}\label{Ricci}
\mathbf{R}=\frac{1}{r_{c}^{2}}\left(8A''-20A'^{2}\right)~.
\end{align}
Given the Ricci tensor and the Ricci scalar one can further compute the non-zero components of the Einstein tensor as,
\begin{align}
\mathbf{G}_{\mu \nu}=3\frac{e^{-2A}}{r_{c}^{2}}\eta _{\mu \nu}\left(-A''+2A'^{2}\right);\qquad \mathbf{G}_{yy}=6A'^{2}~,
\end{align}
such that the bulk gravitational field equations (see \ref{E_bulk}) in the Einstein frame become,
\begin{align}\label{bulk_E}
\mathbf{G}_{ab}=-\Lambda \mathbf{g}_{ab}+\kappa _{5}^{2}\left[\partial _{a}\phi \partial _{b}\phi -\mathbf{g}_{ab}\left\lbrace \frac{1}{2}\mathbf{g}^{cd}\partial _{c}\phi \partial _{d}\phi +V(\phi)\right\rbrace \right]~,
\end{align}
where, $\Lambda$ is the bulk cosmological constant. From \ref{Ricci} it is clear that bulk curvature depends only on the extra coordinate $y$, since $A$ depends on $y$ only. Thus logical consistency of the field equations demand that $\phi$ should also depend only on the extra dimensional coordinate $y$. In which case \ref{bulk_E} reduce to the following three coupled equations for gravity and scalar field as,
\begin{align}
-3A''+6A'^{2}&=-\Lambda r_{c}^{2}-\kappa _{5}^{2}\left[\frac{1}{2}\phi '^{2}+r_{c}^{2}V(\phi)\right]~,
\\
A'^{2}&=-\frac{\Lambda}{6}r_{c}^{2}+\frac{\kappa _{5}^{2}}{12}\phi '^{2}-\frac{r_{c}^{2}\kappa _{5}^{2}}{6}V(\phi)~,
\\
\phi ''-4A'\phi '&=r_{c}^{2}\frac{\partial V}{\partial \phi}~,
\end{align}
where `prime' denotes derivative with respect to $y$. Eliminating $A'$ from the first two equations we obtain,
\begin{align}\label{Aprime}
A''=\frac{\kappa _{5}^{2}}{3}\phi '^{2}~.
\end{align}
In general the solution to the above coupled equations can be obtained by introducing a super-potential $W(\phi)$ which satisfies the following differential equation
\begin{align}
\frac{1}{9}W^{2}=-\frac{\Lambda}{6}r_{c}^{2}+\frac{1}{12\kappa _{5}^{2}}\left(\frac{\partial W}{\partial \phi}\right)^{2}-\frac{r_{c}^{2}\kappa _{5}^{2}}{6}V(\phi)~.
\end{align}
The above differential equation for $W(\phi)$ can be inverted and the potential $V(\phi)$ gets determined in terms of $W(\phi)$ as,
\begin{align}
V(\phi)=-\frac{\Lambda}{\kappa _{5}^{2}}+\frac{1}{2\kappa _{5}^{4}r_{c}^{2}}\left(\frac{\partial W}{\partial \phi}\right)^{2}-\frac{2}{3\kappa _{5}^{2}r_{c}^{2}}W^{2}~.
\end{align}
In such a scenario the coupled equations become separable. One thus obtains separate differential equations for the metric function $A$ and the scalar field $\phi$ in the following form
\begin{align}\label{field_eq}
A'=\frac{1}{3}W;\qquad \phi '=\frac{1}{\kappa _{5}^{2}}\frac{\partial W}{\partial \phi}~.
\end{align}
On the other hand, if one postulates the separability of the coupled field equations, then also the expression of the potential $V(\phi)$ in terms of the super-potential $W(\phi)$ follows. This choice for $A$ and $\phi$ also satisfies the field equation for $\phi$ as well, as one can easily check. The only remaining one corresponds to, \ref{Aprime}, i.e., the $A''$ equation.
\paragraph*{Solutions in Einstein Frame} We have set the stage, it is now time to act. Having obtained the field equations in a sensible form, let us now solve for the bulk metric. In order to satisfy \ref{Aprime} one requires two possible choices for $W(\phi)$ --- (i) $W(\phi)=c\phi$, (ii) $W(\phi)=a-b\phi ^{2}$ for arbitrary choices of $a$, $b$ and $c$. Solving the field equations in both these cases separately leads to,
\begin{itemize}
\item The super-potential is linear in $\phi$, i.e., $W(\phi)=c\phi$. Then from \ref{field_eq} one obtains, $A'=(b/3)\phi$, such that $A''=(c/3)\phi '$, while $\phi '=c/\kappa _{5}^{2}$. This set identically satisfies \ref{Aprime}. The corresponding potential $V(\phi)$ turns out to be,
\begin{align}\label{V1}
V(\phi)=-\frac{\Lambda}{\kappa _{5}^{2}}+\frac{c^{2}}{2\kappa _{5}^{4}r_{c}^{2}}-\frac{2c^{2}}{3\kappa _{5}^{2}r_{c}^{2}}\phi ^{2}~,
\end{align}
with the following solution for $A(y)$ and $\phi (y)$ as,
\begin{align}
\phi (y)&=\phi _{0}+\frac{c}{\kappa _{5}^{2}}y~,
\\
A(y)&=A_{0}+\frac{c\phi _{0}}{3}y+\frac{c^{2}}{6\kappa _{5}^{2}}y^{2}~.
\end{align}
\item The super-potential is quadratic in $\phi$, i.e., $W(\phi)=a-b\phi ^{2}$. From \ref{field_eq} we get, $A'=(1/3)(a-b\phi ^{2})$, hence this yields $A''=-(2b/3)\phi \phi '$, with $\phi '=-(2b/\kappa _{5}^{2})\phi$. These expressions can be easily manipulated to show that \ref{Aprime} is indeed satisfied. Then the potential becomes,
\begin{align}\label{V2}
V(\phi)=\left(-\frac{\Lambda}{\kappa _{5}^{2}}-\frac{2a^{2}}{3r_{c}^{2}\kappa _{5}^{2}}\right)+\left(\frac{b^{2}}{2r_{c}^{2}\kappa _{5}^{4}}+\frac{4ab}{3r_{c}^{2}\kappa _{5}^{2}}\right)\phi ^{2}-\frac{2b^{2}}{3r_{c}^{2}\kappa _{5}^{2}}\phi ^{4}~,
\end{align}
with the following solutions for $A(y)$ and $\phi (y)$,
\begin{align}
\phi (y)&=\phi _{0}\exp \left(-\frac{b}{\kappa _{5}^{2}}y\right)~,
\\
A(y)&=A_{0}+\sqrt{-\frac{\Lambda r_{c}^{2}}{6}}y+\frac{\kappa _{5}^{2}}{6}\phi _{0}^{2}\exp \left(-2\frac{b}{\kappa _{5}^{2}}y\right)~.
\end{align}
\end{itemize}
Thus we have exactly solved the bulk gravitational field equations in the Einstein frame for two choices of the scalar field potential. One of them is quadratic, i.e., $V(\phi)=a+b\phi ^{2}$ while the other corresponds to quartic potential, $V(\phi)=a+b\phi ^{2}+c\phi ^{4}$. Now we need to execute the last two steps in our flowchart, namely, (a) First one should identify a $f(R)$ model, which gives rise to the potentials obtained above and (b) Secondly, one needs to find the conformal factor relating the two frames and hence the metric can be obtained in the Jordan frame.
\paragraph*{Connecting Jordan and Einstein frame} So far we have been working solely in the Einstein frame. In order to obtain the respective solution in the Jordan frame, we have to connect the scalar field to a $f(R)$ model, which can be done through the relations
\begin{align}\label{Neweq}
\kappa _{5}\phi =\frac{2}{\sqrt{3}}\ln f'(R);\qquad \kappa _{5}^{2}V=\frac{f'R-f}{2f'^{5/3}}~.
\end{align}
These relations follow from the original connection between Einstein and Jordan frame discussed in \ref{conditions_01} and \ref{conditions_02} respectively. In principle one should start with the quadratic and quartic potentials obtained previously and hence obtain the corresponding $f(R)$ theory in the Jordan frame by using \ref{Neweq}. However we will take the opposite route, i.e., we will start with some $f(R)$ model and arrive at the respective potentials in the Einstein frame using \ref{Neweq} and map it to those obtained earlier.
\begin{itemize}
\item The simplest model is always the best to start with. For $f(R)$ gravity this corresponds to a situation, where the Einstein-Hilbert term, receives a quadratic correction \footnote{The quadratic correction is well known in the literature, see for example \cite{Liu:2011wi,Barrow:2006xb,Barrow:2009gx,Barrow:1988xi,Maeda:1987xf}.}, i.e., $f(R)=R+\alpha R^{2}$. For this particular model of $f(R)$ gravity, the scalar field and the potential can be written in terms of the Ricci scalar in the Jordan frame as (using \ref{Neweq}),
\begin{align}
\kappa _{5}\phi &=\frac{2}{\sqrt{3}}\ln \left(1+2\alpha R\right);\qquad R=\frac{1}{2\alpha}\left[\exp\left({\frac{\sqrt{3}}{2}\kappa _{5}\phi}\right)-1\right]~,
\\
V(\phi)&=\frac{1}{8\alpha \kappa _{5}^{2}}\left[\exp \left(\frac{\kappa _{5}\phi}{2\sqrt{3}}\right)-2\exp \left(-\frac{\kappa _{5}\phi}{\sqrt{3}} \right)+\exp \left(-\frac{5}{2\sqrt{3}}\kappa _{5}\phi \right)\right]~.
\end{align}
The minima of the potential corresponds to $\partial V/\partial \phi =0$, with $\partial ^{2}V/\partial \phi ^{2}>0$. Both these conditions can be satisfied provided $e^{\kappa _{5}\phi}=1$, or $\phi =0$. Finally, expanding around the minima one obtains the following form for the potential,
\begin{align}\label{vphi_01}
V(\phi)&=\frac{1}{8\alpha \kappa _{5}^{2}}\left[\left(1+\frac{\kappa _{5}\phi}{2\sqrt{3}}+\frac{1}{2}\frac{\kappa _{5}^{2}\phi ^{2}}{12} \right)-2\left(1-\frac{\kappa _{5}\phi}{\sqrt{3}}+\frac{1}{6}\kappa _{5}^{2}\phi ^{2}\right)+\left(1-\frac{5}{2\sqrt{3}}\kappa _{5}\phi +\frac{1}{2}\frac{25\kappa _{5}^{2}\phi ^{2}}{12}\right)\right]
\nonumber
\\
&=\frac{3}{32\alpha}\phi ^{2}~.
\end{align}
Thus we have a quadratic potential for $\phi$, which originate from $R+\alpha R^{2}$ gravity. Matching the potential to that derived in the Einstein frame, given by \ref{V1}, we obtain the following relation: $\Lambda =-9/(128\alpha)$.
\item Let us now consider a more general $f(R)$ gravity model for which $f(R)=R+\alpha R^{2}+\beta R^{4}$, where $\alpha$ and $\beta$ are dimension full constant coefficients, with values such that the model becomes ghost free. Then from \ref{Neweq} we obtain, $R=(\sqrt{3}\kappa _{5}\phi)/(4\alpha)$, such that the potential turns out to be,
\begin{align}
V(\phi)=\frac{\alpha R^{2}+3\beta R^{4}}{2\kappa _{5}^{2}}=\frac{3}{32\alpha}\phi ^{2}+\frac{1}{2}\frac{3^{3}\beta \kappa _{5}^{2}}{4^{4}\alpha ^{4}}\phi ^{4}~,
\end{align}
where we have assumed $\alpha \gg \beta \gg \alpha ^{2}$, consistent with the ghost free criteria for this $f(R)$ model. Comparing this with the potential obtained by solving bulk field equations in the Einstein frame, presented in \ref{V2} we immediately obtain,
\begin{align}
a=\sqrt{-\frac{3r_{c}^{2}\Lambda}{2}};\qquad b=-\frac{4a\kappa _{5}^{2}}{3}\pm \sqrt{\frac{16a^{2}\kappa _{5}^{4}}{9}+\frac{3r_{c}^{2}\kappa _{5}^{4}}{16\alpha}};\qquad \beta =-\frac{4^{5}b^{2}\alpha^{4}}{3^{4}r_{c}^{2}\kappa _{5}^{4}}~.
\end{align}
\end{itemize}
This completes the connection between Einstein and Jordan frames. The potentials obtained in the Einstein frame get mapped to respective $f(R)$ theories.
\paragraph*{Solutions in Jordan Frame} We have now reached the final step of our flowchart, viz., solution to the bulk field equations in the Jordan frame. For this purpose we can use the connection with Einstein frame derived earlier.
\begin{itemize}
\item For $f(R)=R+\alpha R^{2}$, the corresponding scalar field potential in the Einstein frame is quadratic with the mapping being given by \ref{vphi_01}. From which the conformal factor turns out to be, $\Omega=(1+2\alpha R)^{1/3}=[1+(\sqrt{3}\kappa _{5}\phi/2)]^{1/3}$. Hence the bulk solution in the Jordan frame corresponds to,
\begin{align}\label{fr1}
ds^{2}&=\left[1+\frac{\sqrt{3}\kappa _{5}\phi(y)}{2}\right]^{-2/3}\left\lbrace e^{-2A(y)}\eta _{\mu \nu}dx^{\mu}dx^{\nu}+r_{c}^{2}dy^{2} \right\rbrace ~,
\\
\phi (y)&=\phi _{0}+\frac{c}{\kappa _{5}^{2}}y;\qquad A(y)=A_{0}+\frac{c\phi _{0}}{3}y+\frac{c^{2}}{6\kappa _{5}^{2}}y^{2}~,
\nonumber
\end{align}
where $c$ and $\phi_{0}$ are arbitrary constants of integration. Thus for the quadratic $f(R)$ model under consideration, one can map it to the Einstein frame and obtain the respective potential. For this particular case, the field equations in the Einstein frame becomes exactly solvable and hence by conformal transformation one can obtain the corresponding solution in the Jordan frame. Further, from \ref{fr1} it turns out that the warp factor is governed by the factors $c\phi _{0}$ and $c/\kappa _{5}$. Hence in order to have proper suppression of the Planck scale on the visible brane one must have the conditions $c<\kappa _{5}$ and $c\phi _{0}\sim 36$. Hence one arrives at, $\phi_{0}>\kappa _{5}^{-1}$. Further, in this model the radion field varies with extra dimension $y$ as $(a+by)^{2/3}$, where $a$ and $b$ depends on $c,\kappa _{5}$ and $\phi _{0}$.
\item For the other model, i.e., $f(R)=R+\alpha R^{2}+\beta R^{4}$, the scalar field potential in the Einstein frame is quartic. From which one can relate the parameters $\alpha,\beta$ with the respective ones in the Einstein frame. Following the same strategy as above, the conformal factor turns out to yield, $\Omega =(1+2\alpha R+4\beta R^{3})^{1/3}=[1+(\sqrt{3}\kappa _{5}\phi/2)+(3\sqrt{3}\beta \kappa _{5}^{3}\phi ^{3}/16\alpha ^{3})]^{1/3}$. Using which the solution in the Jordan frame becomes,
\begin{align}\label{fr2}
ds^{2}&=\left[1+\frac{\sqrt{3}\kappa _{5}\phi(y)}{2}+\frac{3\sqrt{3}\beta \kappa _{5}^{3}\phi(y)^{3}}{16\alpha ^{3}}\right]^{-2/3}
\left\lbrace e^{-2A(y)}\eta _{\mu \nu}dx^{\mu}dx^{\nu}+r_{c}^{2}dy^{2} \right\rbrace ~,
\\
\phi (y)&=\phi _{0}\exp \left(-\frac{b}{\kappa _{5}^{2}}y\right);\qquad A(y)=A_{0}+\sqrt{-\frac{\Lambda r_{c}^{2}}{6}}y+\frac{\kappa _{5}^{2}}{6}\phi _{0}^{2}\exp \left(-2\frac{b}{\kappa _{5}^{2}}y\right)~,
\nonumber
\end{align}
This demonstrates another $f(R)$ theory for which the corresponding potential in the Einstein frame leads to an exact solution. Using which and the mapping between Einstein and Jordan frame one obtains the respective solution in the Jordan frame. However in contrast with the previous situation, in this case the warp factor behaves exactly like the Randall-Sundrum scenario, since all the corrections are exponentially suppressed (see \ref{fr2}). While the radion field is almost constant due to identical exponential suppression. Hence the $f(R)$ model with quartic correction is more favored in the extra dimensional physics than the earlier one.
\end{itemize}
\paragraph*{Aside: Comment on domain of applicability} After illustrating two examples on how to obtain solutions to higher order field equations, by using the Einstein frame judiciously, we would like to comment on the domain of applicability of this approach.
We should emphasize that we are working in a mesoscopic energy scale, i.e., the energy scale is larger compared to general relativity, such that effect of higher order terms, e.g, $\alpha R^{2}$ cannot be ignored. On the other hand, the energy scale is much smaller compared to the Planck scale so that the additional contributions are still sub-dominant, i.e., $\alpha R<1$. This is important, in particular when one obtains the scalar field in the Einstein frame in terms of the curvature. In order to obtain closed form expression one has to expand the potential near its minimum, this in turn requires one to neglect higher order curvature corrections, e.g., one might neglect $\alpha ^{2}R^{2}$ in comparison with $\alpha R$. In a nutshell, we are working in a high curvature regime such that effect of $f(R)$ gravity can be felt, but not high enough so that the Einstein-Hilbert action becomes sub-dominant.
Another point that requires clarification are the approximations involved in general scenarios. The conformal transformation and hence the conversion of a potential to a corresponding $f(R)$ model is not at all straightforward. In most of the cases the relations turn out to be non-invertible, and one need to resort to approximations. As explained earlier, on physical grounds, one can assume that the higher order terms are sub-leading and hence one can keep only linear order terms. While dealing with complicated potentials most often one needs to resort to these approximations, justified by physical intuitions. However at the Planck scale these approximations brake down, since the assumption that higher orders terms are sub-leading cannot be trusted.
Having discussed two possible scenarios in the context of bulk physics let us now consider brane dynamics. In particular, we will be interested in one spherically symmetric and one cosmological applications.
\section{Jordan to Einstein frame in the brane: Explicit examples}\label{Sec_04}
Both the examples depicted above are related to bulk spacetimes. To complete the discussion we will also derive the metric in the Einstein frame starting from the Jordan frame, but in the brane spacetime. This is to explicitly demonstrate that the technique works both ways --- whenever it is convenient to solve in the Einstein frame, we can solve it and transform back to the Jordan frame, while if the solution is simpler in Jordan frame it can give insight into what happens in scalar coupled gravity, viz., the Einstein frame.
\begin{itemize}
\item In this example, we will start with a particular $f(R)$ model on the brane, solve the effective field equations and obtain a cosmological solution. Then using conformal transformation the corresponding solution in the Einstein frame can be obtained.
\paragraph*{Solution in the Jordan frame} Let us start with the $f(R)$ model given by $f(R)=f_{0}(R-R_{0})^{\alpha}$, where $f_{0}$ and $R_{0}$ are constants and $\alpha \neq 1$. The corresponding solution for the scale factor on the brane can be obtained by solving the effective field equations derived in \cite{Chakraborty:2014xla,Haghani:2012zq}. This leads to power law solution $a(t)\sim t^{n}$, where $n$ is related to $\alpha$ and the matter fields present on the brane.
\paragraph*{Converting back to Einstein frame} We need to convert it back to Einstein frame and hence obtain the corresponding solution in the scalar coupled gravity. For this choice for $f(R)$, we obtain, the scalar field to be,
\begin{align}
\kappa _{5}\phi =\frac{2}{\sqrt{3}}\ln \left(f_{0}\alpha \right)+\frac{2}{\sqrt{3}}\left(\alpha -1\right)\ln \left(R-R_{0}\right)~,
\end{align}
which in turn can be inverted, leading to, $R-R_{0}=\exp[a\kappa _{5}(\phi -\phi_{0})]$, where, $a=\sqrt{3}/(2(\alpha -1))$ and $\kappa _{5}\phi _{0}=(2/\sqrt{3})\ln (f_{0}\alpha)$. Then the potential can be determined readily, using \ref{conditions_02}, leading to,
\begin{align}
V(\phi)=\frac{\alpha -1}{2\kappa _{5}^{2}f_{0}^{2/3}\alpha ^{5/3}}\exp\left[\kappa _{5}a\left(\frac{5}{3}-\frac{2\alpha}{3}\right)\left(\phi -\phi _{0}\right) \right]~.
\end{align}
Hence the power law behavior of the $f(R)$ theory transforms back to exponential potential.
\paragraph*{Solution in Einstein frame} The corresponding cosmological solution in the Einstein frame could be obtained by transforming the Jordan frame solution using the conformal factor. In this particular class of $f(R)$ model, the conformal factor becomes, $\Omega =(f_{0}\alpha)^{1/3}(R-R_{0})^{(\alpha -1)/3}\sim t^{-2(\alpha -1)/3}$. Thus the corresponding solution in the Einstein frame is again cosmological with a new scale factor: $\hat{a}(t)\sim t^{[3n-4(\alpha -1)]/3}$. Hence the cosmological solution in the Einstein frame with exponential potential is still power law.
\item So far we have been dealing with power law $f(R)$ theories. However to show the applicability of our method to more general scenarios, we will consider the following $f(R)$ model: $R-(\alpha/R)$ on the brane. This $f(R)$ model in four dimensional spacetime has been discussed in detail in \cite{Nojiri:2003ft}, however no such solution in the context of effective field equations exist, which by itself would be an interesting future work. However in this work we will contend ourselves in providing basic ingredients regarding this model. Further given a solution in Jordan frame, use of conformal transformation will lead to the corresponding solution in the Einstein frame.
\paragraph*{Solution in the Jordan frame} Let us start with the above mentioned $f(R)$ model. The corresponding solution for the scale factor on the brane can be obtained by solving the effective field equations and can be taken to be $a(t)\sim t^{n}$, where $n$ should be related to $\alpha$ and the matter fields present on the brane.
\paragraph*{Converting back to Einstein frame} We need to convert it back to Einstein frame and hence obtain the corresponding solution in the scalar coupled gravity. For this choice for $f(R)$, we obtain, the scalar field to be,
\begin{align}
\kappa _{5}\phi =\frac{2}{\sqrt{3}}\ln \left(1+\frac{\alpha}{R^{2}}\right)~,
\end{align}
which in turn can be inverted, leading to, $R=\sqrt{\alpha}(\exp[(2/\sqrt{3})\kappa _{5}\phi]-1)^{-1/2}$. Then the potential can be determined readily, using \ref{conditions_02}, leading to,
\begin{align}
V(\phi)=\sqrt{\alpha}\exp\left(-\frac{5}{2\sqrt{3}}\kappa _{5}\phi\right)\sqrt{\exp[(2/\sqrt{3})\kappa _{5}\phi]-1}~.
\end{align}
Expanding for small $\phi$, we obtain, $V(\phi)=\sqrt{\alpha \kappa _{5}\sqrt{3}\phi/2}~(1-(5\kappa _{5}\phi/2\sqrt{3}))$.
Hence negative power law behavior of the $f(R)$ theory transforms back to potential with $\sqrt{\phi}$ as the leading order contribution.
\paragraph*{Solution in Einstein frame} The corresponding cosmological solution in the Einstein frame could be obtained by transforming the Jordan frame solution using the conformal factor. In this particular class of $f(R)$ model, the conformal factor becomes, $\Omega =[1+(\alpha/R^{2})]^{1/3}\sim [1+\alpha t^{4}]^{1/3}$, since $R\sim t^{-2}$. Thus for late times, $\Omega \sim t^{4/3}$. Thus the corresponding late time solution in the Einstein frame is again cosmological with a new scale factor: $\hat{a}(t)\sim t^{n+(8/3)}$, again a power law behavior. Hence the cosmological solution in the Einstein frame with $\sqrt{\phi}$ potential is still a power law.
\item Another explicit spherically symmetric solution on the brane in the Jordan frame has been constructed in \cite{Chakraborty:2014xla} by decomposing the Electric part of the Weyl tensor into dark radiation term $U(r)$ and dark pressure term $P(r)$.
\paragraph*{Solution in Jordan frame} The dark radiation $U(r)$ and dark pressure $P(r)$ acts as auxiliary source to the effective gravitational field equations \cite{Maartens:2001jx}. A possible solution can be obtained when an ``equation of state'' between $U(r)$ and $P(r)$ is specified. For the particular choice $2U(r)+P(r)=0$, we immediately obtain, the corresponding
spherically symmetric solution \cite{Chakraborty:2014xla},
\begin{align}\label{frmetric}
ds^{2}=-f(r)dt^{2}+\frac{dr^{2}}{f(r)}+r^{2}d\Omega ^{2};\qquad f(r)=1-\frac{2GM+Q_{0}}{r}-\frac{3\bar{\kappa}P_{0}}{2r^{2}}+\frac{F(R)-\Lambda _{4}}{3}r^{2}~.
\end{align}
Here $Q_{0}$ and $P_{0}$ corresponds to constants of integration, $\bar{\kappa}$ captures the effect of bulk spacetime, i.e., depends on bulk gravitational constant and $F(R)$, evaluated at the brane location can be constructed from the original $f(R)$ theory by taking appropriate derivative. Assuming that the bulk scalar depends only on the bulk coordinates and for a $f(R)$ theory of the form $f(R)=R+\alpha R^{2}+\beta R^{4}$, the leading order behavior of $F(R)$ is like an effective four dimensional cosmological constant $\Lambda _{4}$.
\paragraph*{Converting back to Einstein frame} The corresponding scalar-tensor solution can be obtained by transforming the metric in \ref{frmetric} using the appropriate conformal factor: $\Omega =(1+2\alpha R+4\beta R^{3})^{1/3}=[1+(3\kappa _{5}\phi/2)+(27\beta \kappa _{5}^{3}\phi ^{3}/64\alpha ^{3})]^{1/3}$. Since the scalar field depends on the extra coordinate only, the conformal factor evaluated on the location of the brane is just a constant.
\paragraph*{Solution in Einstein frame} Since the corresponding conformal factor is just a constant it will scale the metric, which can be absorbed by rescaling of time and radial coordinate by the conformal factor. Hence the solution in the Einstein frame would remain the same.
\end{itemize}
Thus even in the context of brane models solving effective gravitational field equations in one frame and obtaining the solution in other often requires approximations. One has to keep in mind that we are working in the mesoscopic scale, where neither the higher curvature terms are dominant nor are they negligible. This allows one to invert various relations connecting the Einstein frame scalar with Jordan frame curvature, a key aspect while converting solutions in one frame to another.
\section{Conclusions}
A technique for solving field equations of higher curvature gravity theories have been proposed. The technique essentially hinges on the mathematical equivalence of higher curvature gravity theory, e.g., $f(R)$ theories of gravity with scalar-tensor representation. Earlier this equivalence was known only at the level of action principle. It was not clear a priori whether the field equations derived from either the $f(R)$ representation or the scalar-tensor representation would also be equivalent. In this work starting from a five dimensional theory we have explicitly demonstrated --- (a) the bulk gravitational field equations derived from Jordan and Einstein frame are equivalent and (b) the effective field equations on the brane in these two approaches are also equivalent. Using this equivalence we have argued, if one can solve for the field equations in one frame, the solution in the other frame can be easily obtained. Even though for simple models one can perform the above operation exactly, often it
requires suitable approximations. The approximations essentially requires one to work in mesoscopic energy scales, viz., higher than weak scale but less than Planck scale. For practical application of the technique, we have illustrated it in two related situations:
\begin{itemize}
\item The bulk field equation in the Einstein frame have been solved in the context of warped geometry models for two choices of the potential --- quadratic and quartic. Following expectation, the warp factor behaves differently in these two scenarios, but leads to desired exponential warping. These potential through conformal transformations are related to two $f(R)$ models --- (a) $R +\alpha R^{2}$ and (b) $R+\alpha R^{2}+\beta R^{4}$ respectively. From the solution in the Einstein frame we have obtained the solution in $f(R)$ representation as well, having a different warp factor behavior and extra dimension dependent radion field.
\item Secondly, using the known solutions to effective field equations in the $f(R)$ representations we have obtained the corresponding solutions in the scalar-tensor representation. In the cosmological context, the scale factor still exhibit power law behavior, but with a different power. While the spherically symmetric solution results in mere rescaling of the coordinates.
\end{itemize}
In the two examples depicted in this work we have explored practical illustration of the technique for both the bulk and brane spacetimes. Further it turns out that even though in the Einstein frame the brane separation has been fixed at the stabilized value, in the Jordan frame it starts depending on the extra dimension. Interestingly, the warp factor in the two frames are different, leading to different suppression of the Planck scale on the visible brane. This might lead to potential observables, distinguishing the two frames in the context of recent LHC experiments.
\section*{Acknowledgements}
Research of S.C. is funded by a SPM fellowship from CSIR, Government of India. He also thanks IACS, India for warm hospitality; a part of this work was completed there during a visit.
|
\section{Acknowledgement}\label{sec:ackno}
This research is based upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA),via IARPA R\&D Contract No. 2014-14071600012. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon.
\section{Related Work}\label{SOA}
In the past few years, there have been numerous works in using deep features for tasks related to face verification. The DeepFace \cite{deepface14} approach uses a carefully crafted 3D alignment procedure to preprocess face images and feeds them to a deep network that is trained using a large training set. More recently, Facenet \cite{facenet15} uses a large private dataset to train several deep network models using a triplet distance loss function. The training time for this network is of the order of few weeks. Since the release of the IJB-A dataset \cite{ijba15}, there have been several works that have published verification results for this dataset. Previous approaches presented in \cite{wang15} and \cite{parkhi15} train deep networks using the CASIA-WebFace dataset \cite{casia14} and the VGG-Face dataset respectively, requiring substantial training time. This paper proposes a network architecture and a training scheme that needs shorter training time and a small query time. \\
The idea of learning a compact and discriminative representation has been around for decades. Weinberger \emph{et al.} \cite{lmnn05} used a Semi Definite Programming (SDP)-based formulation to learn a metric satisfying pairwise and triplet distance constraints in a large margin framework. More recently, this idea has been successfully applied to face verification by integrating the loss function within the deep network architecture (\cite{facenet15}, \cite{parkhi15}). Joint Bayesian metric learning is also another popular metric used for face verification (\cite{fvf13},\cite{chen15wacv}). These methods either require a large dataset for convergence or learn a metric directly and therefore are not amenable to subsequent operations like discriminative clustering or hashing. Classic methods like t-SNE \cite{tsne}, t-STE \cite{tste} and Crowd Kernel Learning (CKL) \cite{CKL} perform extremely well when used to visualize or cluster a given data collection. They either operate on the data matrix directly or the distance matrix generated from data by generating a large set of pairwise or triplet constraints. While these methods perform very well on a given set of data points, they do not generalize to out-of-sample data. In the current work, we aim to generalize such formulations, to a more traditional classification setting, where domain specific training and testing data is provided. We formulate an optimization problem based on triplet probabilities that performs dimensionality reduction aside from improving the discriminative ability of the test data. The embedding scheme described in this work is a more general framework that can be applied to any setting where labeled training data is available.
\section{Clustering Faces}\label{sec:cluster}
\begin{figure*}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.6\linewidth]{image_clusters/pose_blur/1.png}
\caption{}
\label{fig:sfig1}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.6\linewidth]{image_clusters/pose_blur/2.png}
\caption{}
\label{fig:sfig2}
\end{subfigure}\\
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.6\linewidth]{image_clusters/age/1.png}
\caption{}
\label{fig:sfig3}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.6\linewidth]{image_clusters/age/2.png}
\caption{}
\label{fig:sfig4}
\end{subfigure} \\
\caption{Sample clusters output from the Clustering approach discussed in Section 6 for the data from the split 1 of the IJB-A dataset. Top row (a,b) shows robustness to pose and blur; Bottom row (c,d) contains clusters that are robust to age}
\label{fig:cluster}
\end{figure*}
This section illustrates how the proposed TPE method can be used to cluster a
given data collection. We perform two clustering experiments:
\begin{enumerate}[leftmargin=*]
\item We perform clustering on the entire LFW \cite{lfw} dataset that consists
of 13233 images of 5749 subjects. It should be noted that about 4169 subjects have only one image.
\item We use the IJB-A dataset and cluster the templates corresponding to the query set for each split in the IJB-A verify protocol.
\end{enumerate}
For evaluating the clustering results, we use the metrics defined in
\cite{msu_cluster}. These are summarized below:
\begin{itemize}[leftmargin=*]
\item \textit{Pairwise Precision ($P_{pair}$)}: The fraction of pairs of samples
within a cluster among all possible pairs which are of the same class, over
the total number of same cluster pairs.
\item \textit{Pairwise Recall ($R_{pair}$)}: The fraction of pairs of samples
within a class among all possible pairs which are placed in the same cluster,
over the total number of same-class pairs.
\end{itemize}
Using these metrics, the F$_1$-score is computed as:
\begin{equation}
F_1 = \frac{2*P_{pair}*R_{pair}}{R_{pair}+P_{pair}}
\end{equation}
The simplest way we found to demonstrate the effectiveness of our deep features
and the proposed TPE method, is to use the standard MATLAB implementation of the
agglomerative clustering algorithm with the average linkage metric. We use the
cosine similarity as our basic clustering metric. The simple clustering
algorithm that we have used here has computational complexity of
$O(N^2)$. In its current form, this does not scale to large datasets with
millions of images. We are currently working on a more efficient and scalable (yet approximate) version of this algorithm.
\paragraph{Clustering LFW:-}The images in the LFW dataset are pre-processed as described in Section 5.1. For each image and its flip, the deep features are extracted using the proposed architecture, averaged and normalized to unit $L_2$ norm. We run the clustering algorithm over the entire data in a single shot. The clustering algorithm takes as input a cut-off parameter which acts as a distance threshold (below which any two clusters will not be merged). In our experiments, we vary this cut-off parameter over a small range and evaluate the
resulting clustering using the $F_1$-score. We pick the result that yields the
best $F_1$-score. Table \ref{tab:lfw_cluster} shows the result of our approach
and compares it to a recently released clustering approach based on approximate
Rank-order clustering \cite{msu_cluster}. It should be noted that, in the case
of \cite{msu_cluster}, the clustering result is chosen by varying the number of
clusters and picking the one with the best $F_1$-score. In our approach, we vary
the cut-off threshold which is the property of deep features and hence is a
more intuitive parameter to tune. We see from Table \ref{tab:lfw_cluster} that
aside from better performance, our total cluster estimate is closer to the
ground truth value of 5749 than \cite{msu_cluster}.
\begin{table}
\centering
\begin{tabular}{ | l | l | l|}
\hline
Method & $F_1$-score & Clusters \\ \hline
\cite{msu_cluster} & 0.87 & 6508 \\ \hline
CNN (Ours) & \textbf{0.955} & 5351\\ \hline
\end{tabular}
\captionof{table}{$F_1$-score for comparison of the two clustering schemes on
the LFW dataset. The ground truth cluster number is 5749.}
\label{tab:lfw_cluster}
\end{table}
\begin{table}
\centering
\resizebox{0.45\textwidth}{!}{
\begin{tabular}{ | l | l | l| l |}
\hline
Method & $F_1$-score & Clusters & After Pruning \\ \hline
CNN$_{media}$ & 0.79 (0.02) & 293 (22) & 173 \\ \hline
CNN$_{media}$+TPE & \textbf{0.843 (0.03)} & 258 (17) & 167 \\ \hline
\end{tabular}
}
\captionof{table}{Clustering metrics over the IJB-A 1:1 protocol. The
standard deviation is indicated in brackets. The ground truth subjects per
each split is 167.}
\label{tab:ijba_cluster}
\end{table}
\paragraph{Clustering IJB-A:-}The IJB-A dataset is processed as described in Section 5. In this section, we aim to cluster the query templates provided with each split for the verify protocol. We report the results of two experiments: with the raw CNN features (CNN$_{media}$ in Table 2) and with the projected CNN features, where the projection matrix is learned through the proposed TPE method (CNN$_{media}$+TPE in Table 2). The cut-off threshold required for our clustering algorithm is learned automatically based on the training data, i.e. we choose the threshold that gives the maximum $F_1$-score over the training data. The scores reported in Table \ref{tab:ijba_cluster} are average values over ten splits. As expected, the TPE method improves the clustering performance of raw features. The subject estimate is the number of clusters produced as a direct result of our clustering algorithm. The pruned estimate is obtained by ignoring clusters that have fewer than 3 images.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth,height=0.3\textwidth]{raw_tpe_cluster.png}
\caption{Precision-Recall curve plotted over cut-off threshold varied from 0 to 1.}
\label{fig:pr_curve}
\end{figure}
For a more complete evaluation of our performance over varying threshold values, we plot the Precision-Recall (PR) curve for the IJB-A clustering experiment in Figure \ref{fig:pr_curve}. As can be observed, the PR curve for clustering the IJB-A data using embedded features exhibits a better performance at all operating points. This is a more transparent evaluation than reporting only the $F_1$-score since the latter effectively fixes the operating point but the PR curve reveals the performance at all operating points.
\section{Conclusion and Future Work}\label{conclusion}
In this paper, we proposed a deep CNN-based approach coupled with a
low-dimensional discriminative embedding learned using triplet probability
constraints in a large margin fashion. The proposed pipeline enables a faster
training time and improves face verification performance especially at low
FMRs. We demonstrated the effectiveness of the proposed method on two
challenging datasets: IJB-A and CFP and achieved performance close to
the state of the art while using a deep model which is more compact and trained
using a moderately sized dataset. We demonstrated the robustness of our features
using a simple clustering algorithm on the LFW and IJB-A datasets. For future
work, we plan to use videos directly during training and also embed our TPE
approach into training the deep network. We intend to scale our clustering
algorithm to handle large scale scenarios such as large impostor sets of the
order of millions.
\section{Introduction}
\label{sec:intro}
Recently, with the advent of curated face datasets like Labeled faces in the
Wild (LFW) \cite{lfw} and advances in learning algorithms like Deep neural nets,
there is more hope that the unconstrained face verification problem can be
solved. A face verification algorithm compares two given templates that are
typically not seen during training. Research in face verification has progressed
well over the past few years, resulting in the saturation of performance on the
LFW dataset, yet the problem of unconstrained face verification remains a
challenge. This is evident by the performance of traditional algorithms on the
publicly available IJB-A dataset (\cite{ijba15}, \cite{fvff15}) that was
released recently. Moreover, despite the superb performance of CNN-based
approaches compared to traditional methods, a drawback of such methods is the
long training time needed. In this work, we present a Deep CNN (DCNN)
architecture that ensures faster training, and investigate how much the
performance can be improved if we are provided domain specific data.
Specifically, our contributions are as follows:
\begin{itemize}
\item We propose a deep network architecture and a training scheme that ensures
faster training time.
\item We formulate a triplet probability embedding learning method to improve
the performance of deep features for face verification and subject clustering.
\end{itemize}
During training, we use a publicly available face dataset to train our deep
architecture. Each image is pre-processed and aligned to a canonical view before
passing it to the deep network whose features are used to represent the image. In the case of IJB-A dataset, the data is divided into 10 splits, each split containing a training set and a test set. Hence, to further improve performance, we learn the proposed triplet probability embedding using the training set provided with each split over the features extracted from our DCNN model. During the deployment phase, given a face template, we extract the deep features using the raw CNN model after implementing automatic pre-processing steps such as face detection and fiducial extraction. The deep features are projected onto
a low-dimensional space using the embedding matrix learned during training (note
that the projection involves only matrix multiplication). We use the
128-dimensional feature as the final representation of the given face template.
This paper is organized as follows: Section \ref{SOA} places our work among the
recently proposed approaches for face verification. Section \ref{net} details
the network architecture and the training scheme. The triplet probabilistic
embedding learning method is described in Section \ref{tpe} followed by results
on IJB-A and CFP datasets and a brief discussion in Section
\ref{sec:results}. In Section \ref{sec:cluster}, we demonstrate the ability of
the proposed method to cluster a media collection from LFW and IJB-A
datasets.
\section{Network Architecture}\label{net}
This section details the architecture and training algorithm for the deep
network used in our work. Our architecture consists of 7 convolutional layers
with varying kernel sizes. The initial layers have a larger size rapidly
subsampling the image and reducing the parameters while subsequent layers consist
of small filter sizes, which has proved to be very useful in face recognition
tasks (\cite{parkhi15},\cite{casia14}). Furthermore, we use the Parametric
Rectifier Linear units (PReLUs) instead of ReLUs, since they allow a negative
value for the output based on a learned threshold and have been shown to improve
the convergence rate \cite{prelu15}.
\begin{Table}
\centering
\resizebox{6cm}{2.2cm}{
\begin{tabular}{|c|c|c|}
\hline
Layer & Kernel Size/Stride & \#params\\
\hline
conv1 &11x11/4 & 35K\\
pool1 &3x3/2 & \\
conv2 &5x5/2 & 614K\\
pool2 &3x3/2 & \\
conv3 &3x3/1 & 885K\\
conv4 &3x3/1 & 1.3M\\
conv5 &3x3/1 & 2.3M\\
conv6 &3x3/1 & 2.3M\\
conv7 &3x3/1 & 2.3M\\
pool7 &3x3/2 & \\
fc6 &1024 & 18.8M \\
fc7 &512 & 524K \\
fc8 &10548 & 10.8M \\
Softmax Loss & & Total: 39.8M\\
\hline
\end{tabular}
}
\captionof{table}{Deep Network architecture details}
\label{arch}
\end{Table}
The top three convolutional layers (conv1-conv3) are initialized with the
weights from the AlexNet model \cite{alexnet12} trained on the ImageNet
challenge dataset. Several recent works (\cite{transfer1},\cite{transfer2}) have
empirically shown that this transfer of knowledge across different networks,
albeit for a different objective, improves performance and more significantly
reduces the need to train over a large number of iterations.
The compared methods either learn their deep models from scratch
(\cite{parkhi15},\cite{nan}) or finetune only the last layer of fully
pre-trained models. The former results in large training time and the
latter does not generalize well to the task at hand (face verification) and
hence resulting in sub optimal performance. In the current work, even though we
use a pre-trained model (AlexNet) to initialize the proposed deep network, we do
so only for the first three convolutional layers, since they retain more generic
information (\cite{transfer1}). Subsequent layers learn representations which
are more specific to the task at hand. Thus, to learn more task specific
information, we add 4 convolutional layers each consisting of 512 kernels of
size $3\times3$. The layers conv4-conv7 do not downsample the input thereby
learning more complex higher dimensional representations. This hybrid
architecture proves to be extremely effective as our raw CNN representation
outperforms some very deep CNN models on the IJB-A dataset (Table 2 in Results).
In addition, we achieve that performance by training the proposed deep network using the
relatively smaller CASIA-WebFace dataset.
The architecture of our network is shown in Table \ref{arch}. Layers
conv4-conv7 and the fully connected layers \textit{fc6-fc8} are initialized from
scratch using random Gaussian distributions. PReLU activation functions are added
between each layer. Since the network is used as a feature extractor, the last
layer \textit{fc8} is removed during deployment, thus reducing the number of
parameters to 29M. The inputs to the network are 227x227x3 RGB images. When the
network is deployed, the features are extracted from the \textit{fc7} layer
resulting in a dimensionality of 512. The network is trained using the Softmax
loss function for multiclass classification using the Caffe deep learning
platform \cite{caffe}.
\section{Experimental setup and Results}\label{sec:results}
In this section we evaluate the proposed method on two challenging datasets:
\begin{enumerate}[leftmargin=*]
\item \textbf{IARPA Janus Benchmark-A (IJB-A)} \textbf{\cite{ijba15}}: This dataset contains 500
subjects with a total of 25,813 images (5,399 still images and 20,414 video
frames sampled at a rate of 1 in 60). The faces in the IJB-A dataset contain
extreme poses and illuminations, more challenging than LFW \cite{lfw}. Some sample images from the IJB-A dataset are shown in Figure \ref{fig:ijba-sample}. An additional challenge of the IJB-A verification protocol is that the template comparisons include image to image, image to set and set to set comparisons. In this work, for a given test template of the IJB-A data we perform two kinds of pooling to produce its final representation:
\begin{itemize}[leftmargin=*]
\item \textit{Average pooling} (CNN$_{ave}$): The deep features of the images and/or
frames present in the template are combined by taking a componentwise average
to produce one feature vector. Thus each feature equally contributes to the
final representation.
\item \textit{Media pooling} (CNN$_{media}$): The deep features are combined keeping in
mind the media source they come from. The metadata provided with IJB-A gives
us the \textit{media id} for each item of the template. Thus to get the final
feature vector, we first take an intra-media average and then combine these by
taking the inter-media average. Thus each feature's contribution to the final
representation is weighted based on its source.
\end{itemize}
\item \textbf{Celebrities in Frontal-Profile (CFP) \cite{cfpw}}: This dataset
contains 7000 images of 500 subjects. The dataset is used for evaluating how
face verification approaches handle pose variation. Hence, it consists of 5000
images in frontal view and 2000 images in extreme profile. The data is
organized into 10 splits, each containing equal number of frontal-frontal and
frontal-profile comparisons. Sample comparison pairs of the CFP dataset are
shown in Figure \ref{fig:cfp}.
\end{enumerate}
\begin{figure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=0.6\linewidth]{ijba_sample.png}
\end{subfigure}
\caption{Images from the IJB-A dataset}
\label{fig:ijba-sample}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=.4\linewidth]{ff.png}
\caption{Frontal-Frontal}
\label{fig:sub1}
\end{subfigure}%
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=.4\linewidth]{fp.png}
\caption{Frontal-Profile}
\label{fig:sub2}
\end{subfigure}
\caption{Sample comparison pairs from the CFP dataset}
\label{fig:cfp}
\end{figure}
\subsection{Pre-processing}
In the training phase, given an input image, we use the HyperFace method
\cite{hyperface} for face detection and fiducial point extraction. The HyperFace
detector automatically extracts many faces from a given image. For the IJB-A dataset, since most images contain more than one face, we use the bounding boxes provided along with
the dataset to select the person of interest from the list of automatic
detections. We select the detection that has the maximum area overlap with the
manually provided bounding box. In the IJB-A dataset, there are few images for
which the HyperFace detector cannot find the relevant face. For the missed
cases, we crop the face using the bounding box information provided with the
dataset and pass it to HyperFace to extract the fiducials. We use six fiducial
points (eyes and mouth corners) to align the detected image to a canonical view
using the similarity transform. For the CFP dataset, since the six keypoints cannot
be computed for profile faces we only use three keypoints on one side of the
face for aligning them.
\begin{table*}
\centering
\resizebox{1.02\textwidth}{2cm}{
\begin{tabular}{|c|c||c||c||c||c|c|c|}
\hline
\multirow{3}{*}{Method} &
\multicolumn{3}{c|}{IJB-A Verification (FNMR@FMR)} &
\multicolumn{4}{c|}{IJB-A Identification} \\
\hline
& 0.001 & 0.01 & 0.1 & FPIR=0.01 & FPIR=0.1 & Rank=1 & Rank=10 \\
\hline
GOTS \cite{ijba15} & 0.8 (0.008) & 0.59 (0.014) & 0.37 (0.023) & 0.047 (0.02) & 0.235 (0.03) & 0.443 (0.02) & - \\
VGG-Face \cite{parkhi15} & 0.396 (0.06) & 0.195 (0.03) & 0.063(0.01) & 0.46 (0.07) & 0.67 (0.03) & 0.913 (0.01) & \textbf{0.981 (0.005)} \\
Masi \emph{et al.} \cite{isi3d} & 0.275 & 0.114 & - & - & - & 0.906 & 0.977 \\
NAN \cite{nan} & 0.215 (0.03) & 0.103 (0.01) & 0.041 (0.005) & - & - & - & - \\
Crosswhite \emph{et al.} \cite{str-temp} & \textbf{0.135 (0.02)} & \textbf{0.06 (0.01)} & \textbf{0.017 (0.007)} & \textbf{0.774 (0.05)} & \textbf{0.882 (0.016)} & 0.928 (0.01) & \textbf{0.986 (0.003)} \\
CNN$_{ave}$ (Ours) & 0.287 (0.05) & 0.146 (0.01) & 0.051 (0.006) & 0.626 (0.06) & 0.795 (0.02) & 0.90 (0.01) & 0.974 (0.004) \\
CNN$_{media}$ (Ours) & 0.234 (0.02) & 0.129 (0.01) & 0.048 (0.005) & 0.67 (0.05) & 0.82 (0.013) & 0.925 (0.01) & 0.978 (0.005) \\
CNN$_{media}$+TPE (Ours) & 0.187 (0.02) & 0.10 (0.01) & 0.036 (0.005) & 0.753 (0.03) & 0.863 (0.014) & \textbf{0.932 (0.01)} & 0.977 (0.005) \\
\hline
\end{tabular}
}
\captionof{table}{Identification and Verification results on the IJB-A dataset. For identification, the scores reported are TPIR values at the indicated points. The results are averages over 10 splits and the standard deviation is given in the brackets for methods which have reported them. $'-'$ implies that the result is not reported for that method. The best results are given in bold.}
\label{results}
\end{table*}
\begin{table*}
\centering
\begin{tabular}{|l|l|l|l|l|l|l|}
\hline
\multirow{3}{*}{Algorithm} &
\multicolumn{3}{c|}{Frontal-Frontal} &
\multicolumn{3}{c|}{Frontal-Profile}
\\
\hline
& Accuracy & EER & AUC & Accuracy & EER & AUC \\
\hline
Sengupta \emph{et al.} \cite{cfpw} & 96.40 (0.69) & 3.48 (0.67) & 99.43 (0.31) & 84.91 (1.82) & 14.97 (1.98) & 93.00 (1.55)\\
\hline
Human Accuracy & 96.24 (0.67) & 5.34 (1.79)& 98.19 (1.13) & \textbf{94.57 (1.10)} & \textbf{5.02 (1.07)} & \textbf{98.92 (0.46)} \\
\hline
CNN (Ours) & \textbf{96.93 (0.61)} & \textbf{2.51 (0.81)} & \textbf{99.68 (0.16)} & 89.17 (2.35) & 8.85 (0.99) & 97.00 (0.53) \\
\hline
\end{tabular}
\captionof{table}{Results on the CFP dataset \cite{cfpw}. The numbers are
averaged over ten test splits and the numbers in brackets indicate standard
deviations of those runs. The best results are given in bold.}
\label{cfpw-results}
\end{table*}
\subsection{Parameters and training times}
The training of the proposed deep architecture is done using SGD with momentum,
which is set to 0.9 and the learning rate is set to 1e-3 and decreased uniformly
by a factor of 10 every 50K iterations. The weight decay is set to 5e-4 for all
layers. The training batch size is set to 256. The training time for our deep
network is 24 hours on a single NVIDIA TitanX GPU. For the IJB-A dataset, we use
the training data provided with each split to obtain the triplet embedding which
takes 3 mins per split. This is the only additional splitwise processing that is
done by the proposed approach. During deployment, the average enrollment time per
image after pre-processing, including alignment and feature extraction is 8ms.
\subsection{Evaluation Pipeline}
Given an image, we pre-process it as described in Section 5.1. The deep
features are computed as an average of the image and its flip. Given two deep features to compare, we compute their cosine similarity score. More specifically,
for the IJB-A dataset, given a template containing multiple faces, we
\textit{flatten} the template features by average pooling or media pooling to
obtain a vector representation. For each split, we learn the TPE projection
using the provided training data. Given two templates for comparison, we compute the cosine similarity
score using the projected 128-dimensional representations.
matrix.
\subsection{Evaluation Metrics}
We report two types of results for the IJB-A dataset: Verification and
Identification. For the verification protocol, we report the False Non-Match
Rate (FNMR) values at several False Match Rates (FMR). For the identification
results, we report open set and closed set metrics. For the open set metrics, the True Positive
Identification Rate quantifies the fraction of subjects that are classified
correctly among the ones that exist in probe but not in gallery. For the closed
set metrics, we report the CMC numbers at different values of False Positive
Identification Rates (FPIRs) and Ranks. More details on the evaluation
metrics for the IJB-A protocol can be found in \cite{ijba15}.
For the CFP dataset, following the protocol set in \cite{cfpw}, we report the
Area under the curve (AUC) and Equal Error Rate (EER) values as averages across
splits, in addition to the classification accuracy. To obtain the accuracy for
each split, we threshold our CNN similarity scores where the threshold is set to
the value that provides the highest classification accuracy over the training data
for each split.
\subsection{Discussion}
\subsubsection*{Performance on IJB-A}
Table \ref{results} presents the results for the proposed methods compared to
existing results for the IJB-A Verification and Identification protocol. The
compared methods are described below:
\begin{itemize}[leftmargin=*]
\item Government-of-the-Shelf (GOTS) \cite{ijba15} is the baseline performance
provided along with the IJB-A dataset.
\item Parkhi \emph{et al.} \cite{parkhi15} train a very deep network (22 layers) over the VGG-Face dataset which contains 2.6M images from 2622 subjects.
\item The Neural Aggregation network (NAN) \cite{nan} is trained over large
amount of videos from the CELEB-1000 dataset \cite{celeb1000} starting from the GoogleNet \cite{gnet} architecture.
\item Masi \emph{et al.} \cite{isi3d} use a deep CNN based approach that includes a combination of in-plane aligned images, 3D rendered images to augment their performance. The 3D rendered images are also generated during test time per template comparison. It should be noted that many test images of the IJB-A dataset contain extreme poses, harsh illumination conditions and significant blur.
\item Crosswhite \emph{et al.} use template adaptation \cite{wolf} to tune the performance of their raw features specifically to the IJB-A dataset.
\end{itemize}
\vspace{2mm}
Compared to these methods, the proposed method trains a single CNN model on the
CASIA-WebFace dataset which consists of about 500K images and requires much
shorter training time and has a very fast query time (0.08s after face detection
per image pair). As shown in Table \ref{results}, our raw CNN features after media pooling perform better than most compared methods across both the verification and identification protocols of the IJB-A dataset, with the exception of the template adaptation method by Crosswhite \emph{et al.} \cite{str-temp} which is discussed below. The TPE method provides significant improvement for both identification and verification tasks as shown in Table \ref{results}.\\
The method by Crosswhite \emph{et al.} \cite{str-temp} uses the VGG-Face network \cite{parkhi15} descriptors (4096-d) as the raw features. They use the concept of template adaptation \cite{wolf} to improve their performance as follows: when pooling multiple faces of a given template, they train a linear SVM with the features of this template as positive and a
fixed set of negatives extracted from the training data of the IJB-A splits.
Let's denote the pooled template feature and classifier pair as $(t,w)$. Then, at
query time when comparing two templates $(t_1,w_1)$ and $(t_2,w_2)$, the similarity
score is computed as: $\frac{1}{2} \left( t_1 \cdot w_2 + t_2 \cdot w_1
\right)$. Even when using a carefully engineered fast linear classifier training
algorithm, this procedure increases the run time of the pooling procedure. The query time per template comparison is also higher due to the high dimensionality of the input features. In contrast, the proposed approach requires a matrix multiplication and a vector dot product per
comparison. By using a simple neural network architecture, a relatively smaller
training dataset and a fast embedding method we have realized a faster and more efficient end-to-end system. To improve our performance further, we are currently incorporating the use of video data into our approach.
\subsubsection*{Performance on CFP}
On the CFP dataset, we achieve a new state-of-art on both Frontal-Frontal and
Frontal-Profile comparisons, the latter by a large margin. More specifically,
for the Frontal-Profile case, we manage to reduce the error rate by
\textbf{40.8\%}. It should be noted that for a fair comparison we have used our
raw CNN features without performing TPE. This shows that the raw CNN features we
learn are effective even at extreme pose variations.
\section{Learning a Discriminative Embedding}\label{tpe}
In this section, we describe our algorithm for learning a low-dimensional
embedding such that the resulting projections are more discriminative. Aside
from an improved performance, this embedding provides significant advantages in
terms of memory and enables post-processing operations like visualization and clustering. \\
Consider a triplet $t:=(\bm{v}_i,\bm{v}_j,\bm{v}_k)$, where $\bm{v}_i$ (anchor)
and $\bm{v}_j$ (positive) are from the same class, but $\bm{v}_k$ (negative)
belongs to a different class. Consider a function $S_W:\mathbb{R}^N \times
\mathbb{R}^N \mapsto \mathbb{R}$ that is parameterized by the matrix $\mbf{W}
\in \mathbb{R}^{n \times N}$, that measures the similarity between two vectors
$\bm{v}_i,\bm{v}_j \in \mathbb{R}^N$. Ideally, for all triplets $t$ that exist
in the training set, we would like the following constraint to be satisfied:
\vspace{-5mm}
\begin{align*}
S_\mbf{W}(\bm{v}_i,\bm{v}_j)> S_\mbf{W}(\bm{v}_i,\bm{v}_k)
\tageq \label{eq:constraint}
\end{align*}
Thus, the probability of a given triplet $t$ satisfying (\ref{eq:constraint}) can be written as:
\vspace{-2.5mm}
\begin{align*}
p_{ijk}=\frac{e^{S_\mbf{W}(\bm{v}_i,\bm{v}_j)}}{e^{S_\mbf{W}(\bm{v}_i,\bm{v}_j)} + e^{S_\mbf{W}(\bm{v}_i,\bm{v}_k)}}
\tageq \label{eq:prob}
\end{align*}
The specific form of the similarity function is given as:
$S_\mbf{W}(\bm{v}_i,\bm{v}_j)= (\mbf{W}\bm{v}_i)^T \cdot (\mbf{W} \bm{v}_j)$. In our case, $\bm{v}_i$ and $\bm{v}_j$ are deep features normalized to unit length. To learn the embedding $\mbf{W}$ from a given set of triplets $\mathbb{T}$, we solve the following optimization:
\vspace{-2.5mm}
\begin{align*}
\underset{\mbf{W}}{\text{argmin}} \sum_{ (\bm{v}_i,\bm{v}_j,\bm{v}_k) \in \mathbb{T}} - \log(p_{ijk})
\tageq \label{eq:train1}
\end{align*}
(\ref{eq:train1}) can be interpreted as maximizing the likelihood
(\ref{eq:constraint}) or minimizing the negative log-likelihood (NLL) over the
triplet set $\mathbb{T}$. In practice, the above problem is solved in a
Large-Margin framework using Stochastic Gradient Descent (SGD) and the triplets
are sampled online. The gradient update for $\mbf{W}$ is given as:
\begin{align*}
\mbf{W}_{\tau+1} = \mbf{W}_\tau - \eta * \mbf{W}_\tau *(1-p_{ijk})* (\bm{v}_i(\bm{v}_j-\bm{v}_k)^T \\
+ (\bm{v}_j-\bm{v}_k)\bm{v}_i^T)
\tageq \label{eq:update}
\end{align*}
where $\mbf{W}_\tau$ is the estimate at iteration $\tau$, $\mbf{W}_{\tau+1}$ is the
updated estimate, $(\bm{v}_i,\bm{v}_j,\bm{v}_k)$ is the triplet sampled at the current iteration and $\eta$ is the learning rate.\\
By choosing the dimension of $\mbf{W}$ as $n \times N$ with $n < N$, we achieve
dimensionality reduction in addition to improved performance. For our work, we fix
$n=128$ based on cross validation and $N=512$ is the dimensionality of our deep
features. $\mbf{W}$ is initialized with the first $n$ principal components of
the training data. At each iteration, a random anchor and a random positive data
point are chosen. To choose the negative, we perform hard negative mining, ie.
we choose the data point that has the least likelihood (\ref{eq:prob}) among the
randomly chosen 2000 negative instances at each iteration.
Since we compute the embedding matrix $\mbf{W}$ by optimizing over triplet
probabilities, we call this method Triplet Probability Embedding (TPE). The
technique closest to the one presented in this section, which is used in recent
works (\cite{facenet15},\cite{parkhi15}) computes the embedding \mbf{W} based on
satisfying a hinge loss constraint:
\begin{align*}
\underset{\mbf{W}}{\text{argmin}} \sum_{ (\bm{v}_i,\bm{v}_j,\bm{v}_k) \in \mathbb{T}} \max\{0,\alpha + (\bm{v}_i-\bm{v}_j)^T\mbf{W^TW}(\bm{v}_i-\bm{v}_j)- \\
(\bm{v}_i-\bm{v}_k)^T\mbf{W^TW}(\bm{v}_i-\bm{v}_k) \}
\tageq \label{eq:dist}
\end{align*}
$\alpha$ acts a margin parameter for the loss function. To be consistent with
the terminology used in this paper, we call it Triplet Distance Embedding (TDE). To appreciate the difference between the two approaches, Figure \ref{fig:tde} shows the case where the gradient update for the TDE method
(\ref{eq:dist}) occurs. If the value of $\alpha$ is not appropriately chosen, a triplet is considered good even if the positive and negative are very
close to one another. But under the proposed formulation, both cases referred to
in Figure \ref{fig:tde} will update the gradient but their contribution to the
gradient will be modulated by the probability with which they violate the
constraint in (\ref{eq:constraint}). This modulation factor is specified by the
$(1-p_{ijk})$ term in the gradient update for TPE in (\ref{eq:update}) implying
that if the likelihood of a sampled triplet satisfying (\ref{eq:constraint}) is
high, then the gradient update is given a lower weight and vice-versa. Thus, in
our method, the margin parameter ($\alpha$) is automatically set based on the
likelihood. \\
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth,height=0.30\textwidth]{tde_update.png}
\caption{Gradient update scenarios for the TDE method (\ref{eq:dist}). The notation is explained in the text}
\label{fig:tde}
\end{figure}
To compare the relative performances of the raw features before projection, with
TDE and with TPE (proposed method), we plot the traditional ROC curve (TAR (vs)
FAR) for split 1 of the IJB-A verify protocol for the three methods in Figure
\ref{fig:tl}. The Equal Error Rate (EER) metric is specified for each method. The performance
improvement due to TPE is significant, especially at regions of
FAR$=\{10^{-4},10^{-3}\}$. We observed a similar behaviour for all the ten
splits of the IJB-A dataset.
\begin{figure}
\includegraphics[width=0.5\textwidth,height=0.25\textwidth]{roc_compare.jpg}
\caption{Performance improvement on IJB-A split 1: FAR (vs) TAR plot. EER values are specified in brackets.}
\label{fig:tl}
\end{figure}
\section{Acknowledgement}\label{sec:ackno}
This research is based upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA),via IARPA R\&D Contract No. 2014-14071600012. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation thereon.
\section{Related Work}\label{SOA}
In the past few years, there have been numerous works in using deep features for tasks related to face verification. The DeepFace \cite{deepface14} approach uses a carefully crafted 3D alignment procedure to preprocess face images and feeds them to a deep network that is trained using a large training set. More recently, Facenet \cite{facenet15} uses a large private dataset to train several deep network models using a triplet distance loss function. The training time for this network is of the order of few weeks. Since the release of the IJB-A dataset \cite{ijba15}, there have been several works that have published verification results for this dataset. Previous approaches presented in \cite{wang15} and \cite{parkhi15} train deep networks using the CASIA-WebFace dataset \cite{casia14} and the VGG-Face dataset respectively, requiring substantial training time. This paper proposes a network architecture and a training scheme that needs shorter training time and a small query time. \\
The idea of learning a compact and discriminative representation has been around for decades. Weinberger \emph{et al.} \cite{lmnn05} used a Semi Definite Programming (SDP)-based formulation to learn a metric satisfying pairwise and triplet distance constraints in a large margin framework. More recently, this idea has been successfully applied to face verification by integrating the loss function within the deep network architecture (\cite{facenet15}, \cite{parkhi15}). Joint Bayesian metric learning is also another popular metric used for face verification (\cite{fvf13},\cite{chen15wacv}). These methods either require a large dataset for convergence or learn a metric directly and therefore are not amenable to subsequent operations like discriminative clustering or hashing. Classic methods like t-SNE \cite{tsne}, t-STE \cite{tste} and Crowd Kernel Learning (CKL) \cite{CKL} perform extremely well when used to visualize or cluster a given data collection. They either operate on the data matrix directly or the distance matrix generated from data by generating a large set of pairwise or triplet constraints. While these methods perform very well on a given set of data points, they do not generalize to out-of-sample data. In the current work, we aim to generalize such formulations, to a more traditional classification setting, where domain specific training and testing data is provided. We formulate an optimization problem based on triplet probabilities that performs dimensionality reduction aside from improving the discriminative ability of the test data. The embedding scheme described in this work is a more general framework that can be applied to any setting where labeled training data is available.
\section{Clustering Faces}\label{sec:cluster}
\begin{figure*}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.6\linewidth]{image_clusters/pose_blur/1.png}
\caption{}
\label{fig:sfig1}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.6\linewidth]{image_clusters/pose_blur/2.png}
\caption{}
\label{fig:sfig2}
\end{subfigure}\\
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.6\linewidth]{image_clusters/age/1.png}
\caption{}
\label{fig:sfig3}
\end{subfigure}%
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=.6\linewidth]{image_clusters/age/2.png}
\caption{}
\label{fig:sfig4}
\end{subfigure} \\
\caption{Sample clusters output from the Clustering approach discussed in Section 6 for the data from the split 1 of the IJB-A dataset. Top row (a,b) shows robustness to pose and blur; Bottom row (c,d) contains clusters that are robust to age}
\label{fig:cluster}
\end{figure*}
This section illustrates how the proposed TPE method can be used to cluster a
given data collection. We perform two clustering experiments:
\begin{enumerate}[leftmargin=*]
\item We perform clustering on the entire LFW \cite{lfw} dataset that consists
of 13233 images of 5749 subjects. It should be noted that about 4169 subjects have only one image.
\item We use the IJB-A dataset and cluster the templates corresponding to the query set for each split in the IJB-A verify protocol.
\end{enumerate}
For evaluating the clustering results, we use the metrics defined in
\cite{msu_cluster}. These are summarized below:
\begin{itemize}[leftmargin=*]
\item \textit{Pairwise Precision ($P_{pair}$)}: The fraction of pairs of samples
within a cluster among all possible pairs which are of the same class, over
the total number of same cluster pairs.
\item \textit{Pairwise Recall ($R_{pair}$)}: The fraction of pairs of samples
within a class among all possible pairs which are placed in the same cluster,
over the total number of same-class pairs.
\end{itemize}
Using these metrics, the F$_1$-score is computed as:
\begin{equation}
F_1 = \frac{2*P_{pair}*R_{pair}}{R_{pair}+P_{pair}}
\end{equation}
The simplest way we found to demonstrate the effectiveness of our deep features
and the proposed TPE method, is to use the standard MATLAB implementation of the
agglomerative clustering algorithm with the average linkage metric. We use the
cosine similarity as our basic clustering metric. The simple clustering
algorithm that we have used here has computational complexity of
$O(N^2)$. In its current form, this does not scale to large datasets with
millions of images. We are currently working on a more efficient and scalable (yet approximate) version of this algorithm.
\paragraph{Clustering LFW:-}The images in the LFW dataset are pre-processed as described in Section 5.1. For each image and its flip, the deep features are extracted using the proposed architecture, averaged and normalized to unit $L_2$ norm. We run the clustering algorithm over the entire data in a single shot. The clustering algorithm takes as input a cut-off parameter which acts as a distance threshold (below which any two clusters will not be merged). In our experiments, we vary this cut-off parameter over a small range and evaluate the
resulting clustering using the $F_1$-score. We pick the result that yields the
best $F_1$-score. Table \ref{tab:lfw_cluster} shows the result of our approach
and compares it to a recently released clustering approach based on approximate
Rank-order clustering \cite{msu_cluster}. It should be noted that, in the case
of \cite{msu_cluster}, the clustering result is chosen by varying the number of
clusters and picking the one with the best $F_1$-score. In our approach, we vary
the cut-off threshold which is the property of deep features and hence is a
more intuitive parameter to tune. We see from Table \ref{tab:lfw_cluster} that
aside from better performance, our total cluster estimate is closer to the
ground truth value of 5749 than \cite{msu_cluster}.
\begin{table}
\centering
\begin{tabular}{ | l | l | l|}
\hline
Method & $F_1$-score & Clusters \\ \hline
\cite{msu_cluster} & 0.87 & 6508 \\ \hline
CNN (Ours) & \textbf{0.955} & 5351\\ \hline
\end{tabular}
\captionof{table}{$F_1$-score for comparison of the two clustering schemes on
the LFW dataset. The ground truth cluster number is 5749.}
\label{tab:lfw_cluster}
\end{table}
\begin{table}
\centering
\resizebox{0.45\textwidth}{!}{
\begin{tabular}{ | l | l | l| l |}
\hline
Method & $F_1$-score & Clusters & After Pruning \\ \hline
CNN$_{media}$ & 0.79 (0.02) & 293 (22) & 173 \\ \hline
CNN$_{media}$+TPE & \textbf{0.843 (0.03)} & 258 (17) & 167 \\ \hline
\end{tabular}
}
\captionof{table}{Clustering metrics over the IJB-A 1:1 protocol. The
standard deviation is indicated in brackets. The ground truth subjects per
each split is 167.}
\label{tab:ijba_cluster}
\end{table}
\paragraph{Clustering IJB-A:-}The IJB-A dataset is processed as described in Section 5. In this section, we aim to cluster the query templates provided with each split for the verify protocol. We report the results of two experiments: with the raw CNN features (CNN$_{media}$ in Table 2) and with the projected CNN features, where the projection matrix is learned through the proposed TPE method (CNN$_{media}$+TPE in Table 2). The cut-off threshold required for our clustering algorithm is learned automatically based on the training data, i.e. we choose the threshold that gives the maximum $F_1$-score over the training data. The scores reported in Table \ref{tab:ijba_cluster} are average values over ten splits. As expected, the TPE method improves the clustering performance of raw features. The subject estimate is the number of clusters produced as a direct result of our clustering algorithm. The pruned estimate is obtained by ignoring clusters that have fewer than 3 images.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth,height=0.3\textwidth]{raw_tpe_cluster.png}
\caption{Precision-Recall curve plotted over cut-off threshold varied from 0 to 1.}
\label{fig:pr_curve}
\end{figure}
For a more complete evaluation of our performance over varying threshold values, we plot the Precision-Recall (PR) curve for the IJB-A clustering experiment in Figure \ref{fig:pr_curve}. As can be observed, the PR curve for clustering the IJB-A data using embedded features exhibits a better performance at all operating points. This is a more transparent evaluation than reporting only the $F_1$-score since the latter effectively fixes the operating point but the PR curve reveals the performance at all operating points.
\section{Conclusion and Future Work}\label{conclusion}
In this paper, we proposed a deep CNN-based approach coupled with a
low-dimensional discriminative embedding learned using triplet probability
constraints in a large margin fashion. The proposed pipeline enables a faster
training time and improves face verification performance especially at low
FMRs. We demonstrated the effectiveness of the proposed method on two
challenging datasets: IJB-A and CFP and achieved performance close to
the state of the art while using a deep model which is more compact and trained
using a moderately sized dataset. We demonstrated the robustness of our features
using a simple clustering algorithm on the LFW and IJB-A datasets. For future
work, we plan to use videos directly during training and also embed our TPE
approach into training the deep network. We intend to scale our clustering
algorithm to handle large scale scenarios such as large impostor sets of the
order of millions.
\section{Introduction}
\label{sec:intro}
Recently, with the advent of curated face datasets like Labeled faces in the
Wild (LFW) \cite{lfw} and advances in learning algorithms like Deep neural nets,
there is more hope that the unconstrained face verification problem can be
solved. A face verification algorithm compares two given templates that are
typically not seen during training. Research in face verification has progressed
well over the past few years, resulting in the saturation of performance on the
LFW dataset, yet the problem of unconstrained face verification remains a
challenge. This is evident by the performance of traditional algorithms on the
publicly available IJB-A dataset (\cite{ijba15}, \cite{fvff15}) that was
released recently. Moreover, despite the superb performance of CNN-based
approaches compared to traditional methods, a drawback of such methods is the
long training time needed. In this work, we present a Deep CNN (DCNN)
architecture that ensures faster training, and investigate how much the
performance can be improved if we are provided domain specific data.
Specifically, our contributions are as follows:
\begin{itemize}
\item We propose a deep network architecture and a training scheme that ensures
faster training time.
\item We formulate a triplet probability embedding learning method to improve
the performance of deep features for face verification and subject clustering.
\end{itemize}
During training, we use a publicly available face dataset to train our deep
architecture. Each image is pre-processed and aligned to a canonical view before
passing it to the deep network whose features are used to represent the image. In the case of IJB-A dataset, the data is divided into 10 splits, each split containing a training set and a test set. Hence, to further improve performance, we learn the proposed triplet probability embedding using the training set provided with each split over the features extracted from our DCNN model. During the deployment phase, given a face template, we extract the deep features using the raw CNN model after implementing automatic pre-processing steps such as face detection and fiducial extraction. The deep features are projected onto
a low-dimensional space using the embedding matrix learned during training (note
that the projection involves only matrix multiplication). We use the
128-dimensional feature as the final representation of the given face template.
This paper is organized as follows: Section \ref{SOA} places our work among the
recently proposed approaches for face verification. Section \ref{net} details
the network architecture and the training scheme. The triplet probabilistic
embedding learning method is described in Section \ref{tpe} followed by results
on IJB-A and CFP datasets and a brief discussion in Section
\ref{sec:results}. In Section \ref{sec:cluster}, we demonstrate the ability of
the proposed method to cluster a media collection from LFW and IJB-A
datasets.
\section{Network Architecture}\label{net}
This section details the architecture and training algorithm for the deep
network used in our work. Our architecture consists of 7 convolutional layers
with varying kernel sizes. The initial layers have a larger size rapidly
subsampling the image and reducing the parameters while subsequent layers consist
of small filter sizes, which has proved to be very useful in face recognition
tasks (\cite{parkhi15},\cite{casia14}). Furthermore, we use the Parametric
Rectifier Linear units (PReLUs) instead of ReLUs, since they allow a negative
value for the output based on a learned threshold and have been shown to improve
the convergence rate \cite{prelu15}.
\begin{Table}
\centering
\resizebox{6cm}{2.2cm}{
\begin{tabular}{|c|c|c|}
\hline
Layer & Kernel Size/Stride & \#params\\
\hline
conv1 &11x11/4 & 35K\\
pool1 &3x3/2 & \\
conv2 &5x5/2 & 614K\\
pool2 &3x3/2 & \\
conv3 &3x3/1 & 885K\\
conv4 &3x3/1 & 1.3M\\
conv5 &3x3/1 & 2.3M\\
conv6 &3x3/1 & 2.3M\\
conv7 &3x3/1 & 2.3M\\
pool7 &3x3/2 & \\
fc6 &1024 & 18.8M \\
fc7 &512 & 524K \\
fc8 &10548 & 10.8M \\
Softmax Loss & & Total: 39.8M\\
\hline
\end{tabular}
}
\captionof{table}{Deep Network architecture details}
\label{arch}
\end{Table}
The top three convolutional layers (conv1-conv3) are initialized with the
weights from the AlexNet model \cite{alexnet12} trained on the ImageNet
challenge dataset. Several recent works (\cite{transfer1},\cite{transfer2}) have
empirically shown that this transfer of knowledge across different networks,
albeit for a different objective, improves performance and more significantly
reduces the need to train over a large number of iterations.
The compared methods either learn their deep models from scratch
(\cite{parkhi15},\cite{nan}) or finetune only the last layer of fully
pre-trained models. The former results in large training time and the
latter does not generalize well to the task at hand (face verification) and
hence resulting in sub optimal performance. In the current work, even though we
use a pre-trained model (AlexNet) to initialize the proposed deep network, we do
so only for the first three convolutional layers, since they retain more generic
information (\cite{transfer1}). Subsequent layers learn representations which
are more specific to the task at hand. Thus, to learn more task specific
information, we add 4 convolutional layers each consisting of 512 kernels of
size $3\times3$. The layers conv4-conv7 do not downsample the input thereby
learning more complex higher dimensional representations. This hybrid
architecture proves to be extremely effective as our raw CNN representation
outperforms some very deep CNN models on the IJB-A dataset (Table 2 in Results).
In addition, we achieve that performance by training the proposed deep network using the
relatively smaller CASIA-WebFace dataset.
The architecture of our network is shown in Table \ref{arch}. Layers
conv4-conv7 and the fully connected layers \textit{fc6-fc8} are initialized from
scratch using random Gaussian distributions. PReLU activation functions are added
between each layer. Since the network is used as a feature extractor, the last
layer \textit{fc8} is removed during deployment, thus reducing the number of
parameters to 29M. The inputs to the network are 227x227x3 RGB images. When the
network is deployed, the features are extracted from the \textit{fc7} layer
resulting in a dimensionality of 512. The network is trained using the Softmax
loss function for multiclass classification using the Caffe deep learning
platform \cite{caffe}.
\section{Experimental setup and Results}\label{sec:results}
In this section we evaluate the proposed method on two challenging datasets:
\begin{enumerate}[leftmargin=*]
\item \textbf{IARPA Janus Benchmark-A (IJB-A)} \textbf{\cite{ijba15}}: This dataset contains 500
subjects with a total of 25,813 images (5,399 still images and 20,414 video
frames sampled at a rate of 1 in 60). The faces in the IJB-A dataset contain
extreme poses and illuminations, more challenging than LFW \cite{lfw}. Some sample images from the IJB-A dataset are shown in Figure \ref{fig:ijba-sample}. An additional challenge of the IJB-A verification protocol is that the template comparisons include image to image, image to set and set to set comparisons. In this work, for a given test template of the IJB-A data we perform two kinds of pooling to produce its final representation:
\begin{itemize}[leftmargin=*]
\item \textit{Average pooling} (CNN$_{ave}$): The deep features of the images and/or
frames present in the template are combined by taking a componentwise average
to produce one feature vector. Thus each feature equally contributes to the
final representation.
\item \textit{Media pooling} (CNN$_{media}$): The deep features are combined keeping in
mind the media source they come from. The metadata provided with IJB-A gives
us the \textit{media id} for each item of the template. Thus to get the final
feature vector, we first take an intra-media average and then combine these by
taking the inter-media average. Thus each feature's contribution to the final
representation is weighted based on its source.
\end{itemize}
\item \textbf{Celebrities in Frontal-Profile (CFP) \cite{cfpw}}: This dataset
contains 7000 images of 500 subjects. The dataset is used for evaluating how
face verification approaches handle pose variation. Hence, it consists of 5000
images in frontal view and 2000 images in extreme profile. The data is
organized into 10 splits, each containing equal number of frontal-frontal and
frontal-profile comparisons. Sample comparison pairs of the CFP dataset are
shown in Figure \ref{fig:cfp}.
\end{enumerate}
\begin{figure}
\begin{subfigure}{.5\textwidth}
\centering
\includegraphics[width=0.6\linewidth]{ijba_sample.png}
\end{subfigure}
\caption{Images from the IJB-A dataset}
\label{fig:ijba-sample}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=.4\linewidth]{ff.png}
\caption{Frontal-Frontal}
\label{fig:sub1}
\end{subfigure}%
\begin{subfigure}{.25\textwidth}
\centering
\includegraphics[width=.4\linewidth]{fp.png}
\caption{Frontal-Profile}
\label{fig:sub2}
\end{subfigure}
\caption{Sample comparison pairs from the CFP dataset}
\label{fig:cfp}
\end{figure}
\subsection{Pre-processing}
In the training phase, given an input image, we use the HyperFace method
\cite{hyperface} for face detection and fiducial point extraction. The HyperFace
detector automatically extracts many faces from a given image. For the IJB-A dataset, since most images contain more than one face, we use the bounding boxes provided along with
the dataset to select the person of interest from the list of automatic
detections. We select the detection that has the maximum area overlap with the
manually provided bounding box. In the IJB-A dataset, there are few images for
which the HyperFace detector cannot find the relevant face. For the missed
cases, we crop the face using the bounding box information provided with the
dataset and pass it to HyperFace to extract the fiducials. We use six fiducial
points (eyes and mouth corners) to align the detected image to a canonical view
using the similarity transform. For the CFP dataset, since the six keypoints cannot
be computed for profile faces we only use three keypoints on one side of the
face for aligning them.
\begin{table*}
\centering
\resizebox{1.02\textwidth}{2cm}{
\begin{tabular}{|c|c||c||c||c||c|c|c|}
\hline
\multirow{3}{*}{Method} &
\multicolumn{3}{c|}{IJB-A Verification (FNMR@FMR)} &
\multicolumn{4}{c|}{IJB-A Identification} \\
\hline
& 0.001 & 0.01 & 0.1 & FPIR=0.01 & FPIR=0.1 & Rank=1 & Rank=10 \\
\hline
GOTS \cite{ijba15} & 0.8 (0.008) & 0.59 (0.014) & 0.37 (0.023) & 0.047 (0.02) & 0.235 (0.03) & 0.443 (0.02) & - \\
VGG-Face \cite{parkhi15} & 0.396 (0.06) & 0.195 (0.03) & 0.063(0.01) & 0.46 (0.07) & 0.67 (0.03) & 0.913 (0.01) & \textbf{0.981 (0.005)} \\
Masi \emph{et al.} \cite{isi3d} & 0.275 & 0.114 & - & - & - & 0.906 & 0.977 \\
NAN \cite{nan} & 0.215 (0.03) & 0.103 (0.01) & 0.041 (0.005) & - & - & - & - \\
Crosswhite \emph{et al.} \cite{str-temp} & \textbf{0.135 (0.02)} & \textbf{0.06 (0.01)} & \textbf{0.017 (0.007)} & \textbf{0.774 (0.05)} & \textbf{0.882 (0.016)} & 0.928 (0.01) & \textbf{0.986 (0.003)} \\
CNN$_{ave}$ (Ours) & 0.287 (0.05) & 0.146 (0.01) & 0.051 (0.006) & 0.626 (0.06) & 0.795 (0.02) & 0.90 (0.01) & 0.974 (0.004) \\
CNN$_{media}$ (Ours) & 0.234 (0.02) & 0.129 (0.01) & 0.048 (0.005) & 0.67 (0.05) & 0.82 (0.013) & 0.925 (0.01) & 0.978 (0.005) \\
CNN$_{media}$+TPE (Ours) & 0.187 (0.02) & 0.10 (0.01) & 0.036 (0.005) & 0.753 (0.03) & 0.863 (0.014) & \textbf{0.932 (0.01)} & 0.977 (0.005) \\
\hline
\end{tabular}
}
\captionof{table}{Identification and Verification results on the IJB-A dataset. For identification, the scores reported are TPIR values at the indicated points. The results are averages over 10 splits and the standard deviation is given in the brackets for methods which have reported them. $'-'$ implies that the result is not reported for that method. The best results are given in bold.}
\label{results}
\end{table*}
\begin{table*}
\centering
\begin{tabular}{|l|l|l|l|l|l|l|}
\hline
\multirow{3}{*}{Algorithm} &
\multicolumn{3}{c|}{Frontal-Frontal} &
\multicolumn{3}{c|}{Frontal-Profile}
\\
\hline
& Accuracy & EER & AUC & Accuracy & EER & AUC \\
\hline
Sengupta \emph{et al.} \cite{cfpw} & 96.40 (0.69) & 3.48 (0.67) & 99.43 (0.31) & 84.91 (1.82) & 14.97 (1.98) & 93.00 (1.55)\\
\hline
Human Accuracy & 96.24 (0.67) & 5.34 (1.79)& 98.19 (1.13) & \textbf{94.57 (1.10)} & \textbf{5.02 (1.07)} & \textbf{98.92 (0.46)} \\
\hline
CNN (Ours) & \textbf{96.93 (0.61)} & \textbf{2.51 (0.81)} & \textbf{99.68 (0.16)} & 89.17 (2.35) & 8.85 (0.99) & 97.00 (0.53) \\
\hline
\end{tabular}
\captionof{table}{Results on the CFP dataset \cite{cfpw}. The numbers are
averaged over ten test splits and the numbers in brackets indicate standard
deviations of those runs. The best results are given in bold.}
\label{cfpw-results}
\end{table*}
\subsection{Parameters and training times}
The training of the proposed deep architecture is done using SGD with momentum,
which is set to 0.9 and the learning rate is set to 1e-3 and decreased uniformly
by a factor of 10 every 50K iterations. The weight decay is set to 5e-4 for all
layers. The training batch size is set to 256. The training time for our deep
network is 24 hours on a single NVIDIA TitanX GPU. For the IJB-A dataset, we use
the training data provided with each split to obtain the triplet embedding which
takes 3 mins per split. This is the only additional splitwise processing that is
done by the proposed approach. During deployment, the average enrollment time per
image after pre-processing, including alignment and feature extraction is 8ms.
\subsection{Evaluation Pipeline}
Given an image, we pre-process it as described in Section 5.1. The deep
features are computed as an average of the image and its flip. Given two deep features to compare, we compute their cosine similarity score. More specifically,
for the IJB-A dataset, given a template containing multiple faces, we
\textit{flatten} the template features by average pooling or media pooling to
obtain a vector representation. For each split, we learn the TPE projection
using the provided training data. Given two templates for comparison, we compute the cosine similarity
score using the projected 128-dimensional representations.
matrix.
\subsection{Evaluation Metrics}
We report two types of results for the IJB-A dataset: Verification and
Identification. For the verification protocol, we report the False Non-Match
Rate (FNMR) values at several False Match Rates (FMR). For the identification
results, we report open set and closed set metrics. For the open set metrics, the True Positive
Identification Rate quantifies the fraction of subjects that are classified
correctly among the ones that exist in probe but not in gallery. For the closed
set metrics, we report the CMC numbers at different values of False Positive
Identification Rates (FPIRs) and Ranks. More details on the evaluation
metrics for the IJB-A protocol can be found in \cite{ijba15}.
For the CFP dataset, following the protocol set in \cite{cfpw}, we report the
Area under the curve (AUC) and Equal Error Rate (EER) values as averages across
splits, in addition to the classification accuracy. To obtain the accuracy for
each split, we threshold our CNN similarity scores where the threshold is set to
the value that provides the highest classification accuracy over the training data
for each split.
\subsection{Discussion}
\subsubsection*{Performance on IJB-A}
Table \ref{results} presents the results for the proposed methods compared to
existing results for the IJB-A Verification and Identification protocol. The
compared methods are described below:
\begin{itemize}[leftmargin=*]
\item Government-of-the-Shelf (GOTS) \cite{ijba15} is the baseline performance
provided along with the IJB-A dataset.
\item Parkhi \emph{et al.} \cite{parkhi15} train a very deep network (22 layers) over the VGG-Face dataset which contains 2.6M images from 2622 subjects.
\item The Neural Aggregation network (NAN) \cite{nan} is trained over large
amount of videos from the CELEB-1000 dataset \cite{celeb1000} starting from the GoogleNet \cite{gnet} architecture.
\item Masi \emph{et al.} \cite{isi3d} use a deep CNN based approach that includes a combination of in-plane aligned images, 3D rendered images to augment their performance. The 3D rendered images are also generated during test time per template comparison. It should be noted that many test images of the IJB-A dataset contain extreme poses, harsh illumination conditions and significant blur.
\item Crosswhite \emph{et al.} use template adaptation \cite{wolf} to tune the performance of their raw features specifically to the IJB-A dataset.
\end{itemize}
\vspace{2mm}
Compared to these methods, the proposed method trains a single CNN model on the
CASIA-WebFace dataset which consists of about 500K images and requires much
shorter training time and has a very fast query time (0.08s after face detection
per image pair). As shown in Table \ref{results}, our raw CNN features after media pooling perform better than most compared methods across both the verification and identification protocols of the IJB-A dataset, with the exception of the template adaptation method by Crosswhite \emph{et al.} \cite{str-temp} which is discussed below. The TPE method provides significant improvement for both identification and verification tasks as shown in Table \ref{results}.\\
The method by Crosswhite \emph{et al.} \cite{str-temp} uses the VGG-Face network \cite{parkhi15} descriptors (4096-d) as the raw features. They use the concept of template adaptation \cite{wolf} to improve their performance as follows: when pooling multiple faces of a given template, they train a linear SVM with the features of this template as positive and a
fixed set of negatives extracted from the training data of the IJB-A splits.
Let's denote the pooled template feature and classifier pair as $(t,w)$. Then, at
query time when comparing two templates $(t_1,w_1)$ and $(t_2,w_2)$, the similarity
score is computed as: $\frac{1}{2} \left( t_1 \cdot w_2 + t_2 \cdot w_1
\right)$. Even when using a carefully engineered fast linear classifier training
algorithm, this procedure increases the run time of the pooling procedure. The query time per template comparison is also higher due to the high dimensionality of the input features. In contrast, the proposed approach requires a matrix multiplication and a vector dot product per
comparison. By using a simple neural network architecture, a relatively smaller
training dataset and a fast embedding method we have realized a faster and more efficient end-to-end system. To improve our performance further, we are currently incorporating the use of video data into our approach.
\subsubsection*{Performance on CFP}
On the CFP dataset, we achieve a new state-of-art on both Frontal-Frontal and
Frontal-Profile comparisons, the latter by a large margin. More specifically,
for the Frontal-Profile case, we manage to reduce the error rate by
\textbf{40.8\%}. It should be noted that for a fair comparison we have used our
raw CNN features without performing TPE. This shows that the raw CNN features we
learn are effective even at extreme pose variations.
\section{Learning a Discriminative Embedding}\label{tpe}
In this section, we describe our algorithm for learning a low-dimensional
embedding such that the resulting projections are more discriminative. Aside
from an improved performance, this embedding provides significant advantages in
terms of memory and enables post-processing operations like visualization and clustering. \\
Consider a triplet $t:=(\bm{v}_i,\bm{v}_j,\bm{v}_k)$, where $\bm{v}_i$ (anchor)
and $\bm{v}_j$ (positive) are from the same class, but $\bm{v}_k$ (negative)
belongs to a different class. Consider a function $S_W:\mathbb{R}^N \times
\mathbb{R}^N \mapsto \mathbb{R}$ that is parameterized by the matrix $\mbf{W}
\in \mathbb{R}^{n \times N}$, that measures the similarity between two vectors
$\bm{v}_i,\bm{v}_j \in \mathbb{R}^N$. Ideally, for all triplets $t$ that exist
in the training set, we would like the following constraint to be satisfied:
\vspace{-5mm}
\begin{align*}
S_\mbf{W}(\bm{v}_i,\bm{v}_j)> S_\mbf{W}(\bm{v}_i,\bm{v}_k)
\tageq \label{eq:constraint}
\end{align*}
Thus, the probability of a given triplet $t$ satisfying (\ref{eq:constraint}) can be written as:
\vspace{-2.5mm}
\begin{align*}
p_{ijk}=\frac{e^{S_\mbf{W}(\bm{v}_i,\bm{v}_j)}}{e^{S_\mbf{W}(\bm{v}_i,\bm{v}_j)} + e^{S_\mbf{W}(\bm{v}_i,\bm{v}_k)}}
\tageq \label{eq:prob}
\end{align*}
The specific form of the similarity function is given as:
$S_\mbf{W}(\bm{v}_i,\bm{v}_j)= (\mbf{W}\bm{v}_i)^T \cdot (\mbf{W} \bm{v}_j)$. In our case, $\bm{v}_i$ and $\bm{v}_j$ are deep features normalized to unit length. To learn the embedding $\mbf{W}$ from a given set of triplets $\mathbb{T}$, we solve the following optimization:
\vspace{-2.5mm}
\begin{align*}
\underset{\mbf{W}}{\text{argmin}} \sum_{ (\bm{v}_i,\bm{v}_j,\bm{v}_k) \in \mathbb{T}} - \log(p_{ijk})
\tageq \label{eq:train1}
\end{align*}
(\ref{eq:train1}) can be interpreted as maximizing the likelihood
(\ref{eq:constraint}) or minimizing the negative log-likelihood (NLL) over the
triplet set $\mathbb{T}$. In practice, the above problem is solved in a
Large-Margin framework using Stochastic Gradient Descent (SGD) and the triplets
are sampled online. The gradient update for $\mbf{W}$ is given as:
\begin{align*}
\mbf{W}_{\tau+1} = \mbf{W}_\tau - \eta * \mbf{W}_\tau *(1-p_{ijk})* (\bm{v}_i(\bm{v}_j-\bm{v}_k)^T \\
+ (\bm{v}_j-\bm{v}_k)\bm{v}_i^T)
\tageq \label{eq:update}
\end{align*}
where $\mbf{W}_\tau$ is the estimate at iteration $\tau$, $\mbf{W}_{\tau+1}$ is the
updated estimate, $(\bm{v}_i,\bm{v}_j,\bm{v}_k)$ is the triplet sampled at the current iteration and $\eta$ is the learning rate.\\
By choosing the dimension of $\mbf{W}$ as $n \times N$ with $n < N$, we achieve
dimensionality reduction in addition to improved performance. For our work, we fix
$n=128$ based on cross validation and $N=512$ is the dimensionality of our deep
features. $\mbf{W}$ is initialized with the first $n$ principal components of
the training data. At each iteration, a random anchor and a random positive data
point are chosen. To choose the negative, we perform hard negative mining, ie.
we choose the data point that has the least likelihood (\ref{eq:prob}) among the
randomly chosen 2000 negative instances at each iteration.
Since we compute the embedding matrix $\mbf{W}$ by optimizing over triplet
probabilities, we call this method Triplet Probability Embedding (TPE). The
technique closest to the one presented in this section, which is used in recent
works (\cite{facenet15},\cite{parkhi15}) computes the embedding \mbf{W} based on
satisfying a hinge loss constraint:
\begin{align*}
\underset{\mbf{W}}{\text{argmin}} \sum_{ (\bm{v}_i,\bm{v}_j,\bm{v}_k) \in \mathbb{T}} \max\{0,\alpha + (\bm{v}_i-\bm{v}_j)^T\mbf{W^TW}(\bm{v}_i-\bm{v}_j)- \\
(\bm{v}_i-\bm{v}_k)^T\mbf{W^TW}(\bm{v}_i-\bm{v}_k) \}
\tageq \label{eq:dist}
\end{align*}
$\alpha$ acts a margin parameter for the loss function. To be consistent with
the terminology used in this paper, we call it Triplet Distance Embedding (TDE). To appreciate the difference between the two approaches, Figure \ref{fig:tde} shows the case where the gradient update for the TDE method
(\ref{eq:dist}) occurs. If the value of $\alpha$ is not appropriately chosen, a triplet is considered good even if the positive and negative are very
close to one another. But under the proposed formulation, both cases referred to
in Figure \ref{fig:tde} will update the gradient but their contribution to the
gradient will be modulated by the probability with which they violate the
constraint in (\ref{eq:constraint}). This modulation factor is specified by the
$(1-p_{ijk})$ term in the gradient update for TPE in (\ref{eq:update}) implying
that if the likelihood of a sampled triplet satisfying (\ref{eq:constraint}) is
high, then the gradient update is given a lower weight and vice-versa. Thus, in
our method, the margin parameter ($\alpha$) is automatically set based on the
likelihood. \\
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth,height=0.30\textwidth]{tde_update.png}
\caption{Gradient update scenarios for the TDE method (\ref{eq:dist}). The notation is explained in the text}
\label{fig:tde}
\end{figure}
To compare the relative performances of the raw features before projection, with
TDE and with TPE (proposed method), we plot the traditional ROC curve (TAR (vs)
FAR) for split 1 of the IJB-A verify protocol for the three methods in Figure
\ref{fig:tl}. The Equal Error Rate (EER) metric is specified for each method. The performance
improvement due to TPE is significant, especially at regions of
FAR$=\{10^{-4},10^{-3}\}$. We observed a similar behaviour for all the ten
splits of the IJB-A dataset.
\begin{figure}
\includegraphics[width=0.5\textwidth,height=0.25\textwidth]{roc_compare.jpg}
\caption{Performance improvement on IJB-A split 1: FAR (vs) TAR plot. EER values are specified in brackets.}
\label{fig:tl}
\end{figure}
\section{Clustering faces}
This section illustrates how the proposed TPE method can be used to cluster a given data collection. We utilize the data from the probe set of the split 1 of the IJB-A dataset. We choose a random subset of 30 individuals and 2500 images in total. This provides a rich platform to run quick simulations about the clustering ability of our TPE method. For clustering our data, we use the simple k-means clustering algorithm with multiple restarts with random initialization and choose the one which minimizes sum of squared error. We evaluate the resulting clustering using the F-score which is a popular metric used for retrieval. F-score is computed as:
\begin{equation}
F\mbox{-}score=\frac{2*Precision*Recall}{Precision+Recall}
\end{equation}
We perform two clustering experiments: one based on our raw CNN features and the other using the projected CNN features after TPE. We use cosine distance as the distance measure for k-means and the reported results are averaged over all classes. In both cases, we specify the number of clusters at the start of the k-means algorithm.
\begin{table}
\centering
\begin{tabular}{ | l | l |}
\hline
Method & F-score \\ \hline
CNN+cosine & 0.866 (0.17) \\ \hline
CNN+TPE+cosine & \textbf{0.948 (0.056} \\ \hline
\end{tabular}
\captionof{table}{F-score for comparison of the two clustering schemes. The standard deviation over the classes is specified in brackets}
\label{tab:ami}
\end{table}
\begin{figure*}
\centering
\includegraphics[scale=0.2]{bar_compare.png}
\caption{Comparison of classwise F-scores for the precision experiment}
\label{fig:bar}
\end{figure*}
Figure \ref{fig:bar} shows a classwide comparison of the F-score values. It can be infered from the plot that even though our raw CNN representation gives good performance on most classes, the TPE method improves the performance on classes over which the CNN performs worse (notably over Classes 1,11 aand 12). This small scale experiment illustrates that by being more discriminative in the low dimensional space, our method improves the performance of a simple clustering algorithm.
\section{Clustering faces}
This section illustrates how the proposed TPE method can be used to cluster a given data collection. We utilize the data from the probe set of the split 1 of the IJB-A dataset. We choose a random subset of 30 individuals and 2500 images in total. This provides a rich platform to run quick simulations about the clustering ability of our TPE method. For clustering our data, we use the simple k-means clustering algorithm with multiple restarts with random initialization and choose the one which minimizes sum of squared error. We evaluate the resulting clustering using the F-score which is a popular metric used for retrieval. F-score is computed as:
\begin{equation}
F\mbox{-}score=\frac{2*Precision*Recall}{Precision+Recall}
\end{equation}
We perform two clustering experiments: one based on our raw CNN features and the other using the projected CNN features after TPE. We use cosine distance as the distance measure for k-means and the reported results are averaged over all classes. In both cases, we specify the number of clusters at the start of the k-means algorithm.
\begin{table}
\centering
\begin{tabular}{ | l | l |}
\hline
Method & F-score \\ \hline
CNN+cosine & 0.866 (0.17) \\ \hline
CNN+TPE+cosine & \textbf{0.948 (0.056} \\ \hline
\end{tabular}
\captionof{table}{F-score for comparison of the two clustering schemes. The standard deviation over the classes is specified in brackets}
\label{tab:ami}
\end{table}
\begin{figure*}
\centering
\includegraphics[scale=0.2]{bar_compare.png}
\caption{Comparison of classwise F-scores for the precision experiment}
\label{fig:bar}
\end{figure*}
Figure \ref{fig:bar} shows a classwide comparison of the F-score values. It can be infered from the plot that even though our raw CNN representation gives good performance on most classes, the TPE method improves the performance on classes over which the CNN performs worse (notably over Classes 1,11 aand 12). This small scale experiment illustrates that by being more discriminative in the low dimensional space, our method improves the performance of a simple clustering algorithm.
|
\section{Introduction}
\label{sec:intro}
The electric dipole polarizability $\alpha_D$ in nuclei has been
subject of intense studies, both from the experimental and the
theoretical side. Photo-absorption studies have focused on the
determination of the giant dipole resonances (GDR) in stable nuclei,
originally interpreted as a collective motion of all protons
oscillating against all neutrons~\cite{berman1975}. The discovery of a soft peak at low
energies in neutron-rich and unstable nuclei, i.e. the pygmy dipole
resonance (PDR), has spurred a renewed interest in the electric dipole
response~\cite{kobayashi1989}. For a recent review, we refer the
reader to Ref.~\cite{aumann2013}.
Calculations based on relativistic and non-relativistic density-functional theory pointed out that $\alpha_D$ is very strongly
correlated with the neutron-skin
thickness~\cite{Reinhard2010,Piekarewicz12,Roca-Maza2015}. This can
be contrasted to {\it ab initio} computations based on Hamiltonians
from chiral effective field theory (EFT) that rather found a strong
correlation between the charge and the neutron radii with $\alpha_D$ in $^{48}$Ca~\cite{hagen2015}. In any
case, the dipole polarizability is sensitive to the neutron
distribution, and thereby constrains the neutron equation of state and
the physics of neutron
stars~\cite{brown2000,furnstahl2002,Tsang2012,Hebeler2014}. The
equation of state of asymmetric nuclear matter depends on a few
parameters, such as the slope of the symmetry energy, which correlates
with GDR~\cite{Trippa2008} and PDR~\cite{Carbone2010} features.
Recent experiments measured the dipole polarizability in
$^{208}$Pb~\cite{Tamii2010}, $^{68}$Ni~\cite{Rossi2013}, and
$^{120}$Sn~\cite{krumbholz2015,Hashimoto2015}, and data for $^{48}$Ca
is presently being analyzed by the Darmstadt-Osaka collaboration.
Only scarce data exist on unstable nuclei, but recent activity was
devoted, {\it e.g.}, to $^{22,24}$O~\cite{Leistenschneider2001}.
The dipole polarizability
\begin{equation}\label{polresp}
\alpha_D = 2\alpha \int_{\omega_{ex}}^{\infty} d\omega~\frac{R(\omega)}{\omega}\,,
\end{equation}
where $\alpha$ is the fine structure constant, is an inverse energy weighted sum rule of the dipole response function
$R(\omega)$. Thus, the determination of the low-energy dipole strength
is crucial. Here $\omega$ is the excitation energy and $\omega_{ex}$
is the energy of the first state excited by the dipole referred to the
ground-state. Within one isotopic chain one expects that neutron-rich
nuclei with a significant low-lying dipole strength also exhibit a
larger polarizability than other isotopes. To both interpret recent
data and guide new experiments, it is important to theoretically map
the evolution of $\alpha_D$ as a function of neutron number. Theories
that can reliably address exotic nuclei far from the valley of
stability are needed and \textit{ab initio} methods are best positioned to
deliver both predictive power~\cite{BaccaPastore2014,Roth2014,Quaglioni16} and estimates of
the theoretical uncertainties~\cite{Carlsson16,Wesolowski15,Furnstahl15,Binder15}.
This paper is organized as follows. Section \ref{sec:th} describes the
theoretical approach based on integral transforms and the
coupled-cluster method. In Section~\ref{sec:3nf_calc} we present
results for the nuclei $^4$He, $^{16,22}$O and $^{40}$Ca. First, we
compare different computational approaches with each other. Second,
we present results for the dipole polarizability in these nuclei based
on an interaction from chiral EFT that exhibits accurate saturation
properties~\cite{Ekstroem15}. Third, we study correlations between
the dipole polarizability and charge radii based on a variety of
nucleon-nucleon (NN) interactions and interactions that also
include three-nucleon forces (3NFs). Finally, we summarize our
results in Sect.~\ref{sec:conclusions}.
\section{Theoretical approach}
\label{sec:th}
The electric dipole polarizability in
Eq.~(\ref{polresp}) depends on the dipole response function
\begin{equation}\begin{split}\label{resp} R(\omega) = \sumint_f &\langle \Psi_0
|\hat{\Theta}^\dag |\Psi_f\rangle\langle
\Psi_f|\hat{\Theta}|\Psi_0\rangle\delta(E_f - E_0 - \omega) .
\end{split}\end{equation}
Here $\hat{\Theta}=\sum_{i=1}^AP_i (z_i-Z_{cm})$ is the dipole excitation operator,
where $P_i$ is the proton projection operator and $z_i$/$Z_{cm}$ the nucleon/center of mass z-coordinate, respectively.
$|\Psi_0\rangle$ is the ground state of the nucleus and
$|\Psi_f\rangle$ represents the excited states. The latter can be both
in the discrete and in the continuum region of spectrum, and this is
reflected by the combined discrete and continuum symbol
$\sumint_f$\footnote{For simplicity, in this notation we omit the average on projections of the initial angular momentum}. From Eqs.~(\ref{polresp}) and (\ref{resp}) it is clear
that the dipole polarizability contains the information on $R(\omega)$
at all energies $\omega$, including those in the continuum. A
calculation of $\alpha_D$ would then require to be able to solve the
many-body scattering problem at such energies, which is extremely
difficult for nuclei with mass number larger than four.
To make progress, we rewrite $\alpha_D$ as a sum rule of the response
function. Starting from Eq.~(\ref{polresp}) and using the
completeness of the Hamiltonian eigenstates
$\mathbb{I}=\sumint_{f}|\Psi_f\rangle\langle\Psi_f|$ we obtain
\begin{equation}
\label{sumrule}
\alpha_D =\langle \Psi_0|\hat{\Theta}^\dag
\frac{1}{\hat{H}-E_0} \hat{\Theta} |\Psi_0\rangle\,.
\end{equation}
One way to calculate $\alpha_D$ by means of the sum rule in
Eq.~(\ref{sumrule}) is to represent the Hamiltonian on a finite basis of
$N$ basis functions $|n\rangle$. After diagonalization of the
Hamiltonian matrix $H_{n,n'}$, one obtains its $N$ eigenstates
$|\beta\rangle$ and eigenvalues $E_\beta$, and Eq.~(\ref{sumrule})
becomes
\begin{equation}
\label{sumrule_comp} \alpha_D =\sum_{\beta}^{N} \langle
\Psi_0|\hat{\Theta}^\dag | \beta \rangle \langle \beta
|\frac{1}{E_\beta-E_0} |\beta \rangle \langle \beta | \hat{\Theta}
|\Psi_0\rangle\,.
\end{equation}
Increasing $N$ yields an increasingly more accurate representation of
the eigenfunctions $|\beta\rangle$ and eigenvalues $E_\beta$ of
$\hat{H}$, and eventually the value of $\alpha_D$ would converge. In
practical cases, however, the truncated basis states $|n\rangle$ used
to represent the Hamiltonian are discrete and have a finite
norm. Because the spectrum of $\hat{H}$ has both a
discrete and a continuum part, one may question the use of such a
discrete basis. Similarly to Ref.~\cite{LSR}, we will show that this approach is rigorous and works quite well also within coupled-cluster theory.
\subsection{Integral transforms}
\label{subsec:itf}
Integral transforms reduce the continuum problem of calculating
$R(\omega)$ to the solution of a bound-state-like
problem~\cite{Efros85,Efros94,Efl07}. In such an approach, one first
calculates the integral transform ${\mathcal I}(\sigma)$ of the
response function. In a second step, one might invert the integral
transform to obtain the response function $R(\omega)$, or one might
compute relevant observables (such as the dipole polarizability)
directly from the integral transform. Here, we will use the Stieltjes
integral transform \cite{Efros93} for the direct computation of the
dipole polarizability.
The Stieltjes integral transform reads
\begin{equation}\label{stieltjes}
\mathcal{I}(\sigma) =\int\frac{R(\omega)}{\omega+\sigma}d\omega,
\end{equation}
with $\sigma$ real and positive. Using the completeness on the
Hamiltonian eigenstates and the definition of the response function from
Eq.~(\ref{resp}) yields
\begin{equation}
\label{stieltjes2}\begin{split}
\mathcal{I}(\sigma) &=\langle \Psi_0|\hat{\Theta}^\dag
\frac{1}{\hat{H}-E_0+\sigma} \hat{\Theta} |\Psi_0\rangle\\ &=
\langle\Psi_0|\hat{\Theta}^\dag|\tilde{\Psi}(\sigma)\rangle,
\end{split}
\end{equation}
where we have defined
\begin{equation}\label{psitilde}
|\tilde \Psi(\sigma)\rangle\equiv \frac{1}{\hat{H}-E_0+\sigma} \hat{\Theta} |\Psi_0\rangle.
\end{equation}
The function $|\tilde \Psi(\sigma)\rangle$ is the solution of the
following Schr\"odinger-like equation with a source
\begin{equation}
\label{Schr_like_eq} (\hat{H}-E_0+\sigma)|\tilde
\Psi(\sigma)\rangle = \hat{\Theta} |\Psi_0\rangle.
\end{equation}
Since $\sigma >0$, and for large inter particle distances
$|\Psi_0\rangle \rightarrow 0$, one has that asymptotically -- and for
non singular operators $\hat{\Theta}$ -- $|\tilde \Psi(\sigma)\rangle$
should satisfy a Schr\"odinger equation with eigenvalues smaller than
$E_0$. This implies that $|\tilde \Psi(\sigma)\rangle \rightarrow 0$
asymptotically, namely it has bound state-like asymptotic
conditions. We are therefore allowed to calculate
$\mathcal{I}(\sigma)$ using a bound-state basis expansion, \textit{i.e.} an $L_2$ square integrable basis such as harmonic oscillator functions. Noticing
that Eq.~(\ref{stieltjes2}) differs from Eq.~(\ref{sumrule}) only by
the presence of $\sigma >0$, we proceed as it was described above,
namely using a representation on a bound state basis and increasing
the number $N$ of basis functions up to convergence. Then the value
of $\alpha_D$ can be obtained as
\begin{equation}
\label{polstil}
\alpha_D=2\alpha\lim_{\sigma\to 0^+}\mathcal{I}(\sigma)\,,
\end{equation}
avoiding the continuum problem. The limit taken with positive $\sigma$
is crucial not only to allow the use of a bound state basis, but also
because it avoids poles (we recall that $E_0$ is negative).
For $\sigma<0$ poles will certainly be present, presumably at different places depending on the basis. We indeed observe several poles in the region of $\sigma<0$, while the curve is persistently smooth for $\sigma\geq 0$ in Fig.~\ref{fig:stieltjes}, where we show $\mathcal{I}(\sigma)$ for $^4$He calculated with a realistic interaction~\cite{Ekstroem15}, as detailed later. We choose $^4$He, where calculations are faster and can be benchmarked with few-body methods.
\begin{figure}[!tb]
\includegraphics[width=\linewidth,clip=]{4He_stieltjes_n2losat_noinset.pdf}
\caption{(Color online) The Stieltjes integral transform
$\mathcal{I}(\sigma)$ as a function of $\sigma$ in the case of
$^{4}$He.}
\label{fig:stieltjes}
\end{figure}
Below we will use an implementation of Eq.~(\ref{polstil}) to compute
the dipole polarizability of heavier nuclei. To test this approach, we will also compare it to $\alpha_D$ obtained by the dipole response function as in Eq.~(\ref{polresp}). If one were able to invert the Stieltjes transform, one could obtain $R(\omega)$ to calculate the
integral in Eq.~(\ref{polresp}). Unfortunately, the inversion of this integral transform presents the typical difficulties of an ill-posed problem. In fact in Ref.~\cite{Efros93} it was shown that inversions performed with the regularization method~\cite{Tikhonov} generate
quite different responses, all compatible with the same Stieltjes transform within numerical errors.
Therefore, we will employ the much more suitable Lorentz integral transform (LIT)~\cite{Efros94,Efl07}
\begin{equation}
\label{litdef}
L(\sigma,\Gamma) = \frac{\Gamma}{\pi}\int\frac{R(\omega)}{(\omega -
\sigma)^2 + \Gamma^2}d\omega,
\end{equation}
where $\Gamma,\sigma \in \mathbb{R}$ and $\Gamma >
0$. The Lorentzian kernel $L(\sigma,\Gamma)$ is peaked at $\sigma$ and
has the width $\Gamma$. The LIT can be much more easily inverted to yield the response function, because the width $\Gamma$ introduces a finite resolution.
Thus, the response function is smeared only in a narrow region of space determined by the width $\Gamma$.
The calculation the Lorentz transform proceeds as for the Stieltjes transform, using the definition of response function and the
completeness of the eigenfunctions of the Hamiltonian. One finds
\begin{equation}
\label{lit}
L(z) = \frac{\Gamma}{\pi}\langle\Psi_0|\hat{\Theta}^\dag\frac{1}{(\hat{H} -z^*)}\frac{1}{(\hat{H} -z)}\hat{\Theta}|\Psi_0\rangle,
\end{equation}
with $z = E_0 + \sigma +i\Gamma$. The LIT can be rewritten in a form
that resembles Eq.~(\ref{stieltjes2}) as
\begin{equation}
\label{lit2}\begin{split}
L(z) &= \frac{1}{\pi}\mathfrak{Im}\left[\langle\Psi_0|\hat{\Theta}^\dag\frac{1}{(\hat{H} - z)}\hat{\Theta}|\Psi_0\rangle\right]\\
&=\frac{1}{\pi}\mathfrak{Im}\left[\langle\Psi_0|\hat{\Theta}^\dag|\tilde\Psi(z)\rangle\right]\,.
\end{split}
\end{equation}
Here we defined the function
\begin{equation}
\label{shrodlit}
|\tilde\Psi(z)\rangle \equiv \frac{1}{\hat{H}- z} \hat{\Theta}|\Psi_0\rangle.
\end{equation}
Similarly as for Eq.~(\ref{psitilde}), $|\tilde{\Psi}(z)\rangle$ has a
bound-state-like nature and a finite norm
\begin{equation}\label{norm}
\langle\tilde{\Psi}(z)|\tilde{\Psi}(z)\rangle=L(z)=
\frac{\Gamma}{\pi}\int\frac{R(\omega)}{(\omega - \sigma)^2 + \Gamma^2}d\omega <\infty.
\end{equation}
A couple of remarks are in order here. First, we note that the positive
parameter $\sigma$ enters in the Stieltjes and Lorentz transforms with
a minus and a plus sign respectively. While in
the Stieltjes transform the bound-state-like nature of
$|\tilde{\Psi}\rangle$ is due to that minus sign, in the Lorentz case
it is due to the presence of the imaginary part $\Gamma$. Second,
in the limit $\Gamma \rightarrow 0$
the Lorentzian kernel becomes a delta function
\begin{equation}\label{litdelta}
L(\sigma,\Gamma\to 0) = \int R(\omega)\delta(\omega - \sigma)d\omega=R(\sigma).
\end{equation}
This allows us to estimate the dipole polarizability also using
Eq.~(\ref{litdelta}) together with Eq.~(\ref{polresp})
\begin{equation}\label{poldelta}
\alpha_D = 2\alpha\int \frac{L(\sigma,\Gamma\to 0)}{\sigma}d\sigma.
\end{equation}
However, in L($\sigma$,$\Gamma$) one must be careful in taking smaller and smaller $\Gamma$ since the convergence in the model space expansion becomes increasingly difficult.
\subsection{Coupled-cluster implementation}
\label{subsec:lanzos}
In this Subsection we will compute the dipole polarizability via
Eq.~(\ref{polstil}) with the coupled-cluster method. This
calculation proceeds similarly as done for the LIT in
Refs.~\cite{Bacca13,LITCC}.
Coupled-cluster
theory~\cite{coester1958,coester1960,kuemmel1978,mihaila2000b,dean2004,wloch2005,hagen2010b,binder2013b}
is based on the exponential ansatz for the ground state
\begin{equation}\label{expansatz}
|\Psi_0\rangle = e^{\hat{T}}|0_R\rangle,
\end{equation}
see Refs.~\cite{bartlett2007,hagen2014} for recent reviews. Here,
$|0_R\rangle$ is a reference product state, and the cluster operator
$T$ introduces particle-hole (p-h) excitations into the
reference. Using second quantization, and normal ordering the dipole
excitation operator with respect to the reference state yields the
response function~\cite{LITCC}
\begin{equation}\label{noresp} R(\omega) = \sum_n \langle
0_L|\overline{\Theta}^\dag|n_R\rangle\langle
n_L|\overline{\Theta}|0_R\rangle\delta(\Delta E_n - \Delta E_0 -
\omega).
\end{equation}
Here $\Delta E_n$, $\Delta E_0$ are the correlation
energies of the $n$th-excited state and ground-state respectively,
and solve
\begin{equation}
\begin{split}\label{CCenergy}
\overline{H}|0_R\rangle = \Delta
E_0|0_R\rangle\ \ \ \ &\text{or}\ \ \ \ \langle 0_L|\overline{H} =
\langle 0_L|\Delta E_0,\\ \overline{H}|n_R\rangle = \Delta
E_n|n_R\rangle\ \ \ \ &\text{or}\ \ \ \ \langle n_L|\overline{H} =
\langle n_L|\Delta E_n.
\end{split}
\end{equation}
Here we used similarity-transformed operators via
\begin{equation}\label{simtransf}
\overline{O} = e^{-\hat{T}}\hat{O}_Ne^{+\hat{T}},
\end{equation}
and $\hat{O}_N$ is the normal-ordered form of any
operator $\hat{O}$, \textit{e.g.} $\hat{H}$ or $\hat{\Theta}$. Substituting Eq.~(\ref{noresp}) in Eq.~(\ref{stieltjes}),
and making use of the expressions in Eq.~(\ref{CCenergy}) yields
\begin{equation}
\label{CCstieltjes} \mathcal{I}(\sigma) = \langle
0_L|\overline{\Theta}^\dag\frac{1}{\overline{H} - \Delta E_0 +
\sigma}\overline{\Theta}|0_R\rangle.
\end{equation}
This equation resembles Eq.~(\ref{stieltjes2}), when
operators are replaced by their similarity transformed
counterparts, and one needs to distinguish between left and right states because
of the non-Hermitian nature of the excitation operator $\hat{T}$. We
proceed as in Subsection~\ref{subsec:itf}, and define a state
$|\tilde{\Psi}(\sigma)\rangle$ as the solution of
\begin{equation}\label{LITCC_eq}
(\overline{H} - \Delta E_0 + \sigma)|\tilde{\Psi}_R(\sigma)\rangle =
\overline{\Theta}|0_R\rangle.
\end{equation}
Eq.~(\ref{LITCC_eq}) resembles Eq.~(15) in \cite{LITCC} and can be
solved using the equation-of-motion coupled-cluster method for excited
states \cite{Stanton93}. In this approach, one regards
\begin{equation}\begin{split}\label{cceom1}
|\tilde{\Psi}_R(\sigma)\rangle =
&\hat{\mathcal{R}}(\sigma)|0_R\rangle\equiv \left(r_0(\sigma) +
\sum_{i,a}r^a_i(\sigma) \hat{c}^\dag_a \hat{c}_i +\right.\\ &\left. +
\frac{1}{4}\sum_{i,j,a,b}r^{ab}_{ij}(\sigma)\hat{c}_a^\dag\hat{c}^\dag_b\hat{c}_j\hat{c}_i
+ ...\right)|0_R\rangle\\ \equiv &\sum_\alpha\hat{C}_\alpha
r_\alpha(\sigma)|0_R\rangle\equiv
\hat{\mathbf{C}}\cdot\mathbf{r}(\sigma)|0_R\rangle,
\end{split}
\end{equation}
as an excited state of the similarity-transformed Hamiltonian $\overline{H}$ based on
p-h excitations of the reference. In the last line of Eq.~(\ref{cceom1})
the index $\alpha$ labels the $0$p-$0$h,
$1$p-$1$h, $2$p-$2$h, ... states
\begin{equation}\label{states}
|\Phi_\alpha\rangle \equiv |0_R\rangle, |\Phi_i^a\rangle, |\Phi^{ab}_{ij}\rangle, \ldots .
\end{equation}
We also defined the column vector $\mathbf{r}(\sigma)$ with elements
$r_0(\sigma),r^a_i(\sigma), r^{ab}_{ij}(\sigma), ...$ and a row
vector $\hat{\mathbf{C}}$ whose elements are strings of normal-ordered
creation and annihilation operators. Combining Eq.~(\ref{LITCC_eq}) with
Eq.~(\ref{CCenergy}) and the linear ansatz for
$|\tilde{\Psi}(\sigma)\rangle$, the Stieltjes transform becomes
\begin{equation}\label{matrix_eq}
\mathcal{I}(\sigma) = \langle 0_L|\overline{\Theta}^\dag\hat{\mathcal{R}}(\sigma)|0_R\rangle = \mathbf{S}^L\mathbf{M}(\sigma)^{-1}\mathbf{S}^R .
\end{equation}
Here $\mathbf{S}^L$ and $\mathbf{S}^R$ are row- and column-vectors
respectively with elements
\begin{equation}
\begin{split}
&S^R_\alpha = \langle\Phi_\alpha|\overline{\Theta}|0_R\rangle,\\
&S^L_\alpha = \langle0_L|\overline{\Theta}^\dag|\Phi_\alpha\rangle,
\end{split}
\end{equation}
and $\mathbf{M}$ is a matrix with elements
\begin{equation}
M_{\alpha\beta}(\sigma) = \langle \Phi_{\alpha} |\left[\overline{H},\hat{C}_\beta\right]|0_R\rangle + \sigma\delta_{\alpha\beta}.
\end{equation}
The right-hand side of Eq.~(\ref{matrix_eq}) can be calculated using the Lanczos
procedure. Because we are dealing with non-Hermitian operators, we
have to make use of the generalized Lanczos algorithm for
non-symmetric matrices \cite{Cullum98}. In this approach, one first
defines two pivot vectors
\begin{equation}\label{vecdef}\begin{split}
&\mathbf{v}_0 = \frac{\mathbf{S}^R}{\sqrt{\mathbf{S}^L\cdot\mathbf{S}^R}},\\
&\mathbf{w}_0 = \frac{\mathbf{S}^L}{\sqrt{\mathbf{S}^L\cdot\mathbf{S}^R}},
\end{split}
\end{equation}
and repeated application of the matrix $\mathbf{M}(\sigma)$ generates the Lanczos basis
in which $\mathbf{M}$ is tri-diagonal
\begin{equation}\label{trimat}
\mathbf{M}(\sigma)=
\begin{pmatrix}
a_0 - \sigma& b_0 & 0 & 0 & \cdots\\
b_0 & a_1 -\sigma & b_1 & 0 & \cdots\\
0 & b_1 & a_2 - \sigma & b_2 & \cdots\\
0 & 0 & b_2 & a_3 - \sigma & \cdots\\
\vdots & \vdots & \vdots & \vdots & \ddots
\end{pmatrix}.
\end{equation}
In what follows, we employ the matrix $\mathbf{M}$ in the Lanczos basis.
We note that $\mathbf{S}^L\cdot\mathbf{S}^R = \langle
0_L|\overline{\Theta}^\dag\overline{\Theta}|0_R\rangle$ and find the expression
\begin{equation}\label{stilcontfrac}
\mathcal{I}(\sigma) = \langle 0_L|\overline{\Theta}^\dag\overline{\Theta}|0_R\rangle x_{00}(\sigma),
\end{equation}
for the Stieltjes integral transform. Here
\begin{equation}\label{contfrac}
x_{00}(\sigma) = \mathbf{w}_0\left[\mathbf{M}(\sigma)\right]^{-1}\mathbf{v}_0.
\end{equation}
From the identity $\mathbb{I} =
\mathbf{M}(\sigma)[\mathbf{M}(\sigma)]^{-1}$, one finds the linear
system
\begin{equation}
\sum_\beta {M}_{\alpha\beta}(\sigma)x_{\beta 0}(\sigma) = \delta_{\alpha 0},
\end{equation}
where we defined $x_{\beta 0}(\sigma) =
[\mathbf{M}(\sigma)^{-1}]_{\beta 0}$. Using Cramer's rule to solve the
linear system, we find that $x_{00}(\sigma)$ is given by the continued
fraction
\begin{equation}\label{contfrac2}
x_{00}(\sigma) = \frac{1}{(a_0 - \sigma) - \frac{b_0^2}{(a_1 - \sigma) - \frac{b_1^2}{(a_2 - \sigma) - \cdots}}},
\end{equation}
and finally Eq.~(\ref{stilcontfrac}) becomes
\begin{equation}\label{stilcontfrac2}
\mathcal{I}(\sigma) = \langle 0_L|\overline{\Theta}^\dag\overline{\Theta}|0_R\rangle\left\{\frac{1}{(a_0 - \sigma) - \frac{b_0^2}{(a_1 - \sigma) - \frac{b_1^2}{(a_2 - \sigma) - \cdots}}}\right\}.
\end{equation}
Then, from Eq.~(\ref{polstil}), one finds that the electric dipole
polarizability is the continued fraction
\begin{equation}\label{polstilCC}
\alpha_D =2\alpha\langle 0_L|\overline{\Theta}^\dag\overline{\Theta}|0_R\rangle\lim_{\sigma\to 0^+}\left\{\frac{1}{(a_0 + \sigma) - \frac{b_0^2}{(a_1 + \sigma) - \frac{b_1^2}{(a_2 + \sigma) - \cdots}}}\right\} ,
\end{equation}
which is equivalent to the Lanczos sum rule of Ref.~\cite{LSR}. We note that Eq.~(\ref{polstilCC}) is an exact result if the operators $\hat{T}$ and $\hat{\mathcal{R}}$ are expanded up to $A$p-$A$h excitations in a nucleus with mass number $A$. However, in practical calculations $\hat{T}$ and $\hat{\mathcal{R}}$ are truncated since a full expansion is not feasible due to the very high computational cost. In this paper we truncate $\hat{T}$ and $\hat{\mathcal{R}}$ at singles-and-doubles excitations. However, we remind the reader that such a truncation includes exponentiated $1$p-$1$h and $2$p-$2$h excitations. The exponent yields also products of higher order. As the GDR consists of a superposition of $1$p-$1$h excitations, a truncation at singles-and-doubles only is expected to be a good approximation. Similarly, the dipole polarizability is most sensitive to the GDR.
Summarizing, we presented three different methods to evaluate the electric dipole polarizability: (i) compute the LIT for the dipole response, obtain $R(\omega)$ from its inversion -- with inversions performed as described in Ref.~\cite{Bacca13,Andreasi05,Efros99} -- and compute the dipole polarizability from Eq.~(\ref{polresp}); (ii) use Eq.~(\ref{poldelta}) for $\Gamma\to 0$. (iii) use the continued fraction as in Eq.~(\ref{polstilCC}). The second method is in principle a discretization of the continuum and it will be interesting to compare it with the other two methods.
\section{Results}
\label{sec:3nf_calc}
In Ref.~\cite{Bacca13,LITCC} coupled-cluster results for the dipole
response in $^{4}$He were benchmarked against virtually exact results
from the effective interaction hyperspherical harmonics~\cite{EIHH,Goerke12} method.
Those calculations were based on NN forces~\cite{Entem03} from chiral EFT. In this paper we augment the Hamiltonians to include 3NFs from chiral EFT.
First, we check the convergence of our results with respect to model-space parameters and compare the three different calculational approaches for the dipole polarizability using the NNLO$_{\rm sat}$ interaction~\cite{Ekstroem15}.
Then, we compare to experimental data for $^4$He, $^{16}$O and $^{22}$O. Finally, we explore correlations of the dipole polarizability with the charge radius in $^{16}$O and $^{40}$Ca by employing a variety of Hamiltonians. In addition to NNLO$_{\rm{sat}}$ and the family of interactions from Ref.~\cite{Hebeler11}, we also use a large set of realistic NN potentials~\cite{CDBONN,Entem03,bogner2007,bogner2003,AV18} to probe systematic uncertainties in the underlying Hamiltonians.
When adding 3NFs, we use a Hartree-Fock basis built on 15 major harmonic oscillator shells. We vary the model space size up to $N_{\rm max}=14$ and we truncate the 3NFs matrix elements at $E_{\rm 3max}=N_{\rm max}$ for $^4$He and $^{16,22}$O. For our purposes, this truncation provides well-enough converged results. In fact, for the more challenging neutron-rich $^{22}$O nucleus, increasing $E_{\rm 3max}$ to 16 leads to a variation in energy of only 400~keV. For the calculations in $^{40}$Ca with the NNLO$_{\rm sat}$ interaction and the Hamiltonians from Ref.~\cite{Hebeler11} we employed the same $N_{\rm max}$ and $E_{\rm 3max}$ truncations used for $^{48}$Ca by ~\citeauthor{hagen2015} in Ref.~\cite{hagen2015}.
\subsection{The $^4$He nucleus}
\label{subsec:4He}
\begin{figure}[!tb]
\includegraphics[width=\linewidth,clip=]{4He_pol_conv_n2losat.pdf}
\caption{(Color online) The electric dipole polarizability in $^{4}$He
as a function of the model space size $N_{\rm max}$. Curves for
different values of $\hbar\Omega$, the underlying harmonic
oscillator frequency, are shown.}
\label{fig:fig_pol_conv_4He}
\end{figure}
Figure~\ref{fig:fig_pol_conv_4He} shows the electric dipole
polarizability of $^4$He obtained from the continued fraction of
Eq.(\ref{polstilCC}) with the NNLO$_{\rm{sat}}$ interaction, as a
function of the model space size $N_{\rm max}$. The four curves represent
calculations with different values of oscillator frequency
$\hbar\Omega$. The convergence in $N_{\rm max}$ is excellent and
independence on $\hbar \Omega$ is reached with $N_{\rm max}=14$. The uncertainty
at $N_{\rm max}=14$ for the different values of $\hbar\Omega$ is about $0.1\%$.
\begin{figure}[!tb]
\includegraphics[width=\linewidth,clip=]{4He_polVSen_n2losat.pdf}
\caption{(Color online) The electric dipole polarizability
$\alpha_D(\varepsilon)$ in $^{4}$He as a function of the integration
energy $\varepsilon$: (i) using the LIT and Eq.~(\ref{polresp}) in
blue (band); (ii) using Eq.~(\ref{poldelta}) in red (solid); (iii)
using the continued fraction of Eq.~(\ref{polstilCC}) in black
(dashed). Calculations are performed for $\hbar \Omega=22$ MeV and
$N_{\rm max}=14$.}
\label{fig:fig_pol_vs_e_4He}
\end{figure}
Let us compare the three different ways to calculate the dipole
polarizability for $^4$He as described at the end of
Section~\ref{subsec:itf}. Equations~(\ref{polresp}) and
(\ref{poldelta}) require an integration in energy and we present
$\alpha_D(\varepsilon)$ where $\varepsilon$ is the upper limit of the
integration. Figure~\ref{fig:fig_pol_vs_e_4He} shows the results.
The blue band shows method (i), i.e. $\alpha_D$ is obtained from
integrating Eq.~(\ref{polresp}), and $R(\omega)$ stems from an
inversion of the LIT. Here, the width of the blue band is an estimate
of the uncertainty involved in the inversion procedure. The red solid
line shows method (ii), i.e. $\alpha_D$ obtained from the LIT at
small $\Gamma$ using Eq.~(\ref{poldelta}). The black dashed line shows
the method (iii), i.e. $\alpha_D$ obtained using the continued
fraction in Eq.~(\ref{polstilCC}).
We note that the different methods yield the same dipole
polarizability. The integration methods (i) and (ii) exhibit a
similar dependence on the integration range, the difference being that
the former is smooth while the latter increases in steps. Here, method
(ii) has the advantage of a sharper defined threshold. We also note
that the dependence on the integration range is useful for comparisons
with data for experiments that probe only a limited region of the
energy spectrum.
\subsection{The $^{16}$O nucleus}
\label{sec:16O}
Figure~\ref{fig:fig_pol_conv_16O} shows the electric dipole
polarizability in $^{16}$O as a function of the model space size calculated with
the NNLO$_{\rm{sat}}$ interaction, while Figure~\ref{fig:fig_rch_conv_16O} shows the same for the charge radius, which has been obtained from the point-proton radius taking into account contributions from nucleonic charge radii (see Ref.~\cite{hagen2015} for details). We observe that the curves for different $\hbar\Omega$ values converge very nicely and only a small residual $\hbar \Omega$-dependence remains at the largest model space size $N_{\rm max}=14$. Based on the spread of the different $\hbar\Omega$ curves for $N_{\rm max}=14$, we obtain a conservative error of $2.8\%$ for the electric dipole polarizability and a conservative error of $0.7\%$ for the charge radius.
\begin{figure}[!ht]
\includegraphics[width=\linewidth,clip=]{16O_pol_conv_n2losat.pdf}
\caption{(Color online) Electric dipole polarizability in $^{16}$O as
a function of the model space size $N_{\rm max}$ for different values of
$\hbar\Omega$.}
\label{fig:fig_pol_conv_16O}
\end{figure}
\begin{figure}[!ht]
\includegraphics[width=\linewidth,clip=]{16O_rch_conv_n2losat_new.pdf}
\caption{(Color online) Charge radius in $^{16}$O as
a function of the model space size $N_{\rm max}$ for different values of
$\hbar\Omega$.}
\label{fig:fig_rch_conv_16O}
\end{figure}
Figure~\ref{fig:fig_polVSen_n2losat_16O} compares the results from the
three methods to obtain the polarizability for $^{16}$O. The blue band (i) shows the integration as
in Eq.~(\ref{polresp}) of the weighted response function, and the
width of the band takes into account the uncertainty of the inversion.
The red solid line (ii) refers to the integration of the weighted LIT
with Eq.~(\ref{poldelta}). The black dashed line (iii) is the
reference value calculated with the continued fraction using
Eq.~(\ref{polstilCC}). Again, we find good agreement of the results
for the dipole polarizability.
\begin{figure}[!ht]
\includegraphics[width=\linewidth,clip=]{16O_polVSen_n2losat.pdf}
\caption{(Color online) The electric dipole polarizability
$\alpha_D(\varepsilon)$ in $^{16}$O as a function of the integration
limit $\varepsilon$. The blue band (i) is obtained integrating the
weighted response function as in Eq.~(\ref{polresp}); the red solid
curve (ii) is calculated integrating the weighted LIT at small
$\Gamma$ as in Eq.~(\ref{poldelta}); the black dashed line (iii) is
obtained from the continued fraction of
Eq.~(\ref{polstilCC}). Calculations are performed with $N_{\rm
max}=14$ and $\hbar \Omega=22$ MeV.}
\label{fig:fig_polVSen_n2losat_16O}
\end{figure}
\subsection{The $^{22}$O nucleus}
\label{sec:22O}
\begin{figure}[!ht]
\includegraphics[width=\linewidth,clip=]{22O_pol_conv_n2losat.pdf}
\caption{(Color online) The electric dipole polarizability $\alpha_D$ in $^{22}$O as a function of the model space size $N_{\rm max}$. Different curves for different values of the underlying harmonic oscillator basis frequency $\hbar\Omega$ are shown. }
\label{fig:fig_pol_conv_22O}
\end{figure}
The dipole strength of the neutron-rich nucleus $^{22}$O was measured by~\citeauthor{Leistenschneider2001}~\cite{ Leistenschneider2001} via Coulomb excitation in experiments at GSI.
Figure~\ref{fig:fig_pol_conv_22O} shows the electric dipole polarizability as a function of the model space size of the
calculation.
After having investigated various frequencies, we find that $\hbar\Omega=18$ MeV is the best converging curve. However, the convergence for different $\hbar\Omega$ is slower than what observed in lighter nuclei, resulting in a conservative uncertainty of about $8\%$ at $N_{\rm max}=14$. This might be because the excess neutrons in $^{22}$O are loosely bound, making the wave function more extended and thus the convergence slower.
We note that $\alpha_D$ of $^{22}$O is larger than for $^{16}$O.
\begin{figure}[!ht]
\includegraphics[width=\linewidth,clip=]{22O_polVSen_n2losat.pdf}
\caption{(Color online) The electric dipole polarizability
$\alpha_D(\varepsilon)$ in $^{22}$O as a function of the integration
limit $\varepsilon$. The red solid curve (ii) is calculated
integrating the weighted LIT at small $\Gamma$ as in
Eq.~(\ref{poldelta}); the black dashed line (iii) is obtained from
the continued fraction of Eq.~(\ref{polstilCC}). Calculations are
performed with $N_{\rm max}=14$ and $\hbar \Omega=18$ MeV.}
\label{fig:fig_polVSen_n2losat_22O}
\end{figure}
Finally, in Figure~\ref{fig:fig_polVSen_n2losat_22O} we show a
comparison between the methods (ii) and (iii) to calculate
$\alpha_D$. We used the largest model space and the fastest converging
frequency of $\hbar \Omega=18$~MeV and find good agreement between the
two methods. Because the convergence of the LIT calculations is not
at sub-percent level, we cannot presently obtain stable inversions and
include the method (i) in the comparison. Nevertheless, by looking at
the laddered curve we learn about the convergence of this sum rule as
a function of the energy. This will be used in the following
Subsection to make a comparison with the experimental data from
Ref.~\cite{Leistenschneider2001}.
\subsection{Comparison to experiment}
\label{subsec:correlations}
Table~\ref{compexp} compares theoretical results with experimental
data. We observe that for both $^4$He and $^{16}$O calculations are in
good agreement with the experimental data. For $^4$He the
experimental data is obtained by combining measurements from
Refs.~\cite{Arkatov74,Arkatov80,Pachucki07}. We also present a
comparison with other \textit{ab initio} results obtained with
hyperspherical harmonics~\cite{Gazit06,Ji13} and with the no-core
shell model~\cite{Stetcu09}. Because the experimental errorbar
is quite large, all theoretical calculations are compatible with data.
\begin{table}[htb]
\caption{Theoretical values of $\alpha_D$ for different nuclei
calculated with the NNLO$_{\rm sat}$ interaction in comparison to
experimental data from~\cite{Arkatov74, Arkatov80,Pachucki07} and
other calculations from Refs.~\cite{Stetcu09} (a), \cite{Gazit06}
(b) and \cite{Ji13} (c) for $^4$He, to experimental data from
Ref.~\cite{Ahrens75} for $^{16}$O. For $^{22}$O we compare to the
value obtained integrating the data from
Ref.~\cite{Leistenschneider2001} first over the whole energy range
(d) and then only the first 3 MeV of the strength (e), corresponding
to the low-lying dipole strength. Values are expressed in
fm$^3$. The theoretical uncertainties of our calculations stem from
the $\hbar\Omega$ dependence in the model space with $N_{\rm
max}=14$.}
\label{compexp}
\begin{center}
\footnotesize
\renewcommand{\tabcolsep}{1.8mm}
\begin{tabular}{cll}
\hline\hline
{Nucleus}&{Theory}&{Exp}\\
\hline
{$^4$He}&{0.0735(1)}&{0.074(9)}\\
{ }&{0.0673(5)$^{a}$}&{}\\
{ }&{0.0655$^{b}$}&{}\\
{ }&{0.0651$^{c}$}&{}\\
{ }&{0.0694$^{c}$}&{}\\
\hline
{$^{16}$O}&{0.57(1)}&{0.585(9)}\\
\hline
{$^{22}$O}&{0.86(4)}&{0.43(4)$^{d}$}\\
{}&{0.05(1)}&{0.07(2)$^{e}$}\\
\hline \hline
\end{tabular}
\end{center}
\end{table}
For $^{16}$O the calculation of the dipole polarizability with the
NNLO$_{\rm{sat}}$ interaction overlaps with the experimental
value~\cite{Ahrens75}. This is an improvement compared to
the previous calculation limited to NN interaction
only~\cite{Miorelli15}.
For the $^{22}$O nucleus, to compare our calculations with experimental data we integrate the experimental strength of Ref.~\cite{Leistenschneider2001} up to the available energy range of about 18~MeV above threshold, obtaining $\alpha_D^{exp} = 0.43(4)\ \rm{fm^3}$. This value is much lower than our calculated $\alpha_D^{th} = 0.86(4)\ \rm{fm^3}$ shown in Figure~\ref{fig:fig_polVSen_n2losat_22O}, which corresponds to the integration of the strength up to infinity. The theoretical result exceeds the experimental value by a factor of two and we also find that the integration of the theoretical strength over the first 18~MeV exhausts the 87\% of the polarizability sum rule. On the other hand, \citeauthor{Leistenschneider2001} observed a PDR extending for about 3~MeV above the neutron emission threshold of $S_n=6.85\ \rm{MeV}$. Integrating the data over this interval yields a dipole polarizability $\alpha_D^{\rm exp}(3~{\rm MeV}) = 0.07(2)\ \rm{fm^3}$. While our calculations in Figure~\ref{fig:fig_polVSen_n2losat_22O} does not reproduce the experimental threshold, integration over the first 3 MeV of the strength and considering the different $\hbar\Omega$ frequencies yields $\alpha_D^{th}(\rm{PDR}) = 0.05(1)\ \rm{fm^3}$. This is consistent with the experimental result.\\
In Figure~\ref{fig:fig_resp_comp_4He} we show the response function of $^4$He. The response function is obtained from the inversion of the LIT as described in Refs.~\cite{Bacca13,Andreasi05,Efros99} and the width of the band is an estimate of the inversion uncertainty.
The dark band from Ref.~\cite{LITCC} is the result obtained with coupled cluster with singles-and-doubles (CCSD) using a NN interaction at next-to-next-to-next-to-leading order (N3LO)~\cite{Entem03}. The light band represents the calculation of this work with NNLO$_{\rm sat}$~\cite{Ekstroem15} and it has been obtained by inverting the LIT with $\Gamma =10$ and 20~MeV calculated at $N_{\rm max}=14$ and $\hbar\Omega=22$~MeV. This is also the curve that has been integrated with method (i) in Figure~\ref{fig:fig_pol_vs_e_4He}. We find that the NNLO$_{\rm sat}$ response function, which includes three-nucleon forces, presents a larger peak with respect to other results with three-nucleon forces from Refs.~\cite{Gazit06PRL,Quaglioni2007}.
Finally, the theoretical results are compared with the experimental data by~\citeauthor{nakayama}~\cite{nakayama} (blue circles), ~\citeauthor{Arkatov74}~\cite{Arkatov74,Arkatov80} (white squares), ~\citeauthor{Nilsson}~\cite{Nilsson} (yellow squares), ~\citeauthor{Shima2005}~\cite{Shima2005,Shima2010} (magenta circles) and ~\citeauthor{Tornow2012}~\cite{Tornow2012} (green squares).
\begin{figure}[!ht]
\includegraphics[width=\linewidth,clip=]{4He_polresp_comp.pdf}
\caption{(Color online) $^4$He photo-absorption response function calculated with different methods and interactions (see text for details) compared with experimental data from by~\citeauthor{nakayama}~\cite{nakayama} (blue circles), ~\citeauthor{Arkatov74}~\cite{Arkatov74,Arkatov80} (white squares), ~\citeauthor{Nilsson}~\cite{Nilsson} (yellow squares), ~\citeauthor{Shima2005}~\cite{Shima2005,Shima2010} (magenta circles) and ~\citeauthor{Tornow2012}~\cite{Tornow2012} (green squares). Theoretical curves are shifted on the experimental threshold.}
\label{fig:fig_resp_comp_4He}
\end{figure}
In Figure~\ref{fig:fig_resp_comp_16O} we show the response function for $^{16}$O calculated with a NN interaction using CCSD~\cite{LITCC} (light band) and then with NNLO$_{\rm sat}$ (dark band). The calculations are compared with the experimental data from~\citeauthor{Ahrens75}~\cite{Ahrens75} (triangles with error bars) and~\citeauthor{Ishkhanov2002}~\cite{Ishkhanov2002} (red circles). The response function with NNLO$_{\rm sat}$ has been obtained again by inverting the LIT with both $\Gamma=10$ and 20~MeV and at frequency $\hbar\Omega=22$~MeV. The large error band for the NNLO$_{\rm sat}$ results from the fact that the largest available model space size in our calculation, namely $N_{\rm max}=14$, is smaller than the $N_{\rm max}=18$ used for the N3LO potential. Nevertheless, it is interesting to see that three-nucleon forces enhance the strength, slightly improving the comparison with the experimental data.
\begin{figure}[!ht]
\includegraphics[width=\linewidth,clip=]{16O_polresp_comp.pdf}
\caption{(Color online) $^{16}$O photo-absorption response function calculated with coupled cluster with singles-and-doubles using a NN interaction only~\cite{Entem03,LITCC} (dark band) and NNLO$_{\rm sat}$~\cite{Ekstroem15} (light band). The red circles are the experimental data from~\citeauthor{Ishkhanov2002}~\cite{Ishkhanov2002} while the white triangles with error bars are the experimental results by~\citeauthor{Ahrens75}~\cite{Ahrens75}. Theoretical curves are shifted on the experimental threshold.}
\label{fig:fig_resp_comp_16O}
\end{figure}
Comparing Figure~\ref{fig:fig_pol_vs_e_4He} and~\ref{fig:fig_polVSen_n2losat_16O} with Figure~\ref{fig:fig_resp_comp_4He} and~\ref{fig:fig_resp_comp_16O} respectively, and taking into account the results summarized in Table~\ref{compexp}, it is clear that the polarizability is not very sensitive to the structure and shape of the response function, but rather to the distribution of the dipole strength at low energies.
\subsection{Correlations between $\alpha_D$ and $r_{ch}$}
Let us also attempt to probe systematic theoretical uncertainties that are due to the employed interaction by considering results from different families of Hamiltonians. Such an approach can help to correlate observables of interest, see Refs.~\cite{nogga2004,platter2005,Reinhard2010,Piekarewicz12,Roca-Maza2015,hagen2015,Calci16} for examples. To study such correlations, one needs a considerable number of different interactions, so that one can obtain results spanning a wide range of values for the observables under investigation. For this reason, we choose to use similarity renormalization group (SRG) ~\cite{bogner2007} and V$_{low-k}$ ~\cite{bogner2003} evolutions as a tool to generate a set of phase-shift equivalent two-body interactions. When adding three nucleon forces at next-to-next-to-leading order -- without considering the induced three-body forces -- the low-energy constants were recalibrated on light nuclei observables~\cite{Hebeler11}. Finally, we also consider the newly developed NNLO$_{\rm sat}$ interaction~\cite{Ekstroem15}, which well reproduces radii~\cite{hagen2015}.
Various binding energies from NNLO$_{\rm sat}$ and other interactions of interest are shown in Refs.~\cite{Ekstroem15} and \cite{hagen2010b}.
We note that a correlation between the electric dipole polarizability and the nuclear charge radius $r_{ch}$ is expected from the nuclear droplet models~\cite{Myers77,Lipparini89} in heavy nuclei. In what follows we investigate correlations between the dipole polarizability and charge radius in $^{16}$O and $^{40}$Ca using a variety of interactions. We base our calculations on NN forces and 3NFs from Refs.~\cite{Hebeler11,Ekstroem15}, and also consider computations limited to NN forces alone.
\begin{figure}[htp]
\subfloat{\includegraphics[clip,width=\columnwidth]{16O_pol_vs_rch_v2.pdf}}
\subfloat{\includegraphics[clip,width=\columnwidth]{40Ca_pol_vs_rch_v2.pdf}}
\caption{(Color online) $\alpha_D$ versus $r_{ch}$ in
\textsuperscript{16}O and \textsuperscript{40}Ca. Empty symbols
refer to calculations with NN potentials only: $(a)$~SRG evolved Entem-Machleidt
interaction~\cite{Entem03} with $\Lambda=500$~MeV/c and $\lambda = \infty,3.5,3.0,2.5$ and $2.0 \ \rm{fm^{-1}}$, $(b)$~SRG evolved Entem-Machleidt interaction~\cite{Entem03} with $\Lambda=600$~MeV/c and $\lambda = 3.5,3.0$ and $2.5\ \rm{fm^{-1}}$, $(c)$ SRG evolved CD-BONN~\cite{CDBONN} interaction with $\lambda = 4.0$ and $3.5\ \rm{fm^{-1}}$, $(d)$ V$_{low-k}$ evolved CD-BONN potentials with $\lambda = 3.0,2.5$ and $2.0\ \rm{fm^{-1}}$ and $(e)$ V$_{low-k}$-evolved AV18~\cite{AV18} interaction and $\lambda = 3.0$ and $2.5\ \rm{fm^{-1}}$. The red diamonds $(f)$ refer to calculations that include 3NF: the large one is from NNLO$_{\rm{sat}}$~\cite{Ekstroem15} and the others from chiral interactions as in Ref.~\cite{Hebeler11}. The green bands $(exp)$, show the experimental data~\cite{Ahrens75,Angeli13}.}
\label{fig:pol_vs_rch}
\end{figure}
Figure~\ref{fig:pol_vs_rch} shows $\alpha_D$ -- calculated with method (iii) -- as a function of $r_{ch}$ in $^{16}$O and $^{40}$Ca for various interactions. The charge radii are based on the point-proton radii with contributions from nucleonic charge radii, see Ref.~\cite{hagen2015} for details. Empty symbols correspond to calculations with NN potentials only. In particular, $(a)$ is obtained from SRG evolved Entem and Machleidt~\cite{Entem03} interaction with cutoff $\Lambda=500$~MeV and, in order of decreasing $r_{ch}$ values, $\lambda = \infty,3.5,3.0,2.5$ and $2.0\ \rm{fm^{-1}}$, while for $(b)$ we used the same interaction with cutoff $\Lambda=600$~MeV and, in order of decreasing $r_{ch}$ values, $\lambda = 3.5,3.0$ and $2.5\ \rm{fm^{-1}}$. The points $(c)$ represented with triangles pointing up are calculations with the SRG evolved CD-BONN~\cite{CDBONN} potential with, in order of decreasing $r_{ch}$ value, $\lambda = 4.0$ and $3.5\ \rm{fm^{-1}}$, while the triangles pointing down $(d)$ are calculations with the V$_{low-k}$~\cite{bogner2003} evolved CD-BONN interaction and $\lambda = 3.0,2.5$ and $2.0\ \rm{fm^{-1}}$. The hexagons $(e)$ are calculations with V$_{low-k}$-evolved AV18~\cite{AV18} interaction and $\lambda = 3.0$ and $2.5\ \rm{fm^{-1}}$, in order of decreasing radius. The red diamonds $(f)$ are calculations including 3NFs. The larger red diamond is the value obtained with NNLO$_{\rm{sat}}$~\cite{Ekstroem15}, while the smaller ones are the potentials from Ref.~\cite{Hebeler11} also used for the calculations in $^{48}$Ca in Ref.~\cite{hagen2015}.
The error bars for the calculations represent uncertainties arising both from the coupled-cluster truncation scheme and the model space truncations and are estimated to be of the order of 1\% for the charge radius and 2\% for the polarizability (see Ref.~\cite{hagen2015} for details). Finally, the green bands are the experimental values for the polarizability~\cite{Ahrens75} and the charge radius~\cite{Angeli13}, respectively.
We note that $\alpha_D$ and $r_{ch}$ are strongly correlated. We also note that NN interactions alone systematically underestimate both $\alpha_D$ and $r_{ch}$ while the inclusion of 3NFs improves the agreement with data. The agreement with data is particularly good for the interaction NNLO$_{\rm{sat}}$. We note that one cannot blindly use a correlation between theoretical data points to extrapolate to experimental results. The data based on NN interactions, even when extrapolated with a simple linear or quadratic curve, does not meet the experimental values. In contrast, the results from NN and 3NFs can be interpolated (when e.g. the charge radius is known) to yield a sensible prediction for the dipole polarizability.
\section{Conclusions}
\label{sec:conclusions}
In conclusion, we employed integral transforms to compute the electric dipole polarizability in beta-stable nuclei and rare isotopes. This approach employs bound-state technology but takes the continuum properly into account. We presented in detail the formalism for coupled-cluster calculations of $\alpha_D$ and computed the dipole polarizability in $^4$He, $^{16,22}$O, and $^{40}$Ca. Formulations as the dipole polarizability as an energy-weighted sum rule facilitate the comparison to data in cases where only lower-lying dipole strengths are measured.
The comparison with data reveals the important role of three-nucleon forces, and results based on the NNLO$_{\rm{sat}}$ interaction agree well with data in $^{4}$He and $^{16}$O, and $^{40}$Ca. For the neutron-rich $^{22}$O, the low-lying dipole strength within 3~MeV of threshold theoretical results are consistent with data, while the total theoretical dipole strength is about a factor of two larger than what can be computed from the available data. Further investigation is needed to study the shape of the low-energy strength distribution.
Finally, we studied $^{16}$O and $^{40}$Ca with different two- and three-body interactions and observed a strong correlation between the dipole polarizability and the charge radius. Such a correlation could be useful to predict either of the two observables, when only one of them is experimentally known. Work in this direction is underway for heavier nuclei, such as $^{68}$Ni and $^{90}$Zr.
\begin{acknowledgments}
TRIUMF receives federal funding via a contribution agreement with the National Research Council of Canada. This work was supported in parts by the Natural Sciences and Engineering Research Council (Grant number SAPIN-2015-00031), the US-Israel Binational Science Foundation (Grant
No.~2012212), the Pazy Foundation, the MIUR grant PRIN-2009TWL3MX, the
Office of Nuclear Physics, U.S.~Department of Energy under Grants
Nos.~DE-FG02-96ER40963 (University of Tennessee) and DE-SC0008499
(NUCLEI SciDAC collaboration), and the Field Work Proposal ERKBP57 at
Oak Ridge National Laboratory. Computer time was provided by the
Innovative and Novel Computational Impact on Theory and Experiment
(INCITE) program. This research used resources of the Oak Ridge
Leadership Computing Facility located in the Oak Ridge National
Laboratory, supported by the Office of Science of the U.S.~Department
of Energy under Contract No. DE-AC05-00OR22725, and computational
resources of the National Center for Computational Sciences, the
National Institute for Computational Sciences, and TRIUMF.
\end{acknowledgments}
|
\section{Introduction}\label{intro}
Environment is known to play a role in galaxy evolution,
especially at relatively low redshift ($z \lesssim 1$) and for galaxies
with small to intermediate stellar mass ($\log(\mathcal{M}/\mathcal{M}_\odot)\lesssim 10.5$, see
e.g. \citealp{bolzonella10} and \citealp{davidzon15}). There
are still many open debates on this subject, ranging from the relation
between the 3D galaxy distribution and the underlying dark matter (DM)
structures, to the actual physical processes that shape galaxy
properties, which take place in different environments and on
different time scales. Only through observations at different epochs
we can robustly derive a coherent picture of galaxy evolution.
Ideally, we
need to probe from large to small scales, over a large span in
redshift, and with large statistics, to allow us to robustly measure
galaxy properties in different environments. Several physical
processes that shape galaxy properties take place in galaxy groups and
clusters (e.g. galaxy-galaxy merging, \citealp{toomre1972}, ram
pressure stripping of gas, \citealp{gunn_gott1972}, strangulation,
\citealp{larson1980}, harassment, \citealp{moore1996} and so on),
so the study of local, high-density environment on small scales is
crucial. Moreover, the cosmic web is continuously evolving with time,
and each galaxy can live in very different environments during its
life-time. As a consequence, we need to be able to study how
environment evolves to understand how it shapes
galaxy evolution, and at the same time we need to study how the
properties of the overall galaxy population change with time. For
instance, it would be important to understand if and how we can link
the peak of the star formation rate density ($z=1.5-2$, see
e.g. \citealp{cucciati12_SFRD} and \citealp{madau_dickinson14_CSFH})
with the onset of the relation between star formation rate (SFR) and
local density (\citealp{cucciati06} and \citealp{lin16} at $z\gtrsim1.2$, \citealp{elbaz2007} at
$z\sim1.1$, but see also \citealp{kodama07} and \citealp{spitler12}
for the identification of red massive galaxies in high redshift
clusters). Finally, we need large galaxy samples to remove the
degeneracies among the many parameters that regulate the galaxy
evolution processes (e.g. the complex interplay between local
environment, stellar mass and star formation, like for instance in
\citealp{peng2010_picture}, \citealp{McNaught_Roberts14_GAMA_LF} and
references therein).
In this framework the ability of identifying and isolate a
representative sample of high density structures is of
paramount importance. This would simultaneously require high accuracy
in the redshift measurement, large galaxy density and large
volumes. These requirements cannot be met by standard spectroscopic
surveys performed with current instruments, since the required
observation time would be prohibitively large. In addition, since we
aim to trace galaxy evolution, we should target galaxies at relatively
high redshift. To optimise the targeting of such galaxies using
multi-object spectrographs, some kind of target pre-selection in
near-IR bands is required, but ground-based observations in the
near-IR are limited by the atmospheric transmission windows.
A photometric galaxy survey would provide us with large statistics and
a large span in redshift within a much shorter time-scale of
observations. Many studies in the literature make use of photometric
redshifts to derive local environment. These analyses include galaxy
cluster identification (e.g.
\citealp{mazure07,adami10,bellagamba11,jian14_groups}), detailed studies
of extended structures (e.g., \citealp{guzzo2007_COSMOS,cassata07}),
and a broad analysis of the entire range of density enhancements, from
empty regions to the highest density peaks (see
e.g. \citealp{scoville07_env,scoville13_env,darvish15_env,lin16}). The
price to pay is the much lower ($>1$ dex) redshift precision with respect to
spectroscopic redshifts, which hampers the precision of the 3D
reconstruction of local environment on small scales (e.g.,
\citealp{cooper2005,lai15_photoz} and \citealp{malavasi16}).
Alternatively, one could carry out narrow-band photometric surveys,
with filters aimed to target emission line galaxies in a very narrow
redshift bin (e.g. the HiZELS and NEWFIRM H$\alpha$ surveys, see
\citealp{geach08_HaLF} and \citealp{ly11}). The main drawbacks of
this approach are the very limited redshift bin and the possible
contamination from other emission lines.
Slitless spectroscopy of emission line galaxies (ELG) is a compromise
between redshift precision and being able to probe a large volume
(both wide and deep) in a relative short exposure time. This kind
of spectroscopy uses a prism or grism as dispersing element, so the
spectral resolution can be very high. However, all the sources on
the sky plane are spread out into their spectrum at once, and as a
consequence the information on the angular position of the sources
is limited and spectra can partially overlap. Spectra overlap might
be significant especially in crowded fields, like deep samples, and
in high density regions, that are important targets of environmental
studies. The impact of this contamination can be reduced to some
extent by taking observations of the same field at different
position angles.
If a survey based on slitless spectroscopy is carried out from
space, the collected data do not suffer from atmospheric
absorption. This is the case, for instance, of the slitless surveys
performed with the HST, based on NICMOS \citep{mccarthy99_EL,
yan99_HaLF, shim09_HaLF} and on the WFC3. The WFC3 Infrared
Spectroscopic Parallels program (WISP;
\citealp{atek10_WISP,atek11_WISP}) represents a major improvement with
respect to the previous programme based on NICMOS, given that the WFC3
field of view is much larger than that of NICMOS. These HST
observations will be followed, in the future, by slitless
spectroscopic surveys from two other space missions, the European
Space Agency's Euclid \citep{redbook} and the National Aeronautics and
Space Administration's WFIRST (Wide-Field Infrared Survey Telescope,
\citealp{green12_WFIRST,dressler12_WFIRST}).
Space-based slitless spectroscopy of ELGs thus represents a
potential new advancement in environmental studies, although it also
suffers from some drawbacks. For instance, the deeper the observations,
the greater number of spectra which will overlap on the image. At the same
time, keeping the depth fixed, in over-dense regions the sky is more
crowded than in low density regions, increasing the
contamination from adjacent spectra. A way to tackle this problem is
to repeat observations at the same position but with different
position angles, but this comes at the expense of a longer total
exposure time.
A promising way to study local environment on small scales over large
volumes is through the synergy of photometric and spectroscopic galaxy
samples. For instance, one can anchor the photometric redshifts to a
robustly defined 3D skeleton, built with a (sub)sample of galaxies
with spectroscopic redshifts. A density estimator based on such a
method has been successfully developed for the zCOSMOS
survey \citep{lilly07_zcosmos} by
\cite{kovac2010_density}. \cite{cucciati14} applied a simplified
version of the same method to the VIPERS survey
\citep{guzzo14_vipers,garilli2014_VIPERS}.
The feasibility of such studies has to be evaluated case by case,
according to the selection function of the planned surveys. For
instance, the zCOSMOS and VIPERS samples are both flux-limited at
$i=22.5$, but the zCOSMOS spectroscopic sample has a higher target
sampling rate, and the zCOSMOS ancillary photometric catalogue has a
smaller photometric redshift error. For these two reasons, the
analysis performed with the zCOSMOS sample is robust down to smaller
scales than those reached with VIPERS. The higher sampling rate of the
zCOSMOS survey comes at the price of a much smaller covered area, and
only with a survey as large as VIPERS we can investigate the
properties of rare objects with large statistical significance.
The future Euclid mission will provide the community with both
slitless spectroscopic observations and photometric observations of a
flux limited sample of galaxies. The slitless spectroscopy will
sample ELGs, while the photometric observation will be limited in
$H-$band. The improvements with respect to current ground-based
spectroscopic surveys at $z\lesssim1$ consist of larger and deeper
photometric and spectroscopic catalogues, that might allow
environmental studies up to $z\sim2$ on relatively small scales.
Here, we use galaxy mock catalogues which mimic the Euclid Deep
spectroscopic and photometric surveys to assess the possibility of
studying how environment affects galaxy evolution at high redshift by
exploiting the synergy between spectroscopic and photometric
samples. The approach relies on the ability of sampling high vs. low
density regions rather than reconstructing the whole density field and
its correlation properties as in the galaxy clustering studies. In
this work we shall therefore focus on the ability to discriminate high
density environments from the low density ones.
This paper is organised as follows. In Sect.~\ref{euclid} we present
the mock galaxy catalogues that we use in this work, and in
Sect.~\ref{mocks_comp} some of their properties, such as number counts
and their clustering strength. Sect.~\ref{density} describes how we
estimate the density field using both the photometric and
spectroscopic galaxy catalogues. Our results on the reliability of the
environmental parameterisation are detailed in
Sect.~\ref{reconstruction}, and we show a test case in
Sect.~\ref{clusters}. We summarise our results in
Sect.~\ref{summary}.
We use a cosmology based on a $\Lambda$CDM model with
$\Omega_m = 0.272$, $\Omega_{\Lambda}= 0.728$, $H_0=70.4\,{\rm km\, s}^{-1} {\rm
Mpc}^{-1}$, i.e. the cosmology in which our mock galaxy catalogues
are based. Magnitudes are in the AB system \citep{oke74}.
\section{A test case: the Euclid Deep survey}\label{euclid}
Several next-generation cosmological projects are conceived to
comprise both a spectroscopic and a photometric survey, so to
exploit two independent cosmological probes at once: the baryonic
acoustic oscillations (BAO) and weak gravitational lensing (WL),
respectively.
In this respect, such projects naturally provide the two datasets
required for our strategy: a spectroscopic and a photometric
sample. We will consider the Euclid mission as a realistic case to
illustrate the potential of next generation surveys in environmental
studies.
\subsection{The Euclid Deep survey}\label{euclid_survey}
Full details on the Euclid Mission surveys are given in
the Euclid Definition Study Report (\citealp{redbook}, ``Red Book''
from now on). The pieces of information most relevant for this work
are the following.
The Euclid Mission comprises a Wide and a Deep survey, each of them
based on photometric and spectroscopic (slitless) observations. The Wide survey
will cover a sky area of 15000 deg$^2$, with a flux limit of $H=24$
($5\sigma$ point source) for the NIR photometry and a H$\alpha$ line
flux limit of $\sim3\times10^{-16}$ erg~cm$^{-2}$~s$^{-1}$
($3.5\sigma$ unresolved line flux) for the spectroscopy. Thanks also
to ancillary photometry from other surveys, it will be possible to
compute photometric redshifts ($z_p$) for the sources in the
photometric catalogue. The $z_p$ measurement error is required to be
at most $\sigma_{zp}=0.05(1+z)$, with a maximum 10\% of catastrophic
measurement failures. The spectroscopic redshifts ($z_s$) of the
spectroscopic survey are required to have a maximum measurement error
of $\sigma_{zs}=0.001(1+z)$.
The Deep survey will be two magnitudes deeper than the wide one, but
will cover a much smaller sky area ($\sim40$ deg$^2$).
Namely, the photometric deep survey will be flux limited at $H=26$
($5\sigma$ point source) and the H$\alpha$ flux limit for the
spectroscopic part will be $\sim5\times 10^{-17}$
erg~cm$^{-2}$~s$^{-1}$ ($3.5\sigma$ unresolved line flux). The
required maximum values of photometric and spectroscopic redshift
errors are the same as for the Wide survey.
The upper value for $\sigma_{zp}$, i.e. $0.05(1+z)$, is granted
by the use of the Euclid photometry in three NIR bands ($Y$, $J$,
$H$) covering the wavelength range $0.92-2.0 \mu m$, and the use of
at least four optical bands (from ground-based data, such as the `$g
r i z$' filter set) covering the range $420-930 nm$. The depth of
these observations needs to be at least 24 for the three NIR bands,
and $g=24.4$, $r=24.1$, $i=24.1$ and $z=23.7$ in the optical. These
requirements apply to the Euclid Wide survey. For the Deep survey,
they need to be $\sim2$ magnitudes deeper. To keep $\sigma_{zp}$ as
low as possible, it is also important that the photometry is uniform
within a single field, and across several fields. The requirement
is to reach a relative photometric accuracy of 1.5\% for all
the galaxies in the survey.
According to the Red Book, the total wavelength range covered by the
slitless spectroscopy is $1.1-2.0 ~\mu m$, and it will allow us the
detection of the H$\alpha$ line over the redshift range $0.7\lesssim z
\lesssim 2.05$.
In this work, we focus our attention to the Deep
survey in the redshift range $0.9\leq z \leq 1.8$. Basically, the Deep
survey is conceived to help with the calibration of the Wide survey,
which will be used for one of the Euclid main cosmological
probes. Clearly, the Deep survey will also provide the community with
unprecedented data sets for ancillary science. We want to exploit the
full potentiality of such data in the field of environmental studies,
where it is important to probe small scales, best reached with deep
samples.
\begin{figure} \centering
\includegraphics[width=8cm]{./fig1.ps}
\caption{{\it Top.} Redshift distribution of the galaxies in the
DREF\textsc{mocks}, \textsc{raw}DS\textsc{mocks } and DS\textsc{mocks}, as detailed in the legend, in bins of
$\Delta z = 0.05$. {\it Bottom.} Ratio of the redshift distribution
of the DS\textsc{mocks } over the DREF\textsc{mocks}, expressed in percentage.}
\label{nz_plot}
\end{figure}
\begin{table*}
\caption{List of the galaxy mock catalogues used in
this work, with their main properties: the limiting flux in $H$-band or
H$\alpha$ flux, the redshift error added to the peculiar velocities,
and other specific properties. For more details see
Sect.~\ref{mocks}.}
\label{mocks_tab}
\centering
\begin{tabular}{l c c c c c}
\hline
\hline
Name & sky area & $H$ band flux limit & H$\alpha$ flux limit & Redshift error & Other properties \\
& deg$^2$ & [mag] & [erg cm$^{-2}$ s$^{-1}$] & & \\
\hline
DREF\textsc{mocks } & $2\times2$ & 26 & - & none & -\\
DP\textsc{mocks } & $2\times2$ & 26 & - & $\sigma = 0.05(1+z_{pec})$ & 10\% of catastrophic failures\\
\textsc{raw}DS\textsc{mocks } & $2\times2$ & - & $7\times 10^{-17}$ & $\sigma = 0.001(1+z_{pec})$ & - \\
DS\textsc{mocks } & $2\times2$ & - & $7\times 10^{-17}$ & $\sigma = 0.001(1+z_{pec})$ & C=P=98\%. Adjusted $n(z)$. \\
\hline
\end{tabular}
\end{table*}
\begin{figure} \centering
\includegraphics[width=\hsize]{./fig2a.ps}
\includegraphics[width=\hsize]{./fig2b.ps}
\includegraphics[width=\hsize]{./fig2c.ps}
\includegraphics[width=\hsize]{./fig2d.ps}
\caption{Example of $R.A.-z$ distribution of galaxies in one of the lightcones
used in this work, in the redshift range $1.2<z<1.5$. The projected
$Dec$ covers a range of $0.5^\circ$. {\it Top:} all galaxies in the
DREF\textsc{mock } (flux limit at $H=26$, 100\% sampling rate, the redshift
includes cosmological redshift and peculiar velocity). {\it
Middle-top:} galaxies in the DP\textsc{mock } obtained from the DREF\textsc{mock } in
the top panel. {\it Middle-bottom:} only for reference, in this
panel we show the galaxies in the DREF\textsc{mock } after adding a photometric
redshift error of $\sigma = 0.01(1+z)$, i.e. five time less than the
one we use in our DP\textsc{mock}. {\it Bottom:} galaxies in the DS\textsc{mock }
extracted from the same lightcone as the DREF\textsc{mock } in the top panel. In
all panels, the black horizontal line represents the photometric
redshift error $\pm\sigma = 0.05(1+z)$ and the red line the
spectroscopic redshift error $\pm\sigma = 0.001(1+z)$, both
evaluated at $z=1.35$. In the third panel, the green horizontal line
represents a redshift error of $\pm\sigma = 0.01(1+z)$.}
\label{coni_deep}
\end{figure}
\subsection{Mock galaxy samples}\label{mocks}
For our analysis, we use mock galaxy catalogues that mimic the Euclid
Deep spectroscopic and photometric surveys (their flux limit, their
redshift measurement error etc). We will refer to these catalogues as
DS\textsc{mocks } and DP\textsc{mocks}, respectively, and also as Euclid-like
catalogues. To assess the robustness of our environment
reconstruction, we also need a catalogue as densely populated as the
photometric catalogue, but with no photometric redshift error. The
density estimated in this catalogue will be our reference computation. We call
these reference catalogues DREF\textsc{mocks}.
We use mock galaxy samples built with the semi-analytical
model of galaxy formation and evolution described in
\cite{gonzalez_perez14_SAM}. This model was applied to the dark
matter (DM) halo merging trees derived from a revision of the
Millennium Simulation \citep{springel2005_MILL} run with a cosmology
consistent with the 7 year results of the Wilkinson Microwave
Anisotropy Probe (WMAP7, \citealp{komatsu11_WMAP7}). Namely, the
adopted cosmological model is a $\Lambda$CDM model with $\Omega_m =
0.272$, $\Omega_b = 0.0455$, $h = 0.704$, $\Omega_{\Lambda}= 0.728$, $
n = 0.967$, and $\sigma_8 = 0.810$. The Millennium Simulation
contains $N = 2160^3$ particles of mass $8.6 \times 10^8 h^{-1} M_
{\odot}$ within a comoving box of size 500 $h^{-1}$Mpc on a side.
The construction of the Euclid-like lightcone are fully described in
\cite{merson13_lightcones}. The lightcones are stored in the Virgo
Consortium
Database\footnote{http://galaxy-catalogue.dur.ac.uk:8080/Millennium/}
\citep{lemson06}. These lightcones contain the galaxy properties we
need for our study, i.e. the observed $H-$band magnitude and the
H$\alpha$ flux. For each galaxy, two redshifts are provided: the
cosmological redshift ($z_{cos}$) and a redshift that includes both
the cosmological redshift and the peculiar velocity ($z_{pec}$). We
are interested in the second one, because we perform our analysis in
redshift space. Two Euclid-like lightcones were available at the
time of our analysis: one characterised by a wide area, with
H$\alpha$ and $H-$band limits matching those of the Euclid Wide
survey, and a deeper one, complete down to H$\alpha$ line flux of
$\sim3\times10^{-18}$ erg~cm$^{-2}$~s$^{-1}$ and $H=27$ (meaning
that it comprises all the galaxies which satisfy at least one of the
two selection criteria), covering a smaller, circular region of
$\sim20$deg$^2$. Note that the H$\alpha$ line flux and the $H-$band
limits are both deeper than the selection criteria of the Euclid
Deep survey. In this work we considered the latter lightcone, that
we divided into 4 square-shaped independent lightcones of
$2\times2$ deg$^2$, from which we extracted the mocks used in our
analyses, as follows.
(i) Galaxy mock catalogues with a flux limit of $H=26$, mimicking the
magnitude limit of the Euclid Deep photometric survey. We consider
them as our reference catalogues, and we call them DREF\textsc{mocks}. These
catalogues represent the Euclid Deep photometric survey without
photometric redshift error, and the density field estimated in these
catalogues (in redshift space) sets the reference to assess the
robustness of the density field reconstruction.
(ii) Galaxy mock catalogues with a flux limit of $H=26$, in which
galaxies have photometric redshifts ($z_p$). Namely, we took the
DREF\textsc{mocks } and we added to the redshift $z_{pec}$ of each galaxy a random
value extracted from a Gaussian with $\sigma = 0.05(1 +z_{pec})$. We
also added 10\% of catastrophic redshift measurements, by assigning a
random redshift to 10\% of the galaxies (although this is an
simplistic approximation). From now on we call these mock catalogues
DP\textsc{mocks}.
(iii) Galaxy mock catalogues with an H$\alpha$ flux limit of $7\times
10^{-17}$ erg cm$^{-2}$ s$^{-1}$, mimicking the $5\sigma$ flux limit
of the Euclid Deep spectroscopic survey (more conservative than
the $3.5\sigma$ described in Sect.~\ref{euclid_survey}, see below
for the reason of this choice). Galaxies in this sample will have
spectroscopic redshifts ($z_s$) with a typical measurement error of
$\sigma = 0.001(1 +z)$. To mimic this uncertainty we added to the
redshift $z_{pec}$ of each galaxy a random value extracted from a
Gaussian with $\sigma = 0.001(1 +z_{pec})$. From now on we call these
mock catalogues \textsc{raw}DS\textsc{mocks}.
(iv) Galaxy mock catalogues with H$\alpha$ flux limit as close as
possible to $7\times 10^{-17}$ erg cm$^{-2}$ s$^{-1}$, but with
redshift distribution $n(z)$ adjusted to be consistent with
predictions. Specifically, we take the \textsc{raw}DS\textsc{mocks}, and we modify their $n(z)$
to match the $n(z)$ predicted by \citet[see
Sect.~\ref{mocks_comp_counts}]{pozzetti16_Ha}. This procedure basically
consists in adding in the \textsc{raw}DS\textsc{mocks } also galaxies with H$\alpha$ flux
below $7\times 10^{-17}$ erg cm$^{-2}$ s$^{-1}$, as described in
details in Sect.~\ref{mocks_comp_counts}. We call these mock
catalogues DS\textsc{mocks}.
From now on, we will use the DREF\textsc{mocks } as reference catalogues, and the
DP\textsc{mocks } and DS\textsc{mocks } as ``Euclid-like'' mock galaxy catalogues. We note
that we adopted a very simplistic approach to mimic the Euclid
selection function in these catalogues. In particular, it is
important to explain some of the choices we made in the modelisation
of our mock catalogues.
The Euclid photometric survey will be characterised by masked sky
areas, variations of $\sigma_{zp}$ with galaxy luminosity, and so on,
but we do not take them into account. A full characterisation of
the photometric redshift probability distribution function (PDF) for
the Euclid survey is still under study within the Euclid Consortium,
so we can not mimic, in our mock catalogues, the full $z_P$ PDF nor
the redshift distribution of the catastrophic $z_P$ errors. For this
reason, we decided to use a very simplistic approach, assuming the
PDF to be Gaussian and the catastrophic $z_P$ errors to be $10\%$ at
all redshifts. This way, we do not take the risk of mimicking a
sophisticated $z_p$ observing procedure that is not representative
of the Euclid one.
The spectroscopic sample can be characterised by its
purity and completeness. The completeness, C, is defined as the
fraction of spectroscopic targets for which it is possible to measure
a redshift. The purity, P, is defined as the fraction of real targets
among all the measured redshifts (spurious features due to noise
etc. might be erroneously considered real H$\alpha$ lines). The
spectroscopic Euclid Deep survey is designed, among other goals, to
help with the computation of C and P for the Wide survey, so its C and
P are foreseen to be very close to 100\% at least for galaxies
brighter than the Euclid Wide spectroscopic survey ($3\times10^{-16}$
erg~cm$^{-2}$~s$^{-1}$). For galaxies with H$\alpha$ flux fainter
than the Wide survey limit, C and P have not yet been computed, and
they will possibly depend on redshift and H$\alpha$ flux.
For this reason, we decided to adopt, for our DS\textsc{mocks}, a H$\alpha$
flux limit of $7\times10^{-17}$ erg~cm$^{-2}$~s$^{-1}$, which
corresponds to the flux limit of the Deep survey at $\sim5\sigma$,
instead of the nominal one at $3.5\sigma$, and we consider the Deep
survey to have C=P$=98$\% down to the $\sim5\sigma$ flux limit. We
mimicked C and P by removing 2\% of the galaxies (randomly chosen,
irrespectively of their position or flux), and adding a corresponding
number of fake objects (randomly placed in the survey volume).
We verified the robustness of the density field reconstruction in the
case of a brighter H$\alpha$ flux limit for the DS\textsc{mocks}, to understand
what would change in case our assumption of C=P$=98$\% down to
$7\times10^{-17}$ erg~cm$^{-2}$~s$^{-1}$ is too optimistic. In
Appendix \ref{zade_req} we show the results on the density field
reconstruction if we use DS\textsc{mocks } with H$\alpha$ flux $>10^{-16}$
erg~cm$^{-2}$~s$^{-1}$, and we show that a brighter limit would affect only
marginally our results.
In Table \ref{mocks_tab} we summarise the main properties of the
galaxy mock catalogues listed above. In Fig.~\ref{nz_plot} we show
the redshift distribution $n(z)$ of our mock catalogues. The density
of galaxies in the DS\textsc{mocks } ranges between $\lesssim15$\% to $\sim10$\%
of the density of galaxies in the DP\textsc{mocks}, going from $z=0.9$ to
$z=1.8$. We note that, in the redshift range of interest
($0.9<z<1.8$), basically all the galaxies in the DS\textsc{mocks } also belong
to the DP\textsc{mocks}.
Figure \ref{coni_deep} shows the $R.A.-z$ distribution of galaxies in
one of the lightcones used in this work, in the redshift range
$1.2<z<1.5$. For the given lightcone, we show how galaxies are
distributed in the DREF\textsc{mock}, DP\textsc{mock}, and DS\textsc{mock}. It is clear that the large
photometric redshift error in the DP\textsc{mock } smears out the the small- and
large-scale structures visible in the DREF\textsc{mock}. For a comparison, in
Fig.~\ref{coni_deep} we also show what would happen to the DREF\textsc{mock } when
adding a photometric redshift error of $\sigma_{zp}=0.01(1+z)$, that
roughly corresponds to one of the smallest photometric redshift errors
obtained with real data to date (see e.g. \citealp{ilbert13} and
\citealp{ilbert15} in the COSMOS field, who find
$\sigma_{zp}=0.008(1+z)$ for galaxies with $i<22.5$). At $z\sim1.3$ as
in the Figure, $\sigma_{zp}=0.01(1+z)$ corresponds roughly to the
scale of voids/underdense regions, as is evident from comparing the
green line in the third panel with the typical dimension of the empty
areas in the top panel. This way, the galaxies in the most over-dense
structures do not mix too much when the photometric redshift error is
added, and it is still possible to identify high density regions (see
also \citealp{malavasi16}).
\begin{figure} \centering
\includegraphics[width=\hsize]{./fig3.ps}
\caption{$H-$band number counts in our lightcones (black solid line)
and in some observational data (symbols as in the legend).}
\label{Hband_counts}
\end{figure}
\section{Properties of the mock samples}\label{mocks_comp}
In this section we study the number counts, the redshift distribution
$n(z)$ and the clustering strength of the photometric and
spectroscopic mock catalogues and we compare them with available data
in the literature. This comparison is important, since the consistency
between real and mock samples is a key to guarantee the reliability of
our environmental study forecast.
\subsection{Number counts and redshift
distributions}\label{mocks_comp_counts}
Figure \ref{Hband_counts} shows the $H$-band number counts
(N/deg$^{2}$/mag) in our lightcones, compared with some observational
data
\citep{teplitz98_Hband,yan99_HaLF,thompson99_Hband,chen02_Hband,moy03_Hband,metcalfe06_Hband,
frith06_Hband,cristobal09_Hband,retzlaff10_Hband,bielby12_Hband,jeon14_Hband}. The
counts in the mock catalogues are consistent with the observed galaxy
counts in the range $19\lesssim H \lesssim 23$. Below $H\sim 19$,
predicted counts tend to underestimated the observed ones, up to a
factor of $\lesssim2$ at the brightest magnitudes, but this lack of
objects is mainly related to galaxies at $z\lesssim0.8$, a redshift
range that is not studied in our work. The counts in the mock
catalogues also seem to overestimate the observed counts at
$H\gtrsim23$ by a factor of $\sim50$\%, but at these faint magnitudes
the scatter in the real data is large. Overall, there is a fair
agreement between the counts in the lightcones and in the real data.
The H$\alpha$ counts are not well constrained yet by observations, and
yet the assessment of reliable forecasts of the H$\alpha$ number
counts and redshift distribution is a crucial task for the preparation
of the observational strategy for future missions like Euclid and
WFIRST. We expect a large number of H$\alpha$ emitters in the
redshift range explored by such missions, given that the cosmic star
formation rate (SFR) was higher in the past, with a peak at
$z\sim2$ (see e.g. \citealp{cucciati12_SFRD} and
\citealp{madau_dickinson14_CSFH}). Moreover, the use of star forming
galaxies to study the small scale environment at high redshift, like
in our case, could be particularly effective if it is true (though the
debate is still open in the literature) that SF galaxies reside
preferentially in high densities at $z\gtrsim1-1.5$ \citep{cucciati06,elbaz2007,lin16}.
Several H$\alpha$ samples have been collected in the past years, using
ground-based spectroscopy, grism spectroscopy from space and
narrow-band NIR photometry. We refer the reader to
\cite{pozzetti16_Ha} for the most recent compilation of such samples.
Namely, \cite{pozzetti16_Ha} used these samples to derive three models
of the evolution of the H$\alpha$ luminosity function (LF), to compute
the forecasts for future surveys (see also \citealp{geach10_HaLF} for
a previous modelisation). We adopt `Model 1' of
\citeauthor{pozzetti16_Ha} as the reference model in our analysis,
because it includes the evolution of both $L^*$ and $\phi^*$, while
$\phi^*$ does not evolve in their `Model 2', and their `Model 3' is
based on a shorter redshift range.
We have compared the $n(z)$ of the galaxies in the \textsc{raw}DS\textsc{mocks }
($n_{mock}$) with the predicted H$\alpha$ $n(z)$ of `Model 1'
($n_{mod}$). The \textsc{raw}DS\textsc{mocks } underestimate the predicted $n(z)$ at all
redshifts (for the same flux limit, i.e. $7\times 10^{-17}$ erg
cm$^{-2}$ s$^{-1}$), by an increasing factor going from $z=0.9$ to
$z=1.8$. The \textsc{raw}DS\textsc{mocks } contain 50\% (at $z=0.9$) to 20\% (at $z=1.8$)
of the galaxies predicted by Pozzetti's Model 1 (see their
Fig.6). This is true for all H$\alpha$ fluxes above the Euclid flux
limit (see their Figures 4 and 5). We remark that the Euclid
H$\alpha$ flux limit, even for the Deep survey, corresponds to
values very close to $L^*$ at $z\sim1.5$, so the underestimate of
H$\alpha$ counts in the mock catalogues is not (only) due to a
possibly not-well constrained faint-end slope of the observed
H$\alpha$ LF (see Fig.~2 of \citealp{pozzetti16_Ha} for the typical
uncertainty of $\alpha$ in the literature).
To reproduce the predicted H$\alpha$ $n(z)$, we
added galaxies in the \textsc{raw}DS\textsc{mocks } until $n_{mock}(z)$ and $n_{mod}(z)$ were
the same. Specifically, we picked up these new galaxies from the lightcones from
which the \textsc{raw}DS\textsc{mocks } were extracted, choosing randomly in $R.A.-Dec$
among the galaxies with H$\alpha$ slightly fainter than the flux limit
of the \textsc{raw}DS\textsc{mocks}. For each redshift bin $z_i$, we added fainter and
fainter galaxies until $n_{mock}(z_i)$ matched $n_{mod}(z_i)$. At the
end of the procedure, we had to add galaxies as faint as $2.5\times
10^{-17}$ erg~cm$^{-2}$~s$^{-1}$ at $z\sim1.0$, and even fainter
galaxies at higher redshifts, down to $1.5\times 10^{-17}$
erg~cm$^{-2}$~s$^{-1}$ at $z=1.8$.
Following this procedure, we built the DS\textsc{mocks } from the
\textsc{raw}DS\textsc{mocks}. These modified mock catalogues are those called DS\textsc{mocks } in
Sect.~\ref{mocks}, in which we subsequently modelled the foreseen
purity and completeness of the Euclid Deep spectroscopic survey, as
described in Sect.~\ref{mocks}. We show the redshift distribution of
the \textsc{raw}DS\textsc{mocks } and DS\textsc{mocks } in Fig.~\ref{nz_plot}.
\begin{figure} \centering
\includegraphics[width=\hsize]{./fig4.ps}
\caption{{\it Top.} Redshift-space galaxy two-point correlation
function in different mock catalogues, in the redshift bin
$1.2<z<1.5$. Different colours and style are for different mock
catalogues: the black dotted line is for DREF\textsc{mocks}, the green solid
line for the \textsc{raw}DS\textsc{mocks}, and the dashed blue line for DS\textsc{mocks}. The
error bar in the top right corner shows a difference in $\xi(s)$ of
$\pm10$\%. {\it Bottom.} Ratio $\xi_1 / \xi_2$ between the
redshift-space galaxy 2PCF in the different mock catalogues
shown in the top panel. The denominator ($\xi_2$) is always for
DREF\textsc{mocks}, and the numerator are the \textsc{raw}DS\textsc{mocks } and DS\textsc{mocks } catalogues
(colour-code as in the top panel).}
\label{clustering_all}
\end{figure}
\begin{figure} \centering
\includegraphics[width=\hsize]{./fig5.ps}
\caption{{\it Top.} Redshift-space galaxy two-point correlation
function in the adopted lightcones (before applying any
cut in $H-$band or H$\alpha$) for different bins of H$\alpha$ flux
down to $2\times10^{-17}$ erg~cm$^{-2}$~s$^{-1}$, as in the labels,
in the redshift bin $1.2<z<1.5$. Different colours are for
different H$\alpha$ flux bins. The error bar in the top right corner
shows a difference in $\xi(s)$ of $\pm10$ \%. {\it Bottom.} Ratio
$\xi_1 / \xi_2$ between the redshift-space galaxy 2PCF in the
different mock catalogues shown in the top panel. The denominator
($\xi_2$) is always the 2PCF in the faintest flux bin
($\sim2-3\times 10^{-17}$ erg~cm$^{-2}$~s$^{-1}$), and the numerator
is for the other flux bins in the top panel (colour-code as in the
top panel).}
\label{clustering_bin}
\end{figure}
\subsection{Two-point correlation function}\label{mocks_comp_clus}
Figure \ref{clustering_all} shows the galaxy two-point correlation
function (2PCF) in the DREF\textsc{mocks } and DS\textsc{mocks}. We use DREF\textsc{mocks } instead of
DP\textsc{mocks } for this exercise because our aim is to show the (possibly)
different clustering strength for two samples selected in different
ways ($H-$band or H$\alpha$ flux) rather than to show how photometric
redshift errors would affect the 2PCF (we expect that the small
spectroscopic redshift error does not alter significantly the 2PCF in
the DS\textsc{mocks}).
The 2PCF has been computed using the CosmoBolognaLib presented in
\cite{marulli16_CBL}. The random samples used to compute the 2PCF have
been built adopting the same geometrical selection of the DP\textsc{mocks } and
DS\textsc{mocks}, and their specific $n(z)$. The figure shows that the galaxies
in the photometric catalogue are more clustered than those in the
spectroscopic sample, by a factor of $\sim20\%$. This seems more
evident for scales $r<1 \,h^{-1}{\rm Mpc} $. We note that the two samples are
not necessarily expected to have the same clustering (see e.g. the
different clustering of galaxy populations with different stellar mass
or luminosity, like in
\citealp{li06,meneux08,coil08,christodoulou12_clustering,marulli2013_clustering}).
\cite{geach12_haXi} measured the 2PCF of H$\alpha$ emitters in the
Hi-Z Emission Line Survey (HiZELS) at $z\sim2.2$. The HiZELS sample is
complete down to $5\times 10^{-17}$ erg cm$^{-2}$ s$^{-1}$, that
corresponds to $L_{H\alpha} > 10^{42}$ erg cm$^{-2}$ s$^{-1}$ at z =
2.2. They compared their results with the 2PCF of a galaxy sample
extracted from the lightcones described in \cite{merson13_lightcones}.
The lightcones were built with the semi-analytical model of galaxy
formation and evolution described in \cite{lagos12_galform}, embedded
in the dark matter (DM) halo merging trees derived from the Millennium
Simulation \citep{springel2005_MILL}. The sample extracted from the
lightcones had the same selection function as the HiZELS sample.
\cite{geach12_haXi} find that the mock galaxy sample based on the
semi-analytical model, at the same redshift and $L_{H\alpha}$ as
HiZELS, has less clustering than the HiZELS sample on scales $r < 0.5
\,h^{-1}{\rm Mpc}$, while the two samples have very similar clustering strength
at larger scales.
We measured the 2PCF in the the mock galaxy samples based on the
semi-analytical model of \cite{lagos12_galform}, as in
\citeauthor{geach12_haXi}, but in the same redshift ranges and for the
same H$\alpha$ flux limit as in our DS\textsc{mocks}. We verified that their
2PCF is very similar to the one in our DS\textsc{mocks}, that are based on the
semi-analytical model by \cite{gonzalez_perez14_SAM}. Even if the
redshift explored in \cite{geach12_haXi} is slightly higher than our
maximum redshift ($z=1.8$), their results suggest that the small-scale
clustering in our DS\textsc{mocks } is too weak with respect to real data.
We remark that the clustering properties of the spectroscopic sample
are relevant for the method we use to estimate the density (see the
discussion in Sect.~\ref{zade} and Appendix \ref{zade_Hband}). We
verified that the 2PCF of the fainter galaxies that we added in the
\textsc{raw}DS\textsc{mocks } to match the predicted $n(z)$ is not too different from the
2PCF of the galaxies with H$\alpha$ flux $>7\times 10^{-17}$ erg
cm$^{-2}$ s$^{-1}$. This is shown in Fig.~\ref{clustering_all}, where
we also show the 2PCF for the \textsc{raw}DS\textsc{mocks}, and in
Fig.~\ref{clustering_bin}, where we show the 2PCF for different bins
of H$\alpha$ flux, including fluxes below $7\times 10^{-17}$ erg
cm$^{-2}$ s$^{-1}$. This implies that adding fainter galaxies to reach the
expected number counts does not alter the clustering strength of the
galaxies with brighter H$\alpha$ flux.
\section{Local density and environment}\label{density}
The key quantity that we use to quantify environmental dependencies is
the local density contrast of galaxies, $\delta_N$:
\begin{equation} \displaystyle
\delta_N = N/\langle N\rangle -1,
\label{delta_eq}
\end{equation}
where $N$ is the number of objects in the volume element, and $\langle
N\rangle$ the mean number of objects at a given redshift. Although
$\delta$ is defined as a `density contrast', we will often use simply
`density', for the sake of simplicity\footnote{For the purposes of
this paper, the density contrast $\delta$ is simply the way we
define the environment, we do not need to physically distinguish it
from the local density.}, and we consider it as a proxy of
environment.
We count objects in cylindrical cells of different sizes, all with the
same half-length of $1000 \,{\rm km\, s}^{-1}$ but different radii $R= 1, 2, 3, 5, 8,
10 \,h^{-1}{\rm Mpc}$ ($R_1, R_2, R_3, R_5, R_8, R_{10}$ from now on). The
lengths of the radii are chosen to span from small scales (like
clusters) to relatively large scales. Counts are performed by
randomly throwing cells to over-sample the survey volume. A
cylindrical shape of the cells allows us to adopt a symmetric shape on
the sky and an independent size along the line of sight, chosen to
take into account the peculiar velocities of cluster galaxies.
We only consider cells fully contained within the survey area, which
do not require any statistical correction for edge-induced
incompleteness. We will consider the problems related to boundaries
and gaps in the sky coverage in future work. We refer the reader to
e.g. \cite{cucciati06} for an effective method for boundary
correction in recent spectroscopic surveys, and to \cite{cucciati14}
for some examples of gap-filling methods.
We call `tracer galaxies' (or simply `tracers') the galaxies used to
estimate the local density. The choice of a given tracer
sample is a compromise between the maximisation of the
number density (the denser the sample, the smaller the scales that
can be reached in the density reconstruction) and the homogeneity of
the sample across the surveyed area and along the explored redshift range.
Both the spectroscopic and the photometric Euclid surveys are flux
limited. As a consequence the mean number density of objects decreases
with redshifts. This is a rather typical situation that, however,
does have an impact in environmental studies since it systematically
modifies the minimum scale-length that we can effectively probe. One
commonly adopted solution is to extract a sub-volume limited catalogue
of objects, for instance a luminosity-limited galaxy sample (see,
e.g., \citealp{cucciati10_zCOSMOS,McNaught_Roberts14_GAMA_LF}). This
has the advantage of providing a homogeneous sample of tracers with
the same mean separation at all redshifts. The drawback is to throw
away information at low redshifts (for instance, many faint galaxies
will not be used as tracers), hampering our ability to probe small
scales.
In this work, we want to assess the full potential of the data set at
each redshift to investigate environmental effects, so we consider the
whole flux-limited sample. For this reason, we expect that we will be
able to better reconstruct the local density field on small scales at
lower redshift.
Our aim is to reconstruct the density field in the Euclid photometric
survey, minimising the effects of the large photometric redshift
error. For this we exploit the synergy of the photometric and
spectroscopic Euclid surveys. The first one enjoys a large density of
tracers with a large redshift measurement error, the second one is
sparser, but the redshift measurement is much more precise. To assess
how well we can reconstruct environment, we estimate the density
contrast $\delta_N$ in three different mock galaxy catalogues:
\begin{itemize}
\item[-] counts performed in the DREF\textsc{mocks}, by counting the galaxies
falling within each cylindrical cell; we remind that in the DREF\textsc{mocks},
galaxy redshifts include peculiar velocities; we call this density
contrast $\delta_N^R$;
\item[-] counts performed in the DP\textsc{mocks}, by counting the galaxies
falling within each cylindrical cell; we call this density contrast
$\delta_N^p$; when compared with $\delta_N^R$, it provides us with an
estimate of the information lost due to the large photometric
redshift errors;
\item[-] counts performed in the merged Euclid-like DP\textsc{mocks}+DS\textsc{mocks}, in
which the spatial distribution of the spectroscopic galaxies is
exploited to improve the accuracy of the photometric redshift
estimate (with the ZADE method, see the next section); we call the
resulting density contrast $\delta_N^E$.
\end{itemize}
\begin{figure*} \centering
\includegraphics[width=5.8cm]{./fig6a.ps}
\includegraphics[width=5.8cm]{./fig6b.ps}
\includegraphics[width=5.8cm]{./fig6c.ps}
\includegraphics[width=5.8cm]{./fig6d.ps}
\includegraphics[width=5.8cm]{./fig6e.ps}
\includegraphics[width=5.8cm]{./fig6f.ps}
\caption{Joint probability distribution of $1+\delta_N^R$ (x-axis) and
$1+\delta_N^E$ (y-axis), in equipopulated bins containing 10\% of
the density measurements. $1+\delta_N^R$ and $1+\delta_N^E$ on the
two axes are ranked in increasing order. The colour-coded probability
value is such that in the generic $\delta_N^{E}$ ($\delta_N^{R}$)
bin the sum of the probability for $\delta_N^{R}$ ($\delta_N^{E}$)
is precisely 10\%, so the background colour code ranges from 0\%
probability (white) to 10\% probability (dark violet) as indicated
on the vertical bar on the right. The dotted line is the one-to-one
line, for reference. Black solid lines show, for each bin of the
y-axis, the 25th, 50th and 75th percentile of the distribution of
$1+\delta_N^R$ values. Gray dashed lines show, for each bin of the
x-axis, the 25th, 50th and 75th percentile of the distribution of
$1+\delta_N^E$ values. For a perfect reconstruction, the joint
probability function would be everywhere equal to zero but for
$\delta_N^{R}=\delta_N^{E}$, i.e. along the dotted line in the
plot. From top-left to bottom-right, the panels refers to different
radii ($R_1$, $R_2$, $R_3$, $R_5$, $R_8$ and $R_{10}$); all panels
refer to the redshift range $1.2<z<1.5$ (see
Fig.~\ref{tails_deep1_otherz} and \ref{tails_deep2_otherz} for other
redshift bins).}
\label{tails_deep1}
\end{figure*}
\subsection{The ZADE method}\label{zade}
We use the so called ZADE method to modify the $z_p$ probability
distribution function PDF of the galaxies in the DP\textsc{mocks } exploiting
the 3D distribution of the galaxies in the DS\textsc{mocks}. The ZADE method was
developed for the zCOSMOS survey \citep{lilly07_zcosmos}, and
is fully described in \cite{kovac2010_density}. We use a
simplified version of the ZADE approach, that has already been tested
for the VIPERS survey \citep{guzzo14_vipers} by
\cite{cucciati14}.
In brief, this simplified ZADE method can be broken down into the
following steps.
\begin{itemize}
\item[i)] For each galaxy in the DP\textsc{mocks}, we keep its angular position
({\it R.A.} and {\it Dec}), and set the PDF of its photometric
redshift ($P(z_p)$) to be equal to a normalised Gaussian centred on
$z_p$ and with standard deviation equal to the 1 $\sigma$ error in
the photometric redshift, $\sigma_{zp}$=0.05(1+$z$).
\item[ii)] We compute the spectroscopic redshift distribution $n(z_s)$
of the galaxies in the corresponding DS\textsc{mock } that fall within a
cylinder centred on the position of the given photometric galaxy
({\it R.A.}, {\it Dec}, $z_p$), with radius $R_{ZADE}$ (see below)
and half-length equal to $3\sigma_{pz}$.
\item[iii)] We set a new probability distribution function
PDF$_{ZADE}$ for each galaxy in the DP\textsc{mocks } equal to
$AP(z_p)n(z_s)$, where $A$ is a factor to normalise the integral of
PDF$_{ZADE}$ to unity. PDF$_{ZADE}$ is characterised by several
peaks. The value of PDF$_{ZADE}$ at each redshift peak corresponds
to the weight $w_{ZADE}$ at that given redshift, and for each
photometric galaxy the sum of all its $w_{ZADE}$ is unity by
definition.
\end{itemize}
We note that the $n(z_s)$ is quite discontinuous (because of the
limited number of spectroscopic galaxies) and, for this reason, we
sample it in discretised bins of $\Delta z=0.003$. For consistency,
we sample the Gaussian $P(z_p)$ on the same grid. This way,
PDF$_{ZADE}$ has the form of a histogram, and we call `peaks' all
its bins where it is different from zero. Basically ZADE
transforms a single photometric galaxy into a series of points at the
same {\it R.A.}-{\it Dec} position, and spread along the line of
sight, with each point having a weight given by $w_{ZADE}$. For the
estimation of the density field, the counts in the cylinders are
computed summing up the weights $w_{ZADE}$ of all the peaks that fall
in the given cylinder.
The ZADE approach is particularly suitable to reconstruct the 3D
galaxy distribution if the clustering properties of the spectroscopic
sample used to compute the $n(z_s)$ are the same as the ones in the
photometric sample (see e.g. the zCOSMOS and VIPERS surveys, where the
spectroscopic data set was a random subset of the photometric one). In
the case of Euclid, the photometric and spectroscopic surveys have
different selections (limit in $H$-band and in H$\alpha$ flux,
respectively), so we cannot expect the two samples to have the same
clustering, as discussed in Sect.\ref{mocks_comp_clus}.
Nevertheless, in the next Sections we demonstrate that the ZADE method
is very effective also in the case of the Euclid Deep survey, with
respect of using photometric redshifts alone. We refer the reader to
Appendix \ref{zade_Hband} for an example of how ZADE would work if we
use, at the place of the DS\textsc{mocks}, a random subsample of the DP\textsc{mocks }
(but with spectroscopic redshifts at the place of photometric ones).
We conclude with a remark on $R_{ZADE}$. The dimension of $R_{ZADE}$
is chosen as a compromise between the need to minimise the probed
scale length (to use the clustering strength on small scales) and to
maximise the number of spectroscopic galaxies in the cylinders (to
reduce shot noise). We use $R_{ZADE}=5\,h^{-1}{\rm Mpc}$, but we have checked
the robustness of our density field reconstruction also by varying
$R_{ZADE}$ between 3 and 10 $\,h^{-1}{\rm Mpc}$.
\begin{figure} \centering
\includegraphics[width=7cm]{./fig7.ps}
\caption{As in Fig.~\ref{tails_deep1}, but in this case the
$y-$axis shows $1+\delta_N^P$. The panel refers to $R_3$
over the redshift range $1.2<z<1.5$. For a
perfect reconstruction, the joint probability function would be
everywhere equal to zero but for $\delta_N^{R}=\delta_N^{P}$,
i.e. along the dotted line in the plot. }
\label{zphot_tails_deep}
\end{figure}
\section{Gauging the accuracy of the density field
reconstruction}\label{reconstruction}
In this Section we assess the ability of our method to trace the
underlying galaxy density. First, we aim at separating low- from
high-density environments to perform comparative studies of galaxy
evolution (Sect.~\ref{low_high} and \ref{PC}). Then we move to the
more challenging task of recovering the local galaxy density at a
given position (Sect.~\ref{zade_results}). We discuss our results in
the framework of environmental studies in Sect.~\ref{discussion}.
More specifically, here we compare the density field $\delta_N^E$ obtained using
the ZADE method in the DP\textsc{mocks}+DS\textsc{mocks } to the density
field estimated in the reference mock catalogues $\delta_N^R$. As an
additional check, we
show also the comparison between $\delta_N^{R}$ and the density
$\delta_N^P$ estimated in the DP\textsc{mocks }. The definition of $\delta_N$ is
the one given in Eq.~\ref{delta_eq}.
\begin{figure} \centering
\includegraphics[width=0.9\linewidth]{./fig8a.ps}
\includegraphics[width=0.9\linewidth]{./fig8b.ps}
\caption{Completeness (x-axis) and contamination (y-axis) for our
density reconstruction (see Sect.~\ref{PC} for the definition of
completeness and contamination). Contamination and completeness are
shown for all cylinder radii (different symbols as in the legend),
and for three redshift bins ($0.9<z<1.2$, $1.2<z<1.5$,
$1.5<z<1.8$). The three redshift bins for each radius are connected
with a line, with the lowest redshift bin indicated with a
bigger symbol. For each radius there are two series of points
(filled coloured and empty gray shades) that refer to
contamination coming from two different percentile ranges, as
indicated in the panels. {\it Top}. Completeness for the tail of
the highest $80^{th}$-$100^{th}$ percentiles and contamination from the
$0^{th}$-$20^{th}$ percentiles (filled symbols) and from the $40^{th}$-$60^{th}$
percentile (empty symbols). {\it Bottom}. As the top panel, but the
completeness is computed for the lowest densities (0 to 20 \%), and
its contamination comes from the $40^{th}$-$60^{th}$ percentile range (empty
symbols) and from the $80^{th}$-$100^{th}$ percentile range (filled symbols). See
the text for a caveat on the points for $R_1$ at $1.5<z<1.8$ in the
top panel. }
\label{CP_deep}
\end{figure}
\subsection{High vs low density regions}\label{low_high}
To test how well we can separate high from low
density regions, we do not need to compare $\delta_N^{E}$ and
$\delta_N^{R}$ on a point-by-point basis (as we will do in
Sect.~\ref{zade_results}), but only their ranking.
For this purpose we computed the probability distribution function
(PDF) of $\delta_N^{E}$ and $\delta_N^{R}$ for the cell counts and
sampled these PDFs in bins containing $10\%$ of the counts. We then
computed the joint probability of $\delta_N^{E}$ and $\delta_N^{R}$
P($\delta_N^{E}$, $\delta_N^{R}$) using the same binning. This
function is shown in Fig.~\ref{tails_deep1}, where, by construction,
$P(\delta_N^{R} \, | \, \delta_N^{E}) = \int P(\delta_N^{E},
\delta_N^{R}) \, \delta_N^{R} \, \mathrm{d} \delta_N^{R} = 10\% =
P(\delta_N^{E} \, | \, \delta_N^{R})$. The colour-coded probability
value is such that in the generic $\delta_N^{E}$ ($\delta_N^{R}$) bin
the sum of the probability for $\delta_N^{R}$ ($\delta_N^{E}$) is
precisely 10\%. For a perfect reconstruction, the joint probability
function would be everywhere equal to zero but for
$\delta_N^{R}=\delta_N^{E}$, i.e. along the dotted line in the plot.
From the plots in Fig.~\ref{tails_deep1} we see that:
\begin{itemize}
\item[-] The correlation between the ranked $\delta_N^R$ values and
the ranked $\delta_N^E$ values is stronger (both in terms of a
smaller deviation from the 1:1 correlation and a lower dispersion) for
larger radii.
\item[-] For $R_1$, at low density the values of $\delta_N^R$ are
discretised: the radius of the cylinder is very small with respect
to the mean inter-galaxy separation of the sample, so in many
cylinders we have 0, 1 or 2 galaxies. This is apparent in
Fig.~\ref{tails_deep1}, where $\delta_N^R$ is always binned in bins
of $10\%$. For $R_1$, at low density, we have the
first bin as large as $\sim20$\% . This is the percentage
range where $\delta_N^R=0$. This does not happen for $\delta_N^E$,
because the ZADE method produces objects with fractional weights, so
their counts result in a continuous $\delta_N^E$ distribution.
\item[-] We also explored the redshift dependence of the correlation
between $\delta_N^R$ and $\delta_N^E$ (see Appendix
\ref{zade_otherz}), and we found that the above-mentioned results
hold at all explored redshifts. Moreover, in general, at fixed
radius the correlation is stronger for lower redshift, and this is
especially true for small radii. This happens because at low
redshift the mean galaxy density is higher (the galaxy sample is
flux-limited), so the local density reconstruction suffers from a
lower shot noise. Clearly this effect is more important for small
radii, where the average counts-in-cells are lower and, as a
consequence, their relative variation from low to high redshift is
larger.
\end{itemize}
We notice that the scatter around the $1:1$ correlation in
Fig.~\ref{tails_deep1}, for the lowest\footnote{For all radii but
$R_1$, because of the discretised density values discussed in the
text.} and highest densities, is artificially low, simply because
of the finite range of the density values.
To appreciate more the goodness of the density reconstruction based on
both spectroscopic and photometric redshifts, in
Fig.~\ref{zphot_tails_deep} we show the same as in
Fig.~\ref{tails_deep1} but this time we compare $\delta_N^{R}$ with
the density reconstructed using only photometric redshifts
($\delta_N^{P}$). The almost uniform colour of
Fig.~\ref{zphot_tails_deep} shows that the large photometric errors
erase any correlation between the rankings, making it impossible to
separate high from low density environments using only a photometric
survey with a photometric redshift error of
$\sigma_{zp}=0.05(1+z)$.
\subsection{Completeness and contamination}\label{PC}
Figure \ref{tails_deep1} shows which is the range of $\delta_N^R$
that corresponds to any given selection on $\delta_N^E$. For a more
quantitative analysis, we can choose some specific $\delta_N^E$ ranges
and compute their completeness and contamination, as follows.
For each cylinder radius and redshift, we call $N^E_i$ ($N^R_i$) the
number of cylinders falling in the percentile range $i$ of the
$\delta_N^E$ ($\delta_N^R$) distribution. Also, we call
$N_{i,j}^{E,R}$ the number of cylinders that fall in the percentile
range $i$ of $\delta_N^E$ and in the percentile $j$ of
$\delta_N^R$. We then select a given percentile range $i$ and define
its completeness and contamination as:
\begin{equation} \displaystyle
\text{completeness} = N_{i,i}^{E,R} / N^R_i
\label{comp_eq}
\end{equation}
\begin{equation} \displaystyle
\text{contamination} = N_{i,j}^{E,R} / N^E_i \qquad \text{(with $i \neq j$)}
\label{cont_eq}
\end{equation}
Namely, the completeness indicates the fraction of cylinders that are
placed in the correct percentile. The contamination shows which
fraction of cylinders belonging to the percentile $i$ of $\delta_N^E$
distribution come from a different percentile of the original
$\delta_N^R$ distribution. Ideally one would like a completeness of
100\% and a contamination equal to zero.
Figure \ref{CP_deep} shows the completeness and the contamination for
our Euclid-like mock catalogues. We analyse three percentile ranges
($i$) of the $\delta_N^E$ distribution: $80^{th}$-$100^{th}$,
$40^{th}$-$60^{th}$, and $0^{th}$-$20^{th}$. Basically, they
correspond to the highest, intermediate and lowest density regimes,
that we call $P_H$, $P_I$, and $P_L$, respectively. For $P_H$ and
$P_L$, we show the completeness and the contamination coming from the
other density regimes.
We find that:
\begin{itemize}
\item[-] for all the density regimes, the completeness
(/contamination) increases (/decreases) for lower redshift and for
larger radii; the dependence of the completeness on redshift is
especially evident at small radii for $P_L$; in all cases, the
completeness varies more with radius (at fixed redshift) than with
redshift (at fixed radius);
\item[-] at fixed radius and redshift, $P_H$ is the most complete,
having always a completeness $\gtrsim60\%$ at all radii and redshifts; this is due to
the smaller random error in the density reconstruction (see
Fig.~\ref{delta_deep}) for the highest densities; this difference is
especially true at small radii, where the reconstruction of the low
densities is more difficult;
\item[-] the contamination of $P_L$ and $P_H$ from the opposite
density regime ($P_H$ and $P_L$, respectively) is very low, being
always below 1\% but for $R_1$ for which it gets to a few percent;
this implies that we can always safely separate $P_L$ from $P_H$;
\item[-] for $R_8$ and $R_{10}$ at all redshifts, and for $R_5$ at the
lowest redshift, the contamination in $P_H$ and $P_L$ coming from
$P_I$ is below 2\%; this implies that for these radii and redshift
ranges it is possible to study environment in three different
density regimes, instead of selecting only $P_L$ and $P_H$;
\item[-] for $R_5$, $R_{8}$ and $R_{10}$ the completeness in $P_H$ and
$P_L$ is very similar (always $\gtrsim75$\%), as can be derived by
the similar behaviour at the lowest and highest densities in
Fig.\ref{tails_deep1}; for smaller radii, we still have a good
completeness for $P_H$ ($\gtrsim55$\%, $\gtrsim60$\%, and
$\gtrsim70$\% for $R_1$, $R_{2}$ and $R_{3}$, respectively).
\end{itemize}
A caveat about the contamination from $P_L$: for $R_1$ the values of
$\delta_N^R$ at the lowest densities are discretised (see
Sec.~\ref{low_high}), and in the three redshift bins the lowest
$\delta_N^R$ value is taken by 20\%, 20\% and 30\% of the cylinders
(from low to high redshift). This means that the points at $1.5<z<1.8$
for $R_1$ in the top panel of Fig.~\ref{CP_deep} represents a
contamination from the percentile range 0-30\% instead of from the range
0-20\% as all the other points in the Figure.
In summary, we have shown that we can robustly identify high-density
regions and distinguish them from the lowest densities, at all
explored scales and at all explored redshift. The level of
completeness of the selected sample of high-density regions depends on
the scale and on the redshift, but the contamination from the lowest
densities is always below a few percent for $R_1$ and $<1\%$ for
larger radii. We identify high densities in a more complete
way than low densities, especially at small scales. Moreover, for
$R\geq5\,h^{-1}{\rm Mpc}$ we can separate very well high densities also from
intermediate densities (the contamination is only by a few percent).
\begin{figure} \centering
\includegraphics[width=6cm]{./fig9.ps}
\caption{Comparison of $\delta_N^R$ and $\delta_N^E$ on a
cell-by-cell basis. x-axis: density contrast in the reference
catalogue ($1+\delta_N^R$); y-axis, top panel: density contrast in
the Euclid-like catalogue ($1+\delta_N^{E}$); y-axis, bottom panel:
relative difference
($(\delta_N^{E}-\delta_N^R)/(1+\delta_N^R)$). The thick lines are
the median value of the quantity displayed on the y-axis in each
x-axis bin. Thin lines represent the 16th and 84th percentiles of
its distribution. Points are single cells, when cells per bin are
$<20$ (in which case we do not compute a median and
percentiles). The solid black line in the top panels is
the one-to-one line, and the horizontal lines in the bottom panels
are for reference. Results are shown for different cylinder
radii as in the label ($R_2$, $R_5$ and $R_{10}$) for the redshift bin
$1.2<z<1.5$. See Fig.\ref{delta_deep_otherz} for other redshift
bins.}
\label{delta_deep}
\end{figure}
\begin{figure} \centering
\includegraphics[width=6cm]{./fig10.ps}
\caption{ As in Fig.~\ref{delta_deep}, but in this case we compare
$\delta_N^R$ with $\delta_N^P$. }
\label{zphot_delta_deep}
\end{figure}
\subsection{Density-density comparison}\label{zade_results}
In this Section we move to a more demanding test, i.e. we compare the
values of $\delta_N^{E}$ and $\delta_N^{R}$ on a cell-by-cell basis.
The aim of this density-density comparison is to estimate the systematic
deviation from the true value and the random dispersion around the mean
value for the single cells, to quantify the goodness
of the density reconstruction at a local level.
In Fig.~\ref{delta_deep} we compare $\delta_N^{R}$ with
$\delta_N^{E}$, for different cylinder radii for the redshift bin
$1.2<z<1.5$. We show results only for $R_2$, $R_5$ and $R_{10}$ for
the sake of clarity, and the reader can easily extrapolate the results
for the other radii from those shown here, and from the discussion
below. In Appendix \ref{zade_otherz} we show the same results for the
two other redshift bins.
We note that, for $\delta_N^R \rightarrow -1$, the denominator of the
normalised residuals (the variable in $y$-axis in the bottom panels)
approaches zero and residuals rapidly increase. This is an artifact
related to our definition of residuals, as demonstrated by the upper
panels in which the relation between the two density fields is well
behaved close to zero.
We observe the following.
\begin{itemize}
\item[-] Our density reconstruction $\delta_N^E$ underestimates the
reference density $\delta_N^R$ at the highest densities, for all
cylinder radii and at all redshifts. This underestimation ranges
from $\sim70$\% for the smallest radii ($R_1$ and $R_2$) to
$\sim20$\% for the largest ones ($R_8$ and $R_{10}$).
\item[-] The lowest densities are always overestimated, by a
percentage that rapidly increases for $\delta_N^R \rightarrow -1$
(as explained above).
\item[-] The systematic error depends on the cylinder radius (it is
larger for smaller radii) at fixed redshift, and it mildly depends
on redshift at fixed radius; this indicates that it is more difficult
to reconstruct the absolute value of the local density at small
scales.
\item[-] The random error for the highest densities depends very
mildly (if at all) on the cylinder radius (being possibly slightly
larger for smaller radii) at fixed redshift, while it depends more
evidently on redshift at fixed radius (it varies from
$\lesssim10$\% in the lowest redshift bin to $\sim10-15$\% in the
highest redshift bin).
\item[-] At the highest densities, the random error is smaller than
the systematic error at all radii and redshifts, but for $R_8$ and
for $R_{10}$ at $z>1.5$ where the two errors are similar.
\end{itemize}
\begin{figure*} \centering
\includegraphics[width=5.8cm]{./fig11a.ps}
\includegraphics[width=5.8cm]{./fig11b.ps}
\includegraphics[width=5.8cm]{./fig11c.ps}
\caption{ {\it Left.} $R.A.-Dec$ distribution of galaxies in one of
the DREF\textsc{mocks}, for the redshift bin $1.47<z<1.48$. Blue triangles,
green circles and orange squares locate the position of galaxy
clusters (in this same redshift bin) with total masses of
$\log(M/M_{\odot})\in$ $[13,13.5)$, $[13.5,14)$, and $>14$,
respectively. {\it Middle.} In the same $R.A.$, $Dec$ and redshift
range as in the left panel, the grey-scale colour map represents the
density field estimated in the DREF\textsc{mock } with cylinders of $R=1
\,h^{-1}{\rm Mpc}$. The map is computed on a grid of $1 \,h^{-1}{\rm Mpc}$ per
element. The density value in each grid element is computed
averaging the $\delta_N^R$ values of the cylinders falling in that
given grid element. The grey scale refers to the percentiles of the
$\delta_N^R$ distribution as in the colour bar. Triangles, circles
and squares are as in the left panel. {\it Right.} As in the middle
panel, but here we use $\delta_N^E$ instead of $\delta_N^R$.}
\label{map_deep}
\end{figure*}
In summary, for $R_8$ and $R_{10}$ we are able to compute the absolute
value of the local density field $\delta_N^{R}$ with a small random
error and an acceptable systematic underestimate of $\sim20\%$. For
smaller scales both types of errors increase and the recovery of the
density is mainly hampered by the random ones in low density regions
and by the systematic ones in the high density peaks.
The underestimate of the highest densities in general, and especially
for the smallest radii, is mainly due to the following reasons: i) the
large photometric error, that spread the ZADE weights along the line
of sight on a large distance; ii) the low density of galaxies in the
DS\textsc{mocks } with respect to the density of galaxies in the DP\textsc{mocks } (see
e.g. \citealp{cucciati14}, Appendix A.3, where we verified that the
systematic error in the density reconstruction increases when we
decrease the sampling of the spectroscopic sample); iii) the lower
clustering of the DS\textsc{mocks } with respect to the DP\textsc{mocks } (see the effects
described in Appendix \ref{zade_Hband}).
In general, we expect to overestimate low densities and to
underestimate high densities when using tracers with large redshift
error, like photometric redshifts. This can be more easily seen in
Fig.~\ref{zphot_delta_deep}, where we compare $\delta_N^{R}$ and
$\delta_N^{P}$ values on a cell-by-cell basis.
The large photometric redshift error ($\sigma_{zp}$=0.05(1+$z$))
displaces galaxies far away from their true distance (see the second
panel of Fig.~\ref{coni_deep}), making the galaxy distribution almost
homogeneous in the entire explored volume (voids are filled, and high
density regions are depopulated). As a consequence, the reconstructed
density within cells is everywhere close to the cosmological
mean. This produces a rather small random error, but the reconstructed
$\delta_N^{P}$ is systematically smaller (higher) than the true one
for over-dense (under-dense) regions, and the mismatch increases with
the distance from $\delta_N^{R}=0$. The same trend is found at all
redshifts and for all cell sizes.
\subsection{Discussion}\label{discussion}
The results shown in Sect.~\ref{low_high} and \ref{PC} imply that
using the Euclid Deep survey it will be possible to perform
comparative studies of galaxy evolution in low and high density
regions from large to small scales up to $z\leq1.8$. In particular,
the identification of small-scale ($R\sim1 h^{-1}$Mpc) high-density
regions, with a very low contamination from other environments, will
allow us to study the regions (like e.g. galaxy clusters and groups)
where several physical processes shaping galaxy evolution normally
occur. Moreover, the high completeness and very low contamination in
identifying the lowest density regions on large scales
($R=10h^{-1}$Mpc) indicates that one can potentially use the Euclid
Deep survey to identify cosmic voids (see
e.g. \citealp{micheletti14_voids} for a systematic search of cosmic
voids in spheres of $R\gtrsim15 Mpc$).
More generally, we have shown that in a sample like the Euclid Deep
survey it will be possible to identify high densities and low
densities on different scales. The potentiality of this multi-scale
approach is already been demonstrated in several work in the
literature. On one side, correlating galaxy properties with the local
density on different scales is an effective way to understand how
galaxy evolution depends on the assembly of the large scale structure
(see e.g. \citealp{wilman10, fossati15} and references therein). On
the other side, a multi-scale approach can also be used to identify
the different components of the cosmic web (voids, filaments, walls
and clusters, see
e.g. \citealp{aragon_calvo10_multiscale,smith12_multiscale}).
Clearly, each science case will require a
dedicated feasibility study. For instance, the minimum (maximum) level
of completeness (contamination) required for any given science case
has to be quantified case by case (see e.g. \citealp{cucciati12_MILL}
and \citealp{malavasi16}, where they study whether the contamination
in the density field reconstruction hampers the detection of the
colour-density relation or of the different shape of the galaxy
stellar mass function in low and high densities).
The precise computation of the absolute value of the local density
field is a task much more difficult than the identification of the
lowest and highest densities, especially when we lack a precise
measurement of the 3D position of the tracers of the density field
itself. Nevertheless, our tests show that we are able to measure the
value of $\delta_N$ on scales of $\geq 8\,h^{-1}{\rm Mpc}$ up to $z\sim1.8$.
This achievement, for instance, will allow us to study the evolution
of the galaxy bias on such scales, as compared with surveys at lower
redshift (see e.g. \citealp{diporto14}, \citealp{bel15},
\citealp{cappi15} and references therein). Another application of a
precise local density value is the detection of clusters in their
phase of build-up, although by now this kind of detection has been
systematically tested in the literature at higher redshifts and on
slightly larger scales than those investigate here ($\sim10-17\,h^{-1}{\rm Mpc}$,
$z\gtrsim2$, see e.g. \citealp{chiang13,chiang14}).
We defer to a future work a detailed analysis of the feasibility of
specific environmental studies with the Euclid Deep survey. As a very
simple test case, we show in Sect.~\ref{clusters} the potential
ability of our method of detecting the regions where the most massive
clusters lie.
\section{A test case: the relation between the local environment and galaxy clusters}\label{clusters}
The main goal of this paper is to distinguish in a robust way the
highest density peaks from the low density regions on both small and
large scales, so the precise identification of clusters, voids, and
all the intermediate structures that form the cosmic web is beyond our
aim. Nevertheless it is interesting to see how the density
reconstructed with our mock catalogues compares with some of the
typical components of the large scale structure of the universe. Among
these components, galaxy clusters are certainly more easy to locate
with a precise 3D position, rather than filaments and voids, at least
in simulated DM/galaxy catalogues. Moreover, Fig.~\ref{CP_deep} shows
that our identification of small-scale ($R=1-2 h^{-1}$Mpc) high
density peaks is robust.
In Fig. \ref{map_deep} we show where (relatively massive) clusters are
located in $R.A.-Dec$, in a small redshift bin ($1.47<z<1.48$) of one
of our light-cones, compared to the position of galaxies in the DREF\textsc{mock }
and to the density fields $\delta_N^R$ and $\delta_N^E$. The redshift
bin has been chosen to be at relatively high redshift and to include
at least one of the most massive clusters ($\log(M/M_{\odot}) > 14$),
together with other over-dense regions and voids.
Qualitatively, we see that clusters, and especially the most massive
ones, fall in the highest density regions, and this is true for both
$\delta_N^R$ and $\delta_N^E$. In this example, the density is
estimated on a scale of $R=1\,h^{-1}{\rm Mpc}$ (the smallest scale we study in
this work, to better compare with the typical dimension of galaxy
clusters). More quantitatively, we see in Fig.~\ref{dens_distrib_deep}
that the values of local density evaluated at the positions where
galaxy clusters are located fall in the highest density tail of the
total density distribution for both $\delta_N^R$ and $\delta_N^E$, and
there is even some correlation between the mass of the clusters and the
local density value where they reside.
From Fig.~\ref{dens_distrib_deep}, it is also evident that there are
regions (map elements) with the same measured density as those where
clusters reside, but where there are no clusters. This is both due to
the contamination of our density reconstruction and to the fact that
our method is not fine-tuned to identify clusters as the highest
density peaks. Nevertheless this simple sanity check shows the
potentiality of future galaxy surveys in the study of the large scale
structure of the universe.
\begin{figure} \centering
\includegraphics[width=\hsize]{./fig12.ps}
\caption{Distribution of densities from the density maps presented in
Fig.~\ref{map_deep}: the black dot-dashed histogram is for the DREF\textsc{mock }
($\delta_N^R$) and the black solid histogram is for the
DP\textsc{mock}$+$DS\textsc{mock } ($\delta_N^E$). Blue, green and orange histograms
represent the density distribution of the map elements of
Fig.~\ref{map_deep} where the three classes of galaxy clusters fall
(see the caption of Fig.~\ref{map_deep} for the definition of the
three classes, here we use the same colour code). Also for the
clusters, the dot-dashed histograms refer to $\delta_N^R$ and the solid
histograms to $\delta_N^E$. }
\label{dens_distrib_deep}
\end{figure}
\section{Summary and conclusions}\label{summary}
We studied the feasibility of environmental studies in slitless
spectroscopic surveys from space. In particular, we want to exploit
the synergy between spectroscopic and photometric samples. As a test
case, we use galaxy mock catalogues which mimic the future Euclid Deep
surveys, namely a spectroscopic survey limited in H$\alpha$ flux and a
photometric survey limited in $H-$band. Our main goal is to verify
that it will be possible to disentangle the highest from the lowest
density regions, by means of the parameterisation of the local
environment. We anchored the photometric redshifts of the photometric
sample to the 3D skeleton of the spectroscopic sample, and we measured
the local environment on different scales, from 1 to $10 \,h^{-1}{\rm Mpc}$, in
different redshift bins in the range $0.9<z<1.8$. We found that:
\begin{itemize}
\item[-] We are able to disentangle the highest densities from the
lowest densities at all explored redshifts, and for all radii.
\item[-] For $R_8$ and $R_{10}$ at all redshifts, and for $R_5$ at
$0.9<z<1.2$, it is also possible to select an intermediate-density
regime, robustly separated from the highest and lowest density
tails.
\item[-] When we select the tail of highest densities, the
contamination from poorly-measured low densities decreases for larger
radii and at lower redshift, while the completeness increases.
Completeness is $\gtrsim60\%$ if we select the tails of the
highest 20\% using $R_1$. This percentage becomes 70\%
and 80\% for $R_3$ and $R_8$, respectively. The
contamination is always $\lesssim1\%$, at all redshifts and for all
radii, with the exception of $R_1$ for which it
reaches a few percent at the highest redshift.
\item[-] When we move to the more demanding task of computing the
precise density values, our density reconstruction under-predicts
galaxy counts in the highest densities, for all cylinder radii and
at all redshifts, by a factor ranging from $\sim70$\% for the
smallest radii ($R_1$ and $R_2$) to $\sim20$\% for the largest ones
($R_8$ and $R_{10}$). Moreover, the error budget at low density is
dominated by the random error, while at the highest densities it is
dominated by the systematic error. This is true for all radii, with
the exception of $R_8$ and $R_{10}$ at high density, where the
random and systematic error are comparable.
\item[-] As a qualitative test case, we verified that in a given
redshift bin at $z\sim1.5$, chosen to include at least a small sample
of relatively-high-mass galaxy clusters ($\log(M/M_{\odot}) > 13.5$), the galaxy
clusters reside in the highest tail of the density distribution, when
the density is estimated on a scale of $R=1\,h^{-1}{\rm Mpc}$.
\end{itemize}
Our results show that environmental studies can be efficiently
performed in photometric samples if spectroscopic information is
available for a smaller sample of objects that sparsely samples the
same volume. Namely, we are able to robustly identify the highest
density peaks from the lowest density regions using slitless
spectroscopy coupled with a deep photometric sample, like in the
Euclid Deep survey.
In particular, the redshift range probed by this survey ($0.7\lesssim
z \lesssim 2$) covers the epoch when the relation between local
environment and galaxy properties has been observed to change with
respect to the local universe (\citealp{cucciati06} and
\citealp{lin16} at $z\gtrsim1.2$, \citealp{elbaz2007} at
$z\gtrsim1$). Provided that the spectroscopic redshift information is
crucial to study small-scale environment, only relatively small
ground-based spectroscopic surveys have been able so far to study
galaxy evolution at $z>1$ (e.g. VVDS and DEEP2), while the largest
ones are too shallow (see e.g. zCOSMOS and VIPERS) to probe these
epochs. The possibility to verify these earlier findings at $z>1$ with
more statistics and on smaller scales will be one of the most
important scientific results from the ancillary science foreseen for
the Euclid survey.
\section*{Acknowledgements}
OC and AC acknowledge the support from grants ASI-INAF I/023/12/0
``Attivit\`a relative alla fase B2/C per la missione Euclid'' and MIUR
PRIN 2010-2011 ``The dark Universe and the cosmic evolution of baryons:
from current surveys to Euclid''. OC acknowledges the support from
grant PRIN INAF 2014 ``VIPERS: unveiling the combined evolution of
galaxies and large-scale structure at $0.5<z<1.2$''. OC thanks Sandro
Bardelli, Fabio Bellagamba, Lauro Moscardini and Mauro Roncarelli, for
useful discussions. The Millennium Simulation databases used in this
paper and the web application providing online access to them were
constructed as part of the activities of the German Astrophysical
Virtual Observatory.
\bibliographystyle{mnras}
|
\section{Introduction}
Long-distance quantum communication is one of the most appealing applications of quantum technology. It promises secure classical communication via quantum key distribution and is also essential for distributed quantum computation. High-rate quantum communication over long distances is possible using quantum repeaters, which either employ quantum error correction \cite{Knill96,Zw14,Muralidharan2014} or create long-distance entanglement between two parties from shorter-distance entanglement via swapping and processing operations \cite{Br98, Sa11} (see also \cite{Duan:2001aa,ladd2006,Loock2006,Childress2006,Jiang2007,Collins2007,Silvestre2013,Azuma2015,Pirandola2016,Azuma2016}), thereby overcoming limitations due to noisy and lossy channels and limited local control.
However, in a real-world application such as a quantum internet \cite{Kimble08}, one deals with a multi-user communication network. In such a network, the goal is not only to establish long-distance entangled pairs between fixed communication partners. One rather demands a flexible structure, where any given pair of parties can share entanglement and communicate, and that multipartite entangled states can be established between various communication partners to enable multi-user applications \cite{Ep15}. The latter is of particular importance for applications beyond two party quantum cryptography, for instance in distributed quantum computation \cite{Beals20120686}, secret voting and secret sharing \cite{Hillery05}, clock synchronization \cite{Komar2014}, or remote sensing. Depending on the task, certain kinds of entangled states need to be generated.
Here we address this inherently two-dimensional problem with a 2D-strategy. More specifically, we generalize the idea of the quantum repeater to 2D networks and propose an architecture that enables the direct generation of different kinds of entangled multi-party states \cite{RevModPhys.81.865} that are required for the applications mentioned above over large distances and between arbitrary communication partners, including Greenberger-Horne-Zeilinger (GHZ) states and 2D cluster states \cite{Briegel2001}.
As we show below, our direct 2D approach offers --in certain parameter regimes-- an advantage over a combination of 1D networks, where multipartite entangled states are established by appropriately combining bipartite entangled pairs. It can tolerate more errors, reaches higher fidelities and requires fewer local resources for storage.
More precisely, we show how to establish entangled states of fixed kind and size on larger and larger scales. The procedure combines several elementary (short-distance) states to obtain an equivalent state, but at larger distance. Imperfections in state preparation and local operations lead to a limited fidelity, which can be resolved by using multiparty entanglement purification (MEP) \cite{Du03,Du05,Kr06a,Du07,Glancy2006} to re-establish states with high fidelity from several copies. This repeater cycle can be applied in a nested way, similar as in the 1D repeater \cite{Br98} leading to states of arbitrary distance on the 2D lattice. By combining states of different scale, one obtains a multi-user communication network where all parties can participate and share entanglement. We illustrate this approach using 3-party GHZ states on a triangular lattice, and 8-party 2D cluster states on a rectangular lattice. Apart from the standard operational mode described above (mode I), where entangled states with a fixed number of parties are distributed over long distances, we also consider a variant of the scheme (mode II), where entangled states of growing size, i.e. with a larger number of parties involved, are generated among the 2D network. In this way one can e.g. produce a distributed 2D cluster state shared among all parties of the network. This state can then be used to establish pairwise quantum communication channels, but also for distributed measurement-based quantum computation \cite{Raussendorf2001,Briegel2009}.
The proposed scheme can make use of existing or currently developed platforms for 1D communication networks, as the experimental requirements are essentially the same. Only at the lowest level, the production of entangled pairs needs to be replaced by the preparation of GHZ states. For concreteness, we analyze the performance of an implementation using trapped ions in cavities for realistic noise parameters and show that appealing entangled states can be distributed over a thousand kilometers using present-day or near-future technologies.
The paper is organized as follows. In section \ref{sec:architecture} we introduce different schemes to distribute long-distance multipartite entanglement using a 2D approach. We mainly discuss
methods based on GHZ states and 2D cluster states, but the approach is not exclusive to these states. In the remaining sections we focus on a particular
protocol based on three qubit GHZ states that allows one to establish a two-dimensional communication network. In section \ref{sec:analysis} we analyse this 2D protocol and determine the errors it can tolerate.
In section \ref{sec:measurementbased} we consider a measurement based implementations of the scheme. In section \ref{sec:1dcompare} we give a
comparison between our 2D approach and networks based on 1D repeaters. In section \ref{ions} we discuss a concrete physical implementation based on trapped ions of our 2D repeater scheme without entanglement purification, and compare the performance to 1D strategies. We summarize our findings and conclude in section \ref{sec:summary}, while some technical details and additional
results regarding the use of different MEPs can be found in the appendix.
\section{2D repeater architecture \label{sec:architecture}}
We consider a regular 2D network, where the communication parties are located at the nodes of the lattice and are connected by quantum channels.
The goal is to establish high-fidelity long-distance entangled states shared between multiple communication partners. This is achieved by using entangled states generated over short distances, and connecting and purifying them by means of local operations.
\subsection{Standard operational mode I}
In the standard operational mode, mode I, a certain type of entangled state, e.g. a $m$-qubit GHZ state
\begin{equation}
|GHZ_m\rangle = \tfrac{1}{\sqrt{2}}(|0\rangle^{\otimes m} + |1\rangle^{\otimes m}),
\end{equation}
is produced at larger and larger distances, in such a way that the number of entangled parties (and type of state) is kept constant. This corresponds to a coarse graining of the lattice and the entanglement structure, where intermediate qubits are projected out (similar as in entanglement swapping). For $m=3$ qubit GHZ states, this is shown in Fig. \ref{fig:connect}. This process is performed similarly as in pairwise quantum communication using quantum repeaters. Elementary (multipartite) states are generated via direct transmission of qubits over noisy quantum channels. Several such short-distance entangled states are then connected such that the resulting state is the same as initially, but shared between parties at longer distance. The distance (in all directions) is thereby at least doubled. If the initial state or the local operations used for connection are not perfect, the fidelity of the resulting state is reduced. One can use MEP \cite{Du03,Du05,Kr06a,Du07,Glancy2006} to generate a state with the same fidelity as the initial elementary ones from several copies by means of local operations, thereby resulting in a situation as initially, however with entangled states of longer distance. This defines the 2D repeater cycle, which is applied in a concatenated way to achieve long-distant entanglement. As for the 1D repeaters, this approach leads to a polynomial scaling in the overall resources in the covered area and distance \cite{Br98}.
When using entanglement purification protocols with two-way classical communication, or a probabilistic connection procedure (see below), classical communication between the involved parties in a purification or connection step is required before states can be used at the next repeater level. This classical communication, together with gate times and preparation times of elementary states, determines the achievable rates. Notice that when using deterministic connection operations and deterministic entanglement purification with one-way classical communication, all steps of the protocol can be done simultaneously, and only a final correction operation at the end nodes (that can be done later) is required.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{connect_triangles_and_merge.pdf}
\caption{\label{fig:connect} 2D quantum repeater scheme based on three-party GHZ states. Short-distance GHZ states are connected to form a long-distance GHZ state with reduced fidelity, which is then re-purified to the initial fidelity via entanglement purification.}
\end{figure}
\subsubsection{Flexible quantum communication network}
A flexible quantum communication network where all partners are able to participate (and not just the outermost, far distant ones) can be achieved by using states from different scales, i.e. also the ones that are produced during earlier repeater cycles on short scales. These states can be combined such that GHZ states or pairs shared between {\it any} parties can be generated, see Fig. \ref{fig:illu_layers}. We remark that in such a flexible network, additional purification steps might be required, in particular when GHZ states of different scales are connected.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{multilayers.pdf}
\caption{\label{fig:illu_layers} On a triangular lattice the GHZ states at different scales form a flexible communication network where all parties can participate.}
\end{figure}
\subsubsection{2D cluster states on a rectangular lattice}
The operational mode I is not restricted to the generation of GHZ states, but is also applicable to other entangled states such as graph states \cite{He06}
$|G_m\rangle=\prod_{(k,l)\in E}CZ^{(kl)}|+\rangle^{\otimes m}$,
where the edge set $E$ corresponds to the edges of a corresponding graph and determines the entanglement features of the state, and $CZ^{(kl)}=diag(1,1,1,-1)$ is the controlled-$Z$ operation acting on qubits $k,l$ (see appendix \ref{sec:graphstates}). An example for such a self-similar growing structure for a 2D-cluster type state on a rectangular lattice is shown in Fig. \ref{fig:cluster}.
The procedure to generate more and more coarse grained 2D-cluster states is conceptually very appealing because
the states at every repeater level have the same structure and one can observe the growth of a self-similar structure.
However, from figure \ref{fig:cluster}
it is apparent that many of the qubits are not involved in the protocol at all and are only
measured out to disentangle them. As an alternative scheme not relying on so many redundant qubits the repeater scheme for three-qubit GHZ states
can be used on a quadratic lattice and four GHZ states of an appropriate level can be combined to form the same coarse grained building
block (see figure \ref{fig:quad_ghz}).
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{cluster.pdf}
\caption{\label{fig:cluster} Repeater scheme based on 8-party 2D cluster states. Multipartite entangled states are connected via Bell-type measurements (light red), and some qubits are measured in the $y$ or $z$ basis.}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{quadr_triangles.pdf}
\caption{\label{fig:quad_ghz} The repeater scheme based on GHZ states implemented on a quadratic lattice can be used to construct coarse
grained 2D-cluster states. The L-shaped 3-qubit graph states are LU-equivalent to GHZ states.}
\end{figure}
\subsection{Alternative operational mode II}
Operational mode II corresponds to the growth of entangled states of similar type, but with increasing particle number. In that case, all particles at a specific site are merged into one. For the example of 3-qubit GHZ states on a triangular lattice, this means that starting from three $|GHZ_3\rangle$ states, one $|GHZ_6\rangle$ state shared between all nodes that are included in the larger triangle is generated (rather than a $|GHZ_3\rangle$ state shared between the nodes of the big triangle only). Similarly, the basic 2D-cluster structure is merged into a larger 2D cluster state with open links at the right and bottom (to connect it to neighboring structures), see Fig. \ref{fig:cluster} middle. Again, the fidelity is reduced and re-purification of the resulting states might be required. Notice, however, that now states are {\it not} the same as initially, and in fact MEP becomes less efficient for larger particle numbers. The threshold value, i.e. the tolerable error of local operations, becomes smaller for increasing system size for GHZ states \cite{DB04, He05}, essentially limiting the maximum size $m$. The threshold for 2D cluster states, in turn, is independent of system size \cite{DB04, He05}. This allows for the production of 2D cluster states of arbitrary size, which can e.g. be used as a resource for (distributed) measurement-based quantum computation \cite{Raussendorf2001,Briegel2009}.
\subsection{Connection of states to the next repeater level}
The required connection operations for operational mode I and II can be realized as follows. Two GHZ states $|GHZ_m\rangle \otimes |GHZ_n\rangle$ can be deterministically connected in such a way that (i) two qubits are merged into one, or (ii) both systems are projected out. In case (i) a projection $P_{S}=|00\rangle\langle 00| + |10\rangle \langle 11|$ or $P_S^\perp = |00\rangle\langle 01| + |10\rangle \langle 10|$ that acts on one qubit of the first, and one qubit of the second GHZ state is applied. The first qubit remains and the second is factored out, resulting in $|GHZ_{m+n-1}\rangle$ or $\mathbbm{1}^{\otimes m}\otimes \sigma_x^{\otimes n-1}|GHZ_{m+n-1}\rangle$ depending on the measurement outcome. The local Pauli operators can be corrected and a deterministic merging of two GHZ states is achieved. In case of (ii), one applies in addition a projection in the $X$ basis on the remaining qubit. (Equivalently, a Bell measurement, i.e. a measurement in the basis $\{|\Phi^{\pm}\rangle,|\Psi^\pm\rangle\}$ with $|\Phi^\pm\rangle = (|0\rangle\otimes|0\rangle \pm |1\rangle\otimes|1\rangle)/\sqrt{2}, |\Psi^\pm\rangle = (|0\rangle\otimes|1\rangle \pm |1\rangle\otimes|0\rangle)/\sqrt{2}$, can be directly applied to both qubits.) This leaves the remaining system in a $n+m-2$ particle GHZ state up to local Pauli operations that can be corrected. Notice that the results of the measurements can be used for error detection and the design of probabilistic connection schemes. For general graph states with an open link, i.e. a particle $A$ that is only connected to a single neighbor, a merging operation can be performed by first connecting $A$ via a $CZ$ operation to particle $B$ of the second graph state, and then measuring $A$ in the $Y$ basis \cite{He06}.
Merging all open links of the 2D cluster type state to the neighboring ones as illustrated in figure \ref{fig:cluster} leaves us with a larger 2D cluster state. The coarse-grained 2D cluster state can be obtained by additional $Z$ and $Y$ measurements.
The particular variant that connects three GHZ states to one GHZ state at the next repeater level will be analysed in detail in the sections below.
The protocol that does so deterministically can be summarized as follows
\begin{itemize}
\item Start with three copies of a (probably noisy) GHZ state.
\item Perform Bell measurements on qubits (2,6), (3,8) and (5,9). (see figure \ref{fig:triangle_numbers})
\item Depending on the outcomes of the Bell measurement perform correction operations as outlined in table \ref{tab:corr}.
\end{itemize}
Notice that only two of the Bell measurements are necessary to connect the three GHZ states. The seemingly redundant third
measurement does not only make the protocol symmetric but can actually be used to detect some specific errors. While it is not possible to correct the errors detected this way because the error syndromes are not unique,
it allows to discard the cases where an error is detected and therefore obtain better error thresholds.
However, this also means that the connection procedure only works probabilistically, and the whole procedure has to restart from the beginning if errors are detected.
The cases discarded are those in table \ref{tab:corr} with one or three $\ket{\Psi^\pm}$ outcomes.
Thus one may either use a deterministic connection procedure with slightly worse error thresholds,
or a probabilistic procedure where error thresholds are better, however, the rates may differ.
Unless explicitly stated all results in this paper use the probabilistic scheme.
Notice that as entanglement purification is already a probabilistic procedure, the performance and in particular the scaling of the overall 2D repeater scheme remains unchanged.
\begin{figure}[ht]
\centering
\includegraphics[width=0.5\columnwidth]{connect_triangles_numbers.pdf}
\caption{\label{fig:triangle_numbers} Repeater scheme connecting three GHZ states to one GHZ state on the next higher level.}
\end{figure}
\begin{table}[ht]
\begin{tabular}{ccc|c}
Bell I & Bell II & Bell III & Correction \\
\hline
$\Phi^\pm$ & $\Phi^\pm$ & $\Phi^\pm$ & $\mathbbm{1}$ \\
$\Phi^\pm$ & $\Phi^\pm$ & $\Psi^\pm$ & - \\
$\Phi^\pm$ & $\Psi^\pm$ & $\Phi^\pm$ & - \\
$\Psi^\pm$ & $\Phi^\pm$ & $\Phi^\pm$ & - \\
$\Psi^\pm$ & $\Psi^\pm$ & $\Phi^\pm$ & $X^{(1)}$ \\
$\Psi^\pm$ & $\Phi^\pm$ & $\Psi^\pm$ & $X^{(4)}$ \\
$\Phi^\pm$ & $\Psi^\pm$ & $\Psi^\pm$ & $X^{(7)}$ \\
$\Psi^\pm$ & $\Psi^\pm$ & $\Psi^\pm$ & randomly one of $\{X^{(1)}, X^{(4)}, X^{(7)} \}$ \\
\hline
\end{tabular}
\begin{tabular}{c|c}
\# of $\Phi^-$ or $\Psi^-$ outcomes & Correction \\
\hline
0,2 & - \\
1,3 & \pbox{20cm}{$Z$ on one of the remaining qubits \\ (does not matter which one)} \\
\end{tabular}
\caption{\label{tab:corr} The correction operations for the repeater protocol connecting GHZ states (see figure \ref{fig:triangle_numbers}). }
\end{table}
\section{Analysis of 2D repeaters \label{sec:analysis} }
We now analyze the performance of the 2D repeater when taking noise and imperfections into account.
There are several relevant figures of merit for a repeater scheme. Here we concentrate on error thresholds for local operations and channels as well as reachable fidelities, as this provides information on whether such a scheme is suitable in principle. Another important quantity are achievable rates, which however depend strongly on the specific MEP and details of the implementation. We do not provide a full rate analysis for the general scheme here. For a concrete implementation with trapped ions, however, we also investigate distribution times and reachable distance with limited resources (see Sec. \ref{ions}).
We demonstrate that the usage of 3-party entangled states offers (in certain parameter regimes) an advantage over bipartite schemes w.r.t. error tolerance and achievable fidelity, but also for storage resources. This implies that there exist parameter regimes for channel noise and noisy local operations where a 2D approach allows one to generate GHZ states with a certain fidelity, while this is not possible with a 1D approach. Clearly, in this case also the achievable rates using the 2D approach are higher. In other regimes where both approaches are applicable, the achievable rates using a 1D approach might be higher, as multipartite recurrence type entanglement purification protocols are rather inefficient \cite{Du07}. The situation of direct state distribution (without a repeater scheme) using bipartite and multipartite strategies was investigated in \cite{Kr06} where a similar behavior was found.
\subsection{Error model}
Quantum channels are considered to be lossy and noisy, which prevents a direct transmission of quantum information over longer distances.
In addition, local operations at individual nodes of the network (parties) are considered to be noisy as well.
We model channel errors by a completely positive map (CPM) of the form
${\cal E}_q^{(a)} \rho=q\rho+\tfrac{1-q}{4}\sum_j \sigma_j^{(a)} \rho \sigma_j^{(a)}$,
with channel noise parameter $q$. We describe a noisy operation by ${\hat U}\prod_a{\cal E}_p^{(a)} \rho$, i.e. single-qubit local depolarizing noise (LDN) with error parameter $p$ on all involved particles, followed by the ideal operation described by the superoperator ${\hat U}$ with ${\hat U}\rho = U \rho U^\dagger$.
Clearly, in a physical realization errors may have different form, or channel losses may be dominant. However, this simple error model assuming depolarizing local noise covers the essential features and allows us to illustrate the effect of noise on the performance of such a 2D quantum communication network, similarly as done in \cite{Br98,Du07,Zw12,Zw15} for 1D repeaters. Notice that loss errors can in principle be mapped to depolarizing errors by replacing a lost qubit by a completely mixed state, but there are more efficient or practical ways to deal with loss, e.g. by just repeating the transmission as we consider in the trapped ion implementation below.
\subsection{Error thresholds}
In the following, we will concentrate on operational mode I. In order to analyze whether the repeater works despite imperfections in channels and operations, it is
useful to start by identifying noise thresholds that indicate up to which noise level states remain distillable.
\subsubsection{Error thresholds for entanglement purification}
For noisy entangled pairs and GHZ states, necessary and sufficient conditions for distillability are known \cite{Du99, Du00, DB04, He05}.
In the case of local depolarizing noise with error parameter $q$ that acts on each of the particles, one finds a threshold of $q = 1/\sqrt{3}\approx 0.5774$ [$q\approx 0.5567$] of
local noise per particle for entangled pairs and 3-party GHZ states respectively \cite{Du99, Du00, DB04, He05} (see Appendix \ref{GHZdist}). For GHZ states, the threshold
value for $q$ increases with the size $m$ of the state, while for 2D cluster states the threshold for distillability is independent of $m$ \cite{DB04, He05}. A lower bound
on $q$ for distillability of 2D cluster states is given by $q=0.8281$ \cite{DB04,He05}.
Notice that, perhaps surprisingly, 3-party GHZ states are more robust against local noise than entangled pairs.
This implies that the error thresholds for MEP are more favorable than the ones for bipartite entanglement purification, and the use of 3-party states offers an advantage compared to the use of entangled pairs.
\subsubsection{Error thresholds for repeater cycle}
We now consider a repeater cycle, where the connection of three 3-party GHZ states is followed by appropriate MEP.
All involved operations are considered to be noisy, where we consider single qubit and two-qubit CNOT operations as elementary gates.
To obtain a threshold for the repeater cycle in the gate based model, a specific purification protocol has to be considered.
The threshold is determined by the amount of acceptable noise per gate, such that after connection of three elementary states followed by MEP,
the resulting state is still entangled and has at least the same fidelity as initially. Notice that we make use of the fact that the connection
of three GHZ states offers an intrinsic error detection capability, which allows one to obtain higher fidelities at the price that also the connection
procedure is non-deterministic. For the alternating MEP protocol \cite{Du03,Du05} we find $p_\mathrm{th} \approx 0.9581$. With more advanced
MEP schemes \cite{Reiter} (see appendix \ref{sec:adaptive}), this can be enhanced to $p_\mathrm{th} \approx 0.9490$. Then, the threshold for channel noise $q_\mathrm{min}$ at
the lowest level depends on $p$ and also the specific protocol used, where before the first connection an additional MEP is applied. While the direct multipartite approach
using these MEPs may not be optimal, it should be noted that for $p$ close to $1$ some of them already allow for a better $q_\mathrm{min}$ than is fundamentally possible
for a bipartite approach. Additional explanations and results are
provided in appendix \ref{MEPPReiter}.
\subsubsection{Basic repeater without entanglement purification}
Another interesting quantity to look at is the number of iterations leading to the next repeater level can be performed before entanglement purification becomes necessary. This also provides thresholds for repeater schemes that operate without entanglement purification.
Figure \ref{fig:multi_levels} shows the thresholds for the local noise parameters $p$ and channel noise $q$ for different numbers of connection operations
such that the state remains distillable. The maximal reachable distance
is shown in figure \ref{fig:steps} for $p=q$.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{multi_levels.pdf}
\caption{\label{fig:multi_levels} GHZ states are connected $n$-times in a concatenated way, i.e. $3^n$ states are connected to form a three-party GHZ state of distance $2^n$. The thresholds before state becomes disentangled is shown.}
\end{figure}
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{steps.pdf}
\caption{\label{fig:steps} GHZ states are connected $n$-times in a concatenated way, i.e. $3^n$ states are connected to form a three-party GHZ state of distance $2^n$. The maximum distance $2^n$ before state becomes disentangled is shown, where the given numbers correspond to multiples of elementary distance $L_0$ of the elementary GHZ states.}
\end{figure}
\section{Measurement-based implementation \label{sec:measurementbased} }
One may also consider a measurement-based implementation of entanglement purification and connection \cite{Zw12,Zw13,Zw14,Zw15}. In such an approach, task-specific entangled resource states are used to perform the connection and the purification procedure. Information processing takes place by coupling input particles via Bell-measurements to the particles of the (locally prepared) resource state. Since all the operations used are Clifford operations, the resource states are graph states, and consist only of input and output qubits \cite{Zw12,Zw13,Zw14,Zw15}.
The sole source of noise is given by imperfect resource states (and imperfect Bell measurements), which we model by depolarizing noise with parameter $p$ acting on each of the particles of the resource state as described above. For bipartite entanglement purification and 1D quantum repeaters \cite{Zw12}, it was shown that such an approach offers very high error thresholds, more than $13\%$ noise per particle for fault-tolerant quantum computation \cite{Zw14}, and more than $23\%$ noise per particle for bipartite entanglement distillation \cite{Zw13}. Noise can in fact be moved from resource states to input states under Bell measurements, which leaves us with perfect protocols applied to slightly noisier input states, and noise only acts on output particles \cite{Zw13,Zw15}.
Noisy resource states --that are used to implement the desired connection or purification operations-- are considered to be of the form $\prod_a{\cal E}_p^{(a)}|\psi\rangle\langle \psi|$, i.e. local depolarizing noise acting on each of the particles of the perfect resource state. This leads to an exponentially decreasing fidelity w.r.t. number of qubits, and incorporates that multi-qubit resource states are more difficult to prepare.
The error thresholds can be determined by considering the thresholds of the ideal purification protocol and taking local noise on the output particles into account as we show in detail below. Notice that the thresholds for MEP and the full quantum repeater are fact the same, and are protocol independent. This was shown in \cite{Zw15} for the 1D repeater, and the same argument holds in the 2D case. This follows from the fact that MEP and connection can be merged into a single resource state of minimal size that consists of only input and no output particles. As we have already seen above, the threshold for MEP for a 3-party GHZ state is as large as $p_{c} = pq = 0.5567$, leading to a threshold for the 2D quantum repeater based on tripartite GHZ states of $p_{\rm th} \leq \sqrt{p_c} \approx 0.7461$. That is, local noise of more than $25\%$ per particle is acceptable for resource states, which is even higher than for bipartite strategies.
\subsection{Noise in the measurement-based implementation}
In a measurement based approach any completely positive map $\mathcal{M}$ acting on $n$ qubits
can be probabilistically implemented using the resource state $\rho_\mathcal{M} = \mathbbm{1} \otimes \mathcal{M} \ketbra{\Phi^+}^{\otimes n}$
and utilizing Bell measurements to read in the input state. All the maps we use, including connection and multiparty entanglement purification, consist only of Clifford gates. This implies that they can be deterministically realized in such a measurement-based setup using resource states of minimal size, involving only input and output particles \cite{Zw12,Zw13,Zw14,Zw15}.
Noise that affects this resource state
naturally alters the effective map that is implemented. In the case of a local noise channel $\mathcal{E}$ acting
on each qubit of the resource state the analysis is straightforward because local noise can be shifted
freely between the two qubits on which a Bell measurement is performed \cite{Zw13}:
\begin{equation}
\mathcal{P}_B^{1,2} \mathcal{E}^{(1)} \rho = \mathcal{P}_B^{1,2} \mathcal{E}^{(2)} \rho
\end{equation}
with the superoperator $\mathcal{P}_B$ describing the projection on a Bell state. Therefore the noise on the
qubits functioning as the read-in of the resource state can be transferred directly to the input state and it
is easily checked that the effective map the noisy resource state implements is given by $\mathcal{E}^{\otimes m} \mathcal{M} \mathcal{E}^{\otimes n}$
for a map $\mathcal{M}$ with $n$ input and $m$ output qubits.
Thus we can simply consider local noise channels being applied to all input qubits followed by perfect purification and connection operations and
finally noise on the output qubits that still remain after the procedure. This makes the analysis of error thresholds and performance of such a measurement-based scheme particularly simple.
\subsection{Finite purification steps in the measurement based scenario}
In principle, it is possible to perform several purification rounds, and all repeater steps at all scales with a single (large) resource state at each node. This leads to the asymptotic error threshold of 25\% LDN per particle for a 3D repeater based on 3-party GHZ states announced above. However, here we assume that we only perform one repeater step at once with a particular resource state, and only a limited (small) number of purification steps. This is relevant for a small scale implementation with limited resources, as the required resource state are small.
We consider $m$ purification steps, where each step consists of the application of protocols P1 and P2. This is followed by the connection of the resulting GHZ states. Figure \ref{fig:scenario_b} shows the parameter region where this approach leads to an increase in fidelity after the first level, i.e. where a repeater cycle can be maintained.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{scenario_b.png}
\caption{\label{fig:scenario_b} Parameter regions where implementing the repeater connection and $m$ purification steps using protocols P1 and P2 in a
measurement based way leads to an increase in fidelity.}
\end{figure}
\section{Comparison to architectures based on 1D repeaters \label{sec:1dcompare} }
We remark that one may also use an architecture based on a combination of 1D repeater schemes that allow one to establish Bell states between pairs of parties via bipartite entanglement purification and connection. Depending on the required task, this has to be done along a single connection line in the network (on-demand generation of pairwise entanglement), between several communication partners (generation of pairwise entangled states that are in a final step connected to form the desired multiparty state), or by establishing entangled pairs of various distance, and in different directions (network with pre-prepared bipartite states where all parties of the network can communicate).
There are many (intermediate) strategies how these tasks can be achieved. In order to compare the new 2D approach with 1D architectures, we consider the performance of both schemes under non-ideal conditions where state preparation, channels and gates are noisy, which we describe using the error model from above. This takes into account that the generation of elementary GHZ states is more difficult than the generation of entangled pairs. We use the recurrence entanglement-purification protocol of \cite{De96} in the 1D case, which is usually used in repeater schemes due to its large error thresholds and good performance \cite{Br98,Du07}.
We find that the direct generation of multi-party GHZ states as proposed here has clear advantages compared to strategies based on 1D repeaters with respect to several figures of merit. First, as shown in \cite{Kr06}, higher fidelities can be reached when the goal is to establish entangled states shared between three or more parties, e.g. 3-party GHZ states. Second, the required number of storage particles
per node is smaller, as we discuss below. Third, we have shown that MEP protocols and the whole repeater for the 3-party GHZ state admit higher error thresholds per particle than bipartite protocols.
Networks based on 1D quantum repeaters may still be more efficient in certain parameter regimes or for specific tasks, e.g. the preparation of long-distance bipartite entanglement or certain multipartite entangled states (see \cite{Kr06} for an analysis of direct creation of multipartite states using bipartite and multipartite strategies). While the optimal strategy for a given task may well be a combination of bipartite and multipartite strategies, we have shown that in certain parameter regimes a direct 2D approach outperforms 1D strategies.
\subsection{Storage requirements for multipartite and bipartite networks}
In a full 2D triangular network, where 3-party GHZ states between all nodes are available as illustrated in Fig. \ref{fig:illu_layers}, the storage of final states requires three particles per node and coarse-graining level. In contrast, a system based on 1D repeaters on a triangular lattice requires either 4 or 6 particles, depending on whether one includes connection lines in two or all three directions. In both cases, this does not take into account that several copies might be needed to perform entanglement purification.
The multipartite approach offers an advantage concerning storage, mainly because storing one qubit is enough to have a connection
to multiple parties. The specific advantage however depend on which features one demands from the network. If one wants to consider a full triangular network for 3-party GHZ states - that is after
constructing states of any coarse-graining level the structure formed is again a triangular lattice - it is necessary to use all the states depicted in figure \ref{fig:illu_layers}.
That means all the holes in the Sierpinski triangle structure that appears when looking at different repeater levels have to be filled in with states up to the highest level possible.
For this particular setup a repeater station must be able to store 3 qubits per coarse-graining level it participates in. In contrast, with a bipartite strategy building up a full network
needs connections in all 6 directions at each node, which means 6 qubits per coarse-graining level have to be stored at a repeater station. Notice that this takes into account only the resulting states, while several copies are required to perform entanglement purification.
To guarantee that every party has access to the network in some form, the full scheme described above is not necessary. For example in the bipartite case it suffices to build a network in two
directions to reach every node in the network. However, if one requires the network to be structured in a way that each repeater station is only one connection away from a
repeater station of the next higher level, some connections in the third direction are required as well so the overall scaling does not change. Even when dropping this requirement
that still necessitates four qubits per coarse-graining level to be stored at repeater stations for a network based on 1D repeaters.
Even the not optimized multipartite network still scales better, so there is definitely a
storage advantage for the multipartite strategy. A switch to a quadratic lattice does not change this, although the third direction along the diagonals would probably not be used for the bipartite approach as the longer base distance for these connections would result in higher loss and error rates.
\section{Implementation using trapped ions \label{ions} }
There are intense efforts gearing up all over the world to build small--scale quantum networks and to connect multiple nodes \cite{Sa11}. So far, it has been shown on different platforms how two remote nodes can be entangled and the basic building blocks of a 1D quantum repeater have been experimentally demonstrated \cite{Chou:2005aa, Moehring:2007aa, Hofmann72, Hensen:2015aa}.
In principle, the 2D quantum repeater can be implemented in any system with a quantum light-matter interface and with the ability to perform quantum gates and measurements at each node. While several approaches are very promising, including atomic ensembles in microcells \cite{Borregaard2015} or NV centers in photonic crystal cavities \cite{Li2015,PhysRevA.72.052330}, we concentrate here on a setup with trapped ions, where all relevant building blocks have already been realized \cite{RevModPhys.82.1209, Northup:2014aa}.
While a scalable long-distance implementation requires MEP, a limited approach based on the generation and connection of elementary multipartite states without entanglement purification still allows to obtain networks over considerable distances. We analyze such a quantum network below, which shows concretely how an ion-based 1D repeater \cite{PhysRevA.79.042340} can be extended to 2D networks.
\subsection{Distributing a GHZ state with the 2D quantum repeater}
We concentrate on a scheme based on the generation of 3-party GHZ states using ion-photon entanglement and a suitable linear optics setup for projecting the photons \cite{Wang2009}. An illustration is provided in Fig. \ref{GHZ3}. These elementary GHZ states are connected via swap operations. The swap operation is simply a (deterministic) Bell measurement, which can be implemented with an entangling gate and single-qubit measurements. As a simplified repeater scheme without entanglement purification is considered, the number of states that can be connected is limited.
\begin{figure}[htb]
\centering
\includegraphics[scale=0.3]{GHZ3dir.pdf}
\caption{Illustration of creating a three-qubit GHZ state. The box denotes the linear optics setup and photon detectors, the red dots represent the ions in the cavities.}
\label{GHZ3}
\end{figure}
We would like to mention that one can prepare arbitrary graph states, e.g the 8-qubit cluster state, by creating a Bell pair for each edge in the graph and subsequent local operations/projections at each node.
\subsection{Parameters}
We assume the following parameters: ion-photon entanglement with 99.5\% fidelity, single- (two-) qubit operations with 0.1 (0.5) LDN \cite{Stute2012,Schindler2013}, 90\% photon detector efficiency \cite{detector1,detector2}, 90\% probability of ion emitting single photon and successful frequency conversion to telecom wavelength \cite{Stute2012,Zaske2012,Fernandez13}, and standard telecom
fibers with attenuation length of 22 km. Notice that all these parameters have already been achieved in experiments, except the probability of emitting a photon which is however expected to be reachable with
present-day set-ups \cite{Northup:2014aa}.
For the comparison of the 1D and 2D approach, we remark that this error model takes into account that the generation of GHZ states is more difficult than the generation of entangled pairs.
\subsection{Fidelities}
We compare our intrinsic 2D strategy with a 1D approach. In this setting one prepares Bell pairs between nodes A and B, and between nodes A and C using a 1D quantum repeater. The GHZ state between nodes A, B, and C can be created with an additional ancilla qubit in state $\ket{+}$ at node A and CNOT gates between this qubit, which serves as control qubit, and the two qubits, which are part of the shared Bell pairs, followed by a measurement of the target qubits in the computational basis.
Surprisingly, the 2D approach allows one to obtain slightly higher fidelities for the resulting long-distance GHZ state.
The numbers for various numbers of links are summarized in table \ref{compfid}.
\begin{table}[h]
\vspace{0.3cm}
\centering
\begin{tabular}{| c |c | c | c | }
\hline
& 4 links & 8 links & 16 links \\ \hline
1D & $86.44\%$ & $74.85\%$ & $57.18\%$ \\ \hline
2D & $90.96\%$ & $78.03\%$ & $58.77\%$ \\
\hline
\end{tabular}
\caption{\label{compfid} Fidelities for several numbers of links using the 1D and the 2D approach.}
\end{table}
Notice that this holds for negligible memory errors. For ion traps the dominant error source is collective dephasing \cite{Monz2011}, so that one needs to either assume that the coherence time is large compared to the distribution time or encode quantum information into a decoherence-free-subspace \cite{Zanardi1997,Lidar1998} (see also \cite{DFS}). We have performed an analysis of the influence of memory errors and the usage of a decoherence-free-subspace encoding in \cite{DFS} for a 1D repeater scheme. The techniques are also applicable in the 2D approach presented here.
Note also that an all--optical implementation, generalizing the work of \cite{Azuma2015}, is conceivable, provided one uses a one-way entanglement purification protocol.
\subsection{Distribution times}
The derivation of the distribution time of the GHZ states is analogous to \cite{DFS}. We use the parameters listed above.
\subsubsection{Distribution time using 2D quantum repeater}
In this approach the photons are sent from the corners of the triangle to the center, where the linear optics elements and photon detectors \cite{Wang2009} are placed (for concreteness we assume an equilateral triangle). The distance to the center is then given by $\tfrac{L_0}{\sqrt{3}}$, where $L_0$ is the distance between the nodes. The probability of creating the GHZ state between three elementary nodes is then given by
\begin{equation}
P_{elem}=\frac{1}{4}p_{ion}^3 \eta_d^3\eta_t^3,
\end{equation}
with $\eta_t=\text{exp}(-L_0/(\sqrt{3}L_{att}))$. The time for distributing the GHZ state is then given by (see also \cite{DFS})
\begin{equation}
T=\frac{L_0}{c}\sum_{i=1}^{3^{n}}\frac{1}{1 - (1-P_{elem})^i},
\end{equation}
for a repeater with $2^n$ links. The resulting distribution times are shown in Fig. \ref{figure_times}.
\begin{figure}[ht]
\centering
\includegraphics[width=\columnwidth]{times2DNEW.pdf}
\caption{ \label{figure_times} Distribution time using an ion-trap based implementation as a function of the distance.}
\end{figure}
\subsubsection{Distribution time using 1D quantum repeater}
Here the probability of establishing an elementary Bell pair is given by
\begin{equation}
P_{elem}=\frac{1}{2}p_{ion}^2 \eta_d^2\eta_t,
\end{equation}
with $\eta_t=\text{exp}(-L_0/L_{att})$. The time for distributing the GHZ state is then given by (see also \cite{DFS})
\begin{equation}
T=\frac{1}{p_{\rm suc}}\frac{L_0}{c}\sum_{i=1}^{2\cdot2^{n}}\frac{1}{1 - (1-P_{elem})^i},
\end{equation}
for a repeater with $2^n$ links, where $p_{suc}$ is the total success probability for appropriate Bell measurement outcomes in the connection processes.
\subsubsection{Comparison of 1D and 2D strategy}
The distribution times for 4, 8 and 16 links are plotted in Fig. \ref{comptime}, both for the 1D and the 2D approach. The times for the 1D approach are clearly smaller, and it should also be noted that the 2D approach requires more resources (total number of nodes and ions). However with both approaches one can distribute a GHZ state with fidelity that is large enough to violate a Bell inequality \cite{Mer90,Bel93,Lanyon2014} over a distance of 1000 km in about one second.
\begin{figure}[htb]
\centering
\includegraphics[width=\columnwidth]{times1D2DNEW.pdf}
\caption{Distribution time for a three-qubit GHZ state using the 1D and 2D quantum repeater as a function of distance.}
\label{comptime}
\end{figure}
\section{Summary and Outlook \label{sec:summary}}
We propose a 2D repeater architecture based on self-similar growing structures of multipartite entangled states. The favorable scaling and the high error thresholds of 1D repeater architectures do not only carry over to the 2D approach, 2D repeaters provide in fact higher error thresholds while at the same time, the required resources for storage are reduced. Furthermore, the design of quantum repeaters is not restricted to GHZ or 2D cluster states on 2D lattices. In fact, different lattice structures (including 3D arrays) and target graph states are conceivable.
Notice that MEP protocols for all graph states exist \cite{Du03,Du05,Kr06a,Du07}, such that the procedure of connecting and purifying states at different scales or growing size is universally applicable.
The proposed approach is intrinsically two-dimensional and seems thus ideally suited for real world quantum networks. In particular, such networks offer a high degree of flexibility, with potential two-party and multi-party applications. As outlined above, various platforms offer themselves for an implementation in the near future, opening the way towards real-world application of quantum technology in large-scale quantum networks.
{\em Acknowledgements} This work was supported by the Austrian Science Fund (FWF): P24273-N16, P28000-N27, by the Swiss National Science Foundation (SNSF) through Grant number PP00P2-150579, the Army Research Laboratory Center for Distributed Quantum Information via the project SciNet and the EU via the integrated project SIQS.
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
In this note, we consider finite and simple graphs.
For a graph $G$, we use $V(G)$ and $E(G)$ to denote the set of vertices and edges of $G$, respectively. A {\it subgraph} of $G$ is a graph whose vertex set is a subset of $V(G)$ and whose edge set is a subset of $E(G)$. We say that a subgraph $H$ is an {\it induced subgraph} of $G$ if, for any $x,y \in V(H)$, $xy \in E(H) \mbox{ iff } xy \in E(G)$.
Let $G$ be a graph and $S \subset V(G).$ Then $G[S]$, the subgraph of $G$ induced by $S,$ denotes the graph with vertex set $S$ and edge set $\{uv \in E(G): u,v \in S\}$, and let $G - S=G[(V(G)\setminus S].$ When $S= \{x,y\}$, we often write $G-x-y$ instead of $G-S$. For a positive integer $k$, a {\it proper $k$-coloring} of a graph $G$ is a function $c$ from $V(G)$ to a set of $k$ colors such that $c(u) \ne c(v)$ for any $uv \in E(G).$ A graph $G$ is {\it $k$-colorable} if $G$ has a proper $k$-coloring. We use $\chi(G)$ to denote the smallest integer $k$ such that $G$ is $k$-colorable, which is known as the {\it chromatic number} of $G$. Further we denote by $\omega(G)$ and $\alpha(G)$ the size of the largest clique and independent set in $G$, respectively, and $N(v)$ the set of vertices adjacent to the vertex $v$ in $G$.
We say that a connected graph $G$ with chromatic number $n$ is {\it $n$-double-critical}, if, for any $xy \in E(G)$, $\chi(G-x-y) = n-2$. It is easy to see that the complete graph $K_n$ is $n$-double-critical. The following elegant conjecture was posed by Erd\H{o}s and Lov\'asz \cite{Er66} more than fifty years ago.
\begin{conjecture} \label{DCGC}
$K_n$ is the only $n$-double-critical graph.
\end{conjecture}
It is easy to see that Conjecture \ref{DCGC} holds for $n \le 3.$ With some extra work it can also be verified for $n=4.$ In 1986, Stiebitz \cite{St86} showed that Conjecture \ref{DCGC} is true for $n=5.$ He proved the existence of $K_4$ in every 5-double-critical graph by considering uniquely 3-colorable subgraphs of $G$.
However, this technique does not seem to generalize to finding a larger clique in an $n$-double-critical graph with $n \ge 6.$
The double-critical graph conjecture is a special case of a more general conjecture, the so-called Erd\H os-Lov\'asz Tihany conjecture \cite{Er66}: for any graph $G$ with $\chi(G) > \omega(G)$ and any two integers $k, l \ge 2$ with $k+l = \chi(G)+1$, there exists a partition $(S, T)$ of the vertex set such that $\chi(G[S]) \ge k$ and $\chi(G[T]) \ge l$. The Erd\H os-Lov\'asz Tihany conjecture was proved for various cases: $(k, l)=(2,2),(2,3),(2,4),(3,3),(3,4),(3,5)$ (see: \cite{BJ99, Mo86, St86, St87}). Kostochka and Stiebitz \cite{KM08} showed that it is true for all $(k, l)$ for line graphs. Balogh et al. \cite{BKPS09} generalized this result to quasi-line graphs (a graph is a quasi-line graph if the neighbors of every vertex $v$ can be expressed as the union of two cliques) and graphs of independence number $2$.
A {\it claw} is a $4$-vertex graph with one vertex of degree 3 and the others of degree 1. For convenience, we write $(v;v_1,v_2,v_3)$ to denote a claw in which $v$ has degree 3. A graph $G$ is {\it claw-free} if it does not have a claw as an induced subgraph. Note that the graphs from both families in Balogh et al.'s result \cite{BKPS09} are claw-free. It would be interesting to know whether the double-critical graph conjecture, or the Erd\H os-Lov\'asz Tihany conjecture holds for all the claw-free graphs. As a step in that direction, we prove the following theorem.
\begin{theorem}\label{thm_main}
Let $G$ be a double-critical graph with $\chi(G)=6$. If $G$ is claw-free, then $G \cong K_6$.
\end{theorem}
The rest of this note is organized as follows. In the next section we will prove several lemmas that will be repeatedly used throughout the proof of Theorem \ref{thm_main}. Section \ref{sec_main} contains the proof of Theorem \ref{thm_main}.
\section{Lemmas}
Given a graph $G$ of chromatic number $n$, and a proper $n$-coloring of $G$, all vertices of the same color form a {\it color class}. By definition, each color class is an independent set in $G$.
\begin{lemma}\label{clique}
Let $G$ be an $n$-double-critical graph. If $\omega(G) \ge n-1$, then $G \cong K_n$. Thus if $G \not\cong K_n$, then $\omega(G) \le n-2$.
\end{lemma}
\noindent {\it Proof. }
Suppose $\omega(G) \ge n-1$. Let $v_1,\ldots, v_{n-1}\in V(G)$ be the vertices that induce a copy of $K_{n-1}$. Among all the proper $n$-colorings of $G$ with color classes $V_1, \ldots, V_n$, and $v_i \in V_i$ for $1 \le i \le n-1$, we choose one that minimizes $|V_n|$.
Let $v_n \in V_n$. We claim that $N(v_n) \cap V_i \ne \emptyset$ for $i=1,\ldots,n-1$.
Suppose not. Without loss of generality we may assume that $N(v_n) \cap V_1 = \emptyset$. If $V_n = \{v_n\}$ then the independent sets $V_1 \cup \{v_n\}, V_2, \ldots ,V_{n-1}$ form an proper $(n-1)$-coloring of $G$, contradicting the assumption that $\chi(G)=n$. If $V_n \setminus \{v_n\} \ne \emptyset$, then $V_1 \cup \{v_n\}, V_2, \ldots ,V_n \setminus \{v_n\}$ are the color classes of an $n$-coloring of $G$. Thus we have a contradiction to the minimality of $|V_n|$.
We now show that $v_n$ is adjacent to every vertex in $\{v_1, \cdots, v_{n-1}\}$. Suppose not. Without loss of generality assume that $v_1 \not\in N(v_n)$.
Then, by the above claim, $v_n$ is adjacent to some $y \in V_1 \setminus \{v_1\}$.
However $\chi(G-v_n-y)=n-2$, a contradiction, since $\{v_1,\ldots,v_{n-1}\}$ induces a copy of $K_{n-1}$ in $G-v_n-y$.
Hence $\{v_1,\ldots,v_n\}$ induces $K_n$ in $G$.
If $V(G) = \{v_1,\ldots,v_n\}$, then $G \cong K_n$.
Suppose $V(G) \ne \{v_1,\ldots,v_n\}$, and let $x \in V(G)$ such that $x \notin \{v_1,\ldots,v_n\}$.
Since $G$ is connected, there exists $z \in V(G)$ such that $xz \in E(G)$. However, $G-x-z$ contains a clique on $n-1$ vertices, which contradicts that $\chi(G-x-z)=n-2$, and thus $G \cong K_n$.
\hfill \rule{4pt}{7pt}
\bigskip
\begin{lemma}\label{existence}
Let $G$ be an $n$-double-critical graph that is claw-free. For $xy \in E(G)$, let $V_1, \ldots, V_{n-2}$ be the color classes of an $(n-2)$-coloring of $G-x-y$. Then $N(x) \cap N(y) \cap V_i \ne \emptyset$ for $i \in \{1,\ldots,n-2\}$.
\end{lemma}
\noindent {\it Proof. }
Suppose not. Then, without loss of generality, we may assume that $N(x) \cap N(y) \cap V_{n-2} = \emptyset$. Let $V_x = \{x\} \cup (V_{n-2} \setminus N(x))$ and $V_y = \{y\} \cup V_{n-2} \setminus V_x$. Note that $V_x$ and $V_y$ are independent sets. Now $V_1,\ldots,V_{n-3},V_x,V_y$ are the color classes of an $(n-1)$-coloring of $G$, contradicting the assumption that $\chi(G)=n$.
\hfill \rule{4pt}{7pt}
\bigskip
The {\it degree} of a vertex $v$, denoted $d(v)$, in a graph is the number of edges incident to $v$. We denote by $\Delta(G)$ and $\delta(G)$ the maximum and minimum degree of a vertex in $G$ respectively. The following lemma shows that for $\chi(G)=6$, it suffices to consider 6-double-critical graphs in which every pair of adjacent vertices has $4$ or $5$ common neighbors.
\begin{lemma}\label{neighbor}
Let $G$ be a $6$-double-critical graph that is also claw-free. If $G \not\cong K_6$, then for any $xy \in E(G), 4 \le |N(x) \cap N(y)| \le 5$. If, in addition, $|N(x) \cap N(y)| =5$, then $G[N(x) \cap N(y)] \cong C_5$.
\end{lemma}
\noindent {\it Proof. }
We may assume that $\Delta(G[N(x)\cap N(y)])\le 2$.
For, otherwise, let $v \in N(x)\cap N(y)$ and let $v_1,v_2,v_3 \in N(x)\cap N(y)$ such that $vv_i \in E(G), i=1,2,3$.
Since $(v;v_1,v_2,v_3)$ does not induce a claw in $G$, there exist $i,j \in \{1, 2, 3\}$ such that $i \ne j$ and $v_iv_j \in E(G)$.
Thus $\{v,v_i,v_j,x,y\}$ induces a copy of $K_5$ in $G$.
Hence by Lemma \ref{clique}, $G \cong K_6$.
We may also assume that $\delta(G[N(x)\cap N(y)]) \ge |N(x)\cap N(y)|-3$.
For, otherwise, let $v \in N(x)\cap N(y)$ and let $v_1,v_2,v_3 \in N(x)\cap N(y)$ such that $vv_i \notin E(G), i=1,2,3$.
Then $v_iv_j \in E(G)$ for all distinct $i,j \in \{1,2,3\}$, since $(v;v_i,v_j,x)$ does not induce a claw in $G$.
Therefore $\{v_1,v_2,v_3,x,y\}$ induces a copy of $K_5$ in $G$.
Once again by Lemma \ref{clique}, $G \cong K_6$.
Hence, if $G \not\cong K_6$, $|N(x)\cap N(y)|-3 \le \delta(G[N(x)\cap N(y)]) \le \Delta(G[N(x)\cap N(y)]) \le 2$.
Thus $|N(x)\cap N(y)| \le 5$. On the other hand, by Lemma \ref{existence}, $|N(x)\cap N(y)| \ge 6-2 = 4$.
Now suppose $|N(x) \cap N(y)| =5$. Then $2=|N(x)\cap N(y)|-3 \le \delta(G[N(x)\cap N(y)]) \le \Delta(G[N(x)\cap N(y)]) \le 2$.
Hence $\delta(G[N(x)\cap N(y)]) = \Delta(G[N(x)\cap N(y)])=2$, so every vertex in $G[N(x)\cap N(y)]$ has degree $2$. The only $2$-regular graph on $5$ vertices is $C_5$, thus $G[N(x)\cap N(y)] \cong C_5$.
\hfill \rule{4pt}{7pt}
\bigskip
The following lemma is an easy consequence of $G$ being claw-free.
\begin{lemma}\label{easy}
Let $G$ be a claw-free graph, and $S$ an independent set of $G$. Suppose $x \in V(G) \setminus S$. Then $|N(x) \cap S| \le 2.$
\end{lemma}
\noindent {\it Proof. }
Suppose $|N(x) \cap S| \ge 3.$ Let $x_1,x_2,x_3 \in N(x) \cap S.$ Since $S$ is an independent set, $(x;x_1,x_2,x_3)$ induces a claw in $G$, a contradiction.
\hfill \rule{4pt}{7pt}
\section{Proof of the Main Result} \label{sec_main}
Theorem \ref{thm_main} follows from the two lemmas in this section. From Lemma \ref{neighbor}, we may assume that the number of common neighbors of any two adjacent vertices is either $4$ or $5$. The first lemma settles the case when there exists a pair with $4$ common neighbors.
\begin{lemma}\label{four}
Let $G$ be a $6$-double-critical graph that is claw-free. If $|N(x) \cap N(y)| =4$ for some $xy \in E(G)$, then $G \cong K_6$.
\end{lemma}
\noindent {\it Proof. }
For an arbitrary $xy \in E(G)$, by Lemma \ref{neighbor}, we have $|N(x)\cap N(y)|\ge4$. Thus $d(x) \ge 5$ and $d(y) \ge 5$. Moreover, if $V_1,V_2,V_3,V_4$ denote the color classes of a 4-coloring of $G-x-y$, it follows from Lemma \ref{easy} that $|N(x) \cap V_i| \le 2$ and $|N(y) \cap V_i| \le 2$ for $i \in \{1,2,3,4\}$. Thus $d(x) \le 9$ and $d(y) \le 9$.\\
\textbf{Claim 1.} If $xy \in E(G)$ and $|N(x) \cap N(y)| = 4$, then $d(x),d(y) \in \{7,8\}$.
Let $N(x)\cap N(y)= \{v_1,v_2,v_3,v_4\}$, and let $V_1,V_2,V_3,V_4$ be the color classes of a 4-coloring of $G-x-y$. By Lemma \ref{existence}, we may assume $v_i \in V_i$, $i=1,2,3,4$.
Suppose $d(x) \in \{5,6\}.$ Then we may assume that $N(x)\cap V_i = {v_i}$ for $i \in \{2,3,4\}$. Since $|N(x)\cap N(v_1)| \ge 4$, $v_1v_i \in E(G)$ for $i \in \{2,3,4\}$. If, for every pair of distinct $i,j \in \{2,3,4\}$, $v_iv_j \not\in E(G)$, then $(v_1;v_2,v_3,v_4)$ induces a claw in $G$. Otherwise, there exist distinct $i,j \in \{2,3,4\}$ such that $v_iv_j \in E(G)$. Then $G[\{v_1,v_i,v_j,x,y\}] \cong K_5$. Hence $G \cong K_6$ by Lemma \ref{clique}.
Now suppose that $d(x) = 9.$
Let $N(x) \setminus \{v_1,v_2,v_3,v_4,y\} = \{u_1,u_2,u_3,u_4\}$. By Lemma \ref{easy}, we may assume $u_i \in V_i$ for $i\in \{1,2,3,4\}$.
For any distinct $j,k \in \{1,2,3,4\}$, $u_ju_k \in E(G)$; for otherwise $(x;u_j,u_k,y)$ induces a claw.
Thus $G[\{x,u_1,u_2,u_3,u_4\}] \cong K_5$.
Hence, by Lemma \ref{clique}, $G \cong K_6$.\\
\textbf{Claim 2.} If $xy \in E(G)$ and $|N(x) \cap N(y)| = 4$, then $d(x)=d(y)=8$.
Let $N(x)\cap N(y)= \{v_1,v_2,v_3,v_4\}$, and let $V_1,V_2,V_3,V_4$ be the color classes of a 4-coloring of $G-x-y$. By Lemma \ref{existence}, we may assume $v_i \in V_i, i=1,2,3,4$.
Suppose $d(x) = 7$
and, by Lemma \ref{easy} and by symmetry, let $u_i \in N(x)\cap V_i \setminus v_i$, $i \in \{1,2\}$.
Since $|N(x)\cap N(u_1)| \ge 4$ and $u_1 \not\in N(y)$, $u_1u_2,u_1v_2,u_1v_3,u_1v_4 \in E(G)$.
Similarly, since $|N(x)\cap N(u_2)| \ge 4$ and $u_2 \not\in N(y)$, $u_2v_1,u_2v_3,u_2v_4 \in E(G)$.
We claim that $y$ does not have a neighbor in $V_1\setminus \{v_1\}$ or $V_2 \setminus \{v_2\}$.
For otherwise, suppose there exists $w_1 \in V_1 \setminus {v_1}$ such that $yw_1 \in E(G)$.
Since $|N(y)\cap N(w_1)| \ge 4$, we have $|N(w_1) \cap \{v_2,v_3,v_4\}| \ge 2$. (This is because $d(y) \le 8$, so $y$ has at most two neighbors not from $\{x, v_1, v_2, v_3, v_4, w_1\}$. If $w_1$ is adjacent to at most one vertex from $\{v_2, v_3, v_4\}$, then $|N(y) \cap N(w_1)| \le 3$.) Similarly since $|N(x)\cap N(v_1)| \ge 4$, we have $|N(v_1) \cap \{v_2,v_3,v_4\}| \ge 2$.
Thus there exists $i \in \{2,3,4\}$ such that $v_i \in N(v_1) \cap N(w_1)$.
Note $(v_i;v_1,u_1,w_1)$ induces a claw in $G$, a contradiction.
Therefore by Claim 1 and Lemma \ref{easy}, $d(y)=7$ and $|N(y) \cap V_i| = 2$ for $i \in \{3,4\}$.
Let $w_i \in N(y) \cap V_i \setminus \{v_i\}$ for $i \in \{3,4\}$.
Since $|N(y)\cap N(w_3)| \ge 4$ and $w_3 \not\in N(x)$, $w_3w_4,w_3v_1,w_3v_2,w_3v_4 \in E(G)$,
and similarly since $|N(y)\cap N(w_4)| \ge 4$ and $w_4 \not\in N(x)$, $w_4v_1, w_4v_2, w_4v_3 \in E(G)$.
We may assume that $v_3v_4 \not\in E(G)$; otherwise $G[\{x,u_1,u_2,v_3,v_4\}] \cong K_5$ and, hence, $G \cong K_6$ by Lemma \ref{clique}. Similarly we may assume $v_1v_2 \not\in E(G)$, otherwise $G[\{y,v_1,v_2,w_3,w_4\}] \cong K_5$ and once again $G \cong K_6$ by Lemma \ref{clique}.
Since $|N(x)\cap N(v_1)| \ge 4$ and $v_1v_2 \not\in E(G)$, $v_1v_3,v_1v_4 \in E(G)$.
Similarly since $|N(x)\cap N(v_2)| \ge 4$ and $v_1v_2 \not\in E(G)$, $v_2v_3,v_2v_4 \in E(G)$.
We claim that $u_iw_j \in E(G)$ for all $i \in \{1,2\}$ and $j \in \{3,4\}$. Suppose not. By symmetry, we assume $w_4u_1 \not\in E(G)$.
Since $|N(v_2)\cap N(w_4)| \ge 4$, from the known adjacencies we so far only have $w_3, v_3, y \in N(v_2) \cap N(w_4)$, therefore there exists $w_1 \in V_1 \cup V_3$ such that $w_1v_2,w_1w_4 \in E(G).$
In fact, $w_1 \in V_1$, otherwise then $(w_4;v_3,w_1,w_3)$ induces a claw in $G$, a contradiction. Since $w_4u_1 \not\in E(G)$, $w_1 \ne u_1$.
Note that $w_1v_3 \not\in E(G)$ otherwise $(v_3;v_1,u_1,w_1)$ induces a claw and similarly $w_1v_4 \not\in E(G)$ otherwise $(v_4;v_1,u_1,w_1)$ induces a claw.
Then $(v_2;w_1,v_3,v_4)$ induces a claw in $G$, a contradiction since $G$ is claw-free.
Hence, $G[\{u_1,u_2,w_3,w_4\}] \cong K_4$.
Consider the graph $G-x-v_3$. Since $G$ is 6-double-critical, $\chi(G-x-v_3)=4$. Let $c:V(G)\rightarrow \{1,2,3,4\}$ be a 4-coloring of $G-x-v_3$. Since $G[\{u_1,u_2,w_3,w_4\}] \cong K_4$, we may assume $c(u_1)=1$, $c(u_2)=2$, $c(w_3)=3$, $c(u_4)=4$. Then $c(v_1)=1$, since $v_1u_2,v_1w_3,v_1w_4 \in E(G)$. Similarly $c(v_2)=2$ and $c(v_4)=4$. Then, since $yv_1,yv_2,yw_3,yw_4 \in E(G)$, $y$ cannot be colored by any of the colors 1, 2, 3, 4. Thus $G-x-v_3$ is not 4-colorable, a contradiction.
Therefore $d(x) = 8$. Similarly $d(y) =8$.
This completes the proof of Claim 2. \\
Now let us fix $xy \in E(G)$ with $|N(x) \cap N(y)| = 4$. Let $N(x) \cap N(y) = \{v_1,v_2,v_3,v_4\}$ and let $V_1,V_2,V_3,V_4$ be the color classes of a 4-coloring of $G-x-y$. By Lemma \ref{existence}, we may assume $v_i \in V_i$ for $i \in \{1,2,3,4\}$.
By Claim 2, $d(x)=8$. Thus let $w_i \in N(x)\cap V_i \setminus \{v_i\}$, $i \in \{1,2,3\}$.
Note that $G[\{w_1,w_3,w_3\}]$ is a clique of size 3; otherwise suppose, by symmetry $w_1w_2 \not\in E(G)$, then $(x;y,w_1,w_2)$ induces a claw in $G$.\\
\textbf{Claim 3.} For $i \in \{1,2,3\}$, $\{v_1,v_2,v_3,v_4\} \setminus \{v_i\} \subset N(w_i)$.
Suppose otherwise. Without loss of generality, we may assume $\{v_2,v_3,v_4\} \not\subset N(w_1)$. Since $|N(x)\cap N(w_1)| \ge 4$, $|N(w_1)\cap \{v_2,v_3,v_4\}| = 2$,
and $|N(x) \cap N(w_1)| =4$.
By Claim 2, it suffices to consider the case $d(w_1)=8.$
However, from Lemma \ref{easy}, $$d(w_1) \le 1+|N(w_1)\cap V_2|+|N(w_1)\cap V_3|+|N(w_1)\cap V_4| \le 7,$$
a contradiction. Hence we have Claim 3.\\
Note that for $i \in \{1,2,3\}$, $v_iv_4 \not\in E(G)$ otherwise $\{x,v_i,v_4,w_1,w_2\}$ induces a $K_5$. To avoid the claw $(y; v_i, v_j, v_4)$, we must have $v_iv_j \in E(G)$ for distinct $i, j \in \{1,2,3\}$. In this case $\{x, y, v_1, v_2, v_3\}$ induces a copy of $K_5$, and hence $G \cong K_6$ and the proof of Lemma \ref{four} is complete.
\hfill \rule{4pt}{7pt}
\medskip
Our next lemma settles the remaining case when every pair of adjacent vertices have exactly $5$ common neighbors.
\begin{lemma}\label{five}
Let $G$ be $6$-double-critical graph, and assume that $G$ is claw-free. Suppose $|N(x) \cap N(y)| \ge 5$ for all $xy \in E(G)$.
Then $G \cong K_6$.
\end{lemma}
\noindent {\it Proof. }
We prove this Lemma by way of contradiction. Suppose $G \not \cong K_6$. Then, by Lemma \ref{clique}, $K_5 \not \subset G$.
By Lemma \ref{neighbor} and Lemma \ref{four}, we may assume that $|N(x) \cap N(y)| = 5$ for all $xy \in E(G)$.
Let $N(x)\cap N(y) = \{v_1,v_2,v_3,v_4,v_5\}$, and let $V_1,V_2,V_3,V_4$ be the color classes of a 4-coloring of $G-x-y$. By Lemma \ref{existence}, we may assume that $v_i \in V_i$, $i \in \{1,2,3,4\}$ and $v_5 \in V_1$.
By Lemma \ref{neighbor}, $\{v_1,v_2,v_3,v_4,v_5\}$ induces a $C_5$ in $G$.
Without loss of generality, assume $v_1v_2,v_1v_4,v_2v_5,v_3v_4,v_3v_5 \in E(G)$ and $v_1v_3,v_2v_3,v_2v_4,v_4v_5 \not\in E(G)$.
Since $|N(x) \cap N(v_5)| = 5$, there exist $a,b \in (N(x) \cap N(v_5)) \setminus \{v_1,v_2,v_3,v_4,v_5,y\}$.
Similarly since $|N(y) \cap N(v_5)| = 5$, there exist $c,d \in (N(y) \cap N(v_5)) \setminus \{v_1,v_2,v_3,v_4,v_5,x\}$.
Note that $a,b,c,d \in (V_2 \cup V_3 \cup V_4) \setminus \{v_2,v_3,v_4\}$.
Since $N(x) \cap N(y) = \{v_1,v_2,v_3,v_4,v_5\}$, $a,b\not\in N(y)$ and $c,d \not\in N(x)$. Hence $a,b,c,d$ are pairwise distinct. Moreover $ab \in E(G)$ to avoid the claw $(x;a,b,y)$ in $G$, and $cd \in E(G)$ to avoid the claw $(y;c,d,x) \in E(G)$.
By Lemma \ref{easy} (applied to $v_5$ and $V_i$, for $i \in \{2,3,4\}$) and by the symmetry between $V_2$ and $V_3$, we may assume that $b, d \in V_4$, $a\in V_3$, and $c\in V_2$.
Since $|N(y)\cap N(v_3)| \ge 5$, there exist $z_1,z_2 \in (N(y) \cap N(v_3)) \setminus \{v_1,v_2,v_3,v_4,v_5,x,y\}$. Note that $z_i \not\in V_1$ for $i \in \{1,2\}$; otherwise $(y;v_1,v_5,z_i)$ induces a claw in $G$. Clearly $z_1,z_2 \not\in V_3$ since $V_3$ is independent.
We claim that $d \in \{z_1,z_2\}$. Suppose otherwise, $d \not\in \{z_1,z_2\}$. In this case $z_i \not\in V_4$ for $i \in \{1,2\}$ to avoid the claw $(y;d,v_4,z_i)$ in $G$. Therefore $z_1,z_2 \in V_2$, then $(y;v_2,z_1,z_2)$ induces a claw in $G$, a contradiction.
Since $d \in \{z_1, z_2\}$, we have $dv_3 \in E(G)$.
By a similar argument considering $|N(x) \cap N(v_3)| \ge 5$, $bv_3 \in E(G)$.
Thus $(v_3;b,d,v_4)$ induces a claw in $G$, a contradiction.
\hfill \rule{4pt}{7pt}
|
\section*{Introduction}
Robust approach to pricing and hedging has been an active field of research in mathematical finance over the recent years. In this approach, instead of choosing one model, one considers superhedging simultaneously under a family of models, or pathwise on a set of feasible trajectories.
It generalizes the classical approach where one only considers models which are absolutely continuous with respect to a fixed reference probability measures $\mathbb{P}$.
In such setting, absence of arbitrage is known to be equivalent to the existence of an equivalent martingale measure, result known as the first fundamental theorem of asset pricing, see e.g. \cite{DeSch}, \cite{FS04}.
When the market is complete --- i.e.\ when every contingent claim can be perfectly replicated by a self--financing trading strategy ---
the equivalent martingale measure $\mathbb{Q}$ is unique and option prices are given by their replication cost, which is equal
to the expected value of the discounted payoff under $\mathbb{Q}$.
In an incomplete market, when a perfect replication strategy does not exist,
a safe way of pricing is to use the minimum super--replication cost of the option.
Using duality techniques, this minimal super--replication price can be related to the pricing problem and expressed as the supremum of expectations of the discounted payoff over all martingale measures equivalent to $\mathbb{P}$.
One of the challenges in a robust approach is to extend this dual relationship to non-dominated (robust) context.
In continuous time models under \emph{volatility uncertainty}, such pricing--hedging duality results have been obtained by, among many others, \citet{DM06}, \cite{STZ}, \cite{NN12}, \cite{PRT}. In discrete time, general pricing--hedging duality was shown in e.g. \cite{BN13} and \cite{BFM15}.
Importantly, in a robust setting one often wants to include further market instruments which may be available for trading.
In a setup which goes back to the seminal work of \cite{Hlookback}, one often considers dynamic trading in the underlying and static trading, i.e.\ buy and hold strategies at time zero, in some European options, often call or put options with a fixed maturity. Naturally, such additional assets constraint the set of martingales measures we may use for pricing. General pricing--hedging duality results in variants of this setting, both in continuous and in discrete time, can be found in e.g.\ \cite{ABPS}, \cite{BFHMO16}, \cite{BHLP}, \cite{DS14a}, \cite{HouObloj}, \cite{GTT}, \cite{TT} and we refer to the survey papers \cite{Hreview, Obloj} for more details.
The main focus in the literature so far has been on the duality for (possibly exotic) European payoffs. However, more recently, some focus was put on American options. \cite{CoxHoeggerl}
studied the necessary (and sufficient in some cases) conditions on the American put option prices for the absence of arbitrage.
\cite{Dolinsky14} studied game options (including American options) in a non--dominated discrete--time market, but without statically traded options allowed for super--replication. \cite{Neuberger} considered a discrete--time, discrete space market with presence of statically traded European vanilla options. He observed that the superheding price for an American option may be strictly larger than the supremum of the expected (discounted) payoff over all stopping times and all (relevant) martingale measures. We refer to this situation as \emph{duality gap}. In \cite{Neuberger}, the pricing--hedging duality is then restored by using a \emph{weak} dual formulation. This approach was further exploited, with more general results, in \cite{HobsonNeuberger}. \cite{Bay} studied the same superhedging problem in the setup of \cite{BN13} but only considered \emph{strong} stopping times in their dual formulation, which leads to a duality gap in general.
More recently, and in parallel to an earlier version of this paper, \cite{BayZhou} proved a duality result by considering \emph{randomized} models, under some regularity and integrability conditions on the payoff functions.
Motivated by the above works, we endeavour here to understand the fundamental reasons why pricing--hedging duality for American options holds or fails and offer systematic reasons to mend it in the latter case.
We derive two main general results which we then apply to various specific contexts, both classical and robust.
Our first insight is that by considering a (universal) enlargement of the space, namely the time--space product structure, we can see an American option as a European option and recover the pricing--hedging duality, which may fail in the original formulation.
This may be seen as a weak formulation of the dual (pricing) problem and leads to considering a large family of stopping times.
This formulation of the dual problem is similar in spirit to \cite{Neuberger, HobsonNeuberger, BayZhou},
but our approach leads to a duality results in a more general setting, and/or under more general conditions, see Remark \ref{com:hob} and Subsection \ref{com:bay} and also \cite{HobsonNeuberger2}.
Our second main insight is that the duality gap is caused by the failure of dynamic programming principle.
To recover the duality, under the formulation with \emph{strong} stopping times, it is necessary and sufficient to consider an enlargement which restores the dynamic consistency: it is enough to consider (fictitious) extensions of the market in which all the assets are traded dynamically.
As a byproduct, we find that the dynamic trading strategy on options and the classical semi-static strategy lead to the same superhedging cost in various settings.
The first part of paper, Section \ref{duality}, presents the above two main insights in a very general discrete time framework which covers both classical (dominated) and robust (non--dominated) settings. In the second part of the paper, we apply our general results in the context of two important examples of the robust framework: the setup of \cite{BN13} in Section \ref{sec:main} and the martingale optimal transport setup of \cite{BHLP} in Section \ref{sec:mot}. We obtain suitable pricing--hedging duality for American options in both setups.
In the latter case of martingale optimal transport, there is an infinity of assets to consider and we use measure valued martingales to elegantly describe this setting.
\begin{Example}
\label{ex:intro}
We conclude this introduction with a motivating example showing that the pricing--hedging duality may fail in presence of static trading instruments and how it is recovered when the setup is augmented allowing to trade these dynamically. This example is summarized in Figure \ref{fig:Ex1}. We consider a two period model with stock price process $S$ given by $S_0=S_1=0$ and $S_2\in\{-2,-1,1,2\}$.
The American option process $\Phi$ is defined as $\Phi_1(\{S_1=0\})=1$, $\Phi_2(\{S_2\in\{-2,2\}\})=0$ and $\Phi_2(\{S_2\in\{-1,1\}\})=2$. The (pathwise) superhedging price of $\Phi$ can be easily computed and equals 2. A probability measure $\mathbb{P}$ on the space of four possible paths is uniquely described through a choice of $q_{i}=\mathbb{P}(S_2=i)\geq 0$ for $i\in\{-2,-1,1,2\}$ satisfying $q_2+q_1+q_{-1}+q_{-2}=1$. The martingale condition is equivalent to $2q_2+q_1-q_{-1}-2q_{-2}=0$. Note that there are only two stopping times greater than 0, $\tau_1=1$ and $\tau_2=2$, the market model price given as the double supremum over all stopping times $\tau$ and all martingale measures $\mathbb{Q}$ of $\mathbb{E}^\mathbb{Q}[\Phi_\tau]$ also equals 2 and the two prices agree.
Suppose now that we add a {European option $g$} with a payoff $g=1\!\!1_{\{S_2=|1|\}}-1/2$ and initial price 0, which may be used as a static hedging instrument. With $g$ and $S$, the superhedging price of $\Phi$ drops to $3/2$ (e.g.\ keep 3/2 in cash and buy one option $g$). In presence of $g$, we need to impose a calibration constraint on martingale measures: $q_1+q_{-1}=1/2$. Thus, any calibrated martingale measure can be expressed by $(q_2,q_1, q_{-1},q_{-2})=(q, 3/4-2q, 2q-1/4,1/2-q)$ with $q\in(1/8,3/8)$, and the market model price equals 1. We therefore see that adding a statically traded option breaks the pricing--hedging duality.
\tikzstyle{level 1}=[level distance=2.5cm, sibling distance=2cm]
\tikzstyle{level 2}=[level distance=3cm, sibling distance=1cm]
\tikzstyle{bag} = [text width=5.5em, text centered]
\tikzstyle{bagg} = [text width=3em, text centered]
\tikzstyle{end} = [circle, minimum width=3pt,fill, inner sep=0pt]
\begin{figure}
\begin{center}
\begin{tikzpicture}[grow=right, sloped]
\node[bagg] {({0}, \emph{0})}
child {
node[bagg] {({0}, \textbf{1}) }
child {
node[bag]{({-2}, \textbf{0}, \emph{-0.5})}
edge from parent
node[below] {$q_{-2}$}
}
child {
node[bag]{({-1}, \textbf{2}, \emph{0.5})}
edge from parent
node[above] {$q_{-1}$}
}
child {
node[bag]{({1}, \textbf{2}, \emph{0.5})}
edge from parent
node[above] {$q_{1}$}
}
child {
node[bag] {({2}, \textbf{0}, \emph{-0.5})}
edge from parent
node[above] {$q_{2}$}
}
edge from parent
};
\end{tikzpicture}
\begin{tikzpicture}[grow=right, sloped]
\node[bag] {({0}, \emph{0})}
child {
node[bag] {({0}, \textbf{1}, \emph{-1/2}) }
child {
node[bag]{({-2}, \textbf{0}, \emph{-1/2})}
edge from parent
node[below] {0.5}
}
child {
node[bag] {({2}, \textbf{0}, \emph{-1/2})}
edge from parent
node[above] {0.5}
}
edge from parent
node[below] {0.5}
}
child {
node[bag] {({0}, \textbf{1}, \emph{1/2})}
child {
node[bag]{({-1}, \textbf{2}, \emph{1/2})}
edge from parent
node[below] {0.5}
}
child {
node[bag]{({1}, \textbf{2}, \emph{1/2})}
edge from parent
node[above] {0.5}
}
edge from parent
node[above] {0.5}
};
\end{tikzpicture}
\end{center}
\caption{Prices of the stock are written in regular font, prices of the American option in bold and prices of European option in italic. Model without dynamic trading in the option $g$ is on the left. Model with dynamic trading in $g$ which allows to recover the duality is on the right. \label{fig:Ex1}}
\end{figure}
Let us now show that the duality is recovered when we allow dynamic trading in $g$. We can model this through a process $Y=(Y_t:t=0,1,2)$ given by $Y_2=g$, $Y_1=1/2$ on $\{S_2=|1|\}$, $Y_1=-1/2$ on $\{S_2=|2|\}$ and $Y_0=0$. Note that there exists a (unique) measure $\mathbb{Q}$ such that both $S$ and $Y$ are martingales w.r.t their joint natural filtration $\widehat{\F}$ and in particular $\mathbb{Q}$ is calibrated, $\mathbb{E}^\mathbb{Q}[g]=0$. The filtration $\widehat{\F}$ is richer than the natural filtration of $S$ alone and allows for an additional stopping time $\tau=1\!\!1_{\{Y_1=-1/2\}}+21\!\!1_{\{Y_1=1/2\}}$ and the duality is recovered.
\end{Example}
\section{Pricing--hedging duality for American options}
\label{duality}
We present in this section general results which explain when and why the pricing--hedging duality for American options holds.
We work in a general discrete time setup which we now present.
Let $(\Omega, {\cal F})$ be a measurable space and $\mathbb{F}:=({\cal F}_k)_{k=0,1,...,N}$ be a filtration $\mathbb{F}:=({\cal F}_k)_{k=0,1,...,N}$, where $N\in \mathbb N$ is the time horizon.
We denote by $\mathcal H$ the class of $\mathbb{F}$-predictable processes.
We denote by $\frak P(\Omega)$ the set of all probability measures on $(\Omega, {\cal F})$. We consider a subset ${\cal P}\subset \frak P(\Omega)$.
We will say that a given property holds ${\cal P}$-quasi surely if it holds $\mathbb{P}$-almost surely for all $\mathbb{P}\in{\cal P}$.
We will refer to a set from ${\cal F}$ to be a ${\cal P}$-polar set if it is a null set w.r.t. all $\mathbb{P}\in{\cal P}$. We will write $\mathbb{Q}\lll {\cal P}$ if there exists $\mathbb{P}\in{\cal P}$ such that $\mathbb{Q}\ll \mathbb{P}$.
Given a random variable $\xi$ and a sub-$\sigma$-field ${\cal G} \subset {\cal F}$, we define the conditional expectation $\mathbb{E}^{\mathbb{P}}[ \xi | {\cal G}] := \mathbb{E}^{\mathbb{P}}[ \xi^+ | {\cal G}] - \mathbb{E}^{\mathbb{P}} [ \xi^- | {\cal G}]$ with the convention $\infty -\infty = -\infty$, where $\xi^+ := \xi \vee 0$ and $\xi^- := - (\xi \wedge 0)$.
We consider a market with no transaction costs and with financial assets, some which are dynamically traded and some which are only statically traded.
The former are modeled by an adapted $\mathbb{R}^d$-valued process $S$ with $d\in \mathbb N$. We think of the latter as European options which are traded at time $t=0$ but not necessarily at future times.
We let $g=(g^\lambda)_{\lambda\in \Lambda}$, where $\Lambda$ is a set of an arbitrary cardinality, be the vector of their payoffs which are assumed ${\cal F}$-measurable and $\mathbb{R}$-valued. Up to a constant shift of the payoffs, we may, without loss of generality, assume that all options $g^\lambda$ have zero initial prices. Denote by ${\cal H}$ the set of all $\mathbb{F}$-predictable $\mathbb{R}^d$-valued processes, and by
${\mathfrak h}=\{h\in \mathbb{R}^\Lambda: \exists \textrm{ finite subset $\beta\subset \Lambda$ s.t. $h^\lambda=0$ $\forall \lambda\notin\beta$}\}$. A self--financing strategy trades dynamically in $S$ and statically in finitely many of $g^\lambda$, $\lambda \in \Lambda$ and hence corresponds to a choice of $H\in {\cal H}$ and $h\in {\mathfrak h}$. Its associated final payoff is given by
\begin{eqnarray}
\label{self}
(H \circ S)_N + hg=\sum_{j=1}^d\sum_{k=1}^N H^j_k \Delta S^j_k + \sum_{\lambda\in\Lambda} h^{\lambda} g^{\lambda},
\end{eqnarray}
where $\Delta S^j_k = S^j_{k}-S^j_{k-1}$.
Having defined the trading strategy, we can consider the superhedging price of an option which pays off $\xi$ at time $N$:
\begin{align}\label{eq:def_pi_E}
\pi^E_g(\xi):=\inf\{x: &\exists \ (H, h)\in {\cal H}\times{\mathfrak h} \ \textrm{ s.t. } \ x+(H\circ S)_N+hg\geq \xi \ \mathcal P\textrm{-q.s.} \}.
\end{align}
The inequality is required to hold ${\cal P}$-q.s., i.e.\ it holds $\mathbb{P}$-a.s.\ for any $\mathbb{P}\in {\cal P}$. In particular, if ${\cal P} = \mathfrak{P}(\Omega)$ is the set of all probability measures on ${\cal F}$ and $\{\omega\} \in {\cal F}$ for all $\omega \in \Omega$, then the superreplication in \eqref{eq:def_pi_E} is pathwise on $\Omega$.
To formulate a duality relationship, we need the dual elements given by rational pricing rules, or martingale measures:
\begin{eqnarray}
\mathcal M&=&\{\mathbb Q\in \mathfrak{P}(\Omega): \mathbb{Q} \lll {\cal P} \textrm{ and } \mathbb{E}^\mathbb{Q}[\Delta S_k| {\cal F}_{k-1}]=0, \, \forall k=1,...,N \}\nonumber\\
\label{eq:mcg}
\mathcal M_g&=&\{\mathbb Q\in \mathcal M: \mathbb{E}^\mathbb{Q}[g^\lambda]=0, ~ \forall \lambda\in \Lambda \}.
\end{eqnarray}
\begin{Definition}
Let $\Upsilon$ be a given class of functions defined on $\Omega$,
we say that the \emph{(European) pricing--hedging duality holds for the class $\Upsilon$}
if ${\cal M}_g\neq \emptyset$ and
\begin{eqnarray}\label{eq:PHduality}
\pi^E_g(\xi)=\sup_{\mathbb{Q}\in {\cal M}_g}\mathbb{E}^\mathbb{Q}[\xi],\quad \xi\in\Upsilon,
\end{eqnarray}
\end{Definition}
\begin{Remark}
\label{rem:weak}
Note that the inequality $"\geq"$ in \eqref{eq:PHduality}, called weak pricing--hedging duality, holds automatically from the definition of ${\cal M}_g$ in \eqref{eq:mcg}.
\end{Remark}
A number of papers, including \cite{BN13,BFHMO16}, proved that the above pricing--hedging duality \eqref{eq:PHduality} holds under various further specifications and restrictions on $\Omega$, $\mathbb{F}$, ${\cal P}$ and $\Upsilon$, including in particular an appropriate no--arbitrage condition. We take the above duality for granted here and our aim is to study an analogous duality for American options. We work first in the general setup described above without specifying $\mathbb{F}$ or $\Upsilon$ since our results will apply to any such further specification. Further, many abstract results in this section also extend to other setups, e.g.\ to trading in continuous time.
\subsection{Superhedging of American options}
\label{superhedging}
An American option may be exercised at any time $k \in \mathbb{T} := \{1, \cdots, N\}$ (without loss of generality we exclude exercise at time $0$). It is described by its payoff function $\Phi = (\Phi_k)_{1 \le k \le N}$, where $\Phi_k: \Omega \to \mathbb{R}$ belongs to $\Upsilon$ and is the payoff, delivered at time $N$, if the option is exercised at time $k$. Usually $\Phi_k$ is taken to be ${\cal F}_k$-measurable but here we only assume $\Phi_k$ to be ${\cal F}$-measurable for greater generality which includes, e.g.\ the case of a portfolio containing a mix of American and European options.
We note that when hedging our exposure to an American option, we should be allowed to adjust our strategy in response to an early exercise. In consequence, the superhedging cost of the American option $\Phi$ using semi--static strategies is given by
\begin{align*}
\pi^A_g(\Phi)=\inf \big\{x: &\exists (H^1,...,H^N)\in \mathcal H^N \ \textrm{s.t.} \ H^j_i=H^k_i \ \forall 1\leq i\leq j\leq k\leq N\ \textrm{and} \ h\in \mathbb {\mathfrak h}\\
&\textrm{satisfying} \ x+(H^k\circ S)_N+hg\geq \Phi_k \ \forall k=1,...,N \ \mathcal P\textrm{-q.s.} \big\}
\end{align*}
Classically, pricing of an American option is recast as an optimal stopping problem and a natural extension of \eqref{eq:PHduality} would be
\begin{eqnarray}\label{eq:PHAduality}
\pi^A_g(\Phi)\stackrel{?}{=}\sup_{\mathbb{Q}\in{\cal M}_g}\sup_{\tau\in {\cal T}(\mathbb{F})}\mathbb{E}^\mathbb{Q}[\Phi_\tau],
\end{eqnarray}
where ${\cal T}(\mathbb{F})$ denotes the set of $\mathbb{F}$-stopping times. However, as illustrated with the simple example in the Introduction, this duality may fail. The ``numerical" reason is that the RHS in \eqref{eq:PHduality} may be too small since the set ${\cal M}_g$ is too small. Our aim here is to understand fundamental reasons why the duality fails and hence discuss how and why the right hand side should be modified to obtain equality in \eqref{eq:PHAduality}.
\subsection{American option is a European option on an enlarged space}
\label{ameu}
The first key idea of this paper offers a generic enlargement of the underlying probability space which turns all American options into European options.
Depending on the particular setup, it may take more or less effort to establish \eqref{eq:PHduality} for the enlarged space but this shifts the difficulty back to the better understood and well studied case of European options.
Our reformulation technique --- from an American to European option --- can be easily extended to other contexts, such as the continuous time case.
The enlargement of space is based on construction of random times, previously used e.g. in \cite{JeanblancSong1, JeanblancSong2} to study the existence of random times with a given survival probability and in \cite{ET} to study a general optimal control/stopping problem, and in \cite{GTT2}, \cite{KTT} to study the optimal Skorokhod embedding problem.
Let $\mathbb T:=\{1,..., N\}$ and introduce the probability space $\overline{\Om}:=\Omega \times \mathbb T$ with the canonical time $T: \overline{\Om} \to \mathbb{T}$ given by $T(\overline{\om}):=\theta$, where $\overline{\om}:=(\omega, \theta)$, the filtration $\overline{\F}:=(\overline{{\cal F}}_k)_{k=0,1,...,N}$ with $\overline{{\cal F}}_k={\cal F}_k\otimes \vartheta_k$ and $\vartheta_k=\sigma(T\land (k+1))$, and the $\sigma$-field $\overline{{\cal F}}={\cal F}\otimes \vartheta_N$.
By definition, $T$ is an $\overline{\F}$-stopping time. We denote bt $\overline{\Hc}$ the class of $\overline{\F}$-predictable processes and extend naturally the definition of $S$ and $g^\lambda$ from $\Omega$ to $\overline{\Om}$ as $S(\overline{\om})=S(\omega)$ and $g^\lambda(\overline{\om})=g^\lambda(\omega)$ for $\overline{\om}=(\omega, \theta)\in \overline{\Om}$.
We let $\overline\fcl$ be the class of random variables $\overline{\xi}:\overline{\Om}\to \mathbb{R}$ such that $\overline{\xi}(\cdot,k)\in \Upsilon$ for all $k\in \mathbb{T}$ and we let $\overline{\pi}^E_g(\bar{\xi})$ denote the superreplication cost of $\overline{\xi}$.
We may, and will, identify $\overline\fcl$ with $\Upsilon^N$ via $\overline{\xi}(\overline{\om})=\Phi_\theta(\omega)$.
Finally, we introduce
\begin{eqnarray} \label{eq:mcbg}
\overline{\Pc}&=&\{\overline{\P} \in \frak P(\overline{\Om}): \overline{\P}_{\vert\Omega}\in {\cal P}\},\nonumber\\
\overline{\Mc}&=&\{\overline{\Q} \in \frak P(\overline{\Om}) : \overline{\Q}\lll \overline{\Pc} \textrm{ and } \ \mathbb{E}^{\overline{\Q}}[\Delta S_k|\overline{{\cal F}}_{k-1}]=0 \ \forall k\in\mathbb{T}\}, \nonumber\\
\overline{\Mc}_g&=&\{\overline{\Q} \in\overline{\Mc} : \mathbb{E}^{\overline{\Q}}[g^\lambda]=0 \ \forall \lambda\in \Lambda\}
\end{eqnarray}
\begin{Theorem}
\label{thm:American_European}
For any $\Phi\in \Upsilon^N = \overline\fcl$ we have
\begin{eqnarray}\label{eq:Amer_as_Euro}
\pi^A_g(\Phi)=\overline{\pi}^E_g(\Phi)
:=
\inf\{x: \exists \ (\overline{H}, h)\in \overline{\Hc} \times{\mathfrak h} \ \textrm{ s.t. } \ x+(\overline{H}\circ S)_N+hg\geq \xi \ \overline{\Pc}\textrm{-q.s.} \}.
\end{eqnarray}
In particular, if the European pricing--hedging duality on $\overline{\Om}$ holds for $\Phi$ then
\begin{eqnarray}\label{eq:PHAduality_Euro}
\pi^A_g(\Phi)=\overline{\pi}^E_g(\Phi)=\sup_{\overline{\Q}\in \overline{\Mc}_g}\mathbb{E}^{\overline{\Q}}[\Phi].
\end{eqnarray}
\end{Theorem}
\proof
First note that
$$ \overline{\Hc}=\{\overline{H}=(\overline{H}(\cdot,1),\ldots,\overline{H}(\cdot,N))\in {\cal H}^N: \overline{H}_i(\cdot,j)=\overline{H}_i(\cdot,k) \ \forall 1\leq i\leq j\leq k\leq N
\}
$$
and hence the dynamic strategies used for superhedging in $\pi^A_g$ and in $\overline{\pi}^E_g$ are the same. The equality now follows observing that a set $\Gamma\in \overline{{\cal F}}_N$ is a $\overline{\Pc}$-polar if and only if its $k$-sections $\Gamma_k=\{\omega: (\omega, k)\in \Gamma\}$ are ${\cal P}$-polar for all $k\in \mathbb{T}$. Indeed, for one implication assume that $\overline{\P}(\Gamma)=0$ for each $\overline{\P}\in \overline{\Pc}$. For arbitrary $\mathbb{P}\in {\cal P}$ and $k\in\mathbb{T}$ we can define $\overline{\P}=\mathbb{P}\otimes \delta_k$ which belongs to $\overline{\Pc}$ and hence $\mathbb{P}(\Gamma_k)=0$.\\
To show the reverse implication, assume that $\mathbb{P}(\Gamma_k)=0$ for each $\mathbb{P}\in {\cal P}$ and $k\in\mathbb{T}$. Observe that, for any $\overline{\P}\in\overline{\Pc}$
\begin{align*}
\overline{\P}(\Gamma)=\sum_{k\in \mathbb{T}}\overline{\P}(\Gamma_k\times \{k\})\leq\sum_{k\in \mathbb{T}}\overline{\P}_{\vert\Omega}(\Gamma_k)=0
\end{align*}
since $\overline{\P}_{\vert\Omega}\in {\cal P}$. This completes the proof.
\qed
\begin{Remark}
\label{rem:g}
If the pricing--hedging duality holds w.r.t. filtration $\mathbb{F}$,
then it holds as well for any filtration $\mathbb{H}\supset\mathbb{F}$ such that $\mathbb{H}$
and $\mathbb{F}$ only differ up to ${\cal M}_g$-polar sets.
This change does not affect set ${\cal M}_g$ and may only decrease the superhedging cost as one has more trading strategies available. The duality is not affected by Remark \ref{rem:weak}.
\end{Remark}
\begin{Remark}
\label{rem:weak_relaxed_formulation}
We note that the set $\overline{\Mc}_g$ in \eqref{eq:PHAduality_Euro}, or its projection on $\Omega$, is potentially much larger than ${\cal M}_g$. Indeed, instead of stopping times relative to $\mathbb{F}$, it allows us to consider any \emph{random} time which can be made into a stopping time under some calibrated martingale measure. We can rephrase this as saying that $\overline{\Mc}_g$ is equivalent to a \emph{weak formulation} of the initial problem at the r.h.s. of \eqref{eq:PHAduality}. To make this precise, let us call
a weak stopping term $\alpha$ a collection
$$
\alpha ~=~
\big(
\Omega^{\alpha}, ~{\cal F}^{\alpha}, ~\mathbb{Q}^{\alpha}, \mathbb{F}^{\alpha} = ({\cal F}^{\alpha}_k)_{0 \le k \le N},
~(S^{\alpha}_k)_{0 \le k \le N},
~(g^{\lambda,\alpha})_{\lambda\in \Lambda},~(\Phi_k^\alpha)_{k\in \mathbb{T}},
~\tau^{\alpha}
\big)
$$
with $ \big( \Omega^{\alpha}, ~{\cal F}^{\alpha}, ~\mathbb{Q}^{\alpha}, \mathbb{F}^{\alpha} \big)$ a filtered probability space, $\tau^{\alpha}$ a $\mathbb{T}$-valued $\mathbb{F}^\alpha$-stopping time, an $\mathbb{R}^d$-valued $(\mathbb{Q}^{\alpha}, \mathbb{F}^{\alpha})$-martingale $S^{\alpha}$ and a collection of random variables $g^{\lambda,\alpha},\Phi^\alpha_k$, and such that there is a measurable surjective mapping $\mathtt{i}_\alpha:\Omega^\alpha\to \Omega$ with $\mathbb{Q}=\mathbb{Q}^\alpha\circ \mathtt{i}_\alpha^{-1}\in {\cal M}$ and $\mathtt{i}_\alpha^{-1}({\cal F}_k)\subset{\cal F}_k^\alpha$, $\mathtt{i}_\alpha^{-1}({\cal F})\subset{\cal F}^\alpha$ and finally $\mathcal{L}_{\mathbb{Q}^\alpha}(S^\alpha, g^\alpha,\Phi^\alpha)=\mathcal{L}_\mathbb{Q}(S, g,\Phi)$.
Denote by ${\cal A}_g$ the collection of all weak stopping terms $\alpha$ such that $\mathbb{E}^{\mathbb{Q}^{\alpha}}\big[ g^{\lambda,\alpha} \big] = 0$ for each $\lambda \in \Lambda$. It follows that any $\alpha\in {\cal A}_g$ induces a probability measure $\overline{\Q}\in\overline{\Mc}_g$ and $\mathbb{E}^{\mathbb{Q}^{\alpha}}[\Phi^\alpha_{\tau^{\alpha}}]=\mathbb{E}^{\overline{\Q}}[\Phi]$. Reciprocally, any $\overline{\Q}\in\overline{\Mc}_g$, together with the space $(\overline{\Om}, \overline{{\cal F}}, \overline{\F})$ and $(S, g, \Phi)$, provides a weak stopping term in ${\cal A}_g$. In consequence,
$$
\sup_{\alpha \in {\cal A}_g} \!\mathbb{E}^{\mathbb{Q}^{\alpha}}\!\big[ \Phi_{\tau^{\alpha}} \big]
=\!\!
\sup_{\overline{\Q} \in \overline{\Mc}_g}\! \mathbb{E}^{\overline{\Q}}[ \Phi \big].
$$
In summary, and similarly to number of other contexts, see the introduction in \cite{PhamZhang}, the weak formulation (and not the strong one) offers the right framework to compute the value of the problem. In fact, the set $\overline{\Mc}_g$ is large enough to make the problem static, or European, again. However, while it offers a solution and a corrected version of \eqref{eq:PHAduality}, it does not offer a fundamental insight into why \eqref{eq:PHAduality} may fail and what is the \emph{minimal} way of enlarging the objects on the RHS thereof to preserve the equality. These questions are addressed in the subsequent section.
\end{Remark}
\begin{Remark}
\label{com:hob}
\cite{Neuberger} and \cite{HobsonNeuberger} studied the same superhedging problem
in a Markovian setting, where the underlying process $S$ takes value in a discrete lattice ${\cal X}$.
By considering the weak formulation (which is equivalent to our formulation, as shown in Remark \ref{rem:weak_relaxed_formulation} above), they obtain similar duality results.
Moreover, they only consider $\Phi_k=\phi(S_k)$ where $\phi: \mathbb{R}^d\to\mathbb{R}$. Then the authors show that in the optimization problem $\sup_{\overline{\Q}\in \overline{\Mc}_g}\mathbb{E}^{\overline{\Q}}[\Phi]$ given in \eqref{eq:Amer_as_Euro} one may consider only Markovian martingale measures. The primal and the dual problem then turn to be linear programming problems under linear constraints, which can be solved numerically. Their arguments have also been extended to a more general context, where $S$ takes value in $\mathbb{R}_+$.
Comparing to \cite{Neuberger, HobsonNeuberger}, our idea to weak formulation is very similar to theirs. However our setting is much more general and, when considering specific setups in Sections \ref{sec:main} and \ref{sec:mot}, we rely on entirely different arguments to prove the duality.
\end{Remark}
\subsection{The loss and recovery of the dynamic programming principle and the natural duality for American options}
\label{sec:dpp}
The classical pricing of American options, on which the duality in \eqref{eq:PHAduality} was modelled, relies on optimal stopping techniques which subsume a certain dynamic consistency, or a dynamic programming principle, as explained below. Our second key observation in this paper is that if the pricing--hedging duality \eqref{eq:PHAduality} for American options fails it is because the introduction of static trading of European options $g$ at time $t=0$ destroys the dynamic programming approach, the Bellman optimality principle. Indeed, $\pi^E_g(\xi)$ will typically be lower than the superhedging price at time $t=0$ of the capital needed at time $t=1$ to superhedge from thereon. To reinstate such dynamic consistency, we need to enlarge the model and consider dynamic trading in options in $g$. This will generate a richer filtration than $\mathbb{F}$ and one which will carry enough stopping times to obtain the correct natural duality in the spirit of \eqref{eq:PHAduality}. In particular, if $g=0$ (or equivalently $\Lambda = \emptyset$), then \eqref{eq:PHAduality} should hold. We now first prove this statement and then present the necessary extension when $g$ is non--trivial.
Let $\Upsilon$ be a class of ${\cal F}$-measurable r.v.,
we denote ${\cal E}(\xi) := \sup_{\mathbb{Q}\in {\cal M}}\mathbb{E}^{\mathbb{Q}}[\xi]$,
and suppose that there is a family of operators ${\cal E}_k: \Upsilon\to \Upsilon$ for $k\in\{0,...,N-1\}$
such that ${\cal E}_k(\xi)$ is ${\cal F}_k$-measurable for all $\xi \in \Upsilon$.
We say that the family $({\cal E}_k)$ provides a dynamic programming representation of ${\cal E}$ if
\begin{equation}
\label{dpp_om}
{\cal E}(\xi)={\cal E}_0\circ{\cal E}_1\circ...\circ{\cal E}_{N-1}\left(\xi\right),\ \ \forall \xi\in \Upsilon.
\end{equation}
The family $({\cal E}_k)$ extends to the family $(\overline{\Ec}_k)$ for $k\in\{0,...,N-1\}$ defined for any $\Psi\in\overline\fcl = \Upsilon^N$ by
\begin{align}
\overline{\Ec}_0(\Psi)(\overline{\om})&:={\cal E}_0(\Psi(\cdot, 1))(\omega), \quad \quad \quad\quad \quad \quad\quad \quad \quad ~\mbox{for all}~ \overline{\om} = (\omega, \theta),
\label{eq:def_dpp0}\\
\overline{\Ec}_k(\Psi)(\overline{\om})&:=
\begin{cases}
{\cal E}_k(\Psi(\cdot, \theta))(\omega)&\textrm{if }\theta<k\\
{\cal E}_k(\Psi(\cdot, k))(\omega)\lor {\cal E}_k(\Psi(\cdot, k+1))(\omega)&\textrm{if } \theta\geq k
\label{eq:def_dpp}
\end{cases}
, \quad \textrm{for $1\leq k<N$}.
\end{align}
Assume that $f\lor f'\in \Upsilon$ whenever $f,f'\in \Upsilon$, then $\overline{\Ec}$ maps functionals from $\overline\fcl$ to $\overline\fcl$.
Let us also introduce the ${\cal M}$-Snell envelope process of an American option $\Psi\in \overline\fcl$ by
\begin{eqnarray}\label{eq:opbar}
\overline{\Ec}^k(\Psi):=\overline{\Ec}_k\circ ... \circ \overline{\Ec}_{N-1}(\Psi).
\end{eqnarray}
We say that the family $(\overline{\Ec}_k)$ provides a dynamic programming representation of $\overline{\Ec}(\Psi) := \sup_{\mathbb{Q}\in \overline{\Mc}}\mathbb{E}^{\mathbb{Q}}[\Psi]$ if
\begin{equation}
\label{eq:dpp_omb}
\overline{\Ec}(\Psi) =\overline{\Ec}^0(\Psi),\ \ \forall \Psi\in \overline\fcl.
\end{equation}
Typically we will consider ${\cal E}_k$ to be a supremum over conditional expectations w.r.t. ${\cal F}_k$, see Examples \ref{ex:dom} and \ref{ex:non} below. In these setups ${\cal E}_k$ automatically satisfies
\begin{align}
\label{inequality2}
\sup_{\overline{\Q}\in \overline{\Mc}}\mathbb{E}^{\overline{\Q}}[\Psi]&\leq\overline{\Ec}^0(\Psi), \ \quad \Psi\in \overline\fcl.
\end{align}
\begin{Theorem}
\label{dpp_gen_th}
Assume that $\Lambda=\emptyset$, ${\cal E}_k$ satisfies \eqref{dpp_om}, and that \eqref{inequality2} holds true, and that $f\lor f'\in \Upsilon$ for all $f,f'\in \Upsilon$.
Then, for all $\Phi\in\Upsilon^N = \overline\fcl$,
\begin{equation}\label{eq:dpp_implied}
\sup_{\overline{\Q}\in \overline{\Mc}}\mathbb{E}^{\overline{\Q}}[\Phi]=\sup_{\mathbb{Q}\in {\cal M}}\sup_{\tau\in {\cal T}(\mathbb{F})}\mathbb{E}^{\mathbb{Q}}[\Phi_\tau].
\end{equation}
If, further, the European pricing--hedging duality holds on $\overline{\Om}$ for the class $\overline\fcl$, then
\begin{equation*}
\pi^A(\Phi)=\sup_{\mathbb{Q}\in {\cal M}}\sup_{\tau\in {\cal T}(\mathbb{F})}\mathbb{E}^{\mathbb{Q}}[\Phi_\tau].
\end{equation*}
\end{Theorem}
The second assertion follows instantly from the first one and Theorem \ref{thm:American_European} while the first one follows from Proposition \ref{prop:dpp_Euro_American} below which asserts that \eqref{dpp_om} and \eqref{inequality2} imply an analogue consistency on $\overline{\Om}$ \eqref{eq:dpp_omb}.
This also allows us to identify the optimal stopping time on the RHS of \eqref{eq:dpp_implied}.
We have the following representation
\begin{Proposition}\label{prop:dpp_Euro_American}
Assume that $\Lambda=\emptyset$ and $f\lor f'\in \Upsilon$ for all $f,f'\in \Upsilon$.
Then the dynamic programming representation \eqref{eq:dpp_omb} holds if and only if \eqref{dpp_om} and \eqref{inequality2} hold true.
Moreover, under condition \eqref{eq:dpp_omb}, the $\mathbb{F}$-stopping time
\begin{align}
\label{tau}
\tau^*(\omega)&:= \inf \left\{k\geq 1: {\cal E}_k\left(\Phi(\cdot, k)\right)(\omega)= \overline{\Ec}^{k}(\Phi)(\omega, k)\right\}
\end{align}
provides the optimal exercise policy for $\Phi \in \overline\fcl$:
\begin{equation}
\sup_{\overline{\Q}\in \overline{\Mc}}\mathbb{E}^{\overline{\Q}}[\Phi]=\sup_{\mathbb{Q}\in {\cal M}}\sup_{\tau\in {\cal T}(\mathbb{F})}\mathbb{E}^{\mathbb{Q}}[\Phi_\tau]
=\sup_{\mathbb{Q}\in{\cal M}}\mathbb{E}^{\mathbb{Q}}\left[\Phi_{\tau^*}\right]=\overline{\Ec}^0(\Phi).
\end{equation}
\end{Proposition}
\begin{Remark} \label{rem:infinity_risky_assets}
The proof of Proposition \ref{prop:dpp_Euro_American} will be provided in Section \ref{proofs1}.
The results in Theorem \ref{dpp_gen_th} and Proposition \ref{prop:dpp_Euro_American} are stated on $(\Omega, {\cal F})$, where there are only finitely many dynamic trading risky assets.
However, its proof does not rely on the fact that the number of risky assets is finite,
and the same results holds still true if there are infinitely number of dynamic trading risky assets.
\end{Remark}
Next we give two examples of operators $({\cal E}_k)_{k\leq N}$ satisfying \eqref{dpp_om} and \eqref{inequality2}, therefore, by Proposition \ref{prop:dpp_Euro_American}, also \eqref{eq:dpp_omb}.
\begin{Example}
\label{ex:dom}
Let ${\cal P}=\{\mathbb{P}^*\}$. Then, taking $\Upsilon$ to be the set of all ${\cal F}$-measurable random variables and
\begin{eqnarray}
{\cal E}_k(\xi)={\mathop{\rm ess \, sup}}_{\mathbb{Q}\in{\cal M}_g}\mathbb{E}^\mathbb{Q}[\xi|{\cal F}_k]
\end{eqnarray}
where $\mathop{\rm ess \, sup}$ is taken w.r.t $\mathbb{P}^*$, leads to a family of operators satisfying \eqref{dpp_om}, \eqref{inequality2}, and therefore also \eqref{eq:dpp_omb}.
See the literature on dynamic coherent risk measures for further discussion (e.g. \cite{AP} for an overview).
If in particular we assume that $\Lambda=\emptyset$ then Theorem \ref{dpp_gen_th} recovers the classical superhedging theorem for American options (see e.g. \cite{Myneni}).
\end{Example}
\begin{Example}
\label{ex:non}
Let $(\Omega, d)$ be a Polish space, $\mathcal F$ its Borel $\sigma$-field and $\mathcal P$ a given set of probability measures on $(\Omega, {\cal F})$.
We are given a filtration $\mathbb{F}:=({\cal F}_k)_{k\leq N}$ such that
${\cal F}_0=\{\emptyset, \Omega\}$ and each $\sigma$-field ${\cal F}_k$ is countably generated.
For $k\in \mathbb{T}$ and $\omega\in \Omega$ denote by ${\cal M}^k(\omega)$ the set of measures given by
$${\cal M}^k(\omega):=\{\mathbb{Q}\lll{\cal P} : \ \mathbb{Q}([\omega]_{{\cal F}_k})=1\ \textrm{and}\ \mathbb{E}^\mathbb{Q}[\Delta S_n|{\cal F}_{n-1}]=0 \ \forall n\in\{k+1,..., N\} \}$$
where $[\omega]_{{\cal F}_k}$ denotes the atom of ${\cal F}_k$ which contains $\omega$, i.e.,
\begin{eqnarray} \label{eq:om_k}
[\omega]_{{\cal F}_k}=\bigcap_{F\in {\cal F}_k: \omega\in F}F\ .
\end{eqnarray}
Note that $[\omega]_{{\cal F}_k}\in {\cal F}_k$ since the latter is countably generated. In this setup we define
\begin{equation*}
{\cal E}_k(\xi)(\omega)=\sup_{\mathbb{Q}\in{\cal M}^{k}(\omega)}\mathbb{E}^{\mathbb{Q}}\left[\xi\right].
\end{equation*}
If we furthermore assume that ${\cal E}_k(\xi)\in \Upsilon$ for any $\xi\in \Upsilon$ then the family $({\cal E}_k)_{k\leq N}$ satisfies \eqref{inequality2} which will be proved in Proposition \ref{prop:non}.
We shall prove that under suitable assumptions on $(\Omega, \mathbb{F},{\cal P})$ and $\Upsilon$ also \eqref{dpp_om} holds for this family.
This holds in particular in the setup of \cite{BN13} as shown therein, see (4.12) in \cite{BN13}.
\end{Example}
Let us consider the case with statically traded options: $\Lambda\neq\emptyset$. We saw in Example \ref{ex:intro} that this can break down the dynamic consistency as the universe of traded assets differs at time $t=0$ and times $t\geq 1$. To remedy this, one has to embed the market into a fictitious larger one where both $S$ and all the options $g^\lambda$, $\lambda\in\Lambda$, are traded dynamically.
Let us consider a larger probability space $(\widehat{\Om},\widehat{\Fc})$ which satisfies the following properties.
First, there exists a surjective mapping $\mathtt{i}: \widehat{\Om}\to \Omega$, and it defines a natural extension of $S$, $g$ and $\Phi$ by $S(\widehat{\om})=S(\mathtt{i}(\widehat{\om}))$, $g^\lambda(\widehat{\om})=g^\lambda(\mathtt{i}(\widehat{\om}))$ for all $\lambda\in\Lambda$ and $\Phi(\widehat{\om})=\Phi(\mathtt{i}(\widehat{\om}))$.
Second there exists a family of processes $Y=(Y^\lambda)_{\lambda\in \Lambda}$ on $\widehat{\Om}$, such that $Y^\lambda_0=0$ and $Y^\lambda_N(\widehat{\om})=g^\lambda(\mathtt{i}(\widehat{\om}))$.
Let us denote by $\widehat{S}=(S, Y)$ which will now correspond to dynamically traded assets.
We suppose that there is a filtration $\widehat{\F}:=(\widehat{\Fc})_{k=0,1,...,N}$ such that $\mathtt{i}^{-1}({\cal F}_k)\subset\widehat{\Fc}_k$ and $\widehat{S}$ is $\widehat{\F}$-adapted, and let $\widehat{\mathcal H}$ be the set of $\widehat{\F}$-predictable processes.
Finally, we consider the following sets of probability measures
\begin{align*}
\widehat{\Pc}&:=\{\widehat{\mathbb P}\in \frak P(\widehat{\Om}): \widehat{\mathbb P}\circ\mathtt{i}^{-1}\in {\cal P}\ \}\\
\widehat{\Mc}&:=\{ \widehat{\Q}\lll \widehat{\Pc}: \widehat{S}=(S,Y) \ \ \textrm{is an $(\widehat{\Q}, \widehat{\F})$-martingale}\}.
\end{align*}
Observe that the martingale measures in $\widehat{\Mc}$ are by definition calibrated to market prices of options in $g$.
We furthermore assume that the mapping $\mathtt{I}: \widehat{\Mc}\to{\cal M}_g$ defined by $\mathtt{I}(\widehat{\Q})=\widehat{\Q}\circ\mathtt{i}^{-1}$ is surjective.
The collection $(\widehat{\Om}, \widehat{\Fc}, \widehat{\F}, \mathtt{i}, Y)$ satisfying above properties is called a dynamic extension of
$(\Omega,{\cal F},\mathbb{F}, {\cal P},S,g)$, or in short that $\widehat{\Om}$ is a dynamic extension of $\Omega$.
Note that for any dynamic extension $\widehat{\Om}$ it holds
$$\mathcal{L}_{\widehat{\Q}}(S, Y)=\mathcal{L}_{\mathtt{I}(\widehat{\Q})}(S, Y^{\mathtt{I}(\widehat{\Q})})$$
where $Y^{\lambda, \mathtt{I}(\widehat{\Q})}=(\mathbb{E}^\mathbb{Q}[g^\lambda|{\cal F}_k])_{k\leq N}$.
For any $\widehat{\Q}\in\widehat{\Mc}$ let $\mathtt{I}(\widehat{\Q})=\widehat{\Q}\circ\mathtt{i}^{-1}\in{\cal M}_g$. And conversely, from a given $\mathbb{Q}\in{\cal M}_g$ we may recover its ``parent" measure $\widehat{\Q}\in\widehat{\Mc}$.
We consider a class of functions $\widehat\fcl$ on $\widehat{\Om}$ and assume that $\Upsilon\subset\widehat\fcl$ in the sense that for $f\in\Upsilon$, $f(\widehat{\om}):=f(\mathtt{i}(\widehat{\om}))$ belongs to $\widehat\fcl$.
Then the correspondence between $\widehat{\Mc}$ and ${\cal M}_g$ yields to
$$
\sup_{\mathbb{Q}\in\widehat{\Mc}}\mathbb{E}^\mathbb{Q}\left[\xi\right]
= \sup_{\mathbb{Q}\in{\cal M}_g}\mathbb{E}^\mathbb{Q}\left[\xi\right]\quad \textrm{for any $\xi\in \Upsilon$}.
$$
As introduced at the beginning of Section \ref{ameu}, one can apply the enlargement techniques on space $\widehat{\Om}$ to obtain $\overline{\widehat{\Mc}}$, and one has a similar equality:
\begin{eqnarray}
\label{eq:corr}
\sup_{\mathbb{Q}\in\overline{\widehat{\Mc}}}\mathbb{E}^\mathbb{Q}\left[\Phi\right]
= \sup_{\mathbb{Q}\in\overline{\Mc}_g}\mathbb{E}^\mathbb{Q}\left[\Phi\right]\quad\textrm{for any $\Phi\in \overline\fcl$}.
\end{eqnarray}
Since in the dynamic extension $(\widehat{\Om}, \widehat{\Fc}, \widehat{\F}, \mathtt{i}, Y)$ of $(\Omega,{\cal F},\mathbb{F}, {\cal P},S,g)$ we allow to trade dynamically in $\widehat{S}=(S, Y)$ let us introduce the class of trading strategies $\widehat{H}$ which is a set of $\widehat{\F}$-predictable $\mathbb{R}^{\widehat{\No}}$-valued processes which have only finitely many non--zero coordinates where $\widehat{\No}=\{(i,s): i\in\{1,..,d\}\}\cup\{(\lambda,y):\lambda\in\Lambda\}$, i.e.,
\begin{align*}
\widehat{\mathcal H}=\{\widehat{H}=&{(\widehat{H}^{\widehat{\no}}_k: {\widehat{\no}\in\widehat{\No}})}_{k\leq N}: \ \textrm{$\widehat{\F}$-predictable $\mathbb{R}^{\widehat{\No}}$-valued process s.t.}\\
&\textrm{ $\exists$ finite subset $\widehat{\No}_0 \subset \widehat{\No}$ s.t. $\widehat{H}^{\widehat{\no}}_k=0, ~\forall k,~\forall \widehat{\no}\notin\widehat{\No}_0$ } \}.
\end{align*}
In consequence, a self--financing strategy corresponds to a choice of $\widehat{H}\in \widehat{\mathcal H}$ and yields a final payoff of
\begin{eqnarray}
\label{self}
(\widehat{H} \circ \widehat{S})_N=\sum_{j=1}^d\sum_{k=1}^N \widehat{H}^{(j,s)}_k \Delta S^j_k +\sum_{\lambda\in\Lambda}\sum_{k=1}^N \widehat{H}^{(\lambda,y)}_k \Delta Y^\lambda_k.
\end{eqnarray}
Note that appropriate choice of trading strategies ensures that the sums are finite.
The supehedging costs of a European option $\widehat{\xi}$ and an American option $\widehat{\Psi}=(\widehat{\Psi}_k)_{k\leq N}$ on $\widehat{\Om}$ are given by
\begin{align*}
\widehat{\pi}^E(\widehat{\xi})=\inf\{x: &\exists \widehat{H}\in \widehat{\mathcal H} \ \textrm{s.t.} \ \textrm{satisfying} \ x+(\widehat{H}\circ \widehat{S})_N\geq \widehat{\xi}\ \widehat{\Pc}\textrm{-q.s.} \},\\
\widehat{\pi}^A(\widehat{\Psi})=\inf\{x: &\exists (\widehat{H}^1,...,\widehat{H}^N)\in \widehat{\mathcal H}^N \ \textrm{s.t.} \ \widehat{H}^j_i=\widehat{H}^k_i \ \forall \;1\leq i\leq j\leq k\leq N\ \\
&\textrm{satisfying} \ x+(\widehat{H}^k\circ \widehat{S})_N\geq \widehat{\Psi}_k \ \forall k=1,...,N \ \widehat{\Pc}\textrm{-q.s.} \}.
\end{align*}
\begin{Remark}
Clearly $\widehat{\F}$ is much richer than $\mathbb{F}$.
Beside the price process of the underlying, it captures all possible price processes of vanilla options. Therefore the inequality $\widehat{\pi}^A(\Phi)\leq \pi^A_g(\Phi)$ holds which follows by noting that buy--and--hold strategies are a special case of a dynamic trading strategy and ${\cal P}=\widehat{\Pc}\circ \mathtt{i}^{-1}$.
\end{Remark}
We can now apply Theorem \ref{dpp_gen_th} to the present setting:
\begin{Corollary}
\label{cor:dpp_options}
Let $(\widehat{\Om}, \widehat{\Fc}, \widehat{\F}, \mathtt{i}, Y)$ be the dynamic extension of $(\Omega,{\cal F},\mathbb{F}, {\cal P},S,g)$ with operators $\widehat{\Ec}_k:\widehat\fcl\to\widehat\fcl$ satisfying \eqref{dpp_om} and \eqref{inequality2}.
Assume that the European pricing--hedging duality holds for the class $\Upsilon^N$ on $\overline{\Om}$.
Then for all $\Phi\in\Upsilon^N$
\begin{equation}\label{eq:PHA_dpp}
\pi^A_g(\Phi)= \widehat{\pi}^A(\Phi)=\sup_{\widehat{\Q}\in \widehat{\Mc}}\sup_{\tau\in {\cal T}(\widehat{\F})}\mathbb{E}^{\widehat{\Q}}\left [\Phi_\tau\right].
\end{equation}
\end{Corollary}
\proof
Note that $\pi^A_g\geq \widehat{\pi}^A$ since a buy--and--hold strategy is a special case of a dynamic trading strategy and ${\cal P}=\widehat{\Pc}\circ \mathtt{i}^{-1}$.
Using \eqref{eq:Amer_as_Euro} twice we obtain
$$\overline{\pi}^E_g(\Phi)=\pi^A_g(\Phi)\geq \widehat{\pi}^A(\Phi)=
\overline{\pih}^E\left(\Phi\right)\geq
\sup_{\mathbb{Q}\in\overline{\widehat{\Mc}}}\mathbb{E}^\mathbb{Q}\left[\Phi\right]
= \sup_{\mathbb{Q}\in\overline{\Mc}_g}\mathbb{E}^\mathbb{Q}\left[\Phi\right],
$$
where the penultimate inequality always holds by Remark \ref{rem:weak} and last equality follows by \eqref{eq:corr}. The assumed pricing--hedging duality on $\overline{\Om}$ implies that we have equality throughout and we conclude by
applying Theorem \ref{dpp_gen_th} (with Remark \ref{rem:infinity_risky_assets}) on $\widehat{\Om}$ for the representation of $\sup_{\mathbb{Q}\in\overline{\widehat{\Mc}}}\mathbb{E}^\mathbb{Q}\left[\Phi\right]$.
\qed
\begin{Remark}
Note that if pricing--hedging duality on $\overline{\Om}$ holds then trading vanilla options dynamically or statically makes no difference on superhedging cost.
\end{Remark}
\begin{Example}
\label{ex:product}
We give an example of a dynamic extension of $\Omega$ in the case of finitely many statically traded options, i.e. we assume that $\Lambda=\{1,...,e\}$ for some $e\in\mathbb N$.
Consider the probability space $\widehat{\Om}=\Omega\times \mathbb{R}^{(N-1)\times e}$.
An element $\widehat{\om}$ in $\widehat{\Om}$ can be written as $\widehat{\om}=(\omega, y)$ where $y=(y^1, ..., y^e)\in \mathbb{R}^{(N-1)\times e}$ with $y^i=(y^i_1, ..., y^i_{N-1})$. Define a mapping $\mathtt{i}:\widehat{\Om}\to\Omega$ by $\mathtt{i}(\widehat{\om})=\omega$ which is clearly surjective.
We also introduce the process $Y$ as $Y_k(\widehat{\om})=y_k=(y^1_k, ..., y^e_k)$ for $k\in\{1,...,N-1\}$, $Y_0(\widehat{\om})=0$ and $Y_N(\widehat{\om})=g(\widehat{\om})=g(\omega)$.
Let the filtration $\widehat{\F}:=(\widehat{\Fc})_{k=0,1,...,N}$ be given by $\widehat{\Fc}_k={\cal F}_k\otimes \mathcal Y_k$, $\mathcal Y_k=\sigma(Y_n: n\leq k)$.
In this case it also holds that $\mathtt{I}:\widehat{\Mc}\to {\cal M}_g$ given by $\mathtt{I}(\widehat{\Q})=\widehat{\Q}\circ\mathtt{i}^{-1}=\widehat{\Q}_{\vert\Omega}$ is surjective.
In Section \ref{sec:main}, where the basic setup is taken from \cite{BN13} and hence \eqref{dpp_om} holds on $\Omega$ as recalled above in Example \ref{ex:non}, we show that \eqref{dpp_om} also holds on $\widehat{\Om}$.
\end{Example}
\begin{Remark}
Let us consider the two period ($N=2$) example of \cite{HobsonNeuberger2}, see Figure \ref{fig:hobson} below.
For simplicity, we introduce only one statically traded option $g$ with payoff $1\!\!1_{\{S_2=8\}}$ at time $t=2$ and price $2/5$ at time $t=0$. This already destroys the pricing--hedging duality for the American option $\Phi$.
In \cite{HobsonNeuberger2}, the duality is recovered by considering a (calibrated) mixture of martingale measures.
It is insightful to observe that their mixture model is nothing else but a martingale measure for an augmented setup with dynamic trading in $g$ which, following Corollary \ref{cor:dpp_options}, restores the dynamic programming and the pricing--hedging duality for American options.
To show this, let $Y$ denote the price process of the option $g$: $Y_0=2/5$, $Y_2=g$. Figure \ref{fig:hobson} illustrates a martingale measure $\mathbb{Q}$ along with the intermediate prices $Y_1$ such that the processes $S$ and $Y$ are martingales. With $\tau=1\!\!1_{\{S_1=1, Y_1=0\}}+21\!\!1_{\{S_1=1, Y_1=1/4\}\cup\{S_1=3\}}$ we find $\mathbb{E}^\mathbb{Q}[\Phi_\tau]=18/5$ which is the super--hedging price and the duality is recovered.
\tikzstyle{level 1}=[level distance=3.5cm, sibling distance=2.7cm]
\tikzstyle{level 2}=[level distance=3cm, sibling distance=1cm]
\tikzstyle{bag} = [text width=5em, text centered]
\tikzstyle{end} = [circle, minimum width=3pt,fill, inner sep=0pt]
\begin{figure}
\label{fig:hobson}
\begin{center}
\begin{tikzpicture}[grow=right, sloped]
\node[bag] {({2}, \emph{2/5})}
child {
node[bag] {({1}, \textbf{1}, \emph{0})}
child {
node[bag]{({0}, \textbf{0}, \emph{0})}
edge from parent
node[below] {0.5}
}
child {
node[bag]{({2}, \textbf{0}, \emph{0})}
edge from parent
node[above] {0.5}
}
child {
node[bag]{({4}, \textbf{8}, \emph{1})}
edge from parent
node[above] {0}
}
edge from parent
node[below] {0.4}
}
child {
node[bag] {({1}, \textbf{1}, \emph{1/4}) }
child {
node[bag]{({0}, \textbf{0}, \emph{0})}
edge from parent
node[below] {0.75}
}
child {
node[bag]{({2}, \textbf{0}, \emph{0})}
edge from parent
node[above] {0}
}
child {
node[bag]{({4}, \textbf{8}, \emph{1})}
edge from parent
node[above] {0.25}
}
edge from parent
node[above] {0.1}
}
child {
node[bag] {({3}, \textbf{0}, \emph{3/4})}
child {
node[bag]{({0}, \textbf{0}, \emph{0})}
edge from parent
node[below] {0.25}
}
child {
node[bag]{({2}, \textbf{0}, \emph{0})}
edge from parent
node[above] {0}
}
child {
node[bag]{({4}, \textbf{8}, \emph{1})}
edge from parent
node[above] {0.75}
}
edge from parent
node[above] {0.5}
};
\end{tikzpicture}
\end{center}
\caption{The model on $\widehat{\Om}$ which corresponds to mixture model in \cite{HobsonNeuberger2} attaining the super--hedging price. Prices of the stock are written in regular font, prices of the American option in bold and prices of European option in italic.}
\end{figure}
\end{Remark}
\subsection{Pseudo--stopping times}
It follows from Theorem \ref{thm:American_European} that in general we expect to see
$$
\pi^A_g(\Phi)=\sup_{\overline{\Q}\in \overline{\Mc}_g}\mathbb{E}^{\overline{\Q}}[\Phi]\geq \sup_{\mathbb{Q}\in{\cal M}_g}\sup_{\tau\in {\cal T}(\mathbb{F})}\mathbb{E}^\mathbb{Q}[\Phi_\tau],
$$
where the last inequality may be strict. We showed above that this is linked with the necessity to use random times beyond $\tau\in {\cal T}(\mathbb{F})$. To conclude our general results, we explore this property from another angle and identify the subset(s) of $\overline{\Mc}_g$ which leads to equality in place of inequality above. Introduce
\begin{eqnarray}
~~\underline{\Mc}_g ~:=~
\big\{\overline{\Q} \in \mathfrak{P}(\overline{\Om}) \!\!\!\!\!&:&\!\!\! \overline{\Q} \lll \overline{\Pc}, ~\mathbb{E}^{\overline{\Q}}[g^\lambda] \!=\! 0, \;\;\lambda \in\Lambda \;\;
~ S~\mbox{is an}~(\mathbb{F}, \overline{\Q}) \mbox{-martingale,}\nonumber\\
\mathbb{E}^{\overline{\Q}}[M_T] &=& \mathbb{E}^{\overline{\Q}}[M_0]
~~\mbox{for all bounded $(\mathbb{F}, \overline{\Q})$-martingales}~M
\big\},
\end{eqnarray}
the set of measures which make $S$ an $\mathbb{F}$-martingale and $T$ an $\mathbb{F}$-pseudo--stopping time. These are natural since the martingale part of the Snell envelope can be stopped at the pseudo--stopping time with null expectation.
\begin{Proposition}
\label{pseudo}
Assume that ${\cal M}_g\neq \emptyset$. Then
\begin{eqnarray} \label{eq:equiv_immersion}
\sup_{\overline{\Q} \in \underline{\Mc}_g} \mathbb{E}^{\overline{\Q}}[\Phi]
=\sup_{\mathbb{Q} \in {\cal M}_g} \sup_{\tau \in {\cal T}(\mathbb{F})} \mathbb{E}^{\mathbb{Q}} \big[ \Phi_{\tau}].
\end{eqnarray}
\end{Proposition}
\proof
Let $\overline{\Q} \in \underline{\Mc}_g$ such that $\mathbb{E}^{\overline{\Q}}[|g^\lambda|] < \infty$ and $\mathbb{E}^{\overline{\Q}}[|\Phi_k|] < \infty$ for all $\lambda\in\Lambda$ and $k=1, \cdots, N$.
We next consider the optimal stopping problem $\sup_{\tau \in {\cal T}(\mathbb{F})} \mathbb{E}^{\overline{\Q}} \big[ \Phi_{\tau} \big]$.
Define its Snell envelope $(Z_k)_{0 \le k \le N}$ by
$$
Z_k := \mathrm{esssup}_{\tau \in {\cal T}(\mathbb{F}), \tau \ge k} \mathbb{E}^{\overline{\Q}} \big[\Phi_{\tau} \big| {\cal F}_k \big],
$$
which is an $(\mathbb{F},\overline{\Q})$-supermartingale. Its Doob--Meyer decomposition is given by
$$
Z_k = Z_0 + M_k - A_k, ~~~\mbox{where}~A=(A_k)_{0 \le k \le N}~\mbox{is an $\mathbb{F}$-predictable increasing process},
$$
and $A_0 = M_0 = 0$.
It follows that
\begin{eqnarray}
\mathbb{E}^{\overline{\Q}} \big[ \Phi \big]=\mathbb{E}^{\overline{\Q}} \big[ \Phi_T \big]
~\le~
\mathbb{E}^{\overline{\Q}} [Z_T]
~\le~
Z_0 + \mathbb{E}^{\overline{\Q}}[M_T]
~=~
Z_0.
\end{eqnarray}
We hence obtain that $\sup_{\overline{\Q} \in \underline{\Mc}_g} \mathbb{E}^{\overline{\Q}}[\Phi] \le \sup_{\mathbb{Q} \in {\cal M}_g} \sup_{\tau \in {\cal T}(\mathbb{F})} \mathbb{E}^{\mathbb{Q}} \big[ \Phi_{\tau}]$.
Then \eqref{eq:equiv_immersion} holds since every stopping time $\tau \in {\cal T}(\mathbb{F})$ is a pseudo-stopping time and hence the inverse inequality is trivial.
\qed
\begin{Remark}
The above allows us to see that it is not enough to use randomised stopping times to recover equality in \eqref{eq:PHAduality}.
Such a time corresponds to an $\mathbb{F}$-adapted increasing process $V$ with $V_0=0$ and $V_N=1$. It may be seen as a distribution over all possible stopping times, in our setup a distribution $\eta$ on $\mathbb{T}$ s.t. $\eta(\{k\}):=\Delta V_k=V_k-V_{k-1}$ for each $k\in \mathbb{T}$.
For any pseudo--stopping time $\tau$, the dual optional projection of the process $1\!\!1_{[\![ \tau, N]\!]}$ is a randomised stopping time.
Conversely, for a given $V$, if we take a uniformly distributed random variable $\Theta$ independent from $V$, possibly enlarging probability space, then $\tau:=\inf\{t: V_t\geq \Theta\}$ is $\mathbb{F}$-pseudo--stopping time which generates $V$.
Let ${\cal R}$ be the set of such randomised stopping times. Then, from Proposition \ref{pseudo} and definition of dual optional projection,
\begin{equation*}
\sup_{\mathbb{Q} \in {\cal M}_g} \sup_{\tau \in {\cal T}(\mathbb{F})} \mathbb{E}^{\mathbb{Q}} \big[ \Phi_{\tau}]=\sup_{\mathbb{Q} \in {\cal M}_g} \sup_{V \in {\cal R}} \mathbb{E}^{\mathbb{Q}} \left[ \sum_k\Phi_{k}\Delta V_k\right].
\end{equation*}
\end{Remark}
\begin{Remark}
\cite{NY05} showed that under a progressive enlargement with pseudo--stopping time $\tau$ all martingales from the smaller filtration stopped at $\tau$ remain martingales in the larger filtration.
One can relate this to a more restrictive situation, when all martingales from the smaller filtration remain martingales in the bigger filtration, which is called the immersion property in enlargement of filtration. Clearly each random time satisfying immersion property is a pseudo--stopping time. Thus, keeping the equality \eqref{eq:equiv_immersion} true, the pseudo--stopping time property in the definition of $\underline{\Mc}_g$ above can be replace by a stronger condition characterizing the immersion property
\begin{eqnarray} \label{eq:cond_immersion}
\overline{\Q}[T > k | {\cal F}_n] = \overline{\Q}[ T >k | {\cal F}_k],
~~\mbox{for all}~
0 \le k \le n \le N.
\end{eqnarray}
See Section 3.1.2 of \cite{BJR} for the discrete time context of progressive enlargement of filtration and \cite{AL16} for connections between pseudo--stopping times, the immersion property and projections.
\end{Remark}
\section{A detailed study of the non--dominated setup of\\ \cite{BN13}}
\label{sec:main}
In this section we work in the non--dominated setup introduced in \cite{BN13} which is a special case of Example \ref{ex:non}.
We let $\Omega_0 =\{\omega_0\}$ be a singleton and $\Omega_1$ be a Polish space.
For each $k \in \{1, \cdots, N\}$, we define $\Omega_k := \{\omega_0\}\times\Omega_1^k$ as the $k$-fold Cartesian product.
For each $k$, we denote by ${\cal G}_k := {\cal B}(\Omega_k)$ and by
${\cal F}_k$ its universal completion.
In particular, we notice that ${\cal G}_0$ is trivial and we denote
$$
\Omega := \Omega_N, ~~ {\cal G}:={\cal G}_N~~\mbox{and}~~ {\cal F} := {\cal F}_N.
$$
We shall often see ${\cal F}_k$ and ${\cal G}_k$ as sub-$\sigma$-fields of ${\cal F}_N$,
and hence obtain two filtrations $\mathbb{F} = ({\cal F}_k)_{0 \le k \le N}$ and $\mathbb{G} = ({\cal G}_k)_{0 \le k \le N}$ on $\Omega$.
Recall that a subset of a Polish space $\Omega$ is analytic if it is the image of a Borel subset of another Polish space under a Borel measurable mapping.
We take $\Upsilon$ to be the class of upper semianalytic functions $f: \Omega \to \overline{\R} := [-\infty, \infty]$, i.e.\ such that $\{\omega \in \Omega ~: f(\omega) > c\}$ is analytic for all $c \in \mathbb{R}$.
The price process $S$ is a $\mathbb{G}$-adapted $\mathbb{R}^d$-valued process and the collection of options $g=(g^1, ..., g^e)$ is a $\mathbb{G}$-measurable $\mathbb{R}^e$-valued vector for $e\in \mathbb N$ (thus $\Lambda=\{1,...,e\}$).
Let $k \in \{0, \cdots, N-1\}$ and $\omega \in \Omega_k$,
we are given a non--empty convex set ${\cal P}_k(\omega) \subseteq \mathfrak{P}(\Omega_1)$ of probability measures, which represents the set of all possible models for the $(k+1)$-th period, given state $\omega$ at times $0, 1, \cdots, k$.
We assume that for each $k$,
\begin{eqnarray} \label{eq:AnalyticGraph}
\mbox{graph}({\cal P}_k)
:=
\{ (\omega, \mathbb{P}): \omega \in \Omega_k,
\mathbb{P} \in {\cal P}_k(\omega)
\}
~\subseteq~ \Omega_k \times {\cal P}(\Omega_1)
~~ \text{is analytic.}
\end{eqnarray}
Given a universally measurable kernel $\mathbb{P}_k: \Omega_k \to \mathfrak{P}(\Omega_1)$ for each $k \in \{ 0,1, \cdots, N-1 \}$,
we define a probability measure $\mathbb{P}$ on $\Omega$ by Fubini's theorem:
$$
\mathbb{P}(A)
:=
\int_{\Omega_1} \cdots \int_{\Omega_1} \mathbf{1}_A (\omega_1, \omega_2 \cdots, \omega_N)
\mathbb{P}_{N-1}(\omega_1, \cdots, \omega_{N-1}; d \omega_N) \cdots \mathbb{P}_0(d \omega_1).
$$
We can then introduce the set ${\cal P} \subseteq \mathfrak{P}(\Omega)$ of possible models for the multi--period market up to time $N$:
\begin{eqnarray} \label{eq:def_Pc}
{\cal P} :=
\big\{
\mathbb{P}_0 \otimes \mathbb{P}_1 \otimes \cdots \otimes \mathbb{P}_{N-1} ~: \mathbb{P}_k(\cdot) \in {\cal P}_k(\cdot), k=0, 1, \cdots, N-1
\big\}.
\end{eqnarray}
Notice that the condition \eqref{eq:AnalyticGraph} ensures that ${\cal P}_k$ has always a universally measurable selector: $\mathbb{P}_k : \Omega_k \rightarrow \mathfrak{P}(\Omega_1)$ such that $\mathbb{P}_k(\omega) \in {\cal P}_k(\omega)$ for all $\omega \in \Omega_k$.
Then the set ${\cal P}$ defined in \eqref{eq:def_Pc} is nonempty.
We also denote by ${\cal M}^{k,k+1}(\omega)$ the following set
$${\cal M}^{k,k+1}(\omega)=\{\mathbb{Q}\in \mathfrak P(\Omega_1): \mathbb{Q} \lll {\cal P}_k(\omega) \ \textrm{and} \ \ \mathbb{E}^{\delta_{\omega} \otimes_k \mathbb{Q}}[ \Delta S_{k+1}] = 0\},$$
where $\delta_{\omega} \otimes_k \mathbb{Q} := \delta_{(\omega_1, \cdots, \omega_k)} \otimes \mathbb{Q}$ is a Borel probability measure on $\Omega_{k+1} := \Omega_k \times \Omega_1$.
The following notion of no--arbitrage $\mathrm{NA}({\cal P})$ has been introduced in \cite{BN13}.
$\mathrm{NA}({\cal P})$ holds if for all $(H, h) \in {\cal H} \times \mathbb{R}^e$
$$
(H \circ S)_N + hg \geq 0
\quad {\cal P}\mbox{-q.s.}~
\Longrightarrow~
(H \circ S)_N + hg = 0
\quad {\cal P}\mbox{-q.s.}
$$
Analogously, we will say that $\mathrm{NA}(\overline{\Pc})$ holds if for all $(\overline{H}, h) \in \overline{\Hc} \times \mathbb{R}^e$
\begin{eqnarray}\label{nab}
(\overline{H} \circ S)_N + hg \geq 0\quad \overline{\Pc}\mbox{-q.s.}~
\Longrightarrow~
(\overline{H} \circ S)_N + hg = 0
\quad \overline{\Pc}\mbox{-q.s.}
\end{eqnarray}
Recall also ${\cal M}_g$ and $\overline{\Mc}_g$ have been defined in \eqref{eq:mcg} and \eqref{eq:mcbg}.
As established in \cite{BN13}, the condition $\mathrm{NA}({\cal P})$ is equivalent to the statement that ${\cal P}$ and ${\cal M}_g$ have the same polar sets.
The following lemma extends that result to $\overline{\Om}$.
\begin{Lemma}
\label{lem:na}
$\mathrm{NA}({\cal P})$ $\Longleftrightarrow$ $\mathrm{NA}(\overline{\Pc})$ $\Longleftrightarrow$ $\overline{\Pc}$ and $\overline{\Mc}_g$ have the same polar sets.
\end{Lemma}
\proof
The two conditions $\mathrm{NA}({\cal P})$ and $\mathrm{NA}(\overline{\Pc})$ are equivalent by the same arguments as in proving \eqref{eq:Amer_as_Euro}.
It is enough to show that $\overline{\Pc}$ and $\overline{\Mc}_g$ have the same polar sets if and only if ${\cal P}$ and ${\cal M}_g$ have the same polar sets.
That boils down to proving that a set $\Gamma\in \overline{\Om}$ is an $\overline{\Mc}_g$ polar set if and only if the $k$-section $\Gamma_k=\{\omega: (\omega, k)\in \Gamma\}$ is an ${\cal M}_g$ polar set for each $k\in \mathbb{T}$ which is analogous statement involving ${\cal P}$ and $\overline{\Pc}$ proved in the proof of Theorem \ref{thm:American_European}.
\qed
\subsection{Duality on the enlarged space $\overline{\Om}$}
Our first main result is the following duality under the no--arbitrage condition \eqref{nab}.
\begin{Theorem} \label{theo:main}
Let $\mathrm{NA}(\overline{\Pc})$ hold true.
Then the set $\overline{\Mc}_g$ is nonempty, and, for any upper semianalytic $\Phi: \overline{\Om} \to \mathbb{R}$, one has
\begin{eqnarray} \label{eq:main}
\overline{\pi}^E_g(\Phi) &=& \sup_{\overline{\Q} \in \overline{\Mc}_g} \mathbb{E}^{\overline{\Q}}[ \Phi \big],
\end{eqnarray}
and in particular the pricing-hedging duality \eqref{eq:PHAduality_Euro} holds.
Moreover, there exists $(\overline{H}, h) \in \overline{\Hc} \times \mathbb{R}^e$ such that
$$\overline{\pi}^E_g(\Phi) + (\overline{H} \circ S)_N + h g \ge \Phi, ~~~\overline{\Pc} \mbox{-q.s.}
$$
\end{Theorem}
The proof is delegated to Section \ref{proofs2} and uses the following lemma.
Let us work with operators ${\cal E}_k$ introduced in Example \ref{ex:non}. Observe that
$${\cal E}_k\circ ... \circ {\cal E}_{N-1}(\xi)(\omega)={\cal E}_{k,k+1}\circ ... \circ {\cal E}_{N-1,N}(\xi)(\omega),\quad \xi\in \Upsilon$$
where ${\cal E}_{k,k+1}(\xi)(\omega)=\sup_{{\cal M}^{k,k+1}(\omega)}\mathbb{E}^{\mathbb{Q}}\left[\xi\right]$. By Proposition \ref{prop:dpp_Euro_American}, (4.12) in \cite{BN13} and using that the maximum of upper semianalytic functions is still upper semianalytic we conclude that
\begin{Lemma} \label{lemma:super_hedging}
Consider the case $e=0$, i.e., $\Lambda=\emptyset$.
Let $\Psi\in\overline\fcl$. Then $\overline{\Ec}_k(\Psi)$ in \eqref{eq:def_dpp} is also upper semianalytic and
$$
\sup_{\overline{\Q} \in \overline{\Mc}} \mathbb{E}^{\overline{\Q}}[ \Psi]
~=~
\overline{\Ec}^0(\Phi)
:=
\overline{\Ec}_0 \circ \cdots \circ \overline{\Ec}_{N-1} (\Psi).
$$
\end{Lemma}
\subsection{Dynamic programming principle on $\widehat{\Om}$}
Recall that the family of operators $({\cal E}_k)$ on functionals on $\Omega$ is defined in Example \ref{ex:non},
based on which one obtains a family of operators $(\widehat{\Ec}_k)$ on functionals on $\widehat{\Om}$ as in Example \ref{ex:product}.
\begin{Theorem}
Let $\widehat{\xi}: \widehat{\Om} \to \mathbb{R}^N$ be an upper semianalytic functional.
Then $\overline{\widehat{{\cal E}}}_k(\widehat{\xi})$ is also upper semianalytic and
$$
\sup_{\widehat{\Q} \in \overline{\widehat{\Mc}}_g} \mathbb{E}^{\widehat{\Q}}[ \widehat{\xi}]
~=~
\overline{\widehat{{\cal E}}}^0(\widehat{\xi})
:=
\overline{\widehat{{\cal E}}}_0 \circ \cdots \circ \overline{\widehat{{\cal E}}}_{N-1} (\widehat{\xi}).
$$
In particular, $\mathrm{NA}({\cal P})$ implies that \eqref{eq:PHA_dpp} holds.
\end{Theorem}
\proof
By Lemma \ref{lem:na} $\mathrm{NA}({\cal P})\iff\mathrm{NA}(\overline{\Pc})$, and then by Theorem \ref{theo:main} the pricing--hedging duality on $\overline{\Om}$ in \eqref{eq:main} holds.
Then, by Corollary \ref{cor:dpp_options}, \eqref{eq:PHA_dpp} is implied by the dynamic programming principle on $\widehat{\Om}$ for which it is enough to argue that the $\widehat{\Mc}$ satisfies the same analyticity property as ${\cal P}$, or ${\cal M}$,
and the assertion would follow by (4.12) in \cite{BN13}.
Given $k =0, \cdots, N-1$ and $\omega \in \Omega$,
similar to \eqref{eq:def_Pc},
we define
$$
{\cal M}'^{,k}(\omega) ~:=~ \{\mathbb{Q} := \omega \otimes_k \mathbb{Q}_{k+1} \otimes \cdots \otimes \mathbb{Q}_{N-1} ~: \mathbb{Q}_i \in {\cal M}^{i,i+1}(\omega)
\}.
$$
The collection ${\cal M}'^{,k}(\omega)$ induces a collection $\widehat{\Mc}'^{,k}(\omega,y)$ on $\widehat{\Om}$ by
$$
\widehat{\Mc}'^{,k}(\omega, y) ~:=~ \{\widehat{\Q} := \mathbb{Q} \circ (X, Y^{\mathbb{Q}})^{-1}, ~\mathbb{Q} \in {\cal M}'^{,k}(\omega),~ Y^{\mathbb{Q}}_k = y\}.
$$
By considering its marginal law at time $k+1$, we define
$$
\widehat{\Mc}^k(\omega,y) ~:=~\{ \mathbb{Q} \circ (X_{k+1}, Y^{\mathbb{Q}}_{k+1})^{-1}, ~\mathbb{Q} \in {\cal M}'^{,k}(\omega),~ Y^{\mathbb{Q}}_k = y \}.
$$
We claim that
\begin{eqnarray} \label{eq:claim_analytic}
\mbox{the graph}~
\{ (\omega, y, \widehat{\Q}) ~: \widehat{\Q} \in \widehat{\Mc}^k(\omega, y)\}
~\mbox{is analytic.}
\end{eqnarray}
Then the problem reduces to the same context as in Lemma \ref{lemma:super_hedging} and one obtains immediately the dynamic programming representation for $\sup_{\widehat{\Q} \in \overline{\widehat{\Mc}}_g} \mathbb{E}^{\widehat{\Q}}[\cdot]$ as stated.
To conclude the proof, it is enough to prove the claim \eqref{eq:claim_analytic}.
First, since the graph $\{(\omega, \mathbb{Q}) ~: \mathbb{Q} \in {\cal M}^{k,k+1}(\omega)\}$ is analytic,
it follows by Theorem 2 in \cite{DellacherieMaisonJeu}
that the graph
$\{(\omega, \mathbb{Q}) ~:= {\cal M}'^{,k}(\omega) \}$ is also analytic.
(Notice that the results in \cite{DellacherieMaisonJeu} is given in a Markovian context,
by considering the whole path, we can easily reduce our problem to his Markovian context.)
Next, by Lemma 3.1 of \cite{NN14}, we can choose a version of a family $(Y^{\mathbb{Q}})_{\mathbb{Q} \in {\cal B}}$ such that
$(\omega, \mathbb{Q}) \mapsto Y^{\mathbb{Q}}(\omega)$ is Borel measurable.
It follows that the graph
$\{(\omega, y, \mathbb{Q}) ~: \mathbb{Q} \in \widehat{\Mc}'^{,k}(\omega,y)\}$ is analytic,
thus
$\{(\omega, y, \mathbb{Q}) ~: \mathbb{Q} \in \widehat{\Mc}'^{,k}(\omega,y)\}$ is also analytic.
\qed
\subsection{Comparison with \cite{Bay} and \cite{BayZhou}}
\label{com:bay}
In \cite{Bay}, the authors considered the same superhedging problem $\pi^A_g(\Phi)$ with finite set $\Lambda=\{1,...,e\}$,
and established the duality
\begin{eqnarray} \label{eq:dualBay}
\pi^A_g(\Phi) = \inf_{h \in \mathbb{R}^e} \sup_{\tau \in {\cal T}(\mathbb{F})} \sup_{\mathbb{Q} \in {\cal M}_0} \mathbb{E}^{\mathbb{Q}}[ \Phi_{\tau} - hg],
\end{eqnarray}
under some regularity conditions (see Proposition 3.1 in \cite{Bay}).
Our duality in Theorem \ref{theo:main} is more general and more complete,
and moreover, together with Lemma \ref{lemma:super_hedging}, it induces the above duality \eqref{eq:dualBay}.
In exchange, \cite{Bay} also studied another subhedging problem $\sup_{\tau \in {\cal T}(\mathbb{F})} \inf_{\mathbb{Q} \in {\cal M}} \mathbb{E}^{\mathbb{Q}}[\Phi_{\tau}]$ which we do not consider here
More recently, \cite{BayZhou} consider the ``randomized'' stopping times,
and obtain a more complete duality for $\pi^A_g(\Phi)$.
The dual formulations in \cite{BayZhou} and in our results are more or less in the same spirit (as in \cite{Neuberger, HobsonNeuberger}).
Nevertheless, the duality in \cite{BayZhou} is established under strong integrability conditions and an abstract condition which is checked under regularity conditions (see their Assumption 2.1 and Remark 2.1).
In particular, when ${\cal P}$ is the class of all probability measures on $\Omega$, the integrability condition in their Assumption 2.1 is equivalent to say that $\Phi_k$ and $g^i$ are all uniformly bounded.
In our paper, we only assume that $g^i$ are Borel measurable, $\Phi_k$ are upper semi-analytic and all are $\mathbb{R}$--valued.
Technically, \cite{BayZhou} uses the duality results in \cite{BN13} together with a minimax theorem to prove their results.
Our first main result consists in introducing an enlarged canonical space (together with an enlarged canonical filtration)
to reformulate the main problem as a superhedging problem for European options.
Then by adapting the arguments in \cite{BN13}, we establish our duality under general conditions as in \cite{BN13}.
Moreover, we do not assume that $\Phi_k$ is ${\cal F}_k$-measurable, which permits to study the superhedging problem for a portfolio containing an American option and some European options.
Finally, our setting permits to use an approximation argument to study a new class of martingale optimal transport problem and to obtain a Kantorovich duality.
\section{A martingale (optimal) transport setup}
\label{sec:mot}
In this section we study the duality for American options in presence of a large family of statically traded European options.
We assume that the statically traded options on the market are all vanilla options,
and are arbitrage--free (see \cite{CO} and \cite{CHO}) and numerous enough so that one can recover the marginal distribution of the underlying process $S$ at some maturity times $\mathbb{T}_0 = \{t_1, \cdots, t_{M}\} \subseteq \mathbb{T}$, where $t_{M} = N$.
More precisely, we are given a vector $\mu = (\mu_{1}, \cdots, \mu_{{M}})$ of marginal distributions. We write $\mu(f):=(\int f(x)\mu_1(dx),...,\int f(x)\mu_{M}(dx))$ and we assume that $\mu(|\cdot|) < \infty$ and
\begin{eqnarray}
\label{pcoc}
\mu_i(f) \le \mu_j(f)
~~\mbox{for all}~
i \le j, ~i,j \leq M,
~\mbox{and convex function}~
f: \mathbb{R}^d \to \mathbb{R}.
\end{eqnarray}
Here we work with $\Omega :=\{s_0\}\times\mathbb{R}^{d\times N}$ where $s_0\in \mathbb{R}^d$, $S$ which is a canonical process on $\Omega$ and ${\cal P} := \mathfrak{P}(\Omega)$.
Thus $\overline{\Om} := \Omega \times \mathbb{T}$, $\overline{\Pc}=\frak P(\overline{\Om})$.
The condition \eqref{pcoc} ensures the existence of a calibrated martingale measure, i.e. that the following sets are non--empty
\begin{align*}
{\cal M}_\mu&:=\big\{\mathbb{Q} \in \mathfrak{P}(\Omega) ~: \mathcal{L}_{\mathbb{Q}}(S_{t_i})=\mu_i, ~i \leq M,~\mbox{and}~ S~\mbox{is an}~(\mathbb{Q},\mathbb{F}) \mbox{-martingale} \big\},\\
\overline{\Mc}_\mu&:= \big\{
\overline{\Q} \in \mathfrak{P}(\overline{\Om}) ~: \mathcal{L}_{\overline{\Q}}(S_{t_i})=\mu_i, ~i \leq M,~\mbox{and}~ S~\mbox{is a}~(\overline{\Q},\overline{\F}) \mbox{-martingale}
\big\}.
\end{align*}
Let $\Lambda_0$ be the class of all Lipschitz functions $\lambda: \mathbb{R}^d \to \mathbb{R}$,
and denote $\Lambda := \Lambda_0^{M}$.
The statically traded options $g=(g^\lambda)_{\lambda\in\Lambda}$ are given by $g^\lambda(\omega):=\lambda(\omega)-\mu(\lambda) $ where $\lambda(\omega) := \sum_{i=1}^{M} \lambda_i(\omega_{t_i})$ and
$\mu(\lambda) := \sum_{i=1}^{M} \mu_{t_i}(\lambda_i)$.
Recall that ${\cal M}_g={\cal M}_\mu$.
Since $\Lambda$ is a linear space, the superhedging cost of the American option $\Phi$ using semi--static strategies $\pi^A_g(\Phi)$ defined in Subsection \ref{superhedging} can be rewritten as
\begin{align*}
\pi^A_g(\Phi)=\pi^A_\mu(\Phi)&:=\inf\{\mu(\lambda): \exists (H^1,...,H^N)\in \mathcal H^N \ \textrm{s.t.} \ H^j_i=H^k_i \ \forall 1\leq i\leq j\leq k\leq N\\ &\textrm{and} \ \lambda\in \Lambda\
\textrm{satisfying} \ \lambda(\omega) \!+\! (H^k \!\circ\! S)_N(\omega) \geq \Phi_k(\omega)
\ \mbox{for all}~ k \in \mathbb{T},~ \omega \in \Omega \}.
\end{align*}
Similarly, we denote by $\overline{\pi}^E_{\mu}(\Phi)$ the corresponding superhedging cost for a European option with payoff $\Phi$ defined on $\overline{\Om}$, and one has $\pi^A_\mu(\Phi) = \overline{\pi}^E_{\mu}(\Phi)$ by Theorem \ref{thm:American_European}.
\begin{Example} \label{exam:non_equiv}
Construct an example similar to Example \ref{ex:intro} to highlight that in \eqref{eq:PHAduality} we may have a strict inequality. Consider the case $N=2$, $\mathbb{T}_0 = \mathbb{T} = \{1,2\}$, $\mu_1 = \delta_{\{0\}}$
and $\mu_2 = \frac{1}{4} \big( \delta_{\{-2\}} + \delta_{\{-1\}} + \delta_{\{1\}} + \delta_{\{2\}} \big)$.
Let $\Phi_1(\{S_1=0\}) = 1$, $\Phi_2(\{|S_2|=1\}) = 2$ and $\Phi_2(\{|S_2|= 2\}) =0$.
Then ${\cal M}_\mu$ contains only one probability measure $\mathbb{Q}$, and by direct computation, one has
$$
\mathbb{E}^{\mathbb{Q}} \big[ \Phi_{\tau}] = 1,
~~~\mbox{for all}~~ \tau \in {\cal T}(\mathbb{F}).
$$
Let us now construct a martingale measure $\overline{\Q}_0$ by
$$
\overline{\Q}_0(d \omega, d\theta)
~:=~
\frac{1}{4} \delta_{\{1\}}(d \theta) \otimes \big( \delta_{(0,1)} + \delta_{(0,-1)} \big)(d \omega)
~+~
\frac{1}{4} \delta_{\{2\}}(d \theta) \otimes \big( \delta_{(0,2)} + \delta_{(0,-2)} \big)(d \omega).
$$
Then one can check that $\overline{\Q}_0 \in \overline{\Mc}_\mu$ and it follows that
$$
\sup_{\overline{\Q} \in \overline{\Mc}_\mu}\mathbb{E}^{\overline{\Q}}[\Phi] \ge \mathbb{E}^{\overline{\Q}_0}[\Phi]
~=~ \frac{3}{2}
~>~ 1
~=~
\sup_{\mathbb{Q} \in {\cal M}_\mu} \sup_{\tau \in {\cal T}(\mathbb{F})} \mathbb{E}^{\mathbb{Q}} \big[ \Phi_{\tau}].
$$
Since the superhedging price of $\Phi$ equals to 3/2 as one can consider a superhedging strategy consisting of holding 3/2 in cash and one option $g$ from Example \ref{ex:intro}.
In a similar way as in Example \ref{ex:intro} the duality may be recovered by allowing a dynamic trading options.
\end{Example}
\subsection{Duality on the enlarged space $\overline{\Om}$}
The following theorem shows the duality for $\overline{\Om}$. Its proof is delegated to Section \ref{s:proofs3}.
\begin{Theorem} \label{theo:MOT}
Suppose that $\Phi : \overline{\Om} \to \mathbb{R}$ is bounded from above and upper semicontinuous.
Then there exists an optimal martingale measure $\overline{\Q}^* \in \overline{\Mc}_\mu$ and the pricing--hedging duality holds:
$$
\mathbb{E}^{\overline{\Q}^*}\big[\Phi\big]=\sup_{\overline{\Q} \in \overline{\Mc}_\mu}\mathbb{E}^{\overline{\Q}} \big[\Phi \big] ~~=~~ \overline{\pi}^E_\mu(\Phi).
$$
and in particular \eqref{eq:PHAduality_Euro} holds.
\end{Theorem}
\begin{Remark}
Note that in the above formulation each $\mu_i$ is an element of $\mathfrak{P}(\mathbb{R}^d)$. Instead one could take $\mu_i$ to be an element of $(\mathfrak{P}(\mathbb{R}))^d$. The same statements with analogous proofs would stay in force. This alternative formulation has more transparent financial interpretation since it corresponds only to marginal laws of terminal values of each stock price as opposed to the full distribution, see also \cite{Lim_mm} for a related discussion.
\end{Remark}
\subsection{Dynamic programming principle on $\widehat{\Om}$}
\cite{Eldan} and
\cite{CoxKall} studied the Skorokhod embedding problem and the martingale optimal transport in continuous time using the measure--valued martingales. This point of view allows to obtain the dynamic programming principle with marginal constraint since the terminal constraint is transformed into the initial constraint.
We adopt this perspective which proves to be very useful.\\
As before we work with the set of marginal times $\mathbb{T}_0=\{t_1, ..., t_{M}\}\subset\{1,...,N\}$ such that $t_{M}=N$, and marginal peacock measure $\mu=(\mu_{1},..., \mu_{{M}})$ where each $\mu_{i }$ is a probability measure on $\mathbb{R}^d$.
We let $\frak P_1(\mathbb{R}^d)=\{\eta\in \frak P(\mathbb{R}^d): \eta(|\cdot|)<\infty\}$ be the set of probability measures with finite first moment which we equip with the 1-Wasserstein distance, i.e. $\eta_n \to \eta_0$ if and only if
$$
\int_{\mathbb{R}^d} f(x) \eta_n(dx) \to \int_{\mathbb{R}^d} f(x) \eta_0(dx), ~~\forall f \in {\cal C}_1, $$
where ${\cal C}_1$ denotes the set of all continuous functions on $\mathbb{R}^d$ with linear growth, which makes $\frak P_1(\mathbb{R}^d)$ a Polish space.
Continuing with the construction from Example \ref{ex:product}, $\widehat{\Om}$ has to be an infinitely dimensional space and it is convenient to parametrize it as the canonical space for the measure--valued processes
$$\widehat{\Om} := \{\mu\}\times(\frak P_1(\mathbb{R}^d))^{M\times {N}}$$
and denote $\widehat{X}=(\widehat{X}^1_k, ..., \widehat{X}^{M}_k)_{0\leq k\leq N}$ the canonical process on $\widehat{\Om}$. Let $\widehat{\G} = (\widehat{\Gc}_k)_{0 \le k \le N}$ be the canonical filtration and $\widehat{\F} = (\widehat{\Fc}_k)_{0 \le k \le N}$ its universal completion.
Denote by ${\cal T}(\widehat{\F})$ the collection of all $\widehat{\F}$-stopping times.
For $f \in {\cal C}_1$ we denote the process of its integrals against $\widehat{X}$ as
\begin{align*}
\widehat{X}_k(f)&= (\widehat{X}^1_k(f),...,\widehat{X}^{M}_k(f)),\textrm{ where $\widehat{X}^i_k(f):= \int_{\mathbb{R}^d} f(x) \widehat{X}^{i}_k(dx)$ and}\\
\widehat{X}_k(id) &=(\widehat{X}^1_k(id),...,\widehat{X}^{M}_k(id)), \textrm{ where $\widehat{X}^i_k(id)= \int_{\mathbb{R}^d} x \widehat{X}^{i}_k (dx)$}.
\end{align*}
Define $\mathtt{i}: \widehat{\Om}\to\Omega$ by $\mathtt{i}(\widehat{\om})=(\widehat{X}^{M}_0(id)(\widehat{\om}),...,\widehat{X}^{M}_N(id)(\widehat{\om}))$ which is surjective and naturally extends processes on $\Omega$ to processes on $\widehat{\Om}$.
In particular the price process extends via $S_k(\widehat{\om})=S_k(\mathtt{i}(\widehat{\om}))=\widehat{X}^{M}_k(id)(\widehat{\om})$ and the statically traded options via $g^\lambda(\widehat{\om})=g^\lambda(\mathtt{i}(\widehat{\om}))=\lambda(\mathtt{i}(\widehat{\om}))-\mu(\lambda)$.
Define a family of processes $Y=(Y^\lambda)_{\lambda\in\Lambda}$ by $Y^\lambda=\sum_{i=1}^{M}Y^{\lambda_i}$ where
\begin{align*}
Y^{\lambda_i}_k&=
\begin{cases}
\widehat{X}^i_k(\lambda_i)-\mu_i(\lambda_i)& 0\leq k\leq t_i-1\\
g^{\lambda_i}=\lambda_i(\widehat{X}^i_{t_i}(id))-\mu_i(\lambda_i)&t_i\leq k\leq N
\end{cases}
\end{align*}
Note that $Y^{\lambda_i}_0=0$.
\begin{Definition}
{\rm (a)$\>\>$} A probability measure $\widehat{\Q}$ on $(\widehat{\Om}, \widehat{\Fc})$ is called a measure--valued martingale measure (MVM measure) if the process $(\widehat{X}_k(f))_{0 \le k \le N}$ is a $(\widehat{\Q}, \widehat{\F})$-martingale for all $f \in {\cal C}_1$.\\
{\rm (b)$\>\>$} A MVM measure $\widehat{\Q}$ is terminating if $\widehat{X}^i_{t_i} \in \Delta:=\{\eta\in \frak P(\mathbb{R}^d): \eta=\delta_x, \ x\in \mathbb{R}^d\}$, $\widehat{\Q}$-a.s.\\
{\rm (c)$\>\>$} A MVM measure $\widehat{\Q}$ is consistent if $S_k=\widehat{X}^i_k(id)$ for $k\leq t_i$ and $i=1, \cdots, M$, $\widehat{\Q}$-a.s.
\end{Definition}
Let us denote by
\begin{align*}
\widehat{\Mc}_\mu&=\{\widehat{\Q}\in \mathfrak{P}(\widehat{\Om}\}: \widehat{\Q} ~~\mbox{is terminating, consistent, MVM measure s.t. } \forall \,i\leq M\; \}.
\end{align*}
The following lemma shows that the marginal distribution of $S$ at $t_i$ equals to $\mu_i$, $\widehat{\Mc}_\mu$-q.s. And hence $\widehat{\Q}\circ \mathtt{i}^{-1}\in {\cal M}_\mu$ for any $\widehat{\Q}\in\widehat{\Mc}_\mu$.
\begin{Lemma}
\label{cond:law}
For a measure $\widehat{\Q}\in \widehat{\Mc}_\mu$ the following holds:\\
{\rm (a)$\>\>$} $\mathcal{L}_{\widehat{\Q}}(S_{t_i}| \widehat{\Fc}_k)=\widehat{X}^i_k ~ \widehat{\Q}\textrm{-a.s.}~\textrm{for} ~ k\leq t_i, ~ \textrm{and in particular}~ \mathcal{L}_{\widehat{\Q}}(S_{t_i})=\mu_i.$\\
{\rm (b)$\>\>$} For $k\leq t_j \leq t_i$, $\widehat{X}^j_k\preceq \widehat{X}^i_k$ $\widehat{\Q}$-a.s., i.e., for any convex function $f$
$$\int_{\mathbb{R}^d}f(x)\widehat{X}^j_k(dx)\leq \int_{\mathbb{R}^d}f(x)\widehat{X}^i_k(dx) \quad \widehat{\Q}\textrm{-a.s.}$$
\end{Lemma}
\proof
{\rm (a)$\>\>$}
Let $A\subset \mathbb{R}^d$ and recall that $S_k=\widehat{X}^i_k(id)$ $\widehat{\Q}$-a.s.
Then we have
$$\int_{\mathbb{R}^d}1\!\!1_A(x)\mathcal{L}_{\widehat{\Q}}\left(\widehat{X}^i_{t_i}(id) \Big| \widehat{\Fc}_k\right)(dx)
=\mathbb{E}^{\widehat{\Q}}\left[1\!\!1_{\{\widehat{X}^i_{t_i}(id)\in A\}} \Big|\widehat{\Fc}_k\right]
=\mathbb{E}^{\widehat{\Q}}\left[\widehat{X}^i_{t_i}(1\!\!1_A)\Big|\widehat{\Fc}_k\right]
=\widehat{X}^i_k(1\!\!1_A),$$
where the second equality holds since $\widehat{\Q}$ is terminating and the third one as $\widehat{\Q}$ is MVM measure. Hence the first assertion is proven.\\
{\rm (b)$\>\>$} Let $j\leq i$, $k\leq t_j$ and $f$ be a convex function.
Then
\begin{align*}
\int_{\mathbb{R}^d}f(x)\widehat{X}^i_k(dx)&=
\mathbb{E}^{\widehat{\Q}}\left[f\big(\widehat{X}^i_{t_i}(id)\big)|\widehat{\Fc}_k\right]\\
&\geq \mathbb{E}^{\widehat{\Q}}\left[f\big(\mathbb{E}^{\widehat{\Q}}[\widehat{X}^i_{t_i}(id)|\widehat{\Fc}_{t_j}]\big)|\widehat{\Fc}_k\right]\\
&=\mathbb{E}^{\widehat{\Q}}\left[f\big(\widehat{X}^i_{t_j}(id)\big)|\widehat{\Fc}_k\right]
=\mathbb{E}^{\widehat{\Q}}\left[f\big(\widehat{X}^j_{t_j}(id)\big)|\widehat{\Fc}_k\right]\\
&=\int_{\mathbb{R}^d}f(x)\widehat{X}^j_k(dx)
\end{align*}
where the first and the last equality follow by {\rm (a)$\>\>$}, the penultimate is due to the consistency of $\widehat{\Q}$ and the inequelity follows by conditional Jensen's inequality.
\qed
\vspace{2mm}
Let us denote by $\widehat{\Mc}$ the set of martingale measures for the dynamic extension $(\widehat{\Om}, \widehat{\Fc}, \widehat{\F}, \mathtt{i}, Y)$ of $(\Omega,{\cal F},\mathbb{F}, {\cal P},S,g)$ (see Section \ref{sec:dpp} for the definition of the dynamic extension).
\begin{Lemma}
{\rm (a)$\>\>$} Under any $\widehat{\Q}\in\widehat{\Mc}_\mu$, the processes $S$ and $Y^\lambda$, for $\lambda\in \Lambda$, are $(\widehat{\Q}, \widehat{\F})$-martingales.
In particular, one has $\widehat{\Mc}_\mu\subset\widehat{\Mc}$.
\noindent {\rm (b)$\>\>$} The mapping $\mathtt{I}:\widehat{\Mc}_\mu\to{\cal M}_\mu$,
defined by $\mathtt{I}(\widehat{\Q})=\widehat{\Q}\circ \mathtt{i}^{-1}$,
is surjective.
\end{Lemma}
\proof {\rm (a)$\>\>$} The process $S=\widehat{X}^{M}(id)$ is a $(\widehat{\Q}, \widehat{\F})$-martingale since $\widehat{\Q}$ is MVM measure.
To prove that $Y^\lambda$ is a $(\widehat{\Q}, \widehat{\F})$-martingale for any $\lambda\in\Lambda$, it is enough to show that for any $i\leq M$ and $\lambda \in \Lambda_0$ one has $\mathbb{E}^{\widehat{\Q}}\left[\lambda(\widehat{X}^{M}_{t_i}(id))|\widehat{\Fc}_k\right]=\widehat{X}^i_k(\lambda)$ for any $k<t_i$.
The latter holds since
\begin{align*}
\mathbb{E}^{\widehat{\Q}}\left[\lambda(\widehat{X}^{M}_{t_i}(id))|\widehat{\Fc}_k\right]&=
\mathbb{E}^{\widehat{\Q}}\left[\lambda(\widehat{X}^{i}_{t_i}(id))|\widehat{\Fc}_k\right]=
\mathbb{E}^{\widehat{\Q}}\left[\widehat{X}^{i}_{t_i}(\lambda)|\widehat{\Fc}_k\right]=\widehat{X}^i_k(\lambda),
\end{align*}
where the first equality follows by consistency of $\widehat{\Q}$, the second since $\widehat{\Q}$ is terminating and the last one as $\widehat{\Q}$ is MVM measure.
\vspace{1mm}
{\rm (b)$\>\>$}
Let $\mathbb{Q}\in {\cal M}_\mu$ and define the process $\eta=(\eta^1_k,.., \eta^{M}_k)_{k\leq N}$ by $\eta^i_k=\mathcal{L}_{\mathbb{Q}}(S_{t_i}|{\cal F}_k)$.
Note that by definition there exists a terminating, consistent MVM measure $\widehat{\Q}$ such that $\widehat{\Q}[\widehat{X}=\eta]=1$.
\qed
\vspace{2mm}
For $\widehat{\om} \in \widehat{\Om}$, we define a set $[\widehat{\om}]_{\widehat{\Gc}_k}$ as in \eqref{eq:om_k},
and denote by $\widehat{\Mc}^k_\mu(\widehat{\om})$ the following set of measures:
\begin{align*}
\widehat{\Mc}^k_\mu(\widehat{\om})&:=\Big\{\widehat{\Q}\in \mathfrak{P}(\widehat{\Om}):\widehat{\Q}~\mbox{is terminating and consistent,}\\
&~~~~~~~~~~~~~~~~~~~~~~~~~\widehat{\Q}([\widehat{\om}]_{\widehat{\Gc}_{k}})=1 ~~ \mbox{and $(\widehat{X}_{l})_{k\leq l\leq N}$ is a $(\widehat{\Q},\widehat{\F})$-MVM} \Big\}.
\end{align*}
Let us define a family of operators $\widehat{\Ec}_k$, etc., as in Example \ref{ex:non}:
\begin{align*}
\widehat{\Ec}_k(\widehat{\xi})(\widehat{\om})&=\sup_{\mathbb{Q}\in\widehat{\Mc}^{k}_\mu(\widehat{\om})}\mathbb{E}^{\mathbb{Q}}\left[\widehat{\xi}\right],\quad \widehat{\xi}\in \widehat\fcl,
\end{align*}
and then the extension $\overline{\widehat{{\cal E}}}_k$ as well as $\overline{\widehat{{\cal E}}}^0$ on the enlarged space as in Section \ref{ameu}.
Then we have:
\begin{Theorem}
For all upper semianalytic functionals $\widehat{\xi} : \widehat{\Om} \to \mathbb{R}^N$,
$\overline{\widehat{{\cal E}}}_k(\widehat{\xi})$ is also upper semianalytic and
\begin{eqnarray} \label{eq:claim_DPP_mvm}
\sup_{\overline{\widehat{\Q}} \in \overline{\widehat{\Mc}}_{\mu}} \mathbb{E}^{\overline{\widehat{\Q}}} [ \widehat{\xi}] = \overline{\widehat{{\cal E}}}^0 (\widehat{\xi}).
\end{eqnarray}
In particular the pricing--hedging duality \eqref{eq:PHA_dpp} holds in this MOT context for all functionals $\Phi: \overline{\Om} \to \mathbb{R}^N$ which are upper semicontinuous and bounded from above.
\end{Theorem}
\proof
Notice that the pricing--hedging duality on $\overline{\Om}$ holds by Theorem \ref{theo:MOT}.
Then by Corollary \ref{cor:dpp_options},
it is enough to establish the dynamic programming principle on $\widehat{\Om}$ to prove the pricing-hedging duality \eqref{eq:PHA_dpp}.
Using exactly the same arguments as in (4.12) of \cite{BN13}, to establish the dynamic programming principle on $\widehat{\Om}$, it is enough to argue that $\widehat{\Mc}_\mu$ satisfies that
$$
\{(\widehat{\om}, \widehat{\Q}) ~: \widehat{\Q} \in \widehat{\Mc}_{\mu}^k(\widehat{\om}) \}~~\mbox{is analytic}.
$$
To prove the above analyticity property, we first observe that
$$\widehat{\Ec}_k\circ ... \circ \widehat{\Ec}_{N-1}(\widehat{\xi})(\widehat{\om})=\widehat{\Ec}_{k,k+1}\circ ... \circ \widehat{\Ec}_{N-1,N}(\widehat{\xi})(\widehat{\om}),\quad \widehat{\xi}\in \widehat\fcl$$
where ${\cal E}_{k,k+1}(\widehat{\xi})(\omega)=\sup_{{\cal M}^{k,k+1}_\mu(\omega)}\mathbb{E}^{\mathbb{Q}}\left[\widehat{\xi}\right]$ and
\begin{align*}
\widehat{\Mc}^{k,k+1}_\mu(\widehat{\om})&:=\Big\{\widehat{\Q}\in \mathfrak{P}(\widehat{\Om}):\widehat{\Q}~\mbox{is terminating and consistent,}\\
&~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\widehat{\Q}([\widehat{\om}]_{\widehat{\Gc}_{k}})=1 ~~ \widehat{\om}_k(f) = \mathbb{E}^{\widehat{\Q}}[ \widehat{X}_{k+1}(f) ],
~~\forall f \in {\cal C}_1\Big\}.
\end{align*}
Next, let ${\cal C}_1^0$ denote a countable dense subset of ${\cal C}_1$ under the uniform convergence topology.
Then it is clear that for each $k \in \mathbb{T}$, the set
\begin{align*}
\big\{(\widehat{\om}, \widehat{\Q})\in & \widehat{\Om} \times \mathfrak{P}(\widehat{\Om}) :\ \widehat{\Q} \in \widehat{\Mc}^{k,k+1}_\mu(\widehat{\om}) \big\}
=\big\{(\widehat{\om}, \widehat{\Q}) \in \widehat{\Om} \times \mathfrak{P}(\widehat{\Om}) ~: \widehat{\Q}([\widehat{\om}]_{\widehat{\Gc}_{k}})=1\\
& ~~~~\widehat{\Q}~\mbox{is terminating and consistent,} \ \ \widehat{\om}_k(f) = \mathbb{E}^{\widehat{\Q}}[ \widehat{X}_{k+1}(f) ],
~~\forall f \in {\cal C}_1^0
\big\}
\end{align*}
is a Borel set.
\qed
\section{Proofs for Section \ref{duality}}
\label{proofs1}
\proof[Proof of Proposition \ref{prop:dpp_Euro_American}]
First we prove that \eqref{eq:dpp_omb} implies \eqref{dpp_om}. For a given $\xi$ on $\Omega$ let us define $\Psi$ on $\overline{\Om}$ by
\begin{align*}
\Psi((\omega,k))&=-\infty\ \ \textrm{if} \ k\in \{1,...,N-1\}\\
\Psi((\omega, N))&=\xi(\omega) .
\end{align*}
Definition of $\Psi$ combined with \eqref{eq:dpp_omb} implies that
\begin{align*}
\sup_{\overline{\Q}\in \overline{\Mc}}\mathbb{E}^{\overline{\Q}}[\Psi]&={\cal E}_0\circ{\cal E}_1\circ...\circ{\cal E}_{N-1}\left(\xi\right).
\end{align*}
Moreover, one has that
\begin{align*}
\sup_{\overline{\Q}\in \overline{\Mc}}\mathbb{E}^{\overline{\Q}}[\Psi]&=\sup_{\mathbb{Q}\in{\cal M}}\mathbb{E}^{\mathbb{Q}}\left[\Psi_{N}\right]=\sup_{\mathbb{Q}\in{\cal M}}\mathbb{E}^{\mathbb{Q}}\left[\xi\right]
\end{align*}
since for a measure $\overline{\Q}\in \overline{\Mc}$ such that $\overline{\Q}(\Omega\times \{1,...,N-1\})>0$ the expected value drops to $-\infty$.\\
Now let us prove that \eqref{dpp_om} and \eqref{inequality2} imply \eqref{eq:dpp_omb}.
Define an $\mathbb{F}$-stopping time $\tau^*$ by
\begin{align}
\label{tau}
\tau^*(\omega)&:=\inf \left\{k\geq 1: {\cal E}_k\left(\Psi(\cdot, k)\right)(\omega)= \overline{\Ec}^{k}(\Psi)(\omega, k)\right\}\\
\nonumber
&= \inf \left\{k\geq 1: {\cal E}_k\left(\Psi(\cdot, k)\right)(\omega)\geq {\cal E}_k\left(\overline{\Ec}^{k+1}(\Psi)(\cdot, k+1)\right)(\omega)\right\}.
\end{align}
Note that on $\{k<\tau^*\}$ one has
\begin{align}
\label{before}
{\cal E}_k\left(\Psi(\cdot, k)\right)(\omega)< \overline{\Ec}^{k}(\Psi)(\omega, k)={\cal E}_k\left(\overline{\Ec}^{k+1}(\Psi)(\cdot, k+1)\right)(\omega).
\end{align}
Then
\begin{align*}
\overline{\Ec}_0&\circ...\circ \overline{\Ec}_{N-1}(\Psi)=\\
&={\cal E}_0\left(1\!\!1_{\{\tau^*=1\}}{\cal E}_1\left(\overline{\Ec}^{1}(\Psi)(\cdot, 1)\right)+1\!\!1_{\{\tau^*>1\}}{\cal E}_1\left(\overline{\Ec}^{1}(\Psi)(\cdot, 2)\right)\right)\\
&=...=\\
&={\cal E}_0\circ{\cal E}_1\circ...\circ{\cal E}_{N-1}(\Psi_{\tau^*})\\
&=\sup_{\mathbb{Q}\in{\cal M}}\mathbb{E}^{\mathbb{Q}}\left[\Psi_{\tau^*}\right]
\end{align*}
where the last equality follows from DPP on $\Omega$ \eqref{dpp_om}.
Note as well that
\begin{align}
\label{ineaulity1}
\overline{\Ec}_0&\circ...\circ \overline{\Ec}_{N-1}(\Psi)=\sup_{\mathbb{Q}\in{\cal M}}\mathbb{E}^{\mathbb{Q}}\left[\Psi_{\tau^*}\right]\leq \sup_{\mathbb{Q}\in {\cal M}}\sup_{\tau\in {\cal T}(\mathbb{F})}\mathbb{E}^{\mathbb{Q}}[\Psi_\tau]\leq
\sup_{\overline{\Q}\in \overline{\Mc}}\mathbb{E}^{\overline{\Q}}[\Psi].
\end{align}
Combining \eqref{inequality2}, we then conclude the proof.\qed
\begin{Proposition} \label{prop:non}
The family $(\overline{\Ec}_k)$ given in Example \ref{ex:non} satisfies \eqref{inequality2}.
\end{Proposition}
\proof
In the context of Example \ref{ex:non}, the family $(\overline{\Ec}_k)$ take the following form:
\begin{align}
\overline{\Ec}_0(\Psi)&:=\sup_{\mathbb{Q}\in {\cal M}}\mathbb{E}^\mathbb{Q}[\Psi(\cdot, 1)]\\
\overline{\Ec}_k(\Psi)(\overline{\om})&:=
\begin{cases}
\sup_{\mathbb{Q}\in {\cal M}^k(\omega)}\mathbb{E}^\mathbb{Q}[\Psi(\cdot, \theta)]&\textrm{if }\theta<k\nonumber\\
\sup_{\mathbb{Q}\in {\cal M}^k(\omega)}\mathbb{E}^\mathbb{Q}[\Psi(\cdot, k)]\lor \sup_{\mathbb{Q}\in {\cal M}^k(\omega)}\mathbb{E}^\mathbb{Q}[\Psi(\cdot, k+1)]&\textrm{if } \theta\geq k.\nonumber
\end{cases}
\end{align}
To see that \eqref{inequality2} holds, it is insightful to rewrite $\overline{\Ec}^0$ in a slightly different way, as $\widetilde{{\cal E}}^0$ below.
Let
$$\overline{{\cal G}}^-_k:={\cal G}_k\otimes \sigma(T\land k)\subset \overline{{\cal G}}_k:={\cal G}_k\otimes \sigma(T\land (k+1))\subset {\cal F}_k\otimes \sigma(T\land (k+1))=:\overline{{\cal F}}_k,$$
$$\overline{\Mc}^{k,-}(\overline{\om}):=\{\overline{\Q}\lll\overline{\Pc} : \ \overline{\Q}\left[[\overline{\om}]_{\overline{{\cal G}}^-_k}\right]=1\ \textrm{and}\ \mathbb{E}^{\overline{\Q}}[\Delta S_n|\overline{{\cal F}}_{n-1}]=0 \ \forall n\in\{k+1,..., N\} \},$$
where $[\overline{\om}]_{\overline{{\cal G}}^-_k}$ is defined as in \eqref{eq:om_k}.
Next, for $\Psi\in \overline\fcl$, let us introduce operators
\begin{align*}
\widetilde{{\cal E}}_0(\Psi):=\sup_{\mathbb{Q}\in \overline{\Mc}}\mathbb{E}^\mathbb{Q}[\Psi(\cdot, 1)],\qquad \widetilde{{\cal E}}_k(\Psi)(\overline{\om}):=
\sup_{\overline{\Q}\in \overline{\Mc}^{k,-}(\overline{\om})}\mathbb{E}^{\overline{\Q}}[\Psi],\ k\leq N-1.
\end{align*}
Denote $ \overline{\Ec}^k(\cdot) := \overline{\Ec}_k \circ \cdots \circ \overline{\Ec}_{N-1}(\cdot)$ and $ \widetilde{{\cal E}}^k(\cdot) := \widetilde{{\cal E}}_k \circ \cdots \circ \widetilde{{\cal E}}_{N-1}(\cdot)$, and we claim that
\begin{eqnarray} \label{eq:dpp_two}
\overline{\Ec}^k(\Psi)(\overline{\om})=\widetilde{{\cal E}}^k(\Psi)(\overline{\om}),~0\leq k<N, ~\Psi\in \overline\fcl.
\end{eqnarray}
Note that the conditional regular probabilities of any $\mathbb{Q}\in\overline{\Mc}$ w.r.t. $\overline{{\cal G}}^-_k$, denoted $\mathbb{Q}_{\overline{\om}}$,
satisfy $\mathbb{Q}\left[\{\overline{\om}:\mathbb{Q}_{\overline{\om}}\in \overline{\Mc}^{k,-}(\overline{\om})\}\right]=1$ and one has $\mathbb{E}^\mathbb{Q}[\Psi|\overline{{\cal F}}^-_k]\leq \widetilde{{\cal E}}_k(\Psi)$, $\mathbb{Q}$-a.s., which implies \eqref{inequality2} by the tower property of the conditional expectation and the definition of $\widetilde{{\cal E}}^0$.
Then it is enough to prove the claim \eqref{eq:dpp_two}.
Note that, for $\overline{\om}=(\omega, \theta)$ with $\theta\leq k-1$, a measure $\overline{\Q}\in \overline{\Mc}^{k,-}(\overline{\om})$ satisfies $\overline{\Q}_{\vert\Omega}\in {\cal M}^k(\omega)$ and $\overline{\Q}(\Omega\times\{\theta\})=1$; and a measure $\mathbb{Q}\in {\cal M}_k(\omega)$ satisfies $\mathbb{Q}\otimes \delta_\theta\in \overline{\Mc}^{k,-}(\overline{\om})$. It is thus clear that, in this case, $\overline{\Ec}^k(f)(\overline{\om})=\widetilde{{\cal E}}^k(f)(\overline{\om}).$
As a second step, for $\overline{\om}=(\omega, \theta)$ with $\theta\geq k$, we show that $\overline{\Ec}_{k}(f)(\overline{\om})\leq \widetilde{{\cal E}}_{k}(f)(\overline{\om})$.
Take any $\mathbb{Q}\in{\cal M}^{k}(\omega)$. Then, for $n\in\{k,..., N\}$, $\mathbb{Q}\otimes \delta_n\in \overline{\Mc}^{k,-}(\overline{\om})$ and $\mathbb{Q}\otimes\delta_n(\Omega\times \{n\})=1$. Hence it follows that $\overline{\Ec}_{k}(f)(\overline{\om})\leq \widetilde{{\cal E}}_{k}(f)(\overline{\om})$.
In a final step, we show that, for $\overline{\om}=(\omega, \theta)$ with $\theta\geq k$, $\overline{\Ec}_k(f)(\overline{\om})\geq\widetilde{{\cal E}}_{k}(f)(\overline{\om})$ holds.
Let us start with $k=N-1$.
Take any $\overline{\Q}\in\overline{\Mc}^{N-1,-}(\overline{\om})$ and consider its r.c.p w.r.t. $\overline{{\cal G}}_{N-1}$
(the atom $\{\omega\}\times\{N-1,N\}$ is divided into atoms $\{\omega\}\times\{N-1\}$ and $\{\omega\}\times\{N\}$)
denoted by $\overline{\Q}_N$ and $\overline{\Q}_{N-1}$. Then, clearly, ${\overline{\Q}_N}_{\vert\Omega}$ and ${\overline{\Q}_{N-1}}_{\vert\Omega}$ belong to ${\cal M}^{N-1}(\omega)$,
and $\overline{\Q}_N(\{\omega\}\times\{N\})=1$ and $\overline{\Q}_{N-1}(\{\omega\}\times\{N-1\})=1$. Thus, it follows that $\overline{\Ec}_{N-1}(f)(\overline{\om})\geq \widetilde{{\cal E}}_{N-1}(f)(\overline{\om})$.
Finally, to complete the proof, we need to show that $\overline{\Ec}^{k+1}(f)(\overline{\om})=\widetilde{{\cal E}}^{k+1}(f)(\overline{\om})$ implies $\overline{\Ec}^k(f)(\overline{\om})\geq\widetilde{{\cal E}}^k(f)(\overline{\om})$ for $\overline{\om}=(\omega, \theta)$ with $\theta\geq k$.
First note that $\overline{\Ec}^{k+1}(f)(\overline{\om})=\widetilde{{\cal E}}^{k+1}(f)(\overline{\om})$ is constant on $\theta\in \{k,...,N\}$, i.e.,
\begin{equation}
\label{constant}
\overline{\Ec}^{k+1}(f)((\omega, \theta_1))=\overline{\Ec}^{k+1}(f)((\omega,\theta_2))\ \ \ \textrm{for all $\omega\in\Omega$ and $\theta_1, \theta_2\in \{k,..., N\}$.}
\end{equation}
Take any $\overline{\Q}\in\overline{\Mc}^{k,-}(\overline{\om})$ and consider its r.c.p w.r.t. $\vartheta_N$
(the atom $\{\omega\}\times\{k,...,N\}$ is divided into atoms $\{\omega\}\times\{n\}$ for $n=k,...,N$)
denoted by $\overline{\Q}_n$ for $n=k,...,N$. Then, clearly, ${\overline{\Q}_n}_{\vert\Omega}\in{\cal M}^{k}(\omega)$
and $\overline{\Q}_n([\omega]_k\times\{n\})=1$ where $[\omega]_k$ denotes an atom of ${\cal G}_k$ which contains $\omega$. Thus, combining with \eqref{constant}, it follows that $\overline{\Ec}_{k}(f)(\overline{\om})\geq \widetilde{{\cal E}}_{k}(f)(\overline{\om})$.\qed
\section{Proofs for Section \ref{sec:main}}
\label{proofs2}
We are now in the context of Section \ref{sec:main}, where $\Omega_0 := \{\omega_0\}$ is a singleton, $\Omega_1$ is a nonempty Polish space and $\Omega := \Omega_0 \times \Omega_1^N$.
For technical reason, we introduce a $\Omega_1$-valued canonical process $X = (X_k)_{0 \le k \le N}$ on the enlarged space $\overline{\Om}$ by
$X_k(\overline{\om}) := \omega_k$ for all $\overline{\om} := (\omega, \theta) \in \overline{\Om}$,
and an enlarged filtration $\overline{\G} = (\overline{{\cal G}}_k)_{0 \le k \le N}$ by
$$
\overline{{\cal G}}_0 := \{ \emptyset, \overline{\Om}\}
~~~\mbox{and}~~
\overline{{\cal G}}_k :=\sigma \big\{ X_i, ~\{T \le i\},~ i=1, \cdots, k \big\},
$$
and the universally completed filtration $\overline{\F} = (\overline{{\cal F}}_k)_{0 \le k \le N}$ by defining
$\overline{{\cal F}}_k$ as the universal completion of $\overline{{\cal G}}_k$.
It follows that the random time $T: \overline{\Om} \to \mathbb{T}$ is an $\overline{\G}$-stopping time.
We also define a restricted enlarged space, for every $k = 1, \cdots, N$,
$$
\overline{\Om}_k
~:=~
\Omega_k \times \{1, \cdots, k\}
~=~
\Omega_1^k \times \{1, \cdots, k\}.
$$
\begin{Lemma}
Let $\overline{\P}\in\overline{\Pc}$ be a probability measure on $(\overline{\Om}, \overline{{\cal G}}_N)$,
and $(\overline{\P}_{\overline{\om}})_{\overline{\om} \in \overline{\Om}}$ be a family of regular conditional probability distribution of $\overline{\P}$ w.r.t. $\overline{{\cal G}}_k$.
Then for every $k =0, 1, \cdots, N-1$, one has $\overline{\P}_{\overline{\om}} \circ X_{k+1}^{-1} \in {\cal P}_k(\omega)$ for $\overline{\P}$-a.e. $\overline{\om} = (\omega, \theta) \in \overline{\Om}$.
\end{Lemma}
Let us introduce the following set of measures
\begin{eqnarray*}
\overline{\Mc}_g^{loc}
\!\!\! &:=&\!\!
\{\overline{\Q} ~: \overline{\Q} \lll \overline{\Pc}, ~\mathbb{E}^{\overline{\Q}}[g^i] = 0,~i\in\{1,\cdots, e\} \\
&&~~~~~~~~~~~~~~
~\mbox{and}~
S ~\mbox{is an}~ (\overline{\F}, \overline{\Q}) \mbox{-local martingale}
\}.
\end{eqnarray*}
\begin{Lemma} \label{lemm:weak_duality}
Let $\Phi$ be upper semianalytic and $\overline{\Q} \in \overline{\Mc}^{loc}_g$.
Then for any $x \in \mathbb{R}$ and $(\overline{H}, h) \in \overline{\Hc} \times \mathbb{R}^e$ such that $x + (\overline{H} \circ S)_N(\overline{\om}) + hg(\omega) \ge \Phi(\overline{\om})$, $\overline{\Q}$-a.s. one has
$\mathbb{E}^{\overline{\Q}}\big[ \Phi \big] \le x$.
\end{Lemma}
\proof
The proof follows by exactly the same arguments as in Lemma A.2 of \cite{BN13}, using the discrete time local martingale characterization in Lemma A.1 of \cite{BN13}.
\qed
\vspace{2mm}
Given $\overline{\Q} \in \overline{\Mc}^{loc}_0$ and $\varphi : \Omega \to [0, \infty)$, we denote
\begin{eqnarray*}
\overline{\Mc}^{\varphi,\overline{\Q}}
\!\!\! &:=&\!\!
\{\overline{\Q}' \sim \overline{\Q}~: \mathbb{E}^{\overline{\Q}'}[\varphi] < \infty, ~\mbox{and}~
S ~\mbox{is an}~ (\overline{\F}, \overline{\Q}') \mbox{-martingale}
\}.
\end{eqnarray*}
Then by Lemma \ref{lemm:weak_duality}, one can easily obtain the weak duality:
\begin{eqnarray}\label{eq:weak_duality}
\sup_{\overline{\Q} \in \overline{\Mc}_g} \mathbb{E}^{\overline{\Q}}[\Phi]
\le
\sup_{\overline{\Q} \in \overline{\Mc}_g^{loc}} \mathbb{E}^{\overline{\Q}}[\Phi]
\le
{ \overline{\pi}^E_g(\Phi)}.
\end{eqnarray}
\begin{Lemma} \label{lemm:red_Mcb_varphi}
Let $\Phi$ be upper semianalytic and $\overline{\Q} \in \overline{\Mc}^{loc}_0$ and $\varphi : \Omega \to [1, \infty)$ be such that $|\Phi(\omega,k)| \le \varphi(\omega)$ for all $\overline{\om}=(\omega,k) \in \overline{\Om}$.
Then $\overline{\Mc}^{\varphi,\overline{\Q}} ~\neq~ \emptyset$,
and moreover,
$$\mathbb{E}^{\overline{\Q}}[\Phi] ~\le~ \sup_{\overline{\Q}' \in \overline{\Mc}^{\varphi,\overline{\Q}}} \mathbb{E}^{\overline{\Q}'}[\Phi].$$
\end{Lemma}
\proof
First, by Lemma 3.2 of \cite{BN13}, there exists a probability $\overline{\P}_*$ equivalent to $\overline{\Q}$ on $(\overline{\Om}, \overline{{\cal F}}_N)$ such that $\mathbb{E}^{\overline{\P}_*}[\varphi(X)] < \infty$.
On the filtered probability space $(\overline{\Om}, \overline{{\cal F}}_N, \overline{\F}, \overline{\P}_*)$,
one defines $\overline{\Mc}^{loc}_*$ as the collection of all probability measures $\overline{\Q}' \sim \overline{\Q} \sim \overline{\P}_*$ under which $S$ is an $\overline{\F}$-local martingale.
Denote
$$\pi^{E,\overline{\Q}}_0(\Phi): = \mathrm{inf} ~\big\{ x: \exists \overline{H} \in \overline{\Hc} ~~\mbox{s.t.}~~ x+ (\overline{H} \circ S)_N \geq \Phi, \overline{\Q}\mbox{-a.s.} \big\} ,$$
then by the classical arguments for the dominated discrete time market (such as \cite{Kabanov, KabanovStricker}, see also Lemma A.3 of \cite{BN13}),
one can easily obtain the inequality
$$
\mathbb{E}^{\overline{\Q}}[\Phi] ~\le~ \sup_{\overline{\Q}' \in \overline{\Mc}^{loc}_{*}} \mathbb{E}^{\overline{\Q}'}[\Phi]
~\le~ \overline{\pi}^{E,\overline{\Q}}_0(\Phi)
~\le~ \sup_{\overline{\Q}' \in \overline{\Mc}^{\varphi,\overline{\Q}}} \mathbb{E}^{\overline{\Q}'}[\Phi],
$$
which concludes the proof.
\qed
\vspace{2mm}
Using Theorem 2.2 of \cite{BN13}, one can easily obtain a closedness result for the set of all payoffs which can be super--replicated from initial capital $x=0$, in our context.
Let us denote by $\overline{\Lc}^0_+$ the set of all positive random variables on $\overline{\Om}$,
and define
$$
\overline{\Cc} ~:=~ \big\{ (\overline{H} \circ S)_N + h g ~: \overline{H} \in \overline{\Hc}, h \in \mathbb{R}^e \big\} - \overline{\Lc}^0_+.
$$
\begin{Lemma} \label{lemm:closedness}
Let $\Phi$ be upper semianalytic and $\mathrm{NA}(\overline{\Pc})$. Then the set $\overline{\Cc}$ is closed in the following sense: \\
Let $(W^n)_{n \ge 1} \subset \overline{\Cc}$ and $W$ be a random variable such that $W^n \to W$, $\overline{\Pc}$-q.s., then $W \in \overline{\Cc}$.
\end{Lemma}
\proof
It is a direct consequence of Theorem 2.2 of \cite{BN13},
where the results are given in a general abstract context.
\qed
\subsection{Proof of Theorem \ref{theo:main}: the case $e =0 $, equivalently $\Lambda=\emptyset$}
For each $1 \le i \le j \le N$, we introduce an application from $\Omega_j$ to $\Omega_i$ (resp. $\overline{\Om}_j$ to $\overline{\Om}_i$) by
$$
[\omega]_i \!:=\! (\omega_1, \cdots, \omega_i), ~\mbox{for all}~\omega \in \Omega_j
~\mbox{(resp.}
[\overline{\om}]_i ~:=~ ([\omega]_i, \theta \wedge i),
~\mbox{for all}~
\overline{\om} = (\omega, \theta) \in \overline{\Om}_j).
$$
Note that $\overline{{\cal F}}_k^-$ is the smallest $\sigma$-field on $\overline{\Om}$ generated by $[\cdot]_k: \overline{\Om} \to \overline{\Om}_k$;
or equivalently, an $\overline{{\cal F}}_k^-$-measurable random variable $f$ defined on $\overline{\Om}$ can be identified as a Borel measurable function on $\overline{\Om}_k$.
The canonical processes $X$ and $S$ are naturally defined on the restricted spaces
$\Omega_k$ and $\overline{\Om}_k$.
We next recall a notion of local no--arbitrage condition $\mathrm{NA}({\cal P}_k(\omega))$ introduced at the beginning of Section 4.2 in \cite{BN13}.
Given a fixed $\omega \in \Omega_k$, we can consider $\Delta S_{k+1}(\omega, \cdot) :=S_{k+1}(\omega, \cdot) - S_k(\omega)$ as a random variable on $\Omega_1$, which determines a one--period market on $(\Omega_1, {\cal B}(\Omega_1))$ endowed with a class ${\cal P}_k(\omega)$ of probability measures.
Then $\mathrm{NA}({\cal P}_k(\omega))$ denotes the corresponding no--arbitrage condition in this one--period market, i.e., $\mathrm{NA}({\cal P}_k(\omega))$ holds if for all $H\in \mathbb{R}^d$
$$
H \Delta S_{k+1}(\omega, \cdot) \geq 0
\quad {\cal P}_k(\omega)\mbox{-q.s.}~
\Longrightarrow~
H \Delta S_{k+1}(\omega, \cdot) = 0
\quad {\cal P}_k(\omega)\mbox{-q.s.}
$$
\begin{Lemma} \label{lemm:selectH}
In the context of Section \ref{sec:main},
let $f: \overline{\Om}_{k+1} \to \overline{\R}$ be upper semianalytic,
then $\overline{\Ec}_k(f): \overline{\Om}_k \to \overline{\R}$ is still upper semianalytic.
Moreover, there exist two universally measurable functions $(y_1, y_2): \overline{\Om}_k \to \mathbb{R}^d \times \mathbb{R}^d$
such that
\begin{eqnarray*}
&&
\overline{\Ec}_k(f) (\overline{\om}) ~+~ y_1(\overline{\om}) \Delta S_{k+1}(\omega, \cdot)
~\ge~ f(\omega, \cdot, \theta)
~~{\cal P}_k(\omega) \mbox{-q.s.}\\
&&
\overline{\Ec}_k(f) (\overline{\om}) ~+~ y_2(\overline{\om}) \Delta S_{k+1}(\omega, \cdot)
~\ge~ f(\omega, \cdot, k+1)
~~{\cal P}_k(\omega) \mbox{-q.s. }
\end{eqnarray*}
for all $\overline{\om} = (\omega, \theta) \in \overline{\Om}_k$ such that $\mathrm{NA}({\cal P}_k(\omega))$ holds
and $f(\omega, \cdot, \theta) > - \infty, ~{\cal P}_k(\omega)$-q.s.
$f(\omega, \cdot, k+1) > - \infty, ~{\cal P}_k(\omega)$-q.s.
\end{Lemma}
\proof
Notice that $f_1 \vee f_2$ is upper semianalytic whenever $f_1$ and $f_2$ are both upper semianalytic.
Then the above lemma is an immediate consequence of Lemma 4.10 of \cite{BN13}
as well as the definition of $\overline{\Ec}_k$.
\qed
\vspace{2mm}
Recall that $\overline{\Mc}_0$ (resp. $\overline{\Mc}^{loc}_0$) means $\overline{\Mc}_g$ (resp. $\overline{\Mc}^{loc}_g$) for the case $e=0$.
\vspace{2mm}
\noindent {\bf Proof of Theorem \ref{theo:main} (the case $e=0$).}
First, one has the weak duality as in \eqref{eq:weak_duality}
$$
\sup_{\overline{\Q} \in \overline{\Mc}_g} \mathbb{E}^{\overline{\Q}}[ \Phi \big]
~\le~
\overline{\pi}^E_g(\Phi).
$$
Next, for the inverse inequality, we can assume, without loss of generality, that $\Phi$ is bounded from above.
Indeed, by Lemma \ref{lemm:closedness}, one has
$ \lim_{n \to \infty} \overline{\pi}^E_g(\Phi \wedge n) = \overline{\pi}^E_g(\Phi)$
(see also the proof of Theorem 3.4 of \cite{BN13}).
Besides, the approximation
$\lim_{n \to \infty} \sup_{\overline{\Q} \in \overline{\Mc}_g} \mathbb{E}^{\overline{\Q}} [\Phi \wedge n] = \sup_{\overline{\Q} \in \overline{\Mc}_g} \mathbb{E}^{\overline{\Q}} [\Phi]$
is an easy consequence of the monotone convergence theorem.
When $\Phi$ is bounded from above, by Lemma \ref{lemma:super_hedging}, it is enough to prove that there is some $\overline{H} \in \overline{\Hc}$ such that
\begin{eqnarray} \label{eq:superheding_Ec}
\overline{\Ec}^0[\Phi]
~+~
(\overline{H} \circ S)_N
~\ge~
\Phi
~~~\overline{\Pc}\mbox{-q.s.}
\end{eqnarray}
In view of Lemma \ref{lemm:red_Mcb_varphi}, we know $\overline{\Ec}^k(\Phi)(\overline{\om}) > -\infty$ for all $\overline{\om} \in \overline{\Om}_k$.
Further, by Lemma \ref{lemm:selectH},
there exist two universally measurable functions
$(y^k_1, y^k_2): \overline{\Om}_k \to \mathbb{R}^d \times \mathbb{R}^d$ such that
\begin{eqnarray*}
&&
y^k_1(\overline{\om}) \Delta S_{k+1}(\omega, \cdot) \ge \overline{\Ec}^{k+1}(\Phi)(\omega, \cdot, \theta) - \overline{\Ec}^k(\Phi)(\overline{\om})
~~{\cal P}_k(\omega) \mbox{-q.s.} \\
&&
y^k_2(\overline{\om}) \Delta S_{k+1}(\omega, \cdot) \ge \overline{\Ec}^{k+1}(\Phi)(\omega, \cdot, k+1) - \overline{\Ec}^k(\Phi)(\overline{\om})
~~{\cal P}_k(\omega) \mbox{-q.s.}
\end{eqnarray*}
for all $\overline{\om} = (\omega, \theta) \in \overline{\Om}_k$ such that $\mathrm{NA}({\cal P}_k(\omega))$ holds.
Since $N_k := \{\omega_k : \mathrm{NA}({\cal P}_k(\omega)) ~\mbox{fails}\}$ is ${\cal P}$-polar by Theorem 4.5 of \cite{BN13},
it follows that, with $\overline{H}_{k+1}(\overline{\om}) := y^k_1([\overline{\om}]_k) \mathbf{1}_{\{\theta \le k\}} + y^k_1([\overline{\om}]_k) \mathbf{1}_{\{\theta > k\}}$, one has
$$
\sum_{k=0}^{N-1} \overline{H}_{k+1} \Delta S_{k+1}
~\ge~
\sum_{k=0}^{N-1} \big(\overline{\Ec}^{k+1}(\Phi) - \overline{\Ec}^k(\Phi) \big)
~=~
\Phi - \overline{\Ec}(\Phi),
~~~{\cal P} \mbox{-q.s.}
$$
To conclude, it is enough to notice that the above $\overline{H}$ is an optimal dual strategy for the case $\Phi$ being bounded from above.
The existence of the optimal dual strategy for general $\Phi$ is then a consequence of Lemma \ref{lemm:closedness}.
\qed
\subsection{Proof of Theorem \ref{theo:main}: the case $e \ge 1$, equivalently $\Lambda\neq\emptyset$}
We will adapt the arguments in Section 5 of \cite{BN13} to prove Theorem \ref{theo:main} in the context with finitely many options $e \ge 1$.
For technical reasons, we introduce
$$
\varphi(\omega, \theta) := 1 + |g^1(\omega)| + \cdots + |g^e(\omega)| + \max_{1 \le k \le N} |\Phi_k(\omega)|,
$$
which depends only on $\omega$, and
\begin{eqnarray}
\overline{\Mc}^{\varphi}_g
:=
\{
\overline{\Q} \in \overline{\Mc}_0 ~: \mathbb{E}^{\overline{\Q}}[\varphi] < \infty ~\mbox{and}~
\mathbb{E}^{\overline{\Q}}[g^i] = 0
~\mbox{for}~ i=1, \cdots, e
\}.
\end{eqnarray}
Moreover, in view of Lemma \ref{lemm:red_Mcb_varphi}, one has
$$\sup_{\overline{\Q} \in \overline{\Mc}_g} \mathbb{E}^{\overline{\Q}}[ \Phi] = \sup_{\overline{\Q} \in \overline{\Mc}_g^{\varphi}} \mathbb{E}^{\overline{\Q}}[ \Phi].$$
\noindent {\bf Proof of Theorem \ref{theo:main} (the case $e \ge 1$)}.
The existence of some $\overline{\Q} \in \overline{\Mc}_g$ is an easy consequence of Theorem 5.1 of \cite{BN13} under $\mathrm{NA}({\cal P})$.
Moreover, similarly to \cite{BN13}, there exists an optimal dual strategies by Lemma \ref{lemm:closedness}.
We will then focus on the duality results.
First, the duality ${\overline{\pi}^E_g(\Phi)} = \sup_{\overline{\Q} \in \overline{\Mc}_g} \mathbb{E}^{\overline{\Q}}[\Phi]$ in \eqref{eq:main} has already been proved for the case $e=0$,
we will use the induction arguments:
Suppose that the duality \eqref{eq:main} holds true for the case with $e \geq 0$,
We aim to prove the duality with $e+1$ options:
$$
\overline{\pi}^E_{(g,f)} (\Phi)
~=~
\sup_{\overline{\Q} \in \overline{\Mc}_{(g,f)}^{\varphi}} \mathbb{E}^{\overline{\Q}}[\Phi],
$$
where the additional option has a Borel--measurable payoff function $f \equiv g^{e+1}$
such that $|f| \leq \varphi$, and has an initial price $f_0 = 0$.
By the weak duality in \eqref{eq:weak_duality} and Lemma \ref{lemm:red_Mcb_varphi}, the ``$\ge$'' side of the inequality holds true,
we will focus on the ``$\le$'' side of the inequality:
\begin{eqnarray} \label{eq:ineq_e_1}
\overline{\pi}^E_{(g,f)} (\Phi)
~\le~
\sup_{\overline{\Q} \in \overline{\Mc}_{(f,g)}^{\varphi}} \mathbb{E}^{\overline{\Q}}[\Phi].
\end{eqnarray}
If $f$ is replicable by some semi--static strategy with underlying $S$ and options $(g^1, \cdots, g^e)$ in sense that $\exists \overline{H} \in \overline{\Hc}, h \in \mathbb{R}^e$, s.t. $f=(\overline{H} \circ S)_N + hg, \overline{\Pc}\mbox{-q.s.}$
(or equivalently, $\exists H \in {\cal H}, h \in \mathbb{R}^e$, s.t. $f=(H \circ S)_N + hg, {\cal P}\mbox{-q.s.}$),
then the problem is reduced to the case with $e$ options and the result is trivial.
Let us assume that $f$ is not replicable, and we claim that there exists a sequence $(\overline{\Q}_n)_{n \ge 1} \subset \overline{\Mc}_g^{\varphi}$ such that
\begin{eqnarray} \label{eq:claim_e_1}
\mathbb{E}^{\overline{\Q}_n} [f] ~\longrightarrow~ f_0
~~~\mbox{and}~~
\mathbb{E}^{\overline{\Q}_n} [\Phi] ~\longrightarrow~ \overline{\pi}^E_{(g,f)} (\Phi),
~~~\mbox{as}~~
n \longrightarrow \infty.
\end{eqnarray}
Next, denote by $\overline{\pi}^E_g(f)$ the minimum superhedging cost of European option $f$ using $S$ and $(g^1, \cdots, g^e)$, i.e.
$$\pi^E_g(f)=\overline{\pi}^E_g(f) = \mbox{inf}~ \{ x: \exists \overline{H} \in \overline{\Hc}, h \in \mathbb{R}^e, \mbox{s.t.}~ x+(\overline{H} \circ S)_N + hg \ge f, ~~\overline{\Pc}\mbox{-q.s.} \}.$$
Since $f$ is not replicable, by the second fundamental theorem in Theorem 5.1.(c) of \cite{BN13}, we have that $\overline{\Q} \mapsto \mathbb{E}^{\overline{\Q}}[f]$ is not constant on $\overline{\Mc}_g^{\varphi}$.
Then, under the no--arbitrage condition, one has $0 = f_0 < \overline{\pi}^E_g(f)$.
It follows that
$0 = f_0 < \overline{\pi}^E_g(f) = \sup_{\overline{\Q} \in \overline{\Mc}_g^{\varphi}} \mathbb{E}^{\overline{\Q}} [f(]$.
Thus there exists some $\overline{\Q}_+ \in \overline{\Mc}_g^{\varphi}$, s.t. $0 < \mathbb{E}^{\overline{\Q}_+}[f] < \overline{\pi}^E_g(f)$.
With the same argument on $-f$, we can find another $\overline{\Q}_- \in \overline{\Mc}_g^{\varphi}$ such that
$$
-\overline{\pi}^E_g(-f) < \mathbb{E}^{\overline{\Q}_-} [f] < f_0 < \mathbb{E}^{\overline{\Q}_+} [f] < \overline{\pi}^E_g(f)
$$
Then one can choose an appropriate sequence of weight
$(\lambda_-^n, \lambda_0^n, \lambda_+^n) \in \mathbb{R}_+^3$, such that $\lambda_-^n + \lambda^n_0 + \lambda_+^n = 1$,
$\lambda_{\pm}^n \rightarrow 0$
and
$$
\overline{\Q}_n':= \lambda_-^n \overline{\Q}_- + \lambda^n_0 \overline{\Q}_n + \lambda_+^n \overline{\Q}_+ \in \overline{\Mc}_g,
~~\mbox{and}~
\mathbb{E}^{\overline{\Q}_n'} [f] = f_0 = 0,
$$
i.e. $\overline{\Q}_n' \in \overline{\Mc}_{(g,f)}^{\varphi}$.
Moreover, since $\lambda_{\pm}^n \rightarrow 0$, it follows that
$\mathbb{E}^{\overline{\Q}_n'}[\Phi] \to \overline{\pi}^E_{(g,f)}(\Phi)$
and we hence have the inequality \eqref{eq:ineq_e_1}.
It is enough to prove the claim \eqref{eq:claim_e_1},
for which we suppose without loss of generality that $\overline{\pi}^E_{(g,f)}(\Phi) = 0$.
Assume that \eqref{eq:claim_e_1} fails, then one has
$$
0 ~\notin~ \overline{ \{ \mathbb{E}^{\overline{\Q}} [ (f,\Phi) ]~: \overline{\Q} \in \overline{\Mc}_g^{\varphi} \} }
~\subseteq~
\mathbb{R}^2.
$$
By the convexity of the above set and the separation argument,
there exists $(y,z) \in \mathbb{R}^2$ with $|(y,z)|=1$, such that
\begin{eqnarray}\label{eq:contradition_e_1}
0~>~\sup_{\overline{\Q} \in \overline{\Mc}_g^{\varphi}} \mathbb{E}^{\overline{\Q}} [ yf+z \Phi]~=~\overline{\pi}^E_g(yf + z \Phi)~\ge~\overline{\pi}^E_{(g,f)}(z \Phi).
\end{eqnarray}
The strict inequality $\overline{\pi}^E_{(g,f)}(z \Phi) < 0$ implies that $z \neq 0$.
Now, if $z>0$, we then have $\overline{\pi}^E_{(g,f)} (\Phi) < 0$, which contradicts $\overline{\pi}^E_{(f,g)} (\Phi) =0$.
If $z<0$, then by \eqref{eq:contradition_e_1}, one has $0 > \mathbb{E}^{\overline{\Q}'} [ yf+z \Phi] = \mathbb{E}^{\overline{\Q}'} [ z \Phi ]$ for some $\overline{\Q}' \in \overline{\Mc}_{(g,f)} \subseteq \overline{\Mc}_g$ since $\overline{\Mc}_{(f,g)}$ is nonempty under the $\mathrm{NA}(\overline{\Pc})$ assumption in the case of $e+1$ options.
Then in the case $z<0$, one has $\mathbb{E}^{\overline{\Q}'} [ \Phi ] > 0 = \overline{\pi}^E_{(f,g)}(\Phi)$,
which contradicts the weak duality result \eqref{eq:weak_duality},
and we hence conclude the proof of the duality.
\qed
\section{Proofs for Section \ref{sec:mot}}
\label{s:proofs3}
A first idea how to prove Theorem \ref{theo:MOT} could be the following two steps argument as in \cite{GTT2}.
Firstly, under the condition that $\Phi$ is bounded from above and upper semicontinuous, one could prove that
$$\mu \in \mathfrak{P}((\mathbb{R}^d)^{M}) \mapsto \sup_{\overline{\Q} \in \overline{\Mc}_\mu}\mathbb{E}^{\overline{\Q}} \big[\Phi \big] \in \mathbb{R}$$
is concave and upper semicontinuous,
where we equip $\mathfrak{P}((\mathbb{R}^d)^{M})$ with a Wasserstein kind topology.
Secondly, using Fenchel--Moreau theorem, it follows that
\begin{eqnarray} \label{eq:duality_PD0}
\sup_{\overline{\Q} \in \overline{\Mc}_\mu}\mathbb{E}^{\overline{\Q}} \big[\Phi \big]
~=~
\overline{\pi}^E_{\mu,0}(\Phi)
~:=~
\inf_{\lambda \in \Lambda} \Big\{\mu(\lambda)
+ \sup_{\overline{\Q} \in \overline{\Mc}_0} \mathbb{E}^{\overline{\Q}} \big[ \Phi - \lambda \big]
\Big\}.
\end{eqnarray}
Solving the maximization problem \eqref{eq:duality_PD0}, by using Theorem \ref{theo:main}, concludes the proof of Theorem \ref{theo:MOT}.
However in the following, we will provide another proof, which is based on an approximation argument.
For simplicity, we suppose that $\mathbb{T}_0 = \{N\}$, where the same arguments work for more general $\mathbb{T}_0$.
In preparation, let us provide a technical lemma.
In the context of the martingale optimal transport problem,
we introduce a sequence of payoff functions $(g^i)_{i \ge 1}$ by
$$
g^i(\omega) := f^i(\omega_N) - c^i
~~~\mbox{with}~~
c^i:= \int_{\mathbb{R}^d} f^i(x) \mu(dx),
$$
where $f^i : \mathbb{R}^d \to \mathbb{R}$ is Lipschitz and $(f^i)_{i \ge 1}$ is dense in the space of all Lipschitz functions on $\mathbb{R}^d$ under the uniform convergence topology,
and moreover, it contains all functions in form $(x_j -K_n)^+$, $(-K_n - x_j)^+$ for $j=1, \cdots, d$ and $n \ge 1$, where $(K_n)_{n \ge 1} \subset \mathbb{R}$ is a sequence such that $K_n \to \infty$.
Notice that $\mu$ has finite first order and hence $c^i$ are all finite constants.
Next, let us introduce an approximate dual problem by
\begin{eqnarray*}
\pi^A_{\mu,m}(\Phi)
&:=&
\inf \Big\{
x ~: \exists(\overline{H}, h) \in \overline{\Hc} \times \mathbb{R}^m ~\mbox{s.t.}~
\mbox{for all}~ k \in \mathbb{T}, ~\omega \in \Omega,\\
&&~~~~~~~~~~~~
x + \sum_{i=1}^m h^i g^i(\omega_N) + (\overline{H}^k \circ S)_N(\omega) \ge \Phi_k(\omega)
\Big\}.
\end{eqnarray*}
Similarly,
$$
\overline{\Mc}_{\mu,m} ~:=~
\big\{
\overline{\Q} \in \overline{\Mc} ~: \mathbb{E}^{\overline{\Q}}[g^i] = 0 ~~\mbox{for}~i = 1, \cdots, m
\big\},
$$
and
$$
P_{\mu,m} ~:=~ \sup_{\overline{\Q} \in \overline{\Mc}_{\mu,m}} \mathbb{E}^{\overline{\Q}} \big[ \Phi \big].
$$
\begin{Lemma}\label{lemm:compact1}
Let $(\overline{\Q}_m)_{m \geq 1} \subset \overline{\Mc}$ be a sequence of martingale measures
such that $\overline{\Q}_m \in \overline{\Mc}_{\mu,m}$ for each $m \ge 1$.
Then,
\noindent {\rm (a)$\>\>$} $(\overline{\Q}_m)_{m \geq 1}$ is relatively compact under the weak convergence topology.
\noindent {\rm (b)$\>\>$} The sequence $(S_N, \overline{\Q}_m)_{m \ge 1}$ is uniformly integrable,
and any accumulation point of $(\overline{\Q}_m)_{m \geq 1}$ belongs to ${\cal M}_\mu$.
\end{Lemma}
\proof {\rm (a)$\>\>$} Without loss of generality, we assume that $f_1(x) = \sum_{i=1}^d |x_i|$ so that
$$\sup_{m \ge 1} \mathbb{E}^{\overline{\Q}_m} \big[ \sum_{i=1}^d |S_N^i| \big] < \int_{\mathbb{R}^d} \sum_{i=1}^d|x_i| \mu(dx) < \infty.$$
Let us first prove the relative compactness of $(\overline{\Q}_m)_{m \geq 1}$.
By Prokhorov theorem, it is enough to find, for every $\varepsilon > 0$,
a compact set $D_{\varepsilon} \subset \mathbb{R}^d$ such that $\overline{\Q}_m[ S_k \notin D_{\varepsilon} ] \le \varepsilon$ for all $k = 1, \cdots, N$.
It is then enough to find, for every $\varepsilon > 0$, a constant $K_{\varepsilon} > 0$ such that
$\overline{\Q}_m \big[ |S^i_k| \ge K_{\varepsilon} \big] \le \varepsilon$ for all $i = 1, \cdots, d$ and $k=1, \cdots, N$.
Next, by the martingale property, one has $\mathbb{E}^{\overline{\Q}_m} [|S_k^i|] \le \mathbb{E}^{\overline{\Q}_m}[|S_N^i|]$.
Then for every $\varepsilon>0$, one can choose $K_{\varepsilon} > 0$ such that $\sup_{m \ge 1} \mathbb{E}^{\overline{\Q}_m} \big[ \sum_{i=1}^d |S_N^i| \big] \le K_{\varepsilon} \varepsilon.$
It follows that $\overline{\Q}_m \big[ |S^i_k| \ge K_{\varepsilon} \big] \le \frac{\mathbb{E}^{\overline{\Q}_m}[|S^i_k|]}{K_{\varepsilon}} \le \varepsilon$,
and hence $(\overline{\Q}_m)_{m \ge 1}$ is relatively compact.
\vspace{1mm}
\noindent {\rm (b)$\>\>$} To see that the sequence $(S_N, \overline{\Q}_m)_{m \ge 1}$ is uniformly integrable,
it is enough to notice that $|x_i| \mathbf{1}_{|x_i| \ge 2K_n} \le 2 (|x_i| - K_n) \mathbf{1}_{|x_i| \ge K_n}$, where the latter is a payoff function contained in the sequence $(f_k)_{k \ge 1}$.
\vspace{1mm}
\noindent {\rm (c)$\>\>$} Let $\overline{\Q}_0$ be an accumulation point of $(\overline{\Q}_m)_{m \geq 1}$.
Since the sequence $(f_k)_{k \ge 1}$ is supposed to be dense in the space of all Lipschtiz functions on $\mathbb{R}^d$ under the uniformly convergence topology,
it is easy to obtain that $\overline{\Q}_0 \circ S_N^{-1} = \mu$.
\vspace{1mm}
\noindent {\rm (d)$\>\>$} To conclude the proof,
it is enough to show that the martingale property is preserved for the limiting measure $\overline{\Q}_0$.
By abstracting a subsequence, we assume that $\overline{\Q}_m \to \overline{\Q}_0$ weakly,
and we will prove that for all $1 \leq k_1 < k_2 \leq N$,
for any bounded continuous function $\varphi: (\mathbb{R}^d)^{k_1} \times \mathbb{T} \rightarrow \mathbb{R}$,
one has
\begin{equation} \label{eq:martinglim}
\mathbb{E}^{\overline{\Q}_0} \big[ \varphi \big(S_1, \cdots, S_{k_1}, T \wedge (k_1+1) \big) (S_{k_2}-S_{k_1}) \big]
~=~
0.
\end{equation}
Let $K >0$, and $\chi_K: \mathbb{R}^d \to \mathbb{R}^d$ a continuous function uniformly bounded by $K$ satisfying
$\chi_K(x) = x$ when $\|x\| \le K$, and $\chi_K(x) = 0$ when $\|x \| \ge K+1$.
Then for every $m =0$ or $m \ge 1$, one has
\begin{eqnarray} \label{eq:mart_ineq_1}
\big| \mathbb{E}^{\overline{\Q}_m} \big[ \varphi(S,T) (S_{k_2}-S_{k_1}) \big] \big|
\!\!&\le &\!\!
\big| \mathbb{E}^{\overline{\Q}_m} \big[ \varphi(S,T) \big(\chi_K(S_{k_2}) - \chi_K(S_{k_1}) \big) \big] \big| \nonumber\\
\!\!&&\!\!\!\!\!\!+
|\varphi|_{\infty}
\mathbb{E}^{\overline{\Q}_m} \big[ |S_{k_2}|\mathbf{1}_{|S_{k_2}| \geq K}+|S_{k_1}|\mathbf{1}_{|S_{k_1}| \geq K} \big],
~~~~~
\end{eqnarray}
where we simplify $\varphi(S_1, \cdots, S_{k_1}, T \wedge (k_1+1))$ to $\varphi(S, T)$.
For every $\varepsilon > 0$, by uniformly integrability of $(S_N, \overline{\Q}_m)_{m \ge 1}$, there is $K_{\varepsilon} > 0$
such that
\begin{eqnarray} \label{eq:mart_ineq_2}
|\varphi|_{\infty}
\mathbb{E}^{\overline{\Q}_m}
\big[ |S_{k_2}|\mathbf{1}_{|S_{k_2}| \geq K_{\varepsilon}}
+|S_{k_1}|\mathbf{1}_{|S_{k_1}| \geq K_{\varepsilon}}
\big]
~\le~
\varepsilon,
~~~\mbox{for all}~m=0, 1, \cdots
\end{eqnarray}
Moreover, for $m \ge 1$, $\overline{\Q}_m$ is a martingale measure,
then $\mathbb{E}^{\overline{\Q}_m} \big[ \varphi(S,T) (S_{k_2}-S_{k_1}) \big] = 0$ and hence
$\big| \mathbb{E}^{\overline{\Q}_m} \big[ \varphi(S,T) \big(\chi_K(S_{k_2}) - \chi_K(S_{k_1}) \big) \big] \big| \le \varepsilon$.
Then by taking the limit $m \rightarrow \infty$, it follows that
\begin{eqnarray} \label{eq:mart_ineq_3}
\big| \mathbb{E}^{\overline{\Q}_0} \big[ \varphi(S,T) \big(\chi_K(S_{k_2}) - \chi_K(S_{k_1}) \big) \big] \big|
~\le~
\varepsilon.
\end{eqnarray}
Combining \eqref{eq:mart_ineq_1}, \eqref{eq:mart_ineq_2} and \eqref{eq:mart_ineq_3}, and by the arbitrariness of $\varepsilon > 0$,
it follows that \eqref{eq:martinglim} holds true and we hence conclude the proof.
\qed
\vspace{3mm}
\noindent {\bf Proof of Theorem \ref{theo:MOT}}.
We notice that by Theorem \ref{theo:main},
$$\sup_{\overline{\Q} \in \overline{\Mc}_{\mu,m}} \mathbb{E}^{\overline{\Q}} \big[ \Phi \big] = \pi^A_{\mu,m}(\Phi) \ge \pi^A_\mu(\Phi).$$
Let $(\overline{\Q}_m)_{m \ge 1}$ be a sequence of probability measures such that $\overline{\Q}_m \in \overline{\Mc}_{\mu,m}$ for each $m \ge 1$ and
$$
\limsup_{m \to \infty} \mathbb{E}^{\overline{\Q}_m} \big[ \Phi_T(S) \big]
~=~
\limsup_{m \to \infty} \sup_{\overline{\Q} \in \overline{\Mc}_{\mu,m}} \mathbb{E}^{\overline{\Q}} \big[ \Phi \big].
$$
It follows by Lemma \ref{lemm:compact1} that
there is some $\overline{\Q}_0 \in \overline{\Mc}_\mu$ and a subsequence $\overline{\Q}_{m_k} \to \overline{\Q}_0$
under the weak convergence topology.
Using upper semi--continuity of $\Phi$ and by Fatou's lemma, it follows that $\mathbb{E}^{\overline{\Q}_0} \big[ \Phi_T(S) \big] \ge \limsup_{m \to \infty} \mathbb{E}^{\overline{\Q}_m} \big[ \Phi_T(S) \big]$.
It leads to the inequality
$$
\sup_{\overline{\Q} \in \overline{\Mc}_\mu}\mathbb{E}^{\overline{\Q}} \big[\Phi \big]
~\ge~
\mathbb{E}^{\overline{\Q}_0} \big[ \Phi \big]
~\ge~
\limsup_{m \to \infty} \sup_{\overline{\Q} \in \overline{\Mc}_{\mu,m}} \mathbb{E}^{\overline{\Q}} \big[ \Phi \big]
~=~
\limsup_{m \to \infty} \pi^A_{\mu,m}(\Phi)
~\ge~
\pi^A_\mu(\Phi)
$$
and we hence conclude the proof by the weak duality \eqref{eq:weak_duality}.
\qed
\bibliographystyle{apalike}
|
\section*{Introduction}
On the Cartesian plane there are two methods to define an ordered plane on which the Pasch axiom is false. If we leave out a point from an Euclidean plane we can give such a plane as a simple example. Using this method we automatically omit some further axioms from the axiom system of the Euclidean plane. A line through the missing point cannot fulfil the first and third axioms of congruency (namely, that every segment can be "laid off" on a given side of a given point on a given line in at least one way, and the axiom of the addition of congruent segments), respectively. In addition, the Cantor axiom (and consequently the Dedekind axiom) is also false on this plane.
Another method has been shown by Szczerba (see \cite{szczerba} and \cite{adler}) who defined the property of "betweenness" based on an exotic ordering of the points on a line. Applying this method on the Cartesian plane we can define a plane in which -- except the Pasch axiom -- all axioms of the Euclidean plane hold. We describe this method briefly. Let $f:\mathbb{R}\rightarrow \mathbb{R}$ be an additive onto function and $0<f(1)$. Let $x<^\star y$ if $f(x)<f(y)$. Then $(\mathbb{R},+,\cdot,<^\star)$ is a semi-ordered field, means that it is an ordered additive group but it is not hold necessarily that if $0<^\star x$ and $0<^\star y$ then $0<^\star xy$. Szczerba proved that the Cartesian plane over $(\mathbb{R},+,\cdot,<^\star)$ satisfies the axioms of plane geometry (with the full second-order continuity axiom) except the Pasch axiom. Moreover if the ordering is not the usual one (based on an $\mathbb{R}$-linear mapping $f$) then the corresponding Cartesian plane does not satisfy the Pasch axiom. To define such a non-usual ordering we have to define an additive non-linear function. Szczerba's method based on the Hammel basis of the reals over the rationals and hence it uses the axiom of choice. Addler certified in \cite{adler} that the axiom of choice plays an essential role in Szczerba's proof, namely all models satisfying the axioms of the Euclidean plane except the Pasch axiom are isomorphic to the model constructed by Szczerba. Addler noted also that the construction of additive non-linear functions implies the existence of non-measurable sets in the sense of Lebesgue (such a function is not Lebesgue measurable). Hence if we change the axiom of choice to the axiom of determinacy then all additive functions are linear ones, which means that all models satisfying the remaining axioms of the plane geometry fulfil the Pasch axiom too.
Our purpose is to define a non-Paschian plane based on such well-defined property of betweenness which cannot be derived from an ordering of the points lying on a line. In the new model it is not possible to define the congruence of segments but we can define angle, triangle and angle measure, respectively. With respect to our definitions the plane has an elliptic character, meaning that the sum of the angles of a triangle is greater than $\pi$. It is interesting that the continuity axioms hold as well. Finally, contrary to the plane of Szczerba our construction is independent from the axiom of choice, namely it remains a non-Paschian plane even if we substitute the axiom of choice with the axiom of determinateness.
There are several books on the foundations of geometry. We mention here some of them as references for the interested reader. We propose the books \cite{bonola}, \cite{coxeter}, \cite{hilbert}, \cite{berger}, \cite{gho} to investigate non-Euclidean geometries in general and the papers \cite{gho 1}, \cite{gho 2}, \cite{gho 3} for a deeper investigation of the hyperbolic plane.
\section*{A "betweenness"-based ordered plane}
In this paper the concept of \emph{plane} has a double meaning, on one hand, it is a so-called \emph{set of points} and on the other hand, it is a set of \emph{lines}, thus we denote by $\mathcal{P}$ and $\mathcal{L}$, respectively. We would like to allow on the plane the binary notion of \emph{incidence} to points and lines; and the ternary notion of \emph{betweenness} defined on the set of points which are lying on the same line. We use Hilbert's axiom system (see \cite{hilbert}) which is the most popular and used one in the fundamentals of geometry.
Let us denote the binary relation of {\em incidence} of the point $A$ and the line $a$ by $AIa$ where $A\in \mathcal{P}$ and $a\in \mathcal{L}$. The axioms of incidence are formulated as follows:
\begin{enumerate}
\item[I1.] For any two points $A$ and $B$ there exists a line $a$ which fulfil $AIa$ and $BIa$.
\item[I2.] For any two distinct points there exists only one line that contains both; consequently, if $AIa$ and $BIa$ and also if $AIb$ and $BIb$ are fulfilled then $a\equiv b$. (Hence the line $a$ determined uniquely by $A$ and $B$, and it can be denoted by $AB$.)
\item[I3.] There exist at least two points on a line; for all $a\in \mathcal{L}$ there are $A,B\in \mathcal{P}$ such that $AIa$ and $BIb$. There exist at least three points which do not lie on the same line.
\end{enumerate}
For three points $A,B,C$ the relation that the point \emph{$B$ is between the points $A$ and $C$} we denote by $(ABC)$.
The axioms of betweenness are
\begin{enumerate}
\item[B1.] If $(ABC)$, then $(CBA)$, and there is a line $e$ which contains the points $A,B,C$.
\item[B2.] If $A$ and $C$ are two distinct points, there is at least one point $B\in AC$ for which $(ACB)$.
\item[B3.] Of any three points on a line there exists only one which lies between the other two.
\item[B4.](Pasch's Axiom): Let $A$, $B$, $C$ be three distinct points not collinear and let $a$ be a line not passing through any of the points $A$, $B$, $C$. If there exists a point $D$ lying on the line $a$ for which $(ADB)$ then either there is a point $E$ such that $EIa$ and $(BEC)$ or there exists a point $F$ such that $FIa$ and $(AFC)$.
\end{enumerate}
We note that Hilbert's axiom system has a weaker form, where the existence part of $\mathrm{B}2$ ("there is at least one") is proven by a theorem. As we can see from these axioms (using Tarski's apellations of \cite{tarski}) we have a Lower 2-dimensional axiom ($\mathrm{I}3$), but we do not use Upper 2-dimensional axiom. In fact, the Pasch's Axiom ($\mathrm{B}4$) in the present form restrict the value of the usual concept of dimension to $n=2$ and so a natural definition arise for ordered plane of dimension $2$ if we assume the above seven axioms. If we would like to consider the dimension open then we have to modify the strong axiom $\mathrm{B}4$ on such a way that it will be effect only on "2-dimensional configurations". This motivates the following modification.
\begin{enumerate}
\item[$\mathrm{B}4^\star$.]: Let $A$, $B$, $C$ be three distinct points not collinear and let $a$ be a line not passing through any of the points $A$, $B$, $C$. If there exists a point $D$ lying on the line $a$ for which $(ADB)$ and $a$ intersects the remaining two lines $(BC)$ and $(AC)$ in the points $E$ and $F$, respectively then either $(BEC)$ or $(AFC)$.
\end{enumerate}
Clearly, $\mathrm{B}4^\star$ is a weakening of the requirement of $\mathrm{B}4$. For example, we do not investigate its validity on lines which intersects only two lines from the fixed three ones. Hence we exclude those counterexamples on Pasch's axiom in what a possible point of intersection has been dropped from the plane. We also disregard from those lines which in a case of the extension of the axiom system to the axiom system of a higher dimensional space cannot be considered as a line of the plane of the points $A,B,C$.
On the base of this weakening it is possible that we cannot fit together the betweenness property of three lines, because there is no a fourth one which intersects all of them in distinct points. (The original strong form of Pasch's axiom excludes this possibility.) Hence we introduce a fourth axiom of incidence which is in the case when the original Pasch's axiom holds is a consequence of the axioms.
\begin{enumerate}
\item[I4.]: For any three pairwise intersecting lines $a,b,c$ there is a fourth line $d$ and three distinct points $A,B,C$ such that the respective incidences $AId$, $AIa$, $BId$, $BIb$ and $CId$, $CIc$ are hold.
\end{enumerate}
Our definition of ordered plane is:
\begin{defi}\label{def:orderedplane} The plane is \emph{ordered} if it satisfies the axioms $\mathrm{I}1-\mathrm{I}4$, $\mathrm{B}1-\mathrm{B}3$ and $\mathrm{B}4^\star$, respectively. The plane is \emph{non-Paschian ordered plane} if it is an ordered plane without $\mathrm{B}4^\star$.
\end{defi}
\begin{remark}
In the paper \cite{kreuzer} we can find the characterization of the so-called halfordered planes. Among these types one gives a non-trivial example for non-Paschian ordered plane. In fact, the halfordered plane with the property "from three distinct collinear points are exactly two between the others" characterized as an affine plane with order $5$ with the following betweenness property: If $A,B$ two points on a line and the other points are in the form $C=A+\lambda(B-A)$ where $\lambda\in \mathbb{Z}_5$ then $(BAC)$ holds if and only if $\lambda=4$ (a square in $\mathbb{Z}_5$). If one changes $(ABC)$ with $\neg (ABC)$ one gets an ordered plane with $\neg\mathrm{B}4$. But this is a finite model without any nice metric property as e.g. the existence of the angle of lines with a corresponding angular measure. The purpose of this paper to give another model (with interesting geometric properties) such that which does not arise from any usual affine structure.
\end{remark}
Note that in a classical $2$-dimensional absolute plane (holding axioms $\mathrm{I}1-\mathrm{I}3$, $\mathrm{B}1-\mathrm{B}4$) the concept of segment can be defined, as the set of points between the two end-point of the segment, and the following five statements are valid:
\begin{itemize}
\item[T1.] Every segment has at least one point. (Which is according to our definition not an endpoint of the segment.)
\item[T2.] The so-called degenerated case of the Pasch axiom holds, which says that if a line intersects a segment which is one of the three segments determined by three collinear points, then it intersects at least one of the other two segments.
\item[T3.] For four collinear points $A,B,C,D$ for which $(ABC)$ and $(ACD)$ hold, we also have $(ABD)$. (With other words all points of a segment lying on a segment which contains the endpoints of the first one). Consequently each segment and each line has infinitely many points.
\item[T4.] In a line, a point $O$ determines two (closed) half-lines with the common origin $O$ which is the only common point of them.
\item[T5.] A line $l$ determines two (closed) half-planes with the intersection line $l$ which is the common boundary of the two half-planes.
\end{itemize}
Given a line $l$ and a point $O$ on it. The points $A$ and $B$ on $l$ are locate on \emph{opposite sides} of $O$ if $(AOB)$. The points $A$ and $B$ are on the \emph{same side} of $O$ if $(OAB)$ or $(OBA)$. An (open) \emph{half-line with the origin $O$} contains all the points which are on the same side of $O$. [T4] says that the property "same-side" is an equivalence relation with two classes. Analogously we can define the (open) half-plane as a class of the equivalence relation based on the concepts of "opposite sides" and "same side" of $l$. [T5] requires that for a line $l$ the corresponding equivalence relation has exactly two equivalence classes.
The model of our non-Paschian ordered plane is based on the known model of a finite line, which is the so-called \emph{line-model of five-points}. In this model a betweenness relation is defined satisfying the axioms $\mathrm{B}1,\mathrm{B}2,\mathrm{B}3$. (Thus the model shows that the axioms $\mathrm{B}1,\mathrm{B}2,\mathrm{B}3$ do not imply that a line has infinitely many points (see in \cite{moussong} or \cite{gho}). Using this model we can construct a non-Paschian ordered plane having infinitely many lines with five points and five lines with infinitely many points.
Let $A,B,C,D$ and $E$ be the vertices of a regular pentagon. Clearly, all triangles determined by three vertices of a regular pentagon are isosceles triangles but not equilateral ones. Hence in each triangle we can chose one of the three vertices (in a natural way) which is in between the other two points: the common vertex of the two legs of the triangle. (So if the legs are $AB$ and $BC$ we define the betweenness such that $B$ is between the points $A$ and $C$.)
The properties of a five-points line are the following:
\begin{itemize}
\item It is easy to see that such line fulfils the axioms $\mathrm{B}1,\mathrm{B}2$ and $\mathrm{B}3$.
\item Clearly, the closed segments of the line are the point sets with cardinality three. The two endpoints of a segment are the vertices on the base of the corresponding isosceles triangle. We have two types of segments. One of them called \emph{small segment}, this is formed by two consecutive edges (as legs) of the pentagon. The segment is a \emph{large segment}, if its base is an edge of the pentagon.
\item Theorem [T1] is true but [T2] and [T3] are false, respectively. In fact, from $(ABC)$ follows that $B$ is the only point of the segment $AC$, and from $(CAD)$ follows that the only point of the segment $CD$ is $A$. Since from $(ACD)$ follows that the only point of the segment $AD$ is $C$, thus a line (of an embedding plane) distinct from $ABCDE$ can intersect exactly one edge of the degenerated triangle $ACD$. Moreover $(ABC)$ and $(ACE)$ does not implies $(ABE)$ but also $(EAB)$ holds.
\item Consider the true relations $(ABC)$, $(ACE)$ and $(EAB)$. Assume that these relations arise from an ordering of the points $A,B$ and $C$, this ordering is either $A<B<C$ or $A>B>C$. If we assume the first possibility then $A<B<C$ implies $A<C$ and hence $A<E$ holds, too. Then the third betweenness relation implies that $A>B$ contradicting with our first assumption. The second possibility $A<B<C$ leads to the same contradiction, meaning that there is no ordering of the points which can imply the relation of betweenness.
\item The linear axiom of congruence holds for the equivalence relation on this line in which two large (or two small) segments are congruent to each other, respectively. We can also define the union of two segments. Two segments with one common endpoint have the union as the segment corresponding to the free ends of the given ones. The lengths of the segments can be prescribed as the corresponding elements of the finite field $GF(3)$. ($GF(3)$ is the only finite field with three elements which isomorphic to the ring of integers modulo $3$).
\item Observe that the continuity axiom of Archimedes, Cantor and Dedekind hold, respectively. In fact, Cantor's axiom is true because there is no two segments containing each other. Dedekind's axiom is true, because there are no two sets $\mathcal{A}$, $\mathcal{B}$ in the line forming a Dedekind cut. Finally, the axiom of Archimedes follows from the fact that every large segment is a subset of the union of two small ones; and every small segment is a subset of the union of two big ones.
\end{itemize}
\begin{figure}
\centering
\includegraphics{model}\\
\caption{Two non-intersecting lines}\label{model}
\end{figure}
Let now $a,b,c,d,e$ be five parallel lines orthogonal to the plane of the points $A,B,C,D,E$, and passing through the points $A,B,C,D,E$, respectively (see Fig.\ref{model}). Denote the center of the regular pentagon by $O$ and consider the 2-planes of the Euclidean space through this point. Let the set of points $\mathcal{P}$ be the collection of the points of the $3$-space which belong to the lines $a,b,c,d,e$, and the set of lines $\mathcal{L}$ is the union of $a,b,c,d,e$ with the five-points sets ${X_a,X_b,X_c,X_d,X_e}$ obtained by the intersection of the above mentioned $2$-planes and the lines $a,b,c,d,e$.
The concept of incidence is the same as the Euclidean incidence. It is clear that $\mathrm{I}1-\mathrm{I}3$ are fulfilled. Define the property of betweenness on the following way. If three points $X$,$Y$,$Z$ lie on one of the lines $a,b,c,d$ or $e$ then let us denote by $(XYZ)$ the fact that $Y$ lies between the points $X$ and $Z$ with respect to the Euclidean concept of betweenness. In the case if three points $X_j\in j$, $X_k\in k$ and $X_l\in l$ are such that $\{j,k,l\}\subset \{a,b,c,d,e\}$ - lying on the same plane - we denote by $(X_j,X_k,X_l)$ the fact that $(JKL)$ holds with respect to the concept of betweenness in the five-points model of the points $\{A,B,C,D,E\}$.
\begin{defi}\label{def:angle}
Two planes which intersect the model in the points of two lines $l_1,l_2\in\mathcal{L}$ divide the embedding space into four parts where the opposite parts are congruent to each other, respectively. The common line $m$ of the two planes contains at most one point of the model. In this case let us denote by $P=l_1\cap l_2\in a$ the common point of the lines. We can assign to these lines $l_1,l_2$ two \emph{angles}. They contain all points of the model which belong to the union of the opposite wedge of the space, determined by the given planes, respectively. The legs of these angles are the lines $l_1$ and $l_2$, the common vertex of the angles is $P$. The \emph{angle measures} of the domains are the Euclidean angle measures of dihedral angle. It implies, that the sum the two non-congruent angles is equal to $\pi$.
\end{defi}
\begin{remark} With respect to Def. \ref{def:angle} we have orthogonality in the model. For instance the lines $a$ and $ABCDE$ are orthogonal to each other. We have a natural concept of triangle, three pairwise intersecting lines $l_1$, $l_2$ and $l_3$ (which three have no point in common) determine three points $L_i=l_j\cap l_k$ where $i,j,k\in\{1,2,3\}$ distinct indices. The vertex set of the triangle is $\{L_1,L_2,L_3\}$, the open sides of the triangle are the three segments $\overline{L_iL_j}$, and the triangle domain is the intersection of three angles corresponding to the pairs of lines $l_i,l_j$, respectively (see Fig. \ref{triangle}). The domain of the open triangle depends on the choice of the open angle domains. For each pair of lines we have two possibilities to define it. In Fig. \ref{triangle} $l_1$,$l_2$ and $l_3$ contain the points $X_i$,$Y_i$ and $Z_i$, respectively. The vertices of the triangle are $P=X_a=Y_a$, $Q=Y_b=Z_b$ and $R=X_c=Z_c$, the closed sides are $\mathrm{cl}(\overline{PQ})=\{Y_a,Y_b,Y_d\}$, $\mathrm{cl}(\overline{QR})=\{Z_b,Z_c,Z_e\}$ and $\mathrm{cl}(\overline{PR})=\{X_a,X_c,X_b\}$, respectively. We choose the angle domain of $l_1,l_2$ with the inner points $\mathrm{int}(QPR\angle):=\overline{X_bY_b}\cup \overline{X_cY_c}\cup \overline{X_dY_d}\cup \overline{X_eY_e}$; for $l_2,l_3$ the angle domain $\mathrm{int}(PQR\angle):=b\setminus {Q}\cup (a\setminus \mathrm{cl} \{\overline{Y_aZ_a}\})\cup (c \setminus \mathrm{cl} \{\overline{Y_cZ_c}\})\cup (d \setminus \mathrm{cl} \{\overline{Y_dZ_d}\})\cup (e \setminus \mathrm{cl} \{\overline{Y_eZ_e}\}) $ and for $l_1,l_3$ let the angle domain be $\mathrm{int}(QRP\angle):=\overline{X_aZ_a}\cup \overline{X_bZ_b}\cup \overline{X_dZ_d}\cup \overline{X_eZ_e}$. The interior of the triangle contains the intersection of these angle domains hence $\mathrm{int}(PQR_\triangle)=\overline{Z_bX_b}=\overline{QX_b}$.
\end{remark}
\begin{figure}
\centering
\includegraphics{triangle}\\
\caption{Angles and triangles}\label{triangle}
\end{figure}
\begin{theorem}
This construction defines an elementary model of a non-Paschian ordered plane in the sense of Definition \ref{def:orderedplane}. The three axioms of continuity are true in the corresponding plane. In addition, the sum of the angle measures of the triangles are greater then $\pi$. Moreover there is no universal description of parallelism, there are non-intersecting lines, there is such a pair of points and lines for which the Euclidean axiom of parallels holds, and there is such a pair of points and lines for which the negation of the Euclidean axiom of parallels is true.
\end{theorem}
\begin{proof}
The axioms of incidence $\mathrm{I}1$, $\mathrm{I}2$ and $\mathrm{I}3$ are hold, because three non-collinear points of the Euclidean space uniquely determine a plane of the space. To prove $\mathrm{I}4$ observe that the lines $a,b,c,d,e$ intersect all line of five-points, respectively. This implies that three lines
of five-points can be intersected in three distinct points by at least two from the lines $a,b,c,d,e$. In the remaining cases, when among the examined lines we have one or more from the lines $a,b,c,d,e$ we also can guarantee easily common transversalis. In fact, if $p$ and $q$ are two lines of five-points and the third is $a$ then the line $r$ of five-points determined by $P_b=p\cap b$ and $Q_c=q\cap c$ intersects $a$ in a third point $R_a=r\cap a$. If $p$,$a$ and $b$ gives the examined triplet then arbitrary line of five-points through $p_C=p\cap c$ is usable as a transversalis and in the last case when our choice is the three lines $a$,$b$ and $c$ then we can consider any line of five-points to this purpose.
We saw that the axioms B1,B2 and B3 are true on a line of five-points and the remaining lines $a,b,c,d,e$ have Euclidean ordering. In order to prove that $\mathrm{B}4*$ is false we consider the non-collinear triplet $P,Q,R$ and the line $d$. Since the only point of the segment $\overline{PQ}$ is $Y_d\in d$, the line $d$ and its point $Y_d\in \overline{PQ}$ fulfil the assumption of the axiom $\mathrm{B}4^\star$. On the other hand the segments $\overline{QR}=\{Z_e\}\in e$ and $\overline{RP}= \{X_b\}\in b$ are disjoint to $d$ which is a counterexample to $\mathrm{B}4^\star$.
Observe that drawing a unit sphere around the point $O$ we can realize the angle measures of a triangle of the model as the spherical angles of a spherical triangle on the sphere. Hence the sum of the angles of a triangle of the model is equal to the sum of the angles of a spherical triangle which is greater than $\pi$.
Finally, the lines $a$ and $b$ are non-intersecting ones, for the line $a$ through the point $X_b$ we have only one line which does not intersect the line $a$, it is the line $b$. If $l$ is a line of five-points and a point $X$ does not lie on $l$ then there are only finitely many lines through $X$ intersecting $l$, hence the infinitely many other lines through $X$ do not intersect $l$.
\end{proof}
As a closing note we remark that the above-mentioned construction gives a non-Paschian ordered plane in that case, too, when the axiom of determinateness is used without the axiom of choice to construct the set theory (or to construct any other part of mathematics).
\section*{Acknowledgments}
The author is indebted to the editor for his help to make precise and sharp the axiomatic embeddedness of the model. Many thanks are due to the reviewer his valuable suggestions. He showed me the non-trivial example of Remark 1.
|
\section{Introduction}
One of the most astonishing and counterintuitive ideas in modern physics is that the vacuum is not empty. Quantization of the electromagnetic fields leads to a non-zero ground-state energy density \cite{rugh,zelda}, which can have physical manifestations. In 1948, Hendrik Casimir predicted that adjacent conducting surfaces can confine the spectrum of zero-point modes, leading to a mutual attraction \cite{casimir}. Since then there have been a number of theoretical and experimental developments confirming this prediction, as well as identifying new electromagnetic, Casimir phenomena in various systems \cite{Klimchitskaya2009,Bordag2001,romeo,lam1,lam2,Sushkov2011}, including the onset of repulsive long-ranged forces when the boundary condition for ideal conducting plates is relaxed \cite{lif1,lif2,capasso}.
While the original Casimir force is of purely quantum origin, there are analogous interactions that arise from the confinement of fluctuations in classical systems \cite{kardar,Balibar2005,Gambassi2009}. Perhaps the best known example of these is the critical Casimir force described by Fisher and de Gennes \cite{fish}. A binary liquid mixture close to the critical point experiences large-scale concentration fluctuations in the homogeneous phase. The fluctuations occur over a maximum correlation length which diverges as the system approaches the critical point. When objects are immersed in this mixture, they confine the fluctuations within the fluid between their surfaces, giving rise to a pairwise interaction which depends on the boundary condition at the surface \cite{Hanke1998,Gambassi2011}. For a single particle, the first direct experimental measurements verified theoretical predictions \cite{bechinger}. This interaction can also lead to aggregation and self-assembly of colloidal particles \cite{Bonn2009,Gambassi2010,Vasilyev2014,Edison2015}, a rich problem which includes many-body interactions \cite{Hobrecht2015,Paladugu2016}.
Critical Casimir forces are a universal feature of second-order phase transitions \cite{Gambassi2009,Krech1994,Gambassi2011}. They were first observed in thin $^{4}$He films near the superfluid $\lambda$-point \cite{Garcia1999}, and have also been observed in thin liquid films suspended on an immiscible liquid near the critical mixing point \cite{Fukuto2005,rafai}. In all cases, the force depends on the surface-surface separation, the geometry of the interfaces, and more importantly, on the boundary condition for the order parameter. Symmetric boundary conditions almost always give rise to an attractive force, and asymmetric boundary conditions lead to repulsive forces.
Casimir forces have strengths that are proportional to the driving energy of the fluctuations, and thus proportional to $\hbar$ in quantum systems and temperature in classical systems. However, Casimir-like forces have recently been identified in a number of nonequilibrium systems as a result of the confinement of fluctuations \cite{Hanke2013,Brito2007}. In most of these cases, the fluctuations arise from energy input into the system either by an external field or by the individual particles \cite{grano2}. Examples include driven acoustic noise in a gas \cite{Larraza1998}, soft modes in polymer melts \cite{Obukhov2005}, temperature gradients in liquids \cite{Kirkpatrick2015,Najafi2004}, active matter composed of swimming particles \cite{Ray2014}, and hydrodynamic fluctuations in driven granular fluids \cite{catt,duncan,zuri,villa,reza}. In many of these examples, the Casimir-like forces do not obey a simple scaling function in accordance with Fisher and de Gennes' original argument \cite{fish}. In non-equilibrium systems, the boundary conditions can depend on the order parameter, and the spectrum of fluctuations are not well-defined, in contrast to thermal and quantum systems.
As opposed to classical thermodynamic phase transitions where temperature plays a pivotal role, the jamming transition in granular materials is mainly governed by changes in the density \cite{liu}. In the simplest scenario, a jammed system is composed of frictionless spheres with finite-ranged interactions at $T=0$ (zero kinetic energy). As the density changes, the system reaches a special point, called point $J$, where the spheres begin to touch one another and the system develops a finite pressure \cite{jamm1,corey}. The physics at point $J$ shares many similarities to second-order phase transitions \cite{Drocco2005,Head2009}, yet it is distinctly different. It is a rare example of a random first-order transition in which there are several length scales that diverge as the critical density of the system is approached \cite{Biroli2007,biroli,goodrich2,goodrich3,Reichhardt2014}.
Here we report simulations which show the existence of Casimir-like forces in systems composed of frictionless particles near the $T=0$ jamming transition in two dimensions. The particles interact through a finite-ranged repulsive force, so that there is a well-defined isostatic point as the density is varied. At this isostatic point, the number of mechanical contacts between the particles is equal to the number of motional degrees of freedom. Two equal-sized particles are pinned while the rest of the system, composed of $N$ particles of two sizes, are quenched to mechanical equilibrium, as shown in Fig.\ \ref{cartoon}. Because the pinned particles are fixed in place, they retain a net force, which has two distinct components. The particles experience a short-ranged depletion force which depends on the details of the pair-distribution function, and a long-ranged repulsive force. This long-ranged force increases in magnitude near the jamming transition, as does the fluctuations of this force between different particle configurations.
\begin{figure}
\includegraphics[width=.47 \textwidth]{cartoon.pdf}
\caption{\label{cartoon}Schematic showing a typical simulation. Two pinned particles (yellow) of radius $R$ are held in place while a sea of bidisperse disks are quenched around them. The distance between their surfaces is $D$. In equilibrium, the pinned particles retain a net force, which is equal and opposite.}
\end{figure}
We identify this long-ranged force as a Casimir-like force induced by fluctuations in systems near jamming. The specific dependence of the force on particle separation will be determined by the nature of the fluctuations which give rise to the force. In granular systems, it is well-known that forces are spatially heterogeneous and are clustered in into linear chains \cite{Majmudar2005,Giusti2016}. The distributions of the strength and length of these force chains have been studied \cite{Sanfratello2011,Liu1995,Peters2005,Corwin2005}. These studies, and many others, have shown that both strong and long force chains are exponentially rare, and depend on the state and history of the system. For example, shearing a granular system in one direction will make force chains preferentially align to resist the applied shear, and can induce jamming in unjammed systems \cite{Bi2011}.
While it is not clear how confinement, packing fraction, and system preparation affects the distribution of force chains, our results show that near the jamming transition, the Casimir-like force follows a universal form which scales as $B-A\Delta\phi^{1/2}$, where $\Delta\phi=\phi-\phi_c$ is the distance to the isostatic packing fraction $\phi_c$, and the parameters $B$ and $A$ decrease as the pinned particles move further apart. In our simulations, the system is at finite pressure, so that $\phi$ is strictly greater than $\phi_c$. This Casimir-like force is distinctly different than other Casimir-like forces reported in granular systems \cite{catt,duncan,zuri,villa,reza}. In the latter, hydrodynamic fluctuations are induced by external mechanical excitation, so that there is significant kinetic energy in the system ($T>0$). In our simulations, the Casimir-like force is determined by the critical behavior and inherent fluctuations associated with the $T=0$ jamming transition \cite{Goodrich2014b,Sanfratello2011}.
The origin of this force lies in the distribution of particles with large forces in systems near jamming, and the coordination number of the jammed particles near the boundary of the pinned particles. We find that the inclusion of pinned particles in a jammed system reduces the mean coordination number of the entire system. This reduction is mostly localized near the boundary of the pinned particles. In addition, near jamming, the largest contact forces occur between particles with the minimum number of contacts, which is $z=3$ in two dimensions. These two properties result in the region between the pinned particles having less than the mean number of contacts, and large contact forces, which push the pinned particles apart. Although not explicitly investigated here, we also show how our results may be generalized to higher dimensions.
\section{Methods}
\label{methods}
Our simulations consist of $N$ particles that interact through a finite-ranged, repulsive potential \cite{corey}:
\begin{equation}
\label{pot}
V(r_{ij})=
\begin{cases}
K\dfrac{\epsilon}{\alpha}\left(1-\dfrac{r_{ij}}{\sigma_{ij}}\right)^{\alpha}
& \text{for $r_{ij}<\sigma_{ij}$,}
\\
0 &\text{for $r_{ij}>\sigma_{ij}$,}
\end{cases}
\end{equation}
\noindent
where $\epsilon$ is the energy scale of the interaction, $r_{ij}$ is the distance between the centers of particles $i$ and $j$, $\sigma_{ij}=\sigma_i+\sigma_j$ is the sum of the particle radii, and $\alpha$ defines the type of interaction. Unless otherwise noted, all simulations use $\alpha=2$ (harmonic interactions), however, we obtain similar results with other values such as $\alpha=5/2$ (Hertzian interactions). In order to prevent crystallization, we use a binary mixture of particles with radii $\sigma$ and $1.4\sigma$, with equal numbers of both particles. For these $N$ particles, the constant $K=1$. In addition to these $N$ particles, we include two particles of radius $R$ separated by a distance $D+2R$, so that the surface-surface separation is $D$ (Fig.\ \ref{cartoon}). These particles are pinned in place, and not allowed to move. For large values of $R$, Eq.\ \ref{pot}
produces a very soft potential so that the free particles easily penetrate deep into the boundary of the pinned particles. To avoid this, for particles interacting with the pinned particles, we set $K=(R+1.4\sigma)^{\alpha}/(2.8\sigma)^{\alpha}$. This way the force between pinned particles and free particles is finite in the limit $R\rightarrow\infty$. Near the jamming transition ($\Delta\phi=0$) where the Casimir-like force dominates, our results are insensitive to the choice of $K$ since the overlap distance is much less than both $R$ and $\sigma$.
Each simulation is initiated by fixing the positions of the pinned particles, and placing the remaining $N$ particles down at random in a square box with sides of length $L$ with periodic boundary conditions on all sides. The free particles are then instantaneously quenched to the nearest equilibrium configuration using the FIRE algorithm \cite{fire,Burton2016}. The algorithm is terminated when all the kinetic energy has been removed from the system. Our condition for equilibrium was chosen so that the magnitude of any net force on the free particles is less than $10^{-14}\times\epsilon/\sigma$. Once the system reached equilibrium, we measured both the $x$ and $y$ components of the residual force on the pinned particles ($F_x$ and $F_y$). Since these two particles are the only fixed objects in the system, they must experience equal and opposite forces according to Newton's third law. Thus we report only the force on the upper particle, so a positive force in the $y$-direction corresponds to a net repulsive force between the two pinned particles. The net force components on the pinned particles will fluctuate between different quenched systems, and can be positive or negative, thus we report ensemble averages for the measured forces. Of special interest is the standard deviation of the force distribution in an ensemble, since this is indicative of the inherent force fluctuations in jammed systems, and will be discussed in section \ref{results}.
For small values of $\Delta\phi$, the forces between particles are naturally smaller since there is less overlap. It is more relevant to study the net force normalized by the system pressure and the radius of the pinned particles, $F/(PR)$. The system pressure was measured at each packing fraction, as defined in reference \cite{corey}:
\begin{equation}
P=-\dfrac{1}{2L^2}\sum_{i<j}r_{ij}\dfrac{dV_{ij}}{dr_{ij}}.
\label{pres}
\end{equation}
Here the summation runs over all free and pinned particles in the system. All force measurements reported here will be normalized in this fashion. In addition, the effects of the initial placement of particles, as well as system annealing, will be discussed in the results.
\section{Results and Discussion} \label{results}
There are two distinct components of the force between pinned particles: a short-ranged force which depends on the local packing structure, and a long-ranged Casimir-like force. The relative strength of these two forces will depend on many length scales, such as $\sigma$, $R$, and $D$. In addition, there are multiple characteristic length scales in the jamming transition which diverge as $\Delta\phi\rightarrow 0$. These length scales are associated with the smallest rigid cluster that can be formed at a given $\Delta\phi$ \cite{goodrich2}, and the system's stability to transverse deformations \cite{goodrich3}. For the time being, we associate a general correlation length, $\xi$, with these diverging length scales and we will discuss its relation to both later in the results. Lastly, the size of the bounding box, $L$, will play an important role due to the prescribed periodic boundary conditions. Since the Casimir-like force is long-ranged, we can not ignore the contribution from image particles in neighboring simulation cells.
In general, the normalized force between the two pinned particles will be a function of the ratio of these length scales:
\begin{align}
\left\langle\frac{F_y}{PR}\right\rangle=\Theta\left(\frac{D}{R},\frac{\sigma}{\xi},\frac{R}{\sigma},\frac{R}{L}\right).
\label{fulldep}
\end{align}
In practice, any linearly-independent combination of the ratios of the four length scales would do, however, we have chosen this combination because they are most relevant to our results. A particularly important limiting case, which will be discussed in section \ref{long_ranged}, is when $R/\sigma\rightarrow\infty$ and $R/L\rightarrow 0$. In this limit the details of the jammed medium between the pinned particles are less important, and effects due to the finite size of the simulation domain are negligible. In our simulations these finite-size effects can not be ignored, and will affect the data in different ways.
First, the basic dependencies of the residual force on $D/R$ and $R/\sigma$ are illustrated in Fig.\ \ref{force_and_standev}a-b, which shows the $x$ and $y$-component of the mean force ($\langle F_x\rangle$ and $\langle F_y\rangle$) for two different size ratios at $\phi$ = 0.855 ($\Delta\phi\approx$ 0.015). We observe that $\langle F_{x}\rangle=0$, as expected due to the symmetry of the geometry, as shown in Fig.\ \ref{cartoon}. However, $\langle F_y\rangle$ can vary significantly with particle size, $R/\sigma$, and particle separation, $D/R$. Small pinned particles can experience large positive and negative forces, which decay in amplitude with $D$, although $\langle F_y\rangle$ remains slightly positive for large $D$ (Fig.\ \ref{force_and_standev}a). For large pinned particles, the net force can be negative when the pinned particles are very close ($D/\sigma\sim 1$), as shown in Fig.\ \ref{force_and_standev}b. This is due to depletion. However, for larger values of $D$ there is only a repulsive force, which slowly decays as $D$ increases.
\begin{figure}[t]
\includegraphics[width=.47 \textwidth]{force_and_standev.pdf}
\caption{\label{force_and_standev} (a-b) Components of the mean normalized force between the two pinned particles for two different values of $R/\sigma$: (a) $R/\sigma$ = 1.4, and (b) $R/\sigma$ = 21 ($L/\sigma\approx235$, $\phi = 0.855$, $F_x$ $\square$, $F_y$ $\blacksquare$). Each data point is the average of 200 systems with $N$ = 10,000 particles. PDFs of $F_{y}/(PR)$ are shown in panels (c) and (d), which correspond to the points indicated by the blue arrow in (a) and red arrow in (b). }
\end{figure}
We note here that by symmetry, $\langle F_y\rangle$ will be zero when the distance between the center of the pinned particles is equal to half the system size, which occurs when $D=D^*$, where $D^*$ is given by
\begin{align}
D^*\equiv\frac{L}{2}-2R.
\label{dstar}
\end{align}
This is because of the periodic boundary conditions: each pinned particle feels the same interaction from the image particles in neighboring, tessellated systems. This will be important when comparing systems with different values of $R/\sigma$, and will be discussed in detail. For smaller values of $D/R$, as $D\rightarrow 0$, eventually a point will be reached where only a few jammed particles (or none) can fit between the pinned particles. This depletion-like effect results in a net attractive force since little or no particles exist to push the pinned particles apart. This is true for both small and large values of $R/\sigma$, as shown in Fig.\ \ref{force_and_standev}a-b.
Each data point in Fig.\ \ref{force_and_standev}a-b represents the average of 200 independent systems which have been quenched from a random configuration. Figure \ref{force_and_standev}c-d shows the distribution of forces obtained from these systems for two representative data points, as indicated by the blue and red arrows. We found that the distributions, in general, are not symmetric. In addition, the mean is of the same order as the standard deviation. These characteristics are in contrast to Casimir-like forces in other non-equilibrium systems, such as driven granular gases, where the distributions are symmetric and the mean is less than 5\% of the standard deviation \cite{catt}. In the limits $R/\sigma\rightarrow\infty$ and $D/\sigma\rightarrow\infty$, the distributions will become very narrow due to the dramatic increase in possible jammed system configurations surrounding the pinned particles.
\subsection{Short-Ranged Forces: $R/\sigma\sim 1$}
\begin{figure}[t]
\includegraphics[width=.47 \textwidth]{gofr_force.pdf}
\caption{\label{gofr_force} (a) Normalized force in the $y$-direction versus particle separation for two different system sizes ($R/\sigma$ = 1.4, $\phi = 0.855$. $N=1000$ {\color{black}$\square$}, $N=10,000$ {\color{red}$\fullmoon$}). Each data point represents the average of 800 systems. (b) Data with $N=10,000$ compared with the prediction from Eq.\ \ref{pmf}. }
\end{figure}
First we will focus on the behavior of $\langle F_y\rangle$ when $R/\sigma = 1.4$. Figure \ref{gofr_force}a illustrates this behavior for two system sizes, $N=1000$ and $N=10,000$. The curves are essentially identical, indicating that the important qualitative features of $\langle F_y\rangle$ are independent of the system size, so long as $D\ll L$. Near $D=0$, there is a net attraction between the pinned particles, as evidenced by an overall negative force that further fluctuates between negative and positive values as the distance increases. We attribute these initial fluctuations to a net depletion interaction between the pinned particles due to the amorphous-like structure of the jammed spheres. Similar phenomenology has been observed in experiments with colloidal particles \cite{arjun}. To further strengthen this point, we compare our force values with the result from the force obtained from the mean force potential, which we calculate using the pair correlation function \cite{mcquarrie}:
\begin{equation}
\langle F_y\rangle=k_BT\dfrac{g'(r)}{g(r)}.
\label{pmf}
\end{equation}
Here we have computed $g(r)$ by computing the probability of finding two particles with radii 1.4$\sigma$ at a given value of $D/R$. This method assumes an equilibrium, thermodynamic system with temperature $T$, which defines the fluctuating pathways between different system configurations. Our system configurations are obtained through a quench from $T=\infty$, thus we scaled the amplitude of the predicted force from Eq.\ \ref{pmf} in order to best fit our experimental curve (Fig.\ \ref{gofr_force}b). At this point, the quantitative meaning of the effective temperature used to scale the amplitude is unclear, however, the reasonable agreement between the functional form and the data indicates that the pair distribution function captures the essential features of $\langle F_y\rangle$ for small values of $D/R$. Oscillations in the predicted curve at small values of $D$ are likely due to the differentiation of discrete data from measured values of $g(r)$.
\begin{figure}[t]
\includegraphics[width=.47 \textwidth]{sizerat_force.pdf}
\caption{\label{sizerat_force} Normalized force along the $y$-direction versus particle separation for different size ratios: ($\blacksquare$) $R/\sigma$ = 1.6 , ({\color{red}$\newmoon$}) $R/\sigma$ = 4, ({\color{blue}$\blacktriangle$}) $R/\sigma$ = 8. The packing fraction, $\phi = 0.89$, and the number of particles, $N=1000$, are constant. Each data point represents the average of 800 systems.}
\end{figure}
At larger distances, however, we find that the average force does not tend towards zero as we would expect from liquid-like ordering, rather it remains positive and therefore repulsive. This is indicative of a long-ranged Casimir component to the force, such as those that arise in binary mixtures \cite{fish,bechinger} and in granular systems \cite{catt,grano2}. In order to isolate this component of the force, we gradually increased the size of the pinned particles with respect to the jammed particles. For larger values of $R/\sigma$, the pinned particles have many neighboring particles and the net force does not depend as sensitively on the details of the local packing structure. However, the Casimir-like force becomes more prominent. This can be seen in Fig.\ \ref{sizerat_force}, which shows $\langle F_y \rangle$ for three different values of $R/\sigma$, keeping constant the packing fraction $\phi=0.89$ and the number of particles $N=1000$. As $R/\sigma$ increases, only the positive repulsive force remains, except at very small surface separations when $D/\sigma\sim1$, at which point the packing structure of jammed particles in the region between the pinned particles can strongly influence the mean force. For larger values of $D/R$, the force must tend to zero by the symmetry of the periodic boundary conditions, which can be seen here for $R/\sigma=8$.
\subsection{Long-Ranged Forces: $R/\sigma\gg 1$}
\label{long_ranged}
In order to better understand the Casimir-like component to the total force, we now turn out attention to systems where the pinned particles are significantly larger than the jammed particles ($R/\sigma\geq 20$). For these systems, we performed a detailed study of the dependence of the Casimir-like force on $\Delta\phi$, which controls the distance to the jamming transition. As we approach the critical point, we expect both the size and length scale of the fluctuations to increase \cite{Ellenbroek2006}. In order to obtain systems arbitrarily close to $\Delta\phi=0$, we began with a system at a high packing fraction, and slowly decompressed the system and monitored the decreasing pressure (Eq.\ \ref{pres}).
The pressure scales with $\Delta\phi^{\alpha-1}$ as the jamming transition is approached \cite{corey}. Systems were initially created at a packing fraction of $\phi$ = 0.99, and the packing fraction was decreased slowly in order to obtain a relationship between $P$ and $\phi$. During this process, $R$ was fixed and $\sigma$ was adjusted in order to change the overall packing fraction. Using Newton's root-finding method, we estimated the critical packing fraction where the pressure reaches zero. At each iteration, we reduced $\Delta\phi$ by 1/4, then re-quenched the system, thereby creating a series of packings ranging from $10^{-6}<\Delta\phi<0.15$. This algorithm had a significant effect on the Casimir-like force. Systems which are quenched to the nearest potential energy minimum from $T=\infty$ can rest in a rather high energy state. That is, any amount of annealing can relax the system considerably. This can be seen in Fig.\ \ref{back_and_forth}, which shows a series of decompressions and compressions for $D/R=1$ and $R/\sigma=21$.
We started our systems at $\phi=0.99$ ($\Delta\phi\approx 0.15$), and then reduced the packing fraction in small steps while monitoring the net force on the pinned particles. Initially the force drops, and then begins to increase before reaching a plateau as $\Delta\phi\rightarrow 0$. We attribute the initial drop in the force to an annealing of the system during the initial decompression. The force decreases because free particles in between the two pinned particles are allowed to rearrange and relax. Upon compressing the system back again to large values of $\Delta\phi$, the force drops monotonically. Further decompressions and compressions result in the same qualitative behavior, with the force approaching a steady value of $\langle F_y/(PR)\rangle\approx 0.23$ as $\Delta\phi\rightarrow 0$. Since multiple compressions and decompressions require significant computing time, data reported here has been decompressed, then compressed once (e.g. solid black squares in Fig.\ \ref{back_and_forth}). Further annealing of the systems would result in the same quantitative conclusions, just with smaller overall forces.
\begin{figure}[t]
\includegraphics[width=.47 \textwidth]{back_and_forth.pdf}
\caption{\label{back_and_forth} Mean force in the $y$-direction between two pinned particle upon repeated decompressions and compressions for $D/R=1$ and $R/\sigma=21$. Each data point represents the average of 200 systems with $N$ = 10,000 particles each. Starting from $\Delta\phi=0.15$ ({\color{black}$\square$}), the systems are decompressed towards the jamming transition, then re-compressed. The final state of the system is lowest curve ({\color{Purple}$\bigstar$}). }
\end{figure}
Figure \ref{other_comp}a shows that our results are independent of the type of interaction (the value of $\alpha$ in Eq.\ \ref{pot}). The mean force is shown for both harmonic and Hertzian interactions, with identical results. We also investigated the sensitivity to the initial placement of free particles prior to the quench. For most data, the particles were placed randomly, which means they could easily end up \emph{inside} the pinned particles, and then get pushed out during the quench. Alternatively, we also placed the free particles in a random fashion only on the \emph{outside} of the pinned particles. This had the interesting effect of making the mean force negative (attractive) for large values of $\Delta\phi$. This is likely due to the fact that upon quenching, particles must be pushed into the region between the pinned particles, resulting in a lower average density and depletion-like effect. Nevertheless, upon approach to the jamming transition, the force becomes repulsive. This is shown in Fig.\ \ref{other_comp}b.
\begin{figure}[t]
\includegraphics[width=.47 \textwidth]{other_comp.pdf}
\caption{\label{other_comp} (a) Comparison of the normalized force versus $\Delta\phi$ for both Harmonic ($\blacksquare$) and Hertzian ($\fullmoon$) interactions with $D/R=0.4$ and $R/\sigma=21$. (b) Data for two sequential sets of decompression and compression where all free particles are initially placed outside the boundary of the pinned particles.The order of the events is given by {\color{red}$\newmoon$}, {\color{red}$\fullmoon$}, {\color{blue}$\blacksquare$}, {\color{blue}$\square$}. All data points represent the average of 200 systems with $N$ = 10,000 particles each.}
\end{figure}
For critical Casimir forces, such as those that arise near classical critical points, there are two limiting regimes which have simple scalings in two dimensions for symmetric (Dirichlet) boundary conditions \cite{Hanke1998,Zubaszewska2013}. First, in the limit $R\ll D\ll \xi$, the Casimir force between two equally-sized particles scales as $D^{-5/4}$. In the opposite limit, $D\ll R\ll \xi$, where the Derjaguin approximation is valid, the force scales a $D^{-3/2}$. Thus the Casimir force may be expected to diverge as $D\rightarrow 0$. In practice, this divergence is limited by the assumption that the critical medium is a continuum, i.e. these scalings are valid only when $D$ and $R$ are much larger than the molecular length scale, an assumption which is violated as $D\rightarrow 0$. In our jammed systems, behavior near $D/R\rightarrow 0$ will be significantly affected by the size ratio $R/\sigma$, since this determines the degree to which the jammed particles can be considered a continuum.
For jammed systems, the equivalent limit of $\xi\rightarrow\infty$ corresponds to $\Delta\phi\rightarrow 0$. Figure \ref{force_vs_phi} shows the approach to this limit at various values of particle separation $D/R$, with $R/\sigma$ = 21. In contrast to Fig.\ \ref{back_and_forth} and \ref{other_comp}, data in Fig.\ \ref{force_vs_phi}a is plotted on a linear-linear scale in order to illustrate the sharp increase in the mean force near $\Delta\phi=0$. When the pinned particles are touching each other at $D/R=0$, the mean force is negative for large $\Delta\phi$. This is due to the depletion effect, when $D/\sigma\lesssim 1$. However, as $\Delta\phi$ is decreased, the force sharply increases and becomes repulsive. This sharp increase is seen in all the data, but is weaker as the separation between the pinned particles increases.
Figure \ref{force_vs_phi}b illustrates the relationship between the normalized mean force and the normalized standard deviation of the force distributions, such as those shown in Fig.\ \ref{force_and_standev}c-d. The relationship is always linear, although the slope and intercept vary with parameters such as $D/R$. The data indicates that the size of the fluctuations from system-to-system controls the maximum possible repulsive force. This is consistent with our expectation that Casimir-like forces arise from fluctuations in the system, where the fluctuations here are measured by the standard deviation in the net force over multiple jammed systems.
Near $\Delta\phi=0$, all of the data is consistent with a universal functional form:
\begin{align}
\left\langle \frac{ F_y}{PR}\right\rangle=B-A\Delta\phi^{1/2}.
\label{nearjam1}
\end{align}
We fit all data points below $\Delta\phi=2\times10^{-3}$ to this form to determine the values of $A$ and $B$. A collapse of the data can be seen in Fig.\ \ref{force_vs_phi}c. The data follows a power law with an exponent of 1/2 over three decades in $\Delta\phi$. The noise in the data is likely due to a finite number of systems used for each data point. The respective values of the fitting parameters $A$ and $B$ for each separation distance $D/R$ are shown in the inset. The coefficient $A$ represents the strength of the power law dependence, and increases rapidly for small $D/R$. The exception is when the pinned particles are touching, where depletion forces are present in addition to the Casimir-like force. The constant $B$ has a direct interpretation; it is the limiting value of the force when $\Delta\phi=0$, or alternatively, when the correlation length $\xi=\infty$. $B$ also shows a sharp decrease near $D/R=0$, for the same reasons described above.
\begin{figure}[ht!]
\includegraphics[width=.47 \textwidth]{force_vs_phi.pdf}
\caption{\label{force_vs_phi} (a) Normalized force along the $y$-direction versus $\Delta\phi$ for different surface-to-surface distances: ($\square$) $\frac{D}{R}$ =0 , ({\color{Dandelion}$\fullmoon$}) $\frac{D}{R}$ = 0.2, ({\color{red}$\triangle$}) $\frac{D}{R}$ = 0.4, ({\color{green}$\triangledown$}) $\frac{D}{R}$ = 1, ({\color{blue}$\triangleleft$}) $\frac{D}{R}$ = 1.4, ({\color{violet}$\varhexagon$}) $\frac{D}{R}$ = 2.4, ({\color{CarnationPink} \ding{80}}) $\frac{D}{R}$ = 3. (b) Force versus the standard deviation of the force distribution for the data shown in (a). (c) The same data collapsed onto a universal functional form: $\langle F_y/(PR)\rangle=B-A\Delta\phi^{1/2}$. Each data point represents the average of 200 independent systems with $N$ = 10,000. The inset shows the fitting parameters $A$ and $B$ for each value of $D/R$. For all data, $R/\sigma$ = 21.}
\end{figure}
Critical Casimir forces should depend sensitively on the size of the correlation length $\xi$, which diverges near the critical point \cite{Balibar2005,Hanke1998,Gambassi2009,Krech1994,Gambassi2011}. However, as mentioned previously, the jamming transition resembles a random first-order transition \cite{birolir}, and involves the divergence of multiple length scales, some of which have only recently been illuminated in detail \cite{goodrich3,Mailman2011}. Among these a longitudinal or rigidity length scale, $l^{*}$ \cite{goodrich2, wyart}, related to the maximum size of a stable cluster within the system, and a transverse length scale, $l_{T}$ \cite{goodrich3}, that controls the stability of the system against transverse plane-wave perturbations, have been proposed. Both length scales seem to be relevant to control the stability of the jammed packings but scale differently with $\Delta\phi$. Specifically, $l^*\sim\sigma\Delta\phi^{-1/2}$ and $l_T\sim\sigma\Delta\phi^{-1/4}$. It is not clear which one of these length scales is most important for the Casimir effect in jammed systems. However, given that $l^*$ diverges more quickly, it seems logical that this length scale will play the dominant role. Thus we associate $l^*$ with the correlation length $\xi$.
With this in mind, we can write the resulting normalized force as:
\begin{align}
\left\langle \frac{ F_y}{PR}\right\rangle=B-A\frac{c\sigma}{l^*},
\label{nearjam2}
\end{align}
where $c\approx0.28$ is a numerical constant which has been recently measured in simulations of 2D frictionless disks \cite{goodrich2}. In order to tease apart the dependence of the normalized force on the remaining dimensionless parameters in Eqn.\ \ref{fulldep}, we will focus on the fitting parameters $B$ and $A$. Both will depend on $D/R$, $R/\sigma$, and $R/L$. In this case, the ``thermodynamic limit'' is achieved when $R/\sigma\rightarrow\infty$ and $R/L\rightarrow 0$. For large pinned particles, the limiting values obtained in our simulations are $R/\sigma\approx21,26,32$ and $R/L=0.083,0.105,0.13$. These parameters are coupled together because the ratio $L/\sigma$ is nearly constant for our simulations due to the fixed number of particles ($N=10,000$). Both of these parameters represent a finite-size effect that can not be ignored. However, by varying the separation $D/R$, we can accurately determine the distance-dependence of the Casimir effect.
We are able to obtain a reasonable collapse of the data by plotting $B(R/L)^{1/2}$ and $A(R/L)^{3/4}$ versus $D/D^*$, where $D^*$ is the separation at which the force must go to zero due to the symmetry of the periodic boundary conditions (Eqn.\ \ref{dstar}). The data collapse is shown in Fig.\ \ref{collapse}a-b. The collapse is poor for smaller values of $D/D^*$ because this is where $D/\sigma<5$. The scaling exponents 1/2 ($B$) and 3/4 ($A$) are a natural consequence of the dependence on $D$. If we plot the data on a logarithmic scale (insets in Fig.\ \ref{collapse}), we see that there is a small region where data is consistent with a power-law, bracketed on both sides by finite size effects which cause the data to deviate from the scaling behavior. By fitting the data in this regime, we find exponents in good agreement with the scaling obtained by collapsing the data. These fits are shown by the dashed lines in Fig.\ \ref{collapse}a-b.
\begin{figure}[t]
\includegraphics[width=.47 \textwidth]{collapse.pdf}
\caption{\label{collapse} Normalized fitting parameters $B(L/R)^{1/2}$ (a) and $A(L/R)^{3/4}$ (b) versus $D/D^*$, as defined in Eqn.\ \ref{dstar}. Data is shown for three different values of $R/\sigma$: $R/\sigma=21$ ({\color{red} $\blacktriangle$}), $R/\sigma=26$ ({\color{blue} $\blacksquare$}), and $R/\sigma=32$ ({\color{black} $\newmoon$}). Open symbols are used in (b). The insets show the same data on a logarithmic scale. The dashed lines correspond to fits to the data where $D/\sigma>3.5$ and $D/D^*<6$: $B(R/L)^{1/2}\approx 0.66 (D^*/D)^{0.50\pm0.04}$ and $A(R/L)^{3/4}\approx 1.66 (D^*/D)^{0.78\pm0.05}$.
}
\end{figure}
Thus in this intermediate regime, where the finite-size effects are not dominant, we find:
\begin{align}
\left\langle \frac{ F_y}{PR}\right\rangle\approx0.66\left(\frac{D^*R}{DL}\right)^{1/2}-1.66\left(\frac{D^*R}{DL}\right)^{3/4}\frac{c\sigma}{l^*},
\label{fullexp1}
\end{align}
In the limit $R/L\rightarrow 0$, the distance $D^*\rightarrow L$, and Eqn.\ \ref{fullexp1} becomes
\begin{align}
\left\langle \frac{ F_y}{PR}\right\rangle\approx0.66\left(\frac{R}{D}\right)^{1/2}-0.47\left(\frac{R}{D}\right)^{3/4}\frac{\sigma}{l^*},
\label{fullexp2}
\end{align}
where we have used the numerical value for $c$. Conveniently, as the system size $L\rightarrow\infty$, its dependence drops out of the expression for the force. Although our range of data is not very large, we may say that it is consistent with Eqn.\ \ref{fullexp2}, which gives us some expectation for the dependence of the force on the particle separation when finite-size effects are negligible.
The scaling of critical Casimir forces is often broken into two regimes where $D\ll R$ (Derjaguin) and $D\gg R$ \cite{Hanke1998}. In both cases, the force will scale as $F\propto (R/D)^\beta \theta(D/\xi)$, where the exponent $\beta$ and the function $\theta$ are different in each regime. The universal dependence on $D/\xi$ is a defining feature of traditional Casimir forces \cite{fish}. Equation \ref{fullexp2} is not consistent with this form, so it is unclear if the Casimir-like force between the pinned particles is directly related to quantum or critical Casimir forces. However, as we show in the next section, one of the controlling parameters for both the sign and size of all Casimir forces, namely, the boundary condition at the confining surfaces, varies significantly in jammed systems as the critical point is approached. This is in contrast to the most well-studied examples of critical Casimir forces: binary liquid mixtures near a critical point. Although the correlation length diverges as the temperature approaches the critical point, the hydrophilicity/hydrophobicity of the confining surfaces remains constant. This is perhaps the most important distinction between Casimir-like forces in jammed systems and traditional Casimir forces.
\subsection{Origin of the Casimir Force}
The length scale $l^*$ determines the maximum size of a rigid cluster with no mechanical constraints at the boundaries \cite{goodrich2}. This argument is essentially based on balancing the excess number of contacts within a volume of jammed particles with the number of contacts lost at the free boundaries. At $\Delta\phi=0$, frictionless, jammed systems are isostatic, so there are exactly enough contacts to constrain the degrees of freedom, so $l^*\rightarrow\infty$, meaning that if one contact is broken then the whole system falls apart. Phenomenologically, this is consistent with our simulations. For large volume fractions when $l^*<D$, a rigid cluster can exist in between the pinned particles, so that it need not interact with the boundaries of the pinned particles to maintain a force balance. However, as $l^*\rightarrow \infty$ ($\Delta\phi\rightarrow 0$), the cluster can not maintain its rigidity without interacting with the boundaries of the pinned particles.
The type of interaction will depend on the boundary conditions for the fluctuating fields in the system, as is the case in quantum and critical Casimir forces. In the case of critical Casimir forces in binary liquid mixtures near a critical point, the Casimir force can be positive or negative depending on the preference of each phase to be adjacent to the solid boundary, in effect, the hydrophobicity of the boundary \cite{bechinger}. As we will show using an alteration to the standard Maxwell counting argument for particle contacts, pinned particles in jammed systems require a reduction in the mean contact number per particle, $\langle z\rangle$.
Let's begin by assuming an isostatic ($\Delta\phi=0$), jammed system contains $N$ free particles and $m$ identical pinned particles. Then there are $Nd$ degrees of freedom, where $d$ is the dimension of the system. This must be equal to the number of constraints, or contacts. In the bulk, there are $z_{bulk}$ contacts between free particles, and $z_{bound}$ contacts between free and pinned particles. We are assuming that the pinned particles are sparse, and do not touch. Thus $Nd=z_{bulk}+z_{bound}$. We also assume that each contact $z_{bound}$ represents a single free particle in contact with a single pinned particle, so $z_{bound}=N_{bound}$, where $N_{bound}$ is the number of free particles in contact with the boundary of a pinned particle.
The mean number of contacts for the $N$ particles is given by $\langle z \rangle=(2z_{bulk}+z_{bound})/N$. The factor of 2 accounts for the fact that each bulk contact is shared between two free particles, but bound contacts are not. Thus we find that with the addition of $m$ pinned particles, the coordination number becomes
\begin{equation}
\langle z\rangle=2d-\dfrac{N_{bound}}{N}.
\label{zmean}
\end{equation}
Thus the mean contact number is reduced by an amount $N_{bound}/N$. Far away from a pinned particle, we might expect the bulk system to behave like any other jammed system, so the local mean contact number would be $2d$. For simplicity, let's assume that this reduction in mean contact number is a result of a surface of particles surrounding each pinned particle which has, on average, less contacts than particles in the bulk. We can estimate the thickness of this surface layer in the following way. Instead of being evenly distributed throughout the bulk, this mean contact number reduction will be confined to a layer of thickness $\Delta R$, which surrounds $m$ pinned particles. In two dimensions, the reduction in contact number in this layer, $\langle z \rangle_{layer}$, will be given by total reduction in $\langle z \rangle$ multiplied by the ratio of the total area of the system to the fraction of area included in the localized layers:
\begin{equation}
\langle z \rangle_{layer}=4-\dfrac{N_{bound}}{N}\times\dfrac{L^2-m\pi R^2}{m\pi(R+\Delta R)^2-m\pi R^2}.
\label{zlayer1}
\end{equation}
The area of the total bulk system is reduced by an amount $m\pi R^2$ due to the presence of the $m$ pinned particles. When Casimir-like forces are dominant, i.e. $R/\sigma\gg 1$, we can assume that the boundary of the pinned particles is completely surrounded with free particles, so that $N_{bound}\approx2\pi R m/(2\sigma)$. In addition, at jamming, the system size and particle number are related through the critical packing fraction: $L^2-m \pi R^2\approx N \pi\sigma^2/\phi_{c}$, where $\phi_{c}\approx0.84$. Thus $\langle z \rangle_{layer}$ becomes
\begin{align}
&\langle z \rangle_{layer}=4-\dfrac{\pi\dfrac{\sigma}{R}}{\phi_c\dfrac{\Delta R}{R}\left(2+\dfrac{\Delta R^2}{R^2}\right)}
\label{zlayer2}
\end{align}
In the absence of pinned particles, $\langle z \rangle_{layer}=2d=4$. However, particles need a minimum of $d-1=3$ contacts for stability. Thus we expect $3<\langle z \rangle_{layer}<4$. Near jamming for $R/\sigma=32$, we find that $\langle z \rangle_{layer}\approx3.8$. Using Eqn. \ref{zlayer2}, this suggests that $\Delta R \approx 4.4\sigma$. Thus the thickness of this region of reduced contact number is expected to be a few particles thick. We note that the thickness of the layer is independent of system size ($L$) and the number of pinned particles ($m$).
\begin{figure}[t]
\includegraphics[width=.47 \textwidth]{b_cond.pdf}
\caption{\label{b_cond} (a-b) Coarse-grained color map of the mean contact number per particle, $\langle z \rangle$, around an isolated pinned particle. Data is shown for $R/\sigma=32$ and $D/R=1.9$, and for two different packing fractions: (a) $\Delta\phi=10^{-5}$, and (b) $\Delta\phi=0.15$. The data is averaged over 200 independent systems with $N$ = 10,000 particles each. (c) Plot of $\langle z \rangle$, averaged around the perimeter of the pinned particle, versus the normalized distance, $s/\sigma$, from boundary of the pinned particles. Data is shown for $\Delta\phi=10^{-5}$ ({\color{blue} blue line}), and $\Delta\phi=0.15$ ({\color{red} red line}). For both low and high packing fractions, $\langle z \rangle$ is reduced near the boundary of the pinned particles. Large variations near $s/\sigma=0$ are due to structural layering of particles at the boundary.}
\end{figure}
Confirmation of this prediction can been seen in Fig.\ \ref{b_cond}. Near jamming, when $\Delta\phi$ is small, the mean contact number in the bulk of the system should be $\langle z\rangle=2d$. This is the case far away from the pinned particles, where $s/\sigma\gg 1$, and $s$ is the distance between the center of a free particle and the boundary of a pinned particle. Near the boundary of the pinned particles, $\langle z\rangle$ deviates from this value. The mean contact number decreases to $d-1=3$ at the boundary, which occurs over a few particle radii, as predicted by Eq.\ \ref{zlayer2}. This reduction in $\langle z\rangle$ near the boundary is also true at higher packing fractions, although $\langle z\rangle$ is larger since there are excess contacts in systems far from the jamming point \cite{corey}. Although not shown here, this argument can easily be extended to higher dimensions to obtain similar results by considering the volume of a shell around a $d$-dimensional sphere.
The jamming transition is not a 2nd-order phase transition, yet it shares many similarities. There is no clear choice for a two-point static correlation function from which a growing length scale can be derived \cite{Tewari2009}. Although many quantities have been considered as order parameters for the $T=0$ jamming transition, such as geometric properties of Voronoi tessellations \cite{Morse2016} and point-to-set correlations \cite{Mailman2012}, the number of excess contacts, $\langle z\rangle-2d$, is often considered an order parameter for the system \cite{liu}. This is due to its universal scaling near the jamming transition for $d\geq2$: $\langle z\rangle-2d\propto\Delta\phi^{1/2}$. The local contact number is significantly reduced at the boundary, meaning that particles with less contacts on average will be drawn to the region between the pinned particles. However, if this region of reduced contact number is to stay in mechanical equilibrium with the rest of the system, the contacts should, on average, be stronger than contacts in the bulk of the system. This is the essential reason why the Casimir-like force is strictly positive. The boundary condition requiring a reduced number of contacts necessitates an increased strength per contact. These stronger contact forces are transmitted to the boundaries of the pinned particles, where they push them apart, resulting in a repulsive force.
\begin{figure}[t]
\includegraphics[width=.47 \textwidth]{f_dist.pdf}
\caption{\label{f_dist} PDFs of the mean contact force for particles with different contact numbers. The data comes from systems with no pinned particles, and is averaged over 200 systems with $N$ = 10,000 particles each. (a) $\Delta\phi=10^{-5}$. (b) $\Delta\phi=0.15$.}
\end{figure}
\begin{figure*}[t]
\includegraphics[width=.99 \textwidth]{f_points.png}
\caption{\label{f_points} Locations of the top 0.625\% of particles with the strongest average contact forces ($\langle f_z\rangle$). For each panel, data is shown for 200 independent systems with $N$ = 10,000 and $R/\sigma=32$. (a-c) Three different values of $D/R$ (1.6, 0.8, 0.4) with $\Delta\phi=10^{-5}$. (d-f) The same particle separations with $\Delta\phi=0.15$.}
\end{figure*}
To solidify the link between reduced contact number and stronger forces, in Fig.\ \ref{f_dist} we plot the probability density functions (PDFs) of the average force per contact on a single particle with $z$ contacts:
\begin{align}
\langle f_z\rangle_i&=\dfrac{1}{z}\sum_{j=1}^{z} f_{ij},
\end{align}
where $f_{ij}$ is the magnitude of the contact force between particles $i$ and $j$, and $z$ denotes the number of contacts for particle $i$. The forces are normalized by the mean contact force on all particles in the system:
\begin{align}
\langle f_N\rangle&=\dfrac{1}{N}\sum_{i=1}^{N}\langle f_z\rangle_i.
\end{align}
We separate the PDF for all particles into separate curves for $z$ = 3, 4, 5, 6, and 7 contacts. Figure \ref{f_dist}a shows data near the jamming transition. With this normalization, the total area under all the curves in Fig. \ref{f_dist} is slightly less than unity since we have not considered $z=8$ and larger, which represents a very small fraction of the particles in the systems. An important feature to notice is that particles with the strongest contact forces only have 3 contacts. This may be expected given that these particles are in mechanical equilibrium with the rest of the system, they need stronger forces if they have fewer contacts. Figure \ref{f_dist}b shows the same PDFs at $\Delta\phi=0.15$, far from the jamming transition. For this value of $\Delta\phi$, $\langle z\rangle\approx 5.37$, the same as the data in Fig.\ \ref{b_cond}. The distributions are more narrow with rapidly decreasing tails for strong forces, and particles with $\langle z\rangle$ = 4 and 5 contacts dominate the strongest forces. It is important to note that these distributions are taken from systems with no pinned particles, although they change very little from distributions in the presence of pinned particles.
As stated above, the strongest forces should be localized to the region between the pinned particles. In order to visualize this localization, we look for where the particles with the strongest average contact force are located. Figure \ref{f_points} shows the strongest 0.625\% of particles for 200 independent systems for two different values of $\Delta\phi$, corresponding to Fig.\ \ref{f_dist}, as well as different values of $D/R$. At $\Delta\phi=10^{-5}$ (panels a-c), the data represents particles with $\langle f_z\rangle/\langle f_N \rangle\gtrsim 2.8$, which mostly consists of particles with $z$ = 3 contacts. We find that there is a clear tendency for strong forces to localize in between the pinned particles, especially when the pinned particles are close. This average localization gives rise to the repulsive Casimir effect, which is prominent near jamming. When $\Delta\phi=0.15$, there are strong forces localized to the immediate boundary of the pinned particles, but this localization does not extend far into the region between the particles. The points shown in panels d-f correspond to $\langle f_z\rangle/\langle f_N \rangle\gtrsim 1.6$, where the forces are dominated by particles with $z$ = 4, 5, and 6 contacts. The net effect of this clustering can be seen in Fig.\ \ref{pressure}, which shows a color map of $\langle f_z\rangle$ for closely-spaced particles. Near the jamming transition (Fig.\ \ref{pressure}a), the average contact force in between the particles can be more than 2 standard deviations away from the mean force in the system. This deviation is significantly reduced for larger packing fractions, i.e. far from the jamming transition.
\begin{figure}[t]
\includegraphics[width=.47 \textwidth]{pressure.pdf}
\caption{\label{pressure} Coarse-grained map of $\langle f_z\rangle$ for $D/R=0.4$ with $\Delta \phi=10^{-5}$ (a) and $\Delta \phi=0.15$ (b). The data is averaged over 200 independent systems with $N$ = 10,000 and $R/\sigma=32$. The colors represent mean force ({\color{orange} orange}) and number of standard deviations, $\delta$, from the mean force. }
\end{figure}
Taken together, we interpret the data from Figs.\ \ref{f_dist}, \ref{f_points}, and \ref{pressure} in the following way. Fluctuations in the particle contact number and contact strength occur throughout the bulk, and are larger near jamming, given that the length scale $l^*$ diverges \cite{wyart,goodrich2} and the PDFs in Fig.\ \ref{f_dist} have long tails. The spectrum of fluctuations must be consistent with the boundary condition at the pinned particles. The fixed boundaries of the pinned particles produce a reduction in the mean number of contacts per particle, $\langle z\rangle$. We find that this reduction is localized near the boundary of the pinned particles. As two pinned particles approach one another, the localized layers of reduced $\langle z \rangle$ begin to overlap, resulting in a repulsive interaction due to stronger contacts for particles with $z$ = 3 and $z$ = 4. However, the amplitude of this reduction has an upper bound: $\langle z\rangle$ can not drop below 3 contacts per particle in order to maintain rigidity. Thus, near jamming, the reduction in contact number is spread over a longer distance, i.e. the layer near the boundary is thicker. At larger packing fractions, $\langle z\rangle$ is larger on average, which allows for a sharper decrease in $\langle z\rangle$ near the boundary, and a thinner interaction range for the Casimir effect. Essentially, the excess contacts screen the interaction between the pinned particles.
One important question suggested by our results, and in particular by Fig.\ \ref{b_cond}, is what happens when the pinned particles are allowed to relax with the rest of the system? How would this reduction in contact number near the surface change? We have tested this by allowing the pinned particles to relax to a zero net force, starting from a condition where they were initially pinned. On average, the pinned particles always move further apart due to the repulsive Casimir-like force, before reaching an equilibrium state (zero net force). However, when $R/\sigma\gg1$, the contact number distribution near the boundary is still identical to that of Fig.\ \ref{b_cond}. This is essentially because Eq.\ \ref{zmean} does not require that the particles be pinned. We can see this by considering that large particles immersed in a sea of smaller particles give rise to many contacts ($N_{bound}$), yet they only bring $d$ degrees of freedom per particle to the counting statistics. In this case, where the $m$ pinned particles are allowed to move, we arrive a different expression for the average contact number in the bulk:
\begin{equation}
\langle z\rangle=2d-\dfrac{N_{bound}}{N}+\dfrac{2 d m}{N}.
\label{zmean2}
\end{equation}
If $R/\sigma\gg 1$, so that each large particle has many smaller particles in contact with it boundary ($N_{bound}\gg m$), Eq.\ \ref{zmean2} reduces to Eq.\ \ref{zmean}. If the $m$ particles are the same size as the rest of the jammed particles, then they have the same number of average contacts as any other particle ($N_{bound}=\langle z \rangle m$), and Eq. \ref{zmean2} reduces to $\langle z\rangle=2d$. Thus, there will be a reduction in the number of contacts, which is still localized to the boundary of the large particles, even if the pinned particles are allow to move.
This idea can be seen in Fig.\ \ref{landscape}, which shows an idealization of the potential energy landscape which may be associated with the separation of the large, pinned particles. In general, increasing the separation will decrease the potential energy, so that the ensemble average force between the pinned particles is always repulsive. However, for a single system, if the pinned particles are allowed to relax, they can be quenched to a small, local potential energy minimum. Given some type of excitation, perhaps thermal or an external vibration of the system, the larger particles can escape this local minimum and continue to move further apart. In jammed systems, our results imply that large particles will always tend to move away from each other, or from the fixed boundaries of the system. This may have important implications for the well-known Brazil nut effect, where the largest particles in a system end up near the free surface (away from the bottom boundary of the container) in a driven, granular system.
\begin{figure}[t]
\includegraphics[width=.4 \textwidth]{landscape.pdf}
\caption{\label{landscape} Schematic of the potential energy landscape associated with the position of the large, pinned particles.}
\end{figure}
It is unclear at this point how these results may be related to numerous studies investigating the force on an intruder in a granular medium \cite{Stone2004,Stone2004nat,Candelier2010,Reichhardt2010,Ding2011}. In particular, Stone et al. measure a sharp increase in the total force on an intruder approaching a solid wall. They find that the force is repulsive and depends exponentially on the distance from the wall. A direct comparison of our data with the experiment is complicated by particle-particle, particle-wall, and particle-intruder friction, as well as the free surface of the granular bed, all of which have a potent effect on the experimental measurements. In addition, in our simulations, the force for each value of $D$ is measured by fixing $D$, then quenching the smaller mobile particles. Since we have shown that the preparation of the jammed packing can affect the force, we may expect different results if we first quenched the mobile particles at a large value of $D$, then slowly brought the pinned particles together. We hope to address these questions in future work.
\section{Conclusions}
The jamming transition has been investigated for nearly two decades. We now know that it is a unique example of a random first order transition, and shares similarities with the glass transition. It shares properties with both first and second order phase transitions, such as a finite jump in the order parameter, power law scaling of system properties, diverging correlation lengths, and strong fluctuations near the critical point ($\Delta\phi=0$). The latter two are the main ingredients necessary for macroscopic Casimir forces. Here we have shown how Casimir-like forces arise between two pinned particles in an ambient jammed system of frictionless particles. This force is an ensemble average over many realizations of jammed packings. When the pinned particles are small, they experience sharp fluctuations in their interaction force due to the local structure of the packing, as defined by the pair distribution function. However, at large separations, the particles still retain a small repulsive force.
When the pinned particles are much larger than the ambient jammed particles, the long-ranged Casimir-like force is dominant, and is purely repulsive. Near jamming, the Casimir-like force obeys a universal scaling law, $A-B\Delta\phi^{1/2}$, however the exact values of $A$ and $B$ depend on the preparation of the system. Although we do not currently have an explanation for the exponent value, 1/2, we suspect that this is related to the well-known 1/2 exponent seen in the behavior of the number of excess contacts versus $\Delta\phi$ since the origin of the Casimir effect stems from fluctuations in contact number. The force also depends on the distance between the pinned particles, and is proportional to $(R/D)^{1/2}$ at the jamming transition (Eq.\ \ref{fullexp2}). The repulsive nature of the force and its increased magnitude at short distances can be explained by the confinement of contact number fluctuations. The boundary of the pinned particles requires a decrease in the mean contact number. Regions of particles with fewer contacts have stronger contacts on average, giving rise to an increase in pressure between the pinned particles.
Casimir forces are intimately connected to system fluctuations. Both quantum and thermal fluctuations are easier to define, whereas ensembles of jammed systems depend on the user protocol for generating them. This makes any attempt at an analytic theory more complicated, however, we speculate that the Casimir effect in jammed systems may be intimately related to those found in other nonequilibrium systems \cite{Hanke2013,Brito2007,grano2,Larraza1998,Obukhov2005,catt,duncan,zuri,villa,reza,Ray2014,Kirkpatrick2015,Najafi2004}. Thus we do not expect Casimir-like forces near the jamming transition to follow the same scaling as quantum or critical Casimir forces. Unlike classical critical behavior, the jamming transition is a rare example of a random first-order transition which involves two equally-important diverging length scales, and the boundary condition at interfaces (contact number and contact strength) vary as $\Delta\phi\rightarrow 0$. Nevertheless, there may be some similarities with traditional Casimir forces. The correlation length $l^*$ is associated with the onset of floppy modes which have been shifted upwards in frequency when $\Delta\phi>0$ \cite{liu}. Near jamming, the excess low-frequency modes may serve the same role as Goldstone bosons generated by symmetry breaking in critical Casimir forces. In order to verify this potential hypothesis and fully map out the scaling dependencies of the Casimir-like force on the particle separation, radius, and system size, a more extensive set of simulations would be necessary with many more jammed particles ($N\gg10,000$).
Although our simulations are strictly two-dimensional, Eq.\ \ref{zlayer2} can be derived in any dimension, given knowledge of the critical packing fraction $\phi_c$. Thus we expect our results to be qualitatively valid in three dimensions, although the exact dependence on parameters such as $D/R$ may change. It also remains to be seen how these results extend to finite temperature, especially since a full understanding of the jamming transition at finite temperature is complex, and a subject of active research \cite{DeGiuli2015,Caswell2013,Ikeda2013}. Figure \ref{back_and_forth} would suggest that even upon some form of annealing, the Casimir effect still remains. Also, Fig.\ \ref{landscape} suggests that if the large particles are allowed to move, then they would continue to move farther apart until the system can not escape from a local minimum. We leave these questions open for further study.
Finally, the addition of friction may play a crucial role in the behavior of the Casimir effect, since it strongly affects the average contact number per particle \cite{vanhecke}. Generically, there are still excess contacts above the jamming transition in frictional systems, which would act to ``screen'' the Casimir-like force, as shown in Fig.\ \ref{f_points}. We suspect that a repulsive, Casimir-like force would still dominate near jamming, although it is unclear if the universal behaviors would remain since the jamming transition is more well-defined in the absence of friction \cite{Bi2011}. Although simulations allow precise control over the particle interaction and ensemble generation, it remains to be seen if these ideas can be experimentally observed. Our lab is currently developing a series of ongoing experiments to test these theoretical predictions.
\section{Acknowledgments}
We would like to that Carl Goodrich, Wouter Ellenbroek, Eric Corwin, and Sidney Nagel for helpful discussions. This work was supported by the National Science Foundation through grant NSF DMR-1455086.
\input{casimir_arxiv.bbl}
\end{document}
|
\section{Introduction}
In holographic conformal field theories, states with a simple classical gravity dual interpretation have a remarkable structure of entanglement: according to the holographic entanglement entropy formula~\cite{Ryu:2006bv, Ryu:2006ef, Hubeny:2007xt}, their entanglement entropies for arbitrary regions (at leading order in large $N$) are completely encoded in the extremal surface areas of an asymptotically AdS spacetime. In general, the space of possible entanglement entropies (functions on a space of subsets of the AdS boundary) is far larger than the space of possible asymptotically AdS metrics (functions of a few spacetime coordinates), so this property of geometrically-encodable entanglement entropy should be present in only a tiny fraction of all quantum field theory states~\cite{Lashkari:2014kda}. It is an interesting question to understand better which CFT states have this property\footnote{Even in holographic CFTs, it is clear that not all states will have this property. For example, if $|\Psi_1 \rangle$ and $|\Psi_2 \rangle$ are two such states, corresponding to different spacetimes $M_{\Psi_1}$ and $M_{\Psi_2}$, the superposition $|\Psi_1 \rangle + |\Psi_2 \rangle$ is not expected to correspond to any single classical spacetime but rather to a superposition of $M_{\Psi_1}$ and $M_{\Psi_2}$. Thus, the set of ``holographic states'' is not a subspace, but some general subset.}, and which properties of a CFT will guarantee that families of low-energy states with geometric entanglement exist.
For a hint towards characterizing these holographic states, consider the gravity perspective. A spacetime $M_\Psi$ dual to a holographic state $|\Psi \rangle$ is a solution to the bulk equations of motion. Such a solution can be characterized by a set of initial data on a bulk Cauchy surface (and appropriate boundary conditions at the AdS boundary). The solution away from the Cauchy surface is determined by evolving this initial data forwards (or backwards) in time using the bulk equations. Alternatively, we can think of the bulk solution as being determined by evolution in the holographic radial direction, with ``initial data'' specified at the timelike boundary of AdS. In this case, the existence and uniqueness of a solution is more subtle, but the asymptotic behavior of the fields determines the metric at least in a perturbative sense (e.g. perturbatively in deviations from pure AdS, or order-by-order in the Fefferman-Graham expansion). It is plausible that in many cases, this boundary data is enough to determine a solution nonperturbatively to some finite distance into the bulk, or even for the whole spacetime. Thus, for geometries dual to holographic states, we can say that the bulk spacetime (at least in a perturbative sense) is encoded in the boundary behavior of the various fields.
According to the AdS/CFT dictionary, this boundary behavior is determined by the one-point functions of low-dimension local operators associated with the light bulk fields. On the other hand, the bulk spacetime itself allows us to calculated entanglement entropies (and many other non-local quantities). Thus, the assumption that a state is holographic allows us (via gravity calculations) to determine the entanglement entropies and other non-local properties of the state (again, at least perturbatively) from the local data provided by the one-point functions:
\begin{equation}
\label{Gravrules}
|\Psi \rangle \rightarrow \langle {\cal O}_\alpha(x^\mu)\rangle \rightarrow \phi_\alpha {\rm \; asymptotics} \rightarrow \phi_\alpha (x^\mu, z) \rightarrow {\rm entanglement \; entropies \;} S(A)
\end{equation}
where $\phi$ here indicates all light fields including the metric.\footnote{Here, the region $A$ should be small enough so that the bulk extremal surface associated with $A$ should be contained in the part of the spacetime determined through the equations of motion by the boundary values; we do not need this restriction if we are working perturbatively.}$^,\,$\footnote{Results along these lines in the limit of small boundary regions or constant one-point functions appeared in~\cite{Blanco:2013joa, Banerjee:2014oaa, Banerjee:2014ozp, Giusto:2014aba, Giusto:2015dfa}.}
The recipe (\ref{Gravrules}) could be applied to any state, but for states that are not holographic, the results will be inconsistent with the actual CFT answers. Thus, we have a stringent test for whether a CFT state has a dual description well-described by a classical spacetime: carry out the procedure in (\ref{Gravrules})
and compare the results with a direct CFT calculation of the entanglement entropies; if there is a mismatch for any region, the state is not holographic.\footnote{Another interesting possibility is that the one-point functions could give boundary data that is not consistent with any solution of the classical bulk equations; this possibility exists since the ``initial data'' for the radial evolution problem obeys certain constraints.}
In this paper, our goal is to present some more explicit results for the gravity prediction $S_A^{grav}(\langle {\cal O}_\alpha \rangle)$ in cases where the gravitational equations are Einstein gravity with matter and the region is taken to be a ball-shaped region $B$. We will work perturbatively around the vacuum state to obtain an expression as a power series in the one-point functions of CFT operators. At first-order, the result depends only on the CFT stress tensor expectation value \cite{Casini:2011kv}:
\begin{equation}
\label{firstorder}
S_B(|\Psi \rangle) = S^{vac}_B + 2 \pi \int_B d^{d-1} x {R^2 - r^2 \over 2 R} \langle T_{00} \rangle + {\cal O}(\langle {\cal O}_\alpha \rangle^2) \,.
\end{equation}
This well-known expression is universal for all CFTs since it follows from the first law of entanglement $\delta^{(1)} S_B = \delta \langle H_B \rangle$, where
\begin{equation}
\label{modH}
H_B \equiv -\log \rho_B^{vac} = 2 \pi \int_B d^{d-1} x {R^2 - r^2 \over 2 R} T_{00}
\end{equation}
is the vacuum modular Hamiltonian for a ball-shaped region. Thus, to first-order, the gravity procedure (\ref{Gravrules}) always gives the correct CFT result for ball-shaped regions, regardless of whether the state is holographic.
\subsubsection*{General second-order result for ball entanglement entropy}
Our focus will be on the second-order answer; in this case, it is less clear whether the gravity results from (\ref{Gravrules}) should hold for any CFT or whether they represent a constraint from holography. To obtain explicit formulae at this order, we begin by writing
\begin{equation}
S_B(|\Psi \rangle) = S_B^{vac} + \Delta \langle H_B \rangle - S(\rho_B || \rho_B^{vac})
\end{equation}
which follows immediately from the definition of relative entropy $S(\rho_B || \rho_B^{vac})$ reviewed in Section \ref{Section:background} below. We then make use of a recent result in \cite{Lashkari:2015hha}: to second-order in perturbations from the vacuum state, the relative entropy for a ball-shaped region in a holographic state\footnote{This second-order relative entropy is known as quantum Fisher information.} is equal to the ``canonical energy'' associated with a corresponding wedge of the bulk spacetime. We provide a brief review of this in Section \ref{Section:background} below. On shell, the latter quantity can be expressed as a quadratic form on the space of first-order perturbations to pure AdS spacetime, so we have
\begin{align}
S(\rho_B||\rho_B^{vac}) = \Delta \langle H_B \rangle - \Delta S_B = \frac{1}{2} {\cal E}(\delta \phi_\alpha, \delta \phi_\alpha) + {\cal O}(\delta \phi^3) \,.
\end{align}
Rearranging this, we have a second-order version of (\ref{firstorder}):
\begin{align}
\label{secondorder1}
S_B(|\Psi \rangle)
&= S_B^{vac} + \delta^{(1)}S_B + \delta^{(2)}S_B + \mathcal{O}(\delta \phi^3) \nonumber \\
&= S^{vac}_B + \Delta \langle H_B \rangle - \frac{1}{2} {\cal E}(\delta \phi_\alpha, \delta \phi_\alpha) + {\cal O}(\delta \phi^3) \nonumber \\
&= S^{vac}_B + 2 \pi \int_B d^{d-1} x {R^2 - r^2 \over 2 R} \langle T_{00} \rangle - \frac{1}{2} {\cal E}(\delta \phi_\alpha, \delta \phi_\alpha) + {\cal O}(\delta \phi^3) \,.
\end{align}
As we review in Section~\ref{Section:background} below, the last term can be written more explicitly as
\begin{equation}
\label{secondorder2}
{\cal E}(\delta \phi_\alpha, \delta \phi_\alpha) = \int_\Sigma \omega(\delta g, \lie{\xi} \delta g) - \int_\Sigma \xi^a T^{(2)}_{ab}\epsilon^b\;,
\end{equation}
where $\Sigma$ is a bulk spatial region between $B$ and the bulk extremal surface $\tilde{B}$ with the same boundary, $\omega$ is the ``presymplectic form'' whose integral defines the symplectic form on gravitational phase space, $T^{(2)}_{ab}$ is the matter stress tensor at second-order in the bulk matter fields, and $\xi$ is a bulk Killing vector which vanishes on $\tilde{B}$. The first-order bulk perturbations $\delta \phi_\alpha$ (including the metric perturbation) may be expressed in terms of the boundary one-point functions via bulk-to-boundary propagators
\begin{equation}
\label{propagators}
\delta \phi_\alpha(x, z) = \int_{D_B} K_\alpha( x, z ;x') \langle {\cal O}_\alpha (x') \rangle \,,
\end{equation}
where $D_B$ is the domain of dependence of the ball $B$. Given the one-point functions within $D_B$, we can use \eqref{propagators} to determine the linearized bulk perturbation in $\Sigma$ and evaluate \eqref{secondorder2}.
The expression (\ref{secondorder1}), (\ref{secondorder2}), and (\ref{propagators}) together provide a formal result for the ball entanglement entropy of a holographic state, expanded to second-order in the boundary one-point functions.
\subsubsection*{Explicit results for 1+1 dimensional CFTs}
In order to check the general formula and provide more explicit results, we focus in Section \ref{Sec:Gravity} on the case of 1+1 dimensional CFTs, carrying out an explicit calculation of the gravitational contributions to (\ref{secondorder2}) starting from a general boundary stress tensor. We find the result
\begin{equation}
\label{result2DS2}
\delta^{(2)}S_B^{grav} = -\frac{1}{2} \int_{B'} d x^+_1 \, \int_{B'} d x^+_2 \, K_2 (x^+_1, x^+_2) \langle T_{++} (x^+_1) \rangle \langle T_{++} (x^+_2) \rangle + \left\{+ \leftrightarrow-\right\}
\end{equation}
where the integrals can be taken over any spatial surface $B'$ with boundary $\partial B$, and the kernel is given by
\begin{equation}
\label{result2D}
K_2 (x_1, x_2) = \frac{6 \pi ^2 }{c R^2} \left\{ \begin{array}{ll}
(R-x_1)^2 (R+x_2)^2 & \quad x_1\geq x_2 \\
(R+x_1)^2 (R-x_2)^2 & \quad x_1 < x_2
\end{array} \right. \, ,
\end{equation}
where $c$ is the central charge. In this special case, the conservation equations determine the stress tensor expectation values throughout the region $D_B$ from the expectation values on $B'$, so as in the first-order result (\ref{firstorder}), our final expression involves integrals only over $B'$. This will not be the case for the terms involving matter fields, or in higher dimensions. As a consistency check, we show that the expression (\ref{result2D}) is always negative, as required by its interpretation as the second-order contribution to relative entropy.
We can also check the formula (\ref{result2D}) via a direct CFT calculation by considering states that are obtained from the CFT vacuum by a local conformal transformation. In two dimensions, states with an arbitrary traceless conserved stress-tensor can be obtained, and the entanglement entropy for these states can also be calculated explicitly. We carry out this calculation in section 4, and show that the result (\ref{result2D}) is exactly reproduced.
In Section~\ref{sec:scalar}, we consider the matter terms in (\ref{secondorder2}) providing some explicit results for the quadratic contributions of scalar operator expectation values. Here, as in the generic case, the result takes the form
\begin{equation}
\delta^{(2)}S_B^{matter} = -\frac{1}{2} \int_{D_B} \int_{D_B} G_{\alpha \beta}(x,x') \langle {\cal O}_\alpha(x) \rangle \langle {\cal O}_\beta(x') \rangle
\end{equation}
with integrals over the entire domain of dependence region.
\subsubsection*{Auxiliary de Sitter Space Interpretation}
Recently, in~\cite{deBoer:2015kda} it has been pointed out that the first-order result $\delta^{(1)} S(x^\mu,R)$ for the entanglement entropy of a ball with radius $R$ and center $x^\mu$ can be obtained as the solution to the equation of motion for a free scalar field on an auxiliary de Sitter space $ds^2 = \tfrac{L_{dS}^2}{R^2}(-dR^2 + dx_\mu dx^\mu)$ with the CFT energy density $\langle T_{00}(x^\mu) \rangle$ acting as a source term at $R=0$. In Section~\ref{Sec:EntanglementHolography}, we show that in the 1+1 dimensional case, the stress tensor term (\ref{result2D}) for the entanglement entropy at second-order can also results from solving a scalar field equation on the auxiliary de Sitter space if we add a simple cubic interaction term. In an upcoming paper \cite{helleretal}, it is shown that this agreement extends to all orders for a suitable choice of the scalar field potential. The resulting nonlinear wave equation also reproduces the second-order entanglement entropy near a thermal state in the auxiliary kinematic space recently described in~\cite{Asplund:2016koz}.
Including the contributions from matter fields or moving to higher dimensions, the expression for entanglement entropy involves one-point functions on the entire causal diamond $D_B$, so reproducing these results via some local differential equation will require a more complicated auxiliary space that takes into account the time directions in the CFT. This direction is pursued further in \cite{Czech:2016xec,helleretal}.
\subsubsection*{Discussion}
While the explicit two-dimensional stress tensor contribution (\ref{result2D}) can be obtained by a direct CFT calculation for a special class of states, we emphasize that in general the holographic predictions from (\ref{Gravrules}) are expected to hold only for holographic states in CFTs with gravity duals. It would be interesting to understand better whether all of the second order contributions we considered here are universal for all CFTs or whether they represent genuine constraints/predictions from holography.\footnote{There is evidence in~\cite{Faulkner:2014jva, Speranza:2016jwt, Faulkner:2015csl} that at least some of the contributions at this order can be reproduced by CFT calculations in general dimensions, since they arise from CFT two and three-point functions, though the results there most directly apply to the case where the perturbation is to the theory rather than the state.} In the latter case, and for the results at higher order in perturbation theory, it is an interesting question to understand better which CFT states and/or which CFT properties are required to reproduce the results through direct CFT calculations. This should help us understand better which theories and which states in these theories are holographic.
\section{Background}
\label{Section:background}
Our holographic calculation of entanglement entropy to second-order in the boundary one-point functions makes use of the direct connection between CFT quantum Fisher information and canonical energy on the gravity side, pointed out recently in \cite{Lashkari:2015hha}. We begin with a brief review of these results.
\subsection{Relative entropy and quantum Fisher information}
Our focus will be on ball-shaped subsystems $B$ of the CFT$_{d}$, for which the the vacuum density matrix is known explicitly through (\ref{modH}). More generally, we can write it as
\begin{equation}
\label{modH2}
\rho_B^{vac} = e^{-H_B}\,, \qquad \qquad H_B = \int_{B'} \zeta_B^{\mu} T_{\mu \nu} \epsilon^\nu\,,
\end{equation}
where $T_{\mu \nu}$ is the CFT stress tensor operator and $\epsilon$ is defined as
\begin{equation}
\label{defeps}
\epsilon_\nu = {1 \over (d-1)!} \epsilon_{\nu \nu_1 \cdots \nu_{d-1}} dx^{\nu_1} \wedge \cdots \wedge dx^{\nu_{d-1}}\,,
\end{equation}
so that $n^\mu \epsilon_\mu$ is the volume form on the surface perpendicular to a unit vector $n^\mu$, and $\zeta_B$ is a conformal Killing vector defined in the domain of dependence region $D_B$, with $\zeta_B = 0$ on $\partial B$. For the ball $B$ with radius $R$ and center $x_0^\mu$ in the $t=t_0$ slice, we have
\begin{equation}\labell{defzeta}
\zeta_B = - \frac{2\pi}{ R} (t-t_0) (x^i-x^i_0) \partial_i + \frac{\pi}{R} [R^2 - (t-t_0)^2 - (\vec{x}-\vec{x_0})^2] \, \partial_t \; .
\end{equation}
By the conservation of the current $\zeta_B^\mu T_{\mu} {}^{\nu}$ associated with this conformal Killing vector, the integral in (\ref{modH2}) can be taken over any spatial surface $B'$ in $D_B$ with the same boundary as $B$.
For excited states, the density matrix $\rho_B$ will generally be different than $\rho_B^{vac}$. One measure of this difference is the relative entropy
\begin{eqnarray}
S(\rho_B || \rho^{vac}_B) &=& \textrm{tr}(\rho_B \log \rho_B)-\textrm{tr}(\rho_B \log \rho^{vac}_B) \cr
&=& \Delta \langle H_B \rangle -\Delta S_B \, ,
\label{REdef}
\end{eqnarray}
where $H_B$ is the vacuum modular Hamiltonian given in (\ref{modH2}), $S_B = -{\rm tr}(\rho_B \log \rho_B)$ is the entanglement entropy for the region $B$ and $\Delta$ indicates the difference with the vacuum state.
For a one-parameter family of states near the vacuum, we can expand $\rho_B$ as
\begin{equation}
\rho_B (\lambda) = \rho^{vac}_B+ \lambda \; \delta \rho_1 + \lambda^2 \delta \rho_2 + \mathcal{O}(\lambda^3)\,.
\end{equation}
The first-order contribution to relative entropy vanishes (this is the first law of entanglement $\delta^{(1)} S_B=\delta \lambda H_B\ra$) so the leading contribution to relative entropy appears at second-order in $\lambda$. This quadratic in $\delta \rho_1$ with no contribution from $\delta \rho_2$,
\begin{equation}
\label{REpert}
S(\rho_B(\lambda) || \rho^{vac}_B ) = \lambda^2 \, \langle \delta \rho_1, \delta \rho_1 \rangle_{\rho^{vac}_B} + \mathcal{O}(\lambda^3)\,,
\end{equation}
where
\begin{equation}
\langle \delta \rho, \delta \rho \rangle_\sigma \equiv \frac{1}{2} {\rm tr} \left( \delta \rho \frac{d}{d \lambda} \log(\sigma + \lambda \delta \rho) \Big|_{\lambda=0} \right) \; .
\end{equation}
This quadratic form, which is positive by virtue of the positivity of relative entropy, defines a positive-(semi)definite metric on the space of perturbations to a general density matrix $\sigma$. This is known as the quantum Fisher information metric.
Rearranging (\ref{REdef}) and making use of (\ref{REpert}), we have
\begin{equation}
\label{Sexp}
S_B = S_B^{vac} + \int_{B'} \zeta_B^{\mu} \langle T_{\mu \nu} \rangle \epsilon^\nu - \lambda^2 \langle \delta \rho_1, \delta \rho_1 \rangle_{\rho^{vac}_B} + \mathcal{O}(\lambda^3) \; .
\end{equation}
This general expression is valid for any CFT, but the ${\cal O}(\lambda^2)$ term generally has no simple expression in terms of local operator expectation values. However, for holographic states we can convert this term into an expression quadratic in the CFT one-point functions by using the connection between quantum Fisher information and canonical energy.
\subsection{Canonical energy}
Consider now a holographic state, which by definition is associated with some dual asymptotically AdS spacetime $M$. Near the boundary, we can describe $M$ using a metric in Fefferman-Graham coordinates as
\begin{equation}
\label{FG}
ds^2 = {\ell_{AdS}^2 \over z^2} \left( dz^2 + dx_\mu dx^\mu + z^d \, \Gamma_{\mu \nu}(x, z) dx^\mu dx^\nu \right)
\end{equation}
where $\Gamma_{\mu \nu}(z,x)$ has a finite limit as $z \to 0$ and $\Gamma=0$ for pure AdS.
The relative entropy $S(\rho_B || \rho_B^{vac})$ can be computed at leading order in large $N$ by making use of the holographic entanglement entropy formula, which relates the entanglement entropy for a region $A$ to the area of the minimal-area extremal surface $\tilde{A}$ in $M$ with boundary $\partial A$,
\begin{align}
\label{eq:rel-ent}
S_A & \equiv \frac{ {\rm Area} (\tilde{A})}{4\,G_N} \, .
\end{align}
This yields immediately that $\Delta S_A = ({\rm Area}(\tilde A)_M - {\rm Area}(\tilde A)_{AdS})/(4 G_N)$. The result (\ref{eq:rel-ent}) also allows us to relate the $\Delta \langle H_B \rangle$ term in relative entropy to a gravitational quantity, since it implies that the expectation value of the CFT stress tensor is related to the asymptotic behaviour of the metric through~\cite{Faulkner:2013ica}
\begin{align}
\lambda T_{\mu\nu} \ra &= \frac{d \ell_{AdS}^{d-1}}{16 \pi G_N} \Gamma_{\mu\nu}(x, z=0) \,.
\label{eq:HoloRenT}
\end{align}
Thus, for holographic states, we can write
\begin{equation}
\label{REgrav}
S(\rho_B||\rho_B^{vac}) = \frac{d \ell_{AdS}^{d-1}}{16 \pi G_N} \int_{B} \zeta_B^{\mu} \Gamma_{\mu\nu}(x, 0)\, \epsilon^\nu - \frac{{\rm Area}(\tilde A)_M - {\rm Area}(\tilde A)_{AdS}}{4 G_N} \,.
\end{equation}
For a one-parameter family of holographic states $|\Psi(\lambda) \rangle$ near the CFT vacuum, the dual spacetimes $M(\lambda)$ can be described via a metric and matter fields $\phi_\alpha = (g, \phi^{matter})$ with some perturbative expansion
\begin{align}
g &= g_{AdS} + \lambda \delta g_1 + \lambda^2 \delta g_2 + {\cal O} (\lambda^3)\;, \nonumber \\
\phi^{matter} &= \lambda \delta \phi^{matter}_1 + \lambda^2 \delta \phi^{matter}_2 + {\cal O} (\lambda^3) \;.
\end{align}
By the result (\ref{Sexp}) from the previous section, the second-order contribution to entanglement entropy is equal to the leading order contribution to relative entropy. This is related to a gravitational quantity via (\ref{REgrav}). The main result in \cite{Lashkari:2015hha} is that this second-order quantity can be expressed directly as a bulk integral over the spatial region $\Sigma$ between $B$ and $\tilde{B}$ where the integrand is a quadratic form on the linearized bulk perturbations $\delta g_1 $ and $\delta \phi^{matter}_1$.
\begin{figure}[h!]
\centering
\includegraphics{boundary-front.pdf}
\caption{The Rindler wedge $R_B$ associated to the ball-shaped region $B$ on the boundary. The blue lines indicate the flow of $\zeta_B$, and the red lines $\xi_B$. The surface $\Sigma$ lies between $B$ and the extremal surface $\tilde B$.}\label{fig:rindler-wedge}
\end{figure}
To describe the general result, consider the region $\Sigma$ between $B$ and $\tilde{B}$ in pure AdS spacetime, and define $R_B$ as the domain of dependence of this region, as shown in figure \ref{fig:rindler-wedge}. Alternatively, $R_B$ is the intersection of the causal past and the causal future of $D_B$; it can be thought of as a Rindler wedge of AdS associated with $B$. On $R_B$, there exists a Killing vector which vanishes at $\tilde{B}$ and approaches the conformal Killing vector $\zeta_B$ at the boundary. In Fefferman-Graham coordinates, this is
\begin{equation}
\label{defxi}
\xi_B = - \frac{2\pi}{ R} (t-t_0) [z \partial_z + (x^i-x^i_0) \partial_i ] + \frac{\pi}{R} [R^2 - z^2 - (t-t_0)^2 - (\vec{x}-\vec{x_0})^2] \, \partial_t
\end{equation}
The vector $\xi_B$ is timelike hence defines a notion of time evolution within the region $R_B$; the ``Rindler time'' associated with this Rindler wedge.
The ``canonical energy'', dual to relative entropy at second-order, can be understood as the perturbative energy associated with this time, as explained in \cite{Hollands:2012sf}. This is quadratic in the perturbative bulk fields including the graviton, and given explicitly by
\begin{eqnarray}
\label{canEn}
{\cal E}(\delta g_1, \delta \phi_1) &=& W_\Sigma \left(\delta \phi_1 , \lie{\xi_B} \delta \phi_1 \right) \cr
&=& \int_\Sigma \omega^{full} \left( \delta \phi_1, \lie{\xi_B} \delta \phi_1 \right) \cr
&=& \int_\Sigma \omega \left( \delta g_1, \lie{\xi_B} \delta g_1\right) + \int_\Sigma \omega^{matter} \left( \delta \phi_1, \lie{\xi_B} \delta \phi_1 \right) \cr
&=& \int_\Sigma \omega(\delta g_1, \lie{\xi_B} \delta g_1) - \int_\Sigma \xi^a_B T^{(2)}_{ab} \epsilon^b \; .
\end{eqnarray}
In the first line, $W_\Sigma$ is the symplectic form associated with the phase space of gravitational solutions on $\Sigma$, and $\lie{\xi_B} \delta \phi_1$ is the Lie derivative with respect to $\xi$ on $\delta \phi_1$, the first-order perturbation in metric and matter fields. The symplectic form is equal to the integral over $\Sigma$ of a ``presymplectic'' form $\omega^{full}$ which splits into a gravitational part and a matter part as in the third line. The matter part can be written explicitly in terms of $T^{(2)}_{ab}$, the matter stress tensor at quadratic order in the fields, while the gravitational part $\omega$ is given explicitly by
\begin{eqnarray}
\label{defomega2}
\omega(\gamma^1, \gamma^2) &=& {1 \over 16 \pi G_N} \epsilon_a P^{abcdef} (\gamma^2_{bc} \nabla_d \gamma^1_{ef} - \gamma^1_{bc} \nabla_d \gamma^2_{ef}) \\
P^{abcdef} &=& g^{ae} g^{fb} g^{cd} - {1 \over 2} g^{ad} g^{be} g^{fc} - {1 \over 2} g^{ab} g^{cd} g^{ef} - {1 \over 2} g^{bc} g^{ae} g^{fd} + {1 \over 2} g^{bc} g^{ad} g^{ef} \; . \nonumber
\end{eqnarray}
In deriving (\ref{canEn}) it has been assumed that the metric perturbation has been expressed in a gauge for which the coordinate location of the extremal surface $\tilde{B}$ does not change (so that $\xi_B$ continues to vanish there), and the vector $\xi_B$ continues to satisfy the Killing equation at $\tilde{B}$. Thus, we require that
\begin{eqnarray}
\label{eq:gauge-condition}
\xi_B |_{\tilde B (\lambda)} &= 0,\\
\label{eq:gauge-condition2}
\lie{\xi_B} g(\lambda) |_{\tilde B (\lambda)} &= 0.
\end{eqnarray}
As shown in \cite{Hollands:2012sf}, it is always possible to satisfy these conditions; we will see an explicit example below.
\section{Second-order contribution to entanglement entropy}
\label{Sec:Gravity}
Using the result (\ref{secondorder2}), we can now write down a general expression for the ball entanglement entropy of a general holographic state up to second-order in perturbations to the vacuum state, in terms of the CFT one-point functions. According to (\ref{Sexp}) and (\ref{canEn}), the second-order term in the entanglement entropy for a ball $B$ can be expressed as an integral over the bulk spatial region $\Sigma$ between $B$ and the corresponding extremal surface $\tilde{B}$, where the integrand is quadratic in first-order bulk perturbations.
These linearized perturbations are determined by the boundary behavior of the fields via the linearized bulk equations. In general, to determine the linearized perturbations in the region $\Sigma$ (or more generally in the Rindler wedge $R_B$), we only need to know the boundary behavior in the domain of dependence region $D_B$, as discussed in detail in \cite{Almheiri:2014lwa}. The relevant boundary behaviour of each bulk field is captured by the one-point function of the corresponding operator. We can express the results as
\begin{equation}
\label{bulktoboundary}
(\delta \phi_1)_\alpha(x,z)|_\Sigma = \int_{D_B} d^d x' K_\alpha(x,z ; x') \langle {\cal O}_\alpha(x') \rangle_{CFT}
\end{equation}
where $K_\alpha(x,z ; x')$ is the relevant bulk-to-boundary propagator. As discussed in \cite{Morrison:2014jha, Almheiri:2014lwa,Bousso:2012mh}, $K_\alpha$ should generally be understood as a distribution to be integrated against consistent CFT one-point functions, rather than a function. Since the expression (\ref{bulktoboundary}) is linear in the CFT expectation values, the result (\ref{secondorder2}) is quadratic in these one-point functions and represents our desired second-order result.
To summarize, for a holographic state, the second-order contribution to entanglement entropy in the expansion (\ref{Sexp}) is the leading order contribution to the relative entropy $S(\rho_B||\rho_B^{vac})$. This is dual to canonical energy, given explicitly by:
\begin{equation}
\label{Stwo}
\delta^{(2)} S_B = -\langle \delta \rho_1, \delta \rho_1 \rangle_{\rho^{vac}_B} = - \frac{1}{2} {\cal E}( \delta \phi_1, \delta \phi_1) = -\frac{1}{2} \int_\Sigma \omega(\delta g_1, \lie{\xi_B} \delta g_1) + \frac{1}{2} \int_\Sigma \xi^a_B T^{(2)}_{ab} \epsilon^b \; .
\end{equation}
This is quadratic in the linearized perturbations $\delta \phi_\alpha$ (including the metric perturbation, and these can be expressed in terms of the CFT one-point functions on $D_B$ as (\ref{bulktoboundary}).
\subsection{Example: CFT$_2$ stress tensor contribution}
\label{SubSec:CFT2stress}
In this section, as a sample application of the general formula, we provide an explicit calculation of the quadratic stress tensor contribution to the entanglement entropy for holographic states in two-dimensional conformal field theories. This arises from the first term in (\ref{secondorder2}).
For a general CFT state, the stress tensor is traceless and conserved,
\begin{equation}
\langle T^\mu {}_\mu \rangle = \langle \partial_\mu T^{\mu \nu} \rangle = 0 \; .
\end{equation}
In two dimensions, these constraints can be expressed most simply using light-cone coordinates $x^\pm = x \pm t$, where we have
\begin{equation}
\langle T_{+-} \rangle = \partial_+ \langle T_{--} \rangle = \partial_- \langle T_{++} \rangle = 0 \; .
\end{equation}
Thus, a general CFT stress tensor can be described by the two functions, $\langle T_{++} (x^+)\rangle$ and $\langle T_{--} (x^-)\rangle$.
Assuming that the state is holographic, there will be some dual geometry of the form (\ref{FG}). According to (\ref{eq:HoloRenT}), the stress tensor expectation values determine the asymptotic form of the metric as
\begin{equation}
\Gamma_{++}(x,0) = 8 \pi {G_N \over \ell_{AdS}} \langle T_{++} (x^+)\rangle \qquad \Gamma_{--}(x,0) = 8 \pi {G_N \over \ell_{AdS}} \langle T_{--} (x^-)\rangle
\end{equation}
Now, suppose that our state represents a small perturbation to the CFT vacuum, so that the stress tensor expectation values and the asymptotic metric perturbations are governed by a small parameter $\lambda$:
\begin{equation}
\label{boundary}
\Gamma_{++}(x,0) \equiv \lambda h_+(x^+) \qquad \Gamma_{--}(x,0) \equiv \lambda h_-(x^-) \; .
\end{equation}
Then the metric perturbation throughout the spacetime is determined by this asymptotic behavior by the Einstein equations linearized about AdS. Here, we need only the components in the field theory directions, which give
\begin{equation}
{1 \over z^3} \partial_z (z^3 \partial_z \Gamma_{\mu \nu}) + \partial_\rho \partial^\rho \Gamma_{\mu \nu} = 0 \; .
\end{equation}
The solution in our Fefferman-Graham coordinates with boundary behaviour (\ref{boundary}) is
\begin{equation}
\Gamma^{(1)}_{++}(x,z) = \lambda h_+(x^+) \qquad \Gamma^{(1)}_{--}(x,z) = \lambda h_-(x^-) \;
\end{equation}
with the linearized perturbation $\Gamma^{(1)}_{\mu \nu}$ independent of $z$.
\subsubsection*{Satisfying the gauge conditions}
We would now like to evaluate the metric contribution to (\ref{Stwo})
\begin{equation}
\label{S2grav}
\delta^{(2)}S_B^{grav} = - {1 \over 2} \int_\Sigma \omega^{grav}(\delta g_1, \lie{\xi_B} \delta g_1) \; .
\end{equation}
This formula assumes the gauge conditions (\ref{eq:gauge-condition}) which differ from the Fefferman-Graham gauge conditions we have been using so far. Thus, we must find a gauge transformation to bring our metric perturbation to the appropriate form. In general, we can write
\begin{align}
\gamma_{ab}=h_{ab} +(\lie{V}g)_{ab} = h_{ab} + \nabla_a V_b + \nabla_b V_a \; .
\end{align}
where $\gamma$ is the desired metric perturbation satisfying the gauge condition, and $h$ is the perturbation in Fefferman-Graham coordinates (equivalent to $\Gamma$ for $d=2$).
The procedure for finding an appropriate $V$ and evaluating (\ref{S2grav}) is described in detail in \cite{Lashkari:2015hha}, but we review the main points here. Defining coordinates $(X^A, X^i)$ so that the extremal surface lies at some fixed value of $X^A$ with $X^i$ describing coordinates along the surface, the gauge condition (\ref{eq:gauge-condition}) (equivalent to requiring that the coordinate location of the extremal surface remains fixed) gives
\begin{equation}
\label{eq:V-eqs0}
(\nabla_i \nabla^i V_A + [\nabla_i, \nabla_A] V^i + \nabla_i h^i_A - {1 \over 2} \nabla_A h^i {}_i)|_{\tilde{B}} = 0
\end{equation}
while the condition (\ref{eq:gauge-condition2}) that $\xi_B$ continues to satisfy the Killing equation at $\tilde{B}$ gives
\begin{align}
\left( h_{iA} + \nabla_i V_A + \nabla_A V_i \right) |_{\tilde B } &= 0 \, , \label{eq:V-eqs1}\\
\left( h^A_{\,\,D} - \frac{1}{2} \delta^A_{\,\,D} h^C_{\,\,C} + \nabla^A V_D + \nabla_D V^A - \delta^D_{\,\, D} \nabla_C V^C_{\,\,C} \right)\bigg|_{\tilde B} &= 0 \, .\label{eq:V-eqs2}
\end{align}
To solve these, we first expand our general metric perturbation in a Fourier basis.
\begin{equation}
\label{fourier}
h_{\mu\nu}(t,x, z) =\lambda\int \left[ \delta_\mu^+ \delta_{\nu}^+ \hat h_+(k) e^{i k x^+} + \delta_\mu^- \delta_{\nu}^- \hat h_-(k) e^{i k x^-} \right] d k\,,
\end{equation}
with a gauge choice $h_{za } (t,x,z)= 0$.
For each of the basis elements, we use the equations (\ref{eq:V-eqs0}), (\ref{eq:V-eqs1}) and (\ref{eq:V-eqs2}) to determine $V$ and its first derivatives at the surface $V$. For these calculations, it is useful to define polar coordinates $(z,x) = (r \cos \theta, r \sin \theta)$.
Since the gauge conditions are linear in $V$, the conditions on $V$ for a general perturbation are obtained from these by taking linear combinations as in (\ref{fourier}),
\begin{align}
V_{a} (t,x,z) & =\lambda\int \left[ \hat{V}_{a}^{+}(k) e^{i k x^+} + \hat{V}_{a}^{-}(k) e^{i k x^-} \right] dk \,\,.
\end{align}
After requiring $V_a$ remain finite at $\theta = \pm \frac{\pi}{2}$, we find
\begin{small}
\begin{align
\hat{V}_t^-(k;t,r,\theta) &=\frac{e^{-i k t}}{k^3 r^2 \cos^2\theta} \left(-i\cos(kr) + \sin\theta\sin(kr)
-i\frac{(k^2r^2\cos^2\theta-1)e^{i k r\sin\theta}}{2}\right) \nonumber \\
\hat{V}_r^-(k;t,r,\theta) &=\frac{e^{-i k t}}{k^3r^2\cos^2\theta}\Bigg( \sin(kr) - i\sin\theta\cos(kr) \nonumber \\
& \qquad\qquad\qquad-\frac{(k^2r^2\cos^2\theta\sin\theta+ikr\cos^2\theta+2i\sin\theta)e^{i k r \sin\theta}}{2}\Bigg) \nonumber\\
\partial_t\hat V^-_{\theta} (k;t,r,\theta)& =
\frac{e^{-ikt}}{2\,k^{2}\,r\cos\theta}
\left(
(2 + k^{2}r^{2}\cos^{2}\theta-2\,i k r\,\sin\theta)e^{ikr\sin\theta} - \frac{2\sin(kr)}{k^3r^2}
\right)\,\nonumber
\\
\partial_r \hat V^-_{\theta} (k;t,r,\theta)& =
\frac{e^{-ikt}}{k^{3}r^{2}\cos\theta}\Big(
2i\cos(kr) \nonumber\\
&
\quad+ \left[2kr\sin\theta+r^{3}k^{3}\sin\theta\cos^{2}\theta+i\left(r^{2}k^{2}\cos^{2}\theta-kr^{2}+2\right)\right]e^{ikr\sin\theta} \Big)\nonumber \\
\end{align}
where the $V^{\pm}$ solutions are related through $\hat V_{r}^{+}(k;t, r,\theta)=\hat V_{r}^{-}(k;-t,r,\theta)$ and $\hat V_{t}^{-}(k;t,r,\theta)=-\hat V_{t}^{+}(k;-t,r,\theta)$.
\end{small}
The results here give the behavior of $V$ and its derivatives only at the surface $\tilde{B}$ ($r=R$ in polar coordinates). Elsewhere, $V$ can be chosen arbitrarily, but we will see that our calculation only requires the behavior at $\tilde{B}$.
\subsubsection*{Evaluating the canonical energy}
Given the appropriate $V$, we can evaluate (\ref{S2grav}) using
\begin{eqnarray}
\label{omega}
\omega(g,\gamma,\lie{\xi} \gamma) &=& \omega(h + \lie{V} g, \lie{\xi_B} (h + \lie{V} g)) \\
&=& \omega(g,h,\lie{\xi} h) + \omega(g,h + \lie{V} g,\lie{[\xi,V]} g) - \omega(g,\lie{\xi} h, \lie{V} g) \nonumber
\end{eqnarray}
where
\begin{equation}
[\xi,V]^a = \xi^b \partial_b V^a - V^b \partial_b \xi^a \;
\end{equation}
and we have used that $\lie{\xi} g = 0$. We can simplify this expression using the gravitational identity
\begin{equation}
\label{dchi}
\omega(g, \gamma,\lie{\xi} g) = d \chi(\gamma,X)
\end{equation}
where
\begin{equation}
\label{defchi}
\chi(\gamma,X) = {1 \over 16 \pi G_N} \epsilon_{ab} \left\{\gamma^{ac} \nabla_c X^b - {1 \over 2} \gamma_c{}^c \nabla^a X^b + \nabla^b \gamma^a {}_c X^c - \nabla_c \gamma^{ac} X^b + \nabla^a \gamma^c {}_c X^b \right\} \; .
\end{equation}
Thus, we have
\begin{equation}
\label{defomega}
\omega(g,\gamma,\lie{\xi} \gamma) = \omega(g,h,\lie{\xi} h) + d \rho
\end{equation}
where
\begin{equation}
\label{defrho}
\rho = \chi(h + \lie{V} g,[\xi,V]) - \chi(\lie{\xi} h,V) \; .
\end{equation}
Finally, choosing $V$ so that it vanishes at $B$, we can rewrite (\ref{S2grav}) as
\begin{equation}
\label{calcE}
{\cal E} = \int_\Sigma \omega(g, h, \lie{\xi} h) + \int_{\tilde{B}} \rho(h,V) \; .
\end{equation}
In this final expression, we only need $V$ and its derivatives at the surface $\tilde{B}$.
Thus, we can now calculate the result explicitly for a general perturbation. In the Fourier basis, the final result in terms of the boundary stress tensor is
\begin{align}
\label{eq:k-canonical-energy}
\mathcal {E}=& \int dk_1 \! \int dk_2 \,
\hat K_2(k_1, k_2)\,
\langle T_{++}(k_1) \rangle \langle T_{++} (k_2) \rangle + \{+ \leftrightarrow - \}\,,
\end{align}
where the kernel is
\begin{small}
\begin{align}
\label{eq:canon-kernel}
\hat K_2(k_1, k_2) = \frac{256 \pi^2\, R^4 \,G_N}{\ell_{AdS} K^{3}(K-\kappa)^{3}(K+\kappa)^{3}}
&\left[(K^{5}-2\,(\kappa^{2}+4)K^{3} +\kappa^{4}K)\cos K
\right. \nonumber \\ & \left.
-(5K^{4}-6K^{2}\kappa^{2}+\kappa^{4})\sin K
+8\,K^{3}\cos\kappa \right] \;,
\end{align}
\end{small}
$\!\!$with $K\equiv R(k_1+k_2), \kappa \equiv R(k_1-k_2)$. We note in particular that the result splits into a left-moving part and a right-moving part with no cross term.
Transforming back to position space
\begin{align}
\mathcal{E} &= \int_{B'} dx_1^+ \! \int_{B'} dx_2^+ \, K_2(x_1^+, x_2^+) \, \lambda T_{++}(x_1^+) \ra \lambda T_{++} (x_2^+) \ra + \{+ \leftrightarrow - \}\,,
\end{align}
where the kernel $K_2$ is symmetric under exchange of $x_1^\pm$ and $x_2^\pm$, and has support only on $x_i^\pm \in[-R,R]$. Focusing only on the domain of support, we have
\begin{align}
K_2(x_1, x_2) &= \frac{4 \pi^2 G_N }{R^{2} \ell_{AdS}} \begin{cases}
(R-x_1)^2 (R+x_2)^2& x_1\geq x_2 \\
(R+x_1)^2 (R-x_2)^2 & x_1 < x_2
\end{cases} \,.
\label{eq:CanEnergyKernel}
\end{align}
Using the relation $c = 3 \ell_{AdS} / (2 G_N)$ between the CFT central charge and the gravity parameters, we recover the result (\ref{result2D}) from the introduction.
Like the leading order result in (\ref{Sexp}), the integrals can be taken over any surface $B'$ with boundary $\partial B$. The fact that we only need the stress tensor on a Cauchy surface for $D_B$ is special to the stress tensor in two dimensions, since the conservation relations allow us to find the stress tensor expectation value everywhere in $D_B$ from its value on a Cauchy surface. For other operators, or in higher dimensions, the result will involve integrals over the full domain of dependence. We will see an explicit example in the next subsection.
Positivity of relative entropy requires $\mathcal{E}$ to be positive which requires the kernel to be positive semi-definite. As we show in Appendix \ref{appendix:positivity}, one can demonstrate that the positivity explicitly, providing a check of our results. An alternative proof of positivity is given in Section \ref{Sec:EntanglementHolography}. As a more complete check, we will show in Section \ref{Sec:CFT2} that this result can be reproduced by a direct CFT calculation for the special class of states that can be obtained from the vacuum state by a local conformal transformation.
\subsection{Example: Scalar operator contribution}
\label{sec:scalar}
We now consider an explicit example making use of the bulk matter field term in (\ref{secondorder2}) in order to calculate the terms in the entanglement entropy formula quadratic in the scalar operator expectation values. The discussion for other matter fields would be entirely parallel. This example is more representative, since the formula will involve scalar field expectation values in the entire domain of dependence $D_B$, i.e. a boundary spacetime region rather than just a spatial slice. The results here are similar to the recent work in \cite{Faulkner:2014jva, Speranza:2016jwt, Faulkner:2015csl}, but we present them here to show that they follow directly from the canonical energy formula.
We suppose that the CFT has a scalar operator of dimension $\Delta$ with expectation value $\langle {\cal O}(x) \rangle$. According to the usual AdS/CFT dictionary, this corresponds to a bulk scalar field with mass $m^2 = \Delta(\Delta - d)$ and asymptotic behavior
\begin{equation}
\label{scalarboundary}
\phi(x,z) \to \gamma z^\Delta \langle {\cal O}(x) \rangle \; ,
\end{equation}
where $\gamma$ is a constant depending on the normalization of the operator ${\cal O}$. The leading effects of the bulk scalar field on the entanglement entropy (\ref{Stwo}) come from the matter term in the canonical energy
\begin{equation}
\label{StwoM}
\delta^{(2)} S^{matter}_B = \frac{1}{2} \int_\Sigma \xi^a_B T^{(2)}_{ab} \epsilon^b \; .
\end{equation}
Using the explicit form of $\xi_B$ from (\ref{defxi}) and $\epsilon$ from (\ref{defeps}), this gives (for a ball centered at the origin)
\begin{equation}
\label{StwoM1}
\delta^{(2)} S^{matter}_B= - \frac{\ell_{AdS}^{d-1}}{2} \int_0^R {dz \over z^{d-1}} \int_{x^2 < R^2 - z^2} d^{d-1} x {\pi \over R} (R^2 - z^2 - x^2) T^{(2)}_{00}(x,z) \; .
\end{equation}
This expression is valid for a general bulk matter field. For a scalar field, we have
\begin{equation}
T^{(2)}_{ab} = \partial_a \phi_1 \partial_b \phi_1 - {1 \over 2} g_{ab} (g^{cd} \partial_c \phi_1 \partial_d \phi_1 + m^2 \phi_1^2) \;,
\end{equation}
where $g_{ab}$ is the background AdS metric and $\phi_1$ represents the solution of the linearized scalar field equation on AdS,
\begin{equation}
\label{scalareq}
{1 \over z^{d-1}} \partial_z \left\{ z^{d-1} \partial_z \phi \right\} + \partial_\mu \partial^\mu \phi - {m^2 \over z^2} \phi = 0\;,
\end{equation}
with boundary behavior as in (\ref{scalarboundary}). This solution is given most simply in Fourier space, where we have
\begin{equation}
\label{BBfour}
\phi_1(k,z) = \frac{2^\nu \Gamma(\nu +1)}{(2\pi)^d }\int_{k_0^2 > \vec{k}^2 } d^d k \frac{e^{i k_\mu x^\mu}}{\left(k_0^2 - \vec k^2\right)^{\nu/2}} \, z^{d \over 2} J_{\nu}\left(\sqrt{k_0^2 - \vec{k}^2} z \right) \, \gamma \langle {\cal O}(k) \rangle \; ,
\end{equation}
where $\nu = \Delta - d/2$, but we can formally write a position-space expression using a bulk-to-boundary propagator $K(x, z ; x')$ as~\cite{Hamilton:2006az, Hamilton:2006fh}
\begin{equation}
\label{BBpos}
\phi_1(x,z) = \gamma \int dx' K(x, z ; x') \langle {\cal O}(x') \rangle \; .
\end{equation}
The integral here is over the boundary spacetime, however it has been argued (see, for example \cite{Almheiri:2014lwa,Morrison:2014jha}) that to reconstruct the bulk field throughout the Rindler wedge $R_B$ (and specifically on $\Sigma$), we need only the boundary values on the domain of dependence region. We recall some explicit formulae for this ``Rindler bulk reconstruction'' in Appendix \ref{app:Rindler_reconstruct}. Combining these results, we have a general expression for the scalar field contribution to entanglement entropy at second-order in the scalar one-point functions,
\begin{eqnarray}
\label{StwoM2}
\delta^{(2)} S^{scalar}_B &=& -\frac{\ell_{AdS}^{d-1}}{2} \int_0^R {dz \over z^{d-1}} \int_{x^2 < R^2 - z^2} d^{d-1} x {\pi \over R} (R^2 - z^2 - x^2) \\
&& \qquad \qquad \qquad \qquad \left\{(\partial_0 \phi_1)^2 + (\partial_i \phi_1)^2 + (\partial_z \phi_1)^2 + {m^2 \over z^2} \phi_1^2 \right\} \nonumber
\end{eqnarray}
where $\phi_1$ is given in (\ref{BBfour}) or (\ref{BBpos}) .
As a simple example, consider the case where the scalar field expectation value is constant. In this case it is simple to solve (\ref{scalareq}) everywhere to find that
\begin{equation}
\phi_1(x,z) = \gamma \langle {\cal O} \rangle z^\Delta \; .
\end{equation}
Inserting this into the general formula (\ref{StwoM2}), and performing the integrals, we obtain
\begin{equation}
\delta^{(2)} S^{scalar}_B = -{\pi \ell_{AdS}^{d-1} \over 4 } \gamma^2 \langle {\cal O} \rangle^2 R^{2 \Delta} \Omega_{d-2} {\Delta \Gamma({d \over 2} - {1 \over 2}) \Gamma( \Delta - {d \over 2} + 1) \over \Gamma(\Delta + {3 \over 2})}\,.
\end{equation}
This reproduces previous results in the literature~\cite{Blanco:2013joa, Speranza:2016jwt}.
\section{Stress tensor contribution: direct calculation for CFT$_2$}
\label{Sec:CFT2}
In Section \ref{SubSec:CFT2stress}, we used the equivalence between quantum Fisher information and canonical energy to obtain an explicit expression for the second-order stress tensor contribution to the entanglement entropy for holographic states in two-dimensional CFTs. This is applicable for general holographic states, whether or not other matter fields are present in the dual spacetime (in which case there are additional terms in the expression for entanglement entropy). In special cases where there are no matter fields, the spacetime is locally AdS and we can understand the dual CFT state as being related to the vacuum state by a local conformal transformation. We show in this section that in this special case, we can reproduce the holographic result (\ref{eq:CanEnergyKernel}) through a direct CFT calculation, providing a strong consistency check. We note that the result does not rely on taking the large $N$ limit or on any special properties of the CFT, so the formula holds universally for this simple class of states.
Our approach will be to develop an iterative procedure to express the entanglement entropy as an expansion in the stress tensor expectation value for this special class of states.
We evaluate the entanglement entropy for these states from a correlation function of twist operators obtained by transforming the result for the vacuum state.\footnote{A similar approach was recently used to derive the modular Hamiltonian of these excited states in~\cite{Lashkari:2015dia}.} Similarly, the stress tensor expectation values follow directly from the form of the conformal transformation. Inverting the relationship between the required conformal transformation and the stress tensor expectation value allows us to express the entanglement entropy as a perturbative expansion in the expectation value of the stress tensor. Similar CFT calculations have also been used recently in \cite{helleretal}.
\subsection{Conformal transformations of the vacuum state}
In two-dimensional CFT, under a conformal transformation $w = f(z)$, the stress tensor transforms as
\begin{eqnarray}
T'(w) = \left(\frac{d w}{d z}\right)^{-2} \left( T (z) + \frac{c}{12} \{f(z); z\} \right)\,.
\end{eqnarray}
Here $c$ is the central charge of the CFT and the inhomogeneous part is the Schwarzian derivative
\begin{equation}
\{f(z); z\} \equiv \frac{f'''(z)}{f'(z)} -\frac{3 f''(z)^2}{2f'(z)^2}\,.
\end{equation}
For an infinitesimal transformation $f(z) = z + \lambda\, \epsilon(z)$,
the Schwarzian derivative can be expanded as
\begin{small}
\begin{equation}
\{z+ \lambda \epsilon(z) ; z \} = \lambda\, \epsilon'''(z) -\lambda^2 \left( \epsilon'''(z) \epsilon'(z) + \frac{3}{2} \epsilon''(z)^2 \right) + \lambda^3\left( \epsilon'(z)^2 \epsilon'''(z)+ 3\epsilon'(z) \epsilon''(z)^2 \right) + \cdots
\label{eq:PerturbativeSchwarzian}
\end{equation}
\end{small}
$\!\!$The CFT vacuum is invariant under the $SL(2 ,\mathbb{C})$ subgroup of global conformal transformations.
However, for transformations which are not part of this subgroup, the vacuum state transforms into \emph{excited states}. The action of the full conformal group includes the full Virasoro algebra which involves arbitrary products and derivatives of the stress tensor
\begin{equation}
\text{Id} \sim 1, T, \partial^m T, T^2, T\partial^n T, \cdots \,.
\end{equation}
These states capture the gravitational sector of the gravity dual. Other excited states can be obtained by the action of other primary operators and their descendants.
However we restrict ourselves to the class states that are related to `pure gravity' excitations, which are the states obtained by conformal transformation of the vacuum state.
We denote the excited state as
$| f \rangle = U_f \,| 0 \rangle$
where $U_f$ is the action of a conformal transformation on the vacuum $|0 \rangle$.
The expectation value of the stress tensor for the state perturbed state $|f\rangle$ is
\begin{eqnarray}
\langle f | T(z) | f \rangle = \langle 0 | U_f^\dagger \, T(z) \,U_f | 0\rangle =\langle 0 | T'(w) | 0\rangle = \left(\frac{df}{dz}\right)^{-2} \frac{c}{12} \{f(z) ; z\} \,,
\label{expT}
\end{eqnarray}
where we used that $ \langle 0| T(z)|0 \rangle =0$. The anti-holomorphic component of the stress tensor $\bar T(\bar z)$ is similarly related to the anti-holomophic part of the conformal transformation $\bar f$.
To leading order in a conformal transformation near the identity, this equation relates the conformal transformation to $\langle T(z) \rangle$ by a third-order ordinary differential equation. The three integration constants correspond to the invariance of $\langle T(z) \rangle$ under the global conformal transformations. Thus we have an invertible relationship between the conformal transformations modulo their global part and $\langle T(z) \rangle$, at least near the identity.
\subsection{Entanglement entropy of excited states }
In a two-dimensional CFT, the entanglement entropy can be explicitly computed using the replica method~\cite{Holzhey:1994we, Calabrese:2004eu}.
The computation can be reduced to a correlation function of twist operators $\Phi_\pm$, which are conformal primaries with weight $(h_n, \bar h_n) = \frac{c}{24}(n -1/n, n -1/n)$.
The R\'enyi entropy is
\begin{equation}
\exp\left((1-n)S^{(n)} \right) = \langle \Phi_+ (z_1) \Phi_- (z_2) \rangle = (z_2 - z_1)^{-2 h_n}\,.
\end{equation}
The entanglement entropy is obtained by taking the $n\rightarrow 1 $ limit of $S^{(n)}$.
\begin{equation}
S_{\text{vac}} = \lim_{n \rightarrow 1} S^{(n)} = \lim_{n \rightarrow 1} \, (1-n)^{-1} \log (z_2 -z_1)^{-2h_n} = \frac{c}{12} \log \frac{(z_2 - z_1)^2}{\delta^2} \,.
\end{equation}
For the excited states obtained by conformal transformations $z \rightarrow w = f(z)$\, the R\'enyi entropy is
\begin{eqnarray}
\exp\left((1-n)S_{\text{ex}}^{(n)} \right) &=& \langle f | \Phi_+ (z_1) \Phi_- (z_2) | f \rangle\\
&=&
\left(\frac{df}{dz}\right)^{-h_n}_{z_1} \left(\frac{df}{dz}\right)^{-h_n}_{z_2}
\left(\frac{d\bar f}{d\bar z}\right)^{-\bar h_n}_{\bar z_1}
\left(\frac{d\bar f}{d\bar z}\right)^{-\bar h_n}_{\bar z_2}
\langle 0 | \Phi_+ (z_1) \Phi_- (z_2) | 0 \rangle\,. \nonumber \\
\end{eqnarray}
Here $z_1, z_2$ are the points $f(z_1) = \bar f (\bar z_1) = -R$, $f(z_2) = \bar f(\bar z_2) = R$.
The entanglement entropy of the excited state is
\begin{equation}
S_{\text{ex}} = \lim_{n \rightarrow 1} S_{\text{ex}}^{(n)}
= \frac{c }{12} \log \left| \frac{
f'(z_1) f'(z_2) \bar f'(\bar z_1) \bar f'(\bar z_2)
(z_2 -z_1)^2 }{\delta^2} \right| \,.
\end{equation}
Therefore the change in entanglement entropy respect to the vacuum state is
\begin{align}
\label{deltaSEE}
\delta S \equiv S_{\text{ex}} - S_{\text{vac}}
=& \frac{c}{12} \log \left| \frac{f'(f^{-1}(R)) f'(f^{-1}(-R)) (f^{-1}(R) -f^{-1}(-R))^2 }{(2R)^2} \right| \\
&+ \frac{c}{12} \log \left| \frac{\bar f'(\bar f^{-1}(R)) \bar f'(\bar f^{-1}(-R)) (\bar f^{-1}(R) -\bar f^{-1}(-R))^2 }{(2R)^2} \right|
\,. \nonumber
\end{align}
By inverting \eqref{expT}, the conformal transformation required to reach the state $| f \rangle$ can be expressed as a function of the expectation value of the stress tensor. Plugging this $f$ into \eqref{deltaSEE}, allows us to express the entanglement entropy as a function of the expectation value of the stress tensor alone, as we set out to do.
In practice, we will invert \eqref{expT} order by order in a small conformal transformation and express the entanglement entropy as an expansion in the resulting small stress tensor. The second-order term in this expansion will be the Fisher information metric.
In the following, we will focus on the holomorphic term in \eqref{expT}, noting that the anti-holomorphic part follows identically.\footnote{Note that the potential cross-term between left and right moving contributions vanished in the gravitational computation of $\delta^{(2)} S$.}
\subsection{Perturbative expansion}
Consider a conformal transformation perturbation near the identity transformation
\begin{equation}
w = f(z) = z + \lambda f_1 (z) + \lambda^2 f_2(z) + \lambda^3 f_3(z) + \cdots\,,
\end{equation}
where $\lambda$ is a small expansion parameter.
In this expansion,
\begin{small}
\begin{equation}
\frac{12}{c} \, \langle T(w) \rangle = \lambda\, f_1'''(w) + \lambda^2 \left( -\frac{3}{2}f_1''(w)^2 -3f_1'(w)f_1'''(w)+f_2'''(w)-f_1(w)f_1''''(w)\right) + {\cal O}(\lambda^3)\,,
\label{stressPerturbation}
\end{equation}
\end{small}
$\!\!$and the entanglement entropy is
\begin{small}
\begin{align}
\frac{12}{c} \, S_{\text{ex}} =& \log \left| \frac{f'(z_1) f'(z_2)(z_2 - z_1)^2}{\delta^2} \right |\nonumber \\
=& \log \frac{ (2R)^2}{\delta^2} +\lambda\left[\frac{R \left(f_1'(-R)+f_1'(R)\right)+f_1(-R)-f_1(R)}{R}\right]\nonumber \\
&+ \lambda^2 \Big(-\frac{(f_1(R)-f_1(-R))^2}{4R^2} + \frac{-f_1(-R) f_1'(-R)+f_1(R) f_1'(R)+f_2(-R)-f_2(R)}{R} \nonumber \\
& \phantom{+ \lambda^2 ( ( }-\frac{1}{2} f_1'(-R){}^2-\frac{1}{2} f_1'(R){}^2+f_2'(-R)+f_2'(R)-f_1(-R) f_1''(-R)-f_1(R) f_1''(R) \Big) \nonumber \\
&+ \,{\cal O}(\lambda^3) \,. \label{perturbedSEE}
\end{align}
\end{small}
\subsubsection*{Linear order}
To first-order in $\lambda$, the stress tensor is given by
\begin{equation}
\langle T(z) \rangle = \lambda \frac{c}{12} \, f_1 '''(z) + {\cal O}(\lambda^2) \,,
\end{equation}
so that change in the expectation value of the modular Hamiltonian becomes
\begin{eqnarray}
\delta \langle H_B \rangle &=& \frac{\lambda\, c}{24 R} \int_{-R}^{R} dz \, (R^2 - z^2) f_1'''(z) \nonumber \\
&=& \frac{\lambda\, c }{24R} \left[ (R^2-z^2) f_1''(z) + 2\left( z f_1'(z) - f_1(z) \right) \right]^{R}_{-R} \nonumber \\
&=& \frac{\lambda\, c }{12 R} \left[R( f_1'(R) + f_1'(-R)) - ( f_1(R) - f_1(-R) )\right] \,.
\label{eq:ModHChange}
\end{eqnarray}
From \eqref{deltaSEE} we also have that the first-order change in entanglement entropy is
\begin{eqnarray}
\delta^{(1)} S
= \frac{\lambda\, c }{12 R} \left[R( f_1'(R) + f_1'(-R)) - ( f_1(R) - f_1(-R) )\right] \,.
\end{eqnarray}
Comparing with \eqref{eq:ModHChange} we see that the first law of entanglement holds
\begin{equation}
\delta^{(1)} S = \delta \langle H_B \rangle \,.
\end{equation}
\subsubsection*{Second-order}
The second-order change in entanglement entropy gives the second-order relative entropy as the modular Hamiltonian is linear in the expectation value of the stress tensor. This is the quantum Fisher metric in the state space, which is dual to the canonical energy in gravity~\cite{Lashkari:2015hha}. In this section, we obtain the expression for canonical energy from the CFT side and find an exact match to the results of Section \ref{SubSec:CFT2stress}.
Our procedure so far yields the entanglement entropy of a subregion in terms of a perturbative expansion in small stress tensor expectation value
\begin{eqnarray}
\delta S = && \int_B \frac{d z}{2\pi} \, K_1 (z) \langle T(z) \rangle - \frac{1}{2} \int_B \frac{d z_1}{2\pi} \, \int_B \frac{d z_2}{2\pi} \, K_2 (z_1, z_2) \langle T(z_1) \rangle\langle T(z_2) \rangle +\cdots \nonumber \\
&&+ \, \{ z \leftrightarrow \bar z\} \,.
\label{eq:SEEKernelCFT}
\end{eqnarray}
To obtain $K_2(z_1,z_2)$, we need to invert the relationship in \eqref{stressPerturbation} order by order, the lower order solutions $f_{i-1}, f_{i-2}, \cdots f_1$ becoming sources for the $i$-th order solution.
Taking the explicit expression for $\langle T (z) \rangle$ to simplify solving the differential equations,
\begin{equation}
\langle T (z) \rangle = \lambda \, \left( c_1 e^{i k_1 z } + c_2 e^{i k_2 z } \right)\,,
\end{equation}
is sufficient to extract the Fourier transformed kernel.
The first-order solution is
\begin{equation}
f_1(z) = F_1 + F_2 z + F_3 z^2 + \frac{12 i}{c}\left( \,c_1 \frac{e^{i k_1 z}}{k_1^3}+ \,c_2 \frac{e^{i k_2 z}}{k_2^3}\right)\,,
\end{equation}
where $F_i$ are constants that corresponds to the global part of the conformal transformation and do not effect the final result. We take these constants to be zero for simplicity.
The second-order solution is
\begin{equation}
f_2(z) = -\frac{9}{c^2}\left[ \frac{ 11 i}{16} (c_1^2 \frac{e^{2 i k_1 z } }{k_1^5} + c_ 2^2 \frac{e^{2 i k_2 z}}{k_2^5} ) + i \frac{ c_1 c_2 }{ k_1^3 k_2^3}\frac{ e^{i(k_1+k_2) z} \left(k_1^4+3 k_2 k_1^3+3 k_2^2 k_1^2+3 k_2^3 k_1+k_2^4\right)}{(k_1 +k_2)^3} \right] \,.
\end{equation}
With these solutions, we obtain
\begin{eqnarray}
\tilde K_1 ( k) =&& \frac{2}{k^2} \, \frac{\sin \left(k R\right)-k R \cos \left(k R\right)}{k R}\,,
\end{eqnarray}
as well as
\begin{small}
\begin{align}
\label{eq:CFTk-kernel}
\tilde K_2(k_1, k_2) &=\frac{96 R^4}{c} \frac{ (K^{5}-2(\kappa^{2}+4)K^{3}+\kappa^{4}K)\cos K
-(5K^{4}-6K^{2}\kappa^{2}+\kappa^{4})\sin K+8K^{3}\cos\kappa}{ K^{3}(K-\kappa)^{3}(K+\kappa)^{3}} \,,
\end{align}
\end{small}
$\!\!$with $K \equiv R(k_1 + k_2)$ and $ \kappa \equiv R(k_1 - k_2) $.
Taking the inverse Fourier transformation of $\tilde K_1 ( k) $
\begin{eqnarray}
K_1(z) &=& \int d k \, \tilde K_1 ( k) e^{-i k z}
=\pi \frac{ R^2-z^2 }{R} W(R, z)
\end{eqnarray}
where
\begin{equation}
W(R, x) \equiv \frac{\left(\text{sgn}\left(R+x\right)+ \text{sgn}\left(R-x\right)\right)}{2}
\end{equation}
is a window function with support $x\in [-R,R]$.
The second-order position space kernel is
\begin{align}
K_2(z_1, z_2)
&=\frac{6 \pi ^2 }{c R^2} \begin{cases}
(R-z_1)^2 (R+z_2)^2& -R \leq z_2\leq z_1 \leq R\\
(R+z_1)^2 (R-z_2)^2 & -R \leq z_1 < z_2 \leq R
\end{cases} \,.
\label{eq:secondKernel}
\end{align}
The anti-holomorphic part is the same with $z\rightarrow\bar z$, and the cross term vanishes. With the relation
\begin{equation}
c = \frac{3\ell_{AdS}}{2 G_N}
\label{eq:BrownHenneauxC}
\end{equation}
this reproduces the kernel for canonical energy in \eqref{eq:CanEnergyKernel}.
This result holds for regions defined on any spatial slice of the CFT. If we choose the $t=0$ slice, $z=\bar z=x$ and our result becomes
\begin{eqnarray}
\delta S^{(2)} _{EE} &=& - \frac{1}{2} \int_B d x_1 \, \int_B d x_2 \, K_2 (x_1, x_2) \left[ \langle T_{++}(x_1)\rangle \langle T_{++}(x_2)\rangle
+ \langle T_{--}(x_1)\rangle \langle T_{--}(x_2)\rangle \right] \,.\nonumber
\end{eqnarray}
Changing variables using $x_1=x-r$, $x_2=x+r$, the kernel is simply
\begin{align}
K_2(x,r)=K_2(x,-r)=\frac{12 \pi ^2 }{c R^2} \left[ (R-|r|)^2 -x^2 \right]^2 \Theta\left(R-|r|-|x|\right).
\end{align}
\subsection{Excited states around thermal background}
\label{sec:thermal}
A similar analysis can be applied to perturbations around a thermal state with temperature $ T = \beta^{-1}$. If we denote homogeneous thermal state $| \beta \rangle$, the stress tensor one-point function is
\begin{equation}
\langle \beta | T | \beta \rangle = \frac{\pi^2 c}{6 \beta^2} \,.
\end{equation}
This can be obtained by a conformal transformation from the vacuum with
\begin{equation}
f_\beta ( z ) = \frac{\beta}{2 \pi} \log(z)\,.
\label{eq:ConfTransfToThermal}
\end{equation}
On top of this transformation, one could also apply an infinitesimal conformal transformation to obtain non-homogeneous perturbation around thermal state.
A similar computation as the previous section leads to the first-order kernel
\begin{equation}
K_1^\beta (z) = \frac{2 \beta}{\sinh(\tfrac{2\pi R}{\beta})} \sinh\left(\frac{\pi(R-z)}{\beta}\right) \sinh\left(\frac{\pi(R+z)}{\beta}\right) \,,
\end{equation}
which is the modular hamiltonian of thermal state in 2d CFT.
Furthermore, the second-order kernel is
\begin{align}
K^\beta_2(z_1, z_2)=\frac{24 \beta^2 }{c \, \sinh^2(\tfrac{2\pi R}{\beta})} \begin{cases}
\sinh^2\left(\frac{\pi(R-z_1)}{\beta}\right) \sinh^2\left(\frac{\pi(R+z_2)}{\beta}\right) & -R\leq z_2 \leq z_1 \leq R\\
\sinh^2\left(\frac{\pi(R+z_1)}{\beta}\right) \sinh^2\left(\frac{\pi(R-z_2)}{\beta}\right) & -R\leq z_1 < z_2 \leq R
\end{cases} \,.
\label{secondThermalKernel}
\end{align}
\subsection*{Consistency check : homogeneous BTZ perturbation}
As a check, consider the homogeneous perturbation example, where $\langle T \rangle=\langle \bar T\rangle = \frac{\lambda}{8 G_N}$.\footnote{$\lambda = \frac{2\pi^2}{\beta}$ sets the temperature.}
In AdS$_3$ this is a perturbation towards the planar BTZ geometry
\begin{equation}
ds^2 = \frac{1}{z^2} \left(dz^2 + (1+\lambda z^2/2)^2 dx^2 - (1-\lambda z^2/2)^2 dt^2 \right)
\end{equation}
in Fefferman-Graham coordinates.
Holographic renormalization \eqref{eq:HoloRenT} tells us the stress tensor expectation value of the dual CFT is
\begin{equation}
\langle T_{tt}\rangle= \frac{1}{2\pi} \left( \langle T \rangle + \langle \bar T \rangle \right) = \frac{\lambda}{8 \pi G_N}\,.
\end{equation}
As the black hole corresponds to the thermal state in CFT, the dual state be obtained by the conformal transformation~\eqref{eq:ConfTransfToThermal}.
First, applying \eqref{deltaSEE} for this conformal transformation, the change in entanglement entropy with respect to the vacuum is
\begin{equation}
\delta S =\lambda \frac{ R^2 }{6G} - \lambda^2 \frac{R^4 }{90 G} + \lambda^3 \frac{4 R^6 }{2835 G} + \mathcal{O}(\lambda^4)\,,
\end{equation}
which matches the previous known results \cite{Blanco:2013joa, Lashkari:2015hha}.
The linear order equals $\delta \langle H_B \rangle$ as expected from the entanglement first law.
The second-order term gives the quantum Fisher information or the canonical energy
\begin{align}
{\cal E} =\frac{d^2}{d\lambda^2}(\Delta E - \Delta S) \Big |_{\lambda =0 } = \frac{R^4}{45G_N}\,.
\end{align}
Using the formula using the second-order kernel \eqref{eq:SEEKernelCFT} and \eqref{eq:secondKernel},
we obtain the same canonical energy
\begin{eqnarray}
{\cal E} &=& 2\frac{d^2}{d\lambda^2} \left[\frac{1}{2} \int_B \frac{d z_1}{2 \pi} \, \int_B \frac{d z_2}{2\pi} \, K_2 (x_1, x_2) \langle T\rangle \langle T\rangle \right]_{\lambda =0 } = \frac{R^4}{45G_N}\,.
\end{eqnarray}
\section{Auxiliary de Sitter space interpretation}
\label{Sec:EntanglementHolography}
In \cite{deBoer:2015kda}, it was pointed out that the leading order perturbative expression (\ref{firstorder}) for entanglement entropy, expressed as a function of the center point $x$ and radius $R$ of the ball $B$, is a solution to the wave equation for a free scalar field on an auxiliary de Sitter space, with $\langle T_{00}(x) \rangle$ acting as a source.
It was conjectured that higher order contributions might be accounted for by local propagation in this auxiliary space with the addition of self-interactions for scalar field.
In this section, we show that for two-dimensional CFTs, the second-order result (\ref{result2D}) can indeed be reproduced by moving to a non-linear wave equation with a simple cubic interaction to this scalar field. A slight complication is that we actually require two-scalar fields; one sourced by the holomorphic stress tensor $T_{++}$, and the other sourced by the anti-holomorphic part $T_{--}$; the perturbation to the entanglement entropy is then the sum of these two scalars, $\delta S = \delta S_+ + \delta S_-$, reproducing both terms in (\ref{result2D}).
We will focus on $\delta S_+$ since $\delta S_-$ follows identically.
To reproduce the second-order results for entanglement entropy, we consider an auxiliary de Sitter space with metric
\begin{equation}
ds^2_{dS} = \frac{L_{{dS}}^2}{R^2}\left(-dR^2 + d x^2\right)\,.
\end{equation}
and consider a scalar field $\delta S_+$ with mass $m^2 L_{{dS}}^2= -2 $ and action
\begin{align}
{\cal L } = \frac{1}{2} \nabla_a \left( \delta S_+\right) \nabla^a \left(\delta S_+ \right)
+ \frac{1}{2} m^2 \left(\delta S_+\right)^2
+ \frac{4}{c L_{{dS}}^2} \left(\delta S_+\right)^3 \,.
\label{eq:dSScalarAction}
\end{align}
The equation of motion is
\begin{equation}
\left( \nabla_{{dS}}^2- m^2 \right) \delta S_{+} (R, x) = \frac{12}{c L_{{dS}}^2} \left( \delta S_{+} (R, x) \right)^2\,.
\label{eq:dSwaveEqInteraction}
\end{equation}
As shown in \cite{deBoer:2015kda}, the first-order perturbation (\ref{firstorder}) obeys the linearized wave equation
\begin{equation}
\left( \nabla_{{dS}}^2- m^2 \right) \delta^{(1)} S_+ (R, x) = 0\,.
\end{equation}
We can immediately check that the second-order perturbation (\ref{result2D}) is consistent with the nonlinear equation by acting with the dS wave equation on the second-order kernel \eqref{eq:secondKernel}
\begin{equation}
\left( \nabla_{{dS}}^2 - m^2 \right) K_2 (x_1 -x, x_2- x) = - \frac{24}{c L_{{dS}}^2} K_1 ( x_1 - x) K_1 ( x_2 - x) \,.
\end{equation}
Integration against the CFT stress tensor then gives \eqref{eq:dSwaveEqInteraction}.
Alternatively, introducing the retarded\footnote{These propagators are defined to be non-zero only within the future directed light-cone. This is important in reproducing both the support and the exact form of $K_2(x_1,x_2)$. } bulk-to-bulk propagator~\cite{Xiao:2014uea}
\begin{eqnarray}
G_{dS} (\eta , x ; \eta', x \, ' ) = - \frac{\eta^2 + \eta\,'^2 -( x- x \,')^2}{4 \eta \eta'}
\label{eq:dSbulktobulk}
\end{eqnarray}
and bulk-to-boundary propagator
\begin{eqnarray}
K_{dS} (\eta , x; x \, ' ) &=& \lim_{ \epsilon \rightarrow 0} \left [ -4 \pi \epsilon \, \lim_{\eta'\rightarrow \epsilon}G_{dS} (\eta , x ; \eta', x \, ' ) \right] = \pi \frac{\eta^2 - \left( x - x'\right)^2}{ \eta } \,,
\label{eq:dSboundarytobulk}
\end{eqnarray}
we can show directly that the solution with boundary behavior
\begin{align}
\delta S_+ = \frac{4 \pi}{3} \langle T_{++} \rangle R^2 + {\cal O}(R^3)\,.
\end{align}
For $ R \rightarrow 0$ gives
\begin{small}
\begin{eqnarray}
\delta^{(1)} S_{+} (R, x_0) = \int dx \, K_{dS} (R, x_0; x) \langle T_{++} (x) \rangle
\end{eqnarray}
at first-order and
\begin{eqnarray}
\delta^{(2)} S_{+} (R, x_0) =
\frac{12}{ c L_{dS}^2} \int_{dS} d\eta' dx' \sqrt{|g_{dS}|} \, G_{dS} (R,x_0; \eta' ,x')
\left(\int dx \, K_{dS} (\eta' ,x';x) \langle T_{++} ( x) \rangle \right)^2 \, ,\nonumber \\
\label{EEfromdSTT}
\end{eqnarray}
\end{small}
$\!\!$at second-order, where the latter term comes from the diagram shown in Figure \ref{fig:ds-propagator}.
\begin{figure}[h!]
\centering
\includegraphics{ds-poincare-propogator.pdf}
\caption{Feynman diagram which computes $\delta^{(2)} S$. The $\delta S_+$ field propagates in de Sitter with a cubic interaction given by \eqref{eq:dSScalarAction}. The bold (red) line is the conformal boundary of de Sitter which is identified with a time slice of the CFT. $\delta S_+$ is sourced by the CFT stress tensor on this boundary. }
\label{fig:ds-propagator}
\end{figure}
The integrals can be performed directly to show that these results match with the expressions (\ref{firstorder}) and (\ref{result2D}) respectively.
A useful advantage of writing the second-order result in the form (\ref{EEfromdSTT}) is that it is manifestly negative. More explicitly, we have
\begin{eqnarray}
\delta^{(2)} S_+ (R, x_0) &=&
- \frac{3}{c L_{dS}^2} \int d\eta dy \sqrt{|g_{{dS}}|} \, \frac{R^2 + \eta^2 -( x_0 - y)^2}{R \, \eta}
\left[ \int_{B_y} dx \, K_{{{dS}}} (\eta , y ;x) \langle T_{++} ( x) \rangle \right]^2 \,. \nonumber \\
\end{eqnarray}
where $\sqrt{|g_{{dS}}|}$ and the squared expression are manifestly positive and
the bulk-to-bulk propagator \eqref{eq:dSbulktobulk} is positive over the range of integration where $(y-x_0)^2 \leq (R-\eta)^2$. That this expression is negative is required by the positivity of relative entropy, since we showed above that $-\delta^{(2)} S$ represents the leading order perturbative expression for the relative entropy.
Recently, it has been realized that the modular Hamiltonian in certain non-vacuum states in two dimensional CFTs can be described by propagation in a dual geometry \cite{Asplund:2016koz} matching the kinematic space found previously in \cite{Czech:2014ppa,Czech:2015qta,Czech:2015xna,Czech:2015kbp}. We find that the results of Section \ref{sec:thermal} can be explained by the same interacting theory \eqref{eq:dSScalarAction} on this kinematic space. The kinematic space dual to the thermal state is
\footnote{The kinematic space dual to the BTZ black hole was first described in \cite{Czech:2014ppa,Czech:2015qta}. The explicit form of the metric in the coordinates we are using can be found in \cite{Asplund:2016koz}.}
\begin{align}
ds^2 = \frac{4 \pi^2 L_{dS}^2}{\beta^2 \sinh^2 \left( \frac{2 \pi R}{\beta} \right)} \left(-d R^2 + dx^2\right)\,.
\end{align}
The second-order perturbation to the entanglement entropy from \eqref{secondThermalKernel} obeys the wave equation \eqref{eq:dSwaveEqInteraction}
with the same interactions in this kinematic space.
We could imagine adding additional fields propagating in de Sitter to capture the contributions to the entanglement entropy from scalar operators discussed in Section \ref{sec:scalar}. However, unlike the contribution from the stress tensor, this contribution involves integration of the one-point functions over the full domain of dependence $D_B$. In higher-dimensions, this will also be true for the stress tensor contribution. The $R=0$ boundary of the auxiliary de Sitter space does not include the time direction of the CFT, so any extension of these results to contributions of other operators or higher dimensional cases will require a more sophisticated auxiliary space. Promising work in this direction is discussed in \cite{Czech:2016xec,helleretal}.
\section*{Acknowledgements}
We thank Michal Heller, Ali Izadi Rad, Nima Lashkari, Don Marolf, Robert Myers, and Philippe Sabella-Garnier for helpful discussions.
This research is supported in part by the Natural Sciences and Engineering Research Council of Canada, and by grant 376206 from the Simons Foundation.
|
\section{Introduction} %
The fluctuating electromagnetic field has become a research area of increasing interest and importance, giving rise to phenomena like Casimir or van der Waals interactions \cite{Casimir48,Milonni} and thermal radiation at far and near fields \cite{Rytov3,Polder71,Eckhardt1984,kubo2012statistical}. Also, the rapid development concerning experimental detection and manipulation, including the framework of MEMS, allows exploration of effects down to the nanoscale \cite{lamoreaux1997demonstration,mohideen1998precision,bressi2002measurement,kittel2005near,rousseau2009radiative,shen2009surface,gad2001mems}. In general, fluctuational electrodynamics has been successfully applied to situations in thermal equilibrium, but also to objects in relative motion or at different temperatures \cite{Pendry97,Polder71,Antezza08}. However, such setups have mostly been considered for media described by linear electric and magnetic responses.
The field of nonlinear optics is by itself a growing and fundamentally interesting field, comprising, among others, frequency mixing processes, the optical Kerr effect, Brillouin scattering and Raman effects \cite{boyd2003nonlinear}. Especially considering recent developments concerning metamaterials, where large nonlinear response functions are observed, promising novel materials with interesting and useful properties. Examples include media infused with nanoparticles \cite{fukumi1994gold}, organic materials \cite{carter1985time} or polymers \cite{kuebler2000large}.
Fluctuations in nonlinear systems have been investigated for more than 50 years (mostly for classical systems) \cite{van1958thermal,van1965fluctuations}, also in interacting field theories (see e.g. \cite{kardar2007statistical,chang1975quantum}) and applied to critical Casimir forces \cite{krech1994casimir}. Regarding the fluctuating electromagnetic field, short-range enhancement of van der Waals forces have been predicted \cite{makhnovets2016short} and Casimir forces for systems with nonlinear boundary conditions \cite{fosco2015vacuum} and nonlinear coupling functions \cite{kheirandish2011finite} have also been studied. Fluctuations have also been considered in nonlinear optical cavities \cite{drummond1980quantum}, with intriguing effects regarding heat radiation, studied in the Langevin framework \cite{khandekar2015radiative}. Yet, the direct combination of nonlinear optics and fluctuational electrodynamics is missing in the literature.
In this Letter, we develop fluctuational electrodynamics for systems with nonlinear optical response. Starting from the stochastic nonlinear Helmholtz equation, we develop a perturbative scheme to amend response functions and fluctuations according to the fluctuation dissipation theorem (FDT). We show that the linear response of a system of several objects is not a simple combination of the response functions of the individual objects. Last, we derive and discuss the Casimir force between nonlinear media in equilibrium. We find that an object which is invisible -- in the sense that its linear response is zero -- still feels a Casimir force when brought to a second object due to nonlinear response.
Consider a material described by (linear) dielectric and magnetic responses $\varepsilon$ and $\mu$ as well as a third order nonlinear electric response $\chi^{(3)}$ \footnote{We omit second order nonlinearities for simplicity. These are often absent due to symmetries.}. All response functions can depend on space, thereby allowing the possibility of disconnected objects, e.g. by sharp step functions at the objects' surfaces. The system is -- in Fourier space for time with frequency $\omega$ -- described by a nonlinear Helmholtz equation \cite{boyd2003nonlinear}. In order to include fluctuations, we add a noise source $\mathbf{F}$, whose properties are yet to be determined,
\begin{equation}
\mathbb{H}\mathbf{E}-\mathcal{N}\left[\mathbf{E}\mathbf{E}\mathbf{E}\right]=\mathbf{F}.\label{eq:SNHE}
\end{equation}
The linear Helmholtz operator is $\mathbb{H}=\nabla\times\nabla\times-\mathbb{V}-\frac{\omega^{2}}{c^{2}}\mathbb{I}$, with the electromagnetic potential $\mathbb{V}=\frac{\omega^{2}}{c^{2}}\varepsilon\left(\omega\right)+\nabla\times\left(\mathbb{I}-\frac{1}{\mathbb{\mu\left(\omega\right)}}\right)\nabla\times$, and speed of light $c$. In what follows, operators (e.g. $\mathbb{H}$ or $\mathbb{V}$) are $3\times3$ matrices and depend on two spatial aguments, such that operator products include matrix multiplication and integration over a joint coordinate. The functional $\mathcal{N}$ describes the third order response, i.e., the $i$th component of $\mathcal{N}\left[\mathbf{E}\mathbf{E}\mathbf{E}\right]$ reads more explicitly,
\begin{eqnarray}
\mathcal{N}\left[\mathbf{E}\mathbf{E}\mathbf{E}\right]_{i} (\omega) = \frac{\omega^{2}}{c^{2}}\int\mathrm{d}\omega_{1}\mathrm{d}\omega_{2}\mathrm{d}\omega_{3}\,\delta\left(\omega-\omega_{\sigma}\right)\nonumber \\
\times\chi_{ijkl}^{\left(3\right)}\left(-\omega_{\sigma},\omega_{1},\omega_{2},\omega_{3}\right)\nonumber E_{j}\left(\omega_{1}\right)E_{k}\left(\omega_{2}\right)E_{l}\left(\omega_{3}\right),
\end{eqnarray}
where $\omega_{\sigma}=\omega_{1}+\omega_{2}+\omega_{3}$ and indices denote spacial components. We assume locality of $\chi^{\left(3\right)}$, so that it couples only fields at equal points in space.
The linear system is solved by the Green's function $\mathbb{G}_{1}$ and the field $\mathbf{E}_{1}$, i.e., $\mathbb{H}\mathbb{G}_{1}=\mathbb{I}$ and $\mathbb{H}\mathbf{E}_{1}=0$. The stochastic equation (\ref{eq:SNHE}) describes the fluctuating electromagnetic field, including quantum- and thermal fluctuations. The noise $\mathbf{F}$ is chosen such that Eq.~\eqref{eq:SNHE} yields correct expectation values in equilibrium (see Eq.~\eqref{eq:N} below).
In order to display the FDT in the nonlinear system, we compute response functions and fluctuations, defining $\tilde{\mathbb{G}}$ as the {\it linear response function} of the nonlinear system, i.e.,
\begin{equation}\label{eq:LR}
\tilde{\mathbb{G}}\equiv\left.\frac{\delta\langle \mathbf{E}\rangle }{\delta\mathbf{E}_{\mathrm{in}}}\right|_{\mathrm{eq}}\mathbb{G}_{0},
\end{equation}
with an incoming field $\mathbf{E}_{\mathrm{in}}$ and the vacuum Green's function $\mathbb{G}_{0}$. This linear response function $\tilde{\mathbb{G}}$ obeys FDT \cite{kubo2012statistical}
\begin{equation}\label{eq:FDT2}
\left\langle \mathbf{E}_{\omega}\otimes\mathbf{E}_{\omega^{\prime}}^{*}\right\rangle^{\mathrm{eq}} = \delta\left(\omega-\omega^{\prime}\right)b\left(\omega\right)\mathrm{Im}\tilde{\mathbb{G}}.
\end{equation}
Here, $b\left(\omega\right)=\frac{\hbar}{\pi\varepsilon_{0}}\frac{\omega^{2}}{c^{2}}\left[1-e^{-\frac{\hbar\omega}{k_{\mathrm{B}}T}}\right]^{-1}$ gives the strength of the fluctuations, with Planck's constant $\hbar$, temperature $T$, Boltzmann's constant $k_{\mathrm{B}}$, and permittivity of vacuum $\varepsilon_0$. Note that for linear systems,
$\tilde{\mathbb{G}}=\mathbb{G}_1$, and the familiar FDT \cite{Eckhardt1984} is recovered. Eq.~(\ref{eq:SNHE}) can be formally solved by a so-called Lippmann-Schwinger equation \cite{lippmann1950variational,rahi2009scattering}
\begin{equation}
\mathbf{E}=\mathbf{E}_{1}+\mathbb{G}_{1}\mathbf{F}+\mathbb{G}_{1}\mathcal{N}\left[\mathbf{E}\mathbf{E}\mathbf{E}\right].\label{eq:LS.equation}
\end{equation}
We may treat the linear solution as an incoming field, because, with $\mathbf{E}_{1}=\mathbb{G}_{1}\mathbb{G}_{0}^{-1}\mathbf{E}_{\mathrm{in}}$, the linear response in Eq.~\eqref{eq:LR} may be written as $\tilde{\mathbb{G}}=\left.\frac{\delta\langle \mathbf{E}\rangle }{\delta\mathbf{E}_{1}}\right|_{\mathrm{eq}}\mathbb{G}_{1}$. From Eq.~(\ref{eq:LS.equation}), using, without loss of generality, $\left\langle \mathbf{E}\right\rangle ^{\mathrm{eq}}=0$, and vanishing mean of the noise, $\left\langle \mathbf{F}\right\rangle=0$ \footnote{A finite mean of ${\bf F}$ may be absorbed into the left hand side of Eq.~\eqref{eq:SNHE}.}, we find in first order in $\chi^{(3)}$ for $\tilde{\mathbb{G}}$,
\begin{eqnarray}
\tilde{\mathbb{G}} = \left(\mathbb{I}+\mathbb{G}_{1}\mathbb{N}\right)\mathbb{G}_{1}.\label{eq:Gt}
\end{eqnarray}
The operator $\mathbb{N}$ contains the equilibrium correlation of the field $\mathbf{E}$, which is (perturbatively) expressed using Eq.~\eqref{eq:FDT2}, and reads (again, the spatial $\delta$ function appears because of the locality of the nonlinear response),
\begin{eqnarray}
(\mathbb{N})_{ij}&=& 3\delta^{\left(3\right)}\left(\mathbf{r}-\mathbf{r}^{\prime}\right)\frac{\omega^{2}}{c^{2}}\int_{-\infty}^{\infty}\mathrm{d}\omega^{\prime}b\left(\omega^{\prime}\right)\nonumber \\
& \times & \chi_{ijkl}^{\left(3\right)}\left({\bf r},-\omega,\omega,\omega^{\prime},-\omega^{\prime}\right)\mathrm{Im}{(\mathbb{G}_1)}_{kl}\left(\mathbf{r},\mathbf{r},\omega^{\prime}\right).\label{eq:N<EE>}
\end{eqnarray}
In Eq.~\eqref{eq:N<EE>}, we have the Green's function of the linear system, $\mathbb{G}_1$, since in first order of $\chi^{(3)}$, the solution of the linear problem multiplies $\chi^{(3)}$.
With Eqs.~\eqref{eq:Gt} and \eqref{eq:FDT2}, the fluctuations of the electric field are fixed and known and may thus readily be used to compute equilibrium quantities such as Casimir forces (see below). We additionally derive the correlation function of the noise ${\bf F}$, which may be relevant for out-of-equilibrium scenarios. From Eq.~\eqref{eq:LS.equation}, it follows to be
\begin{align}\label{eq:N}
\left\langle \left(\mathbb{G}_{1}\mathbf{F}\right)_{\omega}\otimes\left(\mathbb{G}_{1}\mathbf{F}\right)_{\omega^{\prime}}^{*}\right\rangle ^{\mathrm{eq}} & = \delta\left(\omega-\omega^{\prime}\right)b\left(\omega\right)\mathrm{Im}\mathbb{G}_{1}\\
& + \delta\left(\omega-\omega^{\prime}\right)b\left(\omega\right)\mathbb{G}_{1}\left(\mathrm{Im}\mathbb{N}\right)\mathbb{G}_{1}^{*}\nonumber.
\end{align}
Eq.~\eqref{eq:FDT2} yields fluctuations of the electromagnetic field, and another form of FDT \cite{Rytov3,Eckhardt1984} expresses the fluctuations of local currents in the bodies. Using $\mathrm{Im}\tilde{\mathbb{G}}=-\tilde{\mathbb{G}} \mathrm{Im}[\tilde{\mathbb{G}}^{-1}]\tilde{\mathbb{G}}^*$, and \cite{kruger2012trace} $\Im[\mathbb{V}]=-\Im[\mathbb{G}^{-1}-\mathbb{G}_0^{-1}]$, we may rewrite Eq.~\eqref{eq:FDT2} in lowest order in $\mathbb{N}$
\begin{align}\label{eq:R}
\left\langle \mathbf{E}_{\omega}\otimes\mathbf{E}_{\omega^{\prime}}^{*}\right\rangle^{\mathrm{eq}} = \delta\left(\omega-\omega^{\prime}\right)b\left(\omega\right)\tilde{\mathbb{G}}\Im\left[\mathbb{V}+\mathbb{N} -\mathbb{G}_0^{-1}\right]\tilde{\mathbb{G}}^*.
\end{align}
Above, $\Im[\mathbb{V}+\mathbb{N} -\mathbb{G}_0^{-1}]$ is identified as the local Rytov currents \cite{Rytov3}, where $\mathrm{Im}\mathbb{G}_0^{-1}$ is the so-called environment dust \cite{Eckhardt1984,kruger2012trace}, and $\Im[\mathbb{V}+\mathbb{N}]$ are the local Rytov-current fluctuations in the objects. These may now be treated with a local equilibrium approximation to access phenomena out of equilibrium (e.g.~heat transfer). We note that the function $\tilde{\mathbb{G}}$ solves $\tilde{\mathbb{H}}\tilde{\mathbb{G}}=\mathbb{I}$, with $\tilde{\mathbb{H}}=\nabla\times\nabla\times-\tilde{\mathbb{V}}-\frac{\omega^{2}}{c^{2}}\mathbb{I}$ with the potential\begin{equation}
\tilde{\mathbb{V}}=\mathbb{V}+\mathbb{N}.\label{eq:V.correction}
\end{equation}
The potential $\tilde{\mathbb{V}}$, consistently also appearing in the Rytov-current \eqref{eq:R}, may be seen as the analog of the renormalized mass in interacting field theories \cite{kardar2007statistical}. The second term $\tilde{\mathbb{V}}$ in general depends globally on all points in space through $\mathbb{G}_1$ in Eq.~\eqref{eq:N<EE>}. This leads in general to a nontrivial spatial (and shape) dependence for fluctuation effects.
Eqs.~(\ref{eq:Gt}) and (\ref{eq:N<EE>}) determine the linear response of the nonlinear system, $\tilde{\mathbb{G}}$, to first order in $\chi^{\left(3\right)}$. With the explicit correlator for the noise in Eq.~\eqref{eq:N} and the Rytov currents in Eq.~\eqref{eq:R}, they present fluctuational electrodynamics for the nonlinear system, and hence our first main result.
It is important to note that Eq.~(\ref{eq:Gt}) displays important properties of linear response functions (which may be familiar from the electromagnetic Green's function of linear systems); $\tilde{\mathbb{G}}=\tilde{\mathbb{G}}^{\mathrm{T}}$ due to time reversal symmetry, and $\tilde{\mathbb{G}}^{*}\left(\omega\right)=\tilde{\mathbb{G}}\left(-\omega\right)$ due to realness of the time domain response. Also $\tilde{\mathbb{G}}_{\omega}$ is analytic for $\mathrm{Im}\,\omega>0$, so that Matsubara summation can be used to obtain equilibrium averages. This important manifestation of causality can directly be seen from Eqs.~(\ref{eq:Gt}) and (\ref{eq:N<EE>}) by noting that both $\mathbb{G}_1$ and $\chi^{\left(3\right)}\left(-\omega,\omega,\omega^{\prime},-\omega^{\prime}\right)$ are analytic for $\mathrm{Im}\,\omega>0$ \footnote{Analyticity in $\omega^\prime$ is however not given, since the imaginary part cannot be positive for both $\omega^\prime$ and $-\omega^\prime$.}.
Interestingly, in Eq.~\eqref{eq:N<EE>}, the third order susceptibility appears only in the form $\chi^{\left(3\right)}\left(-\omega,\omega,\omega^{\prime},-\omega^{\prime}\right)$. Therefore, in equilibrium, only a subset of third order processes can contribute, for example the optical Kerr effect or the Raman effect. General frequency mixing processes do not contribute, however.
\begin{figure}
\includegraphics[width=.9\columnwidth]{fig1}
\caption{Different configurations for a linear response experiment. The electromagnetic potential of nonlinear media is inhomogeneous and dependent on other objects. \label{fig:combination}}
\end{figure}
Turning to a system made of several objects (objects $\alpha$ and $\beta$, see Fig.~\ref{fig:combination}), it is well known that, for purely linear media, the linear response or Green's function of the collection of objects can be found from the results for the isolated objects (see Eq.~\eqref{eq:c} below, which, removing tildes and primes, holds true for linear media). This is the basis of many results found in fluctuational electrodynamics, such as the Lifshitz formula \cite{Lifshitz56}; well known results for Casimir forces or radiative transfer are based on the scattering properties (or $\mathbb{T}$ operators) of the individual objects. In the nonlinear system, this is no longer true: the linear response $\tilde{\mathbb{G}}$ of a collection of objects, as in Fig.~\ref{fig:combination}, takes a nontrivial dependence on the linear responses of the isolated objects ($\tilde{\mathbb{G}}_\alpha$, $\tilde{\mathbb{G}}_\beta$, \dots). Comparing Eq.~\eqref{eq:Gt} for the system containing two objects to its version for isolated objects, we find that
\begin{equation}
\tilde{\mathbb{G}}=\mathbb{G}^{\prime}+\mathbb{G}^{\prime}\sum_i\left[\mathbb{N}-\mathbb{N}_i\right]\mathbb{G}^{\prime},\label{eq:G.correction}
\end{equation}
where we have introduced the naive (as in the linear case) combination $\mathbb{G}^{\prime}$ for several objects. E.g. , for the case of two objects, we recall (see, e.g., Ref.~\cite{kruger2012trace}),
\begin{align}\label{eq:c}
\mathbb{G}^\prime=\tilde{\mathbb{G}}_{\beta}\frac{1}{\tilde{\mathbb{G}}_{\alpha}+\tilde{\mathbb{G}}_{\beta}-\tilde{\mathbb{G}}_{\alpha}\mathbb{G}_0^{-1}\tilde{\mathbb{G}}_{\beta}}\tilde{\mathbb{G}}_{\alpha}.
\end{align}
The operator $\mathbb{N}$ describes, according to Eq.~\eqref{eq:N<EE>}, the total system, and $\mathbb{N}_i$ describes the situation of object $i$ in isolation.
As mentioned before, Eq.~\eqref{eq:G.correction} states that the linear response of a collection of objects is a nontrivial form of the linear response of the individual ones. This is, again, because fluctuations (zero point and thermal) interact with the incoming field through nonlinearities. This effect is absent for linear media as the incoming and the fluctuating fields are decoupled. It is in principle measurable with scattering experiments, and leads to the different behavior of Casimir forces as described below.
As a concrete application, we compute the Casimir force between two parallel, semi-infinite plates at distance $d$. For linear materials, the well known Lifshitz formula \cite{Lifshitz56} gives the Casimir force for this system in terms of the Fresnel coefficients of the individual plates \cite{dzyaloshinskii1961general}. We give here the result for nonlinear materials. The force, or the Casimir energy, may be found in multiple ways, here we compute the equilibrium correlation function of the electric field in the vacuum between the surfaces, which then gives us the Maxwell stress tensor and the force.
In the following, we consider homogeneous and isotropic materials, for which the result is given in the appendix. In order to keep the discussion simple, we restrict to frequency independent material parameters. The Lifshitz force then depends on the linear response of the individual plates ($\varepsilon_\alpha$), as well as on the nonlinear function $\chi^{(3)}_\alpha$. Omitting terms of order $(\chi^{(3)})^2$, it suffices to consider only $\chi^{(3)}$ of one of the two plates to be finite (the effect of $\chi^{(3)}$ of the other plate is found by exchanging the plate indices).
The force evaluation comprises two frequency integrals (see Eq.~\eqref{eq:N<EE>} for the additional integral), both of which are evaluated on the imaginary axis via Matsubara summation. While this is naturally possible for $\omega$ (as mentioned), analyticity for $\Im \omega'>0$ can also be shown if $\chi^{(3)}$ is frequency independent. The pressure $P$ is then split into two terms, the result of the Lifshitz formula for linear media, and a novel term, resulting from $\chi^{(3)}$. In the zero and infinite temperature limits, these terms can be cast as
\begin{eqnarray}
P^{T\rightarrow0} & = & \frac{\hbar c}{d^{4}}I_{\mathrm{lin}}^{T\rightarrow0}+\frac{\chi^{\left(3\right)}}{\varepsilon_{0}}\left(\frac{\hbar c}{d^{4}}\right)^{2}I_{\mathrm{nl}}^{T\rightarrow0},\label{eq:P.0}\\
P^{T\rightarrow\infty} & = & \frac{k_{\mathrm{B}}T}{d^{3}}I_{\mathrm{lin}}^{T\rightarrow\infty}+\frac{\chi^{\left(3\right)}}{\varepsilon_{0}}\left(\frac{k_{\mathrm{B}}T}{d^{3}}\right)^{2}I_{\mathrm{nl}}^{T\rightarrow\infty},\label{eq:P.inf}
\end{eqnarray}
where we used that, for isotropic materials, $\chi_{iikk}^{\left(3\right)} = \chi_{ikki}^{\left(3\right)}=\chi_{ikik}^{\left(3\right)}=\chi^{\left(3\right)}$ for $i\ne k$ and $\chi_{iiii}^{\left(3\right)} = 3\chi^{\left(3\right)}$. The functions $I$ are independent of temperature and distance and we note the following properties of the Casimir force: for $T\to0$, the nonlinear term is proportional to $\hbar^2$, and diverges with $1/d^8$, i.e., it is irrelevant at large distances, and relevant at small $d$. This can be understood intuitively: For small $d$, the intracavity field fluctuations are large, and probe the nonlinear regime. In the high temperature limit, the behavior is very similar, being proportional to $(k_BT)^2$, and diverging with $1/d^6$.
\begin{figure}
\includegraphics[width=1\columnwidth]{fig2}
\caption{Main graph: Nonlinear contribution to the Casimir pressure for different values of $\varepsilon_{lin}$, as labeled, as a function of $\varepsilon_{nl}$. This graph refers to the quantum limit, $T\to0$. Inset shows the corresponding linear term.\label{fig:Low-temperature-limit}}
\end{figure}
\begin{figure}
\includegraphics[width=1\columnwidth]{fig3}
\caption{Main graph: Nonlinear contribution to the Casimir pressure for different values of $\varepsilon_{lin}$, as labeled, as a function of $\varepsilon_{nl}$. This graph refers to the thermal limit, $T\to\infty$. Inset shows the corresponding linear term.\label{fig:High-temperature-limit}}
\end{figure}
\begin{figure}
\includegraphics[width=.9\columnwidth]{fig4}
\caption{Casimir force between a nonlinear object ($\varepsilon_{nl}$ as labeled) and a perfect mirror ($\varepsilon_{\mathrm{lin}}\rightarrow\infty$), in the quantum limit ($T\rightarrow0$), as a function of distance $d$. Here, $\chi^{\left(3\right)}=2\times10^{-16}\,\mathrm{\frac{m^{2}}{V^{2}}}$.\label{fig:Distance-dependence}}
\end{figure}
The numerical results for the functions $I$ are shown in Figs.~\ref{fig:Low-temperature-limit} and \ref{fig:High-temperature-limit} as functions of the dielectric permittivities $\varepsilon_{\mathrm{lin}}$ (corresponding to the plate with $\chi^{(3)}=0$) and $\varepsilon_{\mathrm{nl}}$ (corresponding to the plate with a finite $\chi^{(3)}$). For a given $\chi^{\left(3\right)}$, the nonlinear pressure contribution is maximal if $\varepsilon_{\mathrm{lin}}\rightarrow\infty$, i.e., if the linear plate is a perfect mirror, and if $\varepsilon_{\mathrm{nl}}\rightarrow1$, i.e., if the the nonlinear plate is transparent. These conditions correspond to minimizing the losses through the linear plate and reflections from the nonlinear plate. As $\varepsilon_{\mathrm{nl}}\rightarrow\infty$, the nonlinear term vanishes. We interpret that in this case, the absorption length in the material vanishes, and the waves cannot penetrate the material to probe nonlinearities.
Figure \ref{fig:Distance-dependence} finally gives the total Casimir force as a function of distance $d$, for the quantum limit ($T\to0$). Here, we have taken $\varepsilon_{lin}\to\infty$. The force takes the well known law of $1/d^4$ for large $d$, and crosses over to $1/d^8$ for small $d$. The nonlinear susceptibility $\chi^{(3)}=2\times10^{-16}\,\mathrm{\frac{m^{2}}{V^{2}}}$ was used, a value measured for glass infused with silver nanoparticles \cite{Karvonen2013}. We note that the crossover takes place at a distance of a few nanometers and is in experimental reach.
As already apparent from Fig.~\ref{fig:Low-temperature-limit}, the ratio between nonlinear and linear force can be arbitrarily large if $\varepsilon_{nl}$ approaches unity, i.e., if the nonlinear surface becomes transparent. This is also shown in Fig.~\ref{fig:Distance-dependence}, where we depict the total force for different $\varepsilon_{nl}$. As $\varepsilon_{nl}\to 1$, the nonlinear force becomes more and more dominant.
This has an interesting extreme limit, serving as a simple paradigm for Casimir forces in nonlinear systems. Taking a fully transparent (invisible) object, i.e., having the linear response of vacuum, $\varepsilon_{nl}=1$, and bringing it close to a perfect mirror, the object feels the following total Casimir force,
\begin{align}\label{eq:CT}
&P = \frac{3}{2^{8}\pi^{4}}\varepsilon_{0}\mathrm{Re}\iiiint\mathrm{d}\omega\,\mathrm{d}\omega^{\prime}\mathrm{d}q\mathrm{d}q^{\prime}\chi^{\left(3\right)}\left(-\omega,\omega,\omega^{\prime},-\omega^{\prime}\right)
\nonumber \\ &\times
\frac{a\left(\omega\right)a\left(\omega^{\prime}\right)}{k^{2}k^{\prime2}} N_{q,q^{\prime}}^{\omega,\omega^{\prime}}\left[\frac{e^{2i\left(p+p^{\prime}\right)d}}{\left(p+p^{\prime}\right)p^{\prime}}+\frac{e^{2i\left(p-p^{\prime*}\right)d}}{\left(p-p^{\prime*}\right)p^{\prime}}\right],
\end{align}
where $N_{q,q^{\prime}}^{\omega,\omega^{\prime}}=qq^{\prime}\left[k^{2}\left(4k^{\prime2}-3q^{\prime2}\right)-q^{2}\left(6k^{\prime2}-7q^{\prime2}\right)\right]$ and $a\left(\omega\right)=b\left(\omega\right)-b\left(-\omega\right)$, with the integral ranges of $[0,\infty]$. We used $k=\omega/c$ and $p=\sqrt{k^{2}-q^{2}}$, analogous for primed variables.
As mentioned, the force in Eq.~\eqref{eq:CT} diverges as $1/d^8$ and $1/d^6$ in the quantum and thermal limits, respectively \footnote{The functional behavior of $\chi^{(3)}$ may lead to deviations from these laws.}. Using metamaterials, such extreme material properties may be approached, e.g., by index matching coating \cite{Southwell1991}. Particles in fluids can also be index matched \cite{budwig1994refractive,wiederseiner2011refractive}.
The combination of fluctuational electrodynamics and nonlinear optics offers a variety of unexplored effects. The stochastic Helmholtz equation is then supplied with an adopted noise strength, allowing description of stochastic processes. In equilibrium, the linear response of a collection of objects is a nontrivial function of the linear responses of the isolated particles, and the Casimir force acting between bodies with nonlinear optical properties is amended at small distances. Future work may investigate path integral formulations of this framework and address nonequilibrium situations.
We thank M. Kardar, G. Bimonte, D.S. Dean, T. Emig, N. Graham, and R. L. Jaffe for discussions. This work was supported by Deutsche Forschungsgemeinschaft (DFG) grant No. KR 3844/2-1 and MIT-Germany Seed Fund grant No. 2746830.
|
\section{Introduction}
The radial velocity method to detect extrasolar planets has now reached
long-term precisions of $\sim$ 1 $\rm m\,s^{-1}$, and new instruments are being
built to achieve accuracies in the few tens of $\rm cm\,s^{-1}$ regime
\citep[e.g. ESPRESSO,][]{pepe2010}. A thorough understanding of the current instruments
and their standard extraction, reduction and radial velocity determination pipelines
is essential in achieving their ultimate capabilities, and
identifying the key technical aspects that need refinement in the next
generation of instruments. While 1~$\rm m\,s^{-1}$ precision has been enough to
identify the emerging population of super-Earths in compact planetary systems, and to detect potentially habitable planets around M dwarfs, achieving $\rm sub-m\,s^{-1}$ precision at all time-scales is needed to
efficiently characterize Earth and sub-Earth mass objects in hot-to-warm orbits
around G and K dwarfs. There is active discussion on whether the reported $\sim$ 1$\rm
m\,s^{-1}$ long-term precision limit, measured on stabilised spectrographs, is
caused by stellar processes, instrumental effects or both. While certain stars
are, intrinsically, more Doppler variable than this limit, it is unlikely
that the 0.8~-~0.9~$\rm m\,s^{-1}$ (reported lower noise limit in HARPS-ESO
observations by \citealp{pepe2011}, and of a handful of very quiet M~dwarfs, e.g.
GJ~699 and GJ~588, by \citealp{anglada2012a}) is an universal limit for
Doppler measurements.
Studying the limits of precision of Doppler spectroscopy in the short period domain of M-type dwarfs is one of the primary goals of the ``Cool Tiny Beats survey" \citep[CTB;][]{anglada2014, berdinas2015}. The survey, started in 2013, was designed to obtain high-cadence observations (continuous observing of a single target during several consecutive nights) of a small sample of M-dwarfs with HARPS-ESO (High Accuracy Radial Velocity Planet Searcher) and HARPS-N \citep{mayor2003, pepe2004, cosentino2012}. Our objective was threefold: (1) the detection of small planets in sub-day period orbits, (2) the search for pulsations, and (3) the study of the time-scales of stellar activity.
From the analysis of those data, it was obvious that some kind of instrumental or reduction-process effect was producing intra-night and night-to-night variability common to several targets. After the first runs in each instrument, two of such effects were quickly identified: (1) Doppler shift correlations with the signal-to-noise ratio \citep[SNR,][]{bouchy2009}, and (2) $\sim$0.5~-~1.5~$\rm m\,s^{-1}$ night-to-night jumps in the radial velocity (RV) series. Berdi\~nas et al. in prep. (hereafter, CTB 2016) find that the first one was partly due to the Charge Transfer Efficiency effect, and provide an empirical correction that should at least be valid for M-dwarfs observed from HARPS-ESO. The effect was corrected in HARPS-N (but not in HARPS-ESO) by introducing, in 2013, modifications to the standard Data Reduction Software (DRS) pipeline (Lovis, C., Pepe, F., priv. comm.). The second effect was found to be caused by random and systematic errors in the wavelength solution of unknown origin (Lo Curto, G., priv. comm.). CTB 2016 also provides a solution for this effect that consists in using a mean wavelength solution for each observing run instead of the one derived from individual night calibrations, for both HARPS-ESO and HARPS-N.
While these corrections substantially reduced the common systematic variability of several observed stars, yet another source of systematic noise was strongly affecting HARPS-N data of M-dwarf stars. This variability correlates with the airmass, and was found to produce structured Doppler noise at the level of $\sim$2~$\rm m\,s^{-1}$, with even larger amplitude in the full width at half maximum (FWHM) of the cross-correlation function. Similar trends have been recently reported for circular fiber-fed spectrographs like SOPHIE for other spectral types \citep{bohm2015}.
There are many environmental effects which worsen the higher the airmass is, such as the atmospheric refraction, which increases with the air density at low altitudes splitting the star image in its blue-red components in the zenithal direction; the seeing, which also increases at low altitudes, being the blurring more severe for shorter wavelengths; or the atmospheric extinction, caused by the Rayleigh scattering and by the molecules and dust absorption. In terms of airmasses, the atmospheric dispersion dominates below $\sim50\deg$ on nights with good seeing ($\sim$1 arcsecond in the zenith). To correct for this, HARPS-N includes an atmospheric dispersion corrector (ADC).
The aim of this work is to study the response of the instrument to any of these sources in the intra-night regime, check if these effects lead to variability of the spectra and provide means for correcting them. The paper is organised as follows. We describe the observations and the properties of the targets used for this study (GJ~725A and B) in Section~\ref{sec:observations}. The data reduction and the spectroscopic and image indices to be used in this study are presented in Section~\ref{sec:measurements}. Section~\ref{sec:analysis} comprises the analysis, where, in Section~\ref{sec:corr} we show the common variability detected in the alternating observations between GJ~725A and B, and also we intend to assess statistical correlations between spectroscopic observables and various instrument related quantities. In Section~\ref{sec:experiment} we introduce the least square deconvolution profiles (LSD) and perform two experiments to test the algorithms used in this paper (HARPS-TERRA and DRS data reduction softwares). We propose tentative functions to detrend the line width and the RVs measurements in Section~\ref{sec:detrending}. In Section~\ref{sec:spaprof} we present the analysis of the mean line profiles calculated in the cross-dispersed direction. We discuss a promising $\sim$~2.7 days Doppler signal in GJ~725Bb in Section~\ref{sec:dopplerb}. Finally, Section~\ref{sec:conclusions} present the discussion and conclusions of this work.
\section{Observations}\label{sec:observations}
We used the HARPS-N spectrograph at the 3.6\,m Telescope Nazionale Galileo
(TNG), settled at the Roque de los Muchachos Observatory in La Palma. HARPS-N is
nearly a twin of HARPS-ESO. To guarantee that the optics keep their alignment, both are vacuum-sealed thermally stabilized fibre-fed \'echelle spectrographs. HARPS-N input fibre samples 1~arcsecond circular
apertures on the sky, producing high-resolution optical spectra from 383 up to
693~nm. The high resolution R$\sim$120000 (at the central wavelength, see table~6 in \citealp{cosentino2012}), makes the radial velocities (RVs) measured from the spectra the most long-term accurate currently available.
\begin{table}
\begin{center}
\caption{Observational parameters of the 2014 run.}
\label{tab:runparam}
\begin{tabular}{lcccc}\hline
Nights & Night 1 & Night 2 & Night 3 & Night 4 \\ \hline
$\rm N_{A}$ & 27 & 40 & 16 & 18\\
$\rm N_{B}$ & 27 & 21 & 31 & 37\\
$\rm AM_{max}$ & 1.55 & 1.64 & 1.59 & 1.65 \\
$\rm Seeing$ & 0.86 & 0.87 & 0.74 & 0.78 \\
$\rm Seeing_{min}$ & 0.58 & 0.60 & 0.48 & 0.59 \\
$\rm Seeing_{max}$ & 2.83 & 1.47 & 3.64 & 1.20 \\
$\rm SNR_{A}$ & 78 & 54 & 59 & 71 \\
$\rm SNR_{B}$ & 63 & 49 & 53 & 59 \\\hline
\end{tabular}
\end{center}
Main observational parameters of the 2014 run (2nd to 5th July). Subindices A and B refer to targets GJ~725A and GJ~725B. $\rm N$ indicates the number of exposures. The SNR and seeing values are median measurements of the nights. SNR values correspond to the spectral order 60, centered at 631 nm. Since the TNG seeing monitor was out of order, we show the values given by the nearby Isaac Newton Telescope. 1.17 was the minimum airmass (AM) of all nights.
\end{table}
Our targets were the bright nearby stars: GJ~725A and GJ~725B\footnote{Check
http://www.pas.rochester.edu/$\sim$emamajek/spt/M3V.txt and
http://www.pas.rochester.edu/$\sim$emamajek/spt/M3.5V.txt for a thorough review
of both stars properties.}. We selected this binary system for being a bright,
\citep[$V$=8.91, $V$=9.69, for A and B respectively,][]{jenkins2009, cutri2003},
nearby \citep[d=3.57 pc, d=3.45 pc,][]{anderson2012}, and common proper motion
pair of almost identical spectral type \citep[M3V and M3.5V,][]{reid1995} with
no previous reports of intense magnetic activity or planets. At epoch 2000,
their projected separation was 13.3~arcseconds, implying a minimum separation of
47.5~AU \citep[Hipparcos,][]{vanleeuwen2007}.
We observed the stars in two high-cadence different runs, one in 2013 and
another one in 2014. GJ~725A was the primary target of our five-day long,
August 2013 run, whereas, in 2014, GJ~725A and B were observed every night for 4
consecutive nights alternating exposures between the two of them. This gave us a total
of 314 A data points in 2013 and 101 for A and 116 for B in 2014, after
rejecting those with SNR below 45 and airmass above 2.5 (the HARPS-N atmospheric corrector does not work for higher airmasses, F. Pepe, private comm.). In the 2014 run, except for
the first night, we halted the alternating sequence between A and B at the
meridian passage, observing instead only one of the stars. We did this
to improve the cadence of each target at that specific moment of the night.
Exposure times were set to be between 240 and 480~s. The main observational
parameters of the run are shown in Table~\ref{tab:runparam}.
\section{Measurements}\label{sec:measurements}
For each observation, we obtained a number of measurements (or indices) to perform this study. These indices are detailed below.
\subsection{Spectroscopic measurements}
Spectra were processed and extracted with the standard HARPS-N/TNG Data
Reduction Software (DRS). We obtained the RV
measurements using the maximum likelihood matching technique
implemented in the HARPS-TERRA software \citep{anglada2012a}.
The DRS cross-correlates the observed spectra
with a weighted binary mask formed by ones in the theoretical stellar lines positions and widths and by zeros elsewhere \citep{pepe2002}. The resulting cross-correlation function (CCF) acts
as a proxy for the mean line profile for each spectral order. In our study, we will use
the so-called FWHM index (full width at half maximum of the
Gaussian fitted to the cross-correlation profile). Since the line width might be sensitive
to the presence of spots or magnetic activity \citep[e.g.][]{reiners2013}, this
FWHM is often used as an activity index to decorrelate Doppler
variability.
\subsection{Spectral energy distribution measurements (pSEDs and $K$ index)} \label{sec:sed}
We also studied the distribution of the flux across the spectra.
With this aim, we used the spectra as given by the DRS reduction pipeline for each spectral order (the so-called e2ds data-products), and we calculated how the total flux on each spectral order varied relative to a reference order.
To do this, we firstly added the flux of all pixels in each spectral order, and secondly, we normalised these measurements to the total flux at the spectral order 56, which was selected for having low telluric contamination. We named this
``normalized pseudo spectral energy distribution" (pSED). As we
see for the 2014 data of GJ~725A in the upper panel of Figure~\ref{fig:sed},
the pSED pivots around the normalization order (R). Changes relative to the
first co-added spectrum (T) are shown in the lower panel of Figure~\ref{fig:sed}.
The pSED is expected to vary for a number of environmental effects coupled with the
instrument, as for example the atmospheric dispersion or the chromatic seeing. Account for the atmospheric dispersion is the ADC role, however the flux reaching the telescope will also change as a function of airmass due to the atmospheric extinction (i.e. due to the Rayleigh scattering and to the absorption from molecules and dust in the Earth atmosphere). We corrected the pSED due to atmospheric extinction using the method outlined by
\cite{hayes1975}\footnote{\label{footnote:atmext}See ORM atmospheric extinction
values for a dust-free atmosphere at
http://www.ing.iac.es/Astronomy/observing/manuals/ ps/tech\_notes/tn031.pdf. To
correct from aerosol scattering we used the V-band extinction measurements for
the month of July from table~1 of \cite{garcia-gil2010}.}. Although the
atmospheric extinction is wavelength and airmass dependent, the pSED varied less
than a 4.5 per cent and the wavelength dependence in the pSED remained equally
strong after the correction.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_sed.pdf}
\caption{(Upper panel) Sum of the flux of the 69 spectral orders of GJ~725A (the so-called pseudo spectral energy distribution, pSED) plotted versus wavelength ($\lambda_{c}$ refers to the central wavelength of each spectral order). The pSEDs cross the normalization point (R). The blue line T indicates the initial pSED of the run. (Lower panel) Relative flux compared to T. We limited the study to a linear pSED/T region of $\pm8$ spectral orders from R (blue area). Fluxes were corrected from atmospheric extinction.}
\label{fig:sed}
\end{figure}
To quantify the impact of changes in the pSED, we defined a spectroscopic chromatic index
($K$). This index measures the relative changes in the slope of the pSED as a function of wavelength considering that a linear function fits the pSEDs in the surroundings of R (orders inside the blue box of Figure~\ref{fig:sed})
$K$ is obtained from:
\begin{equation}
\mathrm{pSED}_{\rm t}\,(\lambda) = [1 + K_{\rm t}\,(\lambda_{\rm{c}} - \lambda_{\rm{cR}} )]\, \rm T(\lambda),
\label{eq:k}
\end{equation}
where $t$ indicates the observations, $\lambda_{\rm{cR}}$ is the
central wavelength of the reference order 56, $\lambda_{\rm{c}}$
refers to the central wavelength of any other order,
$\rm{T}\,(\lambda)$ corresponds to the initial
$\rm{pSED}_{\rm{t_{0}}}(\lambda)$ for each star; and finally,
${K}_{\rm{t}}$ is a coefficient with units of 1/$\lambda$,
that we call chromatic index.
\subsection{Measurements on the autoguide camera images (Semimajor axis module, $\Phi$, and Circularity index, $\Theta$)}\label{sec:images}
The images obtained by the autoguide camera were additional products that were
used for trying to find an explanation for the systematics. In HARPS-N, the star
light that is not injected into the fibre is reflected towards the autoguide
camera. To do this, the fiber head is fitted in the central hole of a slightly tilted flat mirror. Therefore, the images recorded by this camera show the shape of the star
at the fibre entrance. Two kind of images are stored, the acquisition images,
taken with the star off-fibre before the science exposure, and the autoguide
images, which are the average of the short images ($\sim 100~\rm ms$) taken during the
science exposure to guide the telescope.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_atmosphere.pdf}
\caption{Schematic view of the sky orientation of a non-de-rotated field of view in an altazimuthal telescope like HARPS-N. During the night, the field rotates following the celestial north axis direction (N), while the zenithal axis (Z) remains perpendicular to the atmosphere and does not rotates with the field. The atmosphere chromatically disperses the light more efficiently when the atmospheric layers are thicker (e.g. for $\rm h_{t_{1}}$ compared with $\rm h_{t_{2}}$). $p_{r}$ refers to the parallactic angle, the angle subtended by Z and N.}
\label{fig:adc}
\end{figure}
It must be realised that, in an equatorial telescope (or in a de-rotated field of view
in an altazimuthal telescope), the direction of the dispersion of the atmosphere
changes during the night (following the zenithal direction), whereas in a
non-de-rotated field of view (the case of HARPS-N@TNG) the zenithal
direction remains constant while the field rotates (see Figure~\ref{fig:adc}).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_autoguide.pdf}
\caption{GJ~725A autoguide camera images taken at the beginning (left), close to the meridian crossing (middle), and at the end (right) of the 2nd July 2014 night. The central hole, which corresponds to the fibre position, seems to be elliptical due the slight tilt of the flat mirror which diverts the light towards the autoguide camera. The solid black lines refers to the elliptical fits performed to account for the elongation of the images. After accounting for the field rotation and the fix angle of the derotator, we measured a good agreement of $p_{h}$ --the angle subtended by the north axis N, and the ellipse semimajor axis Z'-- with $p_{r}$, the real parallactic angle (or angle subtended by the north-zenith axis). The distortion of the images is aligned with the atmospheric dispersion axis (Z), an thus, with the airmass.}
\label{fig:autoguide}
\end{figure}
It is important to note that, as a result of the flat mirror tilt, a stable distortion of the images was expected; i.e. we expected the images to be elongated in the same proportion and direction as we measured for the circular hole of the mirror (see the shape of the central hole in Figure~\ref{fig:autoguide}). However, the autoguide images presented a variable elliptical elongated shape of the star that did not fit the distortion of the mirror hole and did not rotate with the field (within a maximum error of $4^{\circ}$), keeping always aligned with the
zenithal direction. This result, together with the correlation of the elongation with
the airmass (see three autoguide images taken at the beginning, middle and end of
one night, with airmasses 1.30, 1.17, and 1.33 respectively, in
Figure~\ref{fig:autoguide}), led us to hypothesize that the elongated shape was
produced due to an insufficient correction of the atmospheric chromatic
dispersion by the ADC. To validate our hypothesis, we fitted an ellipse to the autoguide images (only the outermost pixels with fluxes within $\pm$1 per cent of the half value between the maximum and the background were considered), and we checked that the elongation was compatible with the zenithal axis. In particular, we measured maximum differences of only $7.8^{\circ}\pm6.2^{\circ}$ between $p_{r}$, the real parallactic angle, and $p_{h}$, its analogous in case the semimajor axis of the ellipse fitted to the image and the zenithal direction were coincident. Furthermore, if this is so, the resulting image is consistent with a superposition of wavelength-dispersed images of the star caused by the atmosphere.
\subsection{Summary of used measurements}
For the sake of clarity, we summarise here the measurements that will be use in this study.
\begin{itemize}
\item{Spectra}
\subitem RV: Extracted with HARPS-TERRA.
\subitem FWHM: Given by the DRS pipeline.
\item{Spectral Energy Distribution}
\subitem \emph{Spectroscopic Chromatic Index} ($K$): Quantifies the changes in the slope of the SED.
\item{Autoguide images}
\subitem \emph{Semimajor axis module} ($\Phi$): Module of the semimajor axis resulting from fitting an ellipse to the autoguide images.
\subitem \emph{Circularity index} ($\Theta$): Seeing difference, in arcseconds, between the x- and y-axis, obtained from the autoguide image headers.
\end{itemize}
\section{Analysis} \label{sec:analysis
\subsection{Systematic effects in the time-series} \label{sec:corr}
GJ~725A was the primary target of the high-cadence 2013 run. These observations showed a
clear $\sim$~1 day Doppler signal highly correlated with the airmass and common to the RVs and the FWHM time-series (see Figure~\ref{fig:run13timeser}). We suspected an instrumental origin, but with only one observed star, we could not distinguish systematic effects from astrophysical ones.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_Run13timeser.pdf}
\caption{RV (top), FWHM (middle), and $K$ (bottom) time-series of
GJ~725A observed during five nights run of 2013. The FWHM vary tens of $\rm
m\,s^{-1}$ peak-to-peak, while the RV peak-to-peak variability is $\sim$
4.5~m\,s$^{-1}$.}
\label{fig:run13timeser}
\end{figure}
As a consequence, we designed a second observational run in 2014 to separate
astrophysical from instrumental effects: given that GJ~725A+B is a
common proper motion pair of almost identical components, obtaining alternating
observations of both components is a simple and efficient way to point out common trends.
For short periods of time, we also halted the alternation strategy to monitor a single star and
increase sensitivity in the high frequency domain.
Correlated variability, common to both stars, can readily be spotted by direct
inspection of the time-series (see Figure~\ref{fig:dataproducts}). For the first
observed night, the RV correlation coefficient between GJ~725A and B time series
was $r=0.30$ and $r=0.87$ for the FWHM.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_dataproducts.pdf}
\caption{GJ~725A (black squares) and GJ~725B (red
dots) RVs (upper panel) and FWHM (lower panel) time-series for
the first HARPS-N run night on 2nd July 2014. For the sake of
clarity, we subtracted the FWHM mean velocities (2989.97 and
2975.62~$\rm m\,s^{-1}$ for GJ~725A and B, respectively), and
we shifted the FWHM (-20~$\rm m\,s^{-1}$) and the RVs
(-2.3~$\rm m\,s^{-1}$) of GJ~725B. We referred the time-axis
to the beginning of the night at 21h 11m 28s UTC. Both stars
show common variability in both quantities confirming
an instrumental origin.}
\label{fig:dataproducts}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_perdifrun.pdf}
\caption{Periodograms for the RV (upper panel) and the FWHM (lower
panel) for the GJ~725A (black line) and GJ~725B (red line) data of the 2014 run.
The blue line corresponds to the periodogram of the two stars differential
time-series. Note the $\sim3$~d peak in the RVs difference over the 0.1 per cent
false alarm probability threshold. We avoided sampling effects by interpolating the GJ~725B times to
the GJ~725A ones. We also excluded high-cadence observations. Common
variabilities result in common peaks, which disappear on the differential
periodograms. The remaining peaks are real GJ~725A or GJ~725B signals.}
\label{fig:difrunper}
\end{figure}
To isolate real signals arising from either star, we computed periodograms of
the time-series of the individual stars, together with periodograms of their
difference (see Figure~\ref{fig:difrunper}).
Periodograms in Figure~\ref{fig:difrunper} follow the procedures outlined in \cite{zechmeister2009}. That is, we used the F-ratio statistic to find what is the period of the sinusoid best-fitting the data \citep[see detailed description in][]{zechmeister2009}.
We used these periodograms because, while being formally equivalent to the Lomb--Scargle periodograms \citep{scargle1982}, they are less susceptible to aliasing and provide more accurate frequencies.
To compute the differential periodograms, we calculated the difference between the GJ~725A and B velocities, evaluating B at the A observing
epochs (we linearly interpolate the GJ~725B time-series). RV and FWHM periodograms of
the individual stars showed common peaks (see black and red lines in Figure~\ref{fig:difrunper}),
and most of these peaks disappear in the differential periodograms (blue lines). This result
suggests common instrumental variability on the data-products of both stars. Note a
promising signal arising over the 0.1 per cent of the false alarm probability \citep[or FAP,][]{cumming2004} threshold at $\sim$3~d in the RV differential periodogram. This signal may corresponds to some real variability on GJ~725B as we discuss later in Section~\ref{sec:dopplerb}. The indices calculated over the autoguide images and the pSED (e.i. $\Phi$, $\Theta$, and $K$) also show common variability on both stars as we show in Figure~\ref{fig:indices}.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_indices.pdf}
\caption{Time-series of $\Phi$, $\Theta$ (images indices), and $K$
(spectroscopic chromatic index) for the 2nd of July 2014. Black squares and red
dots correspond to GJ~725A and GJ~725B, respectively. For the sake of clarity,
we shifted GJ~725B -3~pxl for $\Phi$, -0.35~arcseconds for $\Theta $, and -2.5
$\mu m^{-1}$ for $K$. The time axis is referred to the first observation.}
\label{fig:indices}
\end{figure}
\subsubsection{Analysis of the correlated variability between indices}
We analysed all possible correlations between our spectroscopic, autoguide and
pSED indices using the following procedure. Given two indices $x_i$ and $y_i $,
we modeled the existence of correlations by fitting a linear relation between them:
$y_{i}=a+bx_{i}$, where $a$ and $b$ are the free parameters. Assuming that the
null model containing no correlation is $b=0$ and $a = \bar{y}$,
our significance assessment consists in obtaining the improvement in the
$\chi^2$ statistic as:
\begin{equation}
\Delta\chi^{2}= \sum_{i}\left (
\frac{y_{i}-\left(a+bx_{i}\right)}{\sigma_y{_{i}}}\right)^{2} - \sum_{i}\left
(\frac{y_{i}-\bar{y_{i}}}{\sigma_{y_{i}}}\right)^{2},
\end{equation}
\noindent and then determining whether such an improvement could be caused by a
fortunate arrangement in the noise. Instead of using analytic expressions for
the expected distribution of $\Delta\chi^{2}$ in the presence of noise, we
obtained its empirical distribution using a Monte Carlo approach. For each
correlation under investigation, we generated a large number of synthetic
datasets by randomly swapping the $y_i$ values while keeping the same $x_i$. The
false alarm probability (or FAP) is empirically define here as the number of these random experiments which give a
$\Delta\chi^{2}$ larger than the one obtained from the real data divided by the
total number of experiments. We performed enough experiments (in all cases $10^{3}-10^{4}$ depending on
the obtained FAP) to ensure that the
FAP estimates were accurate enough.
The correlation analysis for the 2014 measurements (which include both
stars) are shown in Figures~\ref{fig:corrimag} and~\ref{fig:corrspec}. The two
squares in the top right corners of the individual plots indicate the FAP
(ticked green squares mean $\rm FAP<~1\%$, dotted orange mean
$1\%\leq\mathrm{FAP}<10\%$, and crossed red mean $\mathrm{FAP}\geq10\%$) for
GJ~725A (black squares data) and GJ~725B (red dots data). We excluded five
outliers on $\Phi$ (three for having a dispersion above 4.5 the measured standard
deviation of the sample, and two produced by sudden seeing increases and
corresponding SNR in the spectra below 40).
The strongest correlations correspond to the FWHM and the spectroscopic $K$-index
(Figure \ref{fig:corrspec}, first column). The correlation is so strong that all
nights tested resulted in compatible fitted parameters for both stars. This
dependence, together with the correlation of $K$ with the image index $\Theta$
(Figure \ref{fig:corrimag}, first panel), seems to indicate that the issues at
the fiber coupling are propagating all the way down to the science spectrum and
affecting its SED and its mean line profile. In other words, the instrumental
profile of the spectrograph seems to vary with wavelength and time. It is important
to note that variable seeing necessarily increases the autoguide image radius
momentarily, but it will increase proportionally in both the x- and y-axis.
As a result, $\Theta$ will remain as a good measurement of the asymmetry in the
fibre injection.
We also measured a less strong correlation between the RVs and the image index
$\Phi$ (Figure \ref{fig:corrimag}, third panel). We used the GJ~725B RVs after subtracting the Doppler signal discussed in Section~\ref{sec:dopplerb}. The correlation is not as strong as that of the FWHM with
$K$, however both stars have the same qualitative behaviour. The $\rm
FAP$ of all observed nights remains below the 10 per cent threshold.
\begin{figure*}
\includegraphics[width=\textwidth]{fig_SuperPanelImageIndex.pdf}
\caption{Correlation of the spectroscopic indices, $\rm{FWHM}$ (left panel), $K$ (central panel), and $\rm{RV}$ (right panel), with the image indices, $\Phi$ and $\Theta$. Each row matches one night. Black squares correspond to GJ~725A and red dots to GJ~725B. Two squares in the top right corners of each plot indicate the FAP for GJ~725A and B, respectively from the left: ticked green squares means $\mathrm{FAP} < 1\%$, dotted orange $1\% \leq \mathrm{FAP} < 10\%$ and crossed red $\mathrm{FAP} \geq 10\%$. Main correlations are: FWHM-$\Theta$, $K$-$\Theta$ and RV-$\Phi$. GJ~725B RVs are residuals to the Doppler signal discussed in Section \ref{sec:dopplerb}.}
\label{fig:corrimag}
\end{figure*}
\begin{figure*}
\includegraphics[width=11cm]{fig_SuperPanelSpecIndex.pdf}
\caption{Best correlation linear models fitted for the spectroscopic indices: $\mathrm{FWHM}-K$ (left panel), $\mathrm{RV}-K$ (central panel), and $\mathrm{FWHM}-\mathrm{RV}$ (right panel). Markers and top right corner squares are coded as in Figure~\ref{fig:corrimag}. Among these indices FWHM-$K$ was the only correlation detected.}
\label{fig:corrspec}
\end{figure*}
\subsubsection{Analysis of the 2013 single star run on GJ~725A} \label{sec:run2013}
Unfortunately, most of the acquisition and some of the guiding images from 2013
were lost due to an error in the acquisition software which was later solved in
March 2014\footnote{\label{fn:technical14}Instrument upgrades can be consulted
in http://www.tng.iac.es/instruments/harps/.}. As a result, we could not obtain
consistent measurements over the autoguide images.
As for the 2014 alternating run, the analysis shows the $K$-index clearly correlated
with the FWHM (see Figure \ref{fig:corrspec13}). During this run we covered a
wider range of airmasses -this study includes data only up to 2.5 in airmass-
compared with the 2014 alternating run (which covered up to 1.6 in airmass).
This might have been the cause of the larger span of $K$ values measured in
2013, but a punctual under-correction of the ADC can not be ruled out. The slope
of the correlation law is equivalent to the value measured on the differential
run. The offset could be due to a focus realignment carried out in between the
two runs (during the technical run of March 2014$^{\ref{fn:technical14}}$.)
\begin{figure*}
\includegraphics[width=15cm]{fig_SuperPanelSpecIndexRun13.pdf}
\caption{Spectral indices correlations for the five consecutive high-cadence nights of GJ~725A observed in August 2013 (black crosses). Alternating run observations taken in 2014 are also shown for comparison. The FAP is coded as in Figure~\ref{fig:corrimag} where the low right square refers to the 2013 data. The wider range of AM sampled in 2013 might have uncovered a chromatic effect also in the RVs (center). Blue line and equation correspond to the tentative linear fit we applied to correct the RV data (see Sec.~\ref{sec:detrending}). On the left plot, the negative slope of the 2013 FWHM$-K$ correlation is compatible with the 2014 results, pointing out towards an instrumental (or software) common origin. The FWHM$-$RV is shown in the right panel.}
\label{fig:corrspec13}
\end{figure*}
The wider range of variability in $K$ seems to better support the existence of a relation
between this index and the RVs as well (see Figure\ref{fig:corrspec13}, central plot). The
variability in both the FWHM and the RV suggested that the mean line profile was changing
during the night. However, we could not rule out a problem in the measurements themselves
(i.e. in the algorithms) producing spurious variability in the time-series. To address this, we
performed a number of tests and applied profile measurements independent from the DRS
in the next section.
\subsection{Independent measurements and validation experiments} \label{sec:experiment}
Systematic variability in the measurements can be caused by \emph{intrinsic}
changes in the line profile (i.e. instrumental), by sub-optimal procedures in the measurement of
these quantities (i.e. algorithmic), or both. For example, even if the line-profile is perfectly
stable, inaccuracies in the fitting of the flux or background subtraction
will produce apparent variability of measurements of the line shape.
In particular, the cross-correlation profiles are obtained by
computing some weighted mean of the profiles of the individual spectral orders.
Because the effective line profile of each individual order is different,
changes in the weights (e.g. if the weights are computed using Poisson
statistics from the photon counts) will produce spurious variability correlated
with these changes. In other words, a different SED implies a different local SNR for each order. On the other hand, an inaccurate continuum or background subtraction will cause different mean line profiles at each spectral order. As a result, we can obtain a modified final mean line profile when we combine a variable weight distribution with different profiles at each order. Furthermore, the RV and FWHM values obtained from the final profile will be also different with the SED. Thus, if our methods do not perform a good continuum or background subtraction, we can measure spurious RV or FWHM variabilities even when the instrument is stable.
Such an effect was earlier reported by \cite{bourrier2014} on 55 Cancri. Following their findings, the DRS flux-normalises the continuum of all the spectra with respect to a reference spectrum for F, G and K stars. However this has not been yet implemented for M stars, because, as it was defined, some zero division problems can occur in the bluer part of the spectra (F. Pepe, private comm.)
In order to validate if the algorithms we used could be the cause of the observed variability in the measurements, we have performed the two experiments detailed below.
\subsubsection{Experiment 1 : RV measurement against pSED variability}
HARPS-TERRA fits the flux iteratively. It carries out a least
squares fit to minimise the difference between the observations and a template
(built from the observations). In particular, the magnitude to minimise, $R$, is given by:
\begin{equation}
R(\lambda) = T[\alpha_{v}\lambda] - f[\lambda]\cdot \rm{FN},
\end{equation}
\begin{equation}
\rm{FN} = \sum_{0}^{M}{\alpha_{m}(\lambda-\lambda_{c})^{m}},
\end{equation}
where $T$ is the template, $f$ is the observed spectra, and FN the flux normalization term. Thus, HARPS-TERRA fits simultaneously the Doppler shift ($\alpha_{v}$) and the others parameters which account for the shape of the continuum ($\alpha_{m}$) at each order (m). As a consequence, the $\alpha_{m}$ parameters are also modeling instrumental flux variations effects such as the atmospheric differential refraction or tracking errors.
In order to validate if this simultaneous continuum fitting technique applied by HARPS-TERRA prevents the RVs from being affected by chromatic correlations, we performed the experiment detailed below.
We selected three epochs, one for each star and run, and we changed the flux of their spectra as:
\begin{equation}
f_{\rm{new}}(\lambda, t) = f_{\rm{orig}}[\lambda, t] \cdot [1 + S(t) * (\lambda-\lambda_{\rm{cR}})],
\label{eq:exp}
\end{equation}
where $S(t)$ randomly takes values between 0 and 2.086 $\mu m^{-1}$. This range was selected to ensure positive new fluxes at least for the last 31 orders that were used in the second experiment (detailed in Section~\ref{sec:exp2}), and to result in flux distortions below 20\% (approximately the SED distortion measured in the linear range used to calculate $K$). We repeated this calculation several times, and, in total, we obtained 100 new synthetic spectra for each of the three selected epochs. Afterwards, we post-processed these synthetic spectra with HARPS-TERRA, recovering in all cases a RV with and RMS below $2\times10^{-6}~m\,s^{-1}$ for the three epochs (see Figure~\ref{fig:fluxexpht}).
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_fluxexpht.pdf}
\caption{RVs plotted versus the $K$-index when the input corresponds to the real observations (left plot) or to a set of 100 flux-distorted spectra calculated from an observed epoch (right plot). HARPS-TERRA measures relatives values and centers the RV at $0~ms^{-1}$. The zero-values (RMS below $2\times10^{-6}~m\,s^{-1}$) on the right plot indicate that the normalization carried out by HARPS-TERRA ensures RVs independent on the flux distortions included deliberately to perform this experiment.}
\label{fig:fluxexpht}
\end{figure}
This result indicates that the RVs measured with this template matching
algorithm are mostly insensitive to changes in the SED. HARPS-TERRA is properly
fitting the flux of the spectra, and thus, our intentionally introduced
wavelength distortions can not be the cause of the observed RVs variability
(neither of the RV$-K$ correlation).
\subsubsection{Experiment 2 : line profile width against pSED variability}\label{sec:exp2}
HARPS-TERRA provides an independent measurement of the RV, but we do not have
any equivalent for the FWHM. With the aim of checking if the strongest FWHM$-K$ correlation could be due to a software issue related with the weighting of the orders (see detailed explanation at the beginning of this section), we estimated the profiles using an independent least-squares deconvolution technique \citep[LSD;][]{donati1997}. In particular, we used the implementation given by \cite{barnes1998, barnes2012}.
The least square routine involves solving the mean-line profile which, when optimally convolved with a line list, gives the best match to the observed spectrum. Generally, the line list used is just a theoretical identification of wavelengths and depths. Instead, we used the high SNR template computed by HARPS-TERRA to obtain a more realistic empirically determined list of lines. This template is built by co-adding all the observed 2-dimensional spectra given by the DRS (the so-called e2ds). Those regions of the spectrum with strong telluric absorption or strong stellar molecular bands were not used for the deconvolution.
Like the DRS, the LSD discards some wavelengths in the blue part of the spectra (only those corresponding to the last 31 orders were used, while the DRS uses the reddest 51 orders). However, in contrast to the DRS, the LSD does not calculate one profile per order. The deconvolution is performed on each spectrum using all lines (including those repeated in adjacent orders) to obtain one single LSD profile per exposure. Our LSD method normalises each spectrum to a template to obtain a individual continuum. That is, each spectrum in turn is divided by the template spectrum and a cubic order polynomial is fitted to the residuals. Later, the continuum of each spectrum is obtained by multiplying this cubic polynomial fit by the template continuum, which was previously obtained by iteratively fitting a 5th degree polynomial to the template. Because each spectrum is normalised to a individual continuum, we did not expect variability related with weighting problems arising from the algorithm or background issues (see beginning of this Section).
The output of the LSD technique is a mean line profile in absorption. These $F_i(t)$ profiles can be inverted (sign changed) after have we subtracted their residual continuum $w(t)$, producing a normalised (positively defined) probability distribution function $f(t)$. Then, the moments of $f(t)$ can then be computed as:
\begin{eqnarray}
f_i(t) &=& w(t)-F_i(t) \\
\hat{M}_{n}(t) &=& \frac{ \sum_{i=1}^{N}{[f_{i}(t)]}\cdot v^{n}}{ \sum_{i=1}^{N}{f_{i}(t)}}
\label{eq:moments}
\end{eqnarray}
Before calculating the moments with Eq.~\ref{eq:moments}, we truncated the profiles in a range of $\pm~20~km s^{-1}$ around the estimated zero-velocity (the velocity of the co-added barycentric corrected high SNR template from which we obtained the line list applied during the deconvolution process), and we excluded the negatives values resulting after the continuum inversion. This truncation was done to avoid including wing distortions in the moments calculation. The profiles were calculated with a mean velocity increment of 0.81 $km s^{-1}$ --value that corresponds to the pixel size of the HARPS-N detector--, and later interpolated by a factor 10 using a spline function. The order zero moment, $M_{0}$, corresponds to the integral of the normalized profile. $M_{1}$ is equivalent to its centroid (all spectra are aligned to zero-velocity before computation of the LSD profile), and $M_{2}$ corresponds to the profile variance and thus, as the FWHM, is a measurement of the width of the line profile.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_fluxexplsd.pdf}
\caption{Moment of second order ($M_{2}$) of the LSD profiles versus the $K$-index for the observed (left plot), and for the synthetic (right plot) spectra. The trend for the synthetic data reproduces the observations pointing out a software issue.}
\label{fig:fluxexplsd}
\end{figure}
With the aim of testing the LSD performance, we calculated the LSD profiles of the flux
modified spectra that we used in the previous section to test the software HARPS-TERRA; i.e.
we took the same three spectra (one for each star and run) and we modified their flux
following the Equation~\ref{eq:exp}. In this case, the second order moment (or $M_2$)
calculated from these contaminated profiles, correlated with $K$ in a way analogous to the
observations (see Figure~\ref{fig:fluxexplsd}). This implies that, even with our method that fits the continuum of each spectrum individually, we find a strong correlation of the
measurement of the line width (through $M_2$) with the artificially introduced slope of the
pSED. In other words, the changes in the distribution of the flux along the CCD can drive the
FWHM variability. Then, the largest correlation involving the
FWHM$-K$ relation can be explained as an inaccurate computation of the profile by the
algorithms. Neither the DRS nor the LSD
method produce accurate enough measurements. Actually, the fact that we need to correct a residual continuum $w(t)$ for each deconvolution profile may arise from a non perfect description of the continuum.
According to our experiment, only the RV measurements obtained by the HARPS-TERRA algorithm, seem to be insensitive to the SED variability. This result indicate that new algorithms that simultaneously fit for the continuum of each echelle order and the line profile need to be developed, or the current methods refined to account for this.
To use the FWHM index produced with the M-dwarf mask of HARPS-N, one should -at the very least- apply some decorrelation procedure. Otherwise, all this
index is tracing are small changes in the measured SED produced by the expected flux loses on fibre-fed spectrographs.
\subsection{Detrending strategies for archival FWHM and RV obtained with HARPS}\label{sec:detrending}
Once we clarified that the strongest FWHM-$K$ correlation has an
algorithm origin, we aimed at defining an empirical function to detrend the FWHM
time-series from chromatic effects. The function should be valid for (at least)
the GJ~725A+B pair and other stars with similar spectral types (note that the
chromatic nature of the effect will likely affect different spectral types
differently). We made use of both, 2013 and 2014 data, to compute the detrending
law.
The FWHM contains additional contributions from other effects. In
particular, since the observations span over a finite amount of
time, the FWHM will be broadened by the change in the barycentric
velocity of the observer. The following procedure was designed to remove
the effect of this barycentric broadening (which can account for
several $\rm m\,s^{-1}$ depending on integration time and coordinates of the star).
The central moments of a distribution can be
denoted as $M_{k}=<(x_{i}-\bar{x})~^{k}>$. Thus, $\mathrm{M}_{2}$
= Var(x), the variance of the distribution which, for two
independent distributions, satisfies the additive property: Var(x +
y) = Var(x) + Var(x) \citep{booknumericalrecipes1992}. The
mean line profiles of the spectral orders can be considered as a
probability density function. In consequence, the FWHM should
also be a variance.
The observed FWHM$_{ \mathrm{obs}}^2$ is the sum of three terms:
\begin{equation}
\mathrm{FWHM}_{ \mathrm{obs}}^2 = \mathrm{FWHM}_{ \mathrm{real}}^{2} + \mathrm{B}^{2} + \mathrm{I}^{2},
\end{equation}
where $\mathrm{FWHM_{\rm real}}$ is the true line profile we want to measure, and B and I are:
\begin{itemize}
\item{The barycentric correction (B):}
It is a squared velocity value we have to subtract to the square of the observed FWHM to correct it from the movement of Earth on its orbit. To first order, let us assume that the velocity $v$ of the
observer changes between the initial instant $t_i$ and the
end-of-integration time $t_f$ by $\mathrm{dV}$. The effect of the
barycentric broadening is then a convolution of the stellar profile with the double of the variance of a boxcar distribution ($F(v)$) of width $\mathrm{dV}$:
\begin{equation}
\mathrm{B}^{2}=2 \int_{-\infty}^{\infty}F(v)v^{2}dv = 2 \frac{\mathrm{dV}^{2}}{12},
\end{equation}
where, being C a constant, $F(v)$ is the boxcar function $f(v)$ normalized to have unit area as follows,
\begin{equation}
\,\,\,\,\,\,F(v) = \frac{1}{\mathrm{C\,dV}} f(v),
\end{equation}
\begin{equation}
\,\,\,\,\,\,f(v) = \left\{
\begin{array}{l l}
\mathrm{C} & \quad \text{if $\frac{-\mathrm{dV}}{2}<\mathrm{v}<\frac{\mathrm{dV}}{2}$}\\
0 & \quad \text{if $\mathrm{v}<\frac{-\mathrm{dV}}{2}$ or $\mathrm{v}>\frac{\mathrm{dV}}{2}$}
\end{array}, \right.
\end{equation}
This correction can be exactly computed using the observation,
the exposure time, and a custom made code used to compute the barycentric correction
(implemented within the HARPS-TERRA software).
\\
\item{Chromatic effect correction (I):}
This term contains the correlation with the chromatic $K$
index as computed in previous sections. We searched for the linear
model which best-fitted the observed FWHM corrected from barycentric term ($\mathrm{FWHM}_\mathrm{B}$), and the $K$-index series,
\begin{equation}
\mathrm{FWHM}_{\rm B}^{2}= \mathrm{FWHM}_{\rm obs}^{2}- \rm B^{2}=\alpha + \beta\; K,
\end{equation}
where $\alpha$, the offset, is a nuisance parameter, which is equal
to $(90.53\pm0.02)\times10^{5}\rm\,m^{2}s^{-2}$ for GJ~725A,
$(89.34\pm0.01)\times10^{5}\rm\,m^{2}s^{-2}$ for GJ~725B, and to
$(91.620\pm0.004)\times10^{5}\rm\,m^{2}s^{-2}$ for GJ~725A observed in the 2013.
Thus, we defined the chromatic correction term as
\begin{equation}
\mathrm{I}^{2} = \beta\;K,
\end{equation}
where $\beta = (- 0.620\pm0.003)\times10^{5}\,\rm m^{2}s^{-2}\mu m$
is the average of the values obtained for the three data sets
independently. The low degree of scatter in the values (coefficient of
variation lower than 10 per cent) means that the average is
representative and thus, we can define a common law which is the
same for both stars and all observed nights (see
Figure~\ref{fig:fwhmcorr}, left panel).
\end{itemize}
Finally, the FWHM corrected from barycentric and chromatic effects can be written as:
\begin{equation}
\mathrm{FWHM}_{\rm real}^2 = \mathrm{FWHM}^{2} - 2\,\frac{dV^2}{12} - \beta \;K,
\end{equation}
and its error, obtained by applying simple error propagation functions, as:
\begin{equation}
\sigma_{\mathrm{FWHM}_{\rm real}^2} =\sqrt{ 4\; \mathrm{FWHM}^{2} \; \sigma^2_{\mathrm{FWHM}} + K^2 \sigma^2_{\beta} }
\end{equation}
Results are shown in Figure~\ref{fig:fwhmcorr}. The same slope for 2013 and 2014 data was expected as it is caused by the same instrumental effect. However, results indicated some systematic residual effects, specially in the 2013 data. This extra variability might be related to some lesser degree contributions specific to each run. However, as the main correlation might have the same origin, we still prefer to apply the same empirical law to detrend all the datasets. We also expected the zero-point offset ($\alpha$) of the two runs to be the same for GJ~725A, but the change could be due to an instrument focus readjustment carry out between the two observational campaigns. In spite of these discrepancies, the improvement is
obvious, the RMS is reduced a factor of 5: GJ~725A from 7.73 to
$1.66\rm\,m\,s^{-1}$, GJ~725B from 9.52 down to
$1.85\rm\,m\,s^{-1}$; and a 87 per cent for the 2013 run in GJ~725A, from 16.71 down to $2.15\rm\,m\,s^{-1}$. We want to point out that the reduction of RMS comes from the presence of highly structured noise,
which makes the new time-series of the FWHM a much more reliable
tracer of the physics occurring in the photosphere of the star.
\begin{figure*}
\includegraphics[width=17cm]{fig_fwhmcorr.pdf}
\caption{Left panel: FWHM (barycentric corrected) and $K$-index correlation. Black squares and red dots match GJ~725A and B, respectively. Blue lines and equation correspond to the best-fitting function. We want to emphasise that the empirical law is unique. Note that the chromatic correction consists of detrending the data by subtracting the slope term. Right panel: Time-series before ($\mathrm{FWHM}_{\rm B}^{2}$, top) and after ($\mathrm{FWHM}_{\rm real}^2$, bottom) the chromatic correction. The FWHM is flattened and corrected from chromatic distortions caused by the atmosphere. The RMS is reduced by a factor of 8 in 2013 and by a factor of 5 in 2014.}
\label{fig:fwhmcorr}
\end{figure*}
On the other hand, the intranight systematic in the RVs can be also modeled with a linear
function. \cite{bourrier2014} detected similar RV correlation for
the 55~Cnc high-cadence data observed with HARPS-N. They used the
ratio of SNR of two orders as the detrending quantity, which is a
simpler version of our chromatic $K$ index. As discussed before, we did not find any significant correlation for the 2014 run.
Therefore, we have only detrended the 2013 data, following:
\begin{equation}
\rm RV_{\rm C} = RV - \omega\,K,
\end{equation}
where $\omega = -0.7\pm0.2\;\rm m\,s^{-1}\mu m^{-1}$. This value of $\omega$ is obtained by fitting the linear correlation law to each night independently to avoid
contamination by additional, longer term variability (e.g. planets or
induced activity happening at time-scales of a few days) Finally, the fitting parameters (see equation in the central panel of Figure~\ref{fig:corrspec}) are the weighted mean of the i-th values.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_rvcorr.pdf}
\caption{2013 RV time series of GJ~725A before (top panel) and after (bottom panel) applying the empirical function to detrend the RV-$K$ correlation. The RMS decreases by a factor 1.3.}
\label{fig:rvcorr}
\end{figure}
Detrended RVs, shown in the bottom panel of Figure~\ref{fig:rvcorr},
have a mean peak-to-peak difference of 3.3~$\rm m\,s^{-1}$. The RMS
is reduced from 1.29 down to 1.00~$\rm m\,s^{-1}$, which is close to
the photon noise of the HARPS-N observations.
We advise caution in using the correlation laws defined here to detrend the FWHM and RVs from HARPS-N observations. As we have seen, instrumental updates and (possibly) spectral
types might produce slightly different values for the correlation laws. As a general rule, in case of aiming at detrending archival data, firstly we strongly recommend to compute the chromatic $K$ index, and secondly, verify if there are chromatic effects; that is, verify if there are correlations between the $K$ index and the main data-products (e.g. RVs, FWHM or other indices). Finally, in case of detecting significant correlations, follow the detrending steps explained in this study. We want to note that, in case of detecting correlations with the RVs, it is preferable to include $K$ as a correlation term in the model used to search for Keplerian signals instead of applying pre-whitening like methods \citep[see an example in][]{anglada2014}.
\subsection{Cross-dispersion profiles} \label{sec:spaprof}
The cross-dispersion profile, or the mean profile obtained from making a cut of the \'echelle
spectrum across the orders, is very sensitive to illumination distortions produced at the
fiber entrance such as changes of the telescope focus, seeing increases or pointing errors.
The reason for this is that each cut across the order is a section of the fiber image; in
fact the fiber size defines the width of the order. Therefore, the cross-dispersion profile
is by itself a measure of the instrumental profile and can be used to measure the
illumination stability in the detector. For example, if the wavelength dependence variations
of the flux of the spectra were due to a non-perfect smoothness of the instrumental profile
(e.i. due to an inefficient fiber scrambling system on HARPS-N; \citealp{avila2004}), we
would expect to find a correlation between the width of the cross-dispersion profile and the
$K$ index.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_diagram.pdf}
\caption{Layout of the cross-dispersed profile calculation process. To extract the cross-dispersed profiles (B), the raw images are cut in the spatial axis following the direction of the blue line (A). Simultaneously with the science observations, we recorded emission lines from a ThAr lamp (see bright spots in the raw image). We blocked the lamp lines (B, shaded areas) before selecting the points used to fit and subtract the floor level (B, blue dots). Finally, after normalizing each peak to its maximum, we cross-correlated the profile with the flat raw image to obtain a mean line profile of the column (C).}
\label{fig:digram}
\end{figure}
To measure this profile, we cut the raw images across the cross-dispersed axis, extracting in that way the flux at each column of the reddest half of the CCD (see diagram in Figure~\ref{fig:digram}). After blocking the lamp emission, which is recorded in the science image simultaneously using a secondary fiber, we fitted and subtracted the floor level using a polynomial. The counts obtained in between consecutive science orders were compatible with the bias level obtained during the standard calibration performed at the beginning of the night. However, we fitted a polynomial instead of subtracting a constant bias level to better account for other possible background effects that can introduce biases in the final width of the cross-dispersion profile. Then, we normalised each order to its maximum and we cross-correlated each column of the raw image with the corresponding flat column. After that, we fitted a Gaussian to this cross-correlated profile, calculating the cross-dispersed FWHM at each column as:
\begin{equation}
\mathrm{FWHM} = 2\sigma\sqrt{2\log{2}},
\end{equation}
where $\sigma$ corresponds to the width of the fitted Gaussian. We note that this function includes a fourth parameter (an offset) in an attempt to model the background flux variabilities.
Because the detector is flat, the image of the fiber onto the detector --and thus the FWHM in the cross-dispersed direction-- increases towards the detector edges. To avoid this effect, we took the differential cross-dispersed FWHM measurements of each column with respect to the first epoch. Furthermore, the detector edges are more sensitive to illumination changes (see table~7 of the HARPS-N User Manual\footnote{http://www.tng.iac.es/instruments/harps/}), and this cause the measurements at the edge columns to be more sensitive to SNR variabilities. Aiming at correcting the SNR dependencies of this origin, we detrended the cross-dispersed FWHM using the fitted offset parameter from the Gaussian model. The final measurement for each exposure ($\rm FWHM_{\rm cd}$) was obtained as the median along the column detrended values.
We have also calculated the centroids in the cross-dispersed axis. This measurement corresponds to the mean parameter of the Gaussian function fitted to the profile. The cross-dispersed centroid (or $\rm CEN_{\rm cd}$) of the exposure was directly defined as the median of all the column values. In this case, we could not use the background level to detrend this measurements because, as it was expected, the centroid of the Gaussian did not move with changes of the background flux level; as a consequence we still had some correlation of the $\rm CEN_{\rm cd}$ with the SNR.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_spatialprof.pdf}
\caption{FWHM (upper panel) and centroids (lower panel) fitted to the mean line profile calculated on the cross-dispersed direction (perpendicular to the spectral axis). No significant correlation with the $K$ index was measured, ruling out a scrambling issue. The $\rm FWHM_{\rm cd}$ values are referred to the first exposure of the series. The average values of the three $\rm CEN_{\rm cd}$ series were subtracted to better compare its scattering.}
\label{fig:spaprof}
\end{figure}
In spite of the corrections explained above, we detected the $\rm FWHM_{\rm cd}$ to be more
dispersed at low SNRs. Therefore, we performed an additional experiment to discard possible
non-linearity effects of the detector as a source of variability. We let the values of the cross-correlated profiles
of the 20\% of the exposures with worse SNR to randomly vary $\pm~0.2$~\%, total amount that could be associated to a non-linearity effect\footnote{See the ESO report
about the EEV44-82 CCDs at: https://www.eso.org/sci/facilities/develop/detectors/optdet/
docs/reports/EEV-report.html}. However, we found no significant changes, discarding the non-linearity as the responsible for any change in the $\rm FWHM_{\rm cd}$.
If the spectral FWHM variability was due to illumination effects (instead of being a pure
software issue), the variation of the cross-dispersed FWHM with $K$ would be significant.
However, as we show in Figure~\ref{fig:spaprof}, the scatter of the $\rm FWHM_{\rm cd}$ is
below $1~mpxl$ ($1~mpxl\sim0.8~m\,s^{-1}$). Moreover, we did not detect any significant
correlation with $K$. The cross-dispersed centroids, whose maximum RMS was only $0.8~m\,s^{-1}$, do not correlate with K either, opposite to what we detected for its spectral counterpart in the 2013 observations.
In consequence, we do not detect significant distortions of the image onto the detector through the measurements in the cross-dispersed direction. Bearing in mind that, in the case of the $\rm CEN_{\rm cd}$ we still deal with some correlation with the SNR, this result indicates that, neither the large variability detected on the spectral FWHM nor the RV-$K$ correlation measured for the 2013 data, are due to a non-perfect smoothness of the inhomogeneities of the injected light spot. This result validates the use of the cross-dispersed profiles as a useful test to account for
illumination issues.
\subsection{Possible GJ~725B planet candidate?} \label{sec:dopplerb}
In this section we discuss a possible planet hosted by GJ~725B. As we pointed out in Figure~\ref{fig:difrunper}, the F-ratio excess detected in the RV differential periodogram could be compatible
with a planet orbiting either GJ~725B or GJ~725A.
To assess the significance of the variability and its
possible fit to a Keplerian orbit, we use likelihood
periodograms as described in \citet{baluev2009} and in \citet{anglada2013}. The model:
\begin{equation}
v_{r}(t) = \gamma+A\cos\left(\frac{2\pi}{P}\right)\Delta t+B\sin\left(\frac{2\pi}{P}\right)\Delta t,
\end{equation}
includes two sinusoids (2
parameters, A and B) plus and velocity offset (+1 parameter, $\gamma$). We used this circular model after have obtained an eccentricity compatible with zero with the complete Keplerian model. Like in classic periodograms, the likelihood periodograms produce a map of peaks for each of the investigated periods. However, in this case the power of the peaks is defined by the model that maximises the logarithmic of the likelihood statistic (\emph{ln-}$\mathcal{L}$). The \emph{ln-}$\mathcal{L}$\ maximisation and $\chi^2$
minimisation are equivalent, but the likelihood
method has the advantage, with respect to
traditional periodograms, of including the so-called jitter as a
free parameter (excess white-noise, +1
parameter). Using the likelihood statistic has a
further advantage: \emph{ln-}$\mathcal{L}$\ gives a probability, and the
ratio of likelihoods $\Delta$\emph{ln-}$\mathcal{L}$$\,=
\mathcal{L}_{\mathrm{null}}-\mathcal{L}_{\mathrm{model}}$
directly estimates the relative probabilities between
models. Properties of the likelihood periodograms are
described in full detail in \cite{baluev2009,
baluev2012}.
The likelihood periodogram analysis associates the peak
detected in the RVs differential periodogram with a $2.7\pm
0.3$~day signal with an amplitude of $1.2\,\rm m\,s^{-1}$. This
means that a model of a sinusoid at that period has a
$\Delta$\emph{ln-}$\mathcal{L}$ = 32.79 when compared with a model with no signal.
This corresponds to a false alarm probability (FAP) $<
1.5\times 10^{-11}$, which is much smaller than the usual 1\%
threshold. A compatible period is obtained when analyzing
GJ~725B RVs independently, indicating that the variability
comes from this star. Besides the common variability in both
stars, no additional significant RV variability is detected on
GJ~725A.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{fig_dopplerb.pdf}
\caption{GJ~725B planet candidate signal. Red dots in the top panel are the GJ~725B RVs for the 2014 run. The black line is the circular Keplerian model best-fitting the data (P~=~2.7~d, $\rm K~=~1.2~\rm m\,s^{-1}$, $\rm FAP~<~0.1\%$ and $\Delta$\emph{ln-}$\mathcal{L}$$~=~49.13$). Residuals to this model are plotted in the lower panel. The RMS decreases from 1.16 to 0.76~$\rm m\,s^{-1}$ on the residuals. Time is referenced to the first exposure of the run.}
\label{fig:doppler}
\end{figure}
The variability in GJ~725B is consistent with a 1.2~\hbox{M$_{\oplus}$}\
planet with an orbital period of $2.7\pm0.3$~d period (or
0.025 AU, see Figure~\ref{fig:doppler}). While the statistical
significance is very high (a sinusoidal model is massively
preferred over no signal at all), its nature cannot be
confirmed because we cannot verify its strict periodicity.
Doppler follow-up observations and/or transit searches will be
needed to confirm it.
\section{Discussion and Conclusions} \label{sec:conclusions}
We obtained high-cadence observations using the HARPS-N spectrograph
on the nearby M-dwarf binary system GJ~725A+B. The two stars alternating
observations unveiled common strong systematic effects in the measurements
of the width of the mean line profile (through the FWHM), and also in the RV
measurements to a lesser degree. The presence of these systematic
effects is a likely component of the floor noise observed in long-term Doppler programs of M-dwarf stars, and seriously affects the use of high-cadence observations for very low amplitude signal searches ($<2\,\rm m\,s^{-1}$).
The systematic effects seem to be related to flux loses due to imperfect corrections of
chromatic effects introduced by the Earth's atmosphere at the telescope-fibre interface. This
suspicion is motivated by several measurements and correlations observed in pre-fibre images
of the star compared to a number of measurements on the spectrum, most notably changes in the
slope of the measured SED. The comparison of measurements between the two stars (and their
almost identical systematic behaviour) shows beyond reasonable doubt that the significant
variability observed with high-cadence on M-dwarfs has an instrumental origin. While
most variability in the width of the line profile is likely to be caused by the algorithms
used, smaller residual RV variability remains unexplained.
HARPS-N includes an ADC which corrects the atmospheric dispersion; however other sources of
error such as the atmospheric extinction or the wavelength dependence with the seeing remain
uncorrected. A direct consequence of a non-optimal correction of the Earth's atmosphere is a
superposition of wavelength-dispersed images of the star at the fiber entrance. This causes
the light injected into the fiber to vary in wavelength as the energy peak (used for
centering the image in the fibre during exposure) changes with the airmass. To measure
whether this was causing distortions in the spectra, we defined the $K$ index, which measures
how the flux is distributed across the detector during the observations (i.e. accounts for
the SED variability). Besides, we measured variability of the $K$ index. This, as well as
with the airmass, resulted to be correlated with the distortions on the pre-fiber images.
However, while these SED variabilities cause changes in the flux, this does not necessarily
imply that the instrumental profile shape changes too (neither the data-products) due to
illumination effects. In fact, to avoid illumination dependencies, all current
high-resolution fiber-fed spectrographs apply scrambling methods \citep{avila2004}.
Therefore, a correlation of the SED variability with the data-products is not expected.
Nevertheless, we found a strong correlation of the SED variability with the line-width (through
the FWHM) and a correlation with the RVs only for our 2013 single star run on GJ~725A.
Further measurements of the spectral order shapes in the cross-dispersed direction
ruled out an inefficient scrambling as the origin for these correlations. These measurements
also validate the use of the cross-dispersed profiles as a useful test to account for
illumination issues on other echelle spectrographs.
We performed a couple of test to validate the algorithms used in this work. Results indicate that, whereas HARPS-TERRA properly corrects and models the continuum, the DRS is sensitive to changes in the slope of the SED. This computational issue explains the strong FWHM-$K$ correlation, however, 2013 RV-$K$ correlation remains unexplained and has to be caused by another effect. Since the width of the mean line profile is a parameter very sensitive to magnetic activity events \citep[e.g.][]{reiners2013}, the FWHM is often used as an activity indicator and a tool to decorrelate Doppler time-series. Thus, there is a value in detrending this index. We outline a procedure to decorrelate the HARPS-N FWHM measurements using the slope of the SED. A tentative decorrelation law is also proposed for the 2013 RVs. The decorrelation laws are likely to be slightly different on each target (especially for targets of different spectral types), so we advise to obtain the SED slopes changes with the $K$ index and perform similar verifications on all individual targets of a given programme.
Ideally, new algorithms that simultaneously fit for the continuum and the line profile need to be developed to account for the FWHM variability. Other option is to refine the current methods, as it was already done in the last DRS version for F, G and K stars, where the continuum is re-normalised with respect to a reference spectrum. Regarding the
variability of the RVs in the 2013 single run in GJ~725A, given the large
correlation with the airmass and the erroneous updating of the ADC movement pointed out during other campaigns, we suspect that the most likely explanation for it was the
sub-optimal performance of the atmospheric correction. The possibility of a systematic ADC
failure only during continuous mode observations (where the pointing procedure is not redone between
same object exposures) was ruled out by the HARPS-N core team (R. Cosentino, priv. comm.). However, our in-situ monitoring of the ADC parameters during posterior HARPS-N campaigns showed that, under some unknown circumstances, the ADC values seem no to be updated. Therefore, we suspect that the ADC failed during this 2013 run.
Remaining correlations with the image distortions might be the origin for the noise floor level which popped up as common signals in the GJ~725A+B alternating run. As a by-product, we have shown that the GJ~725A+B pair is stable enough to be used as a benchmark case for commissioning of future high precision spectrographs, at least down to $\sim1-2\,\rm m\,s^{-1}$. Indeed, CARMENES spectrograph \citep{quirrenbach2014} is planning to use this system for validation purposes. The $\sim1-2\,\rm m\,s^{-1}$ limit is possibly set by a low-mass companion orbiting GJ~725B, whose presence will be further investigated in future campaigns.
Long-term surveys (like the HARPS-GTO program), try to observe all the stars at the same
airmass each night, minimizing sources effects related to airmass and scrambling, and
randomizing possible systematic shifts occurring within a night. However, this is not
possible for programs which require continuous high-cadence observations. Many science
cases, such as the molecule detection on transiting exoplanets
\citep{snellen2010, martins2013}, can benefit from an extremely stabilized instrumental
profile and robust procedures to measure it. Concerning Doppler spectroscopy, we note that an
extra effort should be put into the image stabilization and correction of chromatic effects
such as those reported here ($\sim1\,\rm m\,s^{-1}$) in order to clean spurious signals and
get the most of the new technologies for wavelength calibration at $\rm cm\,s^{-1}$ levels,
like laser frequency combs \citep{probst2014}. The best wavelength determination is useless
if systematic signals populate the periodograms.
\begin{table*}
\begin{center}
\caption{Radial Velocities observed for GJ~725A and GJ~725B with HARPS-N. The columns are: barycentric Julian date, RVs calculated with HARPS-TERRA and their errorbars, FWHM of the cross-correlation function, signal-to-noise at the spectral order centered at 631 nm, airmass, the $\Phi$ and $\Theta$ indices obtained from the autoguide camera images, the $K$-index measured over the spectra, the FWHM and the RV corrected from intranight systematics.}
\label{tab:data}
\begin{tabular}{lcccccccccc}\hline
OBJECT &BJD & RV & FWHM & SNR & AM & $\Phi$ & $\Theta$ & $K$ & $\mathrm{FWHM}^{2}_\mathrm{real}$ & $\mathrm{RV_{C}}$ \\
& & $\rm m\,s^{-1}$ & $\rm m\,s^{-1}$ & at 631~nm& arcseconds & pxl & arcseconds & $\mathrm{\mu m}^{-1}$ & $(\rm{\rm m\,s^{-1}})^{2}$& $\rm m\,s^{-1}$ \\ \hline
GJ~725A & 2456518.36 & 0.88 & 3024.13 & 66.48 & 1.22 & --- & --- & 0.00 & 3024.13 & 0.88 \\
& 2456518.37 & 1.42 & 3030.39 & 72.79 & 1.22 & --- & --- & -0.34 & 3026.87 & 1.18 \\
& 2456518.37 & 1.98 & 3030.63 & 63.74 & 1.21 & --- & --- & -0.50 & 3025.55 & 1.63 \\
& ... & ... & ... & ... & ... & ... & ... & ... & ... &\\
& 2456841.39 & -0.71 & 3002.59 & 67.28 & 1.55 & 6.85 & -0.04 & 0.68 & 3009.60 & ---\\
& 2456841.40 & -0.86 & 2996.76 & 65.87 & 1.48 & 6.93 & 0.00 & 1.26 & 3009.73 & ---\\
& 2456841.41 & -0.36 & 2991.07 & 53.23 & 1.42 & 7.85 & 0.00 & 1.76 & 3009.26 & ---\\
& ... & ... & ... & ... & ... & ... & ... & ... & ... &\\
GJ~725B & 2456841.39 & -0.46 & 2984.25 & 52.24 & 1.51 & 7.24 & 0.01 & 0.66 & 2991.11 & ---\\
& 2456841.40 & -0.24 & 2985.22 & 59.62 & 1.45 & 6.70 & -0.05 & 0.75 & 2993.00 & ---\\
& 2456841.42 & -0.63 & 2978.01 & 58.27 & 1.39 & 6.58 & 0.09 & 1.50 & 2993.54 & ---\\
& ... & ... & ... & ... & ... & ... & ... & ... & ... &\\ \hline
\end{tabular}
\end{center}
This table is available in its entirety in a machine-readable form in the online journal. A portion is shown here for guidance regarding its form and content.
\end{table*}
\section*{acknowledgements}
We thank E. S\'anchez Blanco, M. C. C\'ardenas V\'azquez, N.
Piskunov, F. Pepe, C. Lovis, R. Cosentino, X. Dumusque, D. Staab and C. Haswell for constructive comments and discussions. We acknowledge funding from
AYA2011-30147-C03-01 by MINECO/Spain, FEDER funds/EU, and
2011-FQM-7363 of Junta de Andaluc\'ia/Spain (ZMB, PJA and CR-L); ZMB
acknowledges support from the UGR/Spain and financial funding from
FPI BES-2011-049647 of MINECO/Spain. This study is based on
observations made with the Italian Telescopio Nazionale Galileo
(TNG) operated on the island of La Palma by the Fundaci\'on Galileo
Galilei of the INAF (Istituto Nazionale di Astrofisica) at the
Spanish Observatorio del Roque de los Muchachos of the Instituto de
Astrof\'isica de Canarias. The authors thank the referee T. B\"ohm for his
suggestions that helped improved this paper.
|
\section{Introduction}
\label{sec:motivation}
In this report we address a rather neglected but very important
research area in computational mathematics, namely the numerical
approximation of nonlinear functionals and functional differential
equations (FDEs).
FDEs are arise naturally in many different areas of mathematical physics.
For example, in the context of fluid dynamics,
the Hopf equation \cite{Hopf}
{\color{r}
\begin{equation}
\frac{\partial \Phi([\bm \theta],t)}{\partial t}=
\sum_{k=1}^3\int_V\theta_k(\bm x)\left(i \sum_{j=1}^3\frac{\partial }{\partial x_j}
\frac{\delta^2 \Phi([\bm \theta],t)}{\delta \theta_k(\bm x)\delta\theta_j(\bm x)}
+\nu \nabla^2\frac{\delta \Phi([\bm \theta],t)}{\delta \theta_k(\bm x)}\right)d\bm x,
\label{hopfns}
\end{equation}
was deemed by Monin and Yaglom (\cite{Monin2}, Ch. 10)
to be ``the most compact formulation
of the general turbulence problem'', which is the problem of
determining the statistical properties of the velocity
and the pressure fields of Navier-Stokes equations given
statistical information on the initial state\footnote{Stani\v{s}i\'c \cite{Stanisic} refers to the Hopf equation \eqref{hopfns}
as the ``only exact formulation in the entire field of
turbulence'' (Ch. 12, p. 233).}.
In equation \eqref{hopfns} $V\subset \mathbb{R}^3$ is
a periodic box, $\bm
\theta(\bm x)=(\theta_1(\bm x), \theta_2(\bm x),
\theta_3(\bm x))$
is a vector-valued test function in a suitable divergence-free
space, and $\Phi$ is a nonlinear complex-valued functional
known as Hopf functional \cite{Hopf}. Remarkably,
with such functional available it is possible to
compute any statistical property of
the velocity field that solves the Navier-Stokes
equations (see \cite{Monin2}).
This is of great conceptual importance: the solution to
one single linear functional differential equation can
describe all statistical features of turbulence and there
is no need to refer back to the Navier-Stokes
equations.
From a mathematical viewpoint the Hopf
functional is basically a time-dependent nonlinear operator
in the space of test functions $D(\Phi)$ (domain of the operator
$\Phi$) with range in the complex plane (see Figure \ref{fig:1}).
\begin{figure}
\centerline{
\includegraphics[height=6.5cm]{fig1_new-eps-converted-to.pdf}}
\caption{
Sketch of the mapping at the basis of the
Hopf functional. The domain of the functional $\Phi$,
denoted as $D(\Phi)$ is a suitable
of space of functions while the range of $\Phi$, denoted as
$R(\Phi)$, is a subset of the complex plane $\mathbb{C}$. The
approximation space $D_m$
is a subset of $D(\Phi)$, which is mapped into $R_m$.}
\label{fig:1}
\end{figure}
The operator can be formally defined as a functional integral
\begin{equation}
\Phi([\bm \theta],t)=\int_{\Omega} \exp\left[i\int_{V}
\bm u (\bm x,t;\omega)\cdot \bm \theta(\bm x)d\bm x\right]P([\bm u_0]) \mathcal{D}[\bm u_0]
\label{HcF}
\end{equation}
where $\bm u(\bm x,t; \omega)$ is a stochastic solution to the
Navier-Stokes equations and $P([\bm u_0])$ is the
probability functional of the random initial state
(assuming it exists).
Thus, computing the solution to the Hopf equation \eqref{hopfns}
is equivalent to compute a (complex-valued)
time-dependent nonlinear operator $\Phi$ from an
equation that involves classical
partial derivatives with respect to space and time variables
as well as derivatives with respect to functions, i.e.,
functional derivatives $\delta /\delta
\theta_j(\bm x)$ \cite{Vainberg,Nashed}.}
Another example of functional differential equation
is the Schwinger-Dyson equation of quantum field
theory \cite{Easther,Justin}. Such equation describes the
exact dynamics of the Green functions of a general field theory,
and it allows us to propagate field interactions, either in a
perturbation setting \cite{Okopinska} (weak coupling regime) or
in a strong coupling regime \cite{Swanson}. The Schwinger-Dyson
formalism is also useful in computing the statistical properties of
stochastic dynamical systems. For example, consider Langevin
equation
\begin{equation}
\frac{d \bm \psi(t)}{dt}=\bm G(\bm \psi(t),t)+\bm f(t;\omega),
\label{ODE}
\end{equation}
where $\bm f(t;\omega)$ is random noise.
Define the generating functional \cite{Phythian,Jensen}
\begin{equation}
Z([\bm \xi,\bm \eta])=Z_0 \int \mathcal{D}[\bm \psi] \mathcal{D}[\bm \beta]A([\bm \psi,\bm \beta])\exp\left[\int_0^t d\tau \left(\bm \xi(\tau)\cdot \bm \psi(\tau)+\bm \eta(\tau)\cdot \bm \beta(\tau)
\right)\right],
\label{GF}
\end{equation}
%
where $Z_0$ is a normalization constant and
\begin{equation}
A([\bm \psi,\bm \beta])= C([\bm \beta])\exp\left[-\frac{1}{2}
\int_0^t d\tau \nabla \cdot \bm G(\bm \psi(\tau),\tau) -i
\int_0^t d\tau \bm \beta(\tau)\cdot\left(\frac{d\bm \psi(\tau)}{dt}-\bm G(\bm \psi(\tau),\tau) \right)\right].
\label{AU}
\end{equation}
The functional $C([\bm \beta])$ in \eqref{AU} denotes the (known) characteristic functional of the external random noise $\bm f(t;\omega)$. Clearly, if we have available the stochastic solution to \eqref{ODE}, then we can construct the functional $Z([\bm \xi, \bm \eta])$ and compute all statistical properties we are interested in. On the other hand, it is straightforward to show that $Z([\bm \xi, \bm \eta])$ satisfies the following system of linear FDEs\footnote{
The expression
\begin{equation}
G_k\left(\frac{1}{i}\frac{\delta}{\delta \bm \xi (\tau)},\tau\right)
\end{equation}
in equations \eqref{SD1} and \eqref{SD2} has
to be interpreted in the sense of symbolic operators.
For example, in one dimension, if $G\left(z,t\right)=z+z^2$ then
\begin{equation}
G\left(\frac{1}{i}\frac{\delta}{\delta \xi (\tau)},t\right)Z = \frac{1}{i}\frac{\delta Z}{\delta
\xi(\tau)}-\frac{\delta^2 Z}{\delta \xi(\tau)^2}.
\end{equation} }
(Schwinger-Dyson equations)
\begin{align}
\frac{1}{i}\frac{\partial }{\partial \tau}\frac{\delta Z}{\delta \xi_k(\tau)}
&=\eta_k(\tau)Z + G_k\left(\frac{1}{i}\frac{\delta}{\delta \bm \xi(\tau)},
\tau\right)Z - iD_k\left(\left[\frac{1}{i}\frac{\delta}{\delta \bm
\eta(\tau)}\right],\tau\right)Z,\label{SD1}\\
\frac{1}{i}\frac{\partial }{\partial \tau}\frac{\delta Z}{\delta \eta_k(\tau)}&=- \xi_k(\tau)Z + i\sum_{j=1}^n\frac{\delta}{\delta\eta_j(\tau^+)}\frac{\partial G_k}{\partial \psi_j}\left(\frac{1}{i}\frac{\delta}{\delta \bm \xi(\tau)},\tau\right)Z,\label{SD2}
\end{align}
where
\begin{equation}
D_i([\bm \beta],\tau)=\frac{\delta C([\bm \beta])}{\delta \beta_i(\tau)}.
\end{equation}
The solution to the Schwinger-Dyson equations
\eqref{SD1}-\eqref{SD2} is a nonlinear functional
(i.e., a nonlinear operator) $Z([\bm \xi, \bm \eta])$ which allows
us to compute all statistical properties of the system without any
knowledge of the stochastic process $\bm \psi(t;\omega)$ defined
implicitly by the stochastic ODE \eqref{ODE}. By generalizing
\eqref{GF}, it is possible to derive a functional formalism for any
classical field theory or stochastic system. This yields, in particular,
Schwinger-Dyson-type equations for generating functionals associated
with the solution to stochastic partial differential equations (SPDEs).
If the SPDE admits an action functional, then the construction of the
generating functional as well as the derivation of the corresponding
Schwinger-Dyson equation are rather straightforward (see
\cite{Jensen,Amit,Kleinert}).
The usage of functional differential equations
grew very rapidly during the sixties, when it became
clear that techniques developed for quantum field theory
by Dyson, Feynman, and Schwinger could be applied, at least
formally, to other branches of mathematical physics.
The seminal work of Martin, Siggia, and Rose \cite{Martin} became
a landmark on this subject, since it revealed the possibility of
applying (at least formally) quantum field theoretic methods, such as
functional integrals and diagrammatic
expansions \cite{Phythian,Jensen,Phythian1,Jouvet},
to classical physics. Relevant applications of
these techniques can be found in non-equilibrium statistical mechanics
\cite{Jensen,Phythian,Phythian1,Jouvet,Eyink_1996,Langouche,Ueda,Ueda1},
stochastic dynamics \cite{Hochberg,Venturi_PRS,Klyatskin1},
and turbulence theory
\cite{Frisch,McComb,Eyink,Chen_KraichnanPRL,Giles,Dopazo,Alankus,Lewis, Monin1,Monin2,Rosen_1960,Rosen_1967,Rosen_1969,Hopf_book,Hosokawa}.
An open question that has persisted over the years is: How do we
compute the solution to a functional differential equation?
From the fifties to the eighties,
researchers were of course investigating analytical methods, e.g.,
based on functional power series
\cite{Rosen_1960,Rosen_1967,Volterra,Nelles}, functional integrals
\cite{Kleinert,Lobanov,Popov,Egorov}, transforms with respect to
appropriate measures (\cite{Monin2}, p. 802), and diagrammatic
expansions.
More recently, Waclawczyc and Oberlack \cite{Oberlack,Waclawczyk}
proposed a Lie group analysis and applied it to the Hopf-Burgers equation,
which represents a step forward toward developing new analytical solution
methods. Specifically, invariant solutions of the Hopf-Burgers equation were
found based on the analysis of the infinitesimal generator of
suitable symmetry transformations.
From a numerical viewpoint,
recent advances in computational mathematics -- in particular in
numerical tensor methods \cite{Hackbusch_book} -- open the possibility
to solve functional differential equations on a computer.
In this report, we will present state-of-the-art
mathematical techniques and numerical algorithms to
represent nonlinear functionals and compute the
numerical solution to functional differential equations.
If FDEs are so important, why do they not have a prominent role in
computational mathematics? There are several possible answers to this
question. First of all, FDEs are infinite-dimensional equations, in the sense that
they are, in principle, equivalent to an infinite-dimensional system of PDEs, or
PDEs in an infinite number of variables. This
may have understandably discouraged researchers in numerical analysis to
even attempt a numerical discretization. Most schemes proposed so far are
based on truncations of infinite hierarchies of PDEs obtained, e.g,
from functional power series expansions
\cite{Rosen_1960,Rosen_1967,Rosen_1969,Rosen_1971,Monin2,Ahmadi,Frisch},
or Lundgren-Monin-Novikov hierarchies
\cite{Wacawczyk1,Friedrich,Lundgren,Hosokawa,Rosteck}.
Other approaches are based on a direct discretization of the
functional integral \cite{Kleinert,Egorov,Popov,Langouche}
that defines the field theory (e.g. $Z$ in
equation \eqref{GF}), and its evaluation using Monte Carlo
methods, or source Galerkin
methods \cite{Guralnik1,Guralnik2}.
Dealing with systems of infinitely many PDEs or
very high-dimensional PDEs can indeed be discouraging,
but nowadays it is quite common,
for example when discretizing stochastic systems driven by colored
random noise or stochastic partial differential
equations (SPDEs) \cite{HandyGK,Db_book,Xiao2,db2,DanieleWick}.
Another reason why FDEs have not yet been numerically studied
extensively may be due to a lack of awareness of their existence
within the computational mathematics community. Also, there is no
universal agreement across scientific disciplines as to even the basic
definition of an FDE. For example, most applied mathematics literature
refers to FDEs as ordinary differential equations with memory or delay
terms \cite{Wu,Azbelev,Hale}. In the pure mathematics community,
functional equations have been studied in the context of approximate
homeomorphisms (the Ulam stability problem) \cite{Rassias1,Rassias}, or
more generally within problems where the unknown is a function, e.g.,
Cauchy or d’Alambert functional equations \cite{Sahoo}. The physics
literature, on the other hand, clearly identifies FDEs as those equations
whose unknown is a functional (i.e., a nonlinear operator) and that involve
partial derivatives with respect to independent variables (e.g., space and
time), as well as derivatives with respect to functions (functional
derivatives). These kinds of equations are usually far more challenging
than the functional equations studied by the pure mathematics community,
and indeed there are very few general theorems on the existence and the uniqueness of their solution \cite{Hale,Foias}.
In this report we take the physicist viewpoint
and consider linear functional
differential equations in the form
\begin{equation}
\frac{\partial F([\bm \theta],t)}{\partial t} =L([\bm \theta ],t)F([\bm \theta ],t)+H([\bm \theta θ],t), \qquad F([\bm \theta],0)=F_0([\bm \theta]),
\label{functDE}
\end{equation}
where $F ([\bm \theta], t)$ is a real or complex-valued functional (time-dependent nonlinear operator in a space of functions), $F_0([\bm \theta])$ is a given initial condition, $L([\bm \theta],t)$ is a linear operator in the space of nonlinear functionals, and $H([\bm \theta],t)$ is a known forcing functional.
The linear operator $L([\bm \theta], t)$ usually involves functional derivatives with respect to $\bm \theta(\bm x)$ as well as partial derivatives with respect to independent variables, e.g,
space and time coordinates.
For example, $L([\bm \theta],t)$
could be the linear operator defining
the right hand side of equation \eqref{hopfns}.
We emphasize that the class of equations in the
form \eqref{functDE} is very broad as it encompasses
FDEs describing many physical systems, including statistical properties of
nonlinear SODEs and SPDEs (e.g., Hopf characteristic functional equations
\cite{Hopf,Hopf_book,Monin2,Klyatskin1} or equations for probability
density functionals \cite{Giles,Beran,Dopazo}), functional equations arising
in control theory \cite{Bensoussan_DaPrato},
generalized principles of least actions \cite{Daniele_JMathPhys},
and functional equations of quantum field theory
\cite{Justin,Itzykson,Phythian,Jouvet,Jensen,Langouche}.
To the best of our knowledge, no effective numerical methods have yet
been developed to compute the solution to linear functional differential
equations in the form \eqref{functDE}, and little has been done
for functional differential equations in general, despite their
fundamental importance in many areas of
mathematical physics.
The purpose of this report is to fill this gap and
present state-of-the-art mathematical techniques, including
new classes of numerical algorithms, to approximate
nonlinear functionals and the numerical
solution to functional differential equations in
the form \eqref{functDE}
This report is organized in two parts:
\begin{enumerate}
\item{\bf Approximation of Nonlinear Functionals}
A nonlinear functional is a particular type
of nonlinear operator from a space of functions into
a vector space, e.g., $\mathbb{R}$ or $\mathbb{C}$.
Therefore, the process of approximating a nonlinear functional
is basically the same as approximating a nonlinear operator
\cite{Howlett,Torokhti,Bensoussan_DaPrato}.
In this report we will present various techniques for nonlinear functional
approximation, ranging from polynomial functional series expansions,
to expansions based on stochastic processes, and
functional tensor methods. Within the context of polynomial
functional series expansions we will discuss in particular Lagrange
interpolation in Hilbert and Banach spaces
\cite{Makarov,Prenter,Prenter1,Bertuzzi,Allasia,Chaika},
where the interpolation ``nodes'' are functions in a suitable
function space. We will also discuss series expansions based
on functional tensor methods \cite{Hackbusch_book}.
This class of methods relies on recent developments on
multivariate function approximation such as canonical polyadic (CP)
\cite{Reynolds,Beylkin,Kolda} and hierarchical Tucker (HT)
\cite{Bachmayr,Grasedyck2017} series expansions.
\item {\bf Approximation of Functional Differential Equations}
\label{item:2}
A functional differential equation is an equation whose
solution is a nonlinear functional, i.e., a nonlinear operator.
The equation usually involves functional derivatives of
such functional and partial derivatives
and integrals with respect to independent variables.
The goal of approximation theory for functional differential
equations is therefore to determine an approximation of
such nonlinear functional, e.g., its time evolution given an
initial state.
In this report we will discuss new classes of methods that
extend classical Galerkin, least-squares, and collocation
techniques to functional differential equations.
These methods are based on suitable
representations of the solution functional, e.g.,
in terms of polynomial functionals or functional
tensor networks.
\end{enumerate}
This paper is organized as follows:
In Section \ref{sec:NonlinearFunctionals} we briefly review
what nonlinear functionals, functional derivatives and
provide useful examples of nonlinear functionals in physics.
In Section \ref{sec:rep-Hopf} we address the approximation
of nonlinear functionals and functional derivatives.
In particular, we discuss Lagrange interpolation in
spaces of infinite dimensions (function spaces),
series expansions in terms of polynomial functionals, and
functional tensor methods such as canonical polyadic
and hierarchical Tucker expansions.
In Section \ref{sec:FDEs approximation} we discuss functional
differential equations. We begin by presenting several examples
of FDEs and show how they arise in the context of
well-known physical theories.
In Section \ref{sec:FDEapprox} we address the problem
of computing the numerical solution to an FDE. Specifically, we introduce infinite-dimensional extensions of least squares, Galerkin, and collocation methods.
Finally, in Section \ref{sec:numerical results functionals} and
Section \ref{sec:numerical results functional
equations} we present numerical results on nonlinear functional
approximation and also compute the numerical solution
to a prototype functional advection-reaction problem.
\section{Nonlinear Functionals}
\label{sec:NonlinearFunctionals}
Let $X$ be a Banach space of functions. A nonlinear functional
on $X$ is a nonlinear operator $F$ that takes in an element
$\theta$ of $X$ (i.e., a function), and returns a real or a
complex number. The functional $F$ usually does not operate
on the entire linear space $X$ but rather on a subset set of$X$,
which we denote as $D(F)\subseteq X$ (domain of the functional).
Let us first provide simple examples of nonlinear functionals.
\vspace{0.7cm}
\noindent
{\em Example 1:}
Consider
\begin{equation}
F([\theta])=\int_0^1 x^3 e^{-\theta(x)}dx, \qquad
\theta\in D(F)=C^{(0)}([0,1]),
\label{functional1}
\end{equation}
where $C^{(0)}([0,1])$ is the space of continuous
functions in $[0,1]$.
\vspace{0.7cm}
\noindent
{\em Example 2 (homogeneous polynomial functional of order $n$):} Consider
\begin{equation}
P_n([\theta])=\underbrace{\int_0^1\cdots \int_0^1}_{\textrm{$n$ times}}
K_n(x_1,...,x_n)\theta(x_1)\cdots \theta(x_n)dx_1\cdots dx_n,
\label{functional2}
\end{equation}
where $K_n$ are given kernel functions.
{\color{r}
\vspace{0.7cm}
\noindent
{\em Example 3 (Ginzburg-Landau energy functional):}
The Ginzburg-Landau theory describes
describes phase transitions and critical phenomena in a great
variety of statistical systems ranging from magnetic systems, to diluted polymers and superconductors \cite{KleinertPhi4,Amit}. At the basis of the theory is the energy functional
\begin{equation}
E([\phi]) = \int_{\mathbb{R}^3}
\left[\frac{1}{2}\left(\left|\nabla \phi(\bm x)\right|^2+m^2\phi^2(\bm x)\right)+\frac{\lambda}{4!}\phi^4(\bm x)\right]d\bm x
\end{equation}
where $m$ is the ``mass'' of the field $\phi$ and $\lambda$
is a coupling constant. In this case, the domain of the
functional $E$ can be chosen as $D(E)=C^{(2)}(\mathbb{R}^3)$.
\vspace{0.7cm}
\noindent
{\em Example 4 (Hopf characteristic functional of a
Gaussian random field):}
Consider a scalar Gaussian random field $u(\bm x;\omega)$ defined on a domain $V\subseteq \mathbb{R}^3$ it can be shown (see, e.g., \cite{Klyatskin1}) that the Hopf characteristic functional of such
random field is
\begin{equation}
\Phi([\theta])=\exp\left[i\int_V\mu(\bm x)\theta(\bm x)d\bm x -
\frac{1}{2}\int_V \int_V C(\bm x,\bm y)\theta(\bm x)\theta(\bm y) d\bm xd\bm y\right],
\label{white_0}
\end{equation}
where $\mu(\bm x)$ and $C(\bm x,\bm y)$ denote, respectively,
the mean and the covariance function of the field $u(\bm x;\omega)$.
\vspace{0.7cm}
\noindent
Analysis of nonlinear functionals in Banach spaces
is a well-developed subject \cite{Vainberg,Nashed,Schwartz1964}.
In particular, the classical definition of continuity and
differentiability at a point that holds for real-valued functions
can extended in a more or less straightforward way to nonlinear
functionals. For instance, we say that a functional
$F([\theta])$ is {\em continuous} at a point $\theta(x)\in D(F)$
if for any sequence of functions
$\{\theta_1(x), \theta_2(x), ... \}$ in $D(F)$) converging
to $\theta$ we have that the sequence $\{F([\theta_n])\}$
converges to $\{F([\theta])\}$, i.e.,
\begin{equation}
\lim_{n\rightarrow \infty}
\left|\theta_n(x)-\theta(x)\right|\rightarrow 0 \quad \Rightarrow \quad \lim_{n\rightarrow \infty}
\left|F([\theta_n])-F([\theta])\right|\rightarrow 0.
\label{continuity}
\end{equation}
The functionals we discussed in Examples 1-4 are all continuous.
From the continuity definition \eqref{continuity} it follows,
in particular, that if $F([\theta])$ is continuous on a compact
function space $D(F)$ then $F$ is {\em bounded}.
\vspace{0.7cm}
\noindent
{\em Example 4:}
Another example of a continuous functional
in $D(F)=C^{(\infty)}([-1,1])$ is
\begin{equation}
F([\theta]) = \theta(0)^2+\int_{-1}^1 \sin(\theta(x))dx.
\label{ex4fun}
\end{equation}
In fact, consider any sequence of functions
$\{\theta_n\}$ in $D(F)$. Also choose
an integrable function $g(x)$ such that
$\left|\theta_n(x)\right|\leq g(x)$ for all $n\in \mathbb{N}$.
In these conditions, the Lebesgue dominated convergence
theorem applies and we have
\begin{equation}
\lim_{n\rightarrow\infty} \int_{-1}^1 \sin(\theta_n(x))dx=\int_{-1}^1
\sin(\theta(x))dx,
\end{equation}
i.e.,
\begin{equation}
\theta_n\substack\rightarrow \theta\,
\Rightarrow \, F([\theta_n])\rightarrow F([\theta]).
\end{equation}
}
\subsection{Functional Derivatives}
\label{app:functional derivatives}
Consider a real or a complex valued functional $F$ defined on
the function space $D(F)$ (domain of the functional). For
simplicity, let us assume that $D(F)$ is a space of real valued
functions $\theta(x)$ on the real line.
We say that the functional $F$ is differentiable
at $\theta(x)$ if the limit
\begin{equation}
\lim_{\epsilon\rightarrow 0}\frac{F([\theta(x)+\epsilon \eta(x)])
-F([\theta(x)])}{\epsilon}
\label{gateaux}
\end{equation}
exists and it is finite. The quantity \eqref{gateaux}
is known as {\em G\^ateaux differential} of $F$ in the
direction of $\eta(x)$ (see, e.g., \cite{Vainberg,Schwartz1964}).
Under rather general assumptions such derivative can
be represented as a linear operator \cite{Nashed,Vainberg,Daniele_JMathPhys}
acting on $\eta(x)$. For small $\epsilon$ we have
\begin{equation}
F([\theta(x)+\epsilon \eta(x)])=F([\theta(x)])+\epsilon
L([\theta(x)])\eta(x)+R_1([\epsilon\eta(x);\theta(x)]),
\end{equation}
In this series $L([\theta(x)])$ is a linear operator that sends
the function $\eta(x)$ to a real
or a complex number, while $R_1$ represents a reminder term.
It is clear that $L([\theta(x)])$ involves
integration with respect to $x$, since $L([\theta])\eta(x)$ is a real or complex number. Thus, we look for a representation of $L([\theta])\eta(x)$ in the form
\begin{equation}
L([\theta(x)])\eta(x)=\int \frac{\delta F([\theta])}{\delta \theta(x)}\eta(x) dx,
\label{first_order_L}
\end{equation}
where the kernel function $\delta F([\theta])/\delta \theta(x)$
is a functional of $\theta$ and a function of $x$.
At this point it is convenient to establish
a parallel between functionals and functions in $m$ variables.
Recall that the differential of a scalar field $f(a_1,..,a_m)$
in the direction $\widehat{n}=(n_1,...,n_m)$ is the scalar product of the
gradient $\nabla f$ and $\widehat{n}$, i.e.,
\begin{equation}
df_{\widehat{n}}=\nabla f\cdot \widehat{n}.
\end{equation}
By analogy, $\delta F([\theta])/\delta \theta(x)$ in equation \eqref{first_order_L} can be considered as an infinite-dimensional gradient, known as {\em first-order functional derivative of $F$ with respect
to $\theta(x)$}.
The next question is: how do we compute such functional derivative?
A possible way is to use the definition \eqref{gateaux} and
a compactly supported class of test functions, e.g., functions that
are nonzero only in a small neighbor $\mathcal{I}_x(r)$
of radius $r$ centered at $x$.
Such functions could be compactly supported
elements of a Dirac delta sequence (see Figure \ref{fig:3}),
or even a delta function itself. This allows us to write
\begin{equation}
\frac{\delta F([\theta(x)])}{\delta \theta(x)}=\lim_{\epsilon,r\rightarrow 0}
\frac{F([\theta(x)+\epsilon \alpha(x)])-F([\theta(x)])}
{\displaystyle\epsilon\int_{\mathcal{I}_x(r)}\alpha(y)dy}
\label{A}
\end{equation}
In particular, if we set $\alpha(y)=\delta(x-y)$ then the denominator
in \eqref{A} simply reduces to $\epsilon$, yielding the formula
\begin{align}
\frac{\delta F([\theta])}{\delta \theta(x)}=&\lim_{\epsilon\rightarrow 0}
\frac{F([\theta(y)+\epsilon \delta(x-y)])-F([\theta(y)])}
{\epsilon}.
\label{A2}
\end{align}
\begin{figure}[t]
\centerline{\includegraphics[height=4cm]{fig3-eps-converted-to.pdf}}
\caption{A possible function for the calculation of the kernel
\eqref{A} at $x_k$. If we set
$ \mathcal{I}_{x_k}(r)=\{x\in\mathbb{R}: |x-x_k|\leq r\}$, then
we have $\displaystyle\int_{\mathcal{I}_{x_k}(r)}\alpha(x)dx=1$,
independently on $r$.}
\label{fig:3}
\end{figure}
Functional derivatives of higher order can be defined
in a similar manner. For example, the second order functional
derivative of $F$ is
\begin{equation}
\frac{\delta^2 F([\theta])}{\delta \theta(x)\delta \theta(y)}=
\frac{d}{d\epsilon}\left[\frac{\delta F([\theta(z)+\epsilon \delta(z-y)])}
{\delta\theta(x)}\right]_{\epsilon=0}.
\label{2ND}
\end{equation}
Note that \eqref{2ND} is a function and $x$ and $y$ and a
functional of $\theta(x)$.
\vspace{0.7cm}
\noindent
{\em Example 1:} The first-order functional
derivative of the nonlinear functional \eqref{functional1} can
be obtained as follows. We first compute the G\^ateaux differential
\begin{align}
\frac{d}{d\epsilon}\left. F([\theta+\epsilon \eta])\right|_{\epsilon=0}=&
-\int_{0}^1 x^{3}e^{-\theta(x)}\eta(x)dx\nonumber\\
=&\int_{0}^1 \frac{\delta F([\theta])}{\delta \theta(x)}\eta(x)dx.
\end{align}
Therefore,
\begin{equation}
\frac{\delta F([\theta])}{\delta \theta(x)} = -x^{3}e^{-\theta(x)}.
\end{equation}
Note that the functional derivative is a function of $x$ and a (local)
functional of $\theta(x)$.
\vspace{0.7cm}
\noindent
{\em Example 2 (Functional Derivatives of the Hopf Functional):}
Consider the Hopf characteristic functional of a random function
$u(x;\omega)$ defined in $[0,1]$
\begin{equation}
\Phi([\theta]) = \left<\exp\left[i\int_{0}^1 u(x;\omega)\theta(x)dx\right]\right>.
\label{Hopf76}
\end{equation}
The average operator $\left<\cdot \right>$ here
denotes a functional integral over the probability
functional of $u(x,\omega)$. The G\^ateaux differential
of $\Phi([\theta])$ along $\eta$ is
\begin{align}
\left.\frac{d} {d\epsilon}\Phi([\theta(x)+\epsilon\eta(x)])\right|_{\epsilon=0}
=&i\left<\exp\left[i\int_0^1 u(x;\omega)\theta(x)dx\right]
\int_0^1u(x;\omega)\eta(x)dx\right>\nonumber \\
=&\int_0^1 i\left<\exp\left[i\int_0^1 u(x;\omega)\theta(x)dx\right]
u(x;\omega)\right> \eta(x)dx\nonumber\\
=& \int_0^1 \frac{\delta \Phi([\theta])}{\delta \theta(x)} \eta(x)dx.
\label{first-order_FD}
\end{align}
This implies that
\begin{equation}
\frac{\delta \Phi([\theta])}{\delta \theta(x)}=i\left<u(x;\omega)
\exp\left[i\int_a^b u(x;\omega)\theta(x)dx\right]\right>
\label{kerneldef}
\end{equation}
is the first-order functional derivative of $\Phi([\theta])$ at $\theta(x)$.
Note that \eqref{kerneldef} is itself a functional of $\theta(x)$, which depends also on $x$. As a result,
$\delta\Phi([\theta])/\delta \theta(x)$ has two types of derivatives: an ordinary one
with respect to $x$, and a functional one with respect to $\theta(x)$.
The latter is the the second-order functional derivative
of $\Phi([\theta(x)])$. A simple calculation
shows that
\begin{align}
\frac{\delta^2 \Phi([\theta])}
{\delta \theta(x)\delta \theta(y)}=
i^2\left<u(y;\omega)u(x;\omega)\exp\left[i\int_a^b u(x;\omega)\theta(x)dx\right]\right>.
\label{B}
\end{align}
Proceeding similarly, we can obtain the expression
of higher-order functional derivatives. For instance, the
third-order one is explicitly given as
\begin{align}
\frac{\delta^3 \Phi([\theta])}
{\delta \theta(x)\delta \theta(y)\delta\theta(z)}=
i^3\left<\exp\left[i\int_a^b u(x;\omega)\theta(x)dx\right]
u(y;\omega)u(x;\omega)u(z;\omega)\right>.
\end{align}
Now, suppose we have available $\delta \Phi([\theta])/\delta\theta(x)$.
Based on the definition \eqref{kerneldef}, we see that
\begin{equation}
\left<u(x;\omega)\right>=\frac{1}{i}\frac{\delta \Phi([0])}{\delta\theta(x)}.
\end{equation}
Similarly, higher order moments and cumulants of the
random function $u(x;\omega)$ can be obtained
by computing higher order functional derivatives of $\Phi([\theta(x)])$ and
$\ln \Phi([\theta(x)])$, respectively, and evaluating them at $\theta(x)=0$.
In particular, the second- and third-order correlation functions are,
respectively
\begin{equation}
\left<u(x;\omega)u(y;\omega)\right>=\frac{1}{i^2}
\frac{\delta^2 \Phi([0])}{\delta\theta(x)\delta\theta(y)},
\quad
\left<u(x;\omega)u(y;\omega)u(z;\omega)\right>=\frac{1}{i^3}
\frac{\delta^3 \Phi([0])}{\delta\theta(x)\delta\theta(y)\delta\theta(z)}.
\end{equation}
{\color{r}
\vspace{0.7cm}
\noindent
{\em Example 3:} The Gateaux differential of the
nonlinear functional \eqref{ex4fun} in the direction $\eta(x)$ is
\begin{align}
\left.\frac{d}{d\epsilon}
F([\theta+\epsilon \eta])\right|_{\epsilon=0} = &
\frac{d}{d\epsilon}\left(
\theta(0)+\epsilon \eta(0) +\int_{-1}^1
\sin(\theta(x)+\epsilon \eta(x))dx
\right)_{\epsilon=0}\nonumber\\
&= \int_{-1}^1 \left(\delta(x)+\cos(\theta(x))\right)\eta(x)dx
\end{align}
Therefore the first-order functional derivative is
\begin{equation}
\frac{\delta F([\theta])}{\delta \theta(x)}=\delta(x)+\cos(\theta(x)),
\end{equation}
where $\delta(x)$ at the right hand side is the Dirac delta
function \cite{Kanwal}.
}
\paragraph{Regularity of Functional Derivatives}
The last example clearly shows that functional derivatives
of nonlinear functionals can easily be {\em distributions}
\cite{Kanwal}, e.g., Dirac delta functions.
For example, let $h\in C^{(\infty)}(\mathbb{R})$.
Then any functional in the form $F([\theta])=h((\theta,\theta))$,
where $(,)$ is an inner product in $C^{(\infty)}(\mathbb{R})$,
has a singular second-order functional derivative. In fact,
\begin{equation}
\frac{\delta F([\theta])}{\delta \theta(x)}=2\left.
\frac{\partial h}{\partial a}\right|_{a={(\theta,\theta)}}\theta(x)
\end{equation}
\begin{equation}
\frac{\delta^2 F([\theta])}{\delta \theta(x)\delta \theta(y)}=
2\delta(x-y)\left.\frac{\partial h}{\partial a}\right|_{a={(\theta,\theta)}}+
4\left.\frac{\partial^2 h}{\partial a^2}\right|_{a={(\theta,\theta)}}\theta(x)\theta(y).
\end{equation}
The characteristic functional of zero-mean Gaussian white noise
\begin{equation}
\Phi([\theta])=e^{-(\theta,\theta)/2}
\end{equation}
belongs to this class, i.e., it has a ``singular'' second-order functional
derivative.
On the other hand, the functional
\begin{equation}
G([\theta])=h\left( (K,\theta)\right),
\end{equation}
where $K(x)$ is a given smooth kernel, has smooth functional
derivatives
\begin{equation}
\frac{\delta G([\theta])}{\delta \theta(x)}=\left.
\frac{\partial h}{\partial a}\right|_{a={(K,\theta)}}K(x),\qquad
\frac{\delta^2 G([\theta])}{\delta \theta(x)\delta \theta(y)}=
\left.\frac{\partial^2 h}{\partial a^2}\right|_{a={(K,\theta)}}K(x)K(y).
\end{equation}
\section{Approximation of Nonlinear Functionals}
\label{sec:rep-Hopf}
Approximation theory for nonlinear functionals is
strongly related to approximation theory of nonlinear
operators \cite{Howlett,Torokhti,Bensoussan_DaPrato}.
A nonlinear functional $F$ is in fact
a particular type of nonlinear operator from a space
of functions $D(F)$ (the domain of the functional $F$)
into a vector space, e.g., $\mathbb{R}$ or $\mathbb{C}$.
Thus, the problem of approsimating nonlinear functionals
is basically the same as approximating nonlinear operators.
This topic has been studied extensively by different
scientific communities (see, e.g., \cite{Torokhti,Rugh,Nelles,Shetzen,Galman,
Prenter,Bertuzzi,Makarov,Khlobystov1}) for obvious reasons.
What does it mean to approximate a nonlinear functional?
Consider, as an example, the functional \eqref{functional1}, hereafter rewritten for convenience
\begin{equation}
F([\theta])=\int_0^1 x^3 e^{-\theta(x)}dx, \qquad D(F)=C^{(0)}([0,1]).
\label{functional11}
\end{equation}
Approximating $F([\theta])$ in this case means that
we are aiming at constructing a nonlinear
operator $\widehat{F}([\theta])$ that allows us
to compute an approximation of {\em all} possible integrals in the
form \eqref{functional11}, for arbitrary continuous functions
$\theta\in C^{(0)}([0,1])$. This challenging problem
includes cases in which $F([\theta])$ admits
an analytical solution, e.g., $F([x])$ or $F([\sin(x)])$, as well
as cases where no analytical solution is available, e.g., $F([x^2])$.
Perhaps, the most classical and widely used approach to
represent nonlinear functionals relies on functional power
series\footnote{Functional power series have been
widely used in the turbulence theory to obtain
moment and cumulant expansions (see \cite{Frisch,Monin2,Rosen_1967}).}. The method was originally developed by Volterra \cite{Volterra}, and it represents
the counterpart of power series expansions in the theory of
functions. In practice, the functional of interest is represented in terms
of a series of integral operators involving increasing powers of the test
function and kernels that need to be determined. The canonical form
of the power series expansion is
\begin{equation}
F([\theta])=\sum_{k=0}^\infty P_k([\theta]),\quad \textrm{where}\quad
P_k([\theta])=\int_{-\infty}^\infty\cdots\int_{-\infty}^\infty
K_k(x_1,...,x_k)\theta(x_1)\cdots\theta(x_k)dx_1\cdots dx_k.
\label{power_series}
\end{equation}
\noindent
{\em Remark:}
Functional power series are known to have bad approximation
properties and other issues. For example, they often do not preserve
important properties of the functional, e.g., positive definiteness
or normalization in the case of Hopf functionals.
{\color{r}
\subsection{Functional Approximation in Finite-Dimensional Function Spaces}
\label{sec:Finite_Dim_Approx}
The simplest ways to establish a closed functional
representation is to restrict the domain of the functional
to a finite-dimensional function space spanned by the
basis $\{\varphi_1(x),...,\varphi_m(x)\}$, i.e.,
\begin{equation}
D_m = \textrm{span}\{\varphi_1(x),...,\varphi_m(x)\}.
\label{Dspan}
\end{equation}
In this way, any element in $D_m$ can be
represented as\footnote{
\color{r}
If $D(F)$ is a space of multivariate
functions defined on some subset of $V\subseteq \mathbb{R}^d$
then \eqref{testf} takes the form
\begin{equation}
\theta(\bm x)=\sum_{k=1}^m a_k \varphi_k(\bm x), \qquad \bm x\in V.
\end{equation}
More generally, $\theta(\bm x)$ can be represented by
series expansions based on tensor products,
or more advanced expansions that rely on HDMR \cite{Rabitz,Li1}
or tensor methods \cite{Hackbusch_book,Kolda} (see also Section
\ref{sec:HDMR} and Section \eqref{sec:tensor}).
The latter techniques are recommended when operating
on test function spaces defined on high-dimensional
domains $V$.
}
\begin{equation}
\theta_m(x)=\sum_{k=1}^m a_k \varphi_k(x).
\label{testf}
\end{equation}
Possible choices of $\varphi_k(x)$ are:
\begin{enumerate}
\item {\em Lagrange Characteristic Polynomials}.
Given a set of $m$ distinct interpolation nodes $\{x_j\}$ in the interval
$[a,b]$, for example Gauss-Chebyshev-Lobatto nodes, we set
\begin{equation}
\varphi_j(x)=
\prod_{\substack{i=1\\i\neq j}}^{m}\frac{(x-x_i)}{(x_i-x_j)}.
\end{equation}
\item {\em Jacobi Polynomials.}
The function space $D_m\subseteq D(F)$ can be also represented by
a finite set of Jacobi polynomials $J^{(\alpha,\beta)}_j(x)$ (see
\cite{Gautshi,Hesthaven}), i.e.
\begin{equation}
\varphi_j(x)=J^{(\alpha,\beta)}_j(x).
\end{equation}
As is well known, Jacobi polynomials include
many other families of widely used polynomials such as Gegenbauer,
Legendre and Chebyshev.
\item {\em Trigonometric Polynomials}. If $D(F)$ is the space of
periodic functions in $[0,2\pi]$, then a convenient choice
for $\varphi_k(x)$ may be the set of (nodal) trigonometric
polynomials \cite{Hesthaven}
\begin{equation}
\varphi_j(x)=\frac{1}{m}\sin\left(m\frac{x-x_j}{2}\right)
\cot\left(\frac{x-x_j}{2}\right) \qquad x_j=\frac{2\pi}{m}j\qquad j=0,...,m
\label{nodalpoly}
\end{equation}
or, equivalently, classical Fourier modes
\begin{equation}
\varphi_0(x) = 1,\qquad \varphi_k(x)=\sin(kx)\qquad \varphi_{m+k}(x)=\cos(kx)
\qquad k=1,...,m.
\end{equation}
\end{enumerate}
\noindent
For each specific choice of the basis
set $\{\varphi_1(x),...,\varphi_m(x)\}$, the test function
\eqref{testf} lies on a {\em parametric manifold} of
dimension $m$, i.e., a hyperplane.
Any discretization of the function space $D(F)$ in terms
of a finite-dimensional basis, reduces the
functional $F$ into a multivariate function with
domain $D_m$ and range $F([D_m])$.
Such function depends on as many
variables as the number of degrees of freedom
we consider in the finite-dimensional
approximation of $D(F)$.
\vspace{0.7cm}
\noindent
{\em Example 1:} A substitution of \eqref{testf}
into the Hopf functional \eqref{Hopf76} yields the
complex-valued multivariate function
\begin{equation}
\phi(a_1,...,a_m) = \left<
\exp\left(i\sum_{k=1}^m a_j U_j(\omega)\right)
\right>, \qquad
U_j(\omega) = \int_{0}^1 u(x;\omega)\varphi_j(x)dx,
\label{missdef}
\end{equation}
i.e., the joint characteristic function of the Fourier coefficients
$U_j(\omega)$.
Note that $\phi$ depends on $m$ real variables $(a_1,...,a_m)$.
Such multivariate function can be seen as a $m$-dimensional
parametrization of the mapping $\Phi$ shown in Figure \ref{fig:1}, i.e.,
a parametrization of the nonlinear transformation
$D_m \rightarrow \Phi(D_m)$.
In this setting, approximation of nonlinear functionals is
equivalent to approximation of a real- or complex-valued
multivariate functions. The question of whether a functional
can be approximated by evaluating it in a space spanned by a
finite-dimensional basis is different, and will be addressed
in Section \ref{sec:approximability_of_functionals}.
\vspace{0.7cm}
\noindent
{\em Example 2:} Consider the nonlinear functional \eqref{functional1}
and let $D_m\subseteq D(F)$ be the function space
spanned by a suitable set of orthogonal
polynomials\footnote{Given any positive measure in a one-dimensional
interval, it is always possible to construct a set of polynomials that
is orthogonal with respect to such measure \cite{Gautshi1,Gautshi}.}
in $[0,1]$. Evaluating $F([\theta])$ in $D_m$, i.e., considering test
functions $\theta_m$ in the form \eqref{testf}, yields
the multivariate function
\begin{equation}
f(a_1,....,a_m) = \int_0^1 x^3 \prod_{k=1}^m e^{-a_k\varphi_k(x)}dx.
\end{equation}
This is the exact form of the functional $F$, evaluate in $D_m$, i.e.,
\begin{equation}
F([\theta_m])=f(a_1,....,a_m).
\end{equation}
\subsubsection{Functional Derivatives}
Evaluating a nonlinear functional $F([\theta])$
in a finite-dimensional function space
$D_m$ allows for a simple and effective representation of
functional derivatives. In particular, it can be shown that
\begin{equation}
\left.\frac{\delta F([\theta])}{\delta \theta(x)}
\right|_{\theta\in D_m}=\sum_{k=1}^m \varphi_k(x)\frac{\partial f}{\partial a_k},
\label{ha1}
\end{equation}
\begin{equation}
\left.\frac{\delta^2 F([\theta])}{\delta \theta(x)\delta\theta(y)}
\right|_{\theta\in D_m}=\sum_{j,k=1}^m \varphi_k(x)\varphi_j(y)
\frac{\partial f}{\partial a_j\partial a_k}.
\label{ha2}
\end{equation}
Here, $f(a_1,...,a_m)=F([\theta_m])$ is the function we obtain
by evaluating the functional $F$ in the finite dimensional function
space $D_m$. The meaning of \eqref{ha1} and \eqref{ha2} is the
following: if we evaluate the functional derivatives of
$F$ in the finite dimensional space $D_m$ (recall that the functional derivatives are themselves nonlinear functionals)
then we can represent them in terms of classical partial derivatives
of $f(a_1,...,a_m)$. Note that the basis function
$\varphi_j$ spanning $D_m$ also appear in \eqref{ha1}-\eqref{ha2},
suggesting that the accuracy of the functional derivatives depend on
the choice of such basis functions.
Rather than proving \eqref{ha1} and \eqref{ha2} in a general setting,
let we provide two constructive examples that yield
expressions in the form \eqref{ha1} and \eqref{ha2}.
\vspace{0.7cm}
\noindent
{\em Example 1:} Consider the Hopf functional \eqref{Hopf76}.
By evaluating the analytical expression of the first- and second-order
functional derivatives \eqref{kerneldef}-\eqref{B} in
the finite-dimensional function space \eqref{Dspan} we obtain
\begin{align}
\left.\frac{\delta \Phi(\theta)}{\delta \theta(x)}
\right|_{\theta\in D_m}=& i\left<u(x)e^{i(a_1U_1(\omega)+\cdots
+ a_m U_m(\omega))}\right>,\qquad \\
\left.\frac{\delta^2 \Phi(\theta)}
{\delta \theta(x)\delta \theta(y)}
\right|_{\theta\in D_m}=&-\left<u(x)u(y)
e^{i(a_1U_1(\omega)+\cdots+ a_m U_m(\omega))}\right>,
\end{align}
where $U_j(\omega)$ are random variables
defined in \eqref{missdef}.
By using the definition of the characteristic
function \eqref{missdef} we have that
\begin{align}
\frac{\partial \phi}{\partial a_k}=&\int_{0}^1\varphi_k(x)
\left.\frac{\delta \Phi(\theta)}{\delta \theta(x)}
\right|_{\theta\in D_m}dx,\\
\frac{\partial^2 \phi}{\partial a_k\partial a_j}=&
\int_0^1\int_{0}^1\varphi_k(x)\varphi_j(y)\left.\frac{\delta^2 \Phi(\theta)}
{\delta \theta(x)\delta \theta(y)}
\right|_{\theta\in D_m}dxdy.
\end{align}
This means that the partial derivatives of the characteristic function
are nothing but the projection of the Hopf functional derivatives
onto the space $D_m$. Clearly, if the random
function $u(x;\omega)$ is $D_m$ then the following
inverse formulas hold
\begin{equation}
\left.\frac{\delta \Phi([\theta])}{\delta \theta(x)}
\right|_{\theta,u\in D_m}=\sum_{k=1}^m \varphi_k(x)\frac{\partial \phi}{\partial a_k},
\label{ha111}
\end{equation}
\begin{equation}
\left.\frac{\delta^2 \Phi([\theta])}{\delta \theta(x)\delta\theta(y)}
\right|_{\theta,u\in D_m}=\sum_{j,k=1}^m \varphi_k(x)\varphi_j(y)
\frac{\partial \phi}{\partial a_j\partial a_k}.
\label{ha222}
\end{equation}
\vspace{0.7cm}
\noindent
{\em Example 2:} Consider the sine functional
\begin{equation}
F([\theta])=\sin\left(\int_a^b K(x)\theta(x)dx\right)
\label{sinFunctional}
\end{equation}
where $K(x)$ is a given kernel function.
Evaluating $F$ in $D_m$ yields the multivariate function
\begin{equation}
f(a_1,...,a_m)=\sin\left(\sum_{i=1}^m a_i \int_a^b K(x)\varphi_i(x)dx\right).
\label{fdr}
\end{equation}
Similarly, evaluating the functional derivative of $F$ in $D_m$ yields
\begin{equation}
\left.\frac{\delta F([\theta])}{\delta \theta(x)}\right|_{\theta\in D_m}=
\cos\left(\sum_{i=1}^m a_i \int_a^b K(x)\varphi_i(x)dx\right)K(x).
\label{fde}
\end{equation}
A comparison between \eqref{fdr} and \eqref{fde} immediately yields
\begin{equation}
\frac{\partial f}{\partial a_k}=\int_a^b\left.\frac{\delta F([\theta])}{\delta \theta(x)}\right|_{\theta\in D_N}
\varphi_j(x)dx,
\label{gg1}
\end{equation}
i.e., the gradient of $f$ is the projection of the
functional derivative of $F([\theta])$ (evaluated in $D_m$)
onto $D_m$. On the other hand, if $K(x)$ is a function in $D_m$
then
\begin{equation}
\left.\frac{\delta F([\theta]}{\delta \theta(x)}\right|_{\theta\in D_m}= \sum_{k=1}^m
\frac{\partial f}{\partial a_k} \varphi_k(x).
\label{gg2}
\end{equation}
This clarifies the meaning of the functional derivative
in both finite- and infinite-dimensional ($m\rightarrow \infty$)
cases.
\subsubsection{Distances between Function Spaces and Approximability of Functionals}
\label{sec:approximability_of_functionals}
A key concept when approximating a nonlinear functional $F([\theta])$ by restricting its domain $D(F)$ to a finite-dimensional space functions
$D_m$ is the distance between $D_m$ and $D(F)$. Such distance can be quantified in different ways (see, e.g., \cite{Pinkus}). For example we can define the {\em deviation} of $D_m$ from $D(F)$ as
\begin{equation}
E(D_m,D(F)) = \adjustlimits \sup_{\theta\in D(F)}
\inf_{\theta_m\in D_m} \left\|\theta-\theta_m\right\|
\label{functionaldeviation}
\end{equation}
The number $E$ measure the extent to which the worst
element of $D(F)$ can be approximated from $D_m$. One
may also ask how well we can approximate $D(F)$ with
$m$-dimensional subspaces of $D(F)$ which are allowed to vary within $D(F)$. A measure of such approximation is
given by the Kolmogorov $m$-width
\begin{equation}
d_m(D_m,D(F))= \adjustlimits \inf_{D_m} \sup_{\theta\in D(F)}
\inf_{\theta_m\in D_m} \left\|\theta-\theta_m\right\|_{D(F)}
\label{knw}
\end{equation}
which quantifies the error of the {\em best approximation}
to the elements of $D(F)$ by elements in a vector
subspace $D_m$ of dimension at most $m$.
The Kolmogorov $m$-width can be rigorously defined, e.g., for
nonlinear functionals in Hilbert spaces (\cite{Pinkus}, Ch. 4).
In simpler terms we can define the notion of approximability of
a nonlinear functional as follows. Let $F([\theta])$ be a
continuous nonlinear functional with domain $D(F)$, and consider
a finite-dimensional subspace $D_m\subseteq D(F)$, for example
$D_m=\textrm{span}\{\varphi_1,...,\varphi_m\}$.
We say that $F([\theta])$ is approximable in $D_m$ if for all $
\theta\in D(F)$ and $\epsilon>0$, there exists $m$ (depending on
$\epsilon$) and an element $\theta_m\in D_m$ such that
\begin{equation}
\left\| F([\theta])-F([\theta_m])\right\| \leq \epsilon.
\label{Ferr}
\end{equation}
Clearly if $F$ is continuous and $\theta_m$ is close to $\theta$,
i.e., the deviation \eqref{functionaldeviation} between $D(F)$
and $D_m$ is small, then we expect $\epsilon$ to be small.
It is important to emphasize that the approximation error and the computational complexity of approximating a nonlinear functional
depends on the choice of $D_m$. In particular, a functional may be low-dimensional in one function space and high-dimensional in another. The following example clarifies this question.
\vspace{0.7cm}
\noindent
{\em Example 1:} Consider the sine functional
\begin{equation}
F([\theta])=\sin\left(\int_{0}^{2\pi}\theta(x)dx\right)
\label{Fsin}
\end{equation}
in the space $D(F)$ of periodic functions in $[0,2\pi]$.
If we represent $\theta$ in terms of orthonormal Fourier modes,
i.e., we consider
\begin{equation}
D_{2m+1}= \textrm{span}\left\{\frac{1}{\sqrt{2\pi}},\frac{\sin(x)}{\sqrt{\pi}},..., \frac{\sin(mx)}{\sqrt{\pi}},\frac{\cos(x)}{\sqrt{\pi}},...,\frac{\cos(mx)}{\sqrt{\pi}}\right\}
\label{D2m}
\end{equation}
and
\begin{equation}
\theta_{2m+1}(x)=\frac{a_0}{\sqrt{2\pi}} + \frac{1}{\sqrt{\pi}}
\sum_{j=1}^m a_j \sin(jx)+ \frac{1}{\sqrt{\pi}}
\sum_{j=1}^m b_j \cos(jx)
\end{equation}
In this setting, we obtain
\begin{equation}
F([\theta_{2m+1}])=\sin(\sqrt{2\pi} a_0).
\end{equation}
This means that \eqref{Fsin} is approximable in the function space
\eqref{D2m}. Moreover, the approximation is
{\em exact} and just one-dimensional.
On the other hand, if we set the space
$D_{2m+1}$ to be the span of a normalized nodal Fourier basis
$\{\varphi_0, ...,\varphi_{2m}\}$, e.g., the normalized odd
expansion discussed in \cite{Hesthaven}, then the functional
\eqref{Fsin} technically requires an infinite number of variables
variables. In fact, in this case we have
\begin{equation}
F([\theta_{2m+1}])=\sin\left(\eta \sum_{k=0}^{2m} a_k\right),
\quad \textrm{where}\quad \eta = \int_{0}^{2\pi}\varphi_k(x)dx =
\left(\frac{2\pi}{2m+1}\right)^{\frac{1}{2}}.
\end{equation}
\vspace{0.7cm}
\noindent
{\em Example 2:}
Consider the characteristic functional of zero-mean
Gaussian white noise (see equation \eqref{white_0}),
\begin{equation}
\Phi([\theta])=\exp\left[-\frac{1}{2}\int_0^{2\pi} \theta(x)^2dx\right],
\label{white}
\end{equation}
where $D(\Phi)$ is the space of periodic functions
in $[0,2\pi]$. Let $D_m$ be the space spanned by
any finite orthonormal set of periodic functions.
The deviation between $D(\Phi)$ and $D_m$
this case yields a functional approximation
error of order 1.
To show this in a simple way, evaluate the functional
\eqref{white} in both $D(\Phi)$ and $D_m$. This yields
\begin{equation}
\Phi([\theta])=\exp\left[-\frac{1}{2}\sum_{k=1}^\infty a_k^2\right],\qquad
\Phi([\theta_m])=\exp\left[-\frac{1}{2}\sum_{k=1}^m a_k^2\right].
\end{equation}
If we measure the error between $\Phi([\theta])$
and $\Phi([\theta_m])$ in the uniform operator
norm then we have
\begin{equation}
\left\|\Phi([\theta_m])-\Phi([\theta])\right\|_{\infty} = 1,
\end{equation}
independently on $m$. In other words, \eqref{white}
is not approximable in any finite-dimensional subset of $D(\Phi)$.
This result is consistent with white-noise
theory \cite{Stratonovich}. Recall, in fact, that a
delta-correlated Gaussian process has a flat Fourier
power spectrum. This implies that any finite
truncation of the Fourier series of such process
yields a systematic error that is not small.
\vspace{0.7cm}
\noindent
{\em Remark:} In some cases the effects of
the distance between $D(F)$ and $D_m$ can be
mitigated by the presence of smooth functions
appearing the in functional $F$. For example, consider
the sine functional
\begin{equation}
F([\theta])=\exp\left(\int_{-1}^{1} \sin(x) \theta(x)dx\right)
\end{equation}
and let $D(F)$ be the space of infinitely differentiable functions
in $[-1,1]$. If we expand $\theta$ in terms of Legendre
polynomials $\{\varphi_k\}$, i.e.,
\begin{equation}
\theta_m(x)=\sum_{k=0}^m a_k \varphi_k(x)
\end{equation}
then,
\begin{equation}
\int_{-1}^{1} \sin(x) \theta(x) dx = \sum_{k=0}^m a_k
\int_{-1}^1\sin(x) \varphi_k(x)dx.
\end{equation}
As is well known, the coefficients $\int_{-1}^1\sin(x) \varphi_k(x)dx$
decay to zero exponentially fast with $m$ \cite{Hesthaven}.
This implies that convergence of $F([\theta_m])$ to $F([\theta])$
is exponentially fast in the number of dimensions $m$, that is the
error \eqref{Ferr} goes to zero exponentially fast with $m$.
\vspace{0.7cm}
\noindent
{\em Example 3:}
Consider the Hopf characteristic functional of a zero-mean
correlated Gaussian process in $[0,2\pi]$
\begin{equation}
\Phi([\theta])=\exp\left[-\frac{1}{2}\int_0^{2\pi}\int_{0}^{2\pi} C(x,y)\theta(x)\theta(y)dxdy\right],
\label{hopfc}
\end{equation}
where $C(x,y)$ is a smooth covariance function, and
$D(\Phi)$ is the space of periodic functions in $[0,2\pi]$.
If the projection of $C(x,y)$ onto the span of
an orthonormal set $D_m\subseteq D(F)$
decays with $m$, then the functional $\Phi$ is
approximable in $D_m$. Note that this is indeed
the case if the covariance is smooth (and periodic)
and $\varphi_j(x)$ are the Fourier modes in \eqref{D2m}.
The smoother the covariance the smaller the number
of Fourier modes we need to achieve a certain
accuracy \cite{Hesthaven}, i.e.,
the smaller the number of dimensions.
If $C(x,y)=1$ then the Hopf functional \eqref{hopfc}
is effectively {\em one-dimensional}.
}
\subsection{Functional Interpolation Methods}
\label{sec:Functional Collocation Methods}
In this Section we discuss how to construct an approximation
of a nonlinear functional $F([\theta])$ in terms of
a {\em functional interpolant} $\Pi([\theta])$, i.e., a functional that interpolates $F([\theta])$ at at a given set
of nodes $\{\theta_1(x),...,\theta_m(x)\}\in D(F)$
\begin{equation}
F([\theta_j])=\Pi([\theta_j]) \qquad j=1,...,m.
\end{equation}
Differently from interpolation methods in spaces of
finite dimension (e.g., $d$-dimensional Euclidean spaces),
interpolation here is in a space of functions, i.e.,
the {interpolation nodes} $\theta_k(x)$ are
functions in a Hilbert or a Banach space.
Over the years, the problem of constructing
a functional interpolant through suitable nodes
in Hilbert or Banach spaces has been studied by several
authors and convergence results were established in rather
general cases
\cite{Makarov,Khlobystov0,Prenter,Porter,Kaplitskii,Allasia,
Khlobystov,Khlobystov1,PorterSIAM,Torokhti}.
Before discussing functional interpolation in detail,
let us provide some geometric intuition on what functional
interpolation is and what kind of representations
we should expect.
To this end, let us first recall recall that
a hyperplane in a $d$-dimensional space is a linear
manifold defined uniquely by $d$ interpolation nodes,
each node being a vector of $\mathbb{R}^{d}$.
If we send $d$ to infinity then the hyperplane intuitively
becomes a linear functional, which is therefore defined uniquely
by an infinite number of $\infty$-dimensional nodes,
i.e., an {\em infinite number of functions}.
This suggests that if we consider any finite number of
nodes in a function
space, say $\{\theta_1(x),...,\theta_m(x)\}$, then
we cannot even represent {\em linear functionals} in
an exact way\footnote{We recall that
the variational form of nonlinear PDEs is defined by
linear functionals on test function spaces.
In this setting, classical Galerkin methods
to solve PDEs (see Section \ref{sec:variational}) are basically
identification problems for linear functionals.},
i.e., functionals in the form
\begin{equation}
P_1([\theta])=\int_a^bK_1(x_1)\theta(x_1)dx_1,
\end{equation}
where $K_1(x)$ is a given kernel.
The same conclusion obviously holds
for nonlinear functionals, with the aggravating factor
that the number of test functions theoretically required
for the exact representation grows significantly.
For example, quadratic and cubic forms in $d$-dimensions are identified by
by $d^2$ and $d^3$ interpolation nodes, respectively, where each node is
vector of $\mathbb{R}^{d}$. When we send $d$ to infinity, we intuitively obtain
homogeneous polynomial functionals of second- and third-order, respectively.
These functionals are in the form
\begin{align}
P_2([\theta])&= \int_a^b\int_a^b K_2(x_1,x_2)\theta(x_1)\theta(x_2)dx_1dx_2,
\label{sof}\\
P_3([\theta])&= \int_a^b\int_a^b\int_a^b K_3(x_1,x_2,x_3)
\theta(x_1)\theta(x_2)\theta(x_3)dx_1dx_2dx_3.
\end{align}
Thus, to represent $P_2$ and $P_3$ exactly we need $\infty^2$
and $\infty^3$ nodes in a function space. Why $\infty^2$ and $\infty^3$?
Consider $P_2([\theta])$ and assume that
the kernel function $K_2(x_1,x_2)$ is in a separable
Hilbert space. Represent $K_2$ relative to any complete
orthonormal basis $\{\varphi_k\}_{k=1,...,\infty}$
\begin{equation}
K_2(x_1,x_2)=\sum_{i,j=1}^\infty a_{ij}\varphi_i(x_1)\varphi_j(x_2).
\label{K2e}
\end{equation}
A substitution of \eqref{K2e} into \eqref{sof} yields
\begin{equation}
P_2([\theta])= \sum_{i,j=1}^\infty a_{ij}\int_{a}^b\varphi_i(x)\theta(x)dx
\int_a^b \varphi_j(x)\theta(x)dx.
\label{ssof}
\end{equation}
Without loss of generality we can assume that $K_2$ is symmetric, i.e., that $a_{ij}$
is a symmetric matrix. To represent $P_2([\theta])$ exactly by means of a functional
interpolant we need enough nodes $\{\theta_p(x)\}$ to determine each
$a_{ij}$ in \eqref{K2e} uniquely. Clearly, the choice
$\theta_p(x)=\varphi_p(x)$ ($p=1$,..., $\infty$) is not sufficient for
this purpose, since it allows us to determine only the diagonal
entries $a_{pp}$. Therefore we need to construct a larger set of collocation
nodes, e.g., the set $\theta_p(x)=\varphi_{p_i}(x)+\varphi_{p_j}(x)$, where
$p_i \geq p_j$ and $p_i\in \{1,...,\infty\}$.
This is the $\infty^2$ number of functions we have
mentioned above. Similarly, to identify
$P_3([\theta])$ through functional interpolation we need
$\infty^3$ nodes, e.g., in the form $\theta_p(x)=\varphi_{p_i}(x)
+\varphi_{p_j}(x)+\varphi_{p_k}(x)$, where $p_i\geq p_j\geq p_k$
and $p_k\in \{1,...,\infty\}$.
As we shall see later in this Section, determining a polynomial interpolant
of an unknown functional $F([\theta])$, i.e., determining the kernels
$K_j(x_1,...,x_j)$ in \eqref{power_series} from input-output relations
is a {\em linear problem} that involves high-dimensional systems and
big data.
\subsubsection{Interpolation Nodes in Function Spaces}
\label{sec:interpolationnodes}
Let $F([\theta])$ be a continuous functional
with domain $D(F)$. Within $D(F)$ we define the spaces
of functions
\begin{equation}
S^{(m)}_q=\left\{\theta(x)\in D(F)\,\,|\,\, \theta(x)=a_{i_1}\varphi_{i_1}(x)+\cdots +
a_{i_q}\varphi_{i_q}(x)\right\}.
\label{SNq}
\end{equation}
where $i_j\in\{1,...,m\}$, $a_{i_j}\in \mathbb{R}$
and $\{\varphi_{1}(x),...,\varphi_{m}(x)\}\in D(F)$. The elements of $ S^{(m)}_1$,
$S^{(m)}_2$ and $S^{(m)}_3$ are in the form
\begin{equation}
\begin{array}{cl}
S^{(m)}_1:& \theta(x) = a_{i_1}\varphi_{i_1}(x),\nonumber\\
S^{(m)}_2:& \theta(x) = a_{i_1}\varphi_{i_1}(x)+a_{i_2}\varphi_{i_2}(x),\nonumber\\
S^{(m)}_3:& \theta(x) = a_{i_1}\varphi_{i_1}(x)+a_{i_2}\varphi_{i_2}(x)+
a_{i_3}\varphi_{i_3}(x).\nonumber
\end{array}
\label{snq}
\end{equation}
Clearly, if $\theta_1\in S^{(m)}_j$ and $\theta_2\in S^{(m)}_q$, then
$(\theta_1+\theta_2)\in S^{(m)}_{j+q}$ (if $q+j< m$).
Also, note that the sequence of spaces $S_j^{(m)}$, $j=1, 2, ...$ is
{\em hierarchical} in the sense that the following
chain of embeddings hold
\begin{equation}
S^{(m)}_1\subset S^{(m)}_2\subset \cdots \subset S^{(m)}_m\subset D(F).
\label{embedding}
\end{equation}
\begin{figure}[!t]
\centerline{
\includegraphics[height=6cm]{fig8-eps-converted-to.pdf}
\includegraphics[height=5cm]{fig9-eps-converted-to.pdf}
}
\caption{Network of test functions:
A closed path in the function space $D(\Phi)$ made of four lines
(e.g., the red polygon) is mapped through the functional $\Phi$
into a closed curve in the complex plane. Vertex and edges of the
network are elements of $S_2^{(m)}$. We also show a complete
graph with $n=5$ nodes, where each edge is discretized
with $p=4$ points.}
\label{fig:8}
\end{figure}
The function space $S_q^{(n)}$ admits a simple yet powerful graphical
representation in terms of {\em trajectories of functions}
\cite{Magri,Tonti,Tonti1,Tonti2,Daniele_JMathPhys}. To illustrate
such representation, consider a complex-valued functional $\Phi$.
A trajectory of functions in the space $D(\Phi)$ is a curve
in $D(\Phi)$, which is mapped to a curve
in $\mathbb{C}$ (see Figure \ref{fig:8}).
Furthermore, if the functional is continuous and differentiable,
a smooth curve in $D(F)$ is mapped onto a smooth curve in $\mathbb{C}$.
The set of trajectories in the complex plane associated with $S^{(m)}_1$
is shown in Figure \ref{fig:branching}.
Each curve is parametrized by only one parameter $a_j$ and it
cannot branch into two distinct curves .
On the contrary, if we consider $S_2^{(m)}$ we are adding
one more degree of freedom and each curve departing from $1$ can branch,
but only once. Similarly, we can have three branches in $S_3^{(m)}$, etc.
A remarkable distribution of nodes in $S_2^{(m)}$ is
associated with {\em networks of test functions}, i.e., {\em graphs}
in the function space $D(F)$. Vertex and edges are elements
of $S_2^{(m)}$. A simple example is shown in Figure \ref{fig:8}.
In mathematics such network is called {\em complete graph}, i.e.,
an undirected graph in which every pair of distinct nodes is
connected by a unique edge -- the edge being the trajectory
(straight line) of functions connecting $\theta_i$ to $\theta_j$.
If we discretize each edge with $p$
collocation points (including the endpoints) and we have $m$ nodes then
the number of degrees of freedom is $m(m-1)(p-2)/2+m$.
\begin{figure}[!t]
\centerline{
\includegraphics[height=6cm]{fig10-eps-converted-to.pdf}
\includegraphics[height=5.5cm]{fig11-eps-converted-to.pdf}
}
\caption{Trajectories in the range of a complex valued
functional $\Phi$ corresponding
to functions in the sets $S^{(m)}_1$ (left) and $S^{(m)}_3$ (right).
In the case of $S^{(m)}_1$ each curve is parametrized
by only one parameter $a_j$ and it cannot branch into two distinct curves.
On the contrary, if we consider $S_3^{(m)}$ we are adding
two more degrees of freedom and each curve can branch at most twice.
}
\label{fig:branching}
\end{figure}
Another interesting set of interpolation nodes in $D(F)$
is the one obtained by setting all
coefficients $a_{i_p}$ in \eqref{SNq}
equal to $1$, i.e.,
\begin{equation}
\widehat{S}^{(m)}_q=\left\{\theta_i(x)\in D(F)\,\,|\,\, \theta_i(x)=\varphi_{i_1}(x)+\cdots +
\varphi_{i_q}(x)\right\},\qquad i_s=0,...,m,
\label{SNq1}
\end{equation}
If the set $\{\varphi_j\}$ includes the null element $\{0\}$,
then by symmetry \eqref{SNq1} is equivalent to
\begin{equation}
\widehat{S}^{(m)}_q=\left\{\{0\},\{\varphi_{1},..,\varphi_{m}\},
\{2\varphi_{1},..,2\varphi_{m}\},
\{(\varphi_{1}+\varphi_{2}),...,(\varphi_{1}+\varphi_{m})\},
\{(\varphi_{2}+\varphi_{3}),...,(\varphi_{2}+\varphi_{m})\},...\right\}.
\label{SNq2}
\end{equation}
In this case, the number of elements (cardinality) of $\widehat{S}^{(m)}_q$ is
\begin{equation}
\# \widehat{S}^{(m)}_q=\sum_{j=0}^q\binom{j+m-1}{j},\qquad \textrm{where}
\,\,\, \binom{i}{j}\,\,\,
\textrm{is the binomial coefficient.}
\label{SCardinality}
\end{equation}
For example,
\begin{equation}
\# \widehat{S}^{(10)}_1=11,\qquad \#\widehat{S}^{(10)}_2=66\qquad
\#\widehat{S}^{(10)}_3=286,\qquad \#\widehat{S}^{(10)}_5=3003,\qquad
\#\widehat{S}^{(10)}_{10}=184756.
\label{cardS}
\end{equation}
In addition, the cardinality of $\widehat{S}^{(m)}_q$ satisfies the recursion relation
\begin{equation}
\# \widehat{S}^{(m)}_q= \# \widehat{S}^{(m)}_{q-1}+\binom{q+m-1}{q}.
\end{equation}
The set of functions $\widehat{S}^{(m)}_q$ is sufficient
to uniquely identify a polynomial functional of order $q$ in which each
kernel function is represented relative to tensor product basis with $m$
elements in each variable.
However, $\widehat{S}^{(m)}_q$ is, in general, {\em not} sufficient
to accurately interpolate nonlinear functionals. The main problem is that the
set of nodes \eqref{SNq1} may not be large enough or may not
cover the function space $D(F)$ appropriately.
Another open question is related to the selection of
optimal interpolation nodes in $D(F)$ yielding highly accurate
representations. This question is addressed in Section
\ref{sec:optimal interpolation nodes}.
In a finite-dimensional setting, we can sample
the coefficients $a_{i_p}$ in \eqref{SNq}, e.g., at sparse
grids locations \cite{Barthelmann,Bungartz}, thai is at
unions of appropriate tensorizations of one-dimensional point
sets such as Gauss-Hermite, Clenshaw-Curtis,
Chebyshev or Leja \cite{Akil}. This yields the set
\begin{equation}
\widetilde{S}^{(m)}_q=\left\{\theta_i(x)\in D(F)\,\,|\,\, \theta_i(x)=
a_{i_1}\varphi_{i_1}(x)+\cdots +
a_{i_q}\varphi_{i_q}(x)\right\},\qquad i_s=1,...,m,
\label{SNq3}
\end{equation}
where the vector $(a_{i_1},...,a_{i_q})$ takes discrete values
at sparse grid nodes.
As an example, in Figure \ref{fig:optimal_nodes_2D} we plot three
Clenshaw-Curtis grids and few samples of the corresponding
interpolation nodes in $\widetilde{S}^{(2)}_2$.
For illustration purposes, the basis
elements $\varphi_1(x)$ and $\varphi_2(x)$ here are chosen as
\begin{equation}
\varphi_1(x)=\sin(x)e^{\cos(x)},\qquad \varphi_2(x)=\sin(2x)e^{\cos(2x)}.
\end{equation}
\begin{figure}[t]
\vspace{-0.5cm}
\centerline{\hspace{0.4cm}level 5\hspace{4.8cm} level 7\hspace{4.8cm} level 10}
\centerline{\includegraphics[height=4.5cm]{CC_nodes1-eps-converted-to.pdf}
\includegraphics[height=4.5cm]{CC_nodes2-eps-converted-to.pdf}
\includegraphics[height=4.5cm]{CC_nodes3-eps-converted-to.pdf}
}
\centerline{\includegraphics[height=4cm]{theta_1-eps-converted-to.pdf}
\includegraphics[height=4cm]{theta_2-eps-converted-to.pdf}
\includegraphics[height=4cm]{theta_3-eps-converted-to.pdf}
}
\caption{Clenshaw-Curtis point sets and corresponding nodes in
$\theta_i(x)=a_{1_i}\varphi_1(x)+a_{2_i}\varphi_2(x)$
($i=1, 2, 3$) in the test function space $\widetilde{S}^2_2$. Specifically,
we plot $\theta_i(x)$ corresponding to the nodes $1$, $2$ and $3$ marked
in red in the top right Figure. The basis elements $\varphi_k(x)$ here are
chosen as $\sin(kx)\exp(\cos(kx))$ ($k=1,2$).}
\label{fig:optimal_nodes_2D}
\end{figure}
The construction of sparse grids usually follows
the Smolyak algorithm. Other dimension-adaptive schemes and greedy Leja
rules were recently proposed in \cite{Chkifa,Akil,Akil1}.
\subsubsection{Polynomial Interpolation of Nonlinear Functionals}
\label{sec:functional_polynomial_interpolation}
Let $F:D(F)\rightarrow Y$ ($Y=\mathbb{R}$ or $\mathbb{C}$)
be a nonlinear real- or complex-valued functional on $D(F)$.
Consider the set of polynomial functionals
of degree $n$
\begin{equation}
\Pi_n([\theta])=L_0+L_1([\theta])+L_2([\theta],[\theta])+\cdots +
L_n([\theta],...,[\theta]),\label{pfun}
\end{equation}
where $L_0\in Y$, and $L_k:D(F)^k\rightarrow Y$ are $k$-linear
{\em symmetric} functionals
\begin{equation}
L_k([\theta_1],...,[\theta_p])=\int\cdots\int K_k(x_1,..,x_k)
\theta_1(x_1)\cdots
\theta_k(x_k)dx_1\cdots dx_k,\qquad k=1,...,n.
\label{lP}
\end{equation}
A comparison between equations \eqref{pfun} and \eqref{power_series}
yields
\begin{equation}
P_k([\theta])= L_k([\theta],...,[\theta]),
\end{equation}
and therefore we can equivalently write \eqref{pfun} as
\begin{equation}
\Pi_n([\theta])=\sum_{k=1}^n P_k([\theta])
\label{pfun1}.
\end{equation}
The symmetry assumption on $L_k$ implies that
$K_k(x_1,...,x_k)$ are symmetric kernels, i.e.,
any permutation of $x_1$, ..., $x_n$ leaves $K_k$ unchanged.
It is obviously possible to define polynomial functionals
with non-symmetric kernels. However, such functionals can be {\em always}
written in a symmetric form by rearranging the kernel functions appropriately.
For example, let $H_2(x_1,x_2)$ be non-symmetric. It is easy to verify that
\begin{equation}
\int_a^b \int_a^b H_2(x_1,x_2)\theta(x_1)\theta(x_2)dx_1dx_2=
\int_a^b \int_a^b K_2(x_1,x_2)\theta(x_1)\theta(x_2)dx_1dx_2,\qquad\forall \theta\in D(F),
\end{equation}
where
\begin{equation}
K_2(x_1,x_2)=\frac{1}{2}\left(H_2(x_1,x_2)+H_2(x_2,x_1)\right).
\end{equation}
In other words, the value of the integral does not change if we
replace $H_2(x_1,x_2)$ with its symmetrized version $K_2(x_1,x_2)$.
More generally, we can symmetrize any kernel $H_p(x_1,...,x_p)$
by summing up all terms corresponding to all possible permutations of $(x_1,...,x_p)$
and then dividing up by the factorial of $p$. For example,
\begin{align}
K_3(x_1,x_2,x_3)=\frac{1}{3!}& \left(H_{3}(x_1,x_2,x_3)+H_{3}(x_1,x_3,x_2)+
H_{3}(x_2,x_1,x_3)+H_{3}(x_2,x_3,x_1) + \right.\nonumber\\
&\left.H_{3}(x_3,x_1,x_2)+ H_{3}(x_3,x_2,x_1)\right).
\end{align}
The symmetry of the operators $L_k$ significantly
reduces the number of collocation nodes in the function space $D(F)$ needed
to identify kernels $K_k$, provided these are of finite-rank.
The polynomial functional interpolation problem can be stated as follows:
Given a set of $m$ nodes $\{\theta_1(x),...,\theta_m(x)\}$ in $D(F)$,
find a polynomial functional in the form \eqref{pfun}
satisfying the interpolation conditions
\begin{equation}
\Pi_n([\theta_i])=F([\theta_i]),\qquad i=1,...,m.
\label{interpolation2}
\end{equation}
\paragraph{The Stone-Weierstrass Approximation Theorem}
The possibility of approximating an arbitrary continuous functional
in Hilbert or Banach spaces in terms polynomial functionals
is justified by theorems analogous to the classical
Weierstrass theorem for continuous functions.
We recall that such theorem states that
if $f (x)$ is a continuous, real-valued function on
the closed interval $[a ,b ]$, then given any $\epsilon > 0$
there exists a real polynomial $p(x)$ such
that $| f (x)- p (x) | < \epsilon $ for all $x \in [a ,b ]$.
A remarkable generalization of this result has its roots in the
{\em Stone-Weierstrass theorem} \cite{Stone},
which can be stated as follows: suppose that
$X$ is a compact metric space and $K$ is an algebra of continuous,
real-valued functions on $X$ that separates points\footnote{The
algebra $K$ separates points if for any two distinct
elements $u_1$ ,$u_2 \in X$ there exists $\Pi_n \in K$
such that $\Pi_n ([u_1]) -\Pi_n([x_2]) \neq 0$ of $X$ and that
contains the constant function.}.
Then for any continuous, real-valued functional $F$ on $X$ and
for any $\epsilon > 0$ there exists a polynomial functional $\Pi \in K$
such that $\left\| F ([u]) - \Pi([u]) \right\| < \epsilon$
for all $u \in X$.
The first paper dealing with this subject
is due Frech\'et \cite{Frechet}. He showed that any continuous
functional can be represented by a series of polynomial functionals
whose convergence is uniform in all compact sets
of {\em continuous functions}.
This result was generalized to compact sets of functions in
Hilbert and Banach spaces by Prenter \cite{Prenter1970} and
Istratescu \cite{Istratescu}, respectively. Other relevant work in this
area is \cite{Galman,PorterClark,Bertuzzi,Chaika,Poggio,Porter}.
\subsubsection{Porter Interpolants}
\label{sec:porter}
An effective way to construct
finite-order polynomial functionals with minimal norm
interpolating arbitrary continuous functionals in Hilbert spaces
was proposed by W. Porter in \cite{PorterSIAM}. The key idea relies on
minimizing the norm of \eqref{pfun} subject to the interpolation
conditions \eqref{interpolation2}. A natural way to impose such
conditions is through Lagrange multipliers. This yields the variational
principle
\begin{equation}
\min_{K_1,..,K_n}\left\|\Pi_n\right\|^2+\sum_{j=1}^m
\lambda_i \left(\Pi_n([\theta_i])-F([\theta_i])\right),
\label{minpr}
\end{equation}
The minimum is relative to arbitrary variations of the kernel
functions $K_j(x_1,..., x_j)$. Also, $\left\|\Pi_n\right\|^2$
is the norm of the polynomial functional \eqref{pfun}, which is
defined as
\begin{equation}
\left\|\Pi_n\right\|^2=\sum_{p=0}^n\int\cdots
\int \left|K_p(x_1,...,x_p)\right|^2 dx_1\cdots dx_p< \infty
\end{equation}
The solution to the variational principle \eqref{minpr}
allows us to identify the kernel functions $K_j(x_1,..., x_j)$ and,
correspondingly, the polynomial functional
with {\em minimal norm} interpolating $F([\theta])$ at the $m$
nodes $\{\theta_1,...,\theta_m\}$.
Specifically, we obtain
\begin{align}
K_p(x_1,...,x_p)=\sum_{j,k=1}^m \pi^{(p)}_j(x_1,..,x_p)H^{-1}_{jk} F([\theta_k]),
\qquad i=0,...,n.
\label{Kp}
\end{align}
In this equation,
\begin{align}
\pi^{(0)}_j=1,\qquad
\pi^{(p)}_j(x_1,..,x_p)=\theta_j(x_1)\cdots\theta_j(x_p)\qquad p=1,2, ...
\end{align}
while the (symmetric) matrix $H_{ij}$ is defined as
\begin{equation}
H_{ij}=1+\left(\theta_i,\theta_j\right)+\left(\theta_i,\theta_j\right)^2+
\cdots+\left(\theta_i,\theta_j\right)^n,
\label{tt1}
\end{equation}
where $(,)$ denotes the $L_2(V)$ inner product, $V$ being
the domain of the interpolation nodes $\theta_j(x)$.
The polynomial functional $\Pi_n$ constructed in this way exists
if $F([\theta_k])$ is in the range of the matrix $\bm H$
(see \cite{PorterSIAM} for further details).
If one wants to approximate $F([\theta])$ in terms of
a superimposition of monomials with orders defined by
an index set $\mathcal{I}$ then
\begin{equation}
H_{ij}=\sum_{p\in\mathcal{I}} \left(\theta_i,\theta_j\right)^p
\label{matH}
\end{equation}
The {\em total degree} of the polynomial functional is the largest
number in the index set $\mathcal{I}$.
It is convenient to write Porter's interpolant in terms of basis
functionals $g_i([\theta])$ as
\begin{equation}
\Pi_n \left([\theta]\right)=\sum_{k=1}^m F\left([\theta_k]\right) g_{k}([\theta]),
\label{Porter_interpolant}
\end{equation}
where
\begin{equation}
g_{k}([\theta])=\sum_{j=1}^m H_{jk}^{-1}\sum_{p\in\mathcal{I}}\left(\theta_j,\theta\right)^p.
\label{gi_0}
\end{equation}
The functional interpolant \eqref{Porter_interpolant}-\eqref{gi_0}
has the following properties:
\begin{enumerate}
\item $\{g_k([\theta])\}$ is a set of cardinal basis functionals, i.e., $g_{k}([\theta_q])=\delta_{kq}$. This implies that Porter's
interpolant is a {\em cardinal Lagrangian interpolant}.
{\color{r}
\item If the interpolation nodes $\{\theta_1,...,\theta_m\}$ are
orthonormal with respect to the inner product $(,)$ then
$H_{ii}=\# \mathcal{I}$ (cardinality of the index set $\mathcal{I}$)
and $H_{ij}$ ($i\neq j$) either equal to one or zero,
depending on whether we have $\{0\}$ in the
set $\mathcal{I}$ or not. In every case, $\bm H$ is a matrix with
diagonal entries equal to $\# \mathcal{I}$ and off-diagonal entries
equal to wither zero or one. Such matrix is {\em always invertible}
provided $\mathcal{I}$ does not reduce to the single element $\{0\}$.
}
\item Porter's interpolant is degenerate for $\mathcal{I}=\{0\}$ as the matrix
$\bm H$ is rank one and therefore it is not invertible. The Moore-Penrose
pseudoinverse $\bm H^+$, however, exists and it provides the
correct form of the interpolant. To show this in a simple case, consider
the constant functional $F([\theta])=c\in \mathbb{R}$ and the
zero-order polynomial interpolant at $\{\theta_1,...,\theta_m\}$
\begin{equation}
\Pi_0([\theta])=K_0=\sum_{k,j=1}^m H^{-1}_{jk} F([\theta_j]).
\qquad H_{ij}=1.
\end{equation}
Clearly, $H^{-1}_{jk}$ does not exist since $H_{ij}=1$ and therefore
$\textrm{rank}(H)=1$. However, the Moore-Penrose pseudoinverse
of $\bm H$ has components $H^+_{ij}=1/m^2$, and therefore $\Pi_0([\theta])=c=F([\theta])$.
\item The polynomial functional \eqref{Porter_interpolant}-\eqref{gi_0}
is an interpolant if and only if the matrix \eqref{matH} is invertible, i.e.,
full rank. This depends on both the choice of interpolation
nodes and on the index set $\mathcal{I}$.
The Moore-Penrose pseudoinverse $\bm H^+$, in general, {\em does
not allow} to satisfy the interpolation condition. Indeed, by evaluating
\eqref{Porter_interpolant} at the interpolation nodes $\theta_k$ we obtain
\begin{equation}
\Pi_n \left([\theta_i]\right)=\sum_{k=1}^m F\left([\theta_k]\right)
\sum_{j=1}^m H_{jk}^{-1}H_{ji} = F([\theta_i]).
\end{equation}
However, if we replace $H_{jk}^{-1}$ with $H_{jk}^{+}$
then we get $\Pi_n \left([\theta_i]\right)\neq F \left([\theta_i]\right)$,
since $\bm H^+\bm H\neq \bm I$.
\item Consider the approximation of constant functionals $F([\theta])=c$ by
polynomial functionals of order one, i.e., $\Pi_{1}([\theta])=L_0+L_1([\theta])$.
Assume that the interpolation nodes $\{\theta_k\}_{k=1,...,m}$ are orthogonal functions
with norm that decreases with $m$ as $1/m$ (see Section
\ref{sec:results linear functionals}). Let $H_{ij}=1+h\delta_{ij}$
where $h=\left\|\theta_j\right\|^2$ be the matrix \eqref{matH}
corresponding to the index set $\mathcal{I}=\{0,1\}$.
By using the identity
\begin{equation}
\lim_{h\rightarrow 0}\frac{1}{h}\left(1+h\delta_{ij}\right)^{-1}=\delta_{ij}-\frac{1}{m}
\end{equation}
we obtain
\begin{equation}
\Pi_{1}([\theta])\simeq c\sum_{k,j=1}^m
\left(h\delta_{jk}-\frac{h}{m}\right)
\left(1+(\theta,\theta_j)\right) \qquad \textrm{as}\qquad h\rightarrow 0\quad \textrm{and}
\quad m\rightarrow\infty.
\end{equation}
This means that $\displaystyle\lim_{m\rightarrow \infty} \Pi_{1}([\theta])=c$, i.e., the
polynomial interpolant is {\em consistent} with the
functional $F$ in the sense that
the linear term becomes smaller and smaller as we increase the number
of test functions $\theta_k$. In the limit $m\rightarrow \infty$ (infinite number of
test functions) we see the linear term is absent, and we correctly
recover the constant functional.
\end{enumerate}
By extending these arguments to higher-order polynomial functionals
in Hilbert spaces, one can show that Porter's interpolants of order $n$
{\em converge pointwise to entire functionals or any polynomial
functional of order $n$ or less} as the number of interpolation
nodes $\theta_k$ goes to infinity (see \cite{Khlobystov2} and
Theorem 1 in \cite{Khlobystov0}).
\paragraph{Functional Derivatives} The functional derivatives of Porter
interpolants can be easily determined by computing the functional
derivatives of the basis functionals $g_i([\theta])$ defined in \eqref{gi_0}.
To this end, we first notice that
\begin{equation}
\frac{\delta(\theta,\theta_k)^p}{\delta\theta(x)}=(\theta,\theta_k)^{p-1}p\theta_k(x),
\qquad p\geq1.
\end{equation}
A substitution of this formula into \eqref{gi_0} yields
\begin{equation}
\frac{\delta g_{k}([\theta])}{\delta\theta(x)}=
\sum_{j=1}^m \theta_j(x)H_{jk}^{-1}\sum_{p\in\mathcal{I}}\left(\theta_j,\theta\right)^{p-1}p
\label{dgi_0}
\end{equation}
By evaluating $g_{k}([\theta])$ at the nodes $\theta_j(x)$ we obtain a
functional generalization of the classical differentiation matrix \cite{Hesthaven}
\begin{equation}
D^{(1)}_{ji}(x)=\frac{\delta g_i([\theta_j])}
{\delta \theta(x)}.
\label{Hint}
\end{equation}
Similarly, the second-order functional derivative of $g_k([\theta])$ is
\begin{equation}
\frac{\delta^2 g_{k}([\theta])}{\delta\theta(x)\delta\theta(y)}=
\sum_{j=1}^m \theta_j(x) \theta_j(y)H_{jk}^{-1}
\sum_{p\in\mathcal{I}}p(p-1)\left(\theta_j,\theta\right)^{p-2},
\label{dgi_00}
\end{equation}
and it yields the following second-order functional differentiation matrix
\begin{equation}
D^{(2)}_{ji}(x,y)= \frac{\delta^2 g_{i}([\theta_j])}{\delta\theta(x)\delta\theta(y)}.
\label{H2int}
\end{equation}
At this point it is useful to provide simple examples
of functional interpolation in Hilbert spaces.
\vspace{0.7cm}
\noindent
{\em Example 1:}
Consider the first-order polynomial functional
\begin{equation}
P_1([\theta]) =\int_a^b K_1(x)\theta(x)dx.
\end{equation}
To represent $P_1([\theta])$ in terms of a functional interpolant it is
sufficient to consider the set of orthonormal functions
$\widehat{S}^{(m)}_1=\{\varphi_1,...,\varphi_m\}$ (see Eq. \eqref{SNq1}).
In this case, Porter's cardinal basis functionals \eqref{gi_0} reduce to
\begin{equation}
g_i([\theta])=\left(\varphi_i,\theta\right),
\end{equation}
and the functional interpolant can be written as
\begin{equation}
\Pi_1([\theta])=\sum_{i=1}^m P_1([\varphi_i])g_i([\theta]).
\end{equation}
Clearly, we have that $\Pi_1([\theta])\rightarrow P_1$ as
$m\rightarrow \infty$.
\vspace{0.7cm}
\noindent
{\em Example 2:}
Consider the second-order polynomial functional
\begin{equation}
P_2([\theta]) =\int_a^b\int_a^b K_2(x,y)\theta(x)\theta(y)dxdy.
\end{equation}
To represent $P_2([\theta])$ in terms of a functional interpolant it is
sufficient to consider the set of orthonormal functions
\begin{equation}
\hat{\hat{S}}^{(m)}_2=\{\varphi_1,...,\varphi_m,(\varphi_1+\varphi_2),...,
(\varphi_1+\varphi_m),(\varphi_2+\varphi_3), ..., (\varphi_2+\varphi_m), ...\}
\label{sset}
\end{equation}
which is similar (but not equal) to \eqref{SNq1}.
The matrix \eqref{matH} associated with this set has the structure shown in
Figure \ref{fig:porter_matrix_structure}.
\begin{figure}
\centerline{\footnotesize\hspace{1cm}$\hat{\hat{S}}^{(10)}_1$ \hspace{5.cm}$
\hat{\hat{S}}^{(10)}_2$
\hspace{5.cm} $\hat{\hat{S}}^{(10)}_3$}
\centerline{\includegraphics[height=5.5cm]{porter_matrix_1-eps-converted-to.pdf}
\includegraphics[height=5.5cm]{porter_matrix_2-eps-converted-to.pdf}
\includegraphics[height=5.5cm]{porter_matrix_3-eps-converted-to.pdf}}
\caption{Structure of the $H$-matrix \eqref{matH} associated with
the function set $\hat{\hat{S}}^{(10)}_1$, $\hat{\hat{S}}^{(10)}_2$ and
$\hat{\hat{S}}^{(10)}_3$ defined in Eq. \eqref{sset}.
We consider $\mathcal{I}=1$, $\mathcal{I}=2$
and $\mathcal{I}=3$, respectively.}
\label{fig:porter_matrix_structure}
\end{figure}
The cardinal basis functionals \eqref{gi_0} reduce to
\begin{equation}
g_{k}([\theta])=\sum_{j=1}^{\#\hat{\hat{S}}^{(m)}_2} H_{jk}^{-1}\left(\theta_j,\theta\right)^2,
\end{equation}
where $\#\hat{\hat{S}}^{(m)}_2$ is the number of elements
of $\hat{\hat{S}}^{(m)}_2$,
$\theta_k=\varphi_k$ ($k=1,...,m$), $\theta_{m+1}=\varphi_1+\varphi_2$,
etc. The functional interpolant can be written as
\begin{equation}
\Pi_2([\theta])=\sum_{i=1}^{\#\hat{\hat{S}}^{(m)}_2} P_2([\theta_i])g_i([\theta]),
\end{equation}
and it converges $P_2$ as $m\rightarrow \infty$ (see
Section \ref{sec:numerical results functionals}).
\vspace{0.7cm}
\noindent
{\em Example 3:} Consider the third-order polynomial functional
\begin{equation}
P_3([\theta]) =\int_a^b\int_a^b\int_a^b K_3(x_1,x_2,x_3)
\theta(x_1)\theta(x_2)\theta(x_3)dx_1dx_2dx_3.
\label{p3}
\end{equation}
To represent $P_3([\theta])$ in terms of a functional interpolant it is
sufficient to consider the set of functions $\hat{\hat{S}}^{(m)}_3\subset\hat{S}^{(m)}_3$
defined as
\begin{equation}
\hat{\hat{S}}^{(m)}_3=\hat{\hat{S}}^{(m)}_2\cup \{(\varphi_i+\varphi_j+\varphi_k),
\quad k>j>i\}
\end{equation}
where $\hat{\hat{S}}^{(m)}_2$ is as in \eqref{sset}.
The matrix \eqref{matH} associated with this set has the structure shown in
Figure \ref{fig:porter_matrix_structure}
The cardinal basis functionals \eqref{gi_0} reduce to
\begin{equation}
g_{k}([\theta])=\sum_{j=1}^{\#\hat{\hat{S}}^{(m)}_3}
H_{jk}^{-1}\left(\theta_j,\theta\right)^3.
\end{equation}
The functional interpolant can be written as
\begin{equation}
\Pi_3([\theta])=\sum_{i=1}^{\#\widehat{S}^{(m)}_3} P_3([\theta_i])g_i([\theta]).
\end{equation}
\paragraph{More-Penrose Pseudoinverse and Non-Cardinal Basis Functionals}
We emphasized that the matrix $H_{ij}$
defined in \eqref{matH} may be not invertible in some
cases. This happens, for example, if the interpolation
nodes $\theta_k(x)$ are linearly dependent or if there exist
a symmetry such the inner product of $\theta_k$ and $\theta_j$
yields linearly dependent rows/columns
in \eqref{matH}. In this cases we can still construct a
polynomial functional with {\em minimal norm} which, however,
{\em does not} interpolate $F$ at $\theta_k$.
To this end, we simply use the Moore-Penrose
pseudoinverse of $H_{ij}$ to obtain a representation
of Porter's polynomial functionals in terms of
a {\em non-cardinal basis} $g^+_{k}([\theta])$ as
\begin{equation}
\Pi_n \left([\theta]\right)=\sum_{k=1}^m
F\left([\theta_k]\right) g^+_{k}([\theta]).
\label{Porter_interpolant_noncardinal}
\end{equation}
where
\begin{equation}
g^+_{k}([\theta])=\sum_{j=1}^m H_{jk}^{+}\sum_{p\in\mathcal{I}}
\left(\theta_j,\theta\right)^p.
\label{gi_0_noncardinal}
\end{equation}
In the last equation $H_{ij}^+$ denotes the Moore-Penrose
pseudoinverse of $H_{ij}$.
The approximation properties of polynomial functionals
in the form \eqref{Porter_interpolant_noncardinal} will be
studied in Section \ref{sec:Hopf polynomial} and Section
\ref{sec:sine functional}.
\paragraph{Recursive Porter Interpolation}
The number of interpolation nodes
$\theta_k$ required to represent exactly a
polynomial functional of order $n$ is given in
\eqref{SCardinality}.
For example, if we set $m=10$ elementary functions
and polynomial order $n=12$ such formula
yields $646646$ interpolation nodes!
Such large number of nodes may be an issue
when computing Porter's interpolants. In fact,
computing the inverse of \eqref{matH}
rapidly becomes intractable as we increase
either $m$ or $n$. To overcome this problem
we can split the process of inverting the matrix \eqref{matH}
into a recursive algorithm, e.g., by using Schur
complements and blockwise inversion.
To this end, consider the set of nodes
\begin{equation}
\left\{\{\theta_1,...,\theta_m\},\{\theta_{m+1},...,\theta_{2m}\},
\{\theta_{2m+1},...,\theta_{3m}\}, ...,\{\theta_{(N-1)m+1},...,\theta_{Nm}\}\right\},
\label{partitioned_nodes}
\end{equation}
and define the matrices
\begin{align}
H^{(I,J)}_{ij}=\sum_{p\in\mathcal{I}}(\theta_i,\theta_j)^p\qquad i=(I-1)m+1,..,Im\quad j=(J-1)m+1,..,Jm,
\end{align}
where $\mathcal{I}$ denotes the index set of Porter's monomials
while $I$ and $J$ run from $1$ to $N$. The
$H$-matrix \eqref{matH} corresponding to the
set \eqref{partitioned_nodes} can be represented in a block-wise
form as
\begin{equation}
H_N=\left[
\begin{array}{ccc}
H^{(1,1)} & \cdots & H^{(1,N)}\\
\vdots & & \vdots\\
H^{(N,1)} & \cdots & H^{(N,N)}\\
\end{array}
\right].
\end{equation}
The computational cost of inverting such matrix is prohibitive if $Nm$
is large. However, we can use the following recursive procedure. We first build and invert
$H_1=H^{(1,1)}$, corresponding to the first set of $m$. This allows us to
determine Porter's interpolant on the first set of $m$ nodes in \eqref{partitioned_nodes}.
Next we add the second set, i.e., the nodes $\{\theta_{m+1},...,\theta_{2m}\}$.
The matrix \eqref{matH} corresponding to the nodal set $\{\theta_1,...,\theta_{2m}\}$ is
\begin{equation}
H_2=\left[
\begin{array}{cc}
H_1 & C_1\\
C_1^T & H^{(2,2)}\\
\end{array}
\right]\qquad C_1=H^{(1,2)}
\end{equation}
and it can be inverted by using the block-wise formula \cite{Miao}
\begin{equation}
H^{-1}_2=\left[
\begin{array}{cc}
H_1^{-1}+ B_1A_1B_1^T & -B_1A_1\\
-A_1B_1 & A_1 \\
\end{array}
\right]
\end{equation}
where
\begin{align}
B_1&= H_1^{-1} C_1,\\
A_1&=\left(H^{(2,2)}- C_1^TB_1\right)^{-1} \qquad
\textrm{(inverse of the Schur complement)}.
\end{align}
Now we bring in the third set of nodes $\{\theta_{2m+1},...,\theta_{3m}\}$.
The matrix \eqref{matH} corresponding to the nodal set $\{\theta_1,...,\theta_{3m}\}$ is
\begin{equation}
H_3=\left[
\begin{array}{cc}
H_2 & C_2 \\
C_2^T & H^{(3,3)}\\
\end{array}
\right]\qquad
C_2=\left[
\begin{array}{c}
H^{(1,3)}\\H^{(2,3)}
\end{array}
\right],
\end{equation}
and its inverse is, as before,
\begin{equation}
H^{-1}_3=\left[
\begin{array}{cc}
H_2^{-1}+ B_2A_2B_2^T & -B_2A_2\\
-A_2B_2 & A_2 \\
\end{array}
\right]
\end{equation}
where
\begin{align}
B_2= H_2^{-1} C_2,\qquad A_2=\left(H^{(3,3)}- C_2^TB_2\right)^{-1}.
\end{align}
At this point it is clear that the procedure can be iterated as many times
as needed. This generates a sequence of basis functionals \eqref{gi_0}, and
an interpolating polynomial functional with minimal norm that
passes through an increasing number of nodes.
In this way, we have reduced
the problem of computing Porter's interpolant through a very large number
of nodes into a sequence of matrix inversions of dimension at most $m\times m$.
The storage requirements of the algorithm just described, however,
is not small as because Porter's basis functionals are
ultimately defined by $H_N^{-1}$. An open question is
the identification of interpolation nodes $\theta_j$
leading to minimal complexity/storage requirements
for $H_N^{-1}$.
An alternative interpolation method makes use of residuals.
The main idea is sketched in Figure \ref{fig:recursive_sketch}.
\begin{figure}
\centerline{\includegraphics[height=6cm]{Recursive_sketch-eps-converted-to.pdf}}
\caption{Recursive polynomial interpolation: Minimal residual approach.}
\label{fig:recursive_sketch}
\end{figure}
The functional $F([\theta])$ is interpolated by a polynomial
of order $n$, denoted as $\Pi^{(1)}_n([\theta])$, at just three
nodes $\{\theta_1,\theta_2,\theta_3\}$.
Subtracting $\Pi^{(1)}_n([\theta])$ from $F([\theta])$ yields
the functional residual
\begin{equation}
R^{(1)}_n([\theta]) = F([\theta])-\Pi^{(1)}_n([\theta])
\end{equation}
which is zero at
$\theta_1$, $\theta_2$ and $\theta_3$ because of the interpolation
condition. Now we add three more nodes $\{\theta_4,\theta_5,\theta_6\}$
and construct a Porter's interpolant of $R^{(1)}_n([\theta])$ at
$\{\theta_1,...,\theta_6\}$. We denote such interpolant
by $\Pi^{(2)}_n([\theta])$. The computation of
$\Pi^{(2)}_n([\theta])$ can be carried out as above by using Schur complements
and block-wise inversion of \eqref{matH}. The polynomial functional
$\Pi^{(1)}_n([\theta])+\Pi^{(2)}_n([\theta])$
interpolates $F([\theta])$ at $\{\theta_1,...,\theta_6\}$. The recursive construction
proceeds with the definition of the new residual
\begin{equation}
R^{(2)}_n([\theta])= F([\theta])- \Pi^{(1)}_n([\theta])- \Pi^{(2)}_n([\theta]),
\end{equation}
three more nodes $\{\theta_7,\theta_8,\theta_9\}$,
and a Porter's polynomial $\Pi^{(3)}_n([\theta])$, interpolating
$R^{(2)}_n([\theta])$ at $\{\theta_1,...,\theta_9\}$.
Proceeding recursively with higher-order
residuals up to order $r$, we obtain the polynomial functional
\begin{equation}
Q_n([\theta])= \sum_{k=1}^r \Pi^{(k)}_n([\theta]).
\end{equation}
Clearly $Q_n([\theta])$ interpolates $F([\theta])$ at all
nodes $\theta_k$ and therefore it is completely
equivalent to $\Pi_n([\theta])$, i.e.,
$Q_n([\theta])=\Pi_n([\theta])$.
\paragraph{Hierarchical Matrices}
The algorithm we just described aim at reducing
the computational cost of computing polynomial
functional interpolants by inverting the matrix $\bm H$
defined in \eqref{matH} in a block-wise fashion or recursively.
The structure of such matrix obviously depends on how we select
the interpolation nodes in the function space $D(F)$.
An interesting open question is whether we can determine
sets of nodes for which the interpolation problem
can be solved at a minimal cost. Stated in matrix terms,
can we identity sets of nodes yielding structured matrices that
can be easily inverted, e.g., {\em hierarchical matrices}?
The matrices shown in Figure \ref{fig:porter_matrix_structure}
have indeed a self similar structure which can be used to speed up
their inversion. We leave this question open for future research.
For uniquely solvable interpolation problems we could equivalently
construct Khlobystov polynomial functionals (see
Section \ref{sec:Khlobystov}), and determine the coefficients
of the expansion by using the method of moments.
\subsubsection{Prenter Interpolants}
\label{sec:prenter}
Another method to determine polynomial
interpolants in Hilbert and Banach spaces was introduced by
Prenter in \cite{Prenter}. She proved that
if $F([\theta])$ is a functional in a Hilbert space $D(F)$,
and $\theta_j\in D(F)$ are interpolation nodes,
then there exists a $n$th-order functional
interpolant in the form
\begin{equation}
\Pi_n ([\theta])=\sum_{i=1}^n F([\theta_i])g_i([\theta]),
\label{prenter}
\end{equation}
where
\begin{equation}
g_i([\theta]) = \prod_{\substack{j=1\\j\neq i}}^n \frac{
\left(\theta -\theta_j, \theta_i -\theta_j\right)}
{\left(\theta_i -\theta_j, \theta_i -\theta_j\right)} \qquad \textrm{(cardinal basis)}.
\label{gi_prenterer}
\end{equation}
As before, $(,)$ denotes the inner product in
$L_2(V)$, where $V$ is the domain of $\theta_k(x)$.
Note that each basis element $g_i([\theta])$ is a
polynomial functional of order $n$.
On the other hand, Porter's method yields polynomial basis
functionals of total degree $\max(\mathcal{I})$, where the index set
$\mathcal{I}$ does not depend on the number of collocation points.
Porter \cite{Porter} applied Prenter's theorems to causal systems,
while Bertuzzi, Gandolfi and Germani \cite{Bertuzzi1,Bertuzzi}
extended Prenter's results to causal approximation of
input-output maps in Hilbert spaces. Generalizations to Banach spaces
can be found in Chapter 3 of \cite{Torokhti} (see also \cite{Makarov}
and the references therein).
The functional derivatives of Prenter's polynomial functionals can
be obtained by computing the functional derivatives of \eqref{gi_prenterer}.
This yields\footnote{These expressions can be easily proved by noting that
\begin{equation}
\left.\frac{d}{d\epsilon}g_i([\theta+\epsilon \eta])\right|_{\epsilon=0}=\sum_{\substack {k=1\\k\neq i}}^n
\frac{(\eta,\theta_i-\theta_k)}{\left\|\theta_i-\theta_k\right\|^2}
\prod_{\substack{j=1\\j\neq k,i}}^n
\frac{(\theta-\theta_j,\theta_i-\theta_j)}{\left\|\theta_i-\theta_j\right\|^2}.
\end{equation}
}
\begin{equation}
\frac{\delta g_i([\theta])}{\delta \theta(x)}=\sum_{\substack {k=1\\k\neq i}}^n
\frac{\theta_i(x)-\theta_k(x)}
{\left\|\theta_i-\theta_k\right\|^2}g_{ik}([\theta]),\label{fd1}
\end{equation}
\begin{equation}
\frac{\delta^2 g_i([\theta])}{\delta \theta(x)\delta \theta(y)}=
\sum_{\substack {k=1\\k\neq i}}^n
\frac{\theta_i(x)-\theta_k(x)}
{\left\|\theta_i(x)-\theta_k(x)\right\|^2}
\sum_{\substack {s=1\\s\neq k}}^n
\frac{\theta_i(y)-\theta_s(y)}
{\left\|\theta_i(y)-\theta_s(y)\right\|^2}
g_{iks}([\theta]),\label{fd2}
\end{equation}
where
\begin{equation}
g_{ik}([\theta])=
\prod_{\substack{j=1\\j\neq i,k}}^n
\frac{(\theta-\theta_j,\theta_i-\theta_j)}{\left\|\theta_i-\theta_j\right\|^2},\qquad
g_{iks}([\theta])=
\prod_{\substack{j=1\\j\neq i,k,s}}^n
\frac{(\theta-\theta_j,\theta_i-\theta_j)}{\left\|\theta_i-\theta_j\right\|^2}.
\end{equation}
Note that the functional derivatives \eqref{fd1}-\eqref{fd2} are more
complicated than the ones we obtained in the case Porter's basis \eqref{dgi_0}.
As noted by Allasia and Bracco in \cite{Allasia}, Prenter's interpolants
are badly conditioned as the number of interpolation nodes in
function space increases. This unfortunate feature is common
to many Lagrange interpolants.
\subsubsection{Khlobystov Interpolants}
\label{sec:Khlobystov}
We have seen that any continuous
functional in a Hilbert or a Banach space
can be approximated uniformly in terms of polynomial
functionals \cite{Prenter1,Bertuzzi,Istratescu,Makarov,Howlett,Torokhti}.
Such polynomial functionals may be built based
on an interpolation process (see Section
\ref{sec:porter} and Section \ref{sec:prenter}).
What are the convergence properties of
such expansions?
To address this question, let us consider a polynomial interpolation
problem of a given polynomial functional $\Pi_n([\theta])$ in
the form \eqref{pfun1}. Assume that $D(\Pi_n)$ is a Hilbert
space of test functions, introduce the orthonormal
basis $\{\varphi_1,...,\varphi_m\}$ in $D(\Pi_n)$, and
consider the following expansion
\begin{equation}
\Pi^I_n([\theta])=\sum_{k=1}^n \sum_{i_1=1}^m\cdots \sum_{i_k=1}^m
(\theta,\varphi_{i_1})\cdots(\theta,\varphi_{i_k})a_k(\varphi_{i_1},...,\varphi_{i_k}),
\label{interp4}
\end{equation}
where $(,)$ is an inner product in $D(\Pi_n)$ and
$a_k(\varphi_{i_1},...,\varphi_{i_k})$ are real- or complex-valued coefficients.
Clearly $\Pi^I_n([\theta])$ is an interpolant
of $\Pi_n([\theta])$ at $\varphi_k$, i.e.,
\begin{equation}
\Pi^I_n([\varphi_k])=\Pi_n([\varphi_k]).
\end{equation}
The next step is to write the
coefficients $a_k(\varphi_{i_1},...,\varphi_{i_k})$ in
terms of $\Pi_n([\varphi_j])$. To this end, we can use the method
for finding orthonormal moments of regular polynomial
functionals (Theorem 1 in \cite{Kashpur}). This yields
\begin{align}
n!a_n(\varphi_{i_1},...,\varphi_{i_n})=&
\Pi_n([\varphi_{i_1}+\cdots +\varphi_{i_n}])-
\left\{\Pi_n([\varphi_{i_1}+\cdots +\varphi_{i_{n-1}}])+\nonumber\right.\\
& \left.\Pi_n([\varphi_{i_1}+\cdots +\varphi_{i_{n-2}}+\varphi_{i_{n}}])+\cdots
+\Pi_n([\varphi_{i_2}+\cdots +\varphi_{i_{n}}])\right\}+\nonumber\\
&\left\{ \Pi_n([\varphi_{i_1}+\cdots +\varphi_{i_{n-2}}])+
\Pi_n([\varphi_{i_1}+\cdots +\varphi_{i_{n-3}}+\varphi_{i_{n-1}}])+\cdots+\right.\nonumber\\
&\left.\Pi_n([\varphi_{i_3}+\cdots +\varphi_{i_{n}}])\right\} \cdots
+(-1)^{n-1}\left\{\Pi_n([\varphi_{i_1}])+\cdots +\Pi_n([\varphi_{i_n}])\right\}+\nonumber\\
& (-1)^{n}\Pi_n([0]).
\label{coeffKL}
\end{align}
Once $a_n(\varphi_{i_1},...,\varphi_{i_n})$
are available, we can construct the polynomial
$$\Pi_n-a_n(\varphi_{i_1},...,\varphi_{i_n})
\prod_{z=1}^n(\theta,\varphi_{i_z})$$
and apply \eqref{coeffKL} again to determine
$a_{n-1}(\varphi_{i_1},...,\varphi_{i_{n-1}})$.
After $n$ iterations, we have available all
coefficients to construct the
polynomial functional \eqref{interp4}.
It was shown in \cite{Kashpur} (Theorem 2) that
the following error estimate holds
\begin{equation}
\left\|\Pi^I_n([\theta])-\Pi_n([\theta])\right\|\leq \sum_{k=1}^n\left\|
P_k\right\|\left[\left(\left\|\theta\right\|+\epsilon_m([\theta])\right)^k-
\left\|\theta\right\|^k\right],\label{conver}
\end{equation}
where
\begin{equation}
\left\|P_k\right\|=\int\cdots\int \left|K_k(x_1,...,x_k)\right|dx_1\cdots dx_k,
\end{equation}
\begin{equation}
\epsilon_m([\theta])=\left\|\theta-\sum_{k=1}^m(\theta,\varphi_k)\varphi_k\right\|.
\end{equation}
This means that the interpolant \eqref{interp4} converges
{\em pointwise} to the polynomial functional
$\Pi_n([\theta])$ in \eqref{pfun} as the number of nodes $\{\varphi_k\}$
in the function space $D(\Pi_n)$ goes to infinity\footnote{Note that
$\epsilon_m([\theta])\rightarrow 0$ as $m\rightarrow \infty$.
Thus, the right hand side of the error estimate \eqref{conver}
goes to zero as $m\rightarrow \infty$, i.e.,
$\Pi^I_n([\theta])\rightarrow \Pi_n([\theta])$
pointwise as $m\rightarrow \infty$.}.
As noted by Khlobystov in \cite{Khlobystov0},
Porter's interpolant \eqref{Porter_interpolant} can be written
exactly in the form \eqref{interp4} if we consider test
functions in the form $\theta_i=\varphi_{i_1}+...+\varphi_{i_n}$
(see Lemma 1 in \cite{Khlobystov0}). In this sense, Porter's
interpolants represent the Lagrangian form of
Khlobystov's interpolants.
An interesting and very important question is the approximation
of polynomial functionals of order $n$ in terms interpolants
over nodal sets $\widehat{S}^{(m)}_q$, with $q<n$ (see Eq. \eqref{SNq1}).
A specific example would be the approximation of $P_3([\theta])$
in \eqref{p3} by using an interpolant built on the
set of nodes $\widehat{S}^{(m)}_1$.
In some very special cases we may get uniform convergence
as $m\rightarrow \infty$, for example when $K_3$ is diagonal
\begin{equation}
K_3(x_1,x_2,x_3)=\sum_{j=1}^m a_j\varphi_j(x_1)\varphi_j(x_2)\varphi_j(x_3).
\label{almostCP}
\end{equation}
However, this won't happen in general, and therefore
we will have to accept a systematic truncation error in representing
continuous nonlinear functionals. Such error is similar the
error committed when we approximate infinite-variate functions, e.g.,
by second-order HDMR \cite{Wasilkowski1} (see Section
\ref{sec:HDMR}).
\paragraph{Khlobystov interpolants with Hilbert-Schmidt Kernels}
In this Section we present a procedure to construct interpolants of
polynomial operators in Hilbert spaces with separable kernels.
To this end, let us consider an orthonormal
basis $\{\varphi_1,\varphi_2,...,\varphi_m\}$
and represent each kernel in \eqref{lP} as
\begin{align}
K_1(x_1)=&\sum_{i=1}^m a_{i}\varphi_{i}(x),\label{k1}\\
K_2(x_1,x_2)=&\sum_{i,j=1}^m a_{ij}\varphi_{i}(x_1)\varphi_{j}(x_2),
\label{k2}\\
K_3(x_1,x_2,x_3)=&\sum_{i,j,k=1}^m a_{ijk}\varphi_{i}(x_1)
\varphi_{j}(x_2)\varphi_{k}(x_3),\label{k3}\\
\cdots .\nonumber&
\end{align}
Without loss of generality we can assume that $K_2$, $K_3$, ...
are symmetric, i.e., that the coefficients $a_{ijk\cdots}$ are invariant under
any permutation of the indices $i$, $j$, $k$, etc.
A substitution of \eqref{k1}, \eqref{k2}, etc., into
\eqref{pfun} yields the polynomial functional
\begin{align}
\Pi_n([\theta]) = K_0
+\sum_{i=1}^m a_i(\varphi_i,\theta)
+\sum_{i,j=1}^m a_{ij}(\varphi_i,\theta)(\varphi_j,\theta)
+\sum_{i,j,k=1}^m a_{ijk}(\varphi_i,\theta)(\varphi_j,\theta)
(\varphi_k,\theta)+\cdots.
\end{align}
At this point we pose the following question: How many interpolation
nodes do we need to identify the
coefficients $a_i$, $a_{ij}$, $a_{ijk}$, etc., uniquely?
To clarify the question, consider the following second-order
polynomial functional
\begin{align}
\Pi_2([\theta]) = K_0
+\sum_{i=1}^m a_i(\varphi_i,\theta)
+\sum_{i,j=1}^m a_{ij}(\varphi_i,\theta)(\varphi_j,\theta).
\label{p2interp}
\end{align}
The total number of degrees of freedom (number of independent
coefficients $K_0$, $a_i$ and $a_{ij}$) is
\begin{equation}
1+m+m+\binom{m}{2}=1+\frac{3m+m^2}{2}.
\end{equation}
To determine
such coefficients, we need to test
$\Pi_2([\theta])$ at $(2+3m+m^2)/2$ distinct nodes
\footnote{Note, that testing a second-order polynomial functional
in $\varphi_i(x)$ or $2\varphi_i(x)$ yields different results.}, e.g.,
\begin{equation}
\left\{0,\{\varphi_i\}_{i=1}^m,\{\varphi_i+
\varphi_j\}_{\substack{i,j=1\, (j\geq i)}}^m\right\}.
\label{test_f_s}
\end{equation}
This yields the linear system
\begin{align}
\Pi_2([0]) &= K_0\nonumber\\
\Pi_2([\varphi_p]) &= K_0+a_p+a_{pp}\nonumber\\
\Pi_2([\varphi_p+\varphi_q]) &= K_0+a_p+a_q+ a_{pp}+
a_{qq}+2a_{pq}\nonumber
\end{align}
which can be immediately solved for $a_{pq}$, $a_p$ and $K_0$
\begin{align}
a_{pq}&=\frac{1}{2}\left(\Pi_2([\varphi_p+\varphi_q])-
\Pi_2([\varphi_p])-\Pi_2([\varphi_q])+\Pi_2([0])\right),\label{apq}\\
a_{p}&=-\frac{1}{2}\Pi_2([2\varphi_p])+2\Pi_2([\varphi_p])-\frac{3}{2}
\Pi_2([0]),\label{ap}\\
K_0 &= \Pi_2([0])\label{k0}.
\end{align}
In this way, we can identify the kernels
\eqref{k1}-\eqref{k3} and therefore
any polynomial functional in the form \eqref{p2interp}.
Note that if the basis elements in \eqref{test_f_s} are not
normalized (but still orthogonal) then we simply need to
replace $a_{pq}$ and $a_p$ in \eqref{apq}-\eqref{k0}
with $\left\|\varphi_p\right\|^2\left\|\varphi_q\right\|^2a_{pq}$ and
$\left\|\varphi_p\right\|^2a_{p}$, respectively.
Higher-order polynomial functionals can be constructed
in a similar way. However, the number of degrees of freedom
may increase significantly with the order of the polynomial
(see equation \eqref{SCardinality}).
For example, a twelve-order polynomial functional in which
each kernel is represented relative to $m=10$ basis functions
(tensor product) yields $646646$ degrees of freedom!
\vspace{0.7cm}
\noindent
{\em Remark:}
The fact that we can get analytical expressions for the
coefficients of the polynomial interpolant means that
Porter's matrix \eqref{Hint} is {\em invertible analytically}
for uniquely solvable interpolation problems and orthonormal
bases.
\vspace{0.7cm}
\noindent
The procedure we just described to identify the coefficients
of the symmetric kernel functions $K_n$ relies on
tensor product representations, i.e., series expansions in the form
\begin{align}
K_n(x_1,x_2,...,x_n)=\sum_{i_1,...,i_n=1}^{m}a_{i_1\cdots i_n}
\varphi_{i_1}(x_1)\cdots \varphi_{i_n}(x_n).
\label{eq:tp}
\end{align}
The number of independent coefficients $a_{i_1\cdots i_n}$ is
\begin{equation}
\binom{n+m-1}{n}=\frac{(n+m-1)!}{n!(m-1)!}.
\label{dof_sym}
\end{equation}
Such number can pose serious computational challenges even for moderate
values of $m$ and $n$.
To alleviate this problem one could use
HDMR expansions \cite{Rabitz,Li,Li1}, i.e., represent
$K_n(x_1,x_2,\cdots,x_n)$ in terms of a superimposition of
functions involving a lower number of variables
(interaction terms). This yields, for example
\begin{align}
K_n(x_1,x_2,...,x_n)=f_{0}+\sum_{i=1}^{n}f_{i}(x_{i})
+\sum_{i<j}^{n}f_{ij}(x_{i},x_{j})+\sum_{i<j<k}^{n}f_{ijk}(x_{i},x_{j},x_k)
+ \cdots\,.
\label{eq:anova}
\end{align}
The function $f_{0}$ is a constant. The functions $f_i(x_i)$
(first-order interactions) give us the overall effects of the
variables $x_i$ in $f$ as if they were acting
independently of the other input variables. The functions $f_{ij}(x_i, x_j)$
(second-order interactions) describe the cooperative effect
of the variables $x_i$ and $x_j$. Similarly, higher-order terms
reflect the cooperative effects of an increasing number of variables.
Representing $f_{i}$, $f_{ij}$, $f_{ijk}$
relative to the orthonormal basis $\{\varphi_j(x)\}$
yields the series
\begin{align}
K_n(x_1,x_2,...,x_n)=& f_{0}+\sum_{i=1}^{n}\sum_{s=1}^m a^{i}_s
\varphi_s(x_{i})
+\sum_{i<j}^{n}\sum_{s,q=1}^ma^{ij}_{sq}\varphi_s(x_{i})
\varphi_q(x_{j}) + \nonumber \\
&\sum_{i<j<k}^{n} \sum_{s,q,r=1}^m a^{ijk}_{sqr}\varphi_s(x_{i})
\varphi_q(x_{j})
\varphi_r(x_{j})+\cdots.
\label{eq:anova1}
\end{align}
Given the symmetry of each function $f_{i}$, $f_{ij}$, $f_{ijk}$, the total
number of degrees of freedom of an HDMR expansion of order $Z$ is
\begin{equation}
\sum_{k=0}^Z \binom{n}{k} \binom{m+k-1}{k} = 1+\binom{n}{1}
\binom{m}{1} +\binom{n}{2}\binom{m+1}{2} +
\binom{n}{3}\binom{m+2}{3} +\cdots.
\end{equation}
For example, the second- and third-order
truncations of a $n=10$ dimensional
kernel relative to $m=10$ yield $2576$ and $28976$
degrees of freedom, respectively. On the other hand,
the tensor product representation yields $92378$
degrees of freedom.
{\color{r}
Alternatively, one can use tensor expansions
(see Section \eqref{sec:tensor}), e.g., canonical polyadic or
hierarchical-Tucker series, of each kernel to
reduce the number of degrees of freedom.
The tensor expansion can be fit to data
by interpolation, least-squares or
projection \cite{Nouy2017,NouyHUQ,Nouy2,Nouy}.
}
{\color{r}
\subsection{Functional Approximation by Tensor Methods}
\label{sec:tensor}
Computing high-order polynomial functional expansions
requires representing kernel functions in many
independent variables. To get an feeling
of how serious this problem could be, simply consider that
representing a polynomial functional of order $6$ is as computationally
expensive as representing the solution to the
steady-state Boltzmann equation \cite{Dimarco}, a well-known
challenging problem in computational physics.
Expanding each kernel of the polynomial functional
in terms of HDMR \cite{Rabitz} or canonical polyadic
decompositions -- i.e. separated series expansions \cite{Beylkin} --
can mitigate the dimensionality problem, but
it may not be the most efficient way to proceed.
In this Section we discuss nonlinear functional approximation
based on tensor methods. To introduce these methods, consider
the Hilbert space of functions spanned by the finite-dimensional
basis $\{\varphi_1,...,\varphi_m\}$ (e.g., an orthonormal basis)
and look for an approximant of $F([\theta])$, say $f$,
in the form
\begin{equation}
F([\theta])= f(a_1([\theta]), ..., a_m([\theta]))
+R([\theta]),\qquad a_k([\theta])=(\theta,\varphi_k).
\label{functional-approx}
\end{equation}
In this equation, $f$ is a multivariate function of
$a_k([\theta])$ (linear functionals of $\theta$), $(,)$ denotes an
inner product in the Hilbert Space $D(F)$, e.g.,
\begin{equation}
(\theta,\varphi_k)=\int_a^b \theta(x)\varphi_k(x)dx,
\qquad k=1,...,m
\end{equation}
and $R$ is a (functional) reminder term.
The functionals $a_k([\theta])=(\theta,\varphi_k)$ can be either real or
complex-valued. In the theory of stochastic processes
the set
\begin{equation}
\left\{\theta\in D(F) \,|\, ((\theta,\varphi_1),...,(\theta,\varphi_m))\subseteq\mathbb{R}^m\right\}
\end{equation}
is known as {\em cylindrical set} (see, e.g., \cite{Neerven} p. 56
or \cite{Skorokhod} p. 45). Therefore, according to
Eq. \eqref{functional-approx}, we are looking for an approximant
of $F([\theta])$ in the space of {\em cylindrical functionals}, i.e.,
functionals defined on cylindrical sets.
Thanks to the Stone-Weierstrass theorem,
cylindrical functionals in the form \eqref{cf} can approximate
any continuous functional in a Hilbert or a Banach space.
The representation \eqref{functional-approx}
is very general. For example, it includes the
case where $f$ is a polynomial functional,
e.g., \eqref{interp4} or \eqref{Porter_interpolant},
or the case where the functional is defined
on a finite-dimensional function space (see Section \ref{sec:Finite_Dim_Approx}).
For notational convenience we will often drop the
the functional dependence of $a_k([\theta]$
and write \eqref{functional-approx} as
\begin{equation}
F([\theta]) \simeq f(a_1,...,a_m),\qquad a_j=(\theta,\varphi_j).
\label{cf}
\end{equation}
%
Hereafter, we discuss effective numerical
algorithms to compute an approximation
of the multivariate function $f(a_1,....,a_m)$ based on
{\em high-dimensional model representations (HDMR)}, and {\em tensor
methods}\footnote{If we ask the question ``what is a tensor?''
to an engineer, a physicist or a mathematician we usually get different
answers. The engineer usually says ``a tensor is a matrix''.
On the other hand, the physicist would say that a tensor is a
mathematical object that has the fundamental property
of transforming in a very specific way when the coordinate
system is changed. He or she would point out that
the laws of physics are built upon the {\em principle of general
covariance} \cite{Weinberg,VenturiJPA2009}
(invariance of physical laws
relative to coordinate transformations) which is formulated in
a natural way in terms of tensors. The word ``tensor''
has recently spilled in the multi-linear algebra community
to represent multi-dimensional arrays. Hereafter we will adopt
such terminology, and refer to tensors as multi-dimensional arrays.}
\cite{Kolda,Hackbusch_book}, including
canonical tensor decompositions, hierarchical Tucker formats,
and tensor networks.
Such algorithms rely on optimization (e.g., the alternating
least squares methods \cite{Acar,Reynolds}), or multilinear
algebra techniques such as high-order singular value decomposition \cite{Grasedyck2017}, randomized block sampling,
or generalized Schur decompositions.
\paragraph{Functional Derivatives}
Let us compute the functional derivatives of the cylindrical
functional approximation \eqref{cf}. To this end, we evaluate
the G\^ateaux differential of $f(a_1([\theta]), ..., a_m([\theta])$
in the direction of an arbitrary function $\eta(x)$ to obtain
\begin{align}
\int_a^b \frac{\delta F([\theta])}{\delta \theta(x)}\eta(x)dx&\simeq
\frac{d}{d\epsilon}\left[f((\theta+\epsilon\eta,\varphi_1),...,
(\theta+\epsilon\eta,\varphi_m))\right]_{\epsilon=0}\\
&=\sum_{k=1}^m\frac{\partial f}{\partial a_k}(\eta,\varphi_k).
\end{align}
Setting $\eta(x)=\delta(x-y)$ yields
the following approximation for the first-order functional derivative
\begin{align}
\frac{\delta F([\theta])}{\delta \theta(y)}\simeq \sum_{k=1}^m
\frac{\partial f}{\partial a_k}\varphi_k(y),
\label{ff1}
\end{align}
where $a_k=(\theta,\varphi_k)$. On the other hand,
by projecting \eqref{ff1} onto $\varphi_j$
we obtain
\begin{equation}
\int_a^b \frac{\delta F([\theta])}{\delta \theta(x)}\varphi_j(x)dx\simeq
\frac{\partial f}{\partial a_j}.
\label{ff2}
\end{equation}
This means that partial derivative of $f$ relative to
$a_k=(\theta,\varphi_k)$ approximates
the projection of the functional derivative
of $F$ along $\varphi_k$. Equations \eqref{ff1} and \eqref{ff2}
are consistent with previous results on functional derivatives
in finite-dimensional spaces (see Eqs. \eqref{gg1} and \eqref{gg2}).
By following the same procedure, we can construct functional
derivatives of $f$ of higher-order, e.g.,
\begin{align}
\frac{\delta^2 F([\theta])}{\delta \theta(x)\delta \theta(y)}\simeq&\sum_{k,j=1}^m
\frac{\partial^2 f}{\partial a_k\partial a_j}\varphi_k(x)\varphi_j(y),
\label{ff5}\\
\frac{\delta^3 F([\theta])}{\delta \theta(x)\delta \theta(y)\delta \theta(z)}
\simeq
&\sum_{k,j,i=1}^m
\frac{\partial^3 f}{\partial a_k\partial a_j\partial a_i}
\varphi_k(x)\varphi_j(y)\varphi_i(z).
\label{ff6}
\end{align}
\paragraph{Choice of the Number of Active Dimensions}
The choice of the basis functions $\varphi_k$ and
the number of active dimensions, i.e., the
parameter $m$ in \eqref{functional-approx},
is {\em critical} for the accuracy of the cylindrical representation.
For a fixed basis $\{\varphi_j\}$, smaller values of $m$
yield faster computations but at the same time can
lead to functional approximation problems with poor
approximation errors.
\subsubsection{Canonical Tensor Decomposition}
\label{sec:CP}
The canonical tensor decomposition of the multivariate
function $f(a_1,...,a_m)$ in \eqref{functional-approx}
is a series expansion in the form
\begin{equation}
f(a_1,...,a_m)=\sum_{l=1}^r
\prod_{i=1}^mG^l_i(a_i)\qquad a_j=(\theta,\varphi_j),
\label{functional-SSE}
\end{equation}
where $G^l_i(a_i)$ are one-dimensional
functions usually represented relative to a
known basis $\{\phi_1,...,\phi_Q\}$, i.e.,
\begin{equation}
G^l_i(a_i)=\sum_{k=1}^Q\beta_{ik}^l \phi_k(a_i).
\label{gfun}
\end{equation}
The quantity $r$ in \eqref{functional-SSE} is
called {\em separation rank}.
In the statistics literature, representations of
the form \eqref{functional-SSE}
are known as {\em parallel factorizations} (see
\cite{Kronenberg,Leugarans}).
They are also known as proper generalized
decompositions \cite{ChinestaBook},
canonical polyadic expansions (CP) \cite{Karlsson},
separated series \cite{HeyrimJCP_2014},
and Kruskal tensor formats \cite{Kolda}.
Although there are at present no useful theorems on the size of
the separation rank $r$ needed to represent with accuracy
general classes of functionals $f$, there are cases where the
expansion \eqref{functional-SSE} is {\em exponentially
more efficient} than one would expect a priori.
The basis
functions $\phi_k$ appearing in \eqref{gfun} represent the
variability of the functional $f$ along different directions
$a_k=(\theta,\varphi_k)$ in the test function space $D_m$.
As such, they have to satisfy appropriate boundary conditions.
For example, if $f$ is periodic in the hypercube $[-b,b]^m$
then we could use rescaled trigonometric polynomials
\begin{equation}
\phi_k(a)=l_k\left(\pi\left(\frac{a}{b}+1\right)\right)\qquad a\in[-b,b],
\label{Fourier}
\end{equation}
where
\begin{equation}
l_{k+1}(x)=\frac{1}{Q}\displaystyle
\frac{\sin\left(\displaystyle\frac{Q}{2}\left(x-x_k\right)\right)}
{\sin\left(\displaystyle\frac{x-x_k }{2} \right)},
\qquad x_k=\frac{2\pi}{Q}k\qquad k=0,...,(Q-1).
\end{equation}
For more general boundary conditions we can use a polynomial basis,
e.g., rescaled Legendre orthogonal polynomials
\begin{equation}
\phi_k(a)=L_k\left(\frac{a}{b}\right)\qquad a\in[-b,b],
\label{Legendre}
\end{equation}
where
\begin{align}
L_0(x)=1, \quad
L_1(x)=x,\quad
\cdots, \quad
L_{n+1}(x)=\frac{2n+1}{n+1}xL_{n}(x)-\frac{n}{n+1}L_{n-1}(x), \quad
(n=1,...,Q-1).
\end{align}
The $L_2$ norm of \eqref{Fourier} and \eqref{Legendre} is easily
obtained as
\begin{equation}
\eta_k= \frac{b}{\pi}\frac{2\pi}{Q}\quad \textrm{(Fourier)},\qquad
\eta_k= \frac{2b}{2h+1} \quad \textrm{(Legendre)}, \qquad
k=1,...,Q.
\end{equation}
}
\vspace{0.7cm}
\noindent
{\em Example 1:} Consider the sine functional
\eqref{sinFunctional}, hereafter rewritten for convenience
\begin{equation}
F([\theta])=\sin\left(\int_a^b K(x)\theta(x)dx\right).
\label{sinFunctional1}
\end{equation}
Assuming that the kernel $K(x)$ admits the
finite-dimensional expansion
\begin{equation}
K(x)=\sum_{j=1}^m k_j\varphi_j(x),
\label{kkk}
\end{equation}
and substituting it into \eqref{sinFunctional1} we obtain
\begin{align}
f(a_1,...,a_m)=&\sin\left(\sum_{j=1}^m k_j a_j\right)\nonumber \\
=&\sum_{j=1}^m\sin(k_ja_j)
\prod_{\substack{i=1\\i\neq j}}^m \frac{\sin(k_ia_i+\chi_i-\chi_j)}
{\sin(\chi_i-\chi_j)},
\end{align}
for all $\chi_i$ such that $\chi_i\neq \chi_j$. The last equality was
derived in \cite{Mohlenkamp}, and it shows that the separation rank
of the canonical tensor decomposition of
\eqref{sinFunctional1}-\eqref{kkk} is exactly $r=m$. In other words,
we can represent the nonlinear functional \eqref{sinFunctional1}
exactly in terms of superimposition of $m$ terms.
Furthermore, if we allow the $G^l_i(a_i)$
to be complex-valued\footnote{Constraints on the
functions $G^l_i$ such as positivity can be also imposed, e.g.,
if one is interested in probability functionals.}
then
\begin{equation}
\sin\left(\sum_{j=1}^m k_ja_j\right)=
\prod_{j=1}^m\frac{e^{ik_ja_j}}{2i}-
\prod_{j=1}^m\frac{e^{-ik_ja_j}}{2i},
\end{equation}
i.e., we can reduce the separation rank to $r=2$.
In general, the separation rank depends on the
complexity of the nonlinear function $f(a_1,...,a_m)$.
\paragraph{Functional Derivatives}
With the canonical tensor decomposition \eqref{functional-SSE}
available, it is straightforward to compute an
approximation of the functional derivatives
of $F([\theta])$. Recalling that the canonical tensor
decomposition is a cylindrical representation of the functional
$F([\theta])$, we have the expressions
\eqref{ff1}, \eqref{ff5} and \eqref{ff6}.
The unknowns are the partial derivatives
of $f$ with respect to $a_j([\theta])$, which can be
computed based on \eqref{functional-SSE} as
\begin{equation}
\frac{\partial f}{\partial a_i}=
\sum_{l=1}^r\frac{\partial G^l_i}{\partial a_i}
\prod_{\substack{k=1\\k\neq i}}^m G_k^l(a_k),
\label{pd1}
\end{equation}
and
\begin{equation}
\frac{\partial f}{\partial a_i\partial a_j}=
\begin{cases}
\displaystyle\sum_{l=1}^r\frac{\partial^2 G^l_i}{\partial a_i^2}
\prod_{\substack{k=1\\k\neq i}}^m G_k^l(a_k) & i=j,\\
\displaystyle\sum_{l=1}^r\frac{\partial G^l_i}{\partial a_i}
\frac{\partial G^l_j}{\partial a_j}\prod_{\substack{k=1\\k\neq i,j}}^m G_k^l(a_k)
& i\neq j.
\end{cases}
\label{pd2}
\end{equation}
These derivatives are evaluated at $a_j([\theta])=(\theta,\varphi_j)$.
\paragraph{Alternating Least Squares (ALS) Formulation}
The development of robust and efficient algorithms to
compute \eqref{functional-SSE} to any desirable accuracy is
still a relatively open question (see
\cite{Acar,Espig1,Karlsson,DoostanIaccarino2013,HeyrimJCP_2014}
for recent progresses). Computing the tensor
components $G_k^l(a_k)$ usually relies on (greedy)
optimization techniques such as alternating least
squares (ALS) \cite{Reynolds,Battaglino,Acar,Beylkin}
or regularized Newton methods \cite{Espig1}, which are only locally
convergent \cite{Uschmajew2012} (i.e., the final result may depend
on the initial condition of the algorithm). Hereafter we describe
the simplest form of the ALS algorithm.
To this end, consider the functional residual
\eqref{functional-approx}, with $f$
given in \eqref{functional-SSE}
\begin{equation}
R([\theta])=F([\theta])-\sum_{l=1}^r
\prod_{k=1}^m G^l_k((\theta,\varphi_k)).
\label{Fresidual}
\end{equation}
We aim at computing the tensor components $G_k^l(a_k)$
by minimizing the norm of $R([\theta])$ relative to
independent variations of $G^l_k$.
In particular, if we assume that $G^l_k$ are
in the form \eqref{gfun}, then variations
of $G^l_k$ are generated by variations of $\beta_{hj}^n$.
Therefore, the canonical tensor decomposition of $F([\theta])$
can be computed by the variational principle\footnote{The Euler-Lagrange equations associated with \eqref{als} are nonlinear $\beta_{hj}^n$.}
\begin{equation}
\min_{\beta_{hj}^n}\left\|R([\theta])\right\|_W^2.
\label{als}
\end{equation}
The norm in $\|\cdot \|_W$ may be defined by a weighted
functional integral (see Appendix \ref{sec:functional integrals})
in the form
\begin{equation}
\left\| R([\theta])\right\|^2_W=\int
R([\theta])^2 W([\theta])\mathcal{D}[\theta],
\label{continuous_norm_W}
\end{equation}
where $W([\theta])$ is the functional integration measure, or by a
discrete sum (functional collocation setting)
\begin{equation}
\left\|R([\theta])\right\|^2_W=
\sum_{j=1}^N R([\theta_j])^2 w_j,
\label{discrete_norm_W}
\end{equation}
where $\{\theta_1, ..., \theta_N\}$ are collocation nodes
in the function space $D(F)$, and $w_j$ are
integration weights.
{\color{r}
In the alternating least squares paradigm,
we compute the minimizer of the residual \eqref{Fresidual}
by splitting the non-convex optimization problem \eqref{als} into
a {\em sequence} of convex low-dimensional optimization
problems (see Figure \ref{fig:ALSfigure}).
To illustrate the method, let us define the vectors
\begin{align}
\bm \beta_1&=[\beta^1_{11}, ..., \beta^1_{1Q},...,\beta^r_{11}, ..., \beta^r_{1Q}]^T,\nonumber\\
\bm \beta_2&=[\beta^1_{21}, ..., \beta^1_{2Q},...,\beta^r_{21}, ..., \beta^r_{2Q}]^T,\nonumber\\
\vdots &\nonumber\\
\bm \beta_m&=[\beta^1_{m1}, ..., \beta^1_{mQ},...,\beta^r_{m1}, ..., \beta^r_{mQ}]^T.
\end{align}
Note that $\bm \beta_i$ collects the degrees of freedom
of all functions $\{G^1_{i}, ..., G^r_i\}$ depending
on $a_i=(\theta,\varphi_i)$. Next, we split the optimization
problem \eqref{als} into the following sequence of
convex optimization problems
\begin{equation}
\min_{\bm\beta_1}\left\| R([\theta])\right\|^2_W, \qquad
\min_{\bm\beta_2}\left\| R([\theta])\right\|^2_W, \qquad\cdots, \qquad
\min_{\bm\beta_m}\left\| R([\theta])\right\|^2_W.
\label{als11}
\end{equation}
We emphasize that the system of equations
\eqref{als11} is not equivalent to the full problem \eqref{als}.
In other words, the sequence of low-dimensional
optimization problems \eqref{als11} in general
does not allow us to compute the minimizer of \eqref{als} \cite{Espig,Bezdek,Rohwedder,Uschmajew2012}.
\begin{figure}[t]
\centerline{\includegraphics[height=6cm]{optimization-eps-converted-to.pdf}}
\caption{\color{r}Sketch of the alternating least squares (ALS) algorithm for
the minimization of the functional residual $R([\theta])$.
Convergence analysis of ALS be found in
\cite{Espig,Bezdek,Rohwedder,Uschmajew2012}.}
\label{fig:ALSfigure}
\end{figure}
The Euler-Lagrange equations associated
with \eqref{als11} are in the form
\begin{equation}
\bm A_j \bm \beta_j =\bm f_j\qquad j=1,...,m,
\label{SSYSTEM}
\end{equation}
where
\begin{equation}
\bm A_j=\left[
\begin{array}{cccc}
A^{11}_{j11}\cdots A^{11}_{j1Q} & A^{12}_{j11}\cdots A^{12}_{j1Q}
& \cdots &A^{1r}_{j11}\cdots A^{1r}_{j1Q}\\
\vdots & \vdots & \vdots & \vdots \\
A^{11}_{jQ1}\cdots A^{11}_{jQQ} & A^{12}_{jQ1}\cdots A^{12}_{jQQ}
&\cdots & A^{1r}_{jQ1}\cdots A^{1r}_{jQQ}\\
\vdots & \vdots & \vdots & \vdots \\
A^{r1}_{j11}\cdots A^{r1}_{j1Q} & A^{r2}_{j11}\cdots A^{r2}_{j1Q}
&\cdots & A^{rr}_{j11}\cdots A^{rr}_{j1Q}\\
\vdots & \vdots & \vdots & \vdots \\
A^{r1}_{jQ1}\cdots A^{r1}_{jQQ} & A^{r2}_{jQ1}\cdots A^{r2}_{jQQ}
&\cdots & A^{rr}_{jQ1}\cdots A^{rr}_{jQQ}
\end{array}
\right],\qquad
\bm f_j=\left[
\begin{array}{c}
f_{j1}^1\\
\vdots\\
f_{jQ}^1\\
\vdots\\
f_{j1}^r\\
\vdots\\
f_{jQ}^r\\
\end{array}
\right],
\label{ORDERING}
\end{equation}
and
\begin{equation}
A^{ln}_{jhs}=\int\mathcal{D}[\theta]W([\theta])
\phi_h((\theta,\varphi_j))\phi_s((\theta,\varphi_j))\prod_{\substack{k=1\\k\neq j}}^m
G_k^l((\theta,\varphi_k))G_k^n((\theta,\varphi_k)),
\label{A^{ln}_{jhp}}
\end{equation}
\begin{equation}
f_{jh}^n = \int\mathcal{D}[\theta]W([\theta])F([\theta])
\phi_h((\theta,\varphi_j))\prod_{\substack{k=1\\k\neq j}}^m
G_k^n((\theta,\varphi_k)).
\label{f_{jh}^n}
\end{equation}
\noindent
The matrices $\bm A_j$ are symmetric, positive
definite and of size $rQ\times rQ$. Also, the functional integrals
defining the matrix entries can be simplified and eventually
computed by using techniques for high-dimensional integration,
such as the quasi-Monte Carlo method \cite{Dick}.
Indeed, if we restrict the residual \eqref{Fresidual}
to the finite-dimensional function space
$D_m=\textrm{span}\{\varphi_1,...,\varphi_m\}$ (Section
\ref{sec:Finite_Dim_Approx}), and assume that $F([\theta])$
restricted to $D_m$ is compactly supported within
the hypercube $[-b,b]^m$, then $A^{ln}_{jhp}$ and $f_{jh}^n$
can be written in the form
\begin{equation}
A^{ln}_{jhs}=\int_{-b}^b \phi_h(a_j) \phi_s(a_j)da_j
\prod_{\substack{k=1\\k\neq j}}^m
\int_{-b}^b G_k^l(a_k)G_k^n(a_k)da_k,
\label{Alnjhp1}
\end{equation}
\begin{equation}
f_{jh}^n = \int_{-b}^b \cdots \int_{-b}^b f(a_1,...,a_m)
\phi_h(a_j) \prod_{\substack{k=1\\k\neq j}}^m
G_k^n(a_k) da_1\cdots da_m,
\label{fh1}
\end{equation}
provided we select the integration measure $W([\theta])$
appropriately (see Appendix \ref{sec:functional integrals}).
}
\vspace{0.7cm}
\noindent
{\em Remark:}
Minimizing the residual \eqref{Fresidual}
with respect to $\beta_{hj}^n$
is equivalent to imposing orthogonality
relative to the space spanned by the functionals
\begin{equation}
\phi_h((\theta,\phi_j))\prod^m_{\substack{k=1\\k\neq j}}G^{n}_k((\theta,\varphi_k))
\end{equation}
To show this in a simple and intuitive way, consider the following
example in just two dimensions. Let $f(a_1,a_2)$ be a regular function defined on the unit square $[0,1]^2$. We look for a canonical tensor
decomposition of $f(a_1,a_2)$ in the form
\begin{equation}
f(a_1,a_2)=\sum_{j=1}^r G^{j}_1(a_1)G^{j}_2(a_2),
\quad \textrm{where}\quad G^{j}_k(a_k)=\sum_{p=1}^P \beta^j_{kp}\phi_p(a_k).
\end{equation}
To determine $\beta^j_{kp}$ we first define the residual,
\begin{equation}
R(a_1,a_2)= f(a_1,a_2)-\sum_{j=1}^r G^{j}_1(a_1)G^{j}_2(a_2)
\end{equation}
and then minimize its weighed norm
\begin{equation}
\left\| R\right\|^2_w=\int_0^1\int_0^1R(a_1,a_2)^2 w(a_1,a_2)da_1da_2
\end{equation}
relative to independent variations of $\beta^j_{1p}$ and $\beta^j_{2p}$. This yields
\begin{equation}
\begin{array}{ll}
\displaystyle \int_0^1\int_0^1 w(a_1,a_2) R(a_1,a_2)\phi_p(a_1)G_2^{j}(a_2)da_1da_2=0 &\quad \textrm{linear system for $\beta^j_{1p}$},\vspace{0.2cm}\\
\displaystyle \int_0^1\int_0^1 w(a_1,a_2) R(a_1,a_2)G_1^{j}(a_1)\phi_p(a_2)da_1da_2=0 &\quad \textrm{linear system for $\beta^j_{2p}$.}
\end{array}
\label{orthogR}
\end{equation}
Thus, minimizing the residual with respect to independent
variations of $\beta^j_{1p}$ and $\beta^j_{2p}$ is equivalent impose
Galerkin orthogonality relative to a space spanned by the basis
functions $\phi_p(a_1)G_2^{j}(a_2)$ and $G_1^{j}(a_1)\phi_p(a_2)$,
respectively.
{
\color{r}
\paragraph{Convergence of the ALS Algorithm}
The ALS algorithm we just described is
an alternating optimization scheme, i.e.,
a nonlinear block Gauss--Seidel
method (\cite{Ortega}, \S 7.4).
There is a well--developed local
convergence theory for this type of method
(see \cite{Ortega,Bezdek}).
In particular, it can be shown that ALS is locally
equivalent to the linear block Gauss--Seidel
iteration applied to the Hessian matrix.
This implies that ALS is linearly convergent in the iteration
number \cite{Uschmajew2012},
provided that the Hessian of the residual
is positive definite (except on a trivial null
space associated with the scaling non-uniqueness
of the canonical tensor decomposition). The last assumption
may not be always satisfied\footnote{It was shown in
\cite{Uschmajew2012} that the classical alternating least squares
algorithm does not converge in the iteration number for functionals
in the form \eqref{sinFunctional1}.}. Therefore, convergence
of the ALS algorithm cannot be granted in general.
Another potential issue of the ALS algorithm is
the poor conditioning of the matrices $\bm A_j$
in \eqref{SSYSTEM}, which can addressed by
regularization \cite{Reynolds,Battaglino}.
The canonical tensor decomposition
\eqref{functional-SSE} in $m$ dimensions has relatively
small memory requirements. In fact, the number of degrees of freedom
that we need to store is $r\times m \times Q$, where $r$ is
the separation rank, and $Q$ is the number of degrees of
freedom employed in each tensor component $G_k^l(a_k)$ \eqref{gfun}.
Despite the relatively low-memory requirements, it is often desirable
to employ scalable parallel versions the ALS algorithm
\cite{Karlsson} to compute the canonical
tensor expansion \eqref{functional-SSE}.
}
{
\color{r}
\subsubsection{Tucker Decomposition}
The Tucker decomposition of the cylindrical functional
\eqref{cf} is a series expansion in the form
\begin{equation}
f(a_1,...,a_m)=\sum_{l_1=1}^{r_1}\cdots
\sum_{l_m=1}^{r_m} C_{l_1\cdots l_m}
T^{l_1}_{1}(a_1)\cdots T^{l_m}_{m}(a_m),
\label{Tucker}
\end{equation}
where $C_{l_1...l_m}$ is a $r_1\times \cdots \times r_m$
real- or complex-valued tensor -- often referred to
as {\em core tensor} \cite{Kolda} --
and $T^{l_m}_{m}(a_m)$ are unknown functions.
Tucker decomposition is known
as {\em high-order Schmidt decomposition} in the
context of quantum mechanics \cite{Carteret}.
It is important to emphasize such decomposition is, in general,
are {\em not unique}\footnote{The classical Schmidt decomposition,
i.e., the bi-orthogonal decomposition of bi-variate functions is
not unique either, and defined up to two rotations in Hilbert spaces
\cite{Venturi2,Peres}}.
As pointed out by Kolda and Bader in \cite{Kolda},
this freedom opens the possibility to choosing transformations
that simplify the core structure in some way so that most
of the elements of the core tensor $C_{l_1\cdots l_m}$ are
zero, thereby eliminating interactions between corresponding
components. Diagonalization of the core is, in general,
impossible\footnote{The canonical tensor decompositon
\eqref{functional-SSE} is in the form of a fully diagonal
high-order Schmidt decomposition, i.e,
\begin{equation}
f(a_1,...,a_m)=\sum_{l=1}^{r} C_{l\cdots l}
G^{l}_{1}(a_1)\cdots G^{l}_{m}(a_m).
\end{equation}
The fact that diagonalization of $C_{l_1\cdots l_m}$ is, in general,
impossible in dimension larger than 2 implies that it is impossible
to compute canonical tensor decompositions by standard
linear algebra techniques. Indeed, the best
low-rank approximation problem is {\em ill-posed} for real
tensors with dimension $m>2$ (see, e.g., \cite{Silva,Kolda,Hillar}),
and for complex tensors \cite{Vannieuwenhoven}.}
\cite{Peres}, but it is possible to
try to make as many elements either zero or as small
as possible (see, e.g., \cite{Carteret} or \cite{Moravitz}).
For general tensors we have that the multilinear rank
$r=r_j$ ($j=1,...,m$) is upper semi-continuous, i.e.,
the Tucker expansion is closed.
Several efficient algorithms are currently available to compute
the series expansion \eqref{Tucker}.
For instance, Lathauwer {\em et al.}
proposed in \cite{Lathauwer} a high-order singular value
decomposition method to determine the
components $T_k^{l_k}$ and the core tensor
$C_{l_1\cdots l_m}$ in a discrete setting.
Such algorithm is simple, robust and it yields
{\em quasi-optimal} low-rank
approximations.
To illustrate the procedure to compute the Tucker decomposition
let us first assume that the basis functions $T_k^{l_k}(a_k)$
in \eqref{Tucker} are {\em orthonormal} and {\em known}.
In this case, the expansion \eqref{Tucker}
is simply a tensor product representation of a
multivariate function relative to the basis $T_k^{l_k}$.
The core tensor $C_{l_1\cdots l_m}$ can be immediately
obtained by projecting $f(a_1,...,a_m)$ onto the
basis $T_k^{l_k}(a_k)$, i.e.,
\begin{equation}
C_{l_1\cdots l_m}=\int f(a_1([\theta]),...,a_m([\theta]))
T_1^{l_1}(a_1([\theta]))
\cdots T_m^{l_m}(a_m([\theta]))W([\theta]) \mathcal{D}([\theta]).
\label{FI9}
\end{equation}
Evaluating the functional integral
in $D_m=\textrm{span}\{\varphi_1,...,\varphi_m\}$
and rescaling the integration measure $W([\theta])$
properly (see Appendix \ref{sec:functional integrals})
allows us to rewrite \eqref{FI9}
\begin{equation}
C_{l_1\cdots l_m}=\int_{-b}^b \cdots \int_{-b}^b f(a_1,...,a_m)
T_1^{l_1}(a_1)\cdots T_m^{l_m}(x_m)da_1\cdots da_m,
\end{equation}
where we assumed $f$ to be compactly supported in $[-b,b]^m$.
Next, suppose that each function $T_k^{l_k}(a_k)$ is a
linear combination of $Q$ orthonormal basis functions
$\phi_j$, i.e.,
\begin{equation}
T_k^{l_k}(a_k)=\sum_{s=1}^Q \alpha_{sk}^{l_k} \phi_{s}(a_k)
\qquad \textrm{(fixed $k$).}
\end{equation}
If $\phi_{s}$ is a {\em cardinal basis} associated with
a set of interpolation nodes along $a_k$, then the matrix
$\alpha_{sk}^{l_k}$ represents the set of
functions $T_k^{l_k}(a_k)$ for fixed $k$. We can sort arrange
the matrix $\alpha_{sk}^{l_k}$ in a way where
the $l_k$-column represents the value
of $T_k^{l}(a_k)$ at $Q$ collocation nodes
along $a_k$. This yields the matrix with entries $[T_k]_{il}$
($i=1,...,Q$). If we evaluate the multivariate field $f(a_1,..,a_m)$
at the same collocation nodes we employed to
construct the interpolants of $T_k^{l_k}$, then we can
rewrite \eqref{Tucker} in a full tensor notation as
\begin{equation}
f_{i_1\cdots i_m}=\sum_{l_1=1}^{r_1}\cdots
\sum_{l_m=1}^{r_m} C_{l_1,...,l_m}
[T_1]_{i_1l_1 }\cdots [T_m]_{i_ml_m },
\label{Tucker_matrix}
\end{equation}
where $i_j=1,...,Q$ are indices identifying the interpolation node
along the axis $a_j$.
\vspace{0.7cm}
\noindent
{\em Remark:}
The expansion \eqref{Tucker_matrix}
is a ``matricization'' of continuum series \eqref{Tucker} obtained by
representing each basis element in terms of collocation nodes.
Clearly, we can also set up a matricization of \eqref{Tucker}
based on Galerkin projection. To this end, it is sufficient
to project both the left and the right hand sides
of \eqref{Tucker} onto
the orthonormal basis elements
$\phi_{i_1}(a_1)\phi_{i_2}(a_2)\cdots \phi_{i_m}(a_m)$
to obtain an expression in the form \eqref{Tucker_matrix}.
The meaning of $f_{i_1 \cdots i_n}$ in the case is the
Fourier coefficients of the projection, i.e.,
\begin{equation}
f_{i_1 \cdots i_n}=\int_{-b}^b\cdots\int_{-b}^b f(a_1,...,a_m)
\phi_{i_1}(a_1)\cdots \phi_{i_m}(a_m)dx_1\cdots dx_m.
\end{equation}
Thus, \eqref{Tucker_matrix} represents a multivariate
expansion of Fourier coefficients in a Tucker tensor format.
In such finite-dimensional setting, we basically transformed
the problem of decomposing a multivariate function to
a {\em multi-linear algebra problem}. It is immediate to
see that the discrete Tucker format
a two-dimensional function $f(a_1,a_2)$ is
\begin{equation}
f_{i_1i_2}=\sum_{l_1=1}^{r_1}\sum_{l_2=1}^{r_2} C_{l_1l_2}
[T_1]_{i_1l_1 }[T_2]_{i_2l_2 },
\end{equation}
We can obviously diagonalize the core tensor by using singular value
decomposition.
\vspace{0.7cm}
\noindent
A drawback of the Tucker decomposition is the storage
requirement of the core tensor $C_{l_1\cdots l_m}$, which is $O(r^m)$.
Such problem can be
mitigated by attempting to make zero as many entries of
$C_{l_1\cdots l_m}$ as possible through suitable linear
transformations. Another option is to introduce
further separability properties of the core tensor.
This yields a multitude of possible
expansions, including hierarchical Tucker \cite{Grasedyck,Bachmayr} and
Tucker tensor train \cite{Nouy2017,Rohwedder}.
All these series expansions can be conveniently visualized
by suitable graphs, and as such they fall within the setting
of {\em tensor-networks}.
\subsubsection{Hierarchical Tucker Decomposition}
\label{sec:HT}
We have seen that the canonical tensor
decomposition of a nonlinear functional
has a relatively small number of degrees of freedom,
but its computation is not straightforward.
The most reliable algorithms are based
on regularized optimization and randomized methods
(see \cite{Reynolds,Battaglino} for recent results),
which often yield different series expansions for
different initial conditions.
On the other hand, we have seen the that high-order
Schmidt decomposition suffers from the curse of dimensionality
(dimension of the core tensor), but it is straightforward
to compute by {\em nearly optimal} and {\em robust}
techniques such as high-order singular value
decomposition \cite{Lathauwer}.
The hierarchical Tucker format was introduced
by Hackbush and K\"{u}hn in \cite{Hackbusch2009} (see also \cite{Hackbusch_book,Grasedyck}) to mitigate
the dimensionality problem in the core-tensor of
the classical high-order Schmidt decomposition.
The key idea is to perform a sequence
of classical bi-variate Schmidt decompositions until the
approximation problem is reduced to a product
of one-dimensional functions. To illustrate the method
is a simple way, consider the following cylindrical functional
\begin{equation}
f(a_1,a_2,a_3)=\sum_{i,j=1}^r A^{\{1\}}_{ij}T^{\{1\}}_i(a_1) T^{\{2,3\}}_j(a_2,a_3), \qquad a_j=(\theta,\varphi_j).
\end{equation}
The matrix representation of $f(a_1,a_2,a_3)$ (either Galerkin or collocation form) relative to the given
bases $T^{\{1\}}_i$ and $T^{\{2,3\}}_j$
is $A_{ij}^{\{1\}}$, i.e., it has two indices\footnote{The
second index labels points on the plane $(a_2,a_3)$
(collocation setting), or the projection of $f$ onto the two-dimensional basis function $T_j^{2,3}(a_2,a_3)$ (Galerkin setting).}. We decompose $T^{\{2,3\}}_j$ further by another Schmidt expansion
\begin{equation}
T^{\{2,3\}}_j(a_2,a_3)=\sum_{l,n=1}^r A^{\{2\}}_{jln}
T^{\{2\}}_l(a_2)T^{\{3\}}_n(a_3),
\end{equation}
to obtain
\begin{equation}
f(a_1,a_2,a_3)=\sum_{i,j,l,n=1}^r A^{\{2\}}_{jln}
A^{\{1\}}_{ij}T^{\{1\}}_i(x_1)T^{\{2\}}_l(a_2)T^{\{3\}}_n(a_3).
\end{equation}
The procedure just described is at the basis of the hierarchical
Tucker decomposition and it yields an expansion in the
form \eqref{Tucker}, where the core tensor is factored
as a {\em product of at most three-dimensional tensors}.
To show this in a higher-dimensional case, consider the
six-dimensional cylindrical functional $f(a_1,...,a_6)$.
By performing a sequence of Hilbert-Schmidt
factorizations, while keeping the separation
rank $r$ constant in each factorization, we obtain
\begin{align}
f(a_1,...,a_6)= &\sum_{i_7,i_8=1}^r A^{\{1\}}_{i_7i_8} T^{\{1,2,3\}}_{i_7}
(a_1,a_2,a_3)T^{\{4,5,6\}}_{i_8}(a_4,a_5,a_6),\nonumber\\
=&\sum_{i_7,i_8=1}^r A^{\{1\}}_{i_7i_8}
\sum_{i_1,i_9=1}^r A^{\{2\}}_{i_7i_1i_9}
T^{\{1\}}_{i_1}(a_1)
T^{\{2,3\}}_{i_9}(a_2,a_3)
\sum_{i_4,i_{10}=1}^r A^{\{3\}}_{i_8 i_4i_{10}}
T^{\{4\}}_{i_4}(a_4)T^{\{5,6\}}_{i_{10}}(a_5,a_6),\nonumber\\
&=\sum_{i_1,\cdots, i_{6}=1}^r
C_{i_1\cdots i_6}
T^{\{1\}}_{i_1}(a_1)T^{\{2\}}_{i_2}(a_2)T^{\{3\}}_{i_3}(a_3)
T^{\{4\}}_{i_4}(a_4)T^{\{5\}}_{i_5}(a_5)
T^{\{6\}}_{i_{6}}(a_6),
\label{HTD}
\end{align}
%
where the $6$-dimensional core tensor is explicitly given
\begin{equation}
C_{i_1\cdots i_{6}} =
\sum_{i_7,i_8=1}^r A^{\{1\}}_{i_7i_8}
\sum_{i_9=1}^r A^{\{2\}}_{i_7i_1i_9} A^{\{4\}}_{i_9i_2i_3}
\sum_{i_{10}=1}^r A^{\{3\}}_{i_8 i_4i_{10}}A^{\{5\}}_{i_{10}i_5i_6}.
\end{equation}
Note that such kernel has a recursive structure and it is factored
as a product of at most three-dimensional matrices.
Parallel algorithms to perform basic operations on
hierarchical Tucker expansions such as addition, multiplication,
and rank reduction were recently proposed in
\cite{Grasedyck2017,Etter}. Also, an optimization
framework that leverages on the on recursive subspace
factorizations of Hierarchical Tucker expansions was
developed in \cite{DaSilva2015}.
\subsubsection{Tensor Train Decomposition}
If we single-out one variable at a time and perform a
sequential Schmidt decomposition of the remaining
variables we obtain the so-called {\em tensor-train}
decomposition \cite{OseledetsTT,Rohwedder}.
Tensor train decomposition is a subcase case of hierarchical
Tucker decomposition. For example, the tensor train
decomposition of the four-dimensional
cylindrical functional $f(a_1,...,a_4)$ reads
\begin{align}
f(a_1,...,a_4)= &\sum_{i1,i_2=1}^r A^{\{1\}}_{i_1i_2}
T^{\{1\}}_{i_1}(a_1)T^{\{2,3,4\}}_{i_2}(a_2,a_3,a_4),\nonumber\\
=&\sum_{i_1,i_2=1}^r A^{\{1\}}_{i_1i_2}
T^{\{1\}}_{i_1}(a_1) \sum_{i_3,i_4=1}^r A^{\{2\}}_{i_2i_3i_4}
T^{\{2\}}_{i_3}(a_2) T^{\{3,4\}}_{i_4}(a_3,a_4),\nonumber \\
=&\sum_{i_1,...,i_4=1}^r A^{\{1\}}_{i_1i_2}A^{\{2\}}_{i_2i_3i_4}
T^{\{1\}}_{i_1}(a_1) T^{\{2\}}_{i_3}(a_2)
\sum_{i_5,i_6}^rA^{\{3\}}_{i_4i_5i_6} T^{\{3\}}_{i_5}(a_3)
T^{\{4\}}_{i_6}(a_4),\nonumber \\
=&\sum_{i_1,...,i_6=1}^r A^{\{1\}}_{i_1i_2}A^{\{2\}}_{i_2i_3i_4}
A^{\{3\}}_{i_4i_5i_6}
T^{\{1\}}_{i_1}(a_1) T^{\{2\}}_{i_3}(a_2) T^{\{3\}}_{i_5}(a_3) T^{\{4\}}_{i_6}(a_4).
\label{TTD}
\end{align}
%
Tensor train and hierarchical Tucker expansions
can be conveniently visualized by {\em graphs}.
\subsubsection{Tensor Networks}
During the last decades, the field of Tensor Networks has
undergone an explosion of results in several different directions,
e.g., in the study of quantum many-body systems, and more generally in multivariate functional approximation. Tensor networks
can be conveniently represented by {\em undirected graphs}.
To this end, we adopt the following standard convention:
\begin{itemize}
\item a node in a graph represents a tensor
in as many variables as the number of the
edges connected to it;
\item connecting two tensors by an edge represents
a tensor contraction over the index associated to certain variable.
\end{itemize}
\noindent
In this way, a three-dimensional Tucker format is represented by one
node with three edges emanating from it. If we connect two of such
tensors with one edge, then we obtain a tensor in four variables. In particular, the connection operation here results in a
Tucker format with the core tensor
represented by a product of two three-dimensional
tensors in which we contract one
index (see Figure \ref{Fig:TN}).
\begin{figure}[t]
\centerline{\hspace{-1.2cm}Tucker Format\hspace{6.cm} Tensor Network}
\centerline{\hspace{-1cm}
\includegraphics[width=3.5cm]{tucker_graph-eps-converted-to.pdf}
\hspace{4.2cm}
\includegraphics[width=4.5cm]{connected_tucker_graph-eps-converted-to.pdf}
}
{\footnotesize $\displaystyle \sum_{i_1,i_2,i_3=1}^r
A_{i_1i_2i_3}T^{\{1\}}_{i_1}(a_1) T^{\{2\}}_{i_2}(a_2) T^{\{3\}}_{i_3}(a_3)$}\hspace{1.5cm}
{\footnotesize $\displaystyle \sum_{i_1, ... ,i_4=1}^r
\left(\sum_{j=1}^rA^{\{1\}}_{i_1i_2j}A^{\{2\}}_{ji_3i_4}\right)T^{\{1\}}_{i_1}(a_1) T^{\{2\}}_{i_2}(a_2) T^{\{3\}}_{i_3}(a_3)T^{\{4\}}_{i_4}(a_4)$
}
\caption{\color{r}Examples of tensor networks representing a
cylindrical functional in three and four variables. }
\label{Fig:TN}
\end{figure}
The graph representation of the hierarchical Tucker
and tensor train decomposition is shown in
Figure \ref{fig:HT_TT}.
\begin{figure}[ht]
\centerline{\hspace{-1.2cm}Hierarchical Tucker\hspace{4.2cm} Tensor Train}
\centerline{
\includegraphics[width=6.5cm]{hierarchical_tucker-eps-converted-to.pdf}
\hspace{1cm}
\includegraphics[width=6cm]{tensor_train_6D-eps-converted-to.pdf}
}
\caption{\color{r}Graph representation of the Hierarchical Tucker (HT) and tensor train (TT) decomposition of a six-dimensional cylindrical functional.}
\label{fig:HT_TT}
\end{figure}
A more general tensor network representation of
a cylindrical functional in five variables is shown in Figure
\ref{Fig:TN1}.
\begin{figure}[t]
\centerline{\hspace{0.5cm}Tensor Network\hspace{3.cm}
Equivalent Tucker Format}
\centerline{\hspace{-2cm}
\includegraphics[width=6.0cm]{tensor_network-eps-converted-to.pdf}
\hspace{2cm}
\includegraphics[width=3.0cm]{tucker_graph_5-eps-converted-to.pdf}}
\caption{\color{r}Tensor network representation of a five-dimensional cylindrical functional. After reduction (elimination of the cycle and connection of nodes), it is easy to see the graph is equivalent to
the graph corresponding to a Tucker format of dimension five (right).
The core tensor of such Tucker format is expressed as products and contractions of at most three-dimensional tensors.
}
\label{Fig:TN1}
\end{figure}
Note that each node is connected with at most
three edges (coordination number equal to three)
as we employed a Schmidt decomposition to represent
each function with more than one variable. Clearly,
the graph can be reduced to one
node with five leaves (i.e. a Tucker series)
by eliminating the cycle and clustering all nodes that are connected
by edges \cite{Kressner2014} .
The graph can be therefore reduced to a node
with five leaves -- i.e. a Tucker series, with a particular
structure of the core tensor. At this point the
graph is irreducible (one node with five leaves).
Efficient algorithms that implement basic operations
between tensors, such as addition, orthogonalization, rank
reduction, scalar products, multiplication, and linear
transformations are discussed in
\cite{Kressner2014,Etter,Grasedyck2015,Nouy2017,Kolda}.
}
\subsection{Generalized Lagrangian Interpolation}
\label{sec:gen_Lagr}
We have seen that the classical Lagrange interpolation
problem for multivariate functions can be generalized
to functionals in Hilbert or Banach spaces, i.e.,
objects depending on an infinite number of variables.
Given a real-valued continuous functional
\begin{equation}
F:D(F)\rightarrow \mathbb{R},
\end{equation}
and a set of elements $\{\theta_1(x),...,\theta_n(x)\}$ in $D(F)$,
the Lagrangian interpolant of $F$ is a continuous functional
\begin{equation}
\Pi: D(F) \rightarrow \mathbb{R},
\end{equation}
such that
\begin{equation}
\Pi ([\theta_i])=F([\theta_i]).
\label{interpolation problem}
\end{equation}
A simple representation of $\Pi([\theta])$
can be given in terms of {\em cardinal basis functionals}
$g_i:D(F)\rightarrow \mathbb{R}$ as
\begin{equation}
\Pi ([\theta])=\sum_{i=1}^n F([\theta_i])g_i([\theta]), \quad
\textrm{where}\quad g_i([\theta_j])=\delta_{ij}.
\label{interpolant}
\end{equation}
Similarly to classical Lagrange interpolation for
functions from $\mathbb{R}^n$ to $\mathbb{R}^m$, we
have a great freedom in selecting $g_i([\theta])$. We have
already seen, for example, Porter's and Prenter's cardinal
bases defined in equation \eqref{gi_0} and equation \eqref{gi_prenterer}, respectively.
Now, let
\begin{equation}
\kappa:D(F)\times D(F)\rightarrow\mathbb{R}
\end{equation}
be any continuous (not necessarily symmetric) functional subject to the sole
requirement that
\begin{equation}
\kappa([\theta],[\eta])=0\quad \Leftrightarrow \quad \theta(x)=\eta(x).
\label{cond1}
\end{equation}
Examples of such functional are
\begin{align}
\kappa_1([\theta],[\eta])&=\left\|\theta(x)-\eta(x)\right\|^{p},
\qquad p>0\label{positiveker}\\
\kappa_2([\theta],[\eta])&= \left\|\theta(x)-\eta(x)\right\|+
\left\|\theta(x)\right\|-\left\|\eta(x)\right\|.\\
\kappa_3([\theta],[\eta])&= 1- \exp\left(-\left\|\theta(x)-\eta(x)\right\|^2\right).
\end{align}
By using $\kappa([\theta],[\eta])$ we define the following
cardinal basis associated with the interpolation nodes
$\{\theta_1(x),...,\theta_n(x)\}$ in $D(F)$
(see \cite{Cheney}, Ch. 10)
\begin{equation}
g_i([\theta]) = \prod_{\substack{j=1\\j\neq i}}^n
\frac{\kappa([\theta],[\theta_j])}{\kappa([\theta_i],[\theta_j])}
\qquad i=1,...,n.\label{cardinal_cheeney}
\end{equation}
A substitution of \eqref{cardinal_cheeney} into
\eqref{interpolant} yields a {\em Lagrangian functional interpolant} that depends on the choice of $\kappa([\theta],[\eta])$.
An advantage of such interpolant over Porter's interpolant
is that the basis functionals \eqref{cardinal_cheeney}
are given analytically and do not require
any computation such as the inversion (and storage) of
the $H$-matrix \eqref{matH}. On the other hand, we have no
guarantee that \eqref{cardinal_cheeney} have good
approximation properties away from the interpolation nodes.
The Lagrange interpolation formula \eqref{interpolant}-\eqref{cardinal_cheeney}
can be written in a {\em barycentric form}.
To this end, we first rewrite the numerator
of \eqref{cardinal_cheeney} as
$\Omega([\theta])/\kappa([\theta],[\theta_i])$, where
\begin{equation}
\Omega([\theta])=\prod_{i=k}^n\kappa([\theta],[\theta_k]).
\end{equation}
Then define
the {\em barycentric weights} as
\begin{equation}
w_i = \prod_{\substack{j=1\\j\neq i}}^n
\frac{1}{\kappa([\theta_i],[\theta_j])}.\label{weights}
\end{equation}
Clearly,
\begin{equation}
g_i([\theta])=\Omega([\theta])\frac{w_i}{\kappa([\theta],[\theta_i])}.
\end{equation}
Substituting this into \eqref{interpolant} and applying it
to the constant functional $F([\theta])=1$ yields
\begin{equation}
\Omega([\theta])=\frac{1}{\displaystyle\sum_{i=1}^n\frac{w_i}{\kappa([\theta],[\theta_i])}}.
\end{equation}
Therefore, we obtain the interpolant
\begin{equation}
\Pi([\theta])=
\left(\sum_{i=1}^n\frac{w_i}{\kappa([\theta],[\theta_i])}\right)^{-1}
\sum_{i=1}^n \frac{w_iF([\theta_i])}{\kappa([\theta],[\theta_i]).}
\end{equation}
This expression can be evaluated in $\mathcal{O}(n)$ operations, provided we have available
$w_i$, which needs $\mathcal{O}(n^2)$ operations.
What about updating? Adding a new interpolation node $\theta_{n+1}$ entails two
calculations: 1) divide each $w_j$ by $\kappa([\theta_j],[\theta_{n+1}])$ and
2) compute $w_{n+1}$ using the formula \eqref{weights}.
\vspace{0.7cm}
\noindent
{\em Remark:} The interpolation processes defined by \eqref{interpolant}
and \eqref{cardinal_cheeney} has a variation known as
Shepard's method \cite{Farwig,Lazzaro}. Such method was studied by
Allasia and Bracco \cite{Allasia} for functional interpolation
in Banach spaces and it relies on the cardinal functionals
\begin{equation}
g_i([\theta]) = \frac{1}{\displaystyle1+\sum_{\substack{ k=1\\k\neq i}}^n
\frac{{\kappa}([\theta],[\theta_i])}{{\kappa}
([\theta],[\theta_k])}},
\qquad i=1,...,n.\label{Shepardb}
\end{equation}
The kernel ${\kappa}([\theta],[\eta])$ here satisfies \eqref{cond1} and
the additional positivity requirement
\begin{equation}
{\kappa}([\theta],[\eta])>0\qquad \textrm{if} \qquad
\theta(x)\neq\eta(x).
\label{cond2}
\end{equation}
This implies that the basis functionals \eqref{Shepardb}
have the following properties
\begin{equation}
0\leq g_i([\theta])\leq 1\quad \forall \theta\qquad
g_i([\theta_j])=\delta_{ij}.
\end{equation}
One unfortunate consequence of these conditions is that
the functional derivative of $g_i([\theta])$ (if it exists)
is identically zero at all nodes $\theta_j$.
In fact, $g_i([\theta])$ is always positive, it has
a maximum (equal to 1) at $\theta_i$ and
many minima (equal to 0) at all other nodes.
This yields functional differentiation matrices \eqref{Hint}
that are identically zero.
Functional Lagrangian interpolation can also be defined in terms of
{\em radial basis functionals}
$\kappa(\left\|\theta-\theta_j\right\|)$,
for suitable choices of $\kappa$.
In this cases, the cardinal basis $g_i([\theta])$ can be obtained
by standard techniques such as the ratio of two Vandermonde
determinants. Interpolation using radial
basis functionals can converge towards polynomial functional
interpolants if we take increasingly flat radial bases
(see, e.g., \cite{Schaback,Driscoll,Song}).
\vspace{0.7cm}
\noindent
The Lagrangian interpolation process defined
by \eqref{interpolant}, \eqref{cond1} and
\eqref{cardinal_cheeney} can be generalized further
by considering a family of continuous functionals
\begin{equation}
\kappa_i:D(F)\times D(F)\rightarrow\mathbb{R},\qquad i=1,...,n
\end{equation}
as many as the number of interpolation nodes, satisfying
$\kappa_i([\theta],[\theta_j])\neq 0$ for $i\neq j$,
$\kappa_i([\theta],[\theta_i])=0$ and
$\kappa_i([\theta],[\theta])=0$.
The cardinal basis corresponding
to $\{\kappa_1,...,\kappa_n\}$ is
\begin{equation}
g_i([\theta])=\prod_{\substack{j=1\\j\neq i}}^n \frac{\kappa_i([\theta],[\theta_j])}
{\kappa_i([\theta_i],[\theta_j])}.
\end{equation}
In particular, the choice
\begin{equation}
\kappa_i([\theta],[\eta])=(\theta-\eta,\theta_i-\eta)
\label{kprenter}
\end{equation}
yields Prenter's cardinal basis \eqref{gi_prenterer}. It fact,
\begin{align}
\kappa_i([\theta],[\theta_j]) &=(\theta-\theta_j,\theta_i-\theta_j),\label{kk1}\\
\kappa_i([\theta_i],[\theta_j])&=\left\|\theta_i-\theta_j\right\|^2,\label{kk2}\\
\kappa_i([\theta],[\theta_i]) &=0.\label{kk3}
\end{align}
\noindent
Another generalization of the Lagrange interpolation process
can be obtained along the lines of Porter's functional interpolants. To this end,
let $\kappa([\theta],[\eta])$ be any symmetric functional
such that the matrix
\begin{equation}
H_{ij}=\kappa([\theta_i],[\theta_j])
\label{kH}
\end{equation}
is invertible for every set of distinct nodes $\{\theta_j(x)\}$.
Then the functional
\begin{equation}
\Phi_n([\theta])=\sum_{k=1}^m F(\theta_{k})\underbrace{\sum_{j=1}^m
\kappa([\theta_k],[\theta_j])^{-1}\kappa([\theta],[\theta_j])}_{g_k([\theta])},
\end{equation}
where $\kappa([\theta_k],[\theta_j])^{-1}$ denotes the inverse matrix of
\eqref{kH}, interpolates $F([\theta])$ at $\theta_k(x)$ ($k=1,...,m$).
The question of how to select the symmetric functional $\kappa$
depends on the requirements we impose on the behavior of the functional interpolant away from the interpolation nodes.
In particular,
the choice
\begin{equation}
\kappa([\theta],[\eta])=\sum_{p=0}^n (\theta,\eta)^p
\end{equation}
yields Porter's interpolants, i.e., polynomial functionals of
order $n$ with minimal norm.
\subsubsection{Optimal Interpolation Nodes}
\label{sec:optimal interpolation nodes}
In this Section we briefly address the question
of how to select the interpolation nodes in the function space
$D(F)$ in such a way that the corresponding
functional interpolant, e.g., the Porter's one
\eqref{Porter_interpolant}-\eqref{gi_0},
exhibits good approximation properties.
\paragraph{Adaptive Leja Sequences}
An effective approach to select sub-optimal interpolation nodes
relies on adaptive Leja sequences \cite{Akil,Bos}.
In this setting, given the
set of nodes $\{\theta_1(x),...,\theta_n(x)\}$ in $D(F)$,
the problem is to find a new node $\theta^*(x)$
satisfying the maximization principle
\begin{equation}
\argmax_{\theta^*(x)\in D(F)}\left|\det ( V)\right|
\label{determinant_VDM}
\end{equation}
where $V$ is a Vandermonde-like operator
(infinite-dimensional matrix) (see \cite{Bos,Akil,Akil1,Marchi}
for further details).
This results in the so-called {\em Leja construction} of the
test function $\theta^*(x)$, a greedy version of the well-known
Fekete construction involving nonlinear optimization.
There are also non-determinant versions
of multivariate Leja interpolation nodes, which are related
to {\em potential theory} \cite{Garcia,Ishizaka}.
In this setting, given a symmetric functional
$\kappa([\theta(x)],[\eta(x)])$ one can construct a
{\em greedy $\kappa$-energy sequence} in which
$\theta^*(x)$ is chosen to satisfy
\begin{equation}
\sum_{i=1}^{n}\kappa([\theta^*(x)],[\theta_i(x)])=
\argmin_{\theta^*\in D(F)}\sum_{i=1}^{n} \kappa([\theta^*(x)],[\theta_i(x)]),
\end{equation}
assuming $\{\theta_1(x),...,\theta_n(x)\}$ are available.
The element $\theta^*(x)$ obtained in this way is not unique.
Also, depending on the choice of the kernel
function $\kappa([\theta(x)],[\eta(x)])$, we can
have different sets of nodes. A relevant kernel for
the $\kappa$-energy sequence is the so-called {\em Riesz $p$-kernel}
\begin{equation}
\kappa_p([\theta(x)],[\eta(x)])=
\begin{cases}
\displaystyle \frac{1}{\left\|\theta(x)-\eta(x)\right\|^p} & \textrm{if $p>0$}\\
\displaystyle -\ln(\left\|\theta(x)-\eta(x)\right\|) & \textrm{if $p=0$}\\
\end{cases}
\label{Rieszp}
\end{equation}
Such kernel avoids placing points too close
to each other (the potential $\kappa_p$ is rather large in such regions),
while decreasing monotonically at larger distances.
The asymptotic behavior of greedy $\kappa$-energy sequences has
been studied theoretically by L\'opez-Garc\'ia and Saff \cite{Garcia},
in a finite-dimensional setting.
\paragraph{Minimization of Lebesgue Functionals} Another approach
to determine sub-optimal nodes in the function space $D(F)$
relies on greedy algorithms minimizing {\em Lebesgue-like functionals}.
This problem has been recently addressed in a
finite-dimensional setting by Narayan and Xiu \cite{Akil1,Akil2},
Maday {\em et al.} \cite{Maday,Maday1} and
Van Barel {\em et. al} \cite{Barel}.
The basic idea is simple and can be generalized to
the infinite-dimensional case. Suppose we have available
$\{\theta_1(x)$, ..., $\theta_{n}(x)\}$. We look for a new
node $\theta^{*}(x)$ maximizing a suitable objective
functional, e.g., related to the well-known Lebesgue function
in finite-dimensional interpolation problems.
In particular, following the weighed approach of Narayan
and Xiu \cite{Akil1}, we can look for a new node $\theta^{*}(x)$
satisfying the following (greedy) optimization problem
\begin{equation}
\argmax_{\theta^{*}\in D(F)} \chi([\theta^*]),\qquad
\chi([\theta])=W([\theta])^2\sum_{k=1}^n
\frac{g_k([\theta])^2}{W([\theta_k])^2},
\label{lebesgue function}
\end{equation}
where $W([\theta])$ is a positive
functional and $g_i([\theta])$ is a cardinal basis.
Other choices of $\chi([\theta])$ yield different
sets of nodes. For example,
\begin{equation}
\chi([\theta])=W([\theta])\sum_{k=1}^n
\frac{g_i([\theta])^2}{W([\theta_k])}
\label{fejerOPT}
\end{equation}
yields an infinite-dimensional extension of the Fej\'er points.
The sequence of nodes $\theta_i(x)$ we obtain in this
way {\em strongly depends on the initial set of nodes},
on the weight $W([\theta])$ and
on the cardinal basis $\{g_i(\left[\theta\right])\}$.
Finding the maximum of \eqref{lebesgue function} involves
computing the solution to an optimization
problem in infinite dimensions \cite{Bensoussan_DaPrato}.
A necessary condition for a stationary point
of \eqref{lebesgue function} is
\begin{equation}
\frac{\delta \chi([\theta]) }{\delta \theta(x)}=0.
\label{Euler-Lagrange}
\end{equation}
\vspace{0.7cm}
\noindent
{\em Example 1:} The Euler-Lagrange equation \eqref{Euler-Lagrange}
can be written down explicitly, e.g., for Porter's
interpolants. To this end, simply substitute \eqref{gi_0} in \eqref{fejerOPT} and set $W=1$ to obtain
\begin{equation}
\chi([\theta]) = \sum_{k=1}^n
\sum_{j,z=1}^m H_{jk}^{-1}H_{zk}^{-1} \sum_{p,q=0}^h\left(\theta_j,\theta\right)^p
\left(\theta_z,\theta\right)^q.
\end{equation}
This allows us to write explicitly the conditions identifying
stationary points of $\chi([\theta])$, e.g., in the
function space
\begin{equation}
D_N=\left\{\theta(x)\in D(F)\,\left|
\, \theta(x)=\sum_{k=1}^{N}a_k\varphi_k(x)\right.\right\}.
\label{Gqsig}
\end{equation}
\subsection{High-Dimensional Model Representation}
\label{sec:HDMR}
The high-dimensional model representation (HDMR) \cite{Rabitz,Li,Li1,CaoCG09} of the cylindrical functional \eqref{functional-approx} is a series
expansion in the form
\begin{equation}
f(a_1, ..., a_m)=f_0+
\sum_{k=1}^m f_k(a_k) +
\sum_{\substack{k,j=1\\k< j}}^m
f_{kj}(a_k,a_j)+\cdots.
\label{functional-HDMR}
\end{equation}
The functional $f_{0}$ is simply a constant.
The functionals $f_i(a_i)$, which we shall call
first-order interaction terms, give us the
overall effects of the variables $a_i=(\theta,\varphi_i)$ in $f$
as if they were acting independently of the other variables.
The functions $f_{ij}(a_i, a_j)$ represent the interaction effects
of $a_i$ and $a_j$, and therefore they are usually
called second-order interactions.
Similarly, higher-order terms reflect the cooperative
effects of an increasing number of variables.
The interaction terms in the HDMR expansion can be
easily computed if we assume that the domain of
$f$ is compactly supported and included in the hypercube
$[-b,b]^m$ (see Appendix \ref{sec:functional integrals}).
By introducing the finite-dimensional integration
measure $w(\bm a)=w(a_1,...,a_m)$ and the vector
notation $d\bm a_{-i}=da_{i-1}da_{i+1}\cdots da_m$,
we have
\begin{align}
f_0 &= \int_{[-b,b]^m} w(\bm a) f(\bm a) d\bm a,
\label{anova0}\\
f_i(a_i) &= \int_{[-b,b]^{m-1}} w(\bm a) f(\bm a) d\bm a_{-i} -f_0,
\label{anova1}\\
f_{ij}(a_i,a_j) &= \int_{[-b,b]^{m-2}} w(\bm a) f(\bm a) d\bm a_{-ij}-f_i(a_i)-f_j(a_j) -f_0,
\label{anova2}\\
&\cdots\nonumber .
\end{align}
This procedure generates, by construction,
terms that are orthogonal in the weighted $L_2$
sense. The HDMR series \eqref{functional-HDMR} with
components \eqref{anova0}-\eqref{anova2}
is often called ANOVA-HDMR expansion.
If we consider a Dirac delta measure
$w(\bm a)=\delta(a_1-c_1)\cdots \delta(a_m-c_m)$
with ``anchor point'' $\bm c=(c_1,...,c_m)$, then the
HDMR series is called CUT-HDMR \cite{Dick1,Handy_ANOVA_2010}.
There is also a random-sampling version
of HDMR -- namely the RS-HDMR \cite{Li1} expansion -- in which
the high-dimensional integrals in \eqref{anova0}-\eqref{anova1}
are computed by Monte-Carlo or quasi-Monte Carlo methods.
The HDMR expansion \eqref{functional-HDMR} is usually
truncated at some interaction order, and the interactions
$f_i$, $f_{ij}$ are expanded relative to a certain basis.
For example, the first-order ANOVA-HDMR expansion of
$f$ reads
\begin{equation}
f(a_1,...,a_m)\simeq f_0+
\sum_{k=1}^m f_k(a_k),
\label{functional-HDMR-1}
\end{equation}
where
\begin{equation}
f_k(a_k)=\sum_{i=1}^Q
\alpha_{ik}\phi_i(a_k),
\end{equation}
and $\varphi_i$ are suitable basis
functions in the variable $a_k-(\theta,\varphi_k)$.
\subsection{Cluster Expansion}
\label{sec:cluster_expansion}
The cluster expansion is a functional approximation
method for Hopf characteristic functionals
that leverages on the structure of the functional
itself to provide an approximation that preserves
important properties such as normalization and
marginalization rules.
To illustrate the method, let us
consider the restriction of the Hopf functional to
the finite-dimensional function space $D_m\subseteq D(\Phi)$
(see Example 1 in Section \ref{sec:Finite_Dim_Approx}, and Figure \ref{fig:1}).
\begin{equation}
\phi(a_1,...,a_m)=\left<e^{i(a_1U_1(\omega)+\cdots+ a_m U_m(\omega))}\right>,
\qquad U_k(\omega)=\int_a^b u(x;\omega)\varphi_k(x)dx.
\label{jointCHF}
\end{equation}
To represent the high-dimensional (complex-valued)
function $\phi(a_1,...,a_m)$ one can use techniques such as
HDMR (Section \ref{sec:HDMR}), functional tensor methods
(Section \ref{sec:tensor}), or sparse collocation
\cite{Bungartz,DoostanOwhadi_2011,NovakR_96}.
One of the problems with such techniques is that they
do not preserve important properties of the
characteristic function, normalization $\phi(0,...,0)=1$,
and marginalization rules.
However, in the case of \eqref{jointCHF}, we do know
the {\em structure} of cylindrical functional we are
approximating. Therefore, we can leverage on such
structure to build an expansion that preserves
marginalization rules and other properties.
To this end, let us first define the following
reduced-order characteristic functions
\begin{align}
\phi_n(a_n)&=\left<e^{ia_nU_n(\omega)}\right>,\label{short1}\\
\phi_{nm}(a_n,a_m)&=\left<e^{ia_nU_n(\omega)+ia_mU_m(\omega)}\right>,\label{short2}\\
\phi_{nmk}(a_n,a_m,a_k)&=\left<e^{ia_nU_n(\omega)+ia_mU_m(\omega)+ia_kU_k(\omega)}\right>,\label{short3}\\
&\cdots\nonumber.
\end{align}
Clearly, we have
\begin{equation}
\phi_{nmk}(a_n,0,0)=\phi_{nm}(a_n,0)= \phi_{n}(a_n), \qquad\textrm{(marginalization rule).}
\end{equation}
By using well-known cumulant series representations \cite{Kubo}, we expand \eqref{jointCHF} as
\begin{equation}
\left<\exp\left[i\sum_{j=1}^N a_jU_j(\omega) \right]\right>=
\exp\left[\sum_{\nu_1,...,\nu_N=0}^\infty\left<U_1^{\nu_1}\cdots U_N^{\nu_N}\right>_c
\prod_{k=1}^N\frac{(ia_k)^{\nu_k}}{\nu_k!}\right],
\label{Kubo_cumulant}
\end{equation}
where the last summation is over $\nu_1$, ..., $\nu_N$
excluding the case in which all indices are zero, i.e.,
excluding $\nu_1=\cdots=\nu_N=0$. The cumulant averages
in \eqref{Kubo_cumulant} can be written in terms of
moments of $U_j$. For example, we have
\begin{align}
\left<U_j\right>_c=&\left<U_j\right>,\nonumber\\
\left<U_jU_k\right>_c=&\left<U_jU_k\right>-\left<U_j\right> \left<U_k\right>,\nonumber\\
\left<U_jU_kU_m\right>_c=&\left<U_jU_kU_m\right>-\left<U_j\right> \left<U_kU_m\right>-
\left<U_k\right> \left<U_jU_m\right>-\left<U_m\right> \left<U_kU_j\right>+
2\left<U_j\right>\left<U_k\right>\left<U_m\right>.\nonumber
\end{align}
At this point, it is useful to write the \eqref{Kubo_cumulant} explicitly
for simple cases, e.g., for \eqref{short2} and \eqref{short3}. We have,
\begin{equation}
\phi_{nm}(a_n,a_m)= \phi_n(a_n)\phi_m(a_m)
\exp\left[\sum_{j,k=1}^\infty\left<U_m^{k}U_n^{j}\right>_c
\frac{(ia_m)^{k}(ia_n)^{j}}{k!j!}\right]\quad \textrm{and}
\end{equation}
\begin{align}
\phi_{nmk}(a_n,a_m,a_k)= & \phi_n(a_n)\phi_m(a_m)\phi_k(a_k)
{\color{r}\frac{\phi_{nm}(a_n,a_m)}{\phi_n(a_n)\phi_m(a_m)}
\frac{\phi_{mk}(a_m,a_k)}{\phi_m(a_m)\phi_k(a_k)}
\frac{\phi_{nk}(a_n,a_k)}{\phi_m(a_n)\phi_k(a_k)}}\times\nonumber\\
& \exp\left[\sum_{j,z,q=1}^\infty\left<U_m^{j}U_n^{z}U_k^{q}\right>_c
\frac{(ia_m)^{j}(ia_n)^{z}(ia_k)^{q}}{j!z!q!}\right].
\label{ioio}
\end{align}
By generalizing these results to $N$ variables, we find the following
{\em cluster expansion}
\begin{align}
\phi(a_1,...,a_N)=&\prod_{n=1}^N\phi_n(a_n)
\prod_{n<m}^N\frac{\phi_{nm}(a_n,a_m)}{\phi_n(a_n)\phi_m(a_m)}\times\nonumber\\
&\prod_{n<m<k}^N\frac{\phi_{nmk}(a_n,a_m,a_k)}{\phi_n(a_n)\phi_m(a_m)
\phi_k(a_k)\phi_{nm}(a_n,a_m)\phi_{mk}(a_m,a_k)\phi_{nk}(a_n,a_k)}\times\cdots .
\label{cluster_expansion}
\end{align}
Any truncation of \eqref{cluster_expansion} to a certain order
in the multi-point characteristic functions $\phi_{lmn\cdots}$
yields approximations\footnote{The rationale behind this
approximation relies on the fact that the joint cumulants
of $U_k$ often decay with the order, and therefore
the exponential function (see \eqref{ioio})
tends to $1$ quickly.}.
In particular, the second-order truncation
\begin{align}
\phi(a_1,...,a_N)\simeq&\prod_{n=1}^N\phi_n(a_n)
\prod_{n<m}^N\frac{\phi_{nm}(a_n,a_m)}{\phi_n(a_n)\phi_m(a_m)}
\label{cluster_expansion_3}
\end{align}
involves $N(N-1)/2$ functions $\phi_{ln}$ ($\phi_l$ can be
determined from $\phi_{lm}$), and it can be defined as {\em Gaussian approximation}.
The reason for such definition is that
if $U_1$,..., $U_N$ are jointly Gaussian,
then \eqref{cluster_expansion_3} is exact
(the $n$th-order cumulants of a multivariate Gaussian are zero
if $n\geq 3$).
We can establish an interesting connection between the networks
of test functions we discussed in Section \ref{sec:interpolationnodes}
and the truncation order in \eqref{cluster_expansion}.
\paragraph{Representation in $S^{(m)}_1$}
The representation of the Hopf functional in $S^{(m)}_1$
(see Eq. \eqref{SNq}) is defined in terms of one-dimensional functions
\begin{equation}
\Phi([a_j\varphi_j(x)]),\qquad j=1,...,m.
\end{equation}
The corresponding approximation takes the form
\begin{align}
\Phi([a_1\varphi_1(x)+\cdots +a_m\varphi_m(x)])\simeq&
\prod_{j=1}^m\Phi([a_j\varphi_j(x)]).
\label{cluster_expansion_1}
\end{align}
In statistical physics this is known as {\em mean-field approximation},
and it relies on a statistical independence hypothesis between the
random variables $U_k(\omega)$ in \eqref{jointCHF}.
\paragraph{Representation in $S^{(m)}_2$}
\label{sec:S2N}
The representation of the Hopf functional in $S^{(m)}_2$
involves two-dimensional functions in the form
\begin{equation}
\Phi([a_j\varphi_j(x)+a_k\varphi_k(x)]), \qquad k,j=1,...,m.
\label{twopoint}
\end{equation}
With this set we consider the second-order truncation
of \eqref{cluster_expansion}, i.e.,
\begin{align}
\Phi([a_1\varphi_1(x)+\cdots +a_m\varphi_m(x)])\simeq&\prod_{j=1}^m\Phi([a_j\varphi_j(x)])
\prod_{j<k}^m\frac{\Phi([a_j\varphi_j(x)+a_k\varphi_k(x)])}
{\Phi([a_j\varphi_j(x)])\Phi([a_k\varphi_k(x)])}\nonumber\\
=&\prod_{j<k}^m \Phi([a_j\varphi_j(x)+a_k\varphi_k(x)])
\prod_{j=1}^m\Phi([a_j\varphi_j(x)])^{2-m}.
\label{cluster_expansion_2}
\end{align}
Clearly, the one-dimensional functions $\Phi([a_j\varphi_j(x)])$
can be obtained from \eqref{twopoint} by setting $a_k=0$.
\subsection{Functional Approximation Based on Random Processes }
\label{sec:Stochastic Functional Methods}
So far we discussed representations of nonlinear functionals
based on power series, Lagrangian interpolants, and tensor methods.
In this Section we briefly discuss another way of constructing
polynomial functional expansions by using stochastic processes.
The method was pioneered by N. Wiener in \cite{Wiener66},
Suppose we are given a random function $u(x;\omega)$,
with know statistical properties and a real-valued nonlinear
functional $F([\theta])$. If $u(x;\omega)$ is in the domain of
$F$ then we can evaluate $F([u(x;\omega)])$, which is
a real-valued random variable. The set
\begin{equation}
\{u(x;\omega),F([u(x;\omega)])\}
\end{equation}
can be considered as an {\em infinite collection} of
input-output signals from which we would like to determine
determine a polynomial approximation of $F$.
The key point is that the stochastic signal $u(x;\omega)$ is
equivalent to an infinite collection of functions
that span the domain of $F$, hopefully in a way that is
sufficient to identify $F([\theta])$.
To this end, we look for an expansion in the form
\begin{equation}
F([\theta])=\sum_{k=0}^\infty G_{k}([\theta]),
\label{stochastic_exp}
\end{equation}
where $G_{k}([\theta])$ $k=0, 1, ...$ is a complete
set of orthogonal polynomial functionals \cite{Wiener66,Segall,Ogura,Ernst}.
Orthogonality here is relative to the probability measure $P[u]$ of the
random process $u(x;\omega)$, i.e., relative to the inner product
\begin{equation}
\left(G_{k}([u]),G_{j}([u])\right)_{dP[u]}=\int_{\Omega} G_{k}([u])G_{j}([u])dP[u].
\label{IP2}
\end{equation}
It was shown in \cite{Wiener66} that if $u(x;\omega)$ is Brownian motion, then $G_k$ are Hermite polynomial functionals,
and \eqref{stochastic_exp} becomes the celebrated
{\em Wiener-Hermite expansion} \cite{Wiener66,Cameron,Db_book}.
For completeness, we recall that the first- and
the second-order Hermite polynomial
functionals are defined as (see \cite{Wiener66}, p. 32)
\begin{align}
G_{1}([u])=&\int K_1(x_1)du(x_1;\omega),
\label{D10}\\
G_{2}([u])=&\int \int K_2(x_1,x_2)du(x_1;\omega)du(x_2;\omega)-
\int K_2(x_1,x_1) dx_1,
\label{D2}
\end{align}
where $ K_2(x_1,x_2)$ is subject to the normalization condition
\begin{equation}
\int K_2(x_1,x_2)^2 dx_1dx_2=\frac{1}{2}.
\end{equation}
Clearly, if $u(x;\omega)$ is Brownian motion then
the integrals in \eqref{D10}-\eqref{D2} do not exist
in the ordinary Stieltjes sense because $u(x;\omega)$
is nowhere differentiable. However, we can interpret
the derivative of $u(x;\omega)$ in a distributional
sense to obtain a generalized process, i.e., the white noise.
Wiener has shown that in such generalized setting, the
integral \eqref{D10}-\eqref{D2} are perfectly well defined
for kernel functions $K_1(x_1)$ and $K_2(x_1,x_2)$
in $L_2$. A detailed mathematical analysis of the
Wiener-Hermite expansion can be found in \cite{Ernst,Cameron,Wiener66,Poggio}.
The process of determining the kernels of $G_{1}([\theta])$,
$G_{2}([\theta])$, etc., is known as {\em identification process}
in the theory of nonlinear systems, and it has been
studied extensively for obvious reasons (see \cite{Nelles,Rugh,Shetzen}).
By leveraging on the orthogonality of $G_k$
relative to the inner product \eqref{IP2} one can
show that (see, e.g., Eq. (4.4) in \cite{Wiener66})
\begin{equation}
K_p(x_1,...,x_p)=\frac{1}{p!\epsilon^p}\left(F([u]),G_{p}(H_p,[u])\right)_{dP[u]},\qquad
H_p(z_1,...,z_p)=
\begin{cases}
1 & x_i\leq z_i\leq x_i+\epsilon \\
0 & \textrm{otherwise}
\end{cases}.
\label{Kpp}
\end{equation}
This result relies on the fact that orthogonality of
$G_k$ and $G_j$ in the sense of \eqref{IP2} is not dependent on
the actual kernel functions appearing in $G_k$
and $G_j$. In fact, if we consider the inner product of \eqref{stochastic_exp}
with $G_p(H_p,[u])$ -- arbitrary kernel $H_p(x_1,...,x_n)$ --
then the only term that survives at the right hand side
is $\left(G_p(K_p,[u])G_p(H_p,[u])\right)_{dP[u]}$.
Such inner product can be written as
\begin{equation}
\left(G_p(K_p,[u])G_p(H_p,[u])\right)_{dP[u]}=p!
\int_a^b\cdots \int_a^bK_p(x_1,...,x_p)H_p(x_1,...,x_p)dx_1\cdots dx_p,
\end{equation}
i.e., all contributions of lower-order terms are identically zero.
Thus, if we select $H_p$, properly (e.g., $H_p(x_1,...,x_p)=1$ in
a small hypercube centered at $(x_1,...,x_p)$ with side length
$\epsilon$ and zero otherwise) then we can
extract $K_p(x_1,...,x_p)$ as in equation \eqref{Kpp}.
This means that if we know the response $F([u])$ corresponding
to the stochastic process $u(x;\omega)$, then we can
identify the kernels \eqref{Kpp} and therefore
construct the corresponding polynomial functional
expansion \eqref{stochastic_exp}.
The path integrals in \eqref{Kpp} can be evaluated
numerically, e.g., by using Monte-Carlo or quasi-Monte Carlo
methods (see Appendix \ref{sec:functional integrals}).
To this end, suppose we have available a
collection of $N_s$ of response-excitation signals
$\{u(x;\omega_i),F([u(x;\omega_i)])\}$
($i=1,..., N$), where $u(x;\omega_i)$ is a
realization of the process $u(x;\omega)$ obtained
by sampling the probability functional $P[u]$.
The Monte-Carlo estimate of
the kernels \eqref{Kpp} is simply
\begin{equation}
K_p(x_1,...,x_p)\simeq \frac{1}{p! \epsilon^p N}\sum_{i=1}^N
F([\xi(x,\omega_i)])G_{k}(H_p,[\xi(x,\omega_i)])\qquad p=0,1,...
\label{KppMC}
\end{equation}
where $\xi(x,\omega_i)$ is a realization of the
distributional derivative of $u(x;\omega_i)$.
In particular, if $u(x;\omega)$ is
Brownian motion then $\xi(x;\omega)$ is
{\em spatial white noise}\footnote{A simple numerical
approximation of white
noise $\xi(x;\omega)=du(x;\omega)/dx$ can be obtained as
\begin{equation}
\xi_{\Delta x}(x;\omega) = \frac{u(x+\Delta x;\omega)-u(x;\omega)}{\Delta x},
\qquad \textrm{($u$ Brownian motion).}
\end{equation}
This is a Gaussian process with zero mean, variance $1/\Delta x$ and covariance
\begin{equation}
\left<\xi_{\Delta x}(x;\omega)\xi_{\Delta x}(y;\omega)\right>=
\begin{cases}
\displaystyle\frac{1}{\Delta x}\left(1-\frac{|x-y|}{\Delta x}\right)& \textrm{if $|x-y|\leq \Delta x$}\\
0 & \textrm{otherwise}
\end{cases}
\end{equation}
The Fourier transform of $\left<\xi_{\Delta x}(x;\omega)\xi_{\Delta x}(y;\omega)\right>$
yields a power spectrum which is not flat as it is supposed to be for white noise,
but decays at sufficiently large frequencies.
}.
In this case, we have that the first two
functionals $G_1(H_1,[\xi])$ and $G_2(H_2,\xi)$
in \eqref{KppMC} are given by
\begin{align}
G_{1}(H_1,[\xi])&=\epsilon \xi(x_1;\omega), \nonumber\\
G_{2}(H_2,[\xi])&=\epsilon^2 \left(\xi(x_1;\omega)
\xi(x_2;\omega)-1\right), \nonumber
\end{align}
and therefore
\begin{align}
K_1(x_1)\simeq &\frac{1}{ N}\sum_{i=1}^N F([\xi^{(i)}])
h_1(\xi^{(i)}(x_1;\omega)),\label{KppMC1}\\
K_2(x_1,x_2)\simeq & \frac{1}{ N}\sum_{i=1}^N F([\xi^{(i)}])
h_2(\xi^{(i)}(x_1;\omega),\xi^{(i)}(x_2;\omega)).
\label{KppMC2}
\end{align}
Here $h_1(x_1)=x_1$ and $h_2(x_1,x_2)=(x_1x_2-1)/2$ are the
classical first- and second-order Hermite polynomials.
A very insightful discussion on how to compute the kernels
$K_j$ by using white noise is given by Rugh in \cite{Rugh}, \S 7.4.
The expansion \eqref{stochastic_exp} obtained in this way
can, in principle, be used to compute the value of the functional
$F([\theta])$ corresponding to any deterministic
function $\theta(x)$. However, since the
kernels \eqref{Kpp} are built upon stochastic processes
and their averages, it is not clear in what
sense \eqref{stochastic_exp} will converge for
deterministic input functions $\theta(x)$.
This question was addressed by
Palm and Poggio in \cite{Poggio} (Theorems 4 and 5),
where necessary and sufficient conditions for pointwise
convergence of Wiener-Hermite expansions are provided
(see also \cite{Ernst}).
The functional expansion in terms of orthogonal
polynomial functionals can be generalized
to random processes other than Brownian motion and
white noise\footnote{The construction
of such generalized expansion proceeds as follows.
Starting from the constant functional $F_0$ we look for
\begin{equation}
F_1([u])=C_1 \int K_1(x) du(x;\omega)+ F_0,
\end{equation}
where $u(x;\omega)$ is a generalized random process, and
we make it orthogonal to $F_0$ in the sense
of $dP[u(x;\omega)]$. This yields $C_1$.
Then we construct $F_2([u])$ and we make it
orthogonal to both $F_1([u])$ and $F_0$.
All these conditions are ultimately expressed
analytically in terms of multi-point averages of
the random process $u(x;\omega)$.}.
However, one has to be very careful when
performing such generalizations. There are indeed
random processes that do not allow for a complete
representation of the functional $F([\theta])$
(see \cite{Ernst,Segall,Ogura} for details).
\section{Functional Differential Equations}
\label{sec:FDEs approximation}
A functional differential equation (FDE) is an equation involving
a nonlinear functional (i.e., a nonlinear operator),
derivatives with respect to functions (functional derivatives)
and derivatives with respect to independent variables, e.g.,
space and time variables.
In this report we will study linear FDEs in the form
\begin{equation}
\frac{\partial F([\theta],t)}{\partial t}=L([\theta],t) F([\theta],t)+H([\theta],t)
\qquad
F([\theta],0)=F_0([\theta]),
\label{linfde00}
\end{equation}
where $F_0([\theta])$ is an initial condition,
$H([\theta],t)$ is a known forcing term
and $L([\theta],t)$ is a linear operator
in the space of functionals. This class of FDEs
is very broad, and it includes many well-known
functional equations of theoretical and quantum physics.
The solution to initial value problem \eqref{linfde00}
(assuming it exists) depends on the
{\em initial condition} $F_0$ as well as on the
choice of the {\em function space} $D(F)$ (domain
of the functional). The following
examples clarifies this point.
\vspace{0.7cm}
\noindent
{\em Example 1 (Functional Advection Equation):}
Consider the following FDE
\begin{equation}
\frac{\partial F([\theta],t)}{\partial t}+\int_0^{2\pi}\theta(x)
\frac{\partial }{\partial x}\frac{\delta F([\theta],t)}{\delta \theta(x)}dx=0, \qquad
F([\theta],0)=F_0([\theta]).\label{FDEln}
\end{equation}
Clearly, this equation is in the form \eqref{linfde00}, with
$H([\theta],t)=0$ and
\begin{equation}
L([\theta],t)=-\int_0^{2\pi}\theta(x)
\frac{\partial }{\partial x}\frac{\delta }{\delta \theta(x)} dx.
\label{FDOP}
\end{equation}
The solution to \eqref{FDEln} depends
on the {\em initial condition} $F_0([\theta])$
as well as on the choice of
the {\em function space} $D(F)$. To show this,
let us set
\begin{equation}
F_0([\theta])=\int_0^{2\pi}K(x)\theta(x)dx,
\label{ic0}
\end{equation}
where $K(x)$ is a given kernel function\footnote{
The solution functional $F([\theta],t)$ corresponding to
the initial condition \eqref{ic0} is linear and
homogeneous for all $t\geq0$, i.e., it is in the form
\begin{equation}
F([\theta],t)=\int_0^{2\pi}R(x,t)\theta(x)dx,\qquad R(x,0)=K(x).
\label{K1lin}
\end{equation}
To see this, it is sufficient
to write the first-order Euler scheme
\begin{equation}
F([\theta],\Delta t)=F([\theta],0) -\Delta t \int_0^{2\pi}\theta(x)
\frac{\partial K(x,0)}{\partial x}dx.
\end{equation}
and note that the term within the integral at the right hand side
is a linear functional of $\theta$. This implies that $F([\theta],\Delta t)$ is
a linear functional of $\theta$. By applying this argument over
and over we see that $F([\theta],t)$ is a linear functional of
$\theta$ for all $t\geq0$, i.e., it is in the form \eqref{K1lin}.
A substitution of \eqref{K1lin} into \eqref{FDEln} yields
\begin{equation}
\int_0^{2\pi}\theta(x)\left(\frac{\partial R(x,t)}{\partial t}+
\frac{\partial R(x,t)}{\partial x}\right)dx=0,\qquad \textrm{i.e.} \qquad
\frac{\partial R(x,t)}{\partial t}+
\frac{\partial R(x,t)}{\partial x}=0.\label{adv4}
\end{equation}
Therefore, \eqref{K1lin} is a solution to \eqref{FDEln}-\eqref{ic0}
if $R(x,t)$ solves a simple linear advection equation on the real line,
with initial condition $R(x,0)=K(x)$.
On the other hand, if the initial condition $F_0([\theta])$ is
a nonlinear functional of $\theta$, then $F([\theta],t)$
is nonlinear functional of $\theta(x)$. We will discuss this
case extensively in Section \ref{sec:numerical results functional equations}).
}.
As we will see in Section \ref{sec:ADVR},
if we solve \eqref{FDOP} in the space of periodic functions
\begin{equation}
D(F) =\{\theta(x)\in C^{(\infty)}([0,2\pi])\,|,\ \theta(0)=\theta(2\pi)\},
\end{equation}
then we obtain the solution
\begin{equation}
F([\theta],t)=F_0(\theta),
\end{equation}
i.e., the constant functional. On the other hand,
if we consider the function space
\begin{equation}
D(F) =\{\theta(x)\in {C}^{(\infty)}([0,2\pi])\,|,\ \theta(x)=0\},
\end{equation}
then the solution to \eqref{FDOP} is
\begin{equation}
F([\theta],t)=
\begin{cases}
\displaystyle \int_t^{2\pi}K(x) \theta(x)dx & t\leq 2\pi\\
0 & t > 2\pi
\end{cases}.
\end{equation}
\paragraph{Existence and Uniqueness of the Solution}
A fundamental question that has lasted over the years
is whether the solution to linear FDEs are unique
or not, given the initial state $F_0([\theta])$
and the function space $D(F)$. This is an unsolved mathematical
problem which we hope will be addressed systematically
in near future. At today, there are few general theorems and
results on the existence and the uniqueness of solutions to
functional differential equations (see, e.g.,
\cite{Foias,Hale,Wu}).
Before addressing the question of numerical approximation
functional differential equations, it is useful
to show how such equations look like and, more importantly,
how they arise in the context of well-known mathematical theories.
\subsection{Variational Form of PDEs}
\label{sec:variational}
Perhaps, the simplest example
of a (algebraic) functional equation is the
variational form of a PDE.
To show how such equation
looks like, consider the scalar PDE
\begin{equation}
N(u)=0,\qquad
\label{PDE}
\end{equation}
where $N$ is a nonlinear differential operator subject
to appropriate initial/boundary conditions. For example,
\begin{equation}
N(u)=\frac{\partial u}{\partial t} +u\frac{\partial u}{\partial x}-
\frac{\partial^2 u}{\partial x^2},\qquad u(x,0)=u_0(x), \qquad u(0,t)=u(L,t).
\label{firsteq}
\end{equation}
We multiply \eqref{PDE} by the test function $\theta(x)$
in the space
\begin{equation}
D(F)= \left\{
\theta \in C^{(2)} ([0,2\pi])\,|\, \theta(0)=\theta(2\pi)\right\}
\end{equation}
and integrate over $[0,2\pi]$ we obtain
\begin{equation}
F([\theta])=\left(N(u),\theta\right) =0, \qquad \forall\theta\in D(F).
\end{equation}
This is the starting point of the well-known method
of {\em weighed residuals} \cite{Finlayson},
from which classical Galerkin,
collocation, least-squares and finite volumes
schemes can be derived (see \cite{GKSS_2005}, p.18).
As we will see in Section \ref{sec:rep-Hopf}, to identify
the functional $F$ in this case it is sufficient
to test it relative to a set of linearly independent
functions $\theta_i$ ($i=1, 2,...$), e.g., an
orthonormal basis of $D(F)$.
\subsection{Schwinger-Dyson Equations}
The Schwinger-Dyson equations are functional differential
equations for the generating functional of a field theory.
They arise in both classical statistical physics as
well as in quantum field theory. Hereafter we review the main
aspects of such equations.
\paragraph{Statistical Physics}
The functional integral approach to classical statistical dynamics
\cite{Jensen,Phythian,Jouvet,Langouche} allows us derive formally exact evolution equations for phase space functions
such as the mean and the correlation function of
the solution to SODEs and SPDEs.
The standard approach relies on a {\em generating functional} $Z$.
For stochastic dynamical systems in the form
\begin{equation}
\frac{d{\bm \psi}(t)}{dt}=\bm \Lambda(\bm \psi(t),t)+\bm f(t;\omega),\label{eqofm}
\end{equation}
$Z$ can be written as a functional integral
(see \cite{Phythian,Jensen,Justin,Amit,Stratonovich})
\begin{equation}
Z([\bm \xi,\bm \eta])=Z_0\int \int\mathcal{D}[\bm \psi]
\mathcal{D}[\bm \chi]A([\bm \psi,\bm \chi])
\exp\left(\int_0^t d\tau (\bm \xi(\tau)\cdot\bm \psi(\tau)+
\bm \eta(\tau)\cdot\bm \chi(\tau))\right),
\label{gf}
\end{equation}
where {\color{r}
\begin{equation}
\frac{1}{Z_0} = \int \int\mathcal{D}[\bm \psi]\mathcal{D}[\bm \chi]
A([\bm \psi,\bm \chi]),
\end{equation}}
and
\begin{equation}
A([\bm \psi,\bm \chi]) = C([\bm \chi])\exp\left(-\frac{1}{2}\int_0^t d\tau
\nabla\cdot \bm \Lambda(\bm \psi(\tau),\tau)-i\int_0^t d\tau \bm \chi(\tau)\cdot
\left[ \frac{d{\bm \psi}(\tau)}{dt}-\bm \Lambda(\bm \psi(\tau),\tau)\right]\right).
\end{equation}
Here $C([\bm \chi])$ denotes the (known) characteristic
functional of the external random noise $\bm f(t;\omega)$.
Clearly, if we have available the solution to the stochastic dynamical
system \eqref{eqofm} then we can construct
the functional $Z([\bm \xi,\bm \eta])$, and from it
compute any statistical property we may be interested in.
On the other hand, it is straightforward to show
that $Z([\bm \xi,\bm \eta])$ satisfies a
coupled system of {\em linear} functional differential equations,
known as {\em Schwinger-Dyson equations} in quantum field theory \cite{Itzykson}.
The equations are in the form
\begin{align}
\frac{\partial}{\partial \tau}\left(\frac{1}{i}\frac{\delta Z}{\delta \xi_k(\tau)}\right)&=
\eta_k(\tau)Z+\Lambda_k\left(\frac{1}{i}\frac{\delta}{\delta \bm \xi(\tau)},\tau\right)Z-
iD_k\left(\left[\frac{1}{i}\frac{\delta}{\delta\bm \eta(\tau)}\right],\tau\right)Z,
\label{f1}\\
\frac{\partial}{\partial \tau}\left(\frac{1}{i}\frac{\delta Z}{\delta \eta_k(\tau)}\right)&=
-\xi_k(\tau)Z+i\frac{\delta}{\delta \eta_j(\tau^+)}\frac{\partial \Lambda_k}{\partial \psi_j}
\left(\frac{1}{i}\frac{\delta}{\delta \bm \xi(\tau)},\tau\right)Z
\label{f2},
\end{align}
where
\begin{equation}
D_i([\bm \chi],\tau)=\frac{\delta}{\delta \chi_i(\tau)}\ln C[\bm \chi].
\end{equation}
The quantities $\delta/\delta \xi_k$ and $\delta/\delta \eta_k$ are
first-order functional derivative operators, defined in Appendix
\ref{app:functional derivatives}.
By solving \eqref{f1}-\eqref{f2} we can identify
$Z([\bm \xi,\bm \eta])$ without any knowledge of the
stochastic process $\bm \psi(t;\omega)$.
By generalizing \eqref{gf}, it is possible to derive a functional integral
formalism to classical statistical dynamics yielding Schwinger-Dyson equations
for generating functionals associated with SPDEs (see, e.g., \cite{Jensen,Martin}).
In particular, if the SPDE admits an action functional $A[\phi]$
(see \cite{Daniele_JMathPhys,Eyink,Gomes,Funaki,Yasue,Eyink_1996,Amit}),
then the construction of the generating functional as well as the derivation
of the corresponding Schwinger-Dyson equations are standard.
In this setting, the Schwinger-Dyson functional
differential equations provide a non-perturbative
formulation of the problem of computing the statistical
properties of nonlinear random systems, including
stochastic ODEs and stochastic PDEs.
{\color{r}
\paragraph{Quantum Field Theory}
In quantum field theory, the
Schwinger-Dyson equations govern the dynamics
of the Green functions and they characterize
the propagation field interactions \cite{Easther,Justin,Weinberg}.
Such functional equations can be employed in
a perturbation setting \cite{Okopinska}
(weak coupling regime), but they
show their true strength in the strong coupling
regime \cite{Swanson,Guralnik1}.
The starting point to derive the Schwinger-Dyson
equations is the generating functional of
the correlation functions (Green functions), which
can be often expressed as a functional integral
\begin{equation}
\displaystyle
Z([j(\bm x)]) = \int \mathcal{D}[\phi] e^{i A([\phi]) + i
\int j(\bm x)\phi(\bm x)d\bm x}.
\label{GF0}
\end{equation}
where $A([\phi])$ is the {\em action functional}.
In the context of quantum $\phi^4$-theory \cite{KleinertPhi4}
the action $A$ for a field with mass $m$ is given by
\begin{equation}
A([\phi]) = \int\left[ \frac{1}{2}\left(\nabla \phi(\bm x)\right)^2-
\frac{1}{2} m^2 \phi^2(\bm x) -
\frac{\lambda}{4!}\phi^4(\bm x)\right] d\bm x.
\label{phi4}
\end{equation}
By employing \eqref{GF0},
we can express the Green functions of the
quantum field theory at any order by functional
differentiation\footnote{The so-called {\em connected} Green
functions of the field theory are obtained by functional differentiation
of $\log Z([j])$. The main reason for such definition is that if
we represent the functional derivatives of $\log Z$
in terms of Feynman diagrams then only the
connected diagrams contribute to the expansion.}, i.e.,
\begin{align}
Z([0]) G(\bm x_1,...,\bm x_n)= &\displaystyle\int\mathcal{D}[\phi]\phi(\bm x_1)\cdots \phi(\bm x_n)
e^{i A([\phi])},\nonumber\\
=&\frac{1}{i^n}\left.\frac{\delta^n Z([j(\bm x)])}{\delta j(\bm x_1)\cdots \delta j(\bm x_n)}\right|_{j(\bm x)=0}.
\end{align}
Roughly speaking, the correlation functions are
averages of products of fields $\phi(x_1)\cdots \phi(x_n)$
with respect to the functional measure $\exp{iA([\phi])}$.
From a physical viewpoint they represent the transition
amplitude for the propagation of a particle or
excitation between different points
in space-time.
The generating functional \eqref{GF0} satisfies the
Schwinger-Dyson equation
\begin{equation}
\frac{\delta A([\phi])}{\delta \phi(\bm x)}\left(\left[-i\frac{\delta}{\delta j
(\bm x)}\right]\right)Z([j])+j(\bm x)Z([j])=0.
\end{equation}
For example, if the action $A$ is the one given
in \eqref{phi4}, then the Schwinger-Dyson equation
takes the form
\begin{equation}
\square
\frac{\delta Z([j])}{\delta j(\bm x)}+
m^2 \frac{\delta Z([j])}{\delta j(\bm x)}-
\frac{\lambda}{3!}
\frac{\delta^3 Z([j])}{\delta j(\bm x)^3 }-i j(\bm x) Z([j])=0.
\end{equation}
A substitution of the functional Taylor expansion
of $Z([j])$ into this equation yields an
infinite-dimensional coupled PDE system for the
correlation functions $G(\bm x_1,...,\bm x_n)$.
Here $\bm x_i$ is a quadruple of coordinates.
Solution methods for the Schwinger-Dyson equations
rely on truncated series expansions \cite{Okopinska,Bender},
or renormalized expansions of the generating functional
$Z([\eta])$ (see \cite{Kleinert1} and \cite{Amit}, p. 385),
or numerical algorithms \cite{Easther,Guralnik1,Guralnik2}.
}
\subsection{Hopf Characteristic Functional Equations}
\label{sec:hopf_equations}
The Hopf characteristic functional of a random field
is the functional Fourier transform of the probability
density functional (see \cite{Rosen1,Venturi_PRS} and
Appendix \ref{app:functional fourier transform}).
To introduce this mathematical object in a simple way,
let us consider an integrable random function $u(x;\omega)$
on an interval $x\in[a,b]$.
The Hopf functional associated with $u(x;\omega)$ is defined as
\begin{equation}
\Phi([\theta(x)])=\left<\exp \left[ i\int_a^b u(x;\omega)\theta(x)dx\right]\right>,
\label{hopf_functional}
\end{equation}
where $\theta(x)$ is a deterministic function (test function),
$i$ is the imaginary unit and the average is
defined as a functional integral over the probability
functional of $u(x;\omega)$.
Equation \eqref{hopf_functional} assigns to each
function $\theta(x)$ a complex number $\Phi([\theta(x)])$
(see Figure \ref{fig:1}).
Similarly to the probability density functional, the
Hopf characteristic functional \eqref{hopf_functional} encodes
the {\em full statistical information} of the random function
$u(x;\omega)$, including multi-point moments, joint
characteristic functions and probability density
functions (see \cite{Monin1,Monin2}).
If we consider instead of a random function $u(x;\omega)$
a random vector field $\bm u(\bm x,t;\omega)$, e.g., a
stochastic solution to the Navier-Stokes equations, then
we define\footnote{Lewis and Kraichnan \cite{Lewis,Rosen1} introduced a space-time generalization
of \eqref{hopf_functional2}, namely
\begin{equation}
\Phi([\bm \theta(\bm x,t)])=\left<\exp \left[ i\int_V\int_T \bm u(\bm x,t;\omega)\cdot \bm \theta(\bm x,t)
d\bm x dt\right]\right>.
\label{hopf_functional1}
\end{equation}
This functional allows us to determine joint multi-point statistics
of the random field $\bm u(\bm x,t;\omega)$ at different times.}
\begin{equation}
\Phi([\bm \theta(\bm x)],t)=\left<\exp \left[ i\int_V
\bm u(\bm x,t;\omega)\cdot \bm\theta(\bm x)
d\bm x\right]\right>,
\label{hopf_functional2}
\end{equation}
where $V$ is a spatial domain in $\mathbb{R}^d$ ($d=2,3$).
By the Riemann-Lebesgue lemma we have that
\begin{equation}
\Phi([\bm \theta(\bm x)],t)\rightarrow 0\quad \textrm{as}\quad
\left\|\bm \theta(\bm x)\right\|\rightarrow \infty,
\end{equation}
with a rate that depends on the regularity of
the underlying probability density functional.
The derivation of the Hopf characteristic functional equation
is relatively straightforward if the random field
$\bm u(\bm x,t;\omega)$ satisfies a nonlinear PDE with
polynomial nonlinearities. Hereafter we provide some
examples.
\paragraph{Burgers-Hopf Equation}
Consider the Burgers equation
\begin{equation}
\frac{\partial u}{\partial t}+u\frac{\partial u}{\partial x}=
\nu\frac{\partial^2u}{\partial x^2},
\label{BP}
\end{equation}
in a periodic spatial domain $[0,2\pi]$,
and let the initial condition $u_0(x,\omega)$ be
random.
By differentiating the Hopf functional
\begin{equation}
\Phi([\theta(x)],t)=\left<\exp\left[i\int_{0}^{2\pi}
\theta(x)u(x,t;\omega)dx\right]\right>
\end{equation}
with respect to time
{\color{r}
and using \eqref{BP} we obtain
\begin{align}
\frac{\partial \Phi([\theta],t)}{\partial t} = &
i \int_{0}^{2\pi} \theta(x)
\left<\frac{\partial u(x,t;\omega) }{\partial t}
\exp\left[i\int_{0}^{2\pi}u(x,t;\omega) \theta(x) dx \right]
\right>dx\nonumber\\
=& i\int_{0}^{2\pi} \theta(x)
\left<\left(-\frac{1}{2}\frac{\partial (u(x,t;\omega)^2)}{\partial x}
+\nu\frac{\partial ^2u(x,t;\omega)}{\partial x^2}\right)
\exp\left[i\int_{0}^{2\pi}u(x,t;\omega) \theta(x) dx \right]
\right>dx
\label{PdFeq1}
\end{align}
i.e.,
\begin{align}
\frac{\partial \Phi([\theta],t)}{\partial t}=
\int_a^{b} \theta(x)\left(\frac{i}{2}\frac{\partial}{\partial x}
\frac{\delta^2 \Phi([\theta],t)}{\delta\theta(x)^2}+
\nu\frac{\partial^2}{\partial x^2}\frac{\delta \Phi([\theta],t)}{\delta\theta(x)}\right)dx.
\label{burgersfunctional}
\end{align}}
\noindent
This equation is known as Burgers-Hopf equation and it
has been the subject of numerous investigations
(see, e.g., \cite{Ahmadi,Hosokawa2,Monin2}).
\paragraph{Navier-Stokes-Hopf Equation}
The problem of determining the evolution of the Hopf
characteristic functional for the
Navier-Stokes equations
\begin{equation}
\frac{\partial \bm u}{\partial t}+(\bm u\cdot \nabla)\bm u=-\nabla p +\nu \nabla^2 \bm u,
\qquad \nabla\cdot \bm u=0
\end{equation}
was deemed by Monin and Yaglom as the most compact formulation of the {\em turbulence
problem}, which is the problem of determining the statistical
properties of the velocity and pressure field given statistical
information on the initial condition. The Hopf functional
differential equation corresponding to the Navier-Stokes
equations is derived in \cite{Monin2} Ch. 10 (see also \cite{Beran}, \S 3.1.4), and it is hereafter summarized for convenience
\begin{equation}
\frac{\partial \Phi([\bm \theta],t)}{\partial t}=
\sum_{k=1}^3\int_V\theta_k(\bm x)\left(i \sum_{j=1}^3\frac{\partial }{\partial x_j}
\frac{\delta^2 \Phi([\bm \theta],t)}{\delta \theta_k(\bm x)\delta\theta_j(\bm x)}
+\nu \nabla^2\frac{\delta \Phi([\bm \theta],t)}{\delta \theta_k(\bm x)}\right)d\bm x.
\label{hopfns1}
\end{equation}
Here $V$ is a periodic three-dimensional box and
$\bm \theta(\bm x)$ is chosen in a divergence free
space of test functions
(see Section \eqref{sec:Navier-Stokes-Hopf}).
Functional formulations of non-isothermal turbulent
reactive flow have been also considered, leading to more complicated Hopf equations \cite{Dopazo}.
As we will see in Section \ref{app:equivalenceHopf},
Hopf equations are equivalent to PDEs in an infinite
number of variables or to infinite-dimensional
systems of coupled PDEs, e.g., the
Monin-Lundgren-Novikov hierarchy
\cite{Hosokawa,Montgomery,Wacawczyk1,Lundgren}.
It is interesting to note that the structure
of the Hopf equation somehow resembles the weak form
of a PDE. However, there is a remarkable
difference: in the Hopf equation, both the
solution and its functional derivatives depend on
the test function. In other words, the
test function appears in the functional
equation a nonlinear way. In addition, the
equation involves derivatives with respect to
functions (functional derivatives) which are
not present in classical PDEs.
\vspace{0.7cm}
\noindent
It is straightforward to derive
Hopf functional differential equations corresponding
to linear or nonlinear evolution PDEs with polynomial
nonlinearities, such as the Kuramoto-Sivashinsky equation, the nonlinear wave equation, and Maxwell's equations subject
to random boundary or random initial conditions.
{\color{r}
\paragraph{Hopf Equations Defining Random Processes}
Hopf functional equations arise naturally also in the context of
random processes. For example, the Hopf equation defining
the characteristic functional of a zero-mean Gaussian
process is (see \cite{Klyatskin1}, p. 61)
\begin{equation}
\frac{d}{dt}\Phi([\theta(t)],t)=-\frac{1}{2}\Phi([\theta(t)],t)\theta(t)
\int_0^t C(t,\tau)\theta(\tau)d\tau,\qquad \Phi([\theta(t)],0)=1,
\label{FDEGauss}
\end{equation}
where $C(t,\tau)$ is the covariance function of the process.
In fact, the solution to \eqref{FDEGauss} is the well-known Hopf functional
\begin{equation}
\Phi([\theta(t)],t)=\exp\left[-\frac{1}{2}\int_0^t\int_0^\tau C(\tau_1,\tau_2)
\theta(\tau_1)\theta(\tau_2)\right].
\end{equation}
Similar equations can be derived for orther processes, such as
the telegrapher's random process and general Markov processes.
Note that \eqref{FDEGauss} does not involve any functional
derivative.
}
\subsection{Probability Density Functional Equations}
\label{sec:PDENFUNEQ}
In statistical mechanics, a system of $n$ particles can
be described by the joint probability density function
$p( \bm x_1, ...,\bm x_n, \bm v_1, ..., \bm v_n,t)$
where $( \bm x_i,\bm v_i)$ denotes the position
and the velocity of the $i$-th particle.
Similarly, the phase space associated
with any finite-dimensional approximation of the solution to a
stochastic PDE can be described by the joint probability
density function of the corresponding phase variables, e.g.,
the Fourier coefficients of the series expansion of the
solution \cite{Edwards, Herring,Montgomery}.
Consider a scalar random field $u(x,t;\omega)$.
The full statistical information of the random field
$u(x,t;\omega)$ at time $t$ is encoded in the
probability density functional\footnote{
The probability density functional $P([a(x)],t)$
allows us to compute all moments, cumulants
and joint PDFs by using functional integration. For example,
\begin{equation}
p(b_1,b_2,t)=\int \delta(b_1-a(x_1))\delta(b_2-a(x_2))
P([a(x)],t)\mathcal{D}[a(x)].
\end{equation}
Computing functional integrals often requires a careful definition
of the integration measure as it may be possible to run into convergence issues (see Appendix \ref{sec:functional integrals},
and \cite{Beran} \S 2.2.4).
}
\begin{align}
P([a(x)],t)=&\left<
\delta\left[a(x)-u(x,t;\omega)\right]\right>,
\label{probability_functional}
\end{align}
where $\delta[\cdot]$ denotes a {\em Dirac delta functional}
(see Appendix \ref{app:functional fourier transform}),
and the average $\langle \cdot\rangle$ is a
functional integral over the probability
measure of $u(x,t;\omega)$.
Any well-defined nonlinear stochastic PDE for
the scalar random field $u(x,t;\omega)$
can be rewritten as a {\em linear functional differential
equation} for $P([a(x)],t)$.
Such equation can be obtained by taking an appropriate
continuum limit of a finite-dimensional joint probability density
equation \cite{Venturi_PRS,Venturi_JCP}.
In practice, we can replace the discrete set of
variables $\{a_1=a(x_1), a_2=a(x_2) ,... , a_n=a(x_n)\}$
in the joint PDF equation by a continuous set
denoted by a continuously indexed set $a(x)$, i.e., a
function. Partial derivatives with respect to
$a_i=a(x_i)$ can then be replaced by functional
derivatives with respect to $a(x)$ (see Section
\ref{sec:Finite_Dim_Approx}), etc.
An alternative method to derive the probability
functional equation was proposed by Beran in \cite{Beran}.
The derivation parallels classical the Liouville theory,
where a nonlinear dynamical system is converted into
a linear PDE for the joint probability density function of
of the state vector \cite{Venturi_PRS}. To convert
a nonlinear PDE into a linear equation for the probability
density functional, we need to move one level up and
look for a functional differential equation.
{\color{r}
Perhaps, the simplest way to derive a probability density
functional equation is to inverse Fourier transform the
corresponding Hopf equation, and use functional integration by
parts. To illustrate the procedure, we first differentiate the
Hopf functional $\Phi([\theta],t)$ with respect to
time to obtain\footnote{
\color{r}
In equation \eqref{43R}, we employed
the identity
\begin{equation}
\int \exp\left[i\int_{0}^{2\pi}\theta(x)u(x,t;\omega)dx \right]P([u_0])\mathcal{D}[u_0] = \int \exp\left[i\int_{0}^{2\pi}\theta(x)a(x) dx \right]P([a],t)\mathcal{D}[a],
\end{equation}
where $P([u_0]$ and $P([a],t)$ are, respectively,
the probability functionals of $u_0(x;\omega)$
(initial condition) and $u(x,t;\omega)$.
}
\begin{align}
\frac{\partial \Phi([\theta],t)}{\partial t} = &
\int\exp\left[i \int \theta(x)a(x)dx\right]
\frac{\partial P([a(x)],t)}{\partial t}\mathcal{D}[a].
\label{43R}
\end{align}
Then we set the equality with the functional equation
that defines the evolution of $\Phi([\theta],t)$
for a particular nonlinear PDE.
For instance, if we consider the Burgers equation, then we have
that $\partial \Phi/\partial t$ is given by
equation \eqref{burgersfunctional}, i.e.,
\begin{align}
&\int\exp\left[i \int \theta(x)a(x)dx\right]
\frac{\partial P([a(x)],t)}{\partial t}\mathcal{D}[a] =\nonumber\\
&\int_{0}^{2\pi}
\int \left(-\frac{1}{2}\frac{\partial (a(x)^2)}{\partial x}
+\nu\frac{\partial ^2 a(x)}{\partial x^2}\right)
\frac{\delta}{\delta a(x)}\left(\exp\left[i\int_{0}^{2\pi}a(x) \theta(x) dx \right]\right)dx P([a],t)
\mathcal{D}[a]
\end{align}
%
Performing a functional integration by parts and assuming that
the boundary terms are zero yields
\begin{align}
\frac{\partial P([a(x)],t)}{\partial t} = -\int_{0}^{2\pi}
\frac{\delta}{\delta a(x)}\left(\left[-a(x)\frac{\partial a(x)}{\partial x}
+\nu\frac{\partial ^2 a(x)}{\partial x^2}\right] P([a],t)\right)dx.
\label{PrDFeq2}
\end{align}
\paragraph{The Method of Continuum Limits}
The formal procedure to derive probability density functional
equations we just described can be justified
in a finite dimensional setting. To this end, let us
consider the one-dimensional diffusion problem
\begin{equation}
\frac{\partial u(x,t;\omega)}{\partial t}=\frac{\partial^2 u (x,t;\omega)}{\partial x^2},
\qquad u(x,0;\omega)=u_0(x;\omega)\quad \textrm{(random)}
\label{1dheat}
\end{equation}
in the real line $x\in \mathbb{R}$. The probability
density functional equation of the solution is a subcase of
equation \eqref{PrDFeq2}, namely
\begin{align}
\frac{\partial P([a(x)],t)}{\partial t} = -\int_{-\infty}^{\infty}
\frac{\delta}{\delta a(x)}\left(
\frac{\partial^2 a(x)}{\partial x^2} P([a],t)
\right)dx.
\label{PrDFeq3}
\end{align}
To derive \eqref{PrDFeq3}, we first
discretize \eqref{1dheat} in space, e.g.,
on an spatial grid with evenly spaced
nodes $x_j$ ($j=1,...,n$), the spacing between
the nodes being $\Delta x$.
If we use second-order finite differences
we obtain
\begin{equation}
\frac{d u(x_k,t,\omega)}{dt}= \frac{u(x_{k+1},t;\omega)-2u(x_k,t;\omega)+u(x_{k-1},t;\omega)}{\Delta x^2}.
\end{equation}
Now, let $p(a_1,...,a_n,t)$ be the joint PDF of
$\{u(x_1,t;\omega),...,u(x_n,t;\omega)\}$, i.e.,
\begin{equation}
p(a_1,...,a_n,t) = \left<\prod_{k=1}^n \delta(a_k-u(x_k,t;\omega)\right>.
\end{equation}
We think of $a_k$ as the value of some
function $a(x)$ at $x_k$, that is $a_k=a(x_k)$
(see Figure \ref{fig:ProbDF}).
\begin{figure}[t]
\centerline{
\includegraphics[height=3.7cm]{ProbDF-eps-converted-to.pdf}
}
\caption{Sketch of the variables
$a(x_k)$ representing the random field $u(x,t;\omega)$ at
$x_k$. When we send $\Delta x$ to zero,
the number of variables $a(x_k)$ goes to infinity and
the joint probability density function becomes a probability density
functional (see equation \eqref{probability_functional}).}
\label{fig:ProbDF}
\end{figure}
By using well-known identities involving the Dirac delta function \cite{Khuri} it can be shown that
\begin{align}
&\frac{a(x_{k+1})-2a(x_k)+a(x_{k-1)}}{\Delta x^2}
\left<\prod_{k=1}^n \delta\left(a(x_k)-u(x_k,t;\omega)\right)\right>
=\nonumber\\
&\left<\frac{u(x_{k+1},t;\omega)-2u(x_k,t;\omega)+u(x_{k-1},t;\omega)}{\Delta x^2}
\prod_{j=1}^n \delta\left(a(x_j)-u(x_j,t;\omega)\right)\right>.
\end{align}
This yields,
\begin{align}
\frac{\tilde{\partial}^2 a(x_k)} {\tilde{\partial} x^2}
\left<\prod_{k=1}^n \delta\left(a(x_k)-u(x_k,t;\omega)\right)\right>
=
\left<\frac{\tilde{\partial}^2 u(x_k,t;\omega)}{\tilde{\partial} x^2}
\prod_{j=1}^n \delta\left(a(x_j)-u(x_j,t;\omega)\right)\right>,
\label{secondderivative}
\end{align}
where $\tilde{\partial}^2/\tilde{\partial} x^2$ is
the numerical differentiation operator, i.e., the
approximation of the second-order derivative
operator by using finite differences, or other
differentiation schemes such as pseudospectral
collocation \cite{Hesthaven}. By extending these arguments
to higher-order derivatives, we obtain}
\begin{equation}
\frac{\tilde{\partial}^s a(x_k)}{\tilde{\partial} x^s}
\left<\prod_{k=1}^n \delta\left(a(x_k)-u(x_k,t;\omega)\right)\right>
=
\left<\frac{\tilde{\partial}^s u(x_k,t;\omega)}{\tilde{\partial} x^s}
\prod_{j=1}^n \delta\left(a(x_j)-u(x_j,t;\omega)\right)\right>, \quad
s=1, 2, ...
\end{equation}
\noindent
The joint probability density function of the
state vector $u(x_i,t;\omega)$ ($i=1,...,n$)
satisfies the equation \cite{Venturi_PRS,Venturi_JCP,Venturi_JCP2013}
\begin{align}
\frac{\partial p}{\partial t}=&-\sum_{k=1}^n
\frac{\partial }{\partial a_k}\left<\frac{\tilde{\partial}^2 u(x_i,t;\omega)}
{\tilde{\partial} x^2}
\prod_{j=1}^n \delta\left(a_j-u(x_j,t;\omega)\right) \right>,
\nonumber\\
=&-\sum_{k=1}^n
\frac{\partial }{\partial a_k}
\left(\frac{\tilde{\partial}^2 a(x_i)}{\tilde{\partial} x^2}p\right),
\label{jointPDF}
\end{align}
where $p=p(a_1,...,a_n,t)$. The last equality follows
from \eqref{secondderivative}. By taking the continuum limit,
i.e., by sending $\Delta x$ to zero (and correspondingly $n$
to infinity), we obtain the following functional differential equation
for the probability density functional of the solution
to equation \eqref{1dheat}
\begin{align}
\frac{\partial P([a(x)],t)}{\partial t}=-\int_{-\infty}^\infty
\frac{\delta}{\delta a(x)}\left(\frac{\partial^2 a(x)}{\partial x^2}
P([a(x)],t)\right)dx.
\label{PFDE0}
\end{align}
This equation is in agreement with \eqref{PrDFeq3}.
\vspace{0.7cm}
\noindent
{\em Remark:} The functional equation \eqref{PFDE0}
is linear in $P([a(x)],t)$, but it involves a singular term.
Such term is generated by the derivative
\begin{align}
\frac{\delta}{\delta a(x)}
\left(\frac{\partial^2 a(x)}{\partial x^2}\right)=&
\frac{\delta}{\delta a(x)}\int_{-\infty}^\infty a(y)
\delta''(x-y)dy,\nonumber\\
=&\int_{-\infty}^\infty \delta(x-y)\delta''(x-y)dy.
\label{singular}
\end{align}
The last integral is equivalent to the second derivative of
the Dirac delta function evaluated at zero $x=0$.
Such singularity can also be seen from a purely discrete
viewpoint. To this end, substitute the (second-order)
finite-difference approximation to the second-order derivative into
\eqref{jointPDF}. This yields the equation
\begin{equation}
\frac{\partial p}{\partial t}=-\sum_{k=1}^n
\frac{a(x_{k+1})-2a(x_k)+a(x_{k-1)}}{\Delta x^2}\frac{\partial p}{\partial a_k}-
\frac{2}{\Delta x^2}p.
\end{equation}
It is clear that as $\Delta x$ goes to zero (continuum limit),
the term $2p/\Delta x^2$ generates a singularity.
\paragraph{Regularity of the Probability Functional}
The solution to a probability functional equation
may be an irregular functional. To understand
why, consider Figure \ref{fig:ProbDF}. When we send $\Delta x$ to
zero we have that $x_k$ approaches $x_{k+1}$.
Correspondingly the random variables $u(x_k;t,\omega)$
and $u(x_{k+1};t,\omega)$ tend to be the same random variable.
In this situation, the joint PDF of $u(x_k;t,\omega)$ and
$u(x_{k+1};t,\omega)$ involves a Dirac delta function
as $x_k\rightarrow x_{k+1}$ (see \cite{Lundgren}, p. 970).
In a continuum setting, the phenomenon we just
described happens at each point $x$.
Therefore the probability density functional can
be an irregular mathematical object.
\vspace{0.7cm}
\noindent
{\em Example 1:}
The probability density functional
of a zero mean Gaussian random function $u(x;\omega)$
($x\in \mathbb{R}$) with covariance $C(x,y)$ is
{proportional} to
\begin{equation}
P([a(x)])\sim \exp\left[-\frac{1}{2}\int_{-\infty}^\infty
\int_{-\infty}^\infty
C^{-1}(x,y)a(x)a(y)dxdy\right),
\label{PDFunct_for_Gaussian}
\end{equation}
where $C^{-1}(x,y)$ is the inverse covariance
function.
Such inverse covariance may be obtained by
solving the Fredholm integral equation
of the first kind
\begin{equation}
\int_{-\infty}^{\infty } C(x,y) C^{-1}(y,z)dy=\delta(x-z).
\label{diffinv}
\end{equation}
If $C(x,y)$ is smooth then its differential
inverse $C^{-1}(y,z)$ must have serious
singularities in order for the integral in \eqref{diffinv}
to yield a Dirac delta function (see Table \ref{tab:4}).
If $C(x,y)$ is homogeneous, i.e., if $C(x,y)=C(x-y)$,
then $C^{-1}(y,z)$ is called {\em convolution inverse}
\cite{Hohlfeld,Murthy}.
This method was pioneered by Hirschman and
Widder \cite{Hirschman} in the late forties.
Relevant cases of convolution inverses are summarized
in Table \ref{tab:4}. Note that the convolution
inverses of smooth coavariance functions --
such as the Mat\'e rn covariance in 2D --
are {\em rough} functions involving
Laplacians and bi-harmonic operators
applied to Dirac delta functions.
\begin{table}
\centering
\begin{tabular}{lcccc}
& Covariance & & Inverse Covariance \vspace{0.2cm}\\
\hline\\
\begin{minipage}{3cm}
\centering Exponential (1D)\\\vspace{0.1cm }
\end{minipage}\hspace{1cm}
& $\displaystyle \sigma^2 e^{-|x|/h}$ & &
$\displaystyle \frac{1}{2h \sigma^2}\left(\delta(x)-h\delta''(x)\right)$ \\\\
\begin{minipage}{3cm}
\centering Mat\'ern (2D)\\\vspace{0.1cm }\footnotesize
(polar coordinates)
\end{minipage}\hspace{1cm}
& $\displaystyle \sigma^2 r B_1\left(\frac{r}{h}\right)$ & &
$\displaystyle \frac{1}{4\pi \sigma^2h^2}\left(\frac{\delta(r)}{\pi r}-
2h^2\nabla^2\frac{\delta(r)}{\pi r}+h^4 \nabla^4\frac{\delta(r)}{\pi r}\right)$ \\\\
\begin{minipage}{3cm}
\centering Exponential (3D)\\ \vspace{0.1cm }\footnotesize
(polar coordinates)
\end{minipage}
& $\displaystyle \sigma^2 e^{-r/h}$ &\hspace{0.5cm} &
$\displaystyle \displaystyle \frac{1}{8\pi \sigma^2h^3}\left(\frac{\delta(r)}{\pi r}-
2h^2\nabla^2\frac{\delta(r)}{\pi r}+h^4 \nabla^4\frac{\delta(r)}{\pi r}\right)$\\
\end{tabular}
\caption{Convolution inverses of well-known covariance functions \cite{Oliver}. Here
$B_1$ denotes the modified Bessel function of the first kind.
It is seen that the convolution inverse of smooth functions are rough
functions involving Laplacians and bi-harmonic operators
applied to Dirac delta functions.}
\label{tab:4}
\end{table}
However, such rough functions appear
within integrals in \eqref{PDFunct_for_Gaussian}, and therefore
we expect some regularization. For instance, if
we assume that the covariance function $C(x-y)$ is
exponential (see Table \ref{tab:4}), then from
\eqref{PDFunct_for_Gaussian} we obtain
\begin{equation}
P([a(x)])\sim \exp\left[
-\frac{1}{4h\sigma^2}
\int_{-\infty}^\infty \left(
a(x)^2 + h^2\left[\frac{da(x)}{dx}\right]^2\right)dx
\right].
\label{PDFforGexp}
\end{equation}
Probability density functional equations represent
an excellent starting point to obtain effective
approximations. To this end, one needs
to follow (by analogy) the route taken in
classical statistical mechanics in which we start with
the Liouville equation, and make approximations
in order to derive an computable equation for a quantity of
interest. Such {\em coarse-graining} process for
functional differential equations is discussed
by McComb \cite{McComb} in the context of fluid
turbulence. Probability density functional
equations were derived and studied in the context turbulent
flows by Dopazo and O'Brien \cite{Dopazo}, and
Rosen \cite{Rosen_1960,Rosen_1969,Rosen_1967}.
\vspace{0.7cm}
\noindent
{\em Hopf Functionals and Probability Density Functionals:}
We have seen that Hopf equations and probability density
functional equations are related by a functional Fourier transform.
Therefore, from a purely mathematical viewpoint they are
completely equivalent. However, from the viewpoint
of approximation theory they are not equivalent at all.
Hopf functionals may be hard to resolve due to
high-frequencies related to the complex exponential.
On the other hand, probability density functional equations
may have non-smooth solutions.
The statistical properties of a random field can be equivalently
computed by using the Hopf functional or the probability
density functional. In the first case, we simply need to
take functional derivatives and evaluate them at $\theta(x)=0$
(see Section \ref{app:functional derivatives}).
In the second case, we need to compute functional
integrals, i.e., integrals in an infinite number of variables.
This requires requires a careful definition of the integration
measure (\cite{Beran}, \S 2.2.4).
{\color{r}
\subsection{Effective Fokker-Planck Systems}
Consider the stochastic dynamical system \eqref{eqofm}. Suppose
we are interested in determining an evolution equation for the joint
probability density function of the state vector $\bm \psi(t)$.
To this end, we think of $\bm \psi(t)$ as a nonlinear
functional of the random noise $\bm f(t)$, i.e., we can consider
the map $\bm \psi(t)=\bm \Psi(t;[\bm f(t)])$.
The specific form of $\bm \Psi$ depends on the system,
in particular on the nonlinear map $\bm \Lambda(\bm \psi,t)$ in
\eqref{eqofm}.
The probability density function of the $\bm \Psi(t;[\bm f(t)])$
can be expressed as a functional integral over the probability
density functional of the noise (assuming it exists)
\begin{equation}
p(\bm \psi,t) = \int \delta(\bm \psi -\bm \Psi(t;[\bm f])) P([\bm f])\mathcal{D}[\bm f],
\label{PRS}
\end{equation}
where $\delta$ here is a multivariate Dirac delta function.
From this expression, it is clear that the random noise $\bm f(t)$
determines $p(\bm \psi,t)$, and therefore the structure of
probability density function equation that evolves
$p(\bm \psi,t)$ in time.
Such equation can be derived by using functional calculus.
\cite{Venturi_PRS,Fox,Hanggi4,Moss1}, and it takes the form
\begin{equation}
\frac{\partial p(\bm \psi,t)}{\partial t} +
\nabla_{\bm \psi}\cdot\left(\bm \Lambda (\bm \psi ,t)
p(\bm \psi,t) + \left<\delta\left(\bm \psi-
\bm \Psi(t;[\bm f])\right)\bm f(t)\right>_{\bm f}\right)=0,
\label{PDF}
\end{equation}
where
\begin{equation}
\left<\delta\left(\bm \psi-
\bm \Psi(t;[\bm f])\right)\bm f(t)\right>_{\bm f}=\int
\delta(\bm \psi -\bm \Psi(t;[\bm f])) \bm f(t)P([\bm f])\mathcal{D}[\bm f].
\label{corr}
\end{equation}
The quantity \eqref{corr} represents the {\em correlation}
between two functionals of the random noise, namely
$\delta(\bm \psi -\bm \Psi(t;[\bm f]))$ and the noise itself $\bm f(t)$.
Such correlation can be disentangled
by using functional integral techniques
(see, e.g., \cite{Bochkov,Venturi_PRS,Klyatskin}).
In particular, if $\bm f(t)$ is Gaussian then \eqref{corr}
can be expressed by the well-known Furutsu-Novikov-Donsker
formula \cite{Venturi_PRS,Furutsu,Novikov,Donsker}.
Similarly, if $\bm f(t)$ is Gaussian white noise
then \eqref{eqofm} is a Markovian system and
the correlation \eqref{corr} reduces to a simple
diffusion term \cite{Venturi_PRS}. In this
case \eqref{PDF} coincides with the classical
Fokker-Plank equation \cite{Risken}. We remark that
computing the solution to \eqref{PDF} in the general (non-Markovian) setting
is very challenging. Possible techniques rely on {\em data-driven} models
that employ random paths of the SODE \eqref{eqofm},
e.g., path integral methods \cite{Hanggi3,Pesquera,McKane}.
\paragraph{Non-Markovian Random Processes on Random Graphs}
Consider a {\em non-Markovian} stochastic process
$\bm \psi(t;\omega)\in\mathbb{R}^n$ evolving
on a {\em random graph}
with $n$ nodes (see Figure \ref{fig:graph}).
Such process could model, e.g., the propagation of
epidemics in interacing individuals \cite{Andersson,Daley}.
Introducing uncertainty in the graph allows us to take into
account uncertainty in the interconnections between different
nodes, which is fundamentally important when modeling
dynamics of {\em social networks} and disease propagation.
\begin{figure}
\centerline{
\includegraphics[height=4cm]{graph1-eps-converted-to.pdf}
\includegraphics[height=4cm]{graph2-eps-converted-to.pdf}
\includegraphics[height=4cm]{graph3-eps-converted-to.pdf}
}
\caption{Samples of a random graph with fixed
number of nodes. Introducing uncertainty in the edges that connect
different nodes allows us to take into account uncertainty
in the interactions between different components of a
stochastic model evolving on the graph. This is important
when modeling dynamics of social networks and disease propagation.}
\label{fig:graph}
\end{figure}
We can characterize a random graph mathematically in terms
of random edges defined by {\em mixed} random variables,
i.e., random variables with continuous/discrete probability distribution.
To this end, let $g_{ij}(\omega)$ represent the interaction between
the node $i$ and the node $j$, i.e., the flow of information from $i$
to $j$\footnote{For example, the probability density of $g_{ij}$
could be in the form $p(g_{ij})=\delta(0)/2 + U_{[1,2]}(g_{ij})/2$, in which case
we have $50\%$ of chances that $i$ does not influence $j$ at all
(term $\delta(0)/2$), and $50\%$ of chance that $g_{ij}$ is uniformly
distributed in $[1,2]$. If the graph has a time-evolving random
structure, i.e., some nodes in the network have a time-varying set
of neighbors, then we can introduce time-dependence in the random
variables $g_{ij}$. This makes $g_{ij}$ a set of stochastic processes.}.
Such interaction can be of different types, but roughly speaking
it just characterizes how the random
process $\psi_i(t)$ (at node $i$) influences
the random process $\psi_j(t)$ (at node $j$).
The type of influence is defined by a stochastic model
on the random graph, e.g., a system of
stochastic differential equations in the form
\begin{equation}
\frac{d\bm \psi }{dt}= \bm \Lambda(\bm \psi,t;{\bm G}(\omega)) + \bm f(t),
\label{Eq1}
\end{equation}
where $\Lambda$ is a nonlinear map defining the model,
${\bm G}(\omega)=\{g_{12}(\omega),g_{21}(\omega), ...\}\in\mathbb{R}^{n(n-1)/2}$,
is the random vector representing all interactions among the nodes
in the random graph, and $\bm f(t)$ is colored random noise.
The stochastic system \eqref{Eq1} can be obviously generalized to cases
where we have multiplicative noises,
such as in tumororal cell growth models \cite{Zeng,Wang,Fiasconaro}.
The solution to \eqref{Eq1} (assuming it exists) is a nonlinear function
of the random vector ${\bm G}$ defining the graph,
and a nonlinear functional of the stochastic
process $\bm f(t)$. We write such functional
as $\widehat{\bm \psi}(t;{\bm G},[\bm f(t)])$.
The probability density function of $\bm \psi(t)$
then can be obtained by integrating out the graph
and the noise over the corresponding probability distribution,
i.e.,
\begin{equation}
p(\bm \psi,t)= \int \delta\left(\bm \psi-\widehat{\bm \psi}(t;{\bm G},[\bm f])\right)
p({\bm G})P([\bm f])\mathcal{D}[\bm f] d{\bm G}.
\end{equation}
Note that here we assumed that the noise and the graph are
statistically independent. Integration over noise is
(in general) a functional integral. The exact evolution
equation for $p(\bm X,t)$ can be obtained by using a
functional calculus approach \cite{Venturi_PRS,Fox,Hanggi4,Moss1}. This yields,
\begin{equation}
\frac{\partial p(\bm \psi,t)}{\partial t} +
\int \left[\nabla_{\bm \psi}\cdot\left(\bm \Lambda (\bm \psi ,t;\bm G)
p(\bm \psi,\bm {G},t) \right)+ \nabla_{\bm \psi}\cdot \left<\delta\left(\bm \psi-
\widehat{\bm \psi}(t;\bm {G},[\bm f])\right)\bm f(t)\right>_{\bm f}\right]
d\bm G=0,
\label{PDF_ha}
\end{equation}
where $p(\bm \psi,\bm G,t)$ is the joint probability
density of $\bm \psi(t,\omega)$ and $\bm G(\omega)$,
while $\left<\cdot\right>_{\bm f}$ is defined in
\eqref{corr}. As before, the correlation $\left<\delta\left(\bm \psi-
\widehat{\bm \psi}(t;\bm {G},[\bm f])\right)\bm f(t)\right>_{\bm f}$
can be disentangled by using
functional calculus \cite{Bochkov,Venturi_PRS,Klyatskin}, or
computed by using data-driven methods.
}
\begin{table}
{\small
\begin{center}\begin{tabular}{lcl}
Field Equation & & Functional Differential Equation\\
\hline \\
\begin{minipage}{4cm}
\begin{equation}
\frac{\partial \bm u}{\partial t}+(\bm u\cdot \nabla)\bm u=-\nabla p +\nu \Delta \bm u\nonumber
\end{equation}
\end{minipage}
& &
\begin{minipage}{4cm}
\begin{align}
\frac{\partial \Phi([\theta],t)}{\partial t}=
\sum_{k=1}^3\int_V\theta_k(\bm x)\left(i \sum_{j=1}^3\frac{\partial }{\partial x_j}
\frac{\delta^2 \Phi([\bm \theta],t)}{\delta \theta_k(\bm x)\delta\theta_j(\bm x)}
+\nu \nabla^2\frac{\delta \Phi([\bm \theta],t)}{\delta \theta_k(\bm x)}\right)d\bm x
\nonumber
\end{align}
\end{minipage}
\\\\\\
$\displaystyle \frac{\partial u}{\partial t}+
u\frac{\partial u }{\partial x}=\nu \frac{\partial^2u}{\partial x^2}$
& &
$\displaystyle\frac{\partial P([a],t)}{\partial t} = -\int_{-\infty}^{\infty}
\frac{\delta}{\delta a(x)}\left(\left[-a(x)\frac{\partial a(x)}{\partial x}
+\nu\frac{\partial ^2 a(x)}{\partial x^2}\right] P([a],t)\right)dx$
\\\\
$\displaystyle \square \phi+m^2 \phi =
\frac{\lambda}{6}\phi^3$ &\hspace{1.0cm} &
$\displaystyle
\square
\frac{\delta Z([a])}{\delta a(\bm x)}+
m^2 \frac{\delta Z([a])}{\delta a(\bm x)}-
\frac{\lambda}{3!}
\frac{\delta^3 Z([a])}{\delta a(\bm x)^3 }-i a(\bm x) Z([a])=0$
\end{tabular}
\end{center}
}
\caption{Examples of functional differential equations. In this table,
$\Phi$ denotes the Hopf characteristic functional,
$P$ is the probability density functional, and $Z$ is the generating functional of the quantum $\phi^4$-theory.}
\label{tab:2}
\end{table}
\subsection{Conjugate Flow Action Functionals}
In a recent paper \cite{Daniele_JMathPhys}, we have shown how to
construct an action functional for a non-potential
field theory by using methods of differential geometry and nonlinear
functional analysis \cite{Vainberg,Nashed}. The key idea is to
represent the governing equations of the field theory
relative to a diffeomorphic flow of curvilinear coordinates
which is assumed to be functionally dependent on the
field equations, i.e., on their solution.
Such flow evolves in space and time similarly to a
physical fluid flow of classical mechanics
and it can be chosen to symmetrize the G\^ateaux derivative
of the field equations relative to suitable local bilinear forms.
This is equivalent to require that the governing equations
of the field theory can be derived from a principle of
stationary action on a flow, which we called
the {\em conjugate flow of the theory}. The determining
equations of the conjugate flow are functional
differential equations.
In particular, for a second-order nonlinear scalar field theory
\begin{equation}
f\left(u; u_{,\mu}; u,_{\mu\nu}; \widehat{x}^\mu_{,\nu}; \widehat{x}^\mu_{,\nu\lambda} \right)=0,
\label{fieldeq}
\end{equation}
we obtain
\begin{equation}
R^{\mu\nu}_{,\nu}+R^{\mu\nu}\Gamma^\lambda_{\lambda\nu}=Z^\nu,\label{cfaf}
\end{equation}
where the comma denotes differentiation with respect to the
independent variable $\sigma^\nu$,
$\Gamma^\lambda_{\lambda\nu}$ is the Christoffel symbol of the
second kind and
\begin{align}
Z^\nu&=\frac{\partial f}{\partial u_{,\nu}}+
\frac{\partial f}{\partial \widehat{x}^\mu_{,\nu}}\frac{\delta \widehat{x}^\mu}{\delta u}+
\left(\frac{\partial f}{\partial \widehat{x}^\mu_{,\nu\rho}}+
\frac{\partial f}{\partial \widehat{x}^\mu_{,\rho\nu}}\right)
\left(\frac{\partial}{ \partial \sigma^\rho}\frac{\delta \widehat{x}^\mu}{\delta u}+
\frac{\delta^2\widehat{x}^\mu}{\delta u^2}\frac{\partial u}{\partial \sigma^\rho}\right),\label{Bs}\\
R^{\rho\nu}&=\frac{\partial f}{\partial u_{,\nu\rho}}+
\frac{\partial f}{\partial \widehat{x}^\mu_{,\rho\nu}}\frac{\delta \widehat{x}^\mu}{\delta u}. \label{Fs}
\end{align}
Given a solution to the field equation \eqref{fieldeq}, the
system of functional differential equations \eqref{cfaf}
allows us to identify the {\em conjugate flow} $\widehat{x}^\mu(\sigma^\nu;[u])$, i.e., the functional
relation between the flow $\widehat{x}^\mu$ and
the solution $u$ for which the PDE \eqref{fieldeq}
can be derived from a {\em principle of least action}
(see \cite{Daniele_JMathPhys} for further details).
The identification of transformation groups leaving the
conjugate flow action functional invariant could lead to
the discovery of new conservation laws.
\subsection{Large Deviation Theory and Minimum Action Methods}
Large deviations theory deals with the probabilities of rare
events that are exponentially small as a function of some
parameter. To illustrate the theory,
consider a nonlinear PDE perturbed by space-time additive
random noise of small amplitude $\epsilon$
\begin{equation}
\frac{\partial \bm u}{\partial t}+\bm G(\bm u)=\sqrt{\epsilon}\bm f(\bm x,t;\omega).
\end{equation}
We define the set trajectories connecting two arbitrary states $\{\bm u_1,\bm u_2\}$
\begin{equation}
B=\{\bm u(\bm x,t;\omega) | \quad \bm u(\bm x,0;\omega)=\bm u_1,
\quad \bm u(\bm x,T;\omega)=\bm u_2\}
\end{equation}
If $\bm f$ is white noise, then the Freidlin-Wentzell
theory gives us the following large deviation principle
\begin{equation}
\lim_{\epsilon \rightarrow 0^+} \epsilon \textrm{Pr}(\bm u\in B)=\inf_{\bm u\in B}
\frac{1}{2}\int_0^T \left\|\frac{\partial \bm u}{\partial t}+
\bm G(\bm u)\right\|^2_{L_2}dt
\end{equation}
where $\textrm{Pr}(A)$ denotes the probability of the event $A$,
$\left\|\cdot\right\|^2_{L_2}$ indicates the $L_2$ norm in space.
The large deviation principle is equivalent to minimum action
principle
\begin{equation}
\min_{\substack{\bm u(0,\bm x;\omega)=\bm u_1\\\bm u(T,\bm x;\omega)=\bm u_2}}
S_T([\bm u]),\qquad S_T([\bm u])=\frac{1}{2}\int_0^T
\left\|\frac{\partial \bm u}{\partial t}+\bm G(\bm u)\right\|^2_{L_2}dt.\label{DV}
\end{equation}
The minimizer of \eqref{DV} is called {\em minimum action path},
and it satisfies the functional differential equation
\begin{equation}
\frac{\delta S_T([\bm u])}{\delta \bm u(\bm x,t)}=0 \quad \bm u\in B.
\label{DV1}
\end{equation}
The minimum action path is the most probable transition
path from $\bm u_1$ to $\bm u_2$. A method to
solve \eqref{DV1} is based on a direct discretization
of $S_T(\bm u)$ \cite{E1}.
\section{Approximation of Functional Differential Equations}
\label{sec:FDEapprox}
In this Section we address the numerical
approximation of linear FDEs in the
form \eqref{linfde00}. To this end,
we develop a method of weighted
residuals \cite{Finlayson,GKSS_2005}
in the space of functionals that allows us to
derive functional least squares, functional
Galerkin and functional collocation methods
to FDEs is a unified and straightforward way.
\subsection{The Method of Weighted Functional Residuals}
\label{sec:method of weighted residuals}
The method of weighted residuals illustrates how
the choice of different weight (or test) functionals
can be used to construct different classes of methods
extending Galerkin, least-squares and collocation
methods for PDEs to functional differential equations.
The general framework resembles the classical one for PDEs,
in which one minimizes a residual (least-squares method)
or imposes its orthogonality relative to a suitable space of
test functions (Galerkin or collocation methods).
To describe the weighted residuals technique, let us consider
the linear functional differential equation \eqref{linfde00}.
In approximating its solution numerically we are typically
replacing $F([\theta],t)$ with an approximation
\begin{equation}
\widehat F ([\theta],t) \simeq F([\theta],t),
\label{FDE_exp}
\end{equation}
e.g., a tensor canonical tensor decomposition
(Section \ref{sec:CP}) or a Lagrangian
interpolant (Section \ref{sec:Functional Collocation Methods})
with $N$ degrees of freedom. Substitution of
the approximation \eqref{FDE_exp}
into equation \eqref{linfde00}
yields the (functional) residual
\begin{equation}
R([\theta],t)= \frac{\partial \widehat F ([\theta],t)}{\partial t}-L([\theta],t)
\widehat F([\theta],t)-H([\theta],t).
\label{rres}
\end{equation}
At this point, we introduce the following inner
product in the space of functionals\footnote{
The inner product \eqref{rres} is a functional integral,
which is usually defined in terms of a limiting procedure \cite{Popov,Egorov}. From a mathematical viewpoint,
the limiting procedure defining the functional integral
measure in terms of an infinite products of elementary
measures should be handled with care.
In fact, the classical Lebesgue measure does not exist
in spaces ofinfinite dimension \cite{Marzucchi}.
On the other hand, Gaussian measures
are still well defined in such setting. This is why we included
$W([\theta])$ in \eqref{continuous_ip_W0}. The argument leading
to the result on non-existence of an analogue to the Lebesgue
measure in infinite dimension is related to the argument showing
that the Heine-Borel theorem does not hold in infinite-dimensional
normed linear spaces.} (see Appendix \ref{sec:functional integrals})
\begin{equation}
\left(F,G\right)_W=\int
F([\theta])G([\theta])W([\theta])\mathcal{D}[\theta],\qquad \textrm{(functional integral)}
\label{continuous_ip_W0}
\end{equation}
where $W([\theta])$ is a known weight functional, and consider the set of equations
\begin{equation}
\left(R([\theta],t),h_k([\theta])\right)_W=0\qquad k=1,...,N
\label{WR}
\end{equation}
where $h_k([\theta])$ are test functionals.
There is no particular restriction on $h_k([\theta])$.
For example, they can be
cardinal basis functionals, orthogonal polynomial functionals
or other basis functionals.
The system \eqref{WR} allows us to determine the $N$
degrees of freedom in the functional approximation
$\widehat{F}([\theta],t)$. Specifically,
we are imposing that the residual of the FDE is
orthogonal to the span of the functionals
$\{h_1,...,h_N\}$. The nature of the numerical
scheme is determined by the choice of the test
functionals $h_j([\theta])$ in \eqref{WR}.
Evaluating the functional integrals in \eqref{WR}
is challenging, but there are approximation methods
that allow us to compute them.
For example, several algorithms have been recently
proposed for high-dimensional (possibly infinite-dimensional)
integration \cite{Baldeaux,Wasilkowski,Wasilkowski1,Dick,Dick1}
(see also Apendix \ref{sec:functional integrals} and
Chapter 4 in \cite{Egorov}).
The system of equations \eqref{WR} defines
a {\em functional Galerkin method}\footnote{
Stochastic Galerkin methods \cite{Db_book}
are functional Galerkin methods.
Essentially, these approaches are based
on stochastic representations of the solution functional (Section
\ref{sec:Stochastic Functional Methods}), and functional
inner products involving probability measures.
Stochastic Galerkin methods have been studied extensively
in the theory of turbulence \cite{Meecham,Meecham1,Lee,Bodner},
and in uncertainty quantification \cite{db1,Xiao1}.
}.
\subsubsection{Functional Collocation Methods}
\label{sec:collocation}
In this class of methods the test
functionals $h_j([\theta])$ are
chosen to be Dirac delta functionals \cite{Jouvet} centered
at $\theta_j(x)$, i.e., $h_j([\theta])=\delta[\theta_j(x)-\theta(x)]$.
In this setting, the orthogonality
condition \eqref{WR} can be
written as
\begin{equation}
\int R([\theta],t) \delta[\theta_j(x)-\theta(x)]W([\theta]) \mathcal{D}[\theta]= 0\quad \Rightarrow \quad R([\theta_j],t)=0
\end{equation}
In other words, in the functional collocation method we
impose that the residual $R([\theta],t)$ vanishes
at $N$ collocation nodes in $D(F)$,
i.e., $N$ functions $\{\theta_1(x),...,\theta_N(x)\}$.
This yields a system of $N$ equations for the
unknowns $\{\alpha_1(t), ..., \alpha_N(t)\}$.
The solution we obtain from the functional collocation
method obviously interpolates the exact solution at the
nodes $\theta_i(x)$.
\subsubsection{Functional Least Squares}
\label{sec:functional_least_squares}
In this class of methods we look for an
approximate solution functional that minimizes
the norm of the residual $R([\theta],t)$.
Such norm may be defined in terms of the functional inner
product \eqref{continuous_ip_W0}, i.e.,
$\left\|R\right\|_W^2=(R,R)_W$.
In this case, we obtain the following variational
principle involving a functional integral
\begin{equation}
\min_{\widehat{F}\in D_N(F)}\left\|R([\theta],t)\right\|_W^2=
\min_{\widehat{F}\in D_N(F)} \int
R([\theta],t)^2 W([\theta])\mathcal{D}[\theta].
\label{FI}
\end{equation}
The stationary points of \eqref{FI} corresponding to
variations of the degrees of freedom $\alpha_k(t)$
in \eqref{FDE_exp} satisfy the the Euler-Lagrange
equations
\begin{equation}
\left(R([\theta],t),\frac{\partial R([\theta],t)}{\partial \alpha_k(t)}\right)_W=0,\qquad k=1,...,N.
\label{MinR}
\end{equation}
A comparison between \eqref{WR} and \eqref{MinR} suggests that the test
functionals $h_j([\theta])$ in this case are equation-dependent, i.e.,
they depend on the residual $R([\theta],t)$ through the formula
\begin{equation}
h_k([\theta],t) = \frac{\partial R([\theta],t)}{\partial \alpha_k}.
\end{equation}
\vspace{0.7cm}
\noindent
{\em Remark:}
Error analysis, stability and consistency of
functional Galerkin, functional collocation and functional
least squares methods is an {open question}.
\vspace{0.7cm}
\noindent
{\em Remark:} If we restrict $D(F)$ to a finite-dimensional function space, e.g, the span of a finite-dimensional basis, then weighted residual formulation we just discussed reduces to the weighted residual formulation for multivariate linear PDEs.
\subsection{Temporal Discretization}
\label{sec:ADI_SSE}
The FDE \eqref{linfde00} can be discretized in time with
different numerical schemes such as Adams-Bashforth,
Adams-Multon or BDF methods \cite{Quarteroni}.
Such discretization is quite classical in numerical
analysis, and it represents an important building
block in the development of efficient algorithms
to compute the numerical solution to FDEs.
Hereafter we discuss functional tensor
methods built upon explicit and implicit linear multistep
schemes (see \cite{Quarteroni}, p. 497).
As we will see, such algorithms have significant
advantages over other approaches in terms of accuracy
and computational cost.
Given an evenly-spaced sequence of time
instants $t_k= k\Delta t$ ($k=0,1,...$) we write the formal
solution to to the FDE \eqref{linfde00} as
\begin{equation}
F([\theta],t_{n})=F([\theta],t_{n-1})+\int_{t_{n-1}}^{t_n}
\left(L([\theta],\tau)F([\theta],\tau)+H([\theta],\tau)\right)d\tau.
\label{FSint}
\end{equation}
By approximating the temporal integral with a quadrature rule
we obtain a fully discrete time-integration scheme.
For example, if we replace
\begin{equation}
S([\theta],\tau)=L([\theta],\tau)F([\theta],\tau)+H([\theta],\tau)
\label{SintS}
\end{equation}
by the interpolating polynomial at $t_{n-1}$, ..., $t_{n-q}$, extrapolate in $[t_{n-1},t_n]$, and integrate in time
we obtain the $q$-th order Adams-Bashforth scheme.
Hereafter we provide some examples.
{
\color{r}
\subsubsection{Second-order Adams-Bashforth (AB2) method}
We replace \eqref{SintS} with the polynomial interpolating
$S([\theta],t)$ at $\{t_{n-1},t_{n-2}\}$, extrapolate such polynomial to to $[t_{n-1},t_n]$ and compute the integral in \eqref{FSint}.
This yields the second-order explicit explicit scheme
\begin{equation}
F([\theta],t_{n})=F([\theta],t_{n-1})+\frac{\Delta t}{2}
\left[3S([\theta],t_{n-1})-S([\theta],t_{n-2})\right] +
\Delta t \tau_n([\theta]),
\label{FSint1}
\end{equation}
where
\begin{equation}
\tau_n([\theta])=\frac{5\Delta t^3}{12}\frac{\partial^3 F([\theta],t_{n-2})}{\partial t^3}+\mathcal{O}(\Delta t^4).
\end{equation}
The quantity $\tau_n$ is the {\em local truncation error}
at time $t_n$ (\cite{Quarteroni}, p. 499).
Clearly, if the operator $L([\theta],t)$ is time-independent and $H=0$, then \eqref{FSint} has the simpler form
\begin{equation}
F([\theta],t_{n})=F([\theta],t_{n-1})+\frac{\Delta t}{2}
L([\theta])\left[3F([\theta],t_{n-1})-F([\theta],t_{n-2})\right] +
\Delta t \tau_n([\theta]).
\label{FSintti}
\end{equation}
\subsubsection{Crank-Nicolson Method}
The Crank Nicolson method is an implicit
second-order method of the Adams-Multon family.
The scheme can be easily derived by discretizing
the time integral in \eqref{FSint} with the trapezoidal
rule. This yields
\begin{align}
F([\theta],t_n)=F([\theta],t_{n-1}) + \frac{\Delta t}{2}\left[S([\theta],t_{n})+S([\theta],t_{n-1})\right] + \Delta t\tau_n([\theta]),
\label{CN1}
\end{align}
where
\begin{equation}
\tau_n([\theta])=
-\frac{\Delta t^3}{12}\frac{\partial^3 F([\theta],t_{n-1})}{\partial t^3}+\mathcal{O}(\Delta t^4)
\end{equation}
is the local truncation error (\cite{Quarteroni}, p. 499).
The scheme \eqref{CN1} can be rewritten as
\begin{align}
\left[I-\frac{\Delta t}{2} L([\theta],t_n)\right]F([\theta],t_n)=&
\left[I+\frac{\Delta t}{2} L([\theta],t_{n-1})\right] F([\theta],t_{n-1})+
\nonumber\\
&\frac{\Delta t }{2}\left[H([\theta],t_{n})+H([\theta],t_{n-1})\right]+
\Delta t\tau_{n}([\theta]),
\label{CN2}
\end{align}
The integration process proceeds as follows:
Given $F([\theta],t_{n-1})$ we build the right hand side
of \eqref{CN2} and then solve for $F([\theta],t_n)$.
This involves inverting the following (functional differential)
linear operator
\begin{equation}
A([\theta],t_n)=\left[I-\frac{\Delta t}{2} L([\theta],t_n)\right].
\label{An}
\end{equation}
It is convenient to rewrite \eqref{CN2} as
\begin{equation}
A([\theta],t_n)F([\theta],t_n)=E([\theta],t_n)+\Delta t \tau_n([\theta]),
\label{AFE}
\end{equation}
where
\begin{equation}
E([\theta],t_n)=\left[I+\frac{\Delta t}{2} L([\theta],t_{n-1})\right]
F([\theta],t_{n-1})+ \frac{\Delta t }{2}\left[H([\theta],t_{n})+
H([\theta],t_{n-1})\right] .
\end{equation}
is a known functional, provided $F([\theta],t_{n-1})$ is known (solution at time $t_{n-1}=(n-1)\Delta t$).
}
\subsection{Functional Approximation}
The solution to the FDE \eqref{linfde00} can be approximated at each time step by using the functional approximation methods we discussed
in Section \ref{sec:rep-Hopf}. For example, if we restrict the
domain the solution functional $F$ to the finite-dimensional
Hilbert space spanned by the orthonormal
basis $\{\varphi_1,...,\varphi_m\}$, i.e.,
\begin{equation}
D_m=\textrm{span}\{\varphi_1,...,\varphi_m\}\subseteq D(F),
\label{SoF}
\end{equation}
then the functional becomes a multivariate function.
Alternatively, we can look for an approximant of $F$ in the space
of cylindrical functionals. This yields a functional in the form
(see Section \ref{sec:tensor})
\begin{equation}
f(a_1, ..., a_m, t_n)\simeq F([\theta],t_n),\qquad a_k=(\theta,\varphi_k).
\label{functional-approx_cyl}
\end{equation}
For example, in a canonical tensor decomposition setting we have
\begin{equation}
f(a_1, ..., a_m, t_n)\simeq \sum_{l=1}^r \prod_{j=1}^m
G_j^{l}(a_j,t_n),
\qquad G^{l}_{j}(a_j,t_n)=
\sum_{p=1}^P\beta_{jp}^{l}(t_n)\phi_p(a_j),\qquad
\label{ff9}
\end{equation}
where $r$ is the separation rank.
Replacing with $F([\theta],t)$ with $f(a_1,...,a_m,t)$
in \eqref{FSint1} or \eqref{AFE} yields
a functional residual $R([\theta],t_n)$.
For example, a substitution of \eqref{ff9} into
the Crank-Nicolson scheme \eqref{AFE} yields
\begin{equation}
R([\theta],t_{n})= \sum_{l=1}^r A([\theta],t_n)
G^l_1((\theta,\varphi_1),t_n)\cdots G^l_m((\theta,\varphi_m),t_n) -E([\theta],t_n).
\label{RAFE}
\end{equation}
Note that we have incorporated the local truncation
error $\Delta t \tau_n$ within the residual $R([\theta],t_{n})$.
\subsection{CP-ALS Algorithm for FDEs with Implict Time Stepping}
\label{sec:ALS_formulation}
{\color{r}
We have seen in Section \ref{sec:functional_least_squares}
that the functional least squares fomulation of the FDE
\eqref{linfde00} relies on minimizing the norm of the
residual. Such residual can have different forms. In particular, if we discretize the FDE \eqref{linfde00} in time with
the Crank-Nicolson method and represent its solution
by a canonical polyadic (CP) tensor expansion
then the residual takes the form \eqref{RAFE}.
Its norm can be defined as
\begin{equation}
\left\|R([\theta],t_n)\right\|_W^2=
\left\|\sum_{l=1}^r A([\theta],t_n)
G^l_1((\theta,\varphi_1),t_n)\cdots G^l_m((\theta,\varphi_m),t_n) -
E([\theta],t_n)\right\|_W^2,
\label{resnorm}
\end{equation}
where $\|\cdot \|^2_W$ is induced by the functional
inner product \eqref{continuous_ip_W0}.
Recall that the functions $G_{k}^l((\theta,\varphi_k),t_n)$
are in the form
\begin{equation}
G_{k}^l((\theta,\varphi_k),t_n)=\sum_{s=1}^Q
\beta^l_{ks}(t_n)\phi_s((\theta,\varphi_k)),
\end{equation}
$\beta^l_{ks}(t_n)$ ($l=1,...,r$, $k=1,...,m$, $s=1,...,Q$)
being the degrees of freedom.
We look for a minimizer of \eqref{resnorm}
computed in a {\em parsimonious way}.
The key idea is to split the large scale optimization problem
\begin{equation}
\min_{\beta_{ks}^l(t_n)} \left\|R([\theta],t_n)\right\|_W^2
\end{equation}
into a sequence of optimization problems problems of smaller
dimension, which are solved sequentially and eventually in
parallel \cite{Karlsson}. To this end, we define
\begin{equation}
\bm \beta_k(t_n)=[\beta^1_{k1}(t_n), ..., \beta^1_{kQ}(t_n),...,
\beta^r_{k1}(t_n), ..., \beta^r_{kQ}(t_n)]^T \qquad k=1,...,m.
\end{equation}
Note that the vector $\bm \beta_k(t_n)$ collects
the degrees of freedom representing the solution
functional along $(\theta,\varphi_k)$ at time $t_n$, i.e.,
the set of functions $\{G_k^1((\theta,\varphi_k),t_n),...,G^r_k((\theta,\varphi_k),t_n)\}$.
Minimization of \eqref{resnorm} with respect to
{\em independent} variations of $\bm \beta_k(t_n)$
yields the sequence of {\em convex} optimization problems
\begin{equation}
\min_{\bm \beta_1(t_n)} \left\|R([\theta],t_n)\right\|_W^2, \qquad
\min_{\bm \beta_2(t_n)} \left\|R([\theta],t_n)\right\|_W^2, \qquad
\cdots,\qquad
\min_{\bm \beta_m(t_n)} \left\|R([\theta],t_n)\right\|_W^2.
\label{sequential_min}
\end{equation}
This is the set of equations defining the alternating
least-squares (ALS) method.
The Euler-Lagrange equations identifying the stationary points
of \eqref{sequential_min} are
\begin{equation}
\bm M_h(t_n) \bm \beta_h(t_n) = \bm f_h(t_n), \qquad h=1,...,m
\label{ADIS}
\end{equation}
where,
\begin{equation}
[\bm M_h(t_n)]_{qs}^{zl}=\int
Q_{qh}^z([\theta],t_n)
Q_{sq}^l([\theta],t_n)
W([\theta])\mathcal{D}[\theta],
\label{Mn1}
\end{equation}
\begin{equation}
[\bm f_h(t_n)]^{z}_{q}=\int E([\theta],t_n)
Q_{qh}^z([\theta],t_n)W([\theta])D[\theta],
\label{kn1}
\end{equation}
\begin{equation}
Q_{qh}^z([\theta],t_n)= A([\theta],t_n)\phi_q((\theta,\varphi_h))
\prod_{\substack{j=1\\j\neq h}}^m G_j^z((\theta,\varphi_j),t_n).
\label{QQQ}
\end{equation}
The ordering of the matrix elements $[\bm M_h(t_n)]_{qs}^{zl}$
and the vector $[\bm f_h(t_n)]^{z}_{q}$ is the same as in
\eqref{ORDERING}.
Note that minimizing the norm of the residual \eqref{resnorm} with
respect $\bm \beta_h(t_n)$ is equivalent
to impose orthogonality of \eqref{RAFE} with respect
to the space spanned by the basis functionals
$Q_{qh}^z([\theta],t_n)$.
Indeed, the system \eqref{ADIS} is equivalent to
\begin{equation}
\left(R([\theta],t_n),Q_{qh}^z([\theta],t_n)\right)_W=0
\qquad \textrm{(fixed $h=1,...,m$),}
\label{ALES}
\end{equation}
where $(,)_W$ is the functional inner
product \eqref{continuous_ip_W0}.
The system of equations \eqref{ADIS} is
symmetric and positive definite
\footnote{\color{r}The alternating least-squares formulation
\eqref{ADIS} is very similar to alternating least squares
formulation for nonlinear functionals we studied
in Section \ref{sec:CP}. The main difference is that
here we are trying to determine the canonical tensor decomposition
of a linearly mapped functional, i.e., we solving the
linear system $A([\theta],t_n)F([\theta],t)=E([\theta],t_n)$
with ALS, $F$ being represented as a
canonical tensor decomposition. On the other hand,
in Section \ref{sec:CP} we addressed the problem of
computing the canonical tensor decomposition of $F([\theta])$
given $E([\theta])$ (compare the residual \eqref{RAFE} with \eqref{Fresidual}). These two problems are equivalent
if the linear operator $A$ is invertible.}.
This allows us to use well-known high-performance algorithms
to compute the solution, e.g., the conjugate gradient
method (\cite{Quarteroni}, p.152).
It is important to emphasize that the minimization of the residual
\eqref{resnorm} is basically a {\em fixed point problem}
involving the vector
\begin{equation}
\bm \beta(t_n)=[\bm \beta_1(t_n)\quad \cdots \quad \bm\beta_m(t_n)]
\end{equation}
The ALS method aims at solving such fixed point
problem by splitting it into a sequence of
linear problems \eqref{ADIS} which are solved iteratively
for each $h$. The criterion is to freeze
all $\bm \beta_j(t_n)$ ($j=1,...,m$,
$j\neq h$) when solving for $\bm \beta_h(t_n)$.
We recall that convergence of CP-ALS iterations is, in general,
not granted (see Section \ref{sec:CP}). To overcome
this problem, additional regularization terms may be necessary
\cite{Acar,Battaglino}.
\paragraph{Evaluation of the Functional Integrals}
The ALS coefficients \eqref{Mn1} and \eqref{kn1}
are defined by functional integrals involving cylindrical functionals.
The computation of such integrals is briefly addressed in
Appendix \ref{app:change of variables in SSE}. Hereafter
we summarize the main results. We first restrict
$\theta(x)$ to the space of functions
\eqref{SoF}, i.e., we assume that $\theta(x)$ can be written
as
\begin{equation}
{\theta}(x)=\sum_{k=1}^m a_k \varphi_k(x),\qquad a_k= (\theta,\varphi_k).
\label{KOE}
\end{equation}
In this hypothesis, the basis functionals \eqref{QQQ}
become multivariate functions of $(a_1,...,a_m)$, i.e.,
\begin{equation}
Q^z_{qh}(a_1,....,a_m,t_n)=A(a_1,...,a_m,t_n)\phi_q(a_h)\prod_{\substack{j=1\\j\neq h}}^m
G_j^z(a_j,t_n).
\label{QQ}
\end{equation}
Similarly, the coefficients \eqref{Mn1} and \eqref{kn1}
can be explicitly written as multivariate
integrals\footnote{\color{r} In equations
\eqref{Mn1f} and \eqref{kn1f} we have we set the
weight function $W(a_1,...,a_n)$ equal to one. This is always
possible provided the support of the integrands is compact
(see Appendix \ref{app:change of variables in SSE}).}
\begin{align}
[\bm M_h(t_n)]_{qs}^{zl}=\int_{-b}^b\cdots\int_{-b}^b
Q^z_{qh}(a_1,....,a_m,t_n) Q^l_{sq}(a_1,...,a_m,t_n)da_1\cdots da_m,
\label{Mn1f}
\end{align}
\begin{equation}
[\bm f_h(t_n)]^{z}_{q}=\int_{-b}^b\cdots\int_{-b}^b E(a_1,...,a_m,t_n)
Q^z_{qh}(a_1,..,a_m,t_n) da_1\cdots da_m,
\label{kn1f}
\end{equation}
The quantity $A(a_1,...,a_m,t_n)$ appearing in \eqref{QQ}
is the discrete version of the functional differential operator
$A([\theta],t_n)$, i.e., it is a linear operator in the form
\begin{equation}
A(a_1,...,a_m,t_n)=I-\frac{\Delta t}{2} L(a_1,...,a_m,t_n)
\end{equation}
where $L(a_1,...,a_m,t_n)$ is the finite-dimensional
version of the operator $L([\theta],t_n)$ in \eqref{linfde00}.
In general, the integrals \eqref{Mn1f} and \eqref{kn1f} can
be computed numerically only for a relatively small number of
variables $a_k$ ($k=1,...,m$). However, if we
assume that the operator $A(a_1,...,a_m,t_n)$ is {\em separable}, i.e.,
\begin{equation}
A(a_1,...,a_m,t_n)=\sum_{k=1}^{r_A} A_1^k(a_1,t_n)\cdots A_m^k(a_m,t_n),
\label{A_}
\end{equation}
then the cost of computing such integrals scales {\em linearly} with the
dimension $m$ of the space, since
\eqref{Mn1f} and \eqref{kn1f} can be factored as
products of one-dimensional integrals.
In equation \eqref{A_}, $A_i^k(a_i,t_n)$ are
one-dimensional linear operators, while $r_A$ is the separation
rank of the operator $A(a_1,...,a_m,t_n)$.
%
With the operator decomposition \eqref{A_} available,
we can represent the multivariate fields \eqref{QQQ} as
\begin{equation}
Q_{qs}^l(a_1,...,a_m,t_n)=\sum_{k=1}^{r_A} A_q^k(a_q,t_n)\phi_s(a_q)\prod_{\substack{j=1\\j\neq q}}^m
A_j^k(a_j,t_n)G_j^{l}(a_j,t_n).\label{qqqs}
\end{equation}
This yields the following representation of the matrix coefficients \eqref{Mn1f}
\begin{align}
\left[\bm M_q(t_n)\right]_{sh}^{lz}=&
\int_{-b}^b\cdots \int_{-b}^b Q_{qs}^l(a_1,...,a_m,t_n)
Q_{qh}^z(a_1,...,a_m,t_n)da_1\cdots da_m\nonumber\\
=& \sum_{k,e=1}^{r_A}\int_{-b}^b
A_q^k(a_q,t_n)\phi_s(a_q) A_q^e(a_q,t_n)\phi_h(a_q) da_q \times\nonumber \\
& \prod_{\substack{j=1\\j\neq q}}^m\int_{-b}^b
A_j^k(a_j,t_n)G_j^{l}(a_j,t_n)A_j^e(a_j,t_n)G_j^{z}(a_j,t_n)da_j.
\label{integrals_9}
\end{align}
i.e., sums of products of {one-dimensional
integrals}\footnote{Note that we can express all integrals
at the right hand side of \eqref{integrals_9} in terms of the integrals
\begin{equation}
\int_{-b}^b A_q^k(a_q,t_n)\phi_s(a_q)A_q^e(a_q,t_n)\phi_h(a_q)da_q
\end{equation}
where $q=1,...,m$, $k,e=1,...,r_A$, $s,h=1,...,Q$. In fact,
\begin{equation}
\int_{-b}^b A_j^k(a_j,t_n)G_j(a_j,t_n) A_j^e(a_j,t_n)G^z_j(a_j,t_n) da_j = \sum_{s,h=1}^Q \beta_{js}^l(t_n) \beta_{jh}^z(t_n)
\int_{-b}^b A_j^k(a_j,t_n)\phi_s(a_j)A_j^e(a_j,t_n)\phi_h(a_j)da_j.
\end{equation}
}.
Let us provide a simple example.
\vspace{0.7cm}
\noindent
{\em Example 1:} Consider the time-independent
functional differential operator \eqref{FDOP} and the
associated operator defined in \eqref{An}
\begin{equation}
A([\theta])=I+\frac{\Delta t}{2}\int_{0}^{2\pi} \theta(x)
\frac{\partial }{\partial x}\frac{\delta }{\delta \theta(x)} dx.
\end{equation}
Evaluating $A([\theta])$ in the finite-dimensional function space
\eqref{SoF} yields
\begin{equation}
A(a_1,...,a_m)=I+\frac{\Delta t}{2}\sum_{k=1}^m \left(\sum_{j=1}^m a_j
\int_{0}^{2\pi}\varphi_k(x)\frac{d\varphi_j(x) }{dx}dx\right)
\frac{\partial }{\partial a_k}.
\end{equation}
Therefore, $A(a_1,...,a_m)$ is a separable operator with separation rank $r_A=m^2+1$.
\subsubsection{Collocation Setting}
Consider the sequence of linear systems \eqref{ADIS}, and let
$\{\phi_s(a_k)\}$ be a cardinal basis associated with the
set of collocation nodes $\{a_{k1},...,a_{kQ}\}$.
For simplicity, we consider the same set of nodes in each dimension.
In this assumption, the integrals defining the the matrix entries
\eqref{integrals_9} can be significantly simplified. For example,
\begin{equation}
\int_{-b}^b A_q^k(a)\phi_s(a) A_q^e(a)\phi_h(a)da \simeq
\underbrace{\bm A_q^k \bm W \bm A_q^e}_{\bm K_q^{ke}}
\label{eq:50}
\end{equation}
where $\bm A_q^k$ is the matrix representation of the
operator $A_q^k(a)$ (collocation version), and $\bm W$
is a diagonal matrix of integration weights.
The matrix $\bm K_q^{ke}$ is $Q\times Q$ for all
$k,e=1,...,r_A$ and all $q=1,...,m$. With $\bm K_q^{ke}$
available, it is easy to determine the matrix representation of the
integral
\begin{equation}
\int_{-b}^b A_q^k(a)G^l_j(a) A_q^e(a)G_j^z(a)da=
\sum_{s,h=1}^Q \beta_{js}^l \beta_{jh}^z \int_{-b}^b A_q^k(a)\phi_s(a) A_q^e(a)\phi_h(a)da \simeq
\bm \beta_j^T \bm K_q^{ke} \bm \beta_j
\end{equation}
where $\bm \beta_j$ here is a matrix that has $G_j^z(x_{jp})$
($p=1,..,Q$) as $z$-th column.
This yields the following matrix
\begin{equation}
\bm M_q = \sum_{k,e=1}^{r_A} [\bm R^{ke}_q]^T\otimes \bm K_q^{ke}.
\end{equation}
where
\begin{equation}
\bm R^{ke}_q=\prod_{\substack{j=1\\
j\neq q}}^m \bm \beta_j^T \bm K_q^{ke}\bm \beta_j.
\end{equation}
\subsection{Tensor Formats for FDEs with Explicit Time Stepping}
Consider the following finite-dimensional form of the linear
FDE \eqref{linfde00}
\begin{equation}
\frac{\partial f}{\partial t}=L f +h,
\label{PDEL}
\end{equation}
where $f(a_1,..,a_m,t)$ is a {\em tensor format} that approximates
a functional $F([\theta],t)$ and $L(a_1,...,a_m,t)$ is the
linear operator arising from the discretization of the functional
linear operator $L([\theta],t)$, and $h(a_1,...,a_m,t)$ is the
tensor format that appoximates $H([\theta],t)$.
In particular, consider the case where $h=0$, $L$ is time-independent
and time-integration follows the Adams-Bashforth scheme
\eqref{FSintti}, i.e.,
\begin{equation}
f_{n} = f_{n-1}+\frac{\Delta t}{2} L\left(3f_{n-1}-f_{n-2}\right).
\end{equation}
In the last equation we employed the shorthand notation
$f_n=f(a_1,...,a_m,t_n)$. Assuming, that the
operator $L$ is separable, e.g.,
\begin{equation}
L = -\sum_{i,j=1}^m a_iC_{ij}\frac{\partial }{\partial a_j},
\end{equation}
then a greedy computation of the
tensor format $f_{n}$ involves the following steps:
\begin{enumerate}
\item compute a low rank representation of $\hat{f}_{n-1}=3f_{n-1}/2 - f_{n-2}/2$,
\item compute a low rank representation of
$f_{n}=p_{n-1} +\Delta t L \hat{f}_{n-1}$.
\end{enumerate}
The need for a low rank representation is clear:
any algebraic operation between tensors, including the
application of a linear operator, increases the separation
rank. Therefore, efficient {\em rank reduction methods} are needed to
avoid an explosion of the number of terms when solving
the PDE \eqref{PDEL} with tensor methods.
Among them we recall methods based on alternating least
squares \cite{Karlsson,Battaglino,Etter},
hierarchical Tucker formats \cite{Grasedyck2015,Kolda}
or block coordinate descent methods \cite{Xu}. Disregarding the
particular tensor format employed to represent the solution
functional, we emphasize that the development of rubust and
efficient rank-reduction algorithms is an active area
of research \cite{Bachmayr,NouyHUQ}.
}
\section{Numerical Results: Functionals}
\label{sec:numerical results functionals}
In this Section we provide numerical results and examples on
functional approximation. In particular, we discuss
polynomial functional interpolants and functional
tensor methods.
\subsection{Linear functionals}
\label{sec:results linear functionals}
Consider the linear functional
\begin{align}
F([\theta])=\int^{2\pi}_0 K_1(x) \theta(x)dx
\label{linF}
\end{align}
on the Hilbert space of square integrable periodic functions
in $[0,2\pi]$
\begin{equation}
D(F)=\{\theta\in L_2([0,2\pi])\,|\, \theta(0)=\theta(2\pi)\}.
\label{thespace}
\end{equation}
Our aim is to represent $F([\theta])$ in terms of a functional
interpolant in $D(F)$, i.e.,
\begin{equation}
F([\theta])= \sum_{k=1}^\infty F([\theta_i])g_i([\theta]).
\end{equation}
where $g_i([\theta])$ are cardinal basis functionals
and $\theta_i(x)$ are interpolation nodes in $D(F)$.
In particular, we choose $\theta_i(x)=\varphi_i(x)$
where $\{\varphi_1(x),\varphi_2(x),...\}$ is an orthonormal
basis in $D(F)$.
Assuming that $K_1(x)$ is in $D(F)$, i.e.,
\begin{equation}
K_1(x)=\sum_{k=1}^\infty (K_1,\varphi_k)\varphi_k(x),
\end{equation}
it follows from \eqref{linF} that
\begin{align}
F([\theta])=&\sum_{k=1}^\infty (K_1,\varphi_k)(\varphi_k,\theta),\nonumber\\
=&\sum_{k=1}^\infty F([\varphi_k])(\varphi_k,\theta).
\label{exactF}
\end{align}
This can be written as
\begin{align}
F([\theta])=\sum_{k=1}^\infty F([\varphi_k])g_k([\theta]), \qquad\textrm{where}\qquad
g_k([\theta])=(\varphi_k,\theta).
\label{seriesLF}
\end{align}
Note that this representation coincides with Porter's series
expansion \eqref{Porter_interpolant}-\eqref{gi_0} on the index
set $\mathcal{I}=\{1\}$.
\subsubsection{Polynomial Functional Interpolation}
Let us study numerically an interpolation problem involving
a specific kernel. To this end, we set
\begin{equation}
K_1(x)=e^{\sin(x)}(1+\sin(\cos(x)-2)-\frac{1}{2}\qquad \textrm{(Fig. \ref{fig:kernel1})},
\label{K1}
\end{equation}
\begin{figure}
\centerline{\hspace{0.5cm}(a)\hspace{7cm}(b)}
\centerline{\includegraphics[height=5.5cm]{kernel_f-eps-converted-to.pdf}
\includegraphics[height=5.5cm]{fourier_basis-eps-converted-to.pdf}}
\caption{(a) Kernel function defining the linear functional \eqref{linF}. (b)
Three elements of the orthogonal basis function set \eqref{setbf} ($m=20$).}
\label{fig:kernel1}
\end{figure}
and define the following interpolation nodes in $D(F)$
\begin{equation}
\varphi_{k+1}(x)=\frac{1}{m+1}\displaystyle\frac{\sin\left(\displaystyle\frac{m+1}{2}\left(x-x_k\right)\right)}
{\sin\left(\displaystyle\frac{x-x_k }{2} \right)},\qquad x_k=\frac{2\pi}{m+1}k\qquad k=0,1,...,m
\label{setbf}
\end{equation}
$m$ being an even natural number. These are well-known
nodal trigonometric polynomials (\cite{Hesthaven}, p. 28)
satisfying the orthogonality conditions
\begin{equation}
(\varphi_j,\varphi_k)=\frac{2\pi}{m+1}\delta_{kj}, \qquad k,j=1,..., m+1.
\end{equation}
Clearly, if $K_1(x)$ is in the span of
$\{\varphi_1,...,\varphi_{m+1}\}$ then the exact
representation of the linear functional $F([\theta])$
involves no more than $m+1$ terms, i.e., a truncation
of the series \eqref{seriesLF} to $m+1$ terms is exact.
The kernel \eqref{K1} is not in such span.
Next, we construct a polynomial functional
that interpolates $F([\theta])$ at the nodes \eqref{setbf}.
Specifically we consider Porter's construction
(Section \ref{sec:porter}), which is very easy to implement and
equivalent to Khlobystov polynomials for uniquely solvable
interpolation problems. Such interpolants are dense
in the space of linear functionals if we consider the
set of nodes \eqref{SNq1}, i.e.,
\begin{equation}
\widehat{S}^{(m)}_1=\left\{\theta_i(x)\in D(F)\,\,|\,\,
\theta_i(x)=\varphi_{i}(x) \quad i=0,...,m\right\}.
\end{equation}
This means that to identify linear functionals
it is sufficient to represent them relative to orthogonal bases
(see Section \ref{sec:variational}). As we have noticed
in Section \ref{sec:functional_polynomial_interpolation},
this is not the case for higher-order polynomial functionals
or general nonlinear functionals.
Now, let us consider the general expression of Porter's interpolants
\begin{equation}
F\left([\theta]\right)\simeq\sum_{k=1}^{m+1} F\left([\varphi_k]\right) g_{k}([\theta]),\qquad
g_{k}([\theta])=\sum_{j=1}^{m+1} H_{jk}^{-1}
\sum_{p\in\mathcal{I}}\left(\varphi_j,\theta\right)^p,
\end{equation}
where the matrix $H_{ij}$ is given in \eqref{matH}.
Depending on how we choose the index set $\mathcal{I}$ we
have different expressions for the basis functionals $g_i([\theta])$.
Specifically,
\begin{enumerate}
\item Constant polynomial functionals ($\mathcal{I}=\{0\}$).
This case is degenerate and it requires Moore-Penrose pseudo-inversion of
the matrix \eqref{matH}. This yields the basis functionals
\begin{equation}
g_k([\theta])=\frac{1}{m+1}.
\end{equation}
\item Homogeneous polynomial functionals of first order ($\mathcal{I}=\{1\}$):
\begin{equation}
g_k([\theta])= \left[\frac{2\pi}{m+1}\right]^{-1} \left(\theta,\varphi_k\right).
\label{linPorter}
\end{equation}
\item Quadratic polynomial functionals ($\mathcal{I}=\{1,2\}$) on the basis
set $\left\{\varphi_1,...,\varphi_{m+1}\right\}$. This yields basis
functionals
\begin{equation}
g_k([\theta]) = \left[\left(\frac{2\pi}{m+1}\right)+
\left(\frac{2\pi}{m+1}\right)^2\right]^{-1}
\left(\left(\theta,\varphi_k\right)+\left(\theta,\varphi_k\right)^2\right).
\label{quadPorter}
\end{equation}
\end{enumerate}
In Figure \ref{fig:results_linear} we plot the pointwise error
\begin{equation}
E_m= \sup_{\theta\in \mathcal{G}_q}\left|F([\theta])-
\sum_{k=1}^{m+1}F([\varphi_k])g_k([\theta])\right|
\label{Em}
\end{equation}
versus $m$ for test functions $\theta$ in the Gaussian ensemble
\begin{equation}
\mathcal{G}_q(\sigma)=\left\{\theta(x)\in D(F)\,\left|
\, \theta(x)=\sigma\sum_{k=1}^{q+1}b_k\varphi_k(x),\quad
\{b_1,...,b_{m+1}\}\quad \textrm{i.i.d. Gaussian}\right.\right\},
\label{Gq}
\end{equation}
and interpolants corresponding to the index sets $\mathcal{I}=\{1\}$ and
$\mathcal{I}=\{1,2\}$ (bases
\eqref{linPorter} and \eqref{quadPorter}).
\begin{figure}
\centerline{\hspace{0.5cm}(a)\hspace{8cm}(b)}
\centerline{\includegraphics[height=6.5cm]{error_porter_linear-eps-converted-to.pdf}
\includegraphics[height=6.5cm]{error_porter_linear_1-eps-converted-to.pdf}}
\caption{Linear functionals. Functional interpolation errors obtained by
using Porter's method. Shown are the pointwise
errors \eqref{Em} versus $m$ for the Gaussian ensemble
$\mathcal{G}_{300}(1)$ and Porter's basis functionals
\eqref{linPorter} ($\mathcal{I}=\{1\}$), and
\eqref{quadPorter} ($\mathcal{I}=\{1,2\}$).
The expansion corresponding to $\mathcal{I}=\{1\}$ converges
exponentially for obvious reasons, while the expansion corresponding
to $\mathcal{I}=\{1,2\}$ has a $O(m^{-1/2})$ convergence rate
due to insufficient interpolation nodes (Fig. (b)).}
\label{fig:results_linear}
\end{figure}
The reason why we obtain exponential convergence with the index set
$\mathcal{I}=\{1\}$ is obvious: convergence of the
polynomial interpolant is basically defined by the
convergence of the trigonometric series of $K_1(x)$.
On the other hand, Porter's interpolant corresponding
to $\mathcal{I}={1,2}$ shows an algebraic convergence at rate
$O(m^{-1/2})$. The reason is that the basis
function set $\{\varphi_1,...,\varphi_{m+1}\}$
does not have enough elements to correctly identify
the quadratic part of the interpolant, which is zero in this case.
In fact, the off-diagonal terms in \eqref{sof} or \eqref{ssof}
cannot be identified by using an orthogonal basis,
unless we consider a set of nodes in the form
$\left\{\varphi_i, (\varphi_{i}+\varphi_j)\right\}$
where $i,j=1,...,m+1$, and $j\geq i$.
\subsubsection{Canonical Tensor Decomposition}
We look for a representation of \eqref{linF}
in the form \eqref{functional-SSE}.
The basis functions $G_i^l$ can be equivalently determined
in an alternating Galerkin or least squares setting
by solving the system of equations \eqref{SSYSTEM}
with forcing given by
\begin{equation}
f^n_{jh}=\int_{-b}^b\cdots\int_{-b}^b
\sum_{p=1}^{m+1} k_p a_p
\phi_h(a_j)\prod_{\substack{k=1\\k\neq j}}^{m+1} G^n_k(a_k)da_1\cdots da_{m+1},
\label{fn3}
\end{equation}
\begin{equation}
k_p=\int_{0}^{2\pi} K_1(x)\varphi_p(x)dx.
\end{equation}
All integrals in \eqref{fn3} can be reduced to
products of one-dimensional integrals (linear functionals are
fully separable). In addition, if we use a polynomial basis
for $G^l_i$, then we can represent \eqref{linF} exactly
as a product of constants and
linear polynomials. In fact,
\begin{align}
\sum_{l=1}^rG^l_1(a_1)\cdots G^l_{m+1}(a_{m+1})
=\sum_{l=1}^{m+1} k_l a_l,
\end{align}
which means that the exact separation rank is $r={m+1}$,
\begin{equation}
G^l_l(a_l)=a_l \qquad \textrm{and}\qquad
\prod_{\substack{k=1\\k\neq l}}^{m+1} G_k^l(a_k)=k_l.
\end{equation}
\subsection{Quadratic Functionals}
\begin{figure}[t]
\centerline{$H_2(x,y)$\hspace{5.5cm}$K_2(x,y)$}
\centerline{\includegraphics[height=5cm]{kernel2_f-eps-converted-to.pdf}
\includegraphics[height=5cm]{kernel2_f_sym-eps-converted-to.pdf}}
\caption{Quadratic functional \eqref{nonlinF}. Kernel function \eqref{K2} and its
symmetrized version \eqref{K2sym}. Integrating $H_2(x,y)\theta(y)\theta(x)$
or $K_2(x,y)\theta(y)\theta(x)$ over $x$ and $y$ produces
exactly the same result, independently on $\theta$.}
\label{fig:kernel2}
\end{figure}
Consider the quadratic functional
\begin{align}
F([\theta]) = 5 + \int^{2\pi}_0 K_1(x) \theta(x)dx+
\int^{2\pi}_0\int^{2\pi}_0 H_2(x,y)\theta(x)\theta(y)dxdy
\label{nonlinF}
\end{align}
defined on the space of square integrable periodic functions \eqref{thespace}.
We set $K_1(x)$ as in \eqref{K1} and $H_2(x,y)$ as
\begin{equation}
H_2(x,y)=\sin(\cos(x)+\sin(y))\sin(y)+\frac{1}{2}\cos(\cos(x)).
\label{K2}
\end{equation}
Replacing $H_2(x,y)$ with the symmetrized kernel
(see Figure \ref{fig:kernel2})
\begin{equation}
K_2(x,y)=\frac{1}{2}\left(H_2(x,y)+H_2(y,x)\right)
\label{K2sym}
\end{equation}
does not change the functional \eqref{nonlinF}.
\subsubsection{Polynomial Functional Interpolation}
We first show that Porter's interpolants corresponding
to the index set $\mathcal{I}=\{0,1,2\}$ and Khlobystov
interpolants \eqref{p2interp} coincide when the interpolation
problem is uniquely solvable, e.g., when we consider the set
of nodes \eqref{SNq2}, hereafter rewritten for convenience
\begin{equation}
\widehat{S}^{(m+1)}_2 =
\left\{0,\{\varphi_i\}_{i=1}^{m+1},\{\varphi_i+\varphi_j\}_{j\geq i=1}^{m+1}\right\}.
\label{test_f_s1}
\end{equation}
In Figure \ref{fig:Khlobystov_basis} we plot the elements of such set
for $m=10$.
\begin{figure}[t]
\centerline{\includegraphics[height=5cm]{K_basis_1-eps-converted-to.pdf}
\includegraphics[height=5cm]{K_basis_2-eps-converted-to.pdf}}
\centerline{\includegraphics[height=5cm]{K_basis_3-eps-converted-to.pdf}
\includegraphics[height=5cm]{K_basis_4-eps-converted-to.pdf}}
\caption{Interpolation nodes in the space of periodic functions (each function is a node).
Here we plot the elements in the sets $\widehat{S}^{(m+1)}_p$ (see equation \eqref{SNq2})
for $m=10$ and $p=1,2,3,4$. The total number of elements within each
set is $\#\widehat{S}^{(11)}_1=12$, $\#\widehat{S}^{(11)}_2=78$,
$\#\widehat{S}^{(11)}_3=364$ and $\#\widehat{S}^{(11)}_4=1365$.}
\label{fig:Khlobystov_basis}
\end{figure}
In Figure \ref{fig:results_quadratic} we plot the pointwise error
\begin{equation}
E_m= \sup_{\theta\in \mathcal{G}_{50}(1)}\left|F([\theta])-\Pi_2([\theta])\right|
\label{Em2}
\end{equation}
versus $m$ for functions $\theta(x)$ in the Gaussian ensemble
$\mathcal{G}_{50}(1)$ (see equation \eqref{Gq}). In \eqref{Em2}, $\Pi_2$ represents either
Porter's or Khlobystov interpolant at the set of nodes \eqref{test_f_s1}.
\begin{figure}[t]
\centerline{\includegraphics[height=6.5cm]{error_porter_khlobystov_quadratic-eps-converted-to.pdf}}
\caption{Interpolation of the quadratic functional \eqref{nonlinF}
using Porter and Khlobystov approaches.
Shown are the pointwise errors \eqref{Em2} versus $m$ for the Gaussian ensemble
$\mathcal{G}_{50}(1)$. The small deviation between Porter's and Khlobystov's error plots
observed at $m=24$ is due to inaccuracies in the computation of the inverse of \eqref{matH}.
The number of interpolation nodes required to achieve
accuracy of about $10^{-7}$ ($m=20$) is
$\#\widehat{S}^{(21)}_2=253$ (see equation \ref{SCardinality}).}
\label{fig:results_quadratic}
\end{figure}
It is seen that both methods achieve exponential convergence rate
when interpolating the polynomial functional \eqref{nonlinF}.
This is due to the fact that such functional expansions are
basically approximating the kernel functions
\eqref{K1} and \eqref{K2} in terms of a Fourier spectral basis.
This also establishes the full equivalence of Porter's and Khlobystov's
interpolants for uniquely solvable interpolation problems.
{\color{r}
\subsubsection{Canonical Tensor Decomposition}
Evaluation of \eqref{nonlinF} in the finite-dimensional
function space $D_m=\textrm{span}\{\varphi_1,...,\varphi_{m+1}\}$
yields the multivariate function
\begin{equation}
f(a_1,...,a_m) = 5+\sum_{p=1}^{m+1} k_p a_p+ \sum_{q,p=1}^{m+1}
q_{pq} a_pa_q,
\label{fdimQ}
\end{equation}
where $a_j=(\varphi_j,\theta)$ and
\begin{equation}
k_p= \int_{0}^{2\pi} K_1(x)\varphi_p(x)dx,\qquad
q_{pq}= \int_{0}^{2\pi}\int_{0}^{2\pi} K_2(x,y)\varphi_p(x)\varphi_q(x)dxdy.
\label{cfoQ}
\end{equation}
The forcing term at the right hand side of
\eqref{SSYSTEM} can be written as
\begin{equation}
f^n_{jh}=\int_{-b}^b\cdots\int_{-b}^b
\left(5+\sum_{p=1}^{m+1} k_p a_p+ \sum_{q,p=1}^{m+1}
q_{pq} a_pa_q\right)
\phi_h(a_j)\prod_{\substack{k=1\\k\neq j}}^{m+1} G^n_k(a_k)da_1\cdots da_m,
\label{fn4}
\end{equation}
As before, the integrals \eqref{fn4} can be reduced to
products of one-dimensional integrals. If we use a Legendre basis
$\phi_h$, then we can represent \eqref{nonlinF}
exactly as a product of constants, linear and quadratic
polynomials. To show this, let us consider the
quadratic part of the functional \eqref{nonlinF}.
We have,
\begin{align}
\sum_{l=1}^r G^l_1(a_1)\cdots G^l_{m+1}(a_{m+1})
=\sum_{q,p=1}^{m+1} q_{qp} a_qa_p.
\end{align}
Given the symmetry of $q_{pq}$, the separation
rank for such quadratic part is $(m+1)(m+2)/2$.
More generally, the function \eqref{fdimQ} is separable,
with separation rank $r=1+(m+1)+(m+1)(m+2)/2$. In fact,
it can be written in the form
\begin{equation}
f(a_1,...,a_m)=\sum_{l=1}^r \alpha_l G^l_1(a_1)\cdots G^l_m(a_m).
\label{CTDQ}
\end{equation}
A possible ordering of the series could be the one in Table
\ref{tab:ordering_functional}.
\begin{table}
\color{r}
\centering
\begin{tabular}{c|ccccc}
$q$ & $\alpha_l$ & $G_1^l$ & $G_2^l$ & $\cdots$ & $G_m^l$\\
\hline\\
$1$ & $5$ & $1$ & $1$ & $\cdots$& $1$\\
$2$ & $k_1$ & $a_1$ & $1$ & $\cdots$& $1$\\
$3$ & $k_2$ & $1$ & $a_2$ & $\cdots$& $1$\\
$\vdots $& $\vdots $ & $\vdots$ & $\vdots$ &$\vdots$ &$\vdots$ \\
$m+2 $& $k_{m+1}$ & $1$ & $1$ &$\cdots$ &$a_{m+1}$ \\
$m+3 $& $q_{11}$ & $a_1^2$ & $1$ &$\cdots$ &$1$ \\
$m+4 $& $q_{12}+q_{21}$ & $a_1$ & $a_2$ &$\cdots$ &$1$ \\
$\vdots $& $\vdots $ & $\vdots$ & $\vdots$ &$\vdots$ &$\vdots$ \\
$2m+4 $& $q_{1(m+1)}+q_{(m+1)1}$ & $a_1$ & $1$ &$\cdots$ &$a_{m+1}$\\
$2m+5 $& $q_{22}$ & $1$ & $a_2^2$ &$\cdots$ &$1$\\
$\vdots $& $\vdots $ & $\vdots$ & $\vdots$ &$\vdots$ &$\vdots$ \\
$m+2+(m+1)(m+2)/2 $& $q_{(m+1)(m+1)}$ & $1$ & $1$ &$\cdots$ &$a_{m+1}^2$\\
\end{tabular}
\caption{Ordering of the terms in the canonical tensor
expansion \eqref{CTDQ} of the quadratic
functional \eqref{nonlinF}.}
\label{tab:ordering_functional}
\end{table}
In Figure \ref{fig:HT quadratic} we study the accuracy
of the canonical tensor decomposition
in representing the quadratic functional \eqref{nonlinF}. Specifically,
we plot the relative pointwise error
\begin{equation}
e_m=\left|\frac{F([\theta])-f(a_1,...,a_m)}{F([\theta])}\right|\qquad
a_j=(\theta,\varphi_j)
\label{ptwer}
\end{equation}
at $\theta(x)=\sin(\cos(2x))+\sin(4x)$
versus the number of basis functions $m$ for
different separation ranks.
}
{\color{r}
\subsubsection{Hierarchical Tucker Expansion}
The hierarchical Tucker expansion aims at mitigating
the dimensionality of the core-tensor of multivariate
Schmidt decompositions (Section \ref{sec:HT}). It has
advantages over the canonical tensor decomposition in terms
of robustness and computational efficiency. In Figure \ref{fig:HT quadratic}
we study convergence of the hierarchical Tucker
expansion (see Section \ref{sec:HT}) in representing
the quadratic functional \eqref{nonlinF}. Specifically,
we plot the relative pointwise error \eqref{ptwer}
at $\theta(x)=\sin(\cos(2x))+\sin(4x)$
versus the number of dimensions $m$ for different
separation ranks.
\begin{figure}
\centerline{\hspace{-0.5cm}Canonical Tensor Decomposition\hspace{3.9cm} Hierarchical Tucker}
\centerline{
\includegraphics[height=6cm]{cp_quadratic-eps-converted-to.pdf}\hspace{0.8cm}
\includegraphics[height=6cm]{ht_quadratic-eps-converted-to.pdf}
}
\caption{\color{r}Accuracy of canonical tensor decomposition and
hierarchical Tucker expansions in representing the quadratic
functional \eqref{nonlinF}. Specifically, we plot the relative
pointwise error at $\theta(x)=\sin(\cos(2x))+\sin(4x)$ versus the
number of dimensions $m$, and for different separation ranks.}
\label{fig:HT quadratic}
\end{figure}
}
\subsection{Hopf Characteristic Functionals}
In this Section we study the accuracy of polynomial
functional interpolation and tensor methods
in representing Hopf functionals. To this end, consider
the random function
\begin{equation}
u_0(x;\omega)=\frac{1}{\sqrt{q}}\sum_{k=1}^q\left[\eta_{k}\sin(kx)+\xi_{k}\cos(kx)\right],
\label{Burgers_initial}
\end{equation}
where $\{\eta_k,\xi_k\}_{k=1,...,q}$ are i.i.d. random variables satisfying
\begin{equation}
\langle\xi_{k}\rangle=0,\quad \langle\eta_{k}\rangle=0,\quad \langle\xi_{k}^2\rangle=1,
\quad \langle\eta_{k}^2\rangle=1.
\label{assumptions}
\end{equation}
In Figure \ref{fig:samples_u0} we plot few samples of \eqref{Burgers_initial},
for different values of $q$ in the hypothesis that $\{\xi_k\}$ and
$\{\eta_k\}$ are i.i.d. Gaussian random variables.
\begin{figure}[t]
\centerline{\footnotesize\hspace{0.5cm}$q=5$\hspace{4.5cm}$q=20$\hspace{4.5cm}$q=60$}
\centerline{\includegraphics[height=4.5cm]{samples_u0_m5-eps-converted-to.pdf}
\includegraphics[height=4.5cm]{samples_u0_m20-eps-converted-to.pdf}
\includegraphics[height=4.5cm]{samples_u0_m60-eps-converted-to.pdf}
}
\caption{Samples of the random function \eqref{Burgers_initial} for
different values of $q$ (Gaussian $\eta_k$ and $\xi_k$).}
\label{fig:samples_u0}
\end{figure}
From \eqref{assumptions} it follows that the first two statistical
moments of \eqref{Burgers_initial} are independent of $q$, i.e., we have
\begin{equation}
\left<u_0(x;\omega)\right>=0, \qquad
\left<u_0(x;\omega)^2\right>=1,\quad
\textrm{for all $q$.}
\end{equation}
On the other hand, the covariance function
\begin{equation}
C_0(x,y)=\frac{1}{q}\sum_{k=1}^q[\sin(kx)\sin(ky)
+\cos(kx)\cos(ky)]
\label{testcorrelation}
\end{equation}
does depend on $q$, as shown in Figure \ref{fig:covariance}.
Higher-order moments and cumulants can be computed analytically in a similar way.
\begin{figure}
\centerline{\footnotesize$q=1$\hspace{5cm}$q=5$\hspace{5cm}$q=20$}
\centerline{\includegraphics[height=4.5cm]{covariance_m1-eps-converted-to.pdf}
\includegraphics[height=4.5cm]{covariance_m5-eps-converted-to.pdf}
\includegraphics[height=4.5cm]{covariance_m20-eps-converted-to.pdf}
}
\caption{Covariance function \eqref{testcorrelation} for
different values of $q$.}
\label{fig:covariance}
\end{figure}
The Hopf functional of $u_0(x;\omega)$ is defined as
\begin{equation}
\Phi([\theta(x)])=\left<\exp\left[i\int_{0}^{2\pi}
\theta(x)u_0(x;\omega)dx\right]\right>,
\label{H_initial}
\end{equation}
where the average $\left<\cdot\right>$ is a multi-dimensional integral
with respect to the joint probability density function of $\xi_k$ and $\eta_k$.
By substituting \eqref{Burgers_initial}
into \eqref{H_initial} we obtain
\begin{equation}
\Phi([\theta(x)])=
\prod_{k=1}^q\int e^{ias_k[\theta]}p_{\eta_k}(a)da
\prod_{k=1}^q\int e^{iac_k[\theta]}p_{\xi_k}(a)da, \qquad
\label{H_initial1}
\end{equation}
where
\begin{equation}
s_k[\theta]=\frac{1}{\sqrt{q}}\int_{0}^{2\pi}\theta(x)\sin(kx)dx,\qquad
c_k[\theta]=\frac{1}{\sqrt{q}}\int_{0}^{2\pi}\theta(x)\cos(kx)dx.
\label{coeffskck}
\end{equation}
Depending on the probability density functions
$p_{\eta_k}(a)$ and $p_{\xi_k}(a)$ appearing in \eqref{H_initial1},
we have different expressions of $F([\theta])$.
\paragraph{Gaussian Random Fields} Let us assume
\begin{equation}
p_{\eta_k}(a)=\frac{1}{\sqrt{2\pi}}e^{-a^2/2}\qquad
p_{\xi_k}(a)= p_{\eta_k}(a).
\end{equation}
The integrals in \eqref{H_initial1} are in the form
\begin{equation}
\frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty} e^{ias_k[\theta]-a^2/2}da=
e^{-s^2_k[\theta]/2}.
\end{equation}
Therefore, we obtain
\begin{align}
\Phi([\theta(x)])=&\exp\left[-\frac{1}{2}\sum_{k=1}^q
\left(s^2_k[\theta]+c^2_k[\theta]\right)\right]\label{H_initial3}\\
=&\exp\left[-\frac{1}{2} \int_{0}^{2\pi}\int_{0}^{2\pi}\theta(x)\theta(y)
C_0(x,y)dxdy\right],\label{FDG}
\end{align}
in agreement with well-known results for Hopf characteristic functionals
of Gaussian random fields.
\paragraph{Uniform Random Fields} Let us assume
\begin{equation}
p_{\eta_k}(a)=
\begin{cases}
1/(2\sqrt{3})& a\in[-\sqrt{3},\sqrt{3}]\\
0 & \textrm{otherwise}
\end{cases}\qquad p_{\xi_k}(a)= p_{\eta_k}(a)
\end{equation}
In this way the assumptions \eqref{assumptions} are satisfied.
The integrals in \eqref{H_initial1} are easily obtained as
\begin{equation}
\frac{1}{2\sqrt{3}}\int_{-\sqrt{3}}^{\sqrt{3}} e^{ias_k[\theta]}da=
\frac{\sinh(i\sqrt{3}s_k[\theta])}{i\sqrt{3}s_k[\theta]}=
\frac{\sin(\sqrt{3}s_k[\theta])}{\sqrt{3}s_k[\theta]}.
\end{equation}
A substitution of this formula into \eqref{H_initial1} yields
\begin{align}
\Phi([\theta(x)])
=& \prod_{k=1}^q\frac{\sin(\sqrt{3}s_k[\theta])\sin(\sqrt{3}c_k[\theta])}
{3s_k[\theta]c_k[\theta]}.
\label{H_initial2}
\end{align}
\vspace{0.5cm}
\noindent
Note that in both cases we just discussed
the Hopf functional turns out to be
real-valued. Moreover, $\Phi\rightarrow 0$ as
$\left\|\theta\right\|\rightarrow\infty$ (Riemann-Lebesgue lemma),
at a rate that depends on the regularity of the underlying
probability density functional. In particular, in
the Gaussian case $\Phi$ goes to zero faster than in
the uniform case.
Note also, that both \eqref{FDG} are \eqref{H_initial2} are
{\em entire functionals} (i.e., analytic on the complex plane).
This implies that the polynomial interpolation
process converges {\em pointwise} \cite{Khlobystov2}.
\subsubsection{Effective Dimension}
\label{sec:Hopf_effective_dimensionality}
Let us represent
$\theta(x)$ in the space of periodic functions in $[0,2\pi]$.
Possible bases are the discrete trigonometric polynomials
\eqref{setbf} or the more classical Fourier modes
\begin{equation}
1,\qquad \sin(kx),\qquad \cos(kx), \qquad k=1,2, ... . \label{FB}
\end{equation}
Let us now ask the following question:
what is the effective dimension of the Hopf
functional \eqref{H_initial} in the space of periodic
functions?
Such dimension is clearly determined by the dimension of
the linear functionals $s_k$ and $c_k$ in \eqref{coeffskck}.
If we expand the test function $\theta$ in a
classical Fourier series
\begin{equation}
\theta(x)= a_0 + \sum_{k=1}^N a_k\sin(kx)+ \sum_{k=1}^N b_k\cos(kx)
\end{equation}
and we substitute it into \eqref{coeffskck} then we obtain
\begin{equation}
s_k[\theta]=\frac{\pi}{\sqrt{q}} a_k, \qquad\textrm{and}\qquad c_k[\theta]=
\frac{\pi}{\sqrt{q}} b_k.
\end{equation}
This means that the {\em effective dimension} of the Hopf functional
\eqref{H_initial} is {\em exactly} $2q$.
To show this numerically, we consider the Hopf functional
\eqref{FDG} (Gaussian case) and plot the spectrum of the covariance matrix
\begin{equation}
Z_{ij} = \int_0^{2\pi}\int_0^{2\pi}C_0(x,y)\varphi_i(x)\varphi_j(y)dxdy,
\label{covariance}
\end{equation}
obtained by projecting the covariance function
\eqref{testcorrelation} onto the
the Fourier modes \eqref{FB} or, equivalently,
onto the discrete trigonometric polynomials \eqref{setbf}.
\begin{figure}[t]
\vspace{0.5cm}
\centerline{
\includegraphics[height=6cm]{Gaussian_spectra_fourier1-eps-converted-to.pdf}
}
\caption{
Spectra of the covariance matrix \eqref{covariance}
obtained by projecting the covariance function
\eqref{testcorrelation} onto the Fourier modes \eqref{FB},
or the discrete trigonometric polynomials \eqref{setbf}.
It is seen that the number of active components is exactly
$2q$ and all variables are {\em equally important}. This has
important consequences when we approximate the Hopf
functional using tensor methods or polynomial functionals.
In particular, any approximation in a function space with less
than $2q$ dimensions yields a systematic error.}
\label{fig:covariancespectra}
\end{figure}
As it is clearly seen from
Figure \ref{fig:covariancespectra} the number of
active components is exactly $2q$. Note also that
the spectrum is {flat} in all cases, which means
that all active variables are {equally important}.
This has important consequences when
approximating the functional \eqref{FDG} by
polynomial functionals or SSE. In particular, if we use
functionals involving less than $2q$ components, e.g.,
if $m<2q$ in equation \eqref{functional-SSE}, then we cannot
approximate \eqref{FDG} accurately, no matter how we push the
expansion order.
\subsubsection{Polynomial Functional Interpolation}
\label{sec:Hopf polynomial}
Let us expand \eqref{FDG} in a power series
\begin{align}
\Phi([\theta(x)])=&1-\frac{1}{2} \int_{0}^{2\pi}\int_{0}^{2\pi}
C_0(x_1,x_2)\theta(x_1)\theta(x_2)dx_1dx_2+\nonumber\\
&\frac{1}{4}\int_{0}^{2\pi}\int_{0}^{2\pi}\int_{0}^{2\pi}\int_{0}^{2\pi}C_0(x_1,x_2)C_0(x_3,x_4)
\theta(x_1)\theta(x_2)\theta(x_3)\theta(x_4)dx_1dx_2dx_3dx_4+\cdots.
\end{align}
For small $\theta(x)$ we can truncate the series,
and represent $\Phi$ in terms of an interpolating polynomial functional
of relatively small order, provided we have enough interpolation
nodes nearby $\theta(x)=0$.
This is demonstrated in Figure \ref{fig:Hopf_porter}, where we
study the accuracy of Porter's interpolant through the set of nodes
$\widehat{S}^{(m+1)}_n$ (see Eq. \eqref{SNq2}). Specifically,
we plot the error
\begin{equation}
E_m= \sup_{\theta\in \mathcal{G}_{50}(\sigma)}\left|\Phi([\theta])-\Pi_n([\theta])\right|
\label{Em3}
\end{equation}
versus $m$ for $q=5$, $n=1,2,3,4$ and $\sigma=0.01$.
It is seen that the interpolants converge in both $n$ (polynomial order)
and $m$ (number of basis functions). Convergence in $m$
becomes monotonic for $m\geq10$.
This is related to the fact that the effective dimension of
the Hopf functionals \eqref{H_initial3} and \eqref{H_initial2}
is $10$ for $q=5$ (see Section \ref{sec:Hopf_effective_dimensionality}).
Therefore for $m\geq 10$ we have enough basis functions
to fully resolve them. The accuracy of the polynomial
interpolants then depends only on $n$ (polynomial order),
the number of interpolation nodes and their location.
The number of interpolation nodes in $\widehat{S}^{(m+1)}_n$
is given in \eqref{SCardinality}. For example, the case
$m=20$ and $n=4$ yields $12650$ nodes and an
interpolation matrix \eqref{matH} of size $12650\times12650$\footnote{Recall
that the matrix \eqref{matH} has to be inverted to compute the
cardinal basis \eqref{gi_0}.}.
\begin{figure}[t]
\centerline{\hspace{0.5cm} Gaussian Functional \hspace{4.2cm} Uniform Functional}
\centerline{
\includegraphics[height=6cm]{Errors_Hopf_Gaussian_sig001-eps-converted-to.pdf}
\includegraphics[height=6cm]{Errors_Hopf_Uniform_sig001-eps-converted-to.pdf}
}
\caption{Accuracy of Porter's polynomial functional interpolants in
representing the Hopf functionals \eqref{H_initial3} and \eqref{H_initial2}.
Shown are the pointwise errors \eqref{Em3} versus $m$
for different orders of the polynomial interpolant.
The errors are computed by sampling $20000$ functions from
the set $\mathcal{G}_{50}(\sigma)$ with $\sigma=0.01$
and then computing the supremum \eqref{Em3}.}
\label{fig:Hopf_porter}
\end{figure}
When evaluating the error \eqref{Em3} it is important to
select $\mathcal{G}_{50}(\sigma)$ within the convex hull
of the interpolation nodes, e.g.,
by choosing $\sigma$ small enough. In this way we avoid
inaccuracies due to polynomial extrapolation.
To mitigate this phenomenon, we can also consider different sets
of interpolation nodes, e.g., the set $\widetilde{S}^{(m+1)}_n$
defined in \eqref{SNq3} with $a_{i_q}$ sampled at
Gauss-Hermite sparse-grid points.
In Figure \ref{fig:sparse_grid_sets} we plot the elements
in $\widetilde{S}^{(m+1)}_1$, $\widetilde{S}^{(m+1)}_2$
and $\widetilde{S}^{(m+1)}_3$
for Gauss-Hermite sparse grids of level 5 (see also Table \ref{tab:thenumber}).
\begin{table}
\centering
\begin{tabular}{c|cccccccc}
$m$ & $2$ & $4$ & $6$ & $8$ & $10$ & $16$ & $20$\\
\hline\vspace{-0.4cm}\\
$\#\widetilde{S}^{(m+1)}_2$ & $121$ & $341$ & $673$ & $1117$ & $1673$ & $4013$ & $6133$
\end{tabular}
\caption{Number of interpolation nodes in $\widetilde{S}^{(m+1)}_2$ with
$a_{i_j}$ sampled at Gauss-Hermite sparse grids (level 5).}
\label{tab:thenumber}
\end{table}
\begin{figure}[!ht]
\centerline{\hspace{0.12cm}Nodes in Fourier space \hspace{2.4cm} Nodes in function space}
\centerline{\includegraphics[height=4.9cm]{Herm_points_1-eps-converted-to.pdf}
\includegraphics[height=4.9cm]{SG_basis_1-eps-converted-to.pdf}}
\centerline{\includegraphics[height=4.9cm]{Herm_points_2-eps-converted-to.pdf}
\includegraphics[height=4.9cm]{SG_basis_2-eps-converted-to.pdf}}
\centerline{\includegraphics[height=4.9cm]{Herm_points_3-eps-converted-to.pdf}
\includegraphics[height=4.9cm]{SG_basis_3-eps-converted-to.pdf}}
\caption{Gauss-Hermite sparse grids (left) and corresponding interpolation
nodes in the function space $\widetilde{S}^{(m)}_n$ (right).
Specifically, we plot all elements in
$\widetilde{S}^{(9)}_1$ (first row, $45$ elements),
$\widetilde{S}^{(9)}_2$ (second row, $1117$ elements) and
$\widetilde{S}^{(9)}_3$ (third row, $3949$ elements).}
\label{fig:sparse_grid_sets}
\end{figure}
The symmetry of the nodes $a_{i_j}$ in the Fourier space yields
linearly dependent nodes in $\widetilde{S}^{(m+1)}_n$.
Correspondingly, the interpolation matrix \eqref{matH} is
rank-deficient, i.e., it cannot be inverted in a classical sense.
We can overcome this issue by either removing
some nodes from the set $\widetilde{S}^{(m+1)}_n$, or by
taking the More-Penrose pseudo-inverse of \eqref{matH}. In the latter
case, we obtain a non-cardinal basis \eqref{gi_0_noncardinal} and a
polynomial functional in the form \eqref{Porter_interpolant_noncardinal}.
In Figure \ref{fig:Hopf_porter_SG} we demonstrate convergence
of such polynomial functional (order 1 and 2) to the
Hopf functionals \eqref{FDG} and \eqref{H_initial2}.
Specifically, we plot the pointwise errors \eqref{Em3}
versus $m$ for a set of $20000$ randomly generated evaluation
nodes in $\mathcal{G}_{50}(\sigma)$, with $\sigma=1$.
\begin{figure}[!t]
\centerline{\hspace{0.5cm} Gaussian Functional \hspace{5cm} Uniform Functional}
\centerline{
\includegraphics[height=6.5cm]{Errors_Hopf_Gaussian_sig1_SG-eps-converted-to.pdf}
\includegraphics[height=6.5cm]{Errors_Hopf_Uniform_sig1_SG-eps-converted-to.pdf}
}
\caption{Approximation of the Hopf functionals
\eqref{H_initial3} and \eqref{H_initial2} by using
polynomial functionals of total degree $n=1$ and $n=2$.
The polynomial are in the form \eqref{Porter_interpolant_noncardinal}
(i.e., non-interpolatory) and they are constructed by using the set of nodes
$\widetilde{S}^{(m)}_1$ and $\widetilde{S}^{(m)}_2$
defined in equation \eqref{SNq3}, with $a_{i_j}$ sampled at Gauss-Hermite
sparse grids of level $5$ \cite{Bungartz}.
Shown are the pointwise errors \eqref{Em3} versus $m$ for functions
$\theta(x)$ in the ensemble $\mathcal{G}_{50}(1)$.}
\label{fig:Hopf_porter_SG}
\end{figure}
\subsubsection{Canonical Tensor Decomposition}
We evaluate the Hopf functionals \eqref{FDG} and \eqref{H_initial2}
in the function space spanned by the finite-dimensional
orthonormal basis
\begin{equation}
\varphi_0(x)=\frac{1}{2\pi}, \quad
\varphi_{k-1}(x)=\frac{\sin(kx/2)}{\sqrt{\pi}},\quad
\varphi_{k}(x)=\frac{\cos((1+k/2)x)}{\sqrt{\pi}}\qquad
\textrm{($k$ even)}.
\label{basisE}
\end{equation}
Hereafter we prove that this yields exact rank-one representations
of both Gaussian and uniform Hopf functionals.
\paragraph{Gaussian Functionals}
By evaluating \eqref{FDG} in the space
spanned by \eqref{basisE} we obtain
\begin{equation}
f(a_1,...,a_m) = \exp\left[-\frac{1}{2}\sum_{p,q=1}^mZ_{pq} a_{p}a_q
\right],
\end{equation}
where
\begin{equation}
Z_{ij}= \int_0^{2\pi}\int_0^{2\pi} C_0(x,y)\varphi_i(x)\varphi_j(y)dxdy.
\end{equation}
It can be verified that
\begin{equation}
Z_{ij}=
\begin{cases}
\displaystyle\frac{\pi}{q}\delta_{ij} & \quad i,j\leq 2q\\
0 & \quad \textrm {otherwise}
\end{cases}
\nonumber
\end{equation}
and therefore the Gaussian functional \eqref{FDG}
is {\em rank one} relative to the basis \eqref{basisE}, i.e.,
\begin{equation}
\Phi([\theta])=\exp\left[-\frac{\pi}{2q}\sum_{p=1}^{2q}
a_{p}^2\right], \qquad a_p=(\theta,\varphi_p).
\end{equation}
Thus, a rank one canonical decomposition with $m=2q$ variables
is exact in the case. It's important to remark that the
Gaussian functional \eqref{FDG} is {\em not} rank one
relative to other bases, for example \eqref{setbf}.
\paragraph{Uniform Functionals}
By evaluating \eqref{H_initial2} in the space
spanned by \eqref{basisE} we obtain
\begin{equation}
f(a_1,...,a_m)=\prod_{k=1}^q\frac{\sin(\sqrt{3}s_k(a_1,..,a_m))\sin(\sqrt{3}c_k(a_1,..,a_m))} {3s_k(a_1,..,a_m)c_k(a_1,..,a_m)} ,
\end{equation}
where $s_k(a_1,..,a_m)$ and $c_k(a_1,..,a_m)$ are
defined as (see \eqref{coeffskck})
\begin{align}
s_k(a_1,..,a_m)= &\frac{1}{\sqrt{q}}\sum_{p=0}^m
(\sin(kx),\varphi_p(x)) a_p,\label{re1}\\
c_k(a_1,..,a_m)= &\frac{1}{\sqrt{q}}\sum_{p=0}^m
(\cos(kx),\varphi_p(x)) a_p.\label{re2}
\end{align}
A substitution of \eqref{basisE} into \eqref{re1}-\eqref{re2} yields
\begin{align}
s_k(a_1,..,a_m)=& \sqrt{\frac{\pi}{q}} a_k,\quad k=1,...,q,
\quad (\textrm{zero otherwise}),\label{ss1}\\
c_k(a_1,..,a_m)=& \sqrt{\frac{\pi}{q}} a_{k+q},\quad k=1,...,q,
\quad (\textrm{zero otherwise}).\label{ss2}
\end{align}
This means that, relative to the basis \eqref{basisE}
the functional \eqref{H_initial2} is {\em rank one}, i.e.,
\begin{equation}
\Phi([\theta])=\prod_{k=1}^{2q}
\frac{\sin(\sqrt{3\pi} a_k/\sqrt{q})}{3\sqrt{\pi}a_k/\sqrt{q}}, \qquad
a_k=(\theta,\varphi_k).
\end{equation}
Therefore, a rank one canonical tensor decomposition with
$m=2q$ variables is exact\footnote{Equations
\eqref{ss1} and \eqref{ss2} suggest that
all the $2q$ variables $a_1,...,a_{2q}$ are
equally important.
}. Similarly to the Gaussian case,
the uniform functional is not rank-one relative to other
bases, for example \eqref{setbf}.
\subsubsection{Functional Derivatives}
The first- and second-order functional derivatives
of the Hopf functional \eqref{H_initial} are defined as
\begin{align}
\frac{\delta \Phi([\theta])}{\delta \theta(x)}&=i\left<u_0(x;\omega)\exp\left[i\int_{0}^{2\pi}
\theta(x)u_0(x;\omega)dx\right]\right>,\label{Ai}\\
\frac{\delta^2 \Phi([\theta])}{\delta \theta(x)\delta\theta(y)}&=
-\left<u_0(x;\omega)u_0(y;\omega)\exp\left[i\int_{0}^{2\pi}
\theta(x)u_0(x;\omega)dx\right]\right>.\label{Bi}
\end{align}
Remarkably, these derivatives can
be expressed analytically in terms of simple functions.
To this end, we need a formula to compute
the averages in \eqref{Ai} and \eqref{Bi}.
A lengthy calculation shows that
\begin{equation}
\frac{\delta \Phi([\theta])}{\delta \theta(x)}=
\frac{i}{\sqrt{q}}\left(I^{(c)}[\theta]\sum_{k=1}^q\sin(kx)I^{(s)}_k[\theta]+
I^{(s)}[\theta]\sum_{k=1}^q\cos(kx)I^{(c)}_k[\theta]\right),\label{anA}
\end{equation}
\begin{align}
\frac{\delta^2 \Phi([\theta])}{\delta \theta(x)\delta\theta(y)}=&
-\frac{1}{q}\sum_{k,h=1}^q\left(\sin(kx)\cos(hy)I^{(s)}_{k}[\theta]I^{(c)}_{h}[\theta]+
\cos(kx)\sin(hy)I^{(c)}_{k}[\theta]I^{(s)}_{h}[\theta]\right)\nonumber\\
&-\frac{1}{q}\sum_{\substack{k,h=1\\k\neq h}}^q\left(
I^{(c)}[\theta]\sin(kx)\sin(hy)I^{(s)}_{kh}[\theta]+
I^{(s)}[\theta]\cos(kx)\cos(hy)I^{(c)}_{kh}[\theta]\right)\nonumber \\
&-\frac{1}{q}\sum_{k=1}^q\left(
I^{(c)}[\theta]\sin(kx)\sin(ky)J^{(s)}_{k}[\theta]+
I^{(s)}[\theta]\cos(kx)\cos(ky)J^{(c)}_{k}[\theta]\right),
\label{anB}
\end{align}
where the functionals $I^{(c)}[\theta]$, $I^{(s)}[\theta]$, $I_k^{(c)}[\theta]$, etc.,
are defined in Table \ref{tab:coefficients}.
These expressions are general and they hold for any random function in the
form \eqref{Burgers_initial}, with i.i.d. random variables $\eta_k$ and $\xi_j$.
\begin{table}[!t]
\small
\centerline{\line(1,0){430}\vspace{0.2cm}}
\centerline{Coefficients Appearing in the Functional Derivatives \eqref{anA} and \eqref{anB}}
\centerline{\line(1,0){430}}
\vspace{0.5cm}
\centerline{\small$\displaystyle G_n(z)=\int_{-\infty}^\infty a^n e^{iz a}p(a)da$}
\vspace{1cm}
\centerline{
\begin{tabular}{ll}
\small
Gaussian PDF
& $G_0(z)= e^{-z^2/2}$ \\
& $G_1(z)= ize^{-z^2/2}$ \\
& $G_2(z)=(1-z^2)e^{-z^2/2}$\\\\
Uniform PDF & $\displaystyle G_0(z)= \frac{\sin(\sqrt{3}z)}{\sqrt{3}z}$ \\
& $\displaystyle G_1(z)= i\frac{\sin(\sqrt{3}z)}{\sqrt{3}z^2}
-i\frac{\cos(\sqrt{3}z)}{z}$ \\
& $\displaystyle G_2(z)= \frac{2\cos(\sqrt{3}z)}{z^2}
+\frac{\sqrt{3}\sin(\sqrt{3}z)}{z}
-\frac{2\sin(\sqrt{3}z)}{\sqrt{3}z^3}$
\end{tabular}}
\vspace{1cm}
\centerline{
\begin{tabular}{lcl}
\small
$\displaystyle I^{(s)}[\theta]=\prod_{r=1}^q G_0(s_r[\theta])$ & &
$\displaystyle I^{(c)}[\theta]=\prod_{r=1}^qG_0(c_r[\theta])$\\
$\displaystyle I_k^{(s)}[\theta]=G_1(s_k[\theta])
\prod_{\substack{r=1\\r\neq k}}^qG_0(s_r[\theta])$ & &
$\displaystyle I_k^{(c)}[\theta]=G_1(c_k[\theta])
\prod_{\substack{r=1\\r\neq k}}^qG_0(c_r[\theta])$\\
$\displaystyle J_k^{(s)}[\theta]=G_2(s_k[\theta])
\prod_{\substack{r=1\\r\neq k}}^qG_0(s_r[\theta])$ & &
$\displaystyle J_k^{(c)}[\theta]=G_2(c_k[\theta])
\prod_{\substack{r=1\\r\neq k}}^qG_0(c_r[\theta])$\\
$\displaystyle I_{kh}^{(s)}[\theta]=G_1(s_k[\theta]) G_1(s_h[\theta])
\prod_{\substack{r=1\\r\neq k,h}}^qG_0(s_r[\theta])$ & &
$\displaystyle I_{kh}^{(c)}[\theta]=G_1(c_k[\theta]) G_1(c_h[\theta])
\prod_{\substack{r=1\\r\neq k,h}}^qG_0(c_r[\theta])$ \\\\\hline\\
\end{tabular}}
\caption{Coefficients appearing in the functional derivatives
\eqref{anA} and \eqref{anB}.}
\label{tab:coefficients}
\end{table}
Note that in the case of uniform PDF the coefficients
in Table \ref{tab:coefficients} satisfy
\begin{equation}
\lim_{z\rightarrow 0}G_0(z)=1,\qquad
\lim_{z\rightarrow 0}G_1(z)=0,\qquad
\lim_{z\rightarrow 0}G_2(z)=1.
\end{equation}
and therefore there are no singularities at $z=0$. Also, all coefficients
decay to $0$ when $z$ goes to infinity.
The expressions \eqref{anA} and \eqref{anB} can be simplified
significantly for Gaussian Hopf functionals \eqref{FDG}. In particular, we obtain\footnote{Note that the first-order derivative reduces
to $0$ (mean field) at $\theta=0$ while the second-order
derivative at $\theta=0$ reduces to the opposite of the
correlation function $C_0(x,y)$.}
\begin{align}
\frac{\delta \Phi([\theta])}{\delta \theta(x)}=-
\left[\int_{0}^{2\pi}C_0(x,y)\theta(y)dy\right]
\exp\left[-\frac{1}{2}\int_{0}^{2\pi} \int_{0}^{2\pi}C_0(x,y)\theta(x)\theta(y)dxdy\right]
\label{fD1},
\end{align}
\begin{align}
\frac{\delta^2 \Phi([\theta])}{\delta \theta(x)\delta\theta(y)}=
&\left[-C_0(x,y)+\int_{0}^{2\pi}C_0(x,y)\theta(y)dy\int_{0}^{2\pi}C_0(x,y)\theta(x)dx\right]
\times\nonumber\\
& \exp\left[-\frac{1}{2}\int_{0}^{2\pi} \int_{0}^{2\pi}C_0(x,y)\theta(x)\theta(y)dxdy\right].
\label{fD2}
\end{align}
In Figure \ref{fig:functional_derivatives1} and
Figure \ref{fig:functional_derivatives2} we plot
the first- and second-order functional derivatives of
the Hopf functionals \eqref{H_initial3} and
\eqref{H_initial2}, evaluated at different test functions.
\begin{figure}
\small
\centerline{Gaussian \hspace{4.3cm} Uniform}\vspace{-0.2cm}
\centerline{\line(1,0){300}}
\vspace{0.5cm }
\centerline{$\theta(x)=0$}
\centerline{\includegraphics[height=5cm]{FD1_G0-eps-converted-to.pdf}
\includegraphics[height=5cm]{FD1_U0-eps-converted-to.pdf}}
\vspace{0.5cm}
\centerline{$\theta(x)=40\sin(6x)^2+5\cos(x)^4$}
\centerline{\includegraphics[height=5cm]{FD1_G1-eps-converted-to.pdf}
\includegraphics[height=5cm]{FD1_U1-eps-converted-to.pdf}}
\vspace{0.5cm}
\centerline{$\theta(x)=2\exp(-\sin(4x))+\cos(\cos(4x))$}
\centerline{\includegraphics[height=5cm]{FD1_G2-eps-converted-to.pdf}
\includegraphics[height=5cm]{FD1_U2-eps-converted-to.pdf}}
\caption{First-order functional derivatives of the
Hopf functionals \eqref{H_initial3} (Gaussian) and
\eqref{H_initial2} (Uniform) for $q=10$ evaluated at different test
functions. The Hopf functionals we are considering here are real-valued
and therefore the functional derivatives are real. Note that evaluating
the first-order functional derivative at $\theta(x)=0$ (first row)
yields the mean field $\left<u_0(x)\right>=0$.}
\label{fig:functional_derivatives1}
\end{figure}
\begin{figure}
\small
\centerline{Gaussian \hspace{4.8cm} Uniform}\vspace{-0.2cm}
\centerline{\line(1,0){300}}
\vspace{0.5cm }
\centerline{$\theta(x)=0$}
\centerline{\includegraphics[height=5cm]{FD2_G0-eps-converted-to.pdf}
\includegraphics[height=5cm]{FD2_U0-eps-converted-to.pdf}}
\vspace{0.5cm}
\centerline{$\theta(x)=40\sin(6x)^2+5\cos(x)^4$}
\centerline{\includegraphics[height=5cm]{FD2_G1-eps-converted-to.pdf}
\includegraphics[height=5cm]{FD2_U1-eps-converted-to.pdf}}
\vspace{0.5cm}
\centerline{$\theta(x)=2\exp(-\sin(4x))+\cos(\cos(4x))$}
\centerline{\includegraphics[height=5cm]{FD2_G2-eps-converted-to.pdf}
\includegraphics[height=5cm]{FD2_U2-eps-converted-to.pdf}}
\caption{Second-order functional derivatives of the
Hopf functionals \eqref{H_initial3} (Gaussian) and
\eqref{H_initial2} (Uniform) for $q=10$ evaluated at different test
functions. Both functionals are real-valued
and therefore the functional derivatives are real. Note that evaluating
the second-order functional derivative at $\theta(x)=0$ (first row)
yields the opposite of the correlation function $C_0(x,y)$ in \eqref{testcorrelation}.}
\label{fig:functional_derivatives2}
\end{figure}
\paragraph{Canonical Tensor Expansion of Functional Derivatives}
We have seen in Section \ref{sec:tensor} that the functional
derivatives of any cylindrical approximant of $\Phi$ can be expressed as
\begin{equation}
\frac{\delta {\Phi}([\theta])}{\delta\theta(x)}\simeq \sum_{k=1}^m
\frac{\partial f}{\partial a_k}\varphi_k(x),\label{SSEfd1}
\end{equation}
\begin{equation}
\frac{\delta^2 {\Phi}([\theta])}{\delta\theta(x)\delta\theta(y)}\simeq \sum_{k,j=1}^m
\frac{\partial^2 f}{\partial a_k\partial a_j}\varphi_k(x)\varphi_j(y),
\label{SSEfd2}
\end{equation}
where $a_k =(\theta,\varphi_k)$. In particular, if we consider
a canonical tensor expansion of $\Phi$ then $f$ has the form
\eqref{functional-SSE}, and the partial derivatives in
\eqref{SSEfd1} and \eqref{SSEfd2} can be easily computed.
In Figure \ref{fig:accuracyHopfFDs} we
show that \eqref{SSEfd1} and \eqref{SSEfd2}
provide a very accurate approximation of the functional
derivatives \eqref{fD1} and \eqref{fD2}.
\begin{figure}[!ht]
\centerline{\hspace{1.8cm}Analytical \hspace{3.7cm} Canonical Tensor Decomposition}
\vspace{0.cm}
\centerline{\hspace{-.6cm}
\includegraphics[height=4.3cm]{FD1_Gaussian_EXACT-eps-converted-to.pdf}\hspace{1.8cm}
\includegraphics[height=4.3cm]{FD1_Gaussian_SSE-eps-converted-to.pdf}
}
\centerline{
\includegraphics[height=5.cm]{FD2_Gaussian_EXACT-eps-converted-to.pdf}
\includegraphics[height=5.cm]{FD2_SSE_Gaussian-eps-converted-to.pdf}
}
\caption{First- and second-order functional
derivatives of the Gaussian Hopf functional \eqref{FDG} evaluated
at $\theta(x)=\sin(x)+\sin(2x)+\sin(3x)$ .
Shown are analytical results \eqref{fD1}
and \eqref{fD2} versus numerical result obtained by
the canonical tensor decomposition
(equations \eqref{SSEfd1} and \eqref{SSEfd2}).}
\label{fig:accuracyHopfFDs}
\end{figure}
\subsection{Sine Functional}
\label{sec:sine functional}
Consider the nonlinear functional
\begin{equation}
F([\theta])=\sin\left((K_1,\theta)\right),\qquad
(K_1,\theta)=\int_0^{2\pi}K_1(x)\theta(x)dx,
\label{DINE}
\end{equation}
where $K_1(x)$ is as in equation \eqref{K1}.
We represent $F([\theta])$ in the function space
\begin{equation}
D_m=\textrm{span}\{\varphi_1,...,\varphi_m\}
\end{equation}
spanned by a finite-dimensional orthonormal basis.
This yields the cylindrical representation (see Section \ref{sec:tensor})
\begin{equation}
f(a_1,...,a_m) = \sin\left(\sum_{k=1}^m s_k(\varphi_k,\theta)\right),
\qquad s_k=\int_0^{2\pi}K_1(x)\varphi_k(x)dx.
\label{SINEF}
\end{equation}
In Figure \ref{fig:porterSSEsine} we study the accuracy of
second-order Porter's polynomial functionals and
canonical tensor decomposition \eqref{functional-SSE} in
approximating the functional \eqref{SINEF}.
\begin{figure}[!ht]
\centerline{\hspace{1cm}Polynomial functionals \hspace{3.5cm} Canonical Tensor Decomposition}
\vspace{0.cm}
\centerline{
\includegraphics[height=6.cm]{Porter_sine-eps-converted-to.pdf}
\includegraphics[height=6.cm]{SSE_sine-eps-converted-to.pdf}}
\caption{Approximation of the sine functional \eqref{SINEF}.
Shown are the $L_2([-1,1]^m)$ errors obtained by using Porter's
polynomial functionals \eqref{Porter_interpolant_noncardinal}
of second-order (left) with non-cardinal basis,
and canonical tensor decomposition (right). Specifically,
we study convergence as a function of the number of
interpolation nodes in Porter's method: $m=10$ and sparse grids
level 5 yields $\#\widetilde{S}^{(m+1)}_2=1673$ nodes (see Table \ref{tab:thenumber}). In the canonical tensor decomposition
method we show convergence of \eqref{functional-SSE} as
a function of the separation rank ($r$). The dimension of the
test function space is chosen to be $m=10$.
}
\label{fig:porterSSEsine}
\end{figure}
Specifically, the polynomial functionals are constructed by using
the set of nodes $\widetilde{S}^{(m+1)}_2$ defined in equation
\eqref{SNq2}, where $a_{i_j}$ are sampled
at Gauss-Hermite sparse grids of level 5.
We recall that this set yields a
rank-deficient matrix \eqref{matH}, which requires
More-Penrose pseudoinversion
(see Section \ref{sec:Hopf polynomial}).
Correspondingly, the polynomial functionals do not
interpolate \eqref{SINEF}.
The canonical tensor decomposition, on the other hand,
is based on Legendre polynomials of order $Q=6$.
The functionals $G^{l}_k((\theta,\varphi_k))$
are shown in Figure \ref{fig:G}.
\begin{figure}[!ht]
\centerline{
\includegraphics[height=3.5cm]{G11-eps-converted-to.pdf}
\includegraphics[height=3.5cm]{G21-eps-converted-to.pdf}
\includegraphics[height=3.5cm]{G31-eps-converted-to.pdf}
}
\centerline{
\includegraphics[height=3.5cm]{G12-eps-converted-to.pdf}
\includegraphics[height=3.5cm]{G22-eps-converted-to.pdf}
\includegraphics[height=3.5cm]{G32-eps-converted-to.pdf}
}
\centerline{
\includegraphics[height=3.5cm]{G13-eps-converted-to.pdf}
\includegraphics[height=3.5cm]{G23-eps-converted-to.pdf}
\includegraphics[height=3.5cm]{G33-eps-converted-to.pdf}
}
\caption{Tensor components of the sine functional \eqref{SINEF}
versus $(\theta,\varphi_k)$.}
\label{fig:G}
\end{figure}
\subsubsection{Functional Derivatives}
The first- and second-order functional derivatives of \eqref{DINE}
are easily obtained as
\begin{equation}
\frac{\delta F([\theta])}{\delta \theta(x)}=K_1(x)\cos((K_1,\theta)),
\label{d1sin}
\end{equation}
\begin{equation}
\frac{\delta^2 F([\theta])}{\delta \theta(x)\delta \theta(y)}=
-K_1(x)K_1(y)\sin((K_1,\theta)).
\label{d2sin}
\end{equation}
Note that for each $\theta(x)$, such functional derivatives
are basically a rescaled version of
the functions $K_1(x)$ and $K_1(x)K_1(y)$.
In Figure \ref{fig:sineFD1}
we compare the exact functional derivatives versus those obtained
by the canonical tensor decomposition with separation rank $r=4$.
\begin{figure}[!ht]
\centerline{\hspace{1.3cm}Analytical \hspace{3.8cm} Canonical Tensor Decomposition}
\vspace{0.cm}
\centerline{\hspace{-.7cm}
\includegraphics[height=4.3cm]{sine_FD1_analytical-eps-converted-to.pdf}\hspace{1.8cm}
\includegraphics[height=4.3cm]{sine_FD1_SSE-eps-converted-to.pdf}
}
\centerline{
\includegraphics[height=5.cm]{sine_FD2_exact-eps-converted-to.pdf}\hspace{1.cm}
\includegraphics[height=5.cm]{sine_FD2_SSE-eps-converted-to.pdf}
}
\caption{Sine functional \eqref{DINE}. First- and second-order
functional derivatives evaluated at
$\theta(x)=9(1+\sin(x)+\sin(2x))/10$.
Specifically, we compare the analytical results
\eqref{d1sin} and \eqref{d2sin} versus
results obtained by canonical tensor decomposition
(equations \eqref{ff1} and \eqref{ff5}).}
\label{fig:sineFD1}
\end{figure}
Recall that the functional derivatives can be approximated in the
space of cylindrical functionals as (see Section \ref{sec:tensor})
\begin{equation}
\frac{\delta {F}([\theta])}{\delta\theta(x)}=\sum_{k=1}^m
\frac{\partial f}{\partial a_k}\varphi_k(x),
\label{SSEsinefd1}
\end{equation}
\begin{equation}
\frac{\delta^2 {F}([\theta])}{\delta\theta(x)\delta\theta(y)}=\sum_{k,j=1}^m
\frac{\partial^2 f}{\partial a_k\partial a_j}\varphi_k(x)\varphi_j(y).
\label{SSEsinefd2}
\end{equation}
In the case of canonical tensor expansions,
$\partial f/\partial a_k$ and
$\partial^2 f/\partial a_j\partial a_k$ are
defined in \eqref{pd1} and \eqref{pd2}, respectively.
\section{Numerical Results: Functional Differential Equations}
\label{sec:numerical results functional equations}
Computing the numerical solution to a functional differential
equation is a long standing open problem in mathematical
physics. In this Section we address this problem with reference
to linear functional equations in the
form \eqref{linfde00}. In particular, we study the initial value
problem for a prototype functional advection-reaction equation.
We also develop the numerical discretization of the
Navier-Stokes-Hopf functional equation
(see Section \ref{sec:motivation}), and discuss
its computational complexity.
{
\color{r}
\subsection{Advection-Reaction Functional Differential Equation}
\label{sec:ADVR}
Consider the following advection-reaction functional differential
equation
\begin{equation}
\frac{\partial F([\theta],t)}{\partial t}+\int_{0}^{2\pi} \theta(x)\frac{\partial }{\partial x}\left( \frac{\delta F([\theta],t)}{\delta \theta(x)}\right)dx=H([\theta],t)
\label{FDE3}
\end{equation}
evolving from the initial condition
\begin{equation}
F([\theta],0)=F_0([\theta]).
\label{fffer}
\end{equation}
Here $H([\theta],t)$ is a given functional reaction term.
We assume that $D(F)$ (the domain of the functional
$F$) is a suitable space of functions\footnote{\color{r}We have seen in
Section \ref{sec:FDEs approximation} that the solution
to the initial value problem \eqref{FDE3}-\eqref{fffer}
is strongly dependent on the choice of
function space $D(F)$.}, e.g., the space
of periodic functions in $[0,2\pi]$ or the space of
infinitely differentiable functions in $[0,2\pi]$ such that
$\theta(0)=0$. Evaluation of
$F([\theta],t)$ and $H([\theta],t)$ in
the finite-dimensional subspace
\begin{equation}
D_m = \left\{\theta(x)\in D(F) \,\left|\, \theta(x)=\sum_{k=1}^m a_k \varphi_k(x) \right.
\right\}\subseteq D(F),
\label{Dm}
\end{equation}
where $\{\varphi_1,...,\varphi_m\}$ is an orthonormal basis,
yields the following multivariate functions
\begin{align}
f(a_1,...,a_m,t)=F([\theta],t) \qquad h(a_1,...,a_m,t)=H([\theta]), \qquad \theta\in D_m.
\end{align}
We also recall that the functional derivative $\delta F([\theta],t)/\delta \theta(x)$ can be expressed in $D_m$ as (see Eq. \eqref{gg2})
\begin{equation}
\frac{\delta F([\theta],t)}{\delta \theta(x)}=\sum_{j=1}^m
\frac{\partial f}{\partial a_j}\varphi_j(x).
\end{equation}
A substitution of these expression back into \eqref{FDE3}
yields the following initial value problem for a multivariate first-order PDE
\begin{equation}
\frac{\partial f}{\partial t}+\sum_{j=1}^m\left(\sum_{k=1}^m
C_{jk}a_k\right)\frac{\partial f}{\partial a_j}=h,
\qquad f(a_1,...,a_m,0)=f_0(a_1,...,a_m),
\label{PDE-advR}
\end{equation}
where
\begin{equation}
C_{jk}=\int_{0}^{2\pi} \varphi_k(x)\frac{d\varphi_j(x)}{dx}dx
\label{matC}
\end{equation}
The entries $C_{ij}$ depend on the
choice of $D(F)$, and correspondingly $D_m$. For instance,
if we assume that $D(F)$ is the space of infinitely differentiable
periodic functions in $[0,2\pi]$ then the matrix \eqref{matC}
is centro-skew-symmetric.
\vspace{0.7cm}
\noindent
{\em Example 1:}
Another example of advection FDE is
\begin{equation}
\frac{\partial F}{\partial t}+\int_{0}^{2\pi} \theta(x)\frac{\partial }{\partial x} \left[\frac{\delta F}{\delta \theta(x)}\right]dx=
\int_{0}^{2\pi} \theta(x)\frac{\partial^2 }{\partial x^2}\left[ \frac{\delta F}{\delta \theta(x)}\right]dx.
\label{FDE3_bis}
\end{equation}
The discrete form such equation is
\begin{equation}
\frac{\partial f}{\partial t}+\sum_{k,j=1}^m
a_k\left(C_{jk}-F_{jk}\right)\frac{\partial f}{\partial a_j}=0,
\end{equation}
where $C_{jk}$ is defined in \eqref{matC} while
\begin{equation}
F_{kj}= \int_{0}^{2\pi} \varphi_k(x)\frac{d^2\varphi_j(x)}{dx^2}dx.
\end{equation}
\subsubsection{Analytical Solution}
\label{sec:ADVanalyticalSolution}
The analytical solution to the initial value
problem \eqref{PDE-advR} can be
computed by using the method of
characteristics \cite{Rhee}. To this end, let $\bm C$
be the matrix with entries \eqref{matC}, and $\bm a$
be the vector of coordinates $(a_1, ..., a_m)$.
Then the solution to \eqref{PDE-advR} is
\begin{align}
f(\bm a,t)=f_0\left(e^{-t\bm C }\bm a\right) +
\int_0^t h\left(e^{s\bm C }\bm a\right)ds.
\label{solutionF}
\end{align}
In the particular case where the reaction term $h$
is zero we obtain
\begin{equation}
f(a_1,...,a_m,t) = f_0\left(a_1(t), ...,
a_m(t)\right),
\label{solutionFF}
\end{equation}
where the (inverse) flow map $a_i(t;a_1,...,a_m)$ is
given by\footnote{We recall that the matrix exponential
appearing in \eqref{gg1i} can be represented as
\begin{equation}
e^{-t\bm C}= \bm U^T e^{-t\bm \Lambda} \bm U,
\end{equation}
where $\bm U$ is the matrix of eigenvectors
of $\bm C$ (columnwise) and $\bm \Lambda$ is the
diagonal matrix of eigenvalues.}
\begin{equation}
a_i (t;a_1,...,a_m)=\sum_{j=1}^m Z_{ij}(t)a_j,\qquad
\bm Z(t)=e^{-t\bm C}.
\label{gg1i}
\end{equation}
The solution to the functional equation \eqref{FDE3}
can be obtained by taking the continuum limit of
\eqref{solutionFF}, i.e., by sending $m$ to infinity.
This yields
\begin{equation}
F([\theta],t)=F_0([\theta(x,t)]),
\label{FDEsolution}
\end{equation}
where
\begin{equation}
\theta(x)=\sum_{k=1}^\infty a_k\varphi_k(x),\qquad \theta(x,t)=
\sum_{k=1}^\infty a_k(t)\varphi_k(x).
\label{theta_xt}
\end{equation}
The coefficients $a_k(t)$ are obtained by applying the
semigroup $Z_{ij}(t)$ to $a_k$ (see equation \eqref{gg1i}).
The analytical expression \eqref{FDEsolution} can be
written more rigorously in terms of the action of
a semigroup $U(t)$ \cite{Engel} to $\theta(x)$, i.e.,
\begin{equation}
F([\theta],t)=F_0([U(t)\theta(x)]).
\label{semiG}
\end{equation}
As we will see, such semigroup defines a translation
in the space of functions $D(F)$. Such translation can be
generated by rotations or contractions, depending on
the space of functions $D(F)$ we consider.
Hereafter we discuss this matter in more detail.
\paragraph{Periodic Function Spaces}
Assume that the function
space $D(F)$ (domain of the solution functional $F$),
is the space of infinitely differentiable periodic
functions in $[0,2\pi]$. In this case, the matrix $C_{ij}$
defined in \eqref{matC} is skew-symmetric,
thanks to the periodicity of $\varphi_k$
(just integrate \eqref{matC} by parts ).
Therefore, by the spectral theorem, $C_{ij}$ it has purely
imaginary eigenvalues $\lambda_k=k i$,
$k\in \mathbb{Z}$.
Since $\bm C$ is skew-symmetric we have
that $\exp[-t\bm C]$ is orthogonal, i.e., it defines
an isometry in $\mathbb{R}^m$. Such isometry
generates a translation in the space of periodic
functions with group velocity equal
to one. In other words, we have
\begin{align}
\theta(x,t) =\theta(x-t).
\end{align}
Therefore, the analytical solution to the functional differential
equation \eqref{FDE3} (with $H=0$) in the space of periodic functions $D(F)$ is
\begin{equation}
F([\theta],t)=F_0([\theta(x-t)]).
\label{generalsolution}
\end{equation}
From this equation, we see that if $F_0$ is invariant under
translation, i.e., $F_0([\theta(x-t)])=F_0([\theta(x)])$ then
the solution functional is constantly equal to
the initial condition $F_0([\theta])$. This is discussed in more detail
the following two examples.
\vspace{0.7cm}
\noindent
{\em Example 1:}
Consider the initial condition
\begin{equation}
F_0([\theta])=\sin\left(\int_{0}^{2\pi}\theta(x)dx\right).
\label{FDE0i}
\end{equation}
We have seen in Section \ref{sec:approximability_of_functionals} that this
nonlinear functional is approximable by a one-dimensional
function relative to the standard Fourier (modal)
basis in $[0,2\pi]$. Specifically, we obtained
\begin{equation}
f_0(a_0,...,a_m)= \sin\left(\sqrt{2\pi} a_0\right),
\end{equation}
independently on $m$. This implies that the analytical solution
to the initial value problem \eqref{PDE-advR} is
\begin{equation}
f(a_0,...,a_m,t)=\sin\left(\sqrt{2\pi} \sum_{k=0}^m
Z_{0k}(t) a_k\right), \quad \textrm{where}\quad
Z_{0k}(t)=\left[e^{-t \bm C}\right]_{0k}.
\label{as1}
\end{equation}
It can be shown that if we sort the basis
elements as in \eqref{D2m} then $Z_{0k}(t)=\delta_{k0}$.
Therefore, the solution to \eqref{FDE3}-\eqref{FDE0i} (with $H=0$)
in the space of periodic functions is
\begin{equation}
F([\theta],t)=F_0([\theta]);
\label{thesolutionfunctional}
\end{equation}
i.e., the constant functional.
In general, the solution to a FDE is independent on the
way represent the test function space $D(F)$.
Thus, it shouldn't be surprising that we obtain exactly
the same result if we consider a finite-dimensional
expansion in terms of nodal trigonometric polynomials.
In this case, the initial condition functional can be
discretized as
\begin{equation}
f_0(a_0,...,a_m)= \sin\left(\eta \sum_{k=0}^m a_k\right),\quad
\textrm{where}\quad \eta = \int_{0}^{2\pi} \varphi_k(x) dx = \left(\frac{2\pi}{m+1}\right)
\end{equation}
and the solution is
\begin{equation}
f(a_0,...,a_m,t)= \sin\left(\eta \sum_{k=0}^m a_k(t)\right),
\label{as111}
\end{equation}
where $a_k(t)$ is defined in \eqref{as1}.
Since $\exp(-t\bm C)$ is an orthogonal matrix we have
that
\begin{equation}
\sum_{k=0}^m a_k(t) = \sum_{k=0}^m a_k.
\label{cdfg}
\end{equation}
In the limit $m\rightarrow \infty$ \eqref{as111} and \eqref{cdfg}
imply \eqref{thesolutionfunctional}. Such result can also
be obtained by directly noting that if $\theta$ is
periodic in $[0,2\pi]$ then
\begin{equation}
\int_0^{2\pi} \theta(x)dx = \int_0^{2\pi} \theta(x-t)dx\qquad
\textrm{for all $t\in \mathbb{R}$.}
\label{symF}
\end{equation}
Substituting this into \eqref{FDE0i} and \eqref{generalsolution} yields \eqref{thesolutionfunctional}.
\vspace{0.7cm}
\noindent
{\em Example 2:}
Consider the initial condition
\begin{equation}
F_0([\theta])=\exp\left[-\int_{0}^{2\pi}\theta(x)^2dx\right].
\label{expFun}
\end{equation}
where $\theta(x)$ is in the space of infinitely differentiable
periodic functions in $[0,2\pi]$. Represent $\theta(x)$
in a finite-dimensional space spanned by
any orthonormal periodic basis in $[0,2\pi]$.
This yields the following multivariate function corresponding to $F_0([\theta])$ (see also Section \ref{sec:approximability_of_functionals})
\begin{equation}
f_0(a_1,...,a_m)= \prod_{k=1}^m e^{-a_k^2}.
\label{iic1}
\end{equation}
The analytical solution to the multivariate
PDE \eqref{PDE-advR} with initial condition \eqref{iic1} and
$h=0$ is (see equation \eqref{solutionFF})
\begin{equation}
f(a_1,...,a_m,t)=\prod_{k=1}^m
e^{-a_k(t)^2},
\label{abs1}
\end{equation}
where
\begin{equation}
a_k(t)=\sum_{j=1}^m Z_{kj}(t) a_j\qquad \textrm{and}
\qquad Z_{kj}(t)=\left[e^{-t \bm C}\right]_{kj}.
\label{abs2}
\end{equation}
In a collocation setting where the the
basis functions are normalized nodal trigonometric
polynomials it is easy to see that
\begin{equation}
f(a_1,...,a_m,t) = f(a_1,...,a_m) \qquad \textrm{for all $t\geq 0$}
\end{equation}
In fact, $a_j$ are the rescaled values of
$theta(x)$ at node $x_j\in[0,2\pi]$,
the rescaling coefficient being the norm of the
trigonometric polynomial.
Since $\theta(x)$ is periodic, when $t$
increases we have that values $a_j$ are
just shifted to another location in
$[0,2\pi]$, leaving the product in \eqref{abs1} constant.
In other words, the solution to the FDE \eqref{FDE3} (with $H=0$)
corresponding to the initial condition \eqref{expFun} is
\begin{equation}
F([\theta],t) = F_0([\theta]),
\label{thesolutionfunctional1}
\end{equation}
i.e., the constant functional.
As before, this result can be obtained by
noting that if $\theta(x)$ is periodic in $[0,2\pi]$ then
\begin{equation}
\int_0^{2\pi} \theta(x)^2dx = \int_0^{2\pi} \theta(x-t)^2dx\qquad
\textrm{for all $t\in \mathbb{R}$.}
\label{symF1}
\end{equation}
A substitution of \eqref{symF1} into \eqref{expFun} yields \eqref{thesolutionfunctional1}.
\paragraph{Polynomial Function Spaces}
To obtain non-trivial solutions to equation \eqref{FDE3},
we consider the following space of functions
\begin{equation}
D(F) = \left\{\theta \in {C}^{\infty}([0,2\pi])\,
|\,\theta(0)=0\right\}
\label{Fspace9}
\end{equation}
and the initial condition \eqref{expFun}.
We generate an orthonormal polynomial basis spanning $D(F)$
by orthonormalizing the modified Chebyshev
basis $x T_k(x/\pi-1)$ in $[0,2\pi]$
though the Gram-Schmidt procedure.
The basis functions we obtain are shown
in Figure \ref{fig:basis_eigen}.
\begin{figure}[t]
\centerline{(a) \hspace{5cm} (b)\hspace{5cm}(c)}
\centerline{
\includegraphics[height=5cm]{basis-eps-converted-to.pdf}
\includegraphics[height=5cm]{spectrum_C_m20-eps-converted-to.pdf}
\includegraphics[height=5cm]{spectrum_C_m60-eps-converted-to.pdf}
}
\caption{\color{r}(a) First three orthonormal polynomial basis functions
spanning the function space \eqref{Fspace9}; (b)-(c) spectrum of
the matrix \eqref{matC} for different number of variables $m$.}
\label{fig:basis_eigen}
\end{figure}
This allows us to define the finite-dimensional function space
\begin{equation}
D_m =\left\{\theta\in C^{\infty}([0,2\pi])\,|\, \theta(x) =\sum_{k=1}^m a_k\varphi_k(x) \right\}\subseteq D(F),
\label{finite_dim_FS}
\end{equation}
where $\varphi_k$ are the orthonormal basis functions
shown in Figure \ref{fig:basis_eigen}.
The matrix $C_{ij}$ in this case is not skew-symmetric, and it has
eigenvalues with positive real part (see Figure \ref{fig:basis_eigen}).
This implies that $\bm Z(t)=\exp[-t\bm C]$ is a contraction map
that takes any function $\theta(x)\in D(F)$ and continuously deforms
it to $\theta(x)=0$ (see Figure \ref{fig:function_df}).
\begin{figure}[t]
\centerline{(a)\hspace{6cm} (b)}
\centerline{
\includegraphics[height=5.5cm]{deformation_function_analytical-eps-converted-to.pdf}
\includegraphics[height=5.5cm]{theta_xt_convergence-eps-converted-to.pdf}
}
\caption{\color{r}
(a) Evolution of $\theta(x)= x\exp[-\sin(x/2)^2]\cos(x
+4)/4$ under the (infinite-dimensional)
semigroup generated by the FDE \eqref{FDE3},
i.e, $U(t)\theta(x)$ (see Eq. \eqref{semiG});
(b) Convergence of $\theta_m(x,t)$ to $\theta(x,t)$
as we increase the number of dimensions $m$ at $t=2$.
}
\label{fig:function_df}
\end{figure}
From a dynamical system viewpoint,
$\exp[-t\bm C]\bm a$ is in fact stable spiral.
This means that if we set the initial condition as in
\eqref{abs1}, then we obtain $f(a_1,...,a_m,t)\rightarrow 1$
everywhere as $t\rightarrow \infty$. The speed at which
$f$ goes to one can be bounded by the spectral radius of $\bm C$.
The solution to the functional equation \eqref{FDE3}, with initial
condition \eqref{expFun} is
\begin{equation}
F([\theta],t)= \exp\left[-\int_{0}^{2\pi} \theta(x,t)^2dx\right],
\label{eq:SF}
\end{equation}
where $\theta(x,t)$ is defined in \eqref{theta_xt}.
More explicitly,
\begin{equation}
F([\theta],t)=
\begin{cases}
\displaystyle \exp\left[-\int_t^{2\pi}\theta(x)^2dx\right] & t\in[0,2\pi]\\
\displaystyle 1 & t> 2\pi
\end{cases}.
\label{THESOLUTION}
\end{equation}
\paragraph{Regularity of the Solution Functional}
The characteristic system associated with the
first-order PDE \eqref{PDE-advR} is
\begin{equation}
\left\{
\begin{array}{l}
\displaystyle \frac{da_k}{dt}=\sum_{j=1}^m a_j
C_{kj}\qquad
C_{kj}= \int_{0}^{2\pi}\varphi_j(x)\frac{d \varphi_k(x)}{dx}dx\\
\displaystyle a_k(0)=\int_{0}^{2\pi}\theta(x)\varphi_k(x)dx
\end{array}\right.
\end{equation}
Given any test function $\theta(x)$ in the function
space \eqref{Fspace9}, e.g., the function \eqref{thef},
the semigroup $\exp(-t\bm C)$ pushes forward in time its
Fourier coefficients, yielding the function
$\theta(x,t)$ defined in equation \eqref{theta_xt}. In particular,
if we consider the initial condition \eqref{thef} then $\theta(x,t)$
is shown in Figure \ref{fig:function_df}(a). As easily seen, such
function has a shock discontinuity moving leftwards with
velocity equal to one towards the origin as time increases.
Specifically we have $\theta(x,t)=\theta(x-t)$.
Remarkably, $\theta(x,t)$ is not in $D_m$ if $t>0$. In fact, such
function does not satisfy the boundary condition $\theta(0,t)=0$
($t>0$) and it has a shock discontinuity.
In other words, the semigroup generated by the FDE \eqref{FDE3}
immediately pushes $\theta(x)\in D_m$
out of $D_m$. This has important consequences
when we aim at approximating $\theta(x,t)$ with elements
of $D_m$. In particular, we need to use a high resolution
to resolve the jump at $x=0$ and the shock in $[0,2\pi]$
(see Figure \ref{fig:function_df}(b)).
However, we emphasize that the singularities we
just mentioned do not have any serious effect on
the regularity of the solution functional \eqref{eq:SF}.
In fact, such functional involves integration in
$x$, which is very-well defined for bounded functions
with a finite number of discontinuities
(see Figure \ref{fig:analytical_sol2}). Also, $F([\theta],t)$ is
continuous in $\theta$ and smooth in time,
thanks to the properties of the exponential semigroup.
In particular, from equation \eqref{THESOLUTION} we
see that
\begin{equation}
\theta_m\rightarrow \theta^* \quad
\Rightarrow
\quad
\left|F([\theta_m],t)- F([\theta^*],t)\right|\rightarrow 0
\end{equation}
i.e., the the solution functional is continuous. Moreover,
if we restrict the set of admissible test functions to a
function space $D(F)$ that is ``close enoug h'' to $D_m$, e.g.
in the sense of \eqref{functionaldeviation} or \eqref{knw}, then
we are allowed to say that the functional is approximable $D_m$.
In Figure \ref{fig:analytical_sol2} we show convergence of
$F([\theta_m],t)$ to $F([\theta^*],t)$ as we increase $m$,
where
\begin{equation}
\theta^*(x)= \frac{x}{4}\exp\left[-\sin\left(\frac{x}{2}\right)^2\right]
\cos(x+4),
\label{thef}
\end{equation}
and $\theta_m(x)$ is the projection of $\theta^*(x)$ in the finite
dimensional function space $D_m$ spanned by the orthonormal
polynomial basis shown in Figure \ref{fig:basis_eigen}. The function
\eqref{thef} is shown Figure \ref{fig:function_df} (case $t=0$).
\begin{figure}[t]
\centerline{(a)\hspace{6.5cm} (b)}
\centerline{
\includegraphics[height=5.5cm]{functional-eps-converted-to.pdf}
\includegraphics[height=5.5cm]{functional_error-eps-converted-to.pdf}
}
\caption{\color{r}
Functional convergence with the number of
dimensions $m$. We set $\theta^*$ as in \eqref{thef} and
evaluate $F([\theta_m],t)$ and the $L_{\infty}$ error
$| F([\theta^*],t)-F([\theta_m],t)|$ versus time for different $m$,
where $\theta_m(x)$ is the projection of $\theta^*(x)$ in the finite
dimensional function space $D_m$ spanned by the orthonormal
polynomial basis shown in Figure \ref{fig:basis_eigen}. It is seen that $m=10$ yields reasonably
accurate results.}
\label{fig:analytical_sol2}
\end{figure}
\paragraph{Functional Derivatives}
The first-order functional derivative of the solution
functional \eqref{THESOLUTION} is
\begin{equation}
\frac{\delta F([\theta],t) }{\delta \theta(x)}=
\begin{cases}
\displaystyle -2 \theta(x) \exp\left[-\int_t^{2\pi}\theta(x)^2 dx\right] & t\in[0,2\pi],\\
\displaystyle -2 \theta(x) & t>2\pi.
\end{cases}
\label{FunctionalDerivative}
\end{equation}
Note that the functional derivative at time $t$ evaluated at
$\theta(x)$ is simply a rescaled version of $\theta(x)$,
where the scaling factor grows from
$2\exp\left[\int_0^{2\pi}\theta(x)dx\right]$ (at $t=0$)
to $2$ (at $t=2\pi$).
Such derivative can be expressed in $D_m$ as
\begin{equation}
\left.\frac{\delta F([\theta],t)}{\delta \theta(x)}\right|_{D_m}=
\sum_{j=1}^m\frac{\partial f}{\partial a_j}\varphi_j(x),
\label{ffd3}
\end{equation}
where
\begin{equation}
f(a_1,...,a_m,t)=\prod_{j=1}^m \exp\left[-\left(\sum_{k=1}^m
Z_{jk}(t)a_k\right)^2\right].
\end{equation}
More explicitly,
\begin{equation}
\left.\frac{\delta F([\theta],t)}{\delta \theta(x)}\right|_{D_m}=
-2f(a_1,...,a_m,t) \sum_{p=1}^m \varphi_p(x) \sum_{k,j=1}^m
Z_{jp}(t) Z_{jk}(t)a_k.
\label{FDeRT}
\end{equation}
In Figure \ref{fig:sol2D} we plot the analytical solution \eqref{eq:SF} in
two dimensions.
\begin{figure}
\centerline{\footnotesize$t=0$\hspace{3.2cm}$t=1$\hspace{3.2cm}$t=3$\hspace{3.2cm}$t=6$}
\centerline{
\includegraphics[height=3.5cm]{2D_sol_t0-eps-converted-to.pdf}\hspace{-0.7cm}
\includegraphics[height=3.5cm]{2D_sol_t1-eps-converted-to.pdf}\hspace{-0.7cm}
\includegraphics[height=3.5cm]{2D_sol_t3-eps-converted-to.pdf}\hspace{-0.7cm}
\includegraphics[height=3.5cm]{2D_sol_t6-eps-converted-to.pdf}
}
\caption{\color{r}
Solution to the multivariate PDE \eqref{PDE-advR}
(with $h=0$) in $m=2$ dimensions. The stable spiral at the origin of the characteristic system attracts every curve in the space of modes and ultimately yields $f=1$ everywhere after a transient (see also Figure \ref{fig:analytical_sol2}).}
\label{fig:sol2D}
\end{figure}
The side length of the hypercube cube that encloses
any level set of the solution at time $t$ depends on
the number of dimensions $m$. In particular, if we are
interested in estimating the size of the hypercube that encloses
the set $\{\bm a\in \mathbb{R}^m \,|\, f(\bm a,t)\geq \epsilon\}$, then
we can use the formula \footnote{\color{r}
The inequality \eqref{est:b} can
be obtained by noticing that the solution $f(\bm a,t)$ is in the form
\begin{equation}
f(\bm a,t) = e^{-\bm a^T \bm S(t) \bm a}\qquad \bm S(t)= \bm Z(t)^T \bm Z(t),
\end{equation}
i.e.,
\begin{equation}
-\log(f(\bm a,t)) = \bm a^T \bm S(t)\bm a.
\end{equation}
By diagonalizing $\bm S(t)$ we obtain
\begin{equation}
(\bm V(t) \bm a)^T \bm \Lambda(t) (\bm V(t) \bm a) = - \log(f(\bm a,t)).
\end{equation}
Upon definition of the rotated coordinate system
$\bm y(t)=\bm V(t) \bm a$, we find that the largest
semi-axis of the ellipse representing the $\epsilon$ level
set of $f(\bm a,t)$ is
\begin{equation}
\frac{1}{2}\sqrt{\frac{-\log(\epsilon)}{\lambda_{min}}},
\end{equation}
where $\lambda_{min}$ is the smallest eigenvalue of $\bm S(t)$.
This formula assumes that there is no rotation in the Gaussian
function $f(\bm a,t)$ during the dynamics and therefore it provides a
conservative upper bound $b$ that coincides with the largest semi-axis
of the ellipsoidal level set. On the other hand, there exist a
rotation of the ellipsoid that minimizes the size of the aforementioned
hypercube. Such rotation aligns the largest semi-axis with the diagonal
of the hypercube.
We recall that the diagonal of a hypercube in dimension $m$ has length
$\sqrt{m}\ell$, where $\ell$ is the side length of the hypercube.
Therefore, the upper and the lower bound estimates for $b$, i.e, the
half side length of the hypercube that encloses the $-\log(\epsilon)$
level set of the solution are
\begin{equation}
\frac{1}{2}\sqrt{\frac{-\log(\epsilon)}{ m \lambda_{min}}}\leq b
\leq \frac{1}{2}\sqrt{\frac{-\log(\epsilon)}{\lambda_{min}}}.
\end{equation}
}
\begin{equation}
\frac{1}{2}\sqrt{\frac{-\log(\epsilon)}{ m \lambda_{min}}}\leq b
\leq \frac{1}{2}\sqrt{\frac{-\log(\epsilon)}{\lambda_{min}}}.
\label{est:b}
\end{equation}
where $\lambda_{min}$ is the smallest eigenvalue of the matrix $\bm Z(t)^T \bm Z(t)$ (see Equation \eqref{gg1i}).
In Figure \ref{fig:enclosing_cube} we plot
the upper and the lower bound estimates of
half of the side length of the hypercube that encloses the $10^{-10}$
level set of the solution at $t=1$
versus the number of dimensions $m$. It is seen that
the size of the hypercube increases exponentially fast with $m$.
This has important consequences when it comes to numerical
simulations. In particular, if we perform simulations with
far field boundary conditions, then the size of the computational
domain should be chosen large enough to accommodate the
support of the solution throughout the simulation time interval
of interest.
\begin{figure}[t]
\centerline{\hspace{0.2cm}$t=1$ \hspace{7.cm} $t=3$}
\centerline{
\includegraphics[height=6cm]{range_b-eps-converted-to.pdf}
\includegraphics[height=6cm]{range_b_t3-eps-converted-to.pdf}
}
\caption{\color{r}Upper and lower bound estimates \eqref{est:b} of
the half side length of the hypercube that encloses the $10^{-10}$
level set of the solution at $t=1$ and $t=3$ versus the number of
dimensions $m$. It is seen that $b$ increases
exponentially fast with $m$.
The computational domain should be at least as
big as $[-b,b]^m$, where $b$ is the number given by the
upper curve. For example, if we are aiming at resolving the
numerical solution in the case $m=6$ within the time interval
$[0,3]$, then we should consider the computational
domain $[-b,b]^6$ with $b\simeq 1600$.}
\label{fig:enclosing_cube}
\end{figure}
\subsection{Numerical Discretization}
\label{sec:numericalDiscretizationADVR}
Consider the multivariate PDE \eqref{PDE-advR} with $h=0$,
hereafter written in the operator form
\begin{equation}
\frac{\partial f}{\partial t}=L f,
\label{PDE11}
\end{equation}
where
\begin{equation}
L= - \sum_{j,k=1}^ma_kC_{jk}\frac{\partial }{\partial a_j}.
\label{LadvR1}
\end{equation}
Note that $L$ is a {\em separable} linear
operator with separation rank $r_L=m^2$.
In fact, $L$ can be written in the form
\begin{equation}
L= \sum_{q=1}^{m^2} \alpha_q L^q_1(a_1)\cdots L_m^q(a_m),
\label{eq:operators}
\end{equation}
for suitable one-dimensional linear operators $L^q_j(a_j)$ defined in Table \ref{tab:ordering}.
\begin{table}\color{r}
\centering
\begin{tabular}{c|ccccc}
$q$ & $\alpha_q$ & $L_1^q$ & $L_2^q$ & $\cdots$ & $L_m^q$\\
\hline\\
$1$ & $-C_{11}$ & $a_1\partial_{a_1}$ & $1$ & $\cdots$& $1$\\
$2$ &$-C_{12}$ & $ \partial_{a_1}$ & $a_2$ & $\cdots$& $1$\\
$\vdots $& $\vdots $ & $\vdots$ & $\vdots$ &$\vdots$ &$\vdots$ \\
$m$ &$-C_{1m}$ & $\partial_{a_1}$ & $1$ & $\cdots$& $a_m$\\
$m+1$ &$-C_{21}$ & $a_1$ & $\partial_{a_2}$ & $\cdots$& $1$\\
$m+2$ &$-C_{22}$ & $1$ & $a_2\partial_{a_2}$ & $\cdots$& $1$\\
$\vdots $& $\vdots $ & $\vdots$ & $\vdots$ &$\vdots$ &$\vdots$ \\
$2m$ &$-C_{2m}$ & $1$ & $\partial_{a_2}$ & $\cdots$& $a_m$\\
$\vdots $& $\vdots $ & $\vdots$ & $\vdots$ &$\vdots$ &$\vdots$ \\
$m^2-m+1$ &$-C_{m1}$ & $a_1$ & $1$ & $\cdots$& $\partial_{a_m}$\\
$m^2-m+2$ &$-C_{m2}$ & $1$ & $a_2$ & $\cdots$& $\partial_{a_m}$\\
$\vdots $& $\vdots $ & $\vdots$ & $\vdots$ &$\vdots$ &$\vdots$ \\
$m^2$ &$-C_{mm}$ & $1$ & $1$ & $\cdots$& $a_m\partial_{a_m}$\\
\end{tabular}
\caption{\color{r}Ordering of the linear operators defined
in equation \eqref{eq:operators}.}
\label{tab:ordering}
\end{table}
The solution to the multivariate PDE \eqref{PDE11} can be
represented by using any tensor series expansion. In particular,
hereafter we consider the hierarchical Tucker format (Section \ref{sec:CP}) and the canonical polyadic tensor
decomposition (Section \ref{sec:CTD-ALS}).
\subsubsection{HT and CP Algorithms with Explicit Time Stepping}
\label{sec:HTD}
Let us discretize the PDE \eqref{PDE11} in time by using any explicit
time stepping scheme, for example the second-order Adams-Bashforth scheme (see Section \ref{sec:ADI_SSE})
\begin{equation}
f_{n+2}=f_{n+1}+\frac{\Delta t}{2} L\left(3f_{n+1}-f_n\right).
\label{AB2_}
\end{equation}
In this setting, we see that the only operations needed to
compute $f_{n+2}$ with tensor methods are: i) addition,
ii) application of a (separable) linear operator, and iii) rank
reduction\footnote{On the other hand, if we employ
implicit time-discretization schemes such as \eqref{CNFDE} we
end up solving linear systems with tensor methods. This was
addressed, e.g., in \cite{Ortega,Etter,Reynolds}.}, the last
operation being the most important among all three.
From a computational viewpoint, it would be also very useful if we
could split the tensor operations yielding $f_{n+2}$ into sequences
of tensor operations followed by rank reduction. In this way
we could minimize the storage requirements and the overall
computational cost. For example, we could split \eqref{AB2_}
as
\begin{enumerate}
\item Compute a low rank representation of $w_{n+1}=(3f_{n+1} - f_{n})$,
\item Compute a low rank representation of
$q_{n+1}= L w_{n+1}$,
\item Compute a low rank representation of
$f_{n+2}=f_{n+1} +\Delta t q_{n+1}/2$.
\end{enumerate}
Are we allowed to do so? Unfortunately no.
Splitting tensor operations into
sequences of tensor operations followed by
rank reduction usually yields severe cancellation
errors.
In some cases, this problem can be overcome. For example,
an efficient and robust algorithm that allows us to split
sums and rank reduction operations was recently proposed
in \cite{Kressner2014} in the context of
hierarchical Tucker formats. The algorithm leverages on
the block diagonal structure that arises when adding
hierarchical Tucker format.
Also, the vector resulting from application of the separable
linear operator $L$ to the hierarchical Tucker
format $w_{n+1}$ at point ii) above
can be computed very efficiently if we have available
a hierarchical Tucker representation of the operator $L$.
Such representation can be easily constructed in a Fourier
collocation setting by vectorizing all operators in table
\ref{tab:ordering}. Each one-dimensional
operator is represented relative to a basis trigonometric (nodal)
polynomials (Fourier spectral collocation method) \cite{Hesthaven}
In practice, we simply need to convert the
vectorized canonical polyadic series of $L$ into a hierarchical
Tucker expansion, which is a relative standard operation.
\subsubsection{CP-ALS Algorithm with Implicit Time Stepping}
\label{sec:CTD-ALS}
Let us discretize the PDE \eqref{PDE11} in time by using the
Crank-Nicolson method. This yields
\begin{equation}
\left[I-\frac{\Delta t}{2} L \right]f_{n+1}=
\left[I+\frac{\Delta t}{2} L \right]f_n.
\label{CNFDE}
\end{equation}
This equation can be written in a compact notation as
\begin{equation}
Af_{n+1}=Bf_n,
\end{equation}
where
\begin{equation}
A = I-\frac{\Delta t}{2} L, \quad \textrm{and}\quad B = I+\frac{\Delta t}{2} L.
\end{equation}
By using the definition of $L$ given in \eqref{eq:operators}
(see also Table \ref{tab:ordering}), it is clear that both $A$ and $B$ are
separable operators in the form
\begin{align}
A= 1\cdots 1 +\frac{\Delta t}{2} \sum_{i,j=1}^m a_iC_{ji}
\frac{\partial }{\partial a_j},\label{AL}\\
B= 1\cdots 1 - \frac{\Delta t}{2} \sum_{i,j=1}^m a_iC_{ji}
\frac{\partial }{\partial a_j}.\label{BL}
\end{align}
These operators have separation rank $r_A=r_B=m^2+1$ and can
be conveniently written as
\begin{align}
A= \sum_{q=0}^{m^2} \eta_q E^q_1(a_1)\cdots E^q_m(a_m), \qquad
B= \sum_{q=0}^{m^2} \zeta_q E^q_1(a_1)\cdots E^q_m(a_m),\label{BLs}
\end{align}
where all quantities are defined in Table \ref{tab:AL}.
The difference with Table \ref{tab:ordering} is that
we added one row (the zeroth one) to represent the
identity operator $1\cdots 1$, and we rescaled
all coefficients $C_{ij}$.
\begin{table}\color{r}
\centering
\begin{tabular}{c|cccccc}
$q$ & $\eta_q$ & $\zeta_q$ & $E_{1}^q$ & $E_{2}^q$ & $\cdots$ & $E_{m}^q$\\
\hline\\
$0$ & $1$ &$1$ & $1$ & $1$ & $\cdots$& $1$\\
$1$ & $\Delta t C_{11}/2$ & $-\Delta t C_{11}/2$ & $a_1\partial_{a_1}$ & $1$ & $\cdots$& $1$\\
$2$ &$\Delta t C_{12}/2$ &$-\Delta t C_{12}/2$& $\partial_{a_1}$ & $a_2$ & $\cdots$& $1$\\
$\vdots $& $\vdots $ & $\vdots$ & $\vdots$ &$\vdots$ &$\vdots$ \\
$m$ &$\Delta tC_{1m}/2$ & $-\Delta tC_{1m}/2$ & $\partial_{a_1}$ & $1$ & $\cdots$& $a_m$\\
$m+1$ &$\Delta tC_{21}/2$ & $-\Delta tC_{21}/2$& $a_1$ & $\partial_{a_2}$ & $\cdots$& $1$\\
$m+2$ &$\Delta tC_{22}/2$ & $-\Delta tC_{22}/2$ & $1$ & $a_2\partial_{a_2}$ & $\cdots$& $1$\\
$\vdots $& $\vdots $ & $\vdots$ & $\vdots$ &$\vdots$ &$\vdots$ \\
$2m$ &$\Delta tC_{2m}/2$ &$-\Delta tC_{2m}/2$& $1$ & $\partial_{a_2}$ & $\cdots$& $a_m$\\
$\vdots $& $\vdots $ & $\vdots$ & $\vdots$ &$\vdots$ &$\vdots$ \\
$m^2-m+1$ &$\Delta tC_{m1}/2$&$-\Delta tC_{m1}/2$ & $a_1$ & $1$ & $\cdots$& $\partial_{a_m}$\\
$m^2-m+2$ &$\Delta tC_{m2}/2$ &$-\Delta tC_{m2}/2$& $1$ & $a_2$ & $\cdots$& $\partial_{a_m}$\\
$\vdots $& $\vdots $ & $\vdots$ & $\vdots$ &$\vdots$ &$\vdots$ \\
$m^2$ &$\Delta tC_{mm}/2$&$-\Delta tC_{mm}/2$ & $1$ & $1$ & $\cdots$& $a_m\partial_{a_m}$\\
\end{tabular}
\caption{\color{r}Ordering of the linear operators $A$ and $B$ defined
in \eqref{BLs}.}
\label{tab:AL}
\end{table}
A substitution of the CP decomposition
\begin{equation}
\hat{f}_{n+1} = \sum_{l=1}^r \prod_{k=1}^m{G^l_k(a_k,t_{n+1})}
\label{CP-numerical}
\end{equation}
into equation \eqref{CNFDE} yields the residual
\begin{equation}
R = A\hat{f}_{n+1}-B\hat{f}_n.
\label{rRr}
\end{equation}
Minimization of the $L_2$ norm of \eqref{rRr} with respect to
${\bm\beta}_q(t_{n+1})$ yields the linear systems
of equations
\begin{equation}
\bm M^L_q \bm \beta_q(t_{n+1}) = \bm M^R_q \bm \beta_q(t_{n}),\qquad q=1,...,m
\label{ALS2}
\end{equation}
where
\begin{equation}
\bm M^L_q = \sum_{e,z=0}^{m^2} K^L_{ez}
\left[
{\displaystyle \substack{\vspace{0.2cm}\\\vspace{0.04cm} m\\\bigcirc\\k=1\\k\neq q}}
\hat{\bm \beta}_{k}(t_{n+1})^T \bm
E^{ez}_k \hat {\bm \beta}_{k}(t_{n+1})
\right]^T\otimes \left[{\bm E}^{ez}_q\right]^T,
\label{ML}
\end{equation}
\begin{equation}
\bm M^R_q = \sum_{e,z=0}^{m^2} {K}^R_{ez}
\left[
{\substack{\vspace{0.2cm}\\\vspace{0.04cm} m\\\bigcirc\\k=1\\k\neq q}}
\hat{\bm \beta}_{k}(t_{n})^T
{\bm E}^{ez}_k \hat{\bm \beta}_{k}(t_{n+1})
\right]^T\otimes \left[{\bm E}^{ez}_q\right]^T,
\label{MR}
\end{equation}
and
\begin{equation}
\left[{\bm E}^{ez}_q\right]_{sh}=\int_{-b}^b E_{q}^e(a)\phi_s(a) E_{q}^z(a)\phi_h(a)da.
\label{eq:77}
\end{equation}
In equations \eqref{ML}-\eqref{MR},
$\bigcirc$ denotes the Hadamard matrix
product, $\otimes$ is the Kroneker matrix product,
$\hat {\bm \beta}_{k}$ is the matrix version of $\bm \beta_k$, i.e.,
\begin{equation}
\hat {\bm \beta}_{k}(t_n)=
\left[\begin{array}{ccc}
\beta^1_{k1}(t_n)&\cdots& \beta^r_{k1}(t_n)\\
\vdots &\ddots & \vdots \\
\beta^1_{kQ}(t_n) &\cdots & \beta^r_{kQ} (t_n)
\end{array}
\right],
\end{equation}
$\bm E^{ez}_q$ is the $Q\times Q$ matrix \eqref{eq:77},
and $K^L_{ez}$ and ${K}^R_{ez}$ are entries of the
matrices
\begin{equation}
\bm K^L = \bm \eta \bm \eta^T,\quad \textrm{and}\quad
\bm K^R = \bm \eta \bm \zeta^T,
\end{equation}
where $\bm \eta$ and $\bm \zeta$ are column vectors with entries $\eta_q$ and $\zeta_q$ defined in Table \ref{tab:AL}.
\paragraph{Computing the Matrix System}
There are many of symmetries we can exploit when
constructing the separated series expansion of the
operators $A$ and $B$ in \eqref{BLs}.
Indeed, a closer look at Table \ref{tab:AL} suggests that
if we employ the same series expansion in each
variable $a_j$ (e.g., a trigonometric series)
then the number of operators in \eqref{BLs} that
we effectively need to compute reduces to the
following four
\begin{equation}
1, \qquad a_j, \qquad \frac{\partial}{\partial a_j}, \qquad a_j\frac{\partial}{\partial a_j}.
\end{equation}
This means that the number of terms that are effectively
different in the fundamental matrix \eqref{eq:77} are {\em only 12}
(9 if we are willing to employ matrix transposes).
Specifically,
\begin{equation}
\begin{array}{l l l l}
\displaystyle 1.\, \int_{-b}^b \phi_s \phi_h da \quad &
\displaystyle 2.\,\int_{-b}^b a \phi_s \phi_h da, \quad &
\displaystyle 3.\,\int_{-b}^b \phi_s\frac{d\phi_h}{da} da,\quad &
\displaystyle 4.\,\int_{-b}^b a \phi_s \frac{d\phi_h}{da}da,\\
\displaystyle 5.\, \int_{-b}^b a^2 \phi_s \phi_h da, \quad &
\displaystyle 6\,\int_{-b}^b a^2 \phi_s\frac{d\phi_h}{da} da, \quad &
\displaystyle 7.\,\int_{-b}^b \frac{d\phi_s}{da} \phi_h da, \quad &
\displaystyle 8.\, \int_{-b}^b a \frac{d\phi_s}{da} \phi_h da, \\
\displaystyle 9.\,\int_{-b}^b\frac{d\phi_s}{da}\frac{d\phi_h}{da} da, \quad &
\displaystyle 10.\,\int_{-b}^b a \frac{d\phi_s}{da} \frac{d\phi_h}{da} da, \quad &
\displaystyle 11.\, \int_{-b}^b a^2 \frac{d\phi_s}{da} \phi_h da, \quad &
\displaystyle 12.\,\int_{-b}^b a^2 \frac{d\phi_s}{da}\frac{d\phi_h}{da} da.
\end{array}
\label{eq:all integrals}
\end{equation}
All these integrals can be pre-computed and stored as $Q\times Q$
matrices (see Eq. \eqref{eq:50}). For each $e$, $z$
and $q$, the tensor \eqref{eq:77} corresponds to one of the $12$
integrals above. Such map, denoted as $g^{ez}_q$,
takes in the triple $(e,z,q)$,
where $q\in \{1,...,m\}$ and $e,z\in\{0,...,m^2\}$,
and it returns a number between $1$ and $12$ identifying
which integral in the set \eqref{eq:operators} corresponds
to the tensor entry in \eqref{eq:77}. For example, if we sort
the operators as in Table \ref{tab:AL} then in $m=2$
dimensions we have
\begin{equation}
g^{ez}_{1}=\left[
\begin{array}{ccccc}
1 & 8 & 7& 2 & 1\\
4 & 12 & 10 & 6 & 4\\
3 & 10 & 9 & 4 & 3\\
2 & 11 & 8 & 5 & 2\\
1 & 8 & 7 & 2 & 1
\end{array}
\right]\qquad \qquad
g^{ez}_{2}=\left[
\begin{array}{c c c c c}
1 & 1 & 2 &7 &8\\
1 &1 & 2 & 7 & 8\\
2 &2 & 5 & 8 &11\\
3 &3 & 4 & 9 &10\\
4 &4 & 6 &10 & 12
\end{array}
\right],
\label{dim2}
\end{equation}
while in dimension $m=3$ we have, e.g.,
\begin{equation}
g^{ez}_{1}=\left[
\begin{array}{cccccccccc}
1 & 8 & 7 & 7 & 2 & 1 & 1 & 2 & 1& 1\\
4 & 12 & 10& 10& 6 & 4 & 4 & 6 & 4& 4\\
3 & 10 & 9 & 9 & 4 & 3 & 3 & 4 & 3& 3\\
3 & 10 & 9 & 9 & 4 & 3 & 3 & 4 & 3& 3\\
2 & 11 & 8 & 8 & 5 & 2 & 2 & 5 & 2& 2\\
1 & 8 & 7 & 7 & 2 & 1 & 1 & 2 & 1& 1\\
1 & 8 & 7 & 7 & 2 & 1 & 1 & 2 & 1& 1\\
2 & 11 & 8 & 8 & 5 & 2 & 2 & 5 & 2& 2\\
1 & 8 & 7 & 7 & 2 & 1 & 1 & 2 & 1& 1\\
1 & 8 & 7 & 7 & 2 & 1 & 1 & 2 & 1& 1
\end{array}
\right].
\label{dim3}
\end{equation}
A combinatorial argument shows that the number of
entries equal to $1$, $2$, $3$, etc., in each matrix
$g_{1}^{ez}$, $g_{2}^{ez}$, ..., $g_{m}^{ez}$ is the same
(for fixed $m$). For instance, in \eqref{dim2} we
have $4$ ones, $2$ threes, $3$ fours, $1$
five, etc. This is very useful when we break the sum
in $e$ and $z$ in \eqref{ML} and \eqref{MR} into
multiple sums, and use the associative property of the
tensor product to reduce the number of operations.
By using the map $g_q^{ez}$ we can immediately
identify each matrix $\bm E_q^{ez}$. For example,
in the case $m=2$ we have (see Eq. \eqref{dim2})
\begin{equation}
\left[{\bm E}^{11}_1\right]_{sh}=\int_{-b}^b \phi_s(a)\phi_h(a) da, \qquad
\left[{\bm E}^{12}_1\right]_{sh}=\int_{-b}^b a\frac{d\phi_s(a)}{da}\phi_h(a) da,\quad ...\quad .
\end{equation}
\paragraph{Summary of the Algorithm}
We first compute all integrals in \eqref{eq:all integrals}
and store them in $12$ matrices $Q\times Q$,
$Q$ being the number of degrees of freedom in each
variable (e.g., collocation points of Fourier modes).
We also set up the map between such set
of matrices and any element of the tensor \eqref{eq:77}.
Such map basically takes in the triple $(q,e,z)$,
where $q\in \{1,...,m\}$ and $e,z\in\{0,...,m^2\}$,
and it returns a number between $1$ and $12$ identifying
which integral in the set \eqref{eq:operators}
corresponds to the tensor entry in \eqref{eq:77}. In
a matrix setting, this basically allows us to efficiently
compute each matrix $\bm E_q^{ez}$ appearing
in \eqref{ML} and \eqref{MR}.
Next, we compute the compute the canonical tensor
decomposition of the initial condition $f_0(a_1,...,a_m)$,
by applying the methods we described in Section \ref{sec:CP}.
This gives us the set of vectors
$\{\bm \beta_1(t_0),...,\bm \beta_m(t_0)\}$.
With such vectors available, we can build the matrices
$\bm M^L_1$ and $\bm M^R_1$ defined in \eqref{ML} and
\eqref{MR}. To this end, we need an initial guess
for $\{\bm \beta_1(t_1),...,\bm \beta_m(t_1)\}$ which we can
take to be equal to $\{\bm \beta_1(t_0),...,\bm \beta_m(t_0)\}$,
or a small random perturbation of it.
With $\bm M^L_1$ and $\bm M^R_1$ in place and set, we
can solve the linear system \eqref{ALS2} and
update $\bm \beta_1(t_1)$. At this point
we recompute $\bm M^L_2$ and $\bm M^R_2$ (with the updated
$\bm \beta_1(t_1)$) and solve for $\bm \beta_2(t_1)$.
We repeat this process for $q=3,...,m$ and iterate over
and over among all variables until convergence.
Parallel versions of the ALS algorithm
were recently proposed by Karlsson {\em et al.}
in \cite{Karlsson}.
\subsubsection{Long-Term Integration}
\label{sec:long-term-integration}
A rigorous error analysis of the HT and CP-ALS
algorithms to solve the multivariate
PDE \eqref{PDE11} goes beyond the scope of this report
(see \cite{Bachmayr} for a recent account).
It it useful, however, to point out a few things on
the nature of the discretization error, in particular
on how the temporal local truncation error depends on
the dimension $m$. To this end, let us first recall that
the local truncation error at time $t_{n+1}$ of
the second-order Adams Bashforth
(AB2) scheme applied to the linear PDE \eqref{PDE11} is
$5\Delta t^3 L^3 f(\eta,\bm a)/12$, where $\eta$ is
some time instant between $t_n$ and $t_{n+1}$, $f(\eta,\bm a)$
is the exact solution \eqref{solutionFF}, and $L$ is defined in \eqref{LadvR1}. The operator $L^2$ can be explicitly written as
\begin{align}
L^2 =& \sum_{i,j,l,p=1}^m x_lC_{pl}\frac{\partial }{\partial a_p}\left(
x_jC_{ij}\frac{\partial }{\partial a_j}\right)\nonumber\\
=& \sum_{i,j,l,p=1}^m x_lC_{pl}C_{ij}\left(\delta_{jp}\frac{\partial }{\partial a_j}+x_j\frac{\partial^2}{\partial a_j\partial a_p}\right),
\label{L2}
\end{align}
while $L^3$ has the form
\begin{align}
L^3 =& \sum_{i,j,l,p,q,z=1}^m x_q C_{zq}C_{pl}C_{ij}\delta_{lq}\left(\delta_{jp}\frac{\partial }{\partial a_j}+x_j\frac{\partial^2}{\partial a_j\partial a_p}\right)+\nonumber\\
& \sum_{i,j,l,p,q,z=1}^m x_q x_l C_{zq}C_{pl}C_{ij}\left(\delta_{jp}\frac{\partial^2 }{\partial a_j\partial a_z}+\delta_{jz}\frac{\partial^2}{\partial a_j\partial a_p}+x_j\frac{\partial^3}{\partial a_j\partial a_p\partial a_z}\right).
\label{L3}
\end{align}
From the last expression it is clear that any
inaccuracy in the computation of the derivatives
adds up to the temporal truncation error with
at least with factor $m^6$. Indeed, as shown
in Figure \ref{fig:truncation}, the norm of $L^3 f$
grows faster than $m^6$. This has important consequences
on the accuracy attainable with the AB2 time-integration
scheme applied to high-dimensional linear PDEs.
\begin{figure}
\centerline{
\includegraphics[height=6.5cm]{truncation_error-eps-converted-to.pdf}
}
\caption{\color{r} Norm of $L^3 f$ at initial time versus
$m$. The operator $L^3$, defined \eqref{L3}, has been
discretized in space by using an accurate
Fourier spectral method with $N=600$ points
in each dimension. It is seen that the norm of $L^3 f_0$ grows
faster than $m^6$. This has important consequences
on the local truncation error generated by AB2 time-integrator
applied to the PDF \eqref{PDE11}.}
\label{fig:truncation}
\end{figure}
In fact, suppose that $\left\| L^3 f\right\|$ grows
like $m^6$ not just at initial time but at each time
step (it actually grows faster than that), i.e.,
\begin{equation}
\left\| L^3 f\right\| \sim \kappa m^6,
\label{te}
\end{equation}
where $\kappa$ is a suitable constant. If
we want the AB2 scheme to operate
at constant truncation error for different number of
dimensions $m$ then we need to guarantee that
\begin{equation}
m_1^6 \Delta t_1^3 = m_2^6 \Delta t_2^3.
\label{re}
\end{equation}
For example, if $m_1=2$ and $m_2=10$ then we have
\begin{equation}
\Delta t_2 = \frac{\Delta t_1}{25}.
\end{equation}
In other words, if we want our time integrator to operate at constant
truncation error, then we need to run the simulation in $10$ dimensions
with a time step that is roughly $25$ times smaller
than the one we emply in the simulation in $2$ dimensions.
This can tax the computational resources quite substantially.
In fact, suppose we are interested in integrating our PDE up to $T=1$,
and we set $\Delta t_1=10^{-4}$ in $2$ dimensions, i.e.,
$10^4$ time steps. Assuming that the local truncation error
is the same at each time step (see Eq. \eqref{te}), at the end
of the integration period we accumulated an error of approximately
$10^{-6} \kappa$. The same error is roughly
attained at $T=0.04$ if we integrate the PDE in
$10$ dimensions.
In fact, the time step $\Delta t_2=\Delta t_1/25$ guarantees a
constant local truncation error which adds up to $10^{-6} \kappa$
after just $0.04$ time units. The local truncation error
manifest itself as {\em numerical diffusion}
which eventually dissipates the numerical solution
to zero.
Note that in this simple calculation we did not take into account
the accuracy of the rank reduction process in the CP-ALS
and HT algorithms, which takes place at each time step.
\subsection{Numerical Results}
We solve the multivariate PDE \eqref{PDE11} in
the hypercube $[-b,b]^m$, with $b=60$ and
variable $m$. Such domain is chosen large
enough to accommodate periodic (zero) boundary
conditions in the integration period of interest.
We study both the HT (hierarchical Tucker) and the
CP-ALS schemes we discussed in Section \ref{sec:HTD}
and Section \ref{sec:CTD-ALS}. Specifically, we implemented
a Fourier collocation method with $600$ nodes in each variable
and explicit AB2 time stepping.
To study the accuracy of the numerical solution,
we consider the time-dependent relative error
\begin{equation}
\epsilon_m(t) = \left|\frac{F([\theta_m^*],t)-\hat{F}([\theta_m^*],t)}{F([\theta_m^*],t)}\right|,
\label{relative_error}
\end{equation}
where $F$ is the analytical solution \eqref{eq:SF}, $\hat{F}$
is the numerical solution we obtained by using the
CP-ALS or the HT algorithms in the test function
space $D_m$ (i.e., $m$ with independent variables)
and with separation rank $r$.
The test function $\theta^*$ in \eqref{relative_error}
is defined as
\begin{equation}
\theta_m^{*}(x)= h \sum_{j=1}^m \varphi_j(x),\qquad
h= 0.698835274542439,
\label{point}
\end{equation}
where $\varphi_j(x)$ are the
orthonormal polynomials shown
in Figure \ref{fig:basis_eigen}.
The accuracy of the CP-ALS and the HT algorithms
is studied in Figure \ref{fig:ALS-CP-HT}, where
we plot the relative pointwise error \eqref{relative_error}
for different separation ranks $r$ and for
different number of dimensions.
It is seen that, as expected,
the accuracy of the numerical solution
increases as we increase the separation rank.
Also, as we increase the number of dimensions from
$2$ to $6$ the relative error increases, in agreement
with the results of Section \ref{sec:long-term-integration}
(we emply a constant $\Delta t=2.55\times 10^{-4}$
in all our simulations).
It is worthwhile emphasizing that the CP-ALS
is a randomized algorithm which requires
initialization at each time-step. This means that
results of simulations with the same nominal
parameters may be different. On the other hand, tensor
methods based on multivariate/distributed singular
value decomposition, such as the hierarchical Tucker
decomposition \cite{Grasedyck2017}, do
not suffer from this issue.
The CP-ALS algorithm is faster than HT but, as we just said,
accuracy control may be an issue.
The separation rank of both the CP-ALS
and HT algorithms are computed adaptively up
to the maximum value $r_{max}$ specified
in the legend of Figure \ref{fig:ALS-CP-HT}.
\begin{figure}
\centerline{\hspace{0.5cm}CP-ALS\hspace{6.5cm} HT}
\noindent
\centerline{
\rotatebox{90}{\hspace{2cm}two dimensions}\hspace{0.2cm}
\includegraphics[height=5.5cm]{L-infty_ALS_m2-eps-converted-to.pdf}
\includegraphics[height=5.5cm]{L-infty_HT_m2-eps-converted-to.pdf}
}
\noindent
\centerline{
\rotatebox{90}{\hspace{2cm}three dimensions}\hspace{0.2cm}
\includegraphics[height=5.5cm]{L-infty_ALS_m3-eps-converted-to.pdf}
\includegraphics[height=5.5cm]{L-infty_HT_m3-eps-converted-to.pdf}
}
\centerline{
\rotatebox{90}{\hspace{2cm}six dimensions}\hspace{0.2cm}
\includegraphics[height=5.5cm]{L-infty_ALS_m6-eps-converted-to.pdf}
\includegraphics[height=5.5cm]{L-infty_HT_m6-eps-converted-to.pdf}
}
\centerline{$--$ $r_{max}=4$\hspace{1cm} --- $r_{max}=8$
\hspace{1cm} $-\cdot -$ $r_{max}=12$
}
\caption{\color{r}Accuracy of the numerical solution
to the FDE \eqref{FDE3} in the finite dimensional function space
\eqref{finite_dim_FS} for different number of dimensions.
Specifically, we plot the relative $L_\infty$
error \eqref{relative_error} versus time
we obtained by using the CP-ALS and HT algorithms
we discussed in Section \ref{sec:numericalDiscretizationADVR}.
The separation rank of both CP-ALS and HT is computed adaptively
at each time step up to the maximum value $r_{max}$.}
\label{fig:ALS-CP-HT}
\end{figure}
Note that in two dimensions the CP-ALS algorithm
results in error plots that looks very similar when
$r_{max}=8$ and $r_{max}=12$. This is because the
separation rank is less than $8$ in both cases
throughout the simulation up to $t=1$.
The variability of the results is related to the
random initialization required by the ALS algorithm
at each time step. On the other hand,
in three and six dimensions the error plots we obtain
for $r_max=4$ and $r_max=8$ are of the same
order of magnitude because such separation ranks
are achieved after just few time steps and they
are not sufficient to accurately represent the
multivariate solution.
A similar phenomenon attributable to the separation rank
is observed in the HT simulations. In particular,
the time instant at which the HT tensor series
requires a separation rank higher than $8$ can be clearly
identified, i.e., at $t=0.2$ in two dimensions, at $t=0.1$
in three dimensions, and at $t=0.05$ in six dimensions.
In Figure \ref{fig:CP_MODES} we plot
the time evolution of the first three CP
modes $G_6^1$, $G_6^2$ and $G_6^3$
(see equation \eqref{CP-numerical})
representing the dynamics of the solution functional in
the variable $a_6=(\theta,\varphi_6)$.
\begin{figure}
\centerline{
\rotatebox{90}{\hspace{1.8cm}$t=0.01$ }\hspace{0.7cm}
\includegraphics[height=4.3cm]{G61t0-eps-converted-to.pdf}
\includegraphics[height=4.3cm]{G62t0-eps-converted-to.pdf}
\includegraphics[height=4.3cm]{G63t0-eps-converted-to.pdf}
}
\centerline{
\rotatebox{90}{\hspace{1.8cm}$t=0.5$ }\hspace{0.7cm}
\includegraphics[height=4.3cm]{G61t05-eps-converted-to.pdf}
\includegraphics[height=4.3cm]{G62t05-eps-converted-to.pdf}
\includegraphics[height=4.3cm]{G63t05-eps-converted-to.pdf}
}
\caption{\color{r}Time evolution of the first three CP
modes representing the dynamics along $a_6=(\theta,\varphi_6)$
in the CP expansion \eqref{CP-numerical} of the
solution functional. These modes
are obtained by solving numerically the FDE \eqref{FDE3}
in the function space $D_6$ (see Eq. \eqref{finite_dim_FS})
with the CP-ALS algorithm.}
\label{fig:CP_MODES}
\end{figure}
\vspace{0.7cm}
\noindent
{\em Remark:} If $\theta(x)$ is not in the function space
$D_m$ (see Eq. \eqref{finite_dim_FS}), but can be represented in $D_m$ with accuracy then
the solution to the multivariate PDE \eqref{PDE11}
provides an approximation of the solution
to the full FDE \eqref{FDE3} at such $\theta(x)$.
For instance, consider the following test function
\begin{equation}
\theta(x)= \sin(x).
\end{equation}
The Fourier coefficients of $\theta$ relative to the
orthonormal basis shown in Figure \ref{fig:basis_eigen}
are plotted in Figure \eqref{fig:spectrum_and_accuracy}.
Clearly $\theta(x)$ is not in $D_6$,
but it can be approximated well in $D_6$. This means
that the solution to the FDE at $\sin(x)$ can be approximated
by the solution of the six-dimensional PDE arising when we
evaluate the FDE \eqref{FDE3} $D_6$.
\begin{figure}
\centerline{(a) \hspace{7cm} (b)}
\centerline{
\includegraphics[height=6cm]{sin_spectrum-eps-converted-to.pdf}
\includegraphics[height=6cm]{sin_convergence-eps-converted-to.pdf}
}
\caption{\color{r}(a) Fourier coefficients of $\sin(x)$ relative to the
orthonormal polynomial basis shown in Figure \ref{fig:basis_eigen}. (b)
Comparison between the solution to the FDE
in $D_3$ and $D_{10}$ and the exact solution
\eqref{THESOLUTION} evaluated at $\theta(x)=\sin(x)$.}
\label{fig:spectrum_and_accuracy}
\end{figure}
\paragraph{Computing Functional Derivatives}
In Figure \ref{fig:functional derivative} we compare the
exact first-order functional derivative \eqref{FunctionalDerivative}
at $t=0.4$ with the numerical approximation we obtained
with the HT algorithm in $m=3$ and $m=6$ dimensions.
The separation rank of the tensor series expansions is set
to $r=12$. As expected, the numerical approximation
converges to the exact solution as we increase
the number of dimensions.
\begin{figure}[t]
\centerline{\hspace{0.7cm}$\theta(x)=\sin(x)$ \hspace{5.2cm}
$\theta(x)=xe^{-x^2/4}$}
\noindent
\centerline{
\includegraphics[height=5.5cm]{ffd_sin-eps-converted-to.pdf}
\includegraphics[height=5.5cm]{ffd_exp-eps-converted-to.pdf}
}
\caption{\color{r}First-order functional derivative of the solution
functional \eqref{THESOLUTION}
at $t=0.4$, evaluated at the test functions $\theta(x)=\sin(x)$ and
$\theta(x)=xe^{-x^2/4}$.
We plot the exact result \eqref{FunctionalDerivative} and
the numerical approximation \eqref{FDeRT} we obtained
in the function spaces $D_3$ and $D_6$
(see Eq. \eqref{finite_dim_FS}). The numerical approximation is
computed by solving the mutivariate PDE \eqref{PDE-advR} in
three and six dimensions.}
\label{fig:functional derivative}
\end{figure}
\subsection{The Navier-Stokes-Hopf Functional Equation}
\label{sec:Navier-Stokes-Hopf}
In this Section we discuss approximation of the
Navier-Stokes-Hopf functional equation
\begin{equation}
\frac{\partial \Phi([\bm \theta],t)}{\partial t}=\sum_{k=1}^3\int_V
\theta_k(\bm x)\left[i\sum_{j=1}^3\frac{\partial }{\partial x_j}
\left(\frac{\delta^2 \Phi([\bm \theta],t)}{\delta \theta_k(\bm x)
\delta\theta_j(\bm x)}\right)+\nu\nabla^2 \left(\frac{\delta
\Phi([\bm \theta],t)}{\delta \theta_k(\bm x)}\right) \right]d \bm
x.
\label{h1}
\end{equation}
In this formulation, $V$ is a periodic box
and $\bm \theta(\bm x)$ is chosen in a
divergence-free space of test functions.
The main advantage of using such
divergence-free space is that the pressure term
drops out, just as in the classical Navier-Stokes
equations\footnote{\color{r}
We recall that the pressure functional in
the Navier-Stokes-Hopf equation is defined as
(see \cite{Monin2}, p. 749)
\begin{equation}
\Pi([\bm \theta],\bm x,t) = \left<p(\bm x,t;\omega)\exp
\left[i\int_V \bm \theta(\bm x)\cdot \bm u(\bm x,t;\omega) d\bm
x\right]\right>.
\end{equation}
Therefore,
\begin{align}
\int_V \bm \theta\cdot\left<\nabla p \exp
\left[i\int_V \bm \theta\cdot \bm u d\bm x\right]\right>d\bm x=&
\left<\int_V \bm \theta\cdot \nabla p d\bm x \exp
\left[i\int_V \bm \theta\cdot \bm u d\bm x\right]\right>.
\end{align}
If $\bm \theta$ is divergence-free, i.e.,
$\nabla\cdot \bm \theta = 0$ in $V$, and satisfies the
tangency condition $\bm \theta\cdot \hat{\bm n}=0$
($\hat{\bm n}$ is the outward unit vector orthogonal
to the boundary of $V$) we have
\begin{align}
\int_V \bm \theta(\bm x)\cdot \nabla p(\bm x,t;\omega) d\bm x =
\int_{\partial V} p(\bm x,t;\omega) \bm \theta(\bm x)\cdot
\hat{\bm n} d\bm x = 0.
\label{integg}
\end{align}
The integral \eqref{integg} is zero also
if $\bm \theta$ is divergence-free and the
domain $V$ is a periodic box.
In fact, in this case $\bm \theta(\bm x)$ is periodic,
$p$ is periodic, and therefore the integral
along the boundary $\partial V$ vanishes.}.
To develop a discretization of the Hopf equation \eqref{h1}
it is convenient to first address the question of
how to represent divergence-free spaces of
periodic functions in 2D and 3D.
\subsubsection{Symmetries of the Solution Functional}
The divergence-free constraint in the velocity field
induces a certain number of symmetries in the
Hopf functional.
Let first assume that the flow develops in a bounded
region $V$, i.e., $\bm u\cdot \widehat {\bm n}=0$,
where $\widehat {\bm n}$ is the outward unit vector
normal to the boundary of $ V$.
Such region could be any volume co-moving with the fluid.
In this assumption,
\begin{equation}
\Phi([\bm \theta+\nabla \varphi]) = \Phi([\bm \theta])\label{div}
\end{equation}
for any $\bm \theta(\bm x)$ and any $\varphi (\bm x)$. In fact,
\begin{equation}
\int_V \bm u \cdot \nabla \varphi d\bm x= \int_{\partial V}\varphi \bm u\cdot \hat {\bm n} d\bm x=0.
\label{29}
\end{equation}
If the boundary conditions assigned
to $\bm u$ on $\partial V$ (boundary of $V$) are different from
$\bm u\cdot \hat{\bm n}=0$, then \eqref{div} is still
valid, but not for any $
\varphi(\bm x)$. For example, if $\bm u$ is not
orthogonal to $\hat{\bm n}$
along some part of the boundary $\partial V$, then it is sufficient to set $ \varphi=0$ along such boundary and use the fact that $\nabla \cdot \bm u=0$ to conclude that \eqref{div} is still valid.
Now let us consider the Helmholtz decomposition
\begin{equation}
\bm \theta =\bm \eta + \nabla \varphi,
\label{dec}
\end{equation}
where $\nabla \cdot \bm \eta=0$ in the domain $V$, and $\varphi = 0$ at the boundary $\partial V$. By taking the divergence of \eqref{dec} we conclude that $\varphi$ satisfies the Dirichlet boundary value problem
\begin{equation}
\nabla^2 \varphi = \nabla\cdot \bm \theta\quad \textrm{(in $V$)},\qquad
\varphi=0 \quad \textrm{(in $\partial V$)},
\label{dvp}
\end{equation}
which has a solution. With the decomposition \eqref{dec} available, we have
\begin{equation}
\Phi([\bm \eta+\nabla \varphi ]) = \Phi([\bm \eta]).
\label{symm}
\end{equation}
In fact,
\begin{equation}
\int_V (\bm \eta+\nabla \varphi) \cdot \bm u d\bm x=
\int_V \bm \eta\cdot \bm u d\bm x+
\int_V \nabla \varphi \cdot \bm u d\bm x.
\end{equation}
However, by applying the Gauss theorem
(recall that $\nabla \cdot \bm u=0$)
\begin{equation}
\int_V \nabla \varphi \cdot \bm u dV=\int_{\partial V} \varphi
\bm u\cdot \widehat {\bm n} d\bm x =0
\label{w1}
\end{equation}
Therefore \eqref{symm} holds. The boundary conditions for $\bm \eta=\bm
\theta-\nabla \varphi$ depend on the gradient of the solution
to \eqref{dvp} at the boundary $\partial V$.
Next, consider the unique Helmholtz-Hodge
decomposition (see \cite{Bhatia} or \cite{ChorinMarsden}, p. 36) of the
field $\bm \theta$ in the form \eqref{dec},
where $\nabla \cdot \bm \eta=0$ and $\bm \eta$ is tangent to $\partial V$,
i.e., $\bm \eta\cdot \hat{\bm n}=0$ ($\hat{\bm n}$ outward unit vector normal
to $\partial V$).
If there is no flow across the boundary $\partial V$ ($\bm u\cdot
\hat{\bm n}=0$), or if the boundary conditions of $\varphi$ are chosen such that
the integral at the right hand side of \eqref{w1} is zero then we have the
symmetry \eqref{symm}. For example, if $\varphi$ is constant along the
boundary then the divergence-free requirement on $\bm u$ implies that the
integral at the right hand side of \eqref{w1} is zero. Also, if $\bm \theta$ and $
\bm u$ are periodic on a box then $\varphi$ is periodic and the boundary integral
in \eqref{w1} is zero.
In fact, the field $\varphi$ arising from the Helmoltz-Hodge
decomposition of $\bm \theta$ is the unique solution to the Neuman problem (see \cite{ChorinMarsden}, p. 36)
\begin{equation}
\nabla^2 \varphi = \nabla\cdot \bm \theta\quad \textrm{(in $V$)},\qquad
\frac{\partial \varphi}{\partial n}=\bm \theta \cdot \widehat {\bm n}
\quad \textrm{(in $\partial V$)}.
\end{equation}
Clearly, if $\bm \theta$ is periodic then $\varphi$ is periodic and therefore $\bm \eta$ is periodic too. In other words, If we evaluate $\Phi$ on a space of periodic functions then we have the invariance
\begin{equation}
\Phi([\bm \theta])=\Phi([\bm \eta]),
\end{equation}
where $\bm \eta$ is periodic and divergence free ($\bm \eta$ is the
divergence-free part arising from the unique Helmholtz-Hodge
decomposition of $\bm \theta$.
\subsubsection{Divergence-Free Function Spaces}
\label{sec:DivFreeSF}
There has been a significant research activity in identifying
bases for divergence-free spaces of functions.
For example, Deriaz and Perrier \cite{Deriaz1,Deriaz} have
developed an effective algorithm to construct divergence-free
and curl-free wavelets in 2D and 3D with various types of
boundary conditions.
Other divergence-free bases can be constructed in terms
of radial basis functions \cite{Fuselier}, trigonometric
polynomials \cite{Landriani}, or eigenvalue problems with
appropriate boundary conditions conditions \cite{Venturi_IJHMT,Ozisik}.
Hereafter we discuss how to construct a divergence-free
basis for two-dimensional periodic flows.
To this end, we consider the tensor product basis
\begin{equation}
\psi_n(x,y)=l_{j(n)}(x)l_{i(n)}(y),
\end{equation}
where $j(n)$ and $i(n)$ are suitable sequences of integer number while $l_k(x)$ are trigonometric polynomials. Any scalar-valued periodic function on the square (such as the streamfunction) can be represented as
\begin{equation}
\Psi(x,y)=\sum_{k=1}^M \alpha_k \psi_j(x,y).
\end{equation}
Next define the divergence-free basis
\begin{equation}
\bm \Gamma_k (x,y)=\left(\frac{\partial \psi_k(x,y) }{\partial y},\,
-\frac{\partial \psi_k(x,y) }{\partial x} \right).
\label{basis}
\end{equation}
It is clear that each basis element $\bm \Gamma_k$ is
divergence-free by construction, i.e., $\nabla \cdot \bm \Gamma_k=0$. However, the basis \eqref{basis} is not orthogonal nor normalized relative to the $L_2([-b,b]^2)$ inner product
\begin{align}
\left(\bm \Gamma_i,\bm \Gamma_j\right)=&\int_{-b}^b\int_{-b}^b
\bm \Gamma_i\cdot \bm \Gamma_j dxdy,\nonumber\\
=&\int_{-b}^b\int_{-b}^b
\left(
\frac{\partial \psi_i}{\partial y}\frac{\partial \psi_j}{\partial y}+
\frac{\partial \psi_i}{\partial x}\frac{\partial \psi_j}{\partial x}
\right) dxdy.\nonumber
\end{align}
To generate a divergence-free and orthonormal basis one
could use the Gram-Schmidt orthogonalization.
As we shall see hereafter, such procedure preserves
the divergence-free character of each basis element.
We first normalize $\bm \Gamma_1$ as
\begin{equation}
\widehat{\bm \Gamma}_1(x,y) = \frac{\bm \Gamma_1(x,y)}{\left\|\bm \Gamma_1(x,y)\right\|}, \qquad \left\|\bm \Gamma_1\right\|=
\sqrt{(\bm \Gamma_1,\bm \Gamma_1)}.
\end{equation}
Clearly, $\nabla \cdot \widehat {\bm \Gamma}_1=0$ since the $L_2$ norm of $\bm \Gamma_1$ is just a real number.
Next, define
\begin{equation}
\bm \Theta_2= \bm \Gamma_2- (\bm \Gamma_2,\widehat{\bm
\Gamma}_1)\widehat{\bm \Gamma}_1,\qquad
\widehat{\bm \Gamma}_2(x,y)=\frac{\bm \Theta_2(x,y)}{\left\|\bm
\Theta_2(x,y)\right\|}.
\end{equation}
As before, $\nabla \cdot \widehat{\bm \Gamma}_2=0$. Moreover
$\widehat{\bm \Gamma}_2$ is orthogonal to $\widehat{\bm \Gamma}_1$,
i.e., $(\widehat{\bm \Gamma}_1,\widehat{\bm \Gamma}_2)=0$.
The algorithm proceeds with the computation of
\begin{equation}
\bm \Theta_3= \bm \Gamma_3- (\bm \Gamma_3,\widehat{\bm
\Gamma}_2)\widehat{\bm \Gamma}_2-(\bm \Gamma_3,\widehat{\bm
\Gamma}_1)\widehat{\bm \Gamma}_1 ,\qquad \widehat{\bm \Gamma}_3(x,y)=\frac{\bm \Theta_3(x,y)}{\left\|\bm \Theta_3(x,y)\right\|}.
\end{equation}
Thanks to the fact that $\bm \Gamma_3$, $\widehat{\bm \Gamma}_2$
and $\widehat{\bm \Gamma}_1$ are divergence free, we have
that $\nabla \cdot \widehat{\bm \Gamma}_3=0$.
Moreover $\widehat{\bm \Gamma}_3$ is orthogonal to both
$\widehat{\bm \Gamma}_2$ and $\widehat{\bm \Gamma}_1$.
In other words, $\{\widehat{\bm \Gamma}_1,\widehat{\bm \Gamma}_2,
\widehat{\bm \Gamma}_3\}$ is a divergence-free
orthonormal system. Proceeding in a similar way, we can
construct the divergence-free orthonormal basis we
were looking for, and define the following finite-dimensional
divergence-free space of functions
\begin{equation}
D_M=\textrm{span}\{\bm \Gamma_1, \cdots, \bm \Gamma_M\}.
\label{DM}
\end{equation}
An element of $D_M$ is in the form
\begin{equation}
\bm \theta(x,y)=\sum_{k=1}^Ma_k\hat{\bm \Gamma}_k(x,y).
\label{divfree}
\end{equation}
\subsubsection{Analytical Solution to the Characteristic Function Equation}
A substitution of \eqref{divfree} into the
Navier-Stokes-Hopf equation \eqref{h1} yields the
multivariate (complex-valued) PDE\footnote{\color{r}
We emphasize that equation \eqref{NSCHF} has
exactly the same structure as the characteristic function
equation we obtain for the one-dimensional
Burgers equation. To show this, it is
sufficient to discretize the Burgers-Hopf equation
\begin{equation}
\frac{\partial \Phi([\theta],t)}{\partial t}=\int_{-b}^b\theta(x)\left[i\frac{\partial }{\partial x}\frac{\delta^2 \Phi([\theta],t)}{\delta \theta(x)^2}+\nu\frac{\partial^2 }{\partial x^2} \frac{\delta \Phi([\theta],t)}{\delta \theta(x)} \right]d x,
\label{h2}
\end{equation}
in the finite-dimensional space $D_m$
of periodic functions in $[-b,b]$. To
this end, consider the series expansion
\begin{equation}
\theta(x)=\sum_{k=1}^m a_k\varphi_k(x).
\label{sexP}
\end{equation}
where $\varphi_k(x)$ are orthonormal trigonometric
polynomials. Substituting \eqref{sexP} into
\eqref{h2} and evaluating all functional
derivatives in $D_m$ yields an equation in the form \eqref{NSCHF}.}
\begin{equation}
\frac{\partial \phi}{\partial t}=
i \sum_{p,j,k=1}^M A_{pjk}a_p
\frac{\partial^2\phi}{\partial a_k\partial a_j}+
\nu \sum_{k,p=1}^M a_pB_{pk}\frac{\partial \phi}{\partial a_k},
\label{NSCHF}
\end{equation}
where
\begin{equation}
B_{pk}=\int_V \widehat{\bm \Gamma}_p\cdot
\nabla^2\widehat{\bm \Gamma}_k d \bm x, \qquad
A_{pjk}=\int_V \widehat{\bm \Gamma}_p\cdot
\left[\left(\widehat{\bm \Gamma}_k\cdot \nabla\right)\widehat{\bm \Gamma}_j \right]d \bm x.
\label{coefficients}
\end{equation}
By using integration by parts we can simplify $B_{pk}$ to
\begin{align}
B_{pk}&=\int _{V}\left[\widehat \Gamma^{(x)}_{p} \left(\frac{\partial ^2 \widehat \Gamma^{(x)}_k}{\partial x^2}+\frac{\partial ^2 \widehat \Gamma^{(x)}_k}{\partial y^2}\right)+
\widehat \Gamma^{(y)}_{p} \left(\frac{\partial ^2 \widehat \Gamma^{(y)}_k}{\partial x^2}+\frac{\partial ^2 \widehat \Gamma^{(y)}_k}{\partial y^2}\right)\right]
d\bm x, \nonumber \\
&=-\int _{V}\left[\frac{\partial\widehat \Gamma^{(x)}_{p}}{\partial x} \frac{\partial \widehat \Gamma^{(x)}_k}{\partial x}+\frac{\partial \widehat \Gamma^{(x)}_{p}}{\partial y} \frac{\partial \widehat \Gamma^{(x)}_k}{\partial y}
+ \frac{\partial \widehat \Gamma^{(y)}_{p}}{\partial x} \frac{\partial \widehat \Gamma^{(y)}_k}{\partial x}+\frac{\partial \widehat \Gamma^{(y)}_{p}}{\partial y} \frac{\partial \widehat \Gamma^{(y)}_k}{\partial y}\right]
d\bm x.
\end{align}
Similarly, the coefficients $A_{pjk}$ can be written as
\begin{align}
A_{pkj}&=\int _{V}\left[
\Gamma^{(x)}_{p} \left(\Gamma^{(x)}_k
\frac{\partial \Gamma_j^{(x)}}{\partial x}+\Gamma^{(y)}_k
\frac{\partial \Gamma_j^{(x)}}{\partial y}\right)+
\Gamma^{(y)}_{p} \left(\Gamma^{(x)}_k
\frac{\partial \Gamma_j^{(y)}}{\partial x}+\Gamma^{(y)}_k
\frac{\partial \Gamma_j^{(y)}}{\partial y}
\right)
\right]
d\bm x.
\end{align}
By taking the inverse Fourier transform of the characteristic
function equation \eqref{NSCHF} we obtain
\begin{equation}
\frac{\partial p(\bm u,t)}{\partial t}+\sum_{p=1}^M
\frac{\partial }{\partial u_p}\left[\left(\nu\sum_{k=1}^M u_k B_{pk} -
\sum_{k,j=1}^M u_ku_jA_{pkj} \right) p(\bm u,t)\right]=0.
\label{jointPDFCh}
\end{equation}
This is the multivariate first-order PDE that governs the
evolution of the finite-dimensional approximation of the probability density functional\footnote{Equation \eqref{jointPDFCh}
can be obtained by evaluating the Navier-Stokes
probability density functional equation \cite{Dopazo,Obrien}
in the finite-dimensional test function space \eqref{DM}.}.
The formal solution to \eqref{jointPDFCh} is
\begin{equation}
p(\bm u,t)=p_0\left(\bm U(t,\bm u)\right)\exp\left(-\int_0^t \nabla\cdot
\bm G(\bm u(\tau,\bm U)) d\tau \right),
\label{solutionPDf}
\end{equation}
where
\begin{equation}
G_p(\bm u) = \nu\sum_{k=1}^M u_k B_{pk} - \sum_{k,j=1}^M u_ku_jA_{pkj}, \qquad p=1,...,M.
\end{equation}
The flow map $\bm u(t,\bm U)$ and its inverse $\bm U(t,\bm u)$
are defined by the solution to the ODE system
\begin{equation}
\frac{d\bm u}{dt }=\bm G(\bm u), \qquad \bm u(0)=\bm U.
\label{ODE1}
\end{equation}
By taking the Fourier transform of \eqref{solutionPDf}, we obtain
the following analytical solution to the discretized
Navier-Stokes-Hopf equation \eqref{NSCHF}
\begin{equation}
\phi(\bm a,t)=\int_{-\infty}^\infty\cdots \int_{-\infty}^\infty
e^{i\bm u\cdot \bm a}p_0\left(\bm U(t,\bm u)\right)\exp\left(-\int_0^t \nabla\cdot
\bm G(\bm u(\tau,\bm U)) d\tau \right) d\bm u.
\label{solutionChF}
\end{equation}
Numerical methods to compute the flow map $\bm u(t, \bm U)$ and
its inverse $\bm U(t,\bm u)$ in equations \eqref{solutionPDf} and
\eqref{solutionChF} often rely on ray tracing, i.e., numerical
solutions to the \eqref{ODE1} on a Cartesian mesh
of initial conditions. Such methods are
not efficient in high-dimensions \cite{You2017}, and may be
improved significantly by representing flow maps
using, e.g., tensor formats.
\paragraph{Functional Derivatives}
The the first-order functional derivatives of the Hopf characterisitic functional (evaluated in the space of divergence-free functions) can be approximated as
\begin{align}
\frac{\delta \Phi([\bm \theta],t)}{\delta \theta^{(x)}(\bm x)}=\sum_{k=1}^M
\frac{\partial \phi}{\partial a_k}\widehat\Gamma_k^{(x)}(\bm x),\qquad
\frac{\delta \Phi([\bm \theta],t)}{\delta \theta^{(y)}(\bm x)}=&\sum_{k=1}^M\frac{\partial \phi}{\partial a_k}\widehat\Gamma_k^{(y)}(\bm x).
\end{align}
Similarly, the second-order derivatives can be written as
\begin{align}
\frac{\delta^2 \Phi([\bm \theta],t)}{\delta \theta^{(x)}(\bm x) \delta \theta^{(x)}(\bm y)}=&\sum_{k,j=1}^M
\frac{\partial^2 \phi}{\partial a_k a_j}\widehat\Gamma_k^{(x)}(\bm x)\widehat\Gamma_j^{(x)}(\bm y),\\
\frac{\delta^2 \Phi([\bm \theta],t)}{\delta \theta^{(y)}(\bm x) \delta \theta^{(y)}(\bm y)}=&\sum_{k,j=1}^M
\frac{\partial^2 \phi}{\partial a_k a_j}\widehat\Gamma_k^{(y)}(\bm x)\widehat\Gamma_j^{(y)}(\bm y),\\
\frac{\delta^2 \Phi([\bm \theta],t)}{\delta \theta^{(x)}(\bm x) \delta \theta^{(y)}(\bm y)}=&\sum_{k,j=1}^M
\frac{\partial^2 \phi}{\partial a_k a_j}\widehat\Gamma_k^{(x)}(\bm x)\widehat\Gamma_j^{(y)}(\bm y).
\end{align}
Evaluating these derivatives at $\bm \theta =0$ is equivalent to evaluate the derivatives of the characteristic function at $(a_1,...,a_M)=(0,...,0)$. This yields, the following representation of the mean and cross correlation of the velocity field
\begin{equation}
\left<\bm u(\bm x,t)\right>=\sum_{k=1}^M \left.
\frac{\partial \phi(\bm a,t)}{\partial a_k}\right|_{(0,...,0)} \widehat{\bm \Gamma}_k(\bm x),
\end{equation}
\begin{equation}
\left<u^{(x)}(\bm x,t)u^{(y)}(\bm y,t)\right>=\sum_{k,j=1}^M \left.
\frac{\partial^2 \phi(\bm a,t)}{\partial a_k \partial a_j}\right|_{(0,...,0)} \widehat\Gamma^{(x)}_k(\bm x)\widehat\Gamma^{(y)}_j(\bm y).
\end{equation}
\paragraph{Computational Complexity}
The number of dimensions $M$ appearing
in the characteristic function equation \eqref{NSCHF} and
the joint PDF equation \eqref{jointPDFCh} coincides with the
number of number of degrees of freedom
we employ in the discretization of the velocity field.
For instance, if we consider the classical
two-dimensional Kolmogorov flow \cite{Lucas}
represented on a $128 \times 128$ Fourier basis,
then $M=16384$ ($128^2$). Computing the solution
to such high-dimensional linear PDEs (i.e., \eqref{NSCHF}
or \eqref{jointPDFCh}) obviously requires
parallel algorithms and a highly-efficient tensor methods.
If we employ operator splitting in time, then we can easily take
care of the linear part in \eqref{jointPDFCh} -- i.e., the
one depending linearly on $u_k$ -- by integrating out
the corresponding dynamics with the method
of characteristics.
To this end, let us first
consider an orthonormal divergence-free basis that
diagonalizes the matrix $B_{pk}$ in
\eqref{coefficients}. Such basis exists and it can be
computed by standard linear algebra
techniques \cite{ZhengDong2011}
(simultaneous diagonalization of two quadratic forms).
Relative to the new basis, the joint PDF equation \eqref{jointPDFCh} can be written as
\begin{equation}
\frac{\partial p(\bm u,t)}{\partial t} = -\sum_{p=1}^M
\frac{\partial }{\partial u_p}\left[\left(\nu u_pB_{pp}- \sum_{j,k=1}^M u_ku_jA_{pkj}\right)p(\bm u,t)\right],
\end{equation}
where, with some abuse of notation, we denoted
by $B_{pp}$ and $A_{pkj}$ the entries of $B_{pk}$
and $A_{pkj}$ in \eqref{coefficients}
relative to the new basis. The last equation can be
written in the operator form
\begin{equation}
\frac{\partial p(\bm u,t)}{\partial t} = L(\bm u) p(\bm u,t)+ Q(\bm u)p(\bm u,t),
\label{opjPDF}
\end{equation}
where
\begin{equation}
L(\bm u) = - \sum_{p=1}^M
\nu u_p B_{pp}\frac{\partial }{\partial u_p} - B_{pp}\nu + \sum_{j,p=1}^M u_j
\left(A_{ppj}+A_{pjp}\right), \qquad
Q(\bm u)=
\sum_{k,j,p=1}^M u_ku_jA_{pkj}\frac{\partial }{\partial u_p}.
\end{equation}
Note that $L(\bm u)$ depends linearly on $\bm u$
while $Q(\bm u)$ is quadratic in $\bm u$. Both $L$ and $Q$ are separable linear operators.
The formal solution to \eqref{opjPDF} is\footnote{We recall
that $\exp[t(L+Q)]$ is the Frobenious-Perron
operator associated with the quadratic
dynamical system \eqref{ODE1} (see \cite{Dominy2017}).}
\begin{equation}
p(\bm u,t)=e^{t\left(L(\bm u)+Q(\bm u)\right)}p(\bm u,0).
\label{prop}
\end{equation}
By using operator splitting in time, e.g., the classical
second-order Strang splitting, we can write
\begin{equation}
p(\bm u,t_n)=e^{\Delta t L(\bm u)/2 }e^{\Delta t Q(\bm u) }e^{\Delta t L(\bm u)/2 } p(\bm u,t_{n-1}).
\end{equation}
Clearly, the action of the semigroup $\exp [t L(\bm u)]$ can be
computed exactly since the flow map corresponding
characteristic system associated with the first order linear
PDE $\dot{g} = L g$ is trivial. Specifically, for
any $g_0$ and any $t$ we have
\begin{equation}
e^{t L(\bm u)}g_0(\bm u)=
g_0\left(e^{-t\nu B_{11}}u_1,...,e^{-t\nu B_{MM}}u_M\right)
\exp\left[- t \nu \sum_{j=1}^M B_{jj} + \sum_{j=1}^M \int_{0}^t
e^{t\nu B_{jj}}u_{j0}\sum_{p=1}^M \left(A_{ppj}+A_{pjp}\right)\right].
\end{equation}
The action of the semigroup $\exp[t Q(\bm u)]$ is much more
complicated to compute. Indeed, the flow map of quadratic
dynamical systems is, in general, not known explicitly. For small
$\Delta t$ one may introduce the approximation
$\exp[\Delta Q(\bm u)]\simeq I + \Delta t Q(\bm u)$
(or any higher-order one arising, e.g., from multi-step methods)
and leverage on the separability of the operator $Q(\bm u)$
(rank $M^3$).
}
\subsection*{Acknowledgments}
This research was supported by the Air Force Office
of Scientific Research grant FA9550-16-1-0092.
|
\section{Introduction}
\noindent {\it \textbf{Introduction}}. The Standard Model (SM) of particle physics has long been known to lack an adequate candidate for dark matter. An oft-repeated virtue of its minimal supersymmetric extension (the MSSM) is that a conserved matter parity, or R-parity, will imply that the lightest supersymmetric particle is stable, and therefore a natural dark matter candidate~\cite{Ellis:1983ew}. Unfortunately, as we will argue here, when the MSSM is embedded in a ultraviolet (UV) complete theory, such as string or $M$-theory, the lightest supersymmetric particle will not reside in a visible sector.
Whatever the particular details of any particular string compactification may be, there are certain results that appear to be generic. Of particular importance to this paper are the existence of hidden sectors. We define a hidden sector as containing states in the low-energy effective theory that are uncharged at tree-level under the SM gauge symmetries, but can be charged under their own symmetry group $G_{H}$.
Compactified string/$M$-theory solutions will generically have hidden sectors, containing, at a minimum, the gauge fields and gauginos associated the various group factors contained in $G_H$, when the UV solution is supersymmetric
\cite{Dienes:1996zr}-\cite{Corti:2012kd}.
Such sectors have already been used for model-building purposes, in particular for the breaking of supersymmetry (SUSY), which takes place in a hidden sector and is then mediated to the visible sector.
Hidden sectors will of course interact with our visible sector via gravitational interactions, but can also have other so-called ``portals" to the visible sector~\cite{Batell:2009di}. We argue that the existence of hidden sectors and portals leads to the conclusion that the lightest supersymmetric particle in the visible sector (LVSP) is likely to be unstable, since there will generically be a lighter particle in one of the many hidden sectors (an LHSP), into which the LVSP will decay via the portal. For this decay to occur, a simple list of conditions is required, and we find that they are all quite typically available, leading to the conclusion that the LVSP decays. The conditions are the following:
\begin{enumerate}
\item There exists a hidden sector.
\item There exists a portal connecting the visible sector to the hidden sector.
\item That hidden sector spectrum includes a particle lighter than the LVSP.
\end{enumerate}
We will concentrate on the kinetic mixing portal~\cite{Holdom:1985ag}, as we find this to be the most generic portal arising from string theory~\cite{Dienes:1996zr}. The existence of other portals would only serve to strengthen our argument.\\
\noindent {\it \textbf{Hidden Sectors in String Theory}}. A typical feature of the hidden sectors in string/$M$ theory are their multitude and their richness. The presence of hidden sectors is not optional, but often {\em required} to ensure the mathematical consistency of the theory. While systematic studies remain rare (see the discussion in~\cite{Kane:2015qea}), the genericity of large hidden sectors, with many small-rank groups, has been demonstrated in several contexts. These include the heterotic string in the free-fermionic approach~\cite{Dienes:2006ut} and in the orbifold limit~\cite{Giedt:2000bi}, weakly-coupled Type~II string theory~\cite{Cvetic:2004ui}, and Gepner models~\cite{Dijkstra:2004cc}. $F$-theory models are known to produce similarly rich hidden sectors~\cite{Taylor:2015ppa}. Finally, let us consider $M$-theory compactified on a manifold with G$_2$ holonomy. Whilst, in this case, we are technically far away from being able to perform systematic surveys of gauge groups, the general picture is expected to be somewhat similar to the $F$-theory results of~\cite{Taylor:2015ppa}. This can also be argued from duality with the heterotic and Type-II string theory. Given that the number of hidden sectors is bounded by the third Betti number, which is typically $\mathcal{O}(100)$ \cite{Joyce,Corti:2012kd}, it is expected that having many hidden sectors will also prove generic in $M$-theory.
It is important to note that the size of the hidden sector gauge group is often much larger than that of the observable sector. We may use the rank of the group $G_H$ (typically a product of non-Abelian and Abelian factors) as a proxy for the richness of the hidden sector. Traditional string model-building has centred upon weakly-coupled heterotic or Type~II solutions, which typically give a rank for $G_H$ that is larger than $G_{SM}$, but roughly comparable in size.
In recent years, however, the study of strongly-coupled string theory has been put on a much firmer footing, particularly in the context of $F$-theory. Here the expectation is for the rank of $G_H$ to be much larger than $G_{SM}$, perhaps by orders of magnitude\footnote{Similarly large ranks for $G_H$ have been observed for some time in rational conformal field theory constructions~\cite{Dijkstra:2004cc}.} (see, for example, the specific case studied in~\cite{Taylor:2015xtz}). What is more, the notion of ``generic'' has become increasingly precise in these contexts~\cite{Halverson:2015jua}.
In the next section we will consider the phenomenon of kinetic mixing, which requires the presence of Abelian $U(1)$ factors in the hidden sector.
Given the existence of non-Abelian gauge groups in hidden sectors, it is clear that if they are broken, there can be resulting $U(1)$ factors in the hidden sector. The mechanisms for breaking the hidden sector gauge group can be either via Wilson lines, or via radiative breaking at lower energies. The former case is inherently string-theoretic, in that it requires the presence of non-trivial geometry in the compact space. The latter mechanism is familiar from four-dimensional field theory. Furthermore, there can be $U(1)$'s in the four-dimensional effective field theory that do not stem from non-Abelian groups, and have a string-theoretic origin. A well-studied example is the dimensional reduction of Ramond-Ramond (RR) forms on suitable cycles in Type~II theory~\cite{Jockers:2004yj}.
To be phenomenologically relevant, it is necessary that any such $U(1)$ be non-anomalous, for otherwise the gauge boson would receive a mass of order the string scale through the Green-Schwarz mechanism. In open string theories one typically finds that many of the $U(1)$ factors are anomalous. What is more, many $U(1)$s which are non-anomalous may nevertheless acquire a string-scale mass to satisfy higher-dimensional anomaly cancellation conditions~\cite{Ibanez:1998qp}. Yet the effective mass matrix for the collection of $U(1)$s need not have full rank, and indeed generally does not~\cite{Blumenhagen:2005mu}. The same has been observed in heterotic constructions which generalise the structure group of the gauge bundle~\cite{Anderson:2011ns}.
Furthermore, there are many circumstances in which one expects massless $U(1)$s to emerge. The most obvious are cases in which the $U(1)$ arises from the breaking of a non-Abelian group via Wilson line breaking, or through parallel splitting of stacks of D-branes. Abelian factors arising from the zero modes of closed string RR-forms are guaranteed to be massless on Calabi-Yau surfaces, and can obtain masses only in non-K\"ahler backgrounds~\cite{Grimm:2008ed}. Abelian factors supported by $\overline{D3}$-branes cannot participate in the St\"uckelberg mechanism as the necessary axionic fields are projected out by orientifolding~\cite{Abel:2003ue}. All of these arguments imply that one generically expects light $U(1)$s in the effective field theory.\\
\noindent {\it \textbf{The Kinetic Mixing Portal}}. The kinetic mixing portal was first considered in the context of four-dimensional field theory~\cite{Holdom:1985ag}, in which it arises from the existence, and subsequent integrating out of, heavy bi-fundamental fields, charged under both $U(1)$'s. Such states exist typically in open string theories. For instance, if both $U(1)$'s are supported by $D$-branes which are separated in the extra dimensions, as is the case for all supersymmetric Type I, Type IIA and Type IIB models, then there will be massive open strings which stretch between the two $D$-branes, giving rise to massive bi-fundamentals. There are generalisations of this statement in $M$-theory, $F$-theory and the heterotic string as well.
These bi-fundamentals will lead to a one-loop mixing of the two $U(1)$ symmetries $U(1)_a$ and $U(1)_b$. In the case of interest let $U(1)_a$ correspond to the visible sector $U(1)_Y$, and $U(1)_b$ correspond to a hidden sector $U(1)$. The Lagrangian of the $U(1)$ kinetic sector then reads
\begin{align}
\mathcal{L}_{gauge} = -\frac{1}{4}F^{\mu \nu}_a {F_{a}}_{ \mu \nu} -\frac{1}{4}F^{\mu \nu}_b {F_{b}}_{ \mu \nu} + \frac{\epsilon}{2}F^{\mu \nu}_a {F_{b}}_{ \mu \nu}
\end{align}
where $\epsilon$ parameterises the mixing of the two $U(1)$'s each with field strength tensor $F^{\mu\nu}_i$.
The expected size of $\epsilon$ can be estimated by calculating the two-point polarisation diagram
and is given by
\begin{align}
\epsilon \simeq \frac{g_a g_b }{12\pi^2}Q_a Q_b \log\left(1+\frac{\Delta m_{ab}^2}{M^2}\right)
\label{epsilon}
\end{align}
where $g_{a,b}~(Q_{a,b})$ are the couplings (charges of bi-fundamentals) of $U(1)_{a,b}$, and $\Delta m_{ab}$ is the mass splitting of the bi-fundamental fields charged under both groups and $M$ is the bi-fundamental mass scale.
Clearly if the $U(1)$'s sit in an unbroken non-Abelian gauge symmetry, $\epsilon = 0$. If the matter spectrum is charged under a non-Abelian gauge symmetry with a $U(1)$ factor then the mass degeneracy of the spectrum would naively cause $\epsilon = 0$ also. However, this degeneracy is not stable against radiative corrections, and mass splittings $\Delta m_{ab}$ are generated, thus rendering $\epsilon$ non-zero.
These bi-fundamentals may have masses $M \sim {R \over l_s^2} $, where $R$ is the separation of two stacks of $D_p$ branes connected by the open string. This suggests that the mass should be $M \sim \mathcal{O}(M_{GUT})$. Depending on the size of $\Delta m_{ab}$, $\epsilon$ can take on a wide range of possible values. In particular, if the hidden sector gauge group is broken at a scale $M_{G_H} \sim M$ via a Wilson line, $\epsilon$ can be of $\mathcal{O}( 10^{-3})$ for $\mathcal{O}(1)$ charges $Q_a$ and $Q_b$ and $g_a \sim g_b \sim g_Y$.
On the other hand if the hidden sector gauge group is broken radiatively through field theory dynamics at some much smaller scale $M_{G_H} \ll M$, $\epsilon$ can be as small as $\mathcal{O}( 10^{-26})$ (e.g. if $M_{G_H} \sim 1$ TeV). Crucially however, barring some non-generic external mechanism to prevent $\epsilon$ from being generated, it is always non-zero~\cite{Dienes:1996zr}. Since $\epsilon$ enters via a dimension-4 operator, it is not suppressed by high mass scales.
Explicit calculations of the kinetic mixing parameter~(\ref{epsilon}) in Type~II constructions support these arguments. Typical values for the mixing parameter are found to generally lie in the range $10^{-3} \leq \epsilon \leq 10^{-1}$, with values as low as $\epsilon \sim 10^{-6}$ accessible via tuning~\cite{Abel:2008ai}. Some additional volume suppression in the Type~IIB context can be obtained in various LARGE volume scenarios \cite{LVS}, in which the assumption that $g_b \sim g_Y$ is relaxed. In this case, a compact volume which generates an intermediate string scale $M_s \sim 10^{10} \, {\rm GeV}$ could produce an effective mixing parameter in the range $10^{-8} \leq \epsilon \leq 10^{-6}$~\cite{Goodsell:2009xc}. It is unclear whether such large volumes are generic, even within the context of flux compactifications of Type~IIB string theory, though as we will see below, these values still imply that the LVSP will not be an adequate dark matter candidate.
Finally, we should emphasize that non-vanishing kinetic mixing has also been demonstrated in heterotic contexts, including heterotic $M$-theory~\cite{Lukas:1999nh}, Calabi-Yau compactifications~\cite{Blumenhagen:2005ga} and in certain heterotic orbifold limits~\cite{Goodsell:2011wn}. The genericity of kinetic mixing in string theory, and its typical size ($\epsilon \sim 10^{-3}$) appear to bear out the intuition of Dienes et al. from twenty years ago~\cite{Dienes:1996zr}.\\
\noindent {\it \textbf{The Decay Mode}}. Given the existence of a portal, such as the kinetic mixing portal described above, the stability of the LVSP becomes a simple question of kinematics. Note that the usual argument for LVSP stability is based on discrete symmetries, but that requires that the LVSP mass is sufficiently small compared to other particle masses. The LVSP will not decay if it is lighter than all possible combinations of potential hidden sector decay products permitted by gauge invariance alone. One can ask for a sufficient condition for LVSP stability: {\it ‘’why should the LVSP be lighter than all hidden sector particles?’’} A key point is that this question has no obvious answer, and clearly becomes more and more difficult as one increases the number and complexity of the hidden sectors. If there is no good reason for the LVSP to be light compared to hidden sector particles then, presumably, the LVSP will decay. How does it decay?
With supersymmetry the hidden sector contains the $U(1)$ gauge boson and the associated gaugino. The sector may also contain matter charged under the hidden $U(1)$. If any of these states are lighter than the LVSP, then the LVSP can decay via the portal.
Let us assume that the LVSP is a neutralino, as is common within the MSSM. If a kinetic mixing portal exists to a hidden sector in which the LHSP is also a gaugino, then LVSP decay is expected whenever $\delta m = m_{LVSP}- m_{LHSP} >0$.
If $\delta m >m_Z$, the neutralino LVSP undergoes 2-body decay to a Z boson with lifetime
\begin{align}
\tau^{\chi_i \rightarrow Z\ \chi_j}_{\rm 2-body} \sim 10^{-17} ~ s \times \left( \frac{10^{-3}}{\epsilon} \right)^2 \left( \frac{0.01}{|N_{i3}N_{j3}^*|}\right)^{2} \ ,
\end{align}
where $N_{km}$ is a neutralino mixing matrix element. We have assumed a mostly Bino or Wino LVSP, and have taken $m_{LVSP}=1$ TeV and $m_{LHSP} =100$ GeV for illustrative purposes. Three-body decays can occur if $\delta m < m_Z$, and may dominate \cite{Arvanitaki:2009hb}. Then the characteristic lifetime is
\begin{align}
\tau^{\chi_i \rightarrow Z\ \chi_j}_{\rm 3-body} &\sim 10^{-9}~ \mathrm{s} \times \left( \frac{10^{-3}}{\epsilon} \right)^2 \left( \frac{0.01}{|N_{i3}N_{j3}^*|}\right)^{2} \ ,
\end{align}
where we have taken $m_{LVSP}=1$ TeV and $m_{LHSP} =950$ GeV for illustrative purposes. There are also both two- and three-body decays to a Higgs boson, with $\tau^{\chi_i \rightarrow h\ \chi_j} \sim \frac{H_{ij}^2}{|N_{i3}N_{j3}^*|^2} \tau^{\chi_i \rightarrow Z\ \chi_j}$, where $H_{ij}$ is the neutralino coupling to higgs bosons. Additionally, if the LVSP is mostly Higgsino, $\tau^{\tilde{H}\rightarrow Z\ \chi_j} \sim|N_{i3}N_{j3}^*| ~ \tau^{\tilde{B},\tilde{W}\rightarrow Z\ \chi_j}$ for both the two- and three-body decays.
There can also be decays into a chiral fermion LHSP which can be much lighter; we will describe the resultant parameter space in a follow-up paper. \\
\noindent {\it \textbf{Summary}}. In this paper we have put forward arguments that imply that the lightest supersymmetric particle in the visible sector is not, in fact, stable. At the very least it is metastable, though it is far more likely that it undergoes prompt decays into hidden sector states. The components of the argument are simple: (1) there is at least one hidden sector, (2) there is at least one portal connecting it to the visible sector, and (3) there exists matter in that sector which is lighter than the LVSP. We have illustrated these arguments with a kinetic mixing portal, since this appears to be the most generic outcome from string theory, but the presence of additional portals would only strengthen the argument.
Several aspects of our argument have appeared elsewhere in the literature in various forms. Here we have emphasised the generic nature of these components in string/$M$-theory solutions. Given the multitude of hidden sectors in string compactifications, it is quite likely that there exists at least one sector that satisfies these assumptions. Therefore we conclude that the LVSP will decay. It is the generality of this conclusion that compels us to argue for a paradigm shift in the thinking of phenomenologists when it comes to dark matter. In particular it raises the likelihood that dark matter resides in a hidden sector which might be very difficult to probe.
We conclude by noting that relegating dark matter to some hidden sector has phenomenological consequences. The resulting lifetime will affect collider signatures; we will return to study these in a follow-up paper. The kinetic mixing portal scenario illustrated may be cosmologically perilous, due to long-range forces and millicharged particles, to disruptions in Big Bang nucleosynthesis, to a relic overabundance that conflicts with the known age of the universe~\cite{Goodsell:2009xc,Ibarra:2008kn,Arvanitaki:2009hb}.
None of these challenges negate the conclusion that the LVSP is very likely unstable. It is non-generic to avoid sizeable kinetic mixing and light hidden sector states. Instead, it may turn out that the study of dark matter in string theories will illuminate that corner of the string/M-theory landscape in which our world resides.
\noindent {\it \textbf{Acknowledgements.}} BSA would especially like to thank R. Valandro for discussions on Type IIB examples. We also thank J. Ellis, J. Halverson, A. Pierce, G. Shiu, J. Wells, Y. Zhao and B. Zheng for useful discussions. The work of BSA was supported by the STFC Grant ST/L000326/1. The work of SARE and GLK is supported in part by the U.S. Department of Energy, Office of Science, under grant DE-SC0007859. The work of BN is supported in part by the National Science Foundation, under grant PHY-1314774. The work of MJP is supported in part by STFC.
|
\section{Introduction}
\label{intro}
The mathematical modelling and numerical approximations of free surface water waves propagation and transformations in near-shore areas has received a lot of interest for the last decades, motivated by the perspective of acquiring a better understanding of important physical processes associated with the nonlinear and non-hydrostatic propagation over uneven bottoms. Great improvements have been obtained in the derivation and mathematical understanding of particular asymptotic models able to describe the behavior of the solution in some physical specific regimes. A recent review of the different models that may be derived can be found in \cite{Lannes:2009p4892}. In this work, we focus on the \textit{shallow water} and \textit{large amplitude} regime: the water depth $h_0$ is assumed to be small compared to the typical wave length $\lambda$:
$$
\mbox{(\textit{shallow water regime}) }\quad \mu:=\frac{h_0^2}{\lambda^2}\ll 1.
$$
while there is no assumption on the size of the wave's amplitude $a$:
\begin{equation}
\mbox{(\textit{large amplitude regime}) }\quad \varepsilon:=\frac{a}{h_0}=O(1).
\end{equation}
Under this regime, the classical Nonlinear Shallow Water (NSW) equations can be derived from the full water waves equations by neglecting all the terms of order $O(\mu)$, see for instance \cite{lannes:book}. This model is able to provide an accurate description of important unsteady processes in the surf and swash zones, such as nonlinear wave transformations, run-up and flooding due to storm waves, see for instance \cite{Bonneton:2007p1191}, but it neglects the dispersive effects which are fundamental for the study of wave transformations in the shoaling area and possibly slightly deeper water areas.\\
Keeping the $O(\mu)$ terms in the analysis, the corresponding equations have been derived first by Serre \cite{Serre:1953p5261} in the horizontal surface dimension $d=1$, by Green and Naghdi \cite{CambridgeJournals:387109} for the $d=2$ case, and have
been recently mathematically justified in \cite{AlvarezSamaniego:2008p227}. \\
As far as numerical approximation is concerned, it is only recently that the Green-Naghdi (GN) equations have really received attention and several methods have been proposed including Finite Differences (FD), Finite Volumes (FV), Finite Elements (FE) or Spectral methods, see for instance among others \cite{Antunes-do-carmoSeabra-Santos:1993aa, Cienfuegos:2006p226, Li2014169, Dutykh:2013fk, shi:2012, PandaDawson:2014aa}.\\
After having developed and validated several hybrid FV-FD formulations based on a temporal splitting approach (see \cite{Bonneton20111479, MR2811693, rchgth89} for the $d=1$ case and \cite{lannes_marche:2014} for the $d=2$ case on structured cartesian grids), we have recently introduced in \cite{DuranMarche:2014ab} a new approach based on a fully discontinuous FE method in the $d=1$ case. This approach allows to decouple the hyperbolic and elliptic parts of the model by computing the solutions of the NSW equations supplemented by an additional algebraic source term, which fully accounts for the whole dispersive correction, and which is itself obtained from the resolution of auxiliary elliptic and coercive linear second order problems. With this approach we have paved the way towards:
\begin{enumerate}
\item[$\sharp$] more flexibility, as the proposed discrete formulation handles an arbitrary order of accuracy in space and, although limited in the $d=1$ case in \cite{DuranMarche:2014ab}, it conceptually generalizes to the dimension $d=2$ with unstructured meshes. It can moreover be straightforwardly used to enhance any numerical code based on the NSW equations,
\item[$\sharp$] more efficiency, as we built this approach on the new set of GN equations issued from \cite{lannes_marche:2014} which allows to perform the corresponding matrix assembling and factorization in a preprocessing step, leading to considerable computational time savings.
\end{enumerate}
Although similar decoupling strategies have subsequently been used to implement some hybrid approaches based on the classical GN equations (see \cite{Popinet:2015aa} for the $d=2$ case on cartesian meshes and \cite{FilippiniKazolea:2016aa} for the $d=1$ case), there is still no studies, up to our knowledge, that allows to approximate the solutions of GN equations on arbitrary unstructured meshes in the multidimensional case: this is the main purpose of this work. Starting from the new model recently introduced in \cite{lannes_marche:2014}, we consider a Discontinuous Galerkin (DG) formulation and generalize to the $d=2$ case the decoupling strategy of \cite{DuranMarche:2014ab}:
\begin{enumerate}
\item[$\sharp 1$] we compute the flow variables using a RK-DG approach \cite{Cockburn1998199} by approximating the solutions of the hyperbolic set of NSW equations supplemented with an additional source term that accounts for the whole dispersive correction,
\item[$\sharp 2$] the dispersive source term is obtained from the computation of auxiliary \textit{scalar} linear elliptic problems of second order, approximated through a mixed formulation and \textit{Local Discontinuous Galerkin (LDG)} stabilizing interface fluxes \cite{CockburnShu:1998aa, CastilloCockburn:2000aa}.
\end{enumerate}
\noindent
\noindent
This results in a global formulation of arbitrary order of accuracy in space, which is shown to preserve both the non-negativity of the water height and the motionless steady states up to the machine accuracy. These properties are particularly important in the study of propagating waves reaching the shore. Note that while this multidimensional strategy is clearly promising in terms of flexibility, the emphasize is put here on efficiency. First, the use of the \textit{diagonal constant} model issued from \cite{lannes_marche:2014} allows to considerably simplify the computations associated with the elliptic sub-problems. Indeed, the time evolutions of the velocity vector's components are here fully decoupled, thanks to the simplified analytical structure of the model. The problem therefore reduces to the approximation of \textit{scalar problems} without any third order derivatives, instead of the more complex vectorial ones obtained with the classical GN equations. The corresponding matrix can be assembled and the associated LU factorization stored in a pre-processing step. Secondly, the use of a \textit{nodal} approach \cite{HesthavenWarburton:2002aa} together with the \textit{pre-balanced} formulation of the hyperbolic part of the model (see for instance \cite{duran:2013, duran:marche:dg}), allows to combine
\begin{inparaenum}[(i)]
\item an efficient quadrature free treatment for the integrals which are not involved into the equilibrium balance,
\item a quadrature-based treatment with a lower computational cost, needed to exactly compute the surface and face integrals involved in the preservation of the steady states at rest,
\item a direct nodal product method, in the spirit of pseudo-spectral methods, for the strongly nonlinear terms occurring in the source terms of the elliptic sub-problems.\end{inparaenum}\\
This remainder of this work is organized as follows: we describe the mathematical model and the notations in the next section. Section $3$ is devoted to the introduction of the discrete settings and the DG formulations for both hyperbolic and elliptic sub-problems. This approach is then validated in the last section through convergence analysis and comparisons with data taken from experiments with several discriminating benchmark problems.
\section{The physical model}\label{model}
Let us denote by ${\mathbf{x}}=(x,y)$ the horizontal variables, $z$ the vertical variable and $t$ the time variable. In the following, $\zeta(t,{\mathbf{x}})$ describes the free surface elevation with respect to its rest state, $h_0$ is a reference depth, $-h_0+b({\mathbf{x}})$ is a parametrization of the bottom and $h:=h_0+\zeta-b$ is the water depth, as shown on Figure \ref{domain}.
Denoting by $U_{hor}$ the horizontal component of the
velocity field in the fluid domain, we define the vertically averaged
horizontal velocity ${\mathbf{v}}=(u,v)\in \ensuremath{\mathbb{R}}^2$ as
$$
{\mathbf{v}}(t,{\mathbf{x}})=\frac{1}{h}\int_{-h_0+b}^\zeta U_{hor}(t,{\mathbf{x}},z)dz,
$$
and we denote by ${\mathbf{q}}=h{\mathbf{v}}$ the corresponding horizontal momentum.
\medskip
\begin{figure}[H]
\psfrag{z}{{\footnotesize $z$}}
\psfrag{0}{{\footnotesize $0$}}
\psfrag{h0}[c][c]{\textcolor{white}{7\footnotesize{2}}{\footnotesize $-h_0$}}
\psfrag{N}{{\footnotesize $z=\zeta(t,{\mathbf{x}})$}}
\psfrag{F}{{\footnotesize $z=-h_0+b({\mathbf{x}})$}}
\psfrag{X}{{\footnotesize ${\mathbf{x}}=(x,y)$}}
\hspace{0.45em}\includegraphics[width=0.75\textwidth]{domain.eps}
\caption{Sketch of the domain}
\label{domain}
\end{figure}
\subsection{The Green-Naghdi equations}\label{NDGN}
According to \cite{Bonneton20111479}, the Green-Naghdi equations may be written as follows:
\begin{equation}\label{eq6}
\left\lbrace
\begin{array}{l}
\displaystyle \partial_t \zeta +\nabla\cdot (h {\mathbf{v}})=0,\\
\displaystyle (I+{\mathcal T}[h,b])\big[\partial_t {\mathbf{v}}+({\mathbf{v}}\cdot \nabla){\mathbf{v}}\big]+
g\nabla\zeta
+{\mathcal Q}_1[h,b]({\mathbf{v}})=0,
\end{array}\right.
\end{equation}
where the linear operator ${\mathcal T}[h,b]\cdot$ and the quadratic form ${\mathcal Q}_1[h,b](\cdot)$
are defined for all smooth enough $\ensuremath{\mathbb{R}}^2$-valued function ${\mathbf{w}}$ by
\begin{eqnarray}
\label{eq2}
{\mathcal T}[h,b]{\mathbf{w}} &=&
{\mathcal R}_1[h,b](\nabla\cdot {\mathbf{w}})
+{\mathcal R}_2[h,b](\nabla b\cdot {\mathbf{w}}),\\
\label{eq11}
{\mathcal Q}_1[h,b]({\mathbf{w}}) &=& -2{\mathcal R}_1[h,b](\partial_1{\mathbf{w}}\cdot\partial_2{\mathbf{w}}^\perp+(\nabla\cdot {\mathbf{w}})^2)
+{\mathcal R}_2[h,b]({\mathbf{w}}\cdot({\mathbf{w}}\cdot\nabla)\nabla b),
\end{eqnarray}
(here $\partial_1$ and $\partial_2$ denote space derivatives in the
two horizontal directions) with, for all smooth enough scalar-valued function $w$,
\begin{eqnarray}
\label{eq4}
{\mathcal R}_1[h,b]w&=&-\frac{1}{3h}\nabla(h^3 w)
-\frac{h}{2}w\nabla b,\\
\label{eq5}
{\mathcal R}_2[h,b]w&=&\frac{1}{2h}\nabla(h^2 w)
+w\nabla b.
\end{eqnarray}
We also recall in \cite{Bonneton20111479} that the dispersion properties of (\ref{eq6}) can be improved by adding some terms of order $O(\mu^2)$ to the momentum equation, which consequently does not affect the accuracy of the model. An asymptotically equivalent enhanced family of models parametrized by $\alpha > 0$ is given by
\begin{equation}\label{eq6imp}
\left\lbrace
\begin{array}{l}
\displaystyle \partial_t \zeta +\nabla\cdot (h {\mathbf{v}})=0,\\
\displaystyle (I+\alpha{\mathcal T}[h,b])\left( \partial_t {\mathbf{v}}+({\mathbf{v}}\cdot \nabla){\mathbf{v}}+\frac{\alpha-1}{\alpha}g\nabla\zeta\right)
+\frac{1}{\alpha}g\nabla\zeta
+{\mathcal Q}_1[h,b]({\mathbf{v}})=0.
\end{array}\right.
\end{equation}
or alternatively written in $(h,h{\mathbf{v}})$ variables:
\begin{equation}\label{eq6imphhVbis}
\left\lbrace
\begin{array}{l}
\displaystyle \partial_t h +\nabla\cdot (h {\mathbf{v}})=0,\\
\displaystyle (I+\alpha {\bf T}[h,b])\left( \partial_t (h{\mathbf{v}})
\!+\!\nabla\!\cdot\! (h{\mathbf{v}} \otimes {\mathbf{v}}) + \!\frac{\alpha\!-\!1}{\alpha} gh \nabla \zeta\right) + \frac{1}{\alpha} gh\nabla\zeta
+ h{\mathcal Q}_1[h,b]({\mathbf{v}}) = 0,
\end{array}\right.
\end{equation}
where we have introduced the operator ${\bf T}[h,b]$ defined as follows
$$
{\bf T}[h,b]{\mathbf{w}}=h{\mathcal T}[h,b](\frac{{\mathbf{w}}}{h}).
$$
As shown in \cite{DuranMarche:2014ab}, \eqref{eq6imphhVbis} can be recast as
\begin{subnumcases}{}
\partial_t h +\nabla\cdot (h {\mathbf{v}})=0,\label{disp:1} \\
\partial_t (h{\mathbf{v}}) +\nabla\cdot(h{\mathbf{v}}\otimes {\mathbf{v}}) + gh \nabla \zeta
+ \mathcal{D}_o=0,\label{disp:2}\\
(I\!+\!\alpha {\bf T}[h,b])(\mathcal{D}_o + \frac{1}{\alpha}gh \nabla \zeta ) = \frac{1}{\alpha} gh\nabla\zeta
+ h{\mathcal Q}_1[h,b]({\mathbf{v}}).\label{disp:3}
\end{subnumcases}
which highlights the fact that the dispersive correction of order $O(\mu)$ only acts as a source term $\mathcal{D}_o$ in \eqref{disp:2}, and is obtained as the solution of an auxiliary second-order elliptic sub-problem \eqref{disp:3}.
\begin{remark}\label{remark:benefits}
These formulations have two main advantages:
\begin{enumerate}
\item They do not require the computation of third order derivatives, while this is necessary in the standard formulation of the GN equations,
\item The presence of the operator $(I+\alpha{\bf T}[h,b])^{-1}$ makes the models robust
with respect to high frequency perturbations, which is an interesting property for numerical computations.
\end{enumerate}
\end{remark}
\begin{remark}
These formulations have two main drawbacks:
\begin{enumerate}
\item Solving linear systems arising from discrete formulations of \eqref{eq6imphhVbis} may be computationally expensive as $I+\alpha{\bf T}[h,b]$ is a \emph{matricial} second order differential operator acting on two-dimensional vectors. This structure entails a coupling of the time evolutions of the two components of $h{\mathbf{v}}$ through \eqref{disp:2}.
\item $I+\alpha{\bf T}[h,b]$ is a time dependent operator (through the dependence on $h$) and this is of course the same for any associated discrete formulations: the corresponding matrices have to be assembled at each time step or sub-steps.
\end{enumerate}
\end{remark}
\subsection{A Green-Naghdi model with a simplified analytical structure}\label{model:diag:const}
In \cite{lannes_marche:2014} , some new families of models are introduced to overcome these drawbacks, without loosing the benefits listed in Remark \ref{remark:benefits}. In particular, defining a modified water depth at rest (which therefore does not depend on time):
$$h_b=\max(h_0-b, \varepsilon_0)=\max(h-\zeta,\varepsilon_0),$$ where $\varepsilon_0$ is a numerical threshold introduced to account for possible dry areas in a consistent way, the \textit{one-parameter optimized constant-diagonal} GN equations read as follows:
\begin{equation}\label{eq6impnewter}
\left\lbrace
\begin{array}{l}
\displaystyle \partial_t h +\nabla\cdot (h {\mathbf{v}})=0,\\
\vspace{0.1cm}
\displaystyle \big[1+\alpha\mathds{T}[h_b]\big]\left(\partial_t (h{\mathbf{v}})
+\nabla\cdot (h{\mathbf{v}}\otimes {\mathbf{v}})+\frac{\alpha-1}{\alpha}gh\nabla\zeta\right) +
\frac{1}{\alpha}h\nabla\zeta + \mathds{Q}[h,b](\zeta, {\mathbf{v}})=0,
\
\end{array}\right.
\end{equation}
where for all smooth enough $\ensuremath{\mathbb{R}}$-valued function $w$:
\begin{equation}\label{Tcg}
\mathds{T}[h_b]w= -\frac{1}{3}\nabla\cdot(h_b^3\nabla\frac{w}{h_b}),
\end{equation}
and
\begin{equation}\label{Qcg}
\mathds{Q}[h,b](\zeta, {\mathbf{v}})=h({\mathcal Q}_1[h,b]({\mathbf{v}}) +g{\mathcal Q}_2[h,b](\zeta)) +g{\mathcal Q}_3[h,h_b]\left( \big[1+\alpha \mathds{T}[h_b]\big]^{-1}(h\nabla\zeta)\right),
\end{equation}
is a second order nonlinear operator with
\begin{equation}\label{Q2CG}
{\mathcal Q}_2[h,b](\zeta) = -h(\nabla^\perp h\cdot\nabla) \nabla^\perp\zeta
-\frac{1}{2h}\nabla(h^2 \nabla b\cdot \nabla\zeta) + \big(\frac{h}{2}\Delta\zeta - (\nabla b\cdot \nabla\zeta)\big)\nabla b,
\end{equation}
and for all smooth enough $\ensuremath{\mathbb{R}}$-valued function $w$
\begin{equation}\label{defS}
{\mathcal Q}_3[h,h_b] w= \frac{1}{6}\nabla (h^2-h_b^2)\cdot \nabla w +\frac{h^2-h_b^2}{3}\Delta w -\frac{1}{6}\Delta(h^2-h_b^2)w.
\end{equation}
\begin{remark}
We actually show in \cite{lannes_marche:2014} that it is indeed possible to replace the inversion of $I+\alpha{\bf T}[h,b]$ by the inversion of $1+\alpha\mathds{T}[h_b]$, where $\mathds{T}[h_b]$ depends only on the fluid at rest (i.e. $\zeta=0$), while keeping the asymptotic $O(\mu^2)$ order of the expansion. The interest of working with (\ref{eq6impnewter}) rather than (\ref{eq6imphhVbis}) is that $\mathds{T}[h_b]$ has a simplified \textit{scalar} structure, \textit{i.e.} it can be written in matricial form as
\begin{equation}\label{structTtilde}
\left(\begin{array}{cc}
-\frac{1}{3}\nabla\cdot(h_b^3\nabla\frac{1}{h_b}\cdot) & 0\\
0 & -\frac{1}{3}\nabla\cdot(h_b^3\nabla\frac{1}{h_b}\cdot)
\end{array}\right).
\end{equation}
From a numerical viewpoint, this simplified analytical structure allows to compute each component of the discharge $h{\mathbf{v}}$ separately, and to alleviate the computational cost associated with the dispersive correction of the model, as the discrete version of $1+\alpha\mathds{T}[h_b]$ may be assembled and factorized once and for all, in a preprocessing step.
\end{remark}
\begin{remark}
The other difference with (\ref{eq6imphhVbis}) is the presence of the modified quadratic term $\mathds{Q}[h,b]$, which shares with (\ref{eq6imphhVbis}) the nice property that no computation of third order derivative is needed. The price to pay is the inversion of an extra linear
system, through the computation of ${\mathcal Q}_3[h,h_b]\left( \big[1+\alpha \mathds{T}[h_b]\big]^{-1}(h\nabla\zeta)\right)$. However, this extra cost is largely off-set by the gain obtained by using the time independent scalar operator $\mathds{T}[h_b]$.
\end{remark}
\begin{remark}\label{reminv}
One could replace ${\mathcal Q}_3[h,h_b]\left( \big[I+\alpha\mathds{T}[h_b]\big]^{-1}(h\nabla\zeta)\right)$ by ${\mathcal Q}_3[h,h_b]\left(h\nabla\zeta\right)$ in the second equation of (\ref{eq6impnewter}), keeping the same asymptotic $O(\mu^2)$ order. This would avoid the resolution of this extra linear system but leads to
strong instabilities. We
refer to \cite{lannes_marche:2014} for more comments on this point.
\end{remark}
\begin{remark}
As shown in \cite{DuranMarche:2014ab}, the model \eqref{eq6impnewter} can be recast as follows:
\begin{subnumcases}{\left(\mathcal{CG}_\alpha\right)}
\partial_t h +\nabla\cdot (h {\mathbf{v}})=0,\label{disp:1c} \\
\partial_t (h{\mathbf{v}}) +\nabla\cdot(h{\mathbf{v}}\otimes {\mathbf{v}}) + gh \nabla \zeta
+ \mathcal{D}_c=0,\label{disp:2c}\\
\big[1\!+\!\alpha\mathds{T}[h_b] \big](\mathcal{D}_c + \frac{1}{\alpha} h\nabla\zeta)=
h( \frac{1}{\alpha} \nabla\zeta + {\mathcal Q}_1[h,b]({\mathbf{v}}) +g{\mathcal Q}_2[h,b](\zeta)) +{\mathcal Q}_3[h,h_b]\mathcal{K},\label{disp:3c}\\
\big[1\!+\!\alpha\mathds{T}[h_b]\big]\mathcal{K} = gh\nabla\zeta,\label{disp:4c}
\end{subnumcases}
and we see that the dispersive correction $\mathcal{D}_c$ acting as a source term in \eqref{disp:2c}, is obtained as the solution of auxiliary scalar second-order elliptic sub-problems.
\end{remark}
\begin{remark}
We highlight that $\left(\mathcal{CG}_\alpha\right)$ still does not involve third order derivative computation, which is a very interesting property. It is indeed shown in \cite{FilippiniKazolea:2016aa} that when third order derivatives on the free surface occur, it is important to introduce some sophisticated approximation strategies for the free surface gradient to reduce the dispersion error.
\end{remark}
\subsection{A pre-balanced Green-Naghdi formulation}\label{model:pre:bal}
Paving the way towards the construction of a well-balanced and efficient discrete formulation in \S\ref{discrete:sec}, we now adjust the \textit{pre-balanced} approach of \cite{Liang2009873} for the case $d=1$ and \cite{duran:2013} for the case $d=2$ to the Green-Naghdi equations $\left(\mathcal{CG}_\alpha\right)$. We introduce the {\it total free surface elevation} $\eta=h+b$, denote ${\mathbf{W}}={}^t(\eta, {\mathbf{q}})$ and use the following splitting of the hydrostatic pressure term:
\begin{equation}\label{pre:bal:split}
gh\nabla \zeta = \frac{1}{2}g\nabla(\eta^2 - 2\eta b) + g\eta \nabla b,
\end{equation}
to obtain the \textit{pre-balanced} formulation of \eqref{disp:1c}-\eqref{disp:2c}
\begin{subnumcases}{}
\partial_t \eta +\nabla\cdot {\mathbf{q}}=0,\label{disp:1d} \\
\partial_t {\mathbf{q}} +\nabla\cdot \mathcal{F}({\mathbf{W}}, b)
+ \mathcal{D}_c= \mathcal{B}(\eta,b),\label{disp:2d
\end{subnumcases}
with the flux and topography source terms defined as
\begin{equation}\label{flux:pre:bal}
\mathcal{F}({\mathbf{W}},b) = \frac{{\mathbf{q}}\otimes {\mathbf{q}}}{\eta-b} + \frac{1}{2}g\nabla(\eta^2 - 2\eta b) I_2,\quad\quad
\mathcal{B}(\eta,b)=- g\eta \nabla b.
\end{equation}
We obtain a \textit{one-parameter pre-balanced constant-diagonal Green-Naghdi} model (more simply referred to as $\left(\mathcal{PCG}_\alpha\right)$ model in the following):
\begin{subnumcases}{}
\partial_t \eta +\nabla\cdot {\mathbf{q}}=0,\label{pb:disp:1c} \\
\partial_t {\mathbf{q}} +\nabla\cdot \mathcal{F}({\mathbf{W}}, b)
+ \mathcal{D}_c= \mathcal{B}(\eta,b),\label{pb:disp:2c}\\
\big[1\!+\!\alpha\mathds{T}[h_b] \big](\mathcal{D}_c + \frac{1}{\alpha} gh\nabla\eta)=
h(\frac{1}{\alpha}g\nabla\eta + {\mathcal Q}_1[h,b]({\mathbf{v}}) +g{\mathcal Q}_2[h,b](\eta)) +{\mathcal Q}_3[h,h_b]\mathcal{K},\label{pb:disp:3c}\\
\big[1\!+\!\alpha\mathds{T}[h_b]\big]\mathcal{K} = gh\nabla\eta.\label{pb:disp:4c}
\end{subnumcases}
\section{Discrete formulation}\label{discrete:sec}
\subsection{Settings and notations}
Let $\Omega\subset\ensuremath{\mathbb{R}}^d$, $d=2$, denote an open bounded connected polygonal domain with boundary $\partial\Omega$. We consider a geometrically conforming mesh $\mathcal{T}_{\h}$ defined as a finite collection of nonempty disjoint open triangular elements $T$ of boundary $\partial T$ such that $\closure{\Omega}=\bigcup_{T\in\mathcal{T}_{\h}}\closure{T}$. The meshsize is defined as ${\mathfrak{h}}=\max\limits_{T\in\mathcal{T}_{\h}} \,{\mathfrak{h}}_T$
with ${\mathfrak{h}}_T$ standing for the diameter of the element $T$ and we denote $\vert T \vert$ the area of $T$, $\mathfrak{p}_T$ its perimeter and $\vec{n}_{T}$ its unit outward normal.\\
\noindent
Mesh faces are collected in the set $\mathcal{F}_{\h}$ and the length of a face $F\in\mathcal{F}_{\h}$ is denoted by $\vert F\vert $. A mesh face $F$ is such that either there exist $T_1,T_2\in\mathcal{T}_{\h} $ such that $F\subset\partial T_1\cap\partial T_2$ ($F$ is called an interface and $F\in\Fh^{{\rm i}}$) or there exists $T\in\mathcal{T}_{\h}$ such that $F\subset\partial T \cap\partial\Omega$ ($F$ is called a boundary face and $F\in\Fh^{{\rm b}}$).
For all $T\in\mathcal{T}_{\h}$, ${\mathcal F}_T\mathrel{\mathop:}=\{F\in\mathcal{F}_{\h}\; | \; F\subset\partial T \}$ denotes the set of faces belonging to $\partial T $ and, for all $F\in{\mathcal F}_T$, $\vec{n}_{TF}$ is the unit normal to $F$ pointing out of $T$.\\
\noindent
In what follows, we consider $\mathbb{P}^{k}(\mathcal{T}_{\h})$ the broken bivariate polynomial space defined as follows:
\begin{equation}\label{broken}
\mathbb{P}^{k}(\mathcal{T}_{\h}):= \lbrace{ v \in L^{2}(\Omega) \, \vert \, \;v_{\vert T} \,\in \, \mathbb{P}^{k}(T) \;\;\forall \, T \, \in \,\mathcal{T}_{{\mathfrak{h}}}\rbrace},
\end{equation}
where $\mathbb{P}^{k}(T)$ denotes the space of bivariates polynomials in $T$ of degree at most $k$, and
we define $X_{\mathfrak{h}} = \mathbb{P}^{k}(\mathcal{T}_{\h})\times \left( \mathbb{P}^{k}(\mathcal{T}_{\h})\right)^2$. We denote $N_k = \dim(\mathbb{P}^{k}(T))=(k+1)(k+2)/2$.\\
\noindent
To discretize in time, for a given final computational time $t_{\textrm{max}}$, we consider a partition $(t^n)_{0\leq n\leq N}$ of the time interval $[0, t_{\textrm{max}}]$ with $t^0=0$, $t^N=t_{\textrm{max}}$ and $t^{n+1}-t^{n}=\Delta t^n$. For any sufficiently regular function of time $w$, we denote by $w^n$ its value at discrete time $t^n$. More details on the computations of $\Delta t^n$ and the time marching algorithms are given in \S\ref{Time}.
\subsection{Discrete formulation for the advection-dominated equations}
Postponing to \S\ref{sect:disp} the computation of the projection of the dispersive correction $\mathbb{D}({\mathbf{W}},b)$ on the approximation space, we focus here on the discrete formulation associated to \eqref{pb:disp:1c}-\eqref{pb:disp:2c}, written in a more compact way:
\begin{equation}\label{compact}
\partial_t {\mathbf{W}}+\nabla\cdot \mathbb{F}({\mathbf{W}},b) + \mathbb{D}({\mathbf{W}},b)=\mathbb{B}({\mathbf{W}},b),
\end{equation}
with
\begin{align*}
\mathbb{B}({\mathbf{W}},b) = {}^t \left( 0, {}^t\mathcal{B}(\eta,b) \right), \;\;\;\; \mathbb{D}({\mathbf{W}},b)={}^t\left(0, {}^t\mathcal{D}_c\right)
\;\;\;\;\mbox{and}\;\;\;\;
\mathbb{F}({\mathbf{W}},b) =
\begin{pmatrix} {}^t{\mathbf{q}} \\ \mathcal{F}({\mathbf{W}},b) \end{pmatrix}.
\end{align*}
\subsubsection{The discrete problem}
We seek an approximate solution ${\mathbf{W}}_{\mathfrak{h}}={}^t\left(\eta_{\mathfrak{h}}, {}^t{\mathbf{q}}_{\mathfrak{h}} \right)$ of \eqref{compact} in $X_{\mathfrak{h}}$. Requiring the associated residual to be orthogonal to $\mathbb{P}^{k}(\mathcal{T}_{\h})$, the semi-discrete formulation reduces to the local statement: find $(\eta_{\mathfrak{h}}, {}^t{\mathbf{q}}_{\mathfrak{h}})$ in $X_{\mathfrak{h}}$ such that
\begin{align}
&\int_{T} \frac{d}{dt} {\mathbf{W}}_{\mathfrak{h}}\pi_{{\mathfrak{h}}}d{\mathbf{x}} -\int_{T}\mathbb{F}({\mathbf{W}}_{\mathfrak{h}}, b_{\mathfrak{h}})\cdot\nabla \pi_{{\mathfrak{h}}}d{\mathbf{x}} \,+ \int_{\partial T}(\mathbb{F}({\mathbf{W}}_{\mathfrak{h}}, b_{\mathfrak{h}})\cdot \vec{n}_{T}) \,\pi_{{\mathfrak{h}}}ds\nonumber\\
&\quad\quad\quad\quad\quad + \int_{T} \mathbb{D}({\mathbf{W}}_{\mathfrak{h}},b_{\mathfrak{h}})\pi_{{\mathfrak{h}}}d{\mathbf{x}} = \int_{T} \mathbb{B}({\mathbf{W}}_{\mathfrak{h}},b_{\mathfrak{h}})\pi_{{\mathfrak{h}}}d{\mathbf{x}},\label{discrete:2}
\end{align}
for all $\pi_{\mathfrak{h}}\in\mathbb{P}^{k}(\mathcal{T}_{\h})$ and all element $T\in\mathcal{T}_{\h}$, where $b_{\mathfrak{h}}$ refers to the $L^2$ projection of $b$ on $\mathbb{P}^{k}(\mathcal{T}_{\h})$ and $\mathbb{D}({\mathbf{W}}_{\mathfrak{h}},b_{\mathfrak{h}})$ stands for a polynomial description of $\mathbb{D}({\mathbf{W}},b)$ in $X_{\mathfrak{h}}$, to be obtained in \S\ref{sect:disp}.
\begin{remark}\label{rem:base:T}
Considering a local basis $\{ \phi_i\}_{i=1}^{N_k}$ for a given element $T\in \mathcal{T}_{\h}$,
the local discrete solution ${\mathbf{W}}_{{\mathfrak{h}}\vert T}$ may be expanded as:
\begin{equation}\label{expansion}
{\mathbf{W}}_{{\mathfrak{h}}\vert T}({\mathbf{x}},t) = \sum\limits_{i=1}^{N_k} \tilde{{\mathbf{W}}}_i(t) \phi_i({\mathbf{x}}),\;\;\forall {\mathbf{x}}\in T, \forall t \in [0, t_{max}],
\end{equation}
where $\{\tilde{{\mathbf{W}}}_i(t)\}_{i=1}^{N_k}$ are the local expansion coefficients, defined as $\tilde{{\mathbf{W}}}_i(t)={}^t(\tilde{\eta}_i, {}^t\tilde{{\mathbf{q}}}_i)$ and $\tilde{{\mathbf{q}}}_i = {}^t((\tilde{q}_x)_i, (\tilde{q}_y)_i)$.\\
Many choices are of course possible for the basis. In the following, $\{ \phi_i\}_{i=1}^{N_k}$ refers to the interpolant (nodal) basis on the element $T$ and we choose the Fekete nodes \cite{TaylorWingate:2000aa} as approximation points.\\
\end{remark}
\noindent
Equipped with such local basis, following Remark \ref{rem:base:T}, and splitting the boundary integral into faces integrals, the local statement \eqref{discrete:2} is now equivalent to
\begin{equation}
\begin{split}
\sum\limits_{i=1}^{N_k}& \left(\int_{T} \phi_{i}\phi_{j}\,d{\mathbf{x}}\right) \frac{d}{ dt}\tilde{{\mathbf{W}}}_i(t)
-\int_{T}\mathbb{F}({\mathbf{W}}_{{\mathfrak{h}}},b_{\mathfrak{h}})\cdot\nabla \phi_j\,d{\mathbf{x}} + \sum\limits_{F\in{\mathcal F}_T} \int_{F} \widehat{\mathbb{F}}_{TF}\,\phi_j\,ds\\& + \int_{T}\mathbb{D}({\mathbf{W}}_{{\mathfrak{h}}},b_{\mathfrak{h}})\phi_j\,d{\mathbf{x}}= \int_{T}\mathbb{B}({\mathbf{W}}_{{\mathfrak{h}}},b_{\mathfrak{h}})\phi_j\,d{\mathbf{x}},\;\;\quad 1\leq j\leq N_k,
\label{Weak_formulation2}
\end{split}
\end{equation}
where $ \widehat{\mathbb{F}}_{TF}$ is a stabilizing numerical approximation of the normal interface flux $\mathbb{F}({\mathbf{W}}_{{\mathfrak{h}}},b_{\mathfrak{h}})\cdot \vec{n}_{TF}$, to be defined in the following.
\subsubsection{Interface fluxes and well-balancing}\label{numerical:flux}
We recall in the following a simple choice to approximate the interface fluxes \cite{duran:marche:dg}, leading to a well-balanced scheme that preserves motionless steady states. This modified flux can also be seen as the adaptation of the ideas of \cite{xing:2013} to the {\it pre-balanced} formulation (\ref{disp:1d})-(\ref{disp:2d}).\\
Consider a face $F\in{\mathcal F}_T$ (for the sake of simplicity, we only focus on the case $F\in \mathcal{F}_{\h}^i$ and do not detail the case $F\in \mathcal{F}_{\h}^b$). Let us denote denote ${\mathbf{W}}^{-}$ and ${\mathbf{W}}^{+}$ respectively the {\it interior} and {\it exterior} traces on $F$, with respect to the elements $T$. Similarly, $b^{-}$ and $b^{+}$ stand for the {\it interior} and {\it exterior} traces of $b_{{\mathfrak{h}}}$ on $F$.
We define:
\begin{equation}\label{topo:check}
b^* = \max(b^-,b^+),\quad\quad \check{b} = b^* - \max(0, b^* - \eta^-)
\end{equation}
and
\begin{align}
&\check{h}^- = \max(0, \eta^- - b^*), \quad\quad \check{h}^+ = \max(0, \eta^+ - b^*),\label{def:1}\\
&\check{\eta}^- = \check{h}^- + \check{b}, \quad\quad\quad\quad\quad\;\; \check{\eta}^+ = \check{h}^+ + \check{b},\label{def:2}
\end{align}
leading to the new {\it interior} and {\it exterior} values:
\begin{equation}\label{ext_int}
\check{{\mathbf{W}}}^{-} = {}^t(\check{\eta}^-, \frac{\check{h}^-}{\eta^- - b^-}{\bf q^-}),\;\;\;\; \check{{\mathbf{W}}}^{+} = {}^t(\check{\eta}^+, \frac{\check{h}^+}{\eta^+ - b^+}{\bf q^+}).
\end{equation}
Now we set
\begin{equation}\label{num:flux}
\widehat{\mathbb{F}}_{TF} = \mathbb{F}_{\mathfrak{h}}(\check{{\mathbf{W}}}^{-}, \check{{\mathbf{W}}}^{+}, \check{b}, \check{b},\vec{n}_{TF}) + \widetilde{\mathbb{F}}_{TF},
\end{equation}
as the numerical flux function through the interface $F$,
where:
\begin{enumerate}
\item the numerical flux function $\mathbb{F}_{\mathfrak{h}}$ is the global Lax-Friedrichs flux:
\begin{equation}\label{LFflux}
\mathbb{F}_{\mathfrak{h}}({\mathbf{W}}^-, {\mathbf{W}}^+, b^-, b^+,\vec{n}_{TF}) = \frac{1}{2}( \mathbb{F}({\mathbf{W}}^-, b^-)\cdot \vec{n}_{TF} + \mathbb{F}({\mathbf{W}}^+, b^+)\cdot \vec{n}_{TF} - a({\mathbf{W}}^+-{\mathbf{W}}^-)),
\end{equation}
with $a= \max\limits_{T\in\mathcal{T}_{\h}} \lambda_T$ and
\begin{equation}\label{lambda}
\lambda_T = \max\limits_{\partial T} \left( \left | \frac{{\mathbf{q}}_{{\mathfrak{h}}\vert T}}{\eta_{{\mathfrak{h}}\vert T} - b_{{\mathfrak{h}}\vert T}} \cdot \vec{n}_{T} \right | +\sqrt{g(\eta_{{\mathfrak{h}}\vert T} - b_{{\mathfrak{h}}\vert T})} \right).
\end{equation}
\item $\widetilde{\mathbb{F}}_{TF}$ is a correction term defined as follows:
\begin{equation}
\widetilde{\mathbb{F}}_{TF} = \begin{pmatrix} 0 & 0 \\ g\check{\eta}^-(\check{b} - b^-) & 0 \\0 & g\check{\eta}^-(\check{b} -b^-)\end{pmatrix}\cdot \vec{n}_{TF}.
\end{equation}
\end{enumerate}
\noindent
Note that the modified interface flux (\ref{num:flux}) only induces perturbations of order $k+1$ when compared to the traditional interface fluxes.
\begin{remark}\label{rem:LF}
Let denote in the following by $w_{T}$ the averaged value of the discrete approximation $w_{{\mathfrak{h}}\vert T}$ on the element $T$, for any $\ensuremath{\mathbb{R}}$ or $\ensuremath{\mathbb{R}}^2$-valued function $w$. Let consider a first order scheme for the averaged free-surface:
\begin{equation}\label{schemaDGave2}
\eta^{n+1}_{T} = \eta^{n}_{T} - \frac{\Delta t^n}{\vert T\vert}\sum\limits_{F\in{\mathcal F}_T} \int_{F} \mathbb{F}_{\mathfrak{h}}(\check{{\mathbf{W}}}^{-}, \check{{\mathbf{W}}}^{+}, \check{b}, \check{b},\vec{n}_{TF}) \,ds,
\end{equation}
with $\mathbb{F}_{\mathfrak{h}}$ defined following (\ref{LFflux}) and $\check{{\mathbf{W}}}^{-}$, $\check{{\mathbf{W}}}^{+}, \check{b}$ obtained from (\ref{topo:check})-(\ref{ext_int}), starting from ${\mathbf{W}}^{-}$ and ${\mathbf{W}}^{+}$ defined respectively, for each face $F$, as the first-order \textit{piecewise constant} interior and exterior values ${\mathbf{W}}_{T_1}^n$ and ${\mathbf{W}}_{T_2}^n$, with ${\mathcal T}_F=\{T_1, T_2\}$.\\
Then, assuming that $h^{n}_{T}\geq 0,\;\forall T\in {\mathcal T}_{\mathfrak{h}}$, we have $h^{n+1}_{T}\geq 0$ $\forall T\in {\mathcal T}_{\mathfrak{h}}$, under the condition
\begin{equation}\label{cfl:firstorder}
\lambda_T \frac{ \mathfrak{p}_T}{\vert T\vert} \Delta t^n \leq 1,\;\;\forall T\in {\mathcal T}_{\mathfrak{h}}.
\end{equation}
This is a positivity property of the \textit{Lax-Friedrichs} flux extended to the \textit{pre-balanced} formulation, which is shown in \cite{duran:marche:dg} and is mandatory to obtain a positive high-order DG scheme, see \cite{Zhang20103091}.
\end{remark}
\subsubsection{Positivity}\label{dry}
The enforcement strategy of the water height non-negativity within the DG formulation \cite{xing:2013} requires positivity of the water height at carefully chosen quadrature nodes at the beginning of each time step. Then, the positivity of the water height is ensured providing the use of a positivity preserving first order scheme, like (\ref{LFflux}) (see Remark \ref{rem:LF}) and a suitable time-step restriction.\\
We adjust these ideas to the $\left(\mathcal{PCG}_\alpha\right)$ formulation (\ref{pb:disp:1c})-(\ref{pb:disp:2c}) to enforce the mandatory property that the mean value of the water height $h_{T}=\eta_T-b_T$ on any given element $T\in\mathcal{T}_{\h}$ remains positive during the time marching procedure. The main ideas are summarized for an explicit first order \textit{Euler} scheme in time for the sake of simplicity:
\begin{enumerate}
\item[$\sharp$ 1] considering the broken polynomial space $\mathbb{P}^{k}(T)$, we assume that the face integrals are computed using $(k+1)$-points \textit{Gauss} quadrature (see Remark \ref{quad:pre}). The special quadrature rules introduced in \cite{Zhang:2012fk} are obtained by a transformation of the tensor product of a $\beta$-points \textit{Gauss-Lobatto} quadrature (with $\beta$ the smallest integer such that $2\beta-3\geq k$) and the $(k+1)$-point \textit{Gauss} quadrature. This new quadrature includes all $(k+1)$-point Gauss quadrature nodes for each face $F\in {\mathcal F}_T$, involves positive weights and it is exact for the integration of $\eta^{n}_{{\mathfrak{h}}}$ over $T$. In the following, let us denote $S_{T}^k$ the set of points of this quadrature rule. We show on Figure \ref{quadraturepoints} the resulting quadrature nodes used for $k=2$ and $k=3$ orders of approximation on a reference element.
\item[$\sharp$ 2] for each element $T\in\mathcal{T}_{\h}$, $h_{{\mathfrak{h}}\vert T}^n$ is computed from $\eta_{{\mathfrak{h}}\vert T}^n$ and $b_{{\mathfrak{h}}\vert T}$ and we need to ensure that $h_{{\mathfrak{h}}\vert T}^n({\mathbf{x}}) \geq 0,\; \forall {\mathbf{x}}\in S_{T}^k$, which is a sufficient condition to ensure the non-negativity property for the DG scheme (\ref{discrete:2}), under the CFL-like condition \eqref{cfl2}.\\
This condition is enforced using the accuracy preserving limiter of \cite{Zhang20103091}. Assuming $h_{T}^{n}\geq 0$, we replace $h_{{\mathfrak{h}}\vert T}^n$ by a conservative linear scaling around this element average:
\begin{equation}\label{limiter:robuste}
\hat{h}_{{\mathfrak{h}}\vert T}^n = \theta_{T}(h_{{\mathfrak{h}}\vert T}^n - h_{T}^{n}) + h_{T}^{n},
\end{equation}
with
$$
\theta_{T} = \min \left( \frac{h_{T}^{n}}{h_{T}^{n} -m_{T}} ,1\right), \;\;\;\;\mbox{and}\;\;\;\;
m_{T} = \min\limits_{{\mathbf{x}}\in S_{T}^k} \;h_{{\mathfrak{h}}\vert T}^n({\mathbf{x}}).
$$
\end{enumerate}
\begin{remark}
Note that this approach also ensures that the water height remains positive at the $k+1$ \textit{Gauss} quadrature nodes used to compute the faces integrals in \eqref{Weak_formulation2}.
\end{remark}
\begin{figure}
\begin{center}
\includegraphics[scale=0.22,angle=-90]{triP2}
\includegraphics[scale=0.22,angle=-90]{triP3}
\caption{Nodes locations for the special \textit{Zhang and Shu} quadrature - $\mathbb{P}^2$ and $\mathbb{P}^3$ cases.}
\label{quadraturepoints}
\end{center}
\end{figure}
\subsection{Discrete formulations for the second order elliptic sub-problems}\label{sect:disp}
We are now left with the computation of the projection $\mathbb{D}({\mathbf{W}}_{{\mathfrak{h}}},b_{\mathfrak{h}})$ of the dispersive correction on the approximation space $X_{\mathfrak{h}}$. As its first component identically vanishes, the computation of the discrete version of $\mathcal{D}_c$ is obtained as the solution of the discrete problems associated with
\begin{subnumcases}{}
\big[1\!+\!\alpha\mathds{T}[h_b]\big]\mathcal{K} = gh\nabla\eta,\label{subell1}\\
\big[1\!+\!\alpha\mathds{T}[h_b] \big](\mathcal{D}_c + \frac{1}{\alpha} gh\nabla\eta)= h(\frac{1}{\alpha} g\nabla\eta + {\mathcal Q}_1[h,b]({\mathbf{v}}) +g{\mathcal Q}_2[h,b](\eta)) +{\mathcal Q}_3[h, h_b]\mathcal{K}.\label{subell2}
\end{subnumcases}
Note that both problems ultimately reduce to the construction of a discrete formulation associated with the generic scalar problem
$$\big[1\!+\!\alpha\mathds{T}[h_b]\big] w = f,$$
where the source term $f$ is successively defined as the directional scalar components of the source terms occurring in \eqref{subell1}-\eqref{subell2}, that is to say
\begin{equation}\label{source:term:f}
f=
\begin{cases}
gh\nabla\eta\;\; \mbox{for}\; \eqref{subell1}\\
\displaystyle\frac{1}{\alpha} gh\nabla\eta + h({\mathcal Q}_1({\mathbf{v}}) +g{\mathcal Q}_2(\eta)) +{\mathcal Q}_3[h, h_b]\mathcal{K}\;\;\mbox{for}\; \eqref{subell2},
\end{cases}
\end{equation}
where we keep a vectorial notation for the sake of simplicity.
Remarking now that the following identity holds for any smooth enough scalar-valued function:
\begin{equation}\label{struct:div}
\mathds{T}[h_b]w = -\frac{1}{3}\nabla\cdot\left(H^b\nabla w \right) + \frac{1}{6}\nabla\cdot\left( w\nabla H^b\right),
\end{equation}
with the notation $H^b=h_b^2$, we consider the following mixed formulation, introducing a diffusive flux ${\mathbf{p}}$:
\begin{equation}\label{mixed:disp}
\begin{cases}
w-\alpha\nabla\cdot\left(\frac{1}{3} H^b {\mathbf{p}} - \frac{1}{6}w\nabla H^b\right) = f,\\
{\mathbf{p}} - \nabla w = 0,
\end{cases}
\end{equation}
and the following associated discrete problem: find $(w_{\mathfrak{h}}, {}^t{\mathbf{p}}_{\mathfrak{h}})$ in $X_{\mathfrak{h}}$ such that
\begin{align}
&\int_{T} w_{\mathfrak{h}}\pi_{{\mathfrak{h}}}d{\mathbf{x}} + \frac{\alpha}{3}\int_{T} H^b_{\mathfrak{h}} ({\mathbf{p}}_{\mathfrak{h}}\cdot\nabla \pi_{{\mathfrak{h}}})d{\mathbf{x}} - \frac{\alpha}{3}\int_{\partial T} H^b_{\mathfrak{h}} ({\mathbf{p}}_{\mathfrak{h}}\cdot \vec{n}_T)\pi_{\mathfrak{h}}\,ds\,\label{discrete:1T}\\&\hspace{2cm}-\frac{\alpha}{6}\int_T w_{\mathfrak{h}}(\nabla H^b_{\mathfrak{h}}\cdot\nabla\pi_{\mathfrak{h}})\,d{\mathbf{x}} +\frac{\alpha}{6}\int_{\partial T} w_{\mathfrak{h}}(\nabla H^b_{\mathfrak{h}}\cdot\vec{n}_T)\pi_{\mathfrak{h}}\,ds = \int_T f_{\mathfrak{h}}\pi_{\mathfrak{h}}\,d{\mathbf{x}},\nonumber\\
&\int_{T} {\mathbf{p}}_{\mathfrak{h}}\cdot\phi_{{\mathfrak{h}}}d{\mathbf{x}} + \int_T w_{\mathfrak{h}}\nabla\cdot\phi_{\mathfrak{h}}\,d{\mathbf{x}} - \int_{\partial T} w_{\mathfrak{h}}(\phi_{\mathfrak{h}}\cdot\vec{n}_T)\,ds= 0,\label{discrete:2T}
\end{align}
for all $(\pi_{\mathfrak{h}},\phi_{\mathfrak{h}})\in X_{\mathfrak{h}}$, where $H^b_{\mathfrak{h}}$ and $\nabla H^b_{\mathfrak{h}}$ respectively denote the $L^2$ projections of $h_b^2$ and $\nabla h_b^2$ on $\mathbb{P}^{k}(\mathcal{T}_{\h})$ and $(\mathbb{P}^{k}(\mathcal{T}_{\h}))^2$, and $f_{\mathfrak{h}}$ is a polynomial representation of $f$ in $\mathbb{P}^{k}(\mathcal{T}_{\h})$, with $f$ defined according to \eqref{source:term:f}. We choose to use the \textit{Local Discontinuous Galerkin (LDG)} approach \cite{CockburnShu:1998aa} to compute the associated stabilizing fluxes. For any given element $T\in\mathcal{T}_{\h}$, we have:
\begin{align}
&\int_{\partial T} H^b_{\mathfrak{h}} ({\mathbf{p}}_{\mathfrak{h}}\cdot \vec{n}_T)\pi_{\mathfrak{h}}\,ds=\sum\limits_{F\in{\mathcal F}_T}\int_{F} H^b_{\mathfrak{h}} (\widehat{{\mathbf{p}}}_{TF}\cdot \vec{n}_{TF})\pi_h\,ds,\\
&\int_{\partial T} w_{\mathfrak{h}}(\phi_{\mathfrak{h}}\cdot\vec{n}_T)\,ds = \sum\limits_{F\in{\mathcal F}_T}\int_{F}\widehat{w}_{TF}(\phi_{\mathfrak{h}}\cdot\vec{n}_{TF})\,ds,\\
&\int_{\partial T} w_{\mathfrak{h}}(\nabla H^b_{\mathfrak{h}}\cdot\vec{n}_T)\pi_{\mathfrak{h}}\,ds= \sum\limits_{F\in{\mathcal F}_T}\int_{F} \widehat{w}_{TF} (\nabla H^b_{\mathfrak{h}}\cdot\vec{n}_T)\pi_{\mathfrak{h}}\,ds,
\end{align}
where the interface fluxes are computed as follows:
\begin{align}
&\widehat{w}_{TF} = \avg{w_{\mathfrak{h}}} + ({\boldsymbol \beta}\cdot\vec{n}_{TF})\jump{w_{\mathfrak{h}}},\label{flux:ldg:1}\\
&\widehat{{\mathbf{p}}}_{TF} = \avg{{\mathbf{p}}_{\mathfrak{h}}} - {\boldsymbol \beta}(\jump{{\mathbf{p}}_{\mathfrak{h}}}\cdot\vec{n}_{TF}) + \frac{\xi}{\vert F\vert}\jump{w_{\mathfrak{h}}}\vec{n}_{TF},\label{flux:ldg:2}
\end{align}
with the classical notations for the face average $\avg{w} =(w^{+}+w^{-})/2$ and face jump $\jump{w}= w^{+}-w^{-}$, and $w^{-}$ and $w^{+}$ stand respectively for the interior and exterior traces of $w_{{\mathfrak{h}}}$ with respect to the considered face $F$ (and similar notations for the $\ensuremath{\mathbb{R}}^2$-valued quantity ${\mathbf{p}}$). The penalization parameter $\xi$ and the upwinding parameter ${\boldsymbol \beta}$ are both set to $1$.
\begin{remark}
The choice of the \textit{LDG} fluxes (\ref{flux:ldg:1})-(\ref{flux:ldg:2}) allows to eliminate locally the auxiliary discrete flux ${\mathbf{p}}_{\mathfrak{h}}$, and to globally assemble the matrix corresponding to the discrete formulation of $\big[1\!+\!\alpha\mathds{T}[h_b]\big]$. Note also that, for the sake of simplicity, we also use the upwind flux (\ref{flux:ldg:1}) to approximate the faces contributions associated with the purely advective part of (\ref{struct:div}).
\end{remark}
\begin{remark}
The relative simplicity of the discrete formulation (\ref{discrete:1T})-(\ref{discrete:2T}) is a direct consequence of the simplified analytical structure of the $\left(\mathcal{PCG}_\alpha\right)$ formulation. One can of course adapt this approach to the initial formulation (\ref{eq6imphhVbis}), leading in practice to a more costly and complicated global assembling process, as the operator ${\mathcal T}[h,b]$ directly acts on a $\ensuremath{\mathbb{R}}^2$-valued function.
\end{remark}
For the sake of efficiency, the computations of the source terms integrals $\int_T f_{\mathfrak{h}}\pi_{\mathfrak{h}}\,d{\mathbf{x}}$ occurring in \eqref{discrete:1T} are performed in a collocation way, in the spirit of pseudo-spectral methods. More precisely, we build approximated integrands with polynomial representations $({\mathcal Q}_{j,{\mathfrak{h}}})_{j=1,2,3}$ in $\left( \mathbb{P}^{k}(\mathcal{T}_{\h})\right)^2$ of the operators $({\mathcal Q}_j)_{j=1,2,3}$ using direct products of the discrete flow variables $(\eta_{\mathfrak{h}}, {\mathbf{v}}_{\mathfrak{h}})$ and their derivatives at the \textit{Fekete} nodes. These derivatives are however weakly computed with the DG approach, using LDG fluxes. In practice, this is simply achieved by adjusting the approach of \cite{DuranMarche:2014ab} to the $d=2$ case to easily compute the required first and second order derivatives of $\eta_{\mathfrak{h}}$ and ${\mathbf{v}}_{\mathfrak{h}}$ in each space direction. Again, the second order spatial derivatives are written in mixed form and we use the stabilizing fluxes \eqref{flux:ldg:1}-\eqref{flux:ldg:2}. The corresponding discrete formulations are simplified through local elimination of the diffusive fluxes, leading to the assembling of global matrices for first and second order weak derivatives in each direction. For instance, considering derivatives in the first direction, the mixed form
\begin{equation*}
v + \partial_x w = 0 \quad , \quad u + \partial_x v = 0,\,
\end{equation*}
leads to the following local discrete formulations for $1\leq j\leq N_k$:
\begin{equation}\label{LDG_scheme}
\begin{split}
\sum\limits_{i=1}^{N_k} \tilde{v}_i M_{ij} & =
\sum\limits_{i=1}^{N_k} \tilde{w}_i S_{ij}^x -\sum\limits_{F\in{\mathcal F}_T} \int_{F} \widehat{w}_{TF} \phi_{j} n_{TF}^x ds\, , \\
\sum\limits_{i=1}^{N_k} \tilde{u}_i M_{ij} & =
\sum\limits_{i=1}^{N_k} \tilde{v}_i S_{ij}^x -\sum\limits_{F\in{\mathcal F}_T} \int_{F} \widehat{v}_{TF} \phi_{j} n_{TF}^x\, ds,
\end{split}
\end{equation}
where $n_{TF}^x$ is the first component of $\vec{n}_{TF}$ and with
\begin{equation*}
M_{ij} := \int_{T} \phi_{i} \phi_{j}\,d{\mathbf{x}} \quad , \quad S^x_{ij} := \int_{T} \phi_{i} \partial_x \phi_{j}\, d{\mathbf{x}}.
\end{equation*}
These systems can be globally rewritten as follows:
\begin{equation}\label{LDG_Matrice}
\begin{split}
&{\mathbb{M}} \tilde{V} = {\mathbb{S}}_{x} \tilde{W} - \left({\mathbb{E}}_{x} - {\mathbb{F}}_{x} \right) \tilde{W} \, , \\
&{\mathbb{M}} \tilde{U} = {\mathbb{S}}_{x} \tilde{V} - \left({\mathbb{E}}_{x} + {\mathbb{F}}_{x} \right) \tilde{V} - \frac{\xi}{{\mathfrak{h}}} {\mathbb{F}}_{x} \tilde{W},
\end{split}
\end{equation}
where:
\begin{inparaenum}[(i)]
\item $\tilde{U}, \tilde{V}$ and $\tilde{W}$ are $N_k\times \card \mathcal{T}_{\h}$ vectors gathering the expansion coefficients $\tilde{u}, \tilde{v}$ and $\tilde{w}$ for all mesh elements,
\item ${\mathbb{M}}$ and ${\mathbb{S}}_x$ are the square $\card{\mathcal{T}_{\h}} \times N_k$ global mass and stiffness matrices, with a block-diagonal structure,
\item ${\mathbb{E}}_x$ and ${\mathbb{F}}_x$ are square $\card{\mathcal{T}_{\h}} \times N_k$ matrices, globally assembled by gathering all the mesh faces contributions, accounting respectively for the average $\avg{\cdot}$ and jump $\jump{\cdot}$ operators occurring in the definition \eqref{flux:ldg:1}-\eqref{flux:ldg:2} of the \textit{LDG} fluxes.
\end{inparaenum} Note that each interface $F\in{\mathcal F}^i_{\mathfrak{h}}$ contributes to four blocks of size $N_k$ in the global matrices and each boundary face $F\in{\mathcal F}^b_{\mathfrak{h}}$ contributes to one block. Equipped with these global structures, first and second order global differentiation matrices can be straightforwardly assembled, leading to the following compact notations:
\begin{align}
&\tilde{V} = {\mathbb{D}}_x \tilde{W},\;\;\mbox{with}\;\;{\mathbb{D}}_x = {\mathbb{M}}^{-1} \left({\mathbb{S}}_x - {\mathbb{E}}_x + {\mathbb{F}}_x \right),\\
&\tilde{U} = {\mathbb{D}}_{xx} \tilde{W},\;\;\mbox{with}\;\; {\mathbb{D}}_{xx} = {\mathbb{M}}^{-1} \Big(\left({\mathbb{S}}_x-{\mathbb{E}}_x-{\mathbb{F}}_x\right){\mathbb{D}}_x - \frac{\xi}{{\mathfrak{h}}} {\mathbb{F}}_x \Big).
\end{align}
Similar constructions are performed for ${\mathbb{D}}_y$ and ${\mathbb{D}}_{yy}$.
\begin{remark}
The use of such a direct nodal products method helps to reduce the computational cost by reducing the accuracy of quadrature but may threaten the numerical stability for strongly nonlinear or marginally resolved problems with the possible introduction of what is known in the field of spectral methods as aliasing driven instabilities. Several well-known methods may help to alleviate this issue, like the use of stabilization filtering methods. In this work however, and considering the test cases studied in \S\ref{section:numerical}, we did not need to use any additional stabilization mechanism.
\end{remark}
\subsection{Time-marching, boundary conditions and wave-breaking}
\subsubsection{Time discretization}\label{Time}
The time stepping is carried out using the explicit third-order SSP-RK scheme \cite{gottliedtadmor}. Up to $k=3$, we consider RK-SSP schemes of order $k+1$. A fourth order SSP-RK scheme is used for $k\geq 3$. For instance, writing the semi-discrete equations as $\frac{d}{dt}{\mathbf{W}}_{{\mathfrak{h}}}+\mathcal{A}_{{\mathfrak{h}}}({\mathbf{W}}_{{\mathfrak{h}}}) = 0$, advancing from time level $n$ to $n + 1$ is computed as follows with the third-order scheme:
\begin{equation} \label{RK3}
\left \lbrace \begin{array}{ll}
\vspace{0.2cm}
{\mathbf{W}}_{{\mathfrak{h}}}^{n,1}={\mathbf{W}}_{{\mathfrak{h}}}^{n}- \Delta t^n \tilde{\mathcal{A}}_{{\mathfrak{h}}}({\mathbf{W}}_{{\mathfrak{h}}}^{n}) \, , \\
\vspace{0.2cm}
{\mathbf{W}}_{{\mathfrak{h}}}^{n,2}=\frac{1}{4}(3{\mathbf{W}}_{{\mathfrak{h}}}^{n}+{\mathbf{W}}_{{\mathfrak{h}}}^{n,1})- \frac{1}{4}\Delta t^n \tilde{\mathcal{A}}_{{\mathfrak{h}}}({\mathbf{W}}_{{\mathfrak{h}}}^{n,1}) \, ,\\
{\mathbf{W}}_{{\mathfrak{h}}}^{n+1}=\frac{1}{3}({\mathbf{W}}_{{\mathfrak{h}}}^{n}+2{\mathbf{w}}_{{\mathfrak{h}}}^{n,2})- \frac{2}{3}\Delta t^n \tilde{\mathcal{A}}_{{\mathfrak{h}}}({\mathbf{W}}_{{\mathfrak{h}}}^{n,2}) \, .
\end{array} \right.
\end{equation}
with the formal notation $\tilde{\mathcal{A}}_{\mathfrak{h}} = \mathcal{A}_{\mathfrak{h}} \circ \Lambda\Pi_{\mathfrak{h}}$, ${\mathbf{W}}_{{\mathfrak{h}}} \leftarrow \Lambda\Pi_{\mathfrak{h}} {\mathbf{W}}_{{\mathfrak{h}}}$ being the limitation operator possibly acting on the approximated vector solution (see \S\ref{Limitation}), and $\Delta t^n$ is obtained from the CFL condition \eqref{cfl2}.
\subsubsection{Boundary conditions}\label{boundary}
For the test cases studied in the next section, boundary conditions are imposed weakly, by enforcing suitable reflecting relations at virtual exterior nodes, at each boundaries, allowing to compute the corresponding interface stabilizing fluxes. Solid-wall (reflective) but also periodic conditions (as the computational domains geometries of the cases studied in \S\ref{section:numerical} are rectangular) can be enforced following this simple process. \\
These simple boundary conditions possibly have to be complemented with \textit{ad-hoc} absorbing boundary conditions, allowing the dissipation of the incoming waves energy together with an efficient damping of possibly non-physical reflections, and generating boundary conditions that mimic a wave generator of free surface waves. We have implemented relaxation techniques and we enforce periodic waves combined with generation/absorption by mean of a generation/relaxation zone, following the ideas of \cite{madsen2003}, using the relaxation functions described in \cite{Wei:1995p885}, and the computational domain is locally extended to include sponge layers which may also include a generating layer. The length of these layers has to be calibrated from the incoming waves (generally 2 or 3 wavelengths).
\subsubsection{Wave-breaking and limiting}\label{Limitation}
Obviously, vertically averaged models cannot reproduce the surface wave overturning and are therefore inherently unable to fully model wave breaking. Moreover, if the GN equations can accurately reproduce most phenomena exhibited by non-breaking waves in finite depth, including the steepening process occurring just before breaking, they do not intrinsically account for the energy dissipation mechanism associated with the conversion into turbulent kinetic energy observed during broken waves propagation. \\
Several methods have been proposed to embed wave breaking in depth averaged models. Many of them focus on the inclusion of an energy dissipation mechanism through the activation of extra terms in the governing equations when wave breaking is likely to occur, and the reader is referred to \cite{tissier2} for a recent review of these approaches. More recently, hybrid strategies have been elaborated for weakly nonlinear models \cite{Tonelli:2009p1244, Kazolea:2014kx} and for fully nonlinear models \cite{Bonneton20111479, tissier2, tissier1, shi:2012}. Roughly speaking, the idea is to switch from GN to NSW equations when the wave is ready to break by locally suppressing the dispersive correction and the various approaches may differ by the level of sophistication of the detection criteria and the switching strategies.\\
Denoting that, from a numerical point of view, such an approach allows to avoid numerical instabilities by turning off the computation of higher-order derivatives and nonlinear and non-conservative terms in the vicinity of appearing singularities, we use in \cite{DuranMarche:2014ab} a purely numerical smoothness detector to identify the potential instability areas, switch to NSW equations in such areas by simply locally neglecting the $\mathcal{D}_c$ source term and use a limiter strategy to stabilize the computation, letting breaking fronts propagate as moving bores.\\
As the aim of the present work is only to introduce and validate our fully discontinuous discrete formulation in the multidimensional case, we do not focus on the development of new wave breaking strategies and we simply adjust this simple approach to the $d=2$ case to possibly stabilize the computations performed in \S\ref{section:numerical}. \\
We detect the troubled elements (following the terminology of \cite{Qiu:siam:2005}) using the criterion proposed in \cite{Krivodonova2004323}, and based on the strong superconvergence phenomena exhibited at element's outflow boundaries. More precisely, for any $T\in{\mathcal T}_{\mathfrak{h}}$, denoting by $\partial T_{in}$ the inflow part of $\partial T$, we use the following criterion:
\begin{equation}\label{sensor}
\mathbb{I}_T= \frac{\sum\limits_{F\in\partial T_{in}}\displaystyle\int_{F}(h^- - h^+)\,ds}{{\mathfrak{h}}_T^{(k+1)/2}\vert \partial T_{in}\vert \| h_{{\mathfrak{h}}\vert T}\| },
\end{equation}
which has already successfully been used as a \textit{troubled cells} detector in purely hyperbolic shallow water models, see among others \cite{Ern:2008p3536, kesser2012, zhu:2013}.\\
If $\mathbb{I}_T\geq 1$ then we apply a slope limiter on each scalar component of ${\mathbf{W}}_{{\mathfrak{h}}\vert T}$, based on the \textit{maxmod} function, see \cite{Burbeau2001111}. This limiting strategy is not recalled here, as we straightforwardly reproduce the implementation described in \cite{duran:marche:dg} for the NSW equations in \textit{pre-balanced} form. Of course, waves about to break do not embed any free surface singularities yet, but our numerical investigations have shown that when the wave has steepened enough, the free surface gradient becomes large enough to activate the criteria. This simple approach, although quite rough and far less sophisticated than recent strategies introduced for instance in \cite{tissier2, Kazolea:2014kx}, allows us to obtain good results in the various cases of \S\ref{section:numerical}.
\subsection{Main properties}
\noindent
We have the following result:
\begin{proposition} \label{well_balancing}
The discrete formulation (\ref{Weak_formulation2}) together with the interface fluxes discretization (\ref{num:flux}) and a first order \textit{Euler} time-marching algorithm has the following properties:
\begin{enumerate}
\item it preserves the motionless steady states, providing that the integrals of \eqref{Weak_formulation2} are exactly computed for the motionless steady states. In other terms, we have for all $n\in\ensuremath{\mathbb{N}}$:
\begin{equation} \label{WB} \Big( \left \lbrace \begin{array}{ll}
\eta_{{\mathfrak{h}}}^{n} \equiv \eta^e \\
\textbf{q}_{{\mathfrak{h}}}^{n} \equiv 0 \end{array} \right. \Big) \Rightarrow \quad \Big(\left \lbrace \begin{array}{ll}
\eta_{{\mathfrak{h}}}^{n+1} \equiv \eta^e \\
\textbf{q}_{{\mathfrak{h}}}^{n+1} \equiv 0 \end{array} \right. \Big),
\end{equation}
with $\eta^e$ constant,
\item assuming moreover that $h_{T}^{n} \geq 0,\;\forall T\in{\mathcal T}_{\mathfrak{h}}$ and $h_{{\mathfrak{h}}\vert T}^n({\mathbf{x}})\geq 0,\;\forall {\mathbf{x}}\in S_{T}^k,\;\forall T\in{\mathcal T}_{\mathfrak{h}}$, then we have $h_{T}^{n+1} \geq 0, \;\forall T\in{\mathcal T}_{\mathfrak{h}}$
under the condition
\begin{equation}\label{cfl2}
\lambda_T \frac{ \mathfrak{p}_T}{\vert T\vert} \Delta t^n \leq \frac{2}{3} \hat{\omega}_1^\beta,
\end{equation}
where $\hat{\omega}_1^\beta$ is the first quadrature weight of the $\beta$-point \textit{Gauss-Lobatto} rule used in \S\ref{dry}.
\end{enumerate}
\end{proposition}
\begin{proof}
\begin{enumerate}
\item assuming that the following equilibrium${\mathbf{W}}_{\mathfrak{h}}={\mathbf{W}}_{\mathfrak{h}}^e=(\eta^e, 0)$ holds, we have to show that $\forall T\in\mathcal{T}_{\h}$ and $1\leq j\leq N_k$:
\begin{equation}
\int_{T}\mathbb{F}({\mathbf{W}}_{{\mathfrak{h}}}^e,b_{\mathfrak{h}})\cdot\nabla \phi_j\,d{\mathbf{x}} - \sum\limits_{F\in{\mathcal F}_T} \int_{F} \mathbb{F}_{TF}^e\,\phi_j\,ds - \int_{T}\mathbb{D}({\mathbf{W}}_{{\mathfrak{h}}}^e,b_{\mathfrak{h}})\phi_j\,d{\mathbf{x}}+\int_{T}\mathbb{B}({\mathbf{W}}_{{\mathfrak{h}}}^e,b_{\mathfrak{h}})\phi_j\,d{\mathbf{x}}=0,\\
\label{wb:1}
\end{equation}
where $\mathbb{F}_{TF}^e$ is the interface numerical flux obtained at equilibrium.
Looking at (\ref{num:flux}), and highlighting that for each interface $F$ we have $\check{\eta}^- = \check{\eta}^+=\eta^e$ and therefore $\check{{\mathbf{W}}}^{-}=\check{{\mathbf{W}}}^{+}$, it is easy to check that $ \mathbb{F}_{TF}^e= \mathbb{F}({\mathbf{W}}^-, b^-)\cdot\vec{n}_{TF}$, thanks to the consistency of the numerical flux function $\mathbb{F}_{\mathfrak{h}}$. Consequently, we have
\begin{align}\label{wb:4}
\int_{T}\mathbb{F}({\mathbf{W}}_{{\mathfrak{h}}}^e,b_{\mathfrak{h}})\cdot\nabla \phi_j d{\mathbf{x}} - \sum\limits_{F\in{\mathcal F}_T} \int_{F} \mathbb{F}_{TF}^e\,\phi_j\,ds &= -\int_{T}\nabla\cdot\mathbb{F}({\mathbf{W}}_{{\mathfrak{h}}}^e,b_{\mathfrak{h}})\phi_jd{\mathbf{x}},\\
&=-\int_{T} \mathbb{B}({\mathbf{W}}_{{\mathfrak{h}}}^e,b_{\mathfrak{h}})\phi_jd{\mathbf{x}}.
\end{align}
assuming that the integrals are computed exactly (see Remark \ref{quad:pre}), and
observing that we have $\nabla_{\mathfrak{h}}\cdot \mathbb{F}({\mathbf{W}}_{{\mathfrak{h}}}^e,b_{\mathfrak{h}})=\mathbb{B}({\mathbf{W}}_{{\mathfrak{h}}}^e,b_{\mathfrak{h}})$. For the integrals associated with the dispersive source term, it is straightforward to check from the definitions \eqref{eq11} and \eqref{Q2CG} of $({\mathcal Q}_j)_{j=1,2}$ that ${\mathcal Q}_{1,{\mathfrak{h}}}=0$ whenever ${\mathbf{q}}_{\mathfrak{h}}=0$ and that ${\mathcal Q}_{2,{\mathfrak{h}}}=0$ whenever $\eta_{\mathfrak{h}}=\eta^e$. Moreover, using again the fact that $\nabla_{\mathfrak{h}}\eta^e=0$, the polynomial representation $\mathcal{K}_{\mathfrak{h}}$ of $\mathcal{K}$, obtained as the solution of the discrete problem \eqref{discrete:1T}-\eqref{discrete:2T} associated with \eqref{subell1}, taking $f_{\mathfrak{h}}=0$ as source term and suitable boundary conditions, identically vanishes leading to ${\mathcal Q}_{3,{\mathfrak{h}}}=0$ using \eqref{defS}. Note that this result holds no matter what method we use to compute polynomials representations $({\mathcal Q}_{j,{\mathfrak{h}}})_{j=1,2,3}$.
\item the proof of \cite{Zhang:2012fk}, adjusted to the pre-balanced framework in\cite{duran:marche:dg}, can be straightforwardly reproduced, as the dispersive source term $\mathbb{D}({\mathbf{W}}_{{\mathfrak{h}}},b_{\mathfrak{h}})$ has no influence on the mass conservation equation.
Note that such positivity preservation can be extended to the third order SSP scheme of \S\ref{Time}, see \cite{Zhang:2012fk}.
\end{enumerate}
\end{proof}
\begin{remark}\label{quad:pre}
The face integrals of \eqref{Weak_formulation2} can be \textit{exactly} computed at motionless steady states with a $k+1$ point Gauss quadrature rule thanks to the \textit{pre-balanced} splitting \eqref{pre:bal:split}. Indeed, with the resulting formulation of the advective flux \eqref{flux:pre:bal}, the integrands on faces are of order $2k$ at equilibrium. In the same way, the cubature rules used for the surface integrals have to be exact only for bivariate polynomials of order $2k-1$.
\end{remark}
\section{Numerical validations}\label{section:numerical}
In this section, we validate the previous discrete formulation through several benchmarks. Unless stated otherwise, we use periodic boundary conditions in each direction, we set $\alpha=1.159$, $\varepsilon_0=0.1$ and the time step restriction is computed according to \eqref{cfl2}. Accordingly with the non-linear stability result of the previous section, we do not suppress the dispersive effects in the vicinity of dry areas. Some accuracy analysis are performed in the first two cases using the broken $L^2$ norm defined as follows for any arbitrary scalar valued piecewise polynomial function $w_{\mathfrak{h}}$ defined on $\mathcal{T}_{\h}$:
$$
\norm{w_{\mathfrak{h}}}_{L^2(\mathcal{T}_{\h})} = \left( \sum\limits_{T\in\mathcal{T}_{\h}}\norm{w_{{\mathfrak{h}}\vert T}}_{L^2(T)}^2 \right)^{\frac{1}{2}}.
$$
\subsection{Preservation of motionless steady state}
This preliminary test case is devoted to check the ability of the formulation to preserve motionless steady states and accuracy validation. The computational domain is the [-1,1] $\times$ [-1,1] square, and we use an unstructured mesh of $8466$ elements. The bottom elevation involves a bump and a hollow having same dimensions, respectively located at ${\mathbf{x}}_1=(x_1,y_1)=(-\dfrac{1}{3} , -\dfrac{1}{3})$ and ${\mathbf{x}}_2=(x_2,y_2)=(\dfrac{1}{3} , \dfrac{1}{3})$, leading to the following analytic profile :
\begin{equation}
b(r_1,r_2)= 1 + d \, e^{-(r_{1}/L)^{2}} - d \, e^{-(r_{2}/L)^{2}} \, ,
\end{equation}
where $r_{1,2}$ are respectively the distances from ${\mathbf{x}}_1$ and ${\mathbf{x}}_2$ and we set $d = 0.45$ and $L = 0.15$. The reference water depth is $h_{0} = 1.5\,m$, leading to the configuration depicted on Fig. \ref{WB_fig1}. Our numerical investigations confirm that this initial condition is preserved up to the machine accuracy for any value of polynomial order $k$. For instance, the $L^{2}$ numerical errors obtained at $t=50\,s$ using a $k=2$ approximation are respectively $3.0e$-$16$, $7.4e$-$16$ and $7.9e$-$16$ for $\eta$, $hu$ and $hv$. \\
Keeping the same computational domain and topography profile, we also perform a convergence study, using a reference solution obtained with $k=4$ and a regular triangulation with space steps $\Delta x = \Delta y= 2^{-9}\,m$. The initial free surface is set to:
\begin{equation}
\zeta(t=0, {\mathbf{x}}) = a \,e^{-(\norm{{\mathbf{x}}}/L)^2} \, \quad , \quad a = 0.075\,h_0.
\end{equation}
The computations are performed on a sequence of regular triangular meshes with increasing refinement ranging from $2^{-4}\,m$ to $2^{-9}\,m$ and polynomial expansions of degrees ranging from $k=1$ to $k=4$, while keeping the time step constant and small enough to ensure that the leading error orders are provided by the spatial discretization. The $L^{2}$ errors computed at $t_{\textrm{max}}=0.02\,s$ lead to the convergence curves shown on Fig. \ref{order_h} for the free surface elevation and Fig. \ref{order_q} for the discharge. The corresponding convergence rates obtained by linear regression are reported on each curve. We note that while the $L^2$-errors are typically larger on the discharge than on the free surface, we asymptotically reach some convergence rates between $\mathcal{O}({\mathfrak{h}}^{k+\frac{1}{2}})$ and $\mathcal{O}({\mathfrak{h}}^{k+1})$ for both variables. We observe that for this problem, the resulting convergence rates are greater than the $\mathcal{O}({\mathfrak{h}}^{k+\frac{1}{2}})$ optimal estimate one would be expected, see \cite{ZhangShu:2004aa}.
\begin{figure}
\begin{center}
\includegraphics[scale=0.25,angle=0]{WB2.eps}
\caption{Test 1 - Motionless steady states preservation: topography and free surface}
\label{WB_fig1}
\end{center}
\end{figure}
\begin{figure}
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={$\Delta x$},
ylabel={$L^2$-error},
xtick={0,-1,-2,-3,-4},
legend pos=south east,
ymajorgrids=true,
grid style=dashed,
yticklabel style = {font=\tiny},
xticklabel style = {font=\tiny},
]
\addplot[
color=blue,
mark=square,
] table [y=a, x expr=1/\thisrow{dx}] {conv_h1.txt};
\addplot[
color=red,
mark=diamond,
] table [y=b, x expr=1/\thisrow{dx}] {conv_h1.txt};
\addplot[
color=brown,
mark=o,
] table [y=c, x expr=1/\thisrow{dx}] {conv_h1.txt};
\addplot[
color=magenta,
mark=star,
] table [y=d, x expr=1/\thisrow{dx}] {conv_h4.txt};
\addplot[color=black,
mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=a,
variance list={10, 10, 10, 1000}}}]{conv_h1_line.txt}
coordinate [pos=0.25] (A)
coordinate [pos=0.1] (B);
\xdef\slope{\pgfplotstableregressiona}
\draw (A) -| (B)
node [pos=0.1,anchor=south]
{\pgfmathprintnumber{\slope}};
\addplot[color=black,
mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=b,
variance list={10, 10, 10,1000}}}]{conv_h1_line.txt}
coordinate [pos=0.1] (C)
coordinate [pos=0.25] (D)
;
\xdef\slopeq{\pgfplotstableregressiona}
\draw (C) -| (D)
node [pos=0.2,anchor=north]
{\pgfmathprintnumber{\slopeq}};
\addplot[color=black, mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=c,
variance list={10, 10, 10, 1000}}}]{conv_h1_line.txt}
coordinate [pos=0.25](E)
coordinate [pos=0.1](F);
\xdef\slopec{\pgfplotstableregressiona}
\draw (E) -| (F)
node [pos=0.2,anchor=south]
{\pgfmathprintnumber{\slopec}};
\addplot[color=black, mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=d,
variance list={10, 10, 10, 1000}}}]{conv_h4_line.txt}
coordinate [pos=0.1](G)
coordinate [pos=0.23](H);
\xdef\sloped{\pgfplotstableregressiona}
\draw (G) -| (H)
node [pos=0.2,anchor=north]
{\pgfmathprintnumber{\sloped}};
\legend{$k=1$,$k=2$,$k=3$,$k=4$}
\end{loglogaxis}
\end{tikzpicture}
\caption{Test 1 - $L^2$-error for the free surface elevation vs. $\Delta x$ for $k=1,2,3$ and $k=4$ at $t_{\textrm{max}}=0.02\,s$. }\label{order_h}
\end{figure}
\begin{figure}
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={$\Delta x$},
ylabel={$L^2$-error},
xtick={0,-1,-2,-3,-4},
legend pos=south east,
ymajorgrids=true,
grid style=dashed,
yticklabel style = {font=\tiny},
xticklabel style = {font=\tiny},
]
\addplot[
color=blue,
mark=square,
] table [y=a, x expr=1/\thisrow{dx}] {conv_u1.txt};
\addplot[
color=red,
mark=diamond,
] table [y=b, x expr=1/\thisrow{dx}] {conv_u1.txt};
\addplot[
color=brown,
mark=o,
] table [y=c, x expr=1/\thisrow{dx}] {conv_u1.txt};
\addplot[
color=magenta,
mark=star,
] table [y=d, x expr=1/\thisrow{dx}] {conv_u4.txt};
\addplot[color=black,
mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=a,
variance list={10, 10, 10, 1000, 1000}}}]{conv_u1_line.txt}
coordinate [pos=0.25] (A)
coordinate [pos=0.1] (B);
\xdef\slope{\pgfplotstableregressiona}
\draw (A) -| (B)
node [pos=0.2,anchor=south]
{\pgfmathprintnumber{\slope}};
\addplot[color=black,
mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=b,
variance list={10, 10, 10,1000, 1000}}}]{conv_u1_line.txt}
coordinate [pos=0.1] (C)
coordinate [pos=0.25] (D)
;
\xdef\slopeq{\pgfplotstableregressiona}
\draw (C) -| (D)
node [pos=0.2,anchor=north]
{\pgfmathprintnumber{\slopeq}};
\addplot[color=black, mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=c,
variance list={10, 10, 10, 1000, 1000}}}]{conv_u1_line.txt}
coordinate [pos=0.22](E)
coordinate [pos=0.1](F);
\xdef\slopec{\pgfplotstableregressiona}
\draw (E) -| (F)
node [pos=0.2,anchor=south]
{\pgfmathprintnumber{\slopec}};
\addplot[color=black, mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=d,
variance list={10, 10, 10, 1000}}}]{conv_u4_line.txt}
coordinate [pos=0.1](G)
coordinate [pos=0.23](H);
\xdef\sloped{\pgfplotstableregressiona}
\draw (G) -| (H)
node [pos=0.2,anchor=north]
{\pgfmathprintnumber{\sloped}};
\legend{$k=1$,$k=2$,$k=3$,$k=4$}
\end{loglogaxis}
\end{tikzpicture}
\caption{Test 1 - $L^2$-error for the discharge vs. $\Delta x$ for $k=1,2,3$ and $k=4$ at $t_{\textrm{max}}=0.02\,s$}\label{order_q}
\end{figure}
\begin{figure}[H]
\begin{center}
\includegraphics[scale=0.25,angle=0]{Soliton_NS.eps}
\caption{Test 2 - Solitary wave propagation over a flat bottom: initial free surface}
\label{soliton:init}
\end{center}
\end{figure}
\begin{figure}[H]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={$\Delta x$},
ylabel={$L^2$-error},
xtick={0,-1,-2,-3,-4},
legend pos=south east,
ymajorgrids=true,
grid style=dashed,
yticklabel style = {font=\tiny},
xticklabel style = {font=\tiny},
]
\addplot[
color=blue,
mark=square,
] table [y=a, x expr=1/\thisrow{dx}] {t2_conv_h1.txt};
\addplot[
color=red,
mark=diamond,
] table [y=b, x expr=1/\thisrow{dx}] {t2_conv_h1.txt};
\addplot[
color=brown,
mark=o,
] table [y=c, x expr=1/\thisrow{dx}] {t2_conv_h1.txt};
\addplot[
color=magenta,
mark=star,
] table [y=d, x expr=1/\thisrow{dx}] {t2_conv_h4.txt};
\addplot[color=black,
mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=a,
variance list={10, 10, 10, 10, 1000}}}]{t2_conv_h1_line.txt}
coordinate [pos=0.25] (A)
coordinate [pos=0.1] (B);
\xdef\slope{\pgfplotstableregressiona}
\draw (A) -| (B)
node [pos=0.1,anchor=south]
{\pgfmathprintnumber{\slope}};
\addplot[color=black,
mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=b,
variance list={10, 10, 10, 10, 1000}}}]{t2_conv_h1_line.txt}
coordinate [pos=0.1] (C)
coordinate [pos=0.25] (D)
;
\xdef\slopeq{\pgfplotstableregressiona}
\draw (C) -| (D)
node [pos=0.2,anchor=north]
{\pgfmathprintnumber{\slopeq}};
\addplot[color=black, mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=c,
variance list={10, 1000, 10, 10}}}]{t2_conv_h1_line.txt}
coordinate [pos=0.25](E)
coordinate [pos=0.1](F);
\xdef\slopec{\pgfplotstableregressiona}
\draw (E) -| (F)
node [pos=0.2,anchor=south]
{\pgfmathprintnumber{\slopec}};
\addplot[color=black, mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=d,
variance list={10, 10, 10, 1000}}}]{t2_conv_h4_line.txt}
coordinate [pos=0.1](G)
coordinate [pos=0.23](H);
\xdef\sloped{\pgfplotstableregressiona}
\draw (G) -| (H)
node [pos=0.2,anchor=north]
{\pgfmathprintnumber{\sloped}};
\legend{$k=1$,$k=2$,$k=3$,$k=4$}
\end{loglogaxis}
\end{tikzpicture}
\caption{Test 2 - $L^2$-error for the free surface elevation vs. $\Delta x$ for $k=1,2,3$ and $k=4$ at $t_{\textrm{max}}=0.2\,s$. }\label{order_h_solit}
\end{figure}
\begin{figure}[H]
\centering
\begin{tikzpicture}
\begin{loglogaxis}[
xlabel={$\Delta x$},
ylabel={$L^2$-error},
xtick={0,-1,-2,-3,-4},
legend pos=south east,
ymajorgrids=true,
grid style=dashed,
yticklabel style = {font=\tiny},
xticklabel style = {font=\tiny},
]
\addplot[
color=blue,
mark=square,
] table [y=a, x expr=1/\thisrow{dx}] {t2_conv_u1.txt};
\addplot[
color=red,
mark=diamond,
] table [y=b, x expr=1/\thisrow{dx}] {t2_conv_u1.txt};
\addplot[
color=brown,
mark=o,
] table [y=c, x expr=1/\thisrow{dx}] {t2_conv_u1.txt};
\addplot[
color=magenta,
mark=star,
] table [y=d, x expr=1/\thisrow{dx}] {t2_conv_u4.txt};
\addplot[color=black,
mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=a,
variance list={10, 10, 10, 10, 1000}}}]{t2_conv_u1_line.txt}
coordinate [pos=0.25] (A)
coordinate [pos=0.1] (B);
\xdef\slope{\pgfplotstableregressiona}
\draw (A) -| (B)
node [pos=0.1,anchor=south]
{\pgfmathprintnumber{\slope}};
\addplot[color=black,
mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=b,
variance list={10, 10, 10, 10, 10}}}]{t2_conv_u1_line.txt}
coordinate [pos=0.25] (C)
coordinate [pos=0.1] (D)
;
\xdef\slopeq{\pgfplotstableregressiona}
\draw (C) -| (D)
node [pos=0.2,anchor=south]
{\pgfmathprintnumber{\slopeq}};
\addplot[color=black, mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=c,
variance list={10, 1000, 10, 10}}}]{t2_conv_u1_line.txt}
coordinate [pos=0.25](E)
coordinate [pos=0.1](F);
\xdef\slopec{\pgfplotstableregressiona}
\draw (E) -| (F)
node [pos=0.2,anchor=south]
{\pgfmathprintnumber{\slopec}};
\addplot[color=black, mark=no,]
table[
x expr=1/\thisrow{dx},
y={create col/linear regression={y=d,
variance list={10, 10, 10, 10}}}]{t2_conv_u4_line.txt}
coordinate [pos=0.1](G)
coordinate [pos=0.23](H);
\xdef\sloped{\pgfplotstableregressiona}
\draw (G) -| (H)
node [pos=0.2,anchor=north]
{\pgfmathprintnumber{\sloped}};
\legend{$k=1$,$k=2$,$k=3$,$k=4$}
\end{loglogaxis}
\end{tikzpicture}
\caption{Test 2 - $L^2$-error for the discharge vs. $\Delta x$ for $k=1,2,3$ and $k=4$ at $t_{\textrm{max}}=0.2\,s$. }\label{order_q_solit}
\end{figure}
\subsection{Solitary wave propagation}
We consider now the time evolution of a solitary wave profile define as follows:
\begin{equation}\label{Soliton}
\left \lbrace \begin{array}{ll}
h(x,t) = h_{0} + \varepsilon h_0 \,sech^{2}\left( \kappa (x - ct) \right) \, ,\\
u(x,t) = c \left( 1-\dfrac{h_{0}}{h(x,t)} \right) \, ,\end{array} \right.
\end{equation}
with $\kappa=\sqrt{\dfrac{3\varepsilon}{4h_{0}^{2}(1+\varepsilon)}}$ \, , \, and $c=\sqrt{gh_{0}(1+\varepsilon)}$. Note that if such profiles are exact solutions of the original Green-Naghdi equations \eqref{eq6}, or equivalently \eqref{eq6imp} with $\alpha=1$, these are only solutions of the $\left(\mathcal{CG}_\alpha\right)$ model up to $\mathcal{O}(\mu^{2})$ terms. However, for small enough values of $\varepsilon$, such profiles are expected to propagate over flat bottoms without noticeable deformations. We use a rectangular computational domain of $200\,m$ length and $25\,m$ width, and a relatively coarse unstructured mesh with faces' length ranging from $0.7\,m$ to $1.5\,m$. The reference water depth is set to $h_{0} = 1 m$, the relative amplitude is set to $\varepsilon = 0.2$ and the initial free surface, shown on Fig. \ref{soliton:init}, is centred at $x_0 = 50\, m$. Cross sections along the x-direction centerline, obtained with $k=3$ at several times during the propagation, are shown on Fig. \ref{Sol_fig2}.\\
To further investigate the convergence properties of our approach, we now consider a sequence of regular meshes with mesh size ranging from $2^{-4}\,m$ to $2^{-9}\,m$ and polynomial expansions of degrees ranging from $k=1$ to $k=4$. The corresponding $L^{2}$ errors computed at $t_{\textrm{max}}=0.2\,s$ are used to plot the convergence curves on Fig. \ref{order_h_solit} for the free surface elevation and Fig. \ref{order_q_solit} for the discharge. The corresponding convergence rates obtained by linear regression are also reported and we observe an irregular behavior with rates varying between $\mathcal{O}({\mathfrak{h}}^k)$ and $\mathcal{O}({\mathfrak{h}}^{k+1})$, generally close to $\mathcal{O}({\mathfrak{h}}^\frac{1}{2})$ but with a sub-optimal convergence observed on the free surface for $k=4$. Such a behavior is also observed with slightly larger amplitude waves.
\begin{figure}
\begin{center}
\includegraphics[scale=0.4,angle=0]{Soliton.eps}
\caption{Test 2 - Solitary wave propagation: cross section of the free surface elevation at several times during the propagation.}
\label{Sol_fig2}
\end{center}
\end{figure}
\subsection{Run-up of a solitary wave}
We investigate now the ability of the scheme in handling dry areas with a test based on the experiments of \textit{Synolakis} \cite{Synolakis:1987p4276}. We study the propagation, shoaling, breaking and run-up of a solitary wave over a topography with constant slope $s=1/19.85$. The reference water depth is set to $h_0 = 1\,m$, and a solitary wave profile is considered as initial condition (\ref{Soliton}), with a relative amplitude $\varepsilon=0.28$. The simulation involves a $50m \times 5m$ basin, regularly meshed with a space step $\Delta x = 0.25\,m$ and $k=2$. We show on Fig. \ref{Synolakis_1} some cross sections of the solution, taken at various times during the propagation and compared with the experimental data.
\begin{figure}
\begin{center}
\begin{fig}
\includegraphics[scale=0.65,angle=0]{Synolakis.eps}
\end{fig}
\caption{Test 3 - Solitary wave breaking over a sloping beach: free surface profiles comparison
between numerical results (solid lines) and experimental data (crosses) at several times during the propagation ($t^{\ast} = t(g/h_0)^{1/2}$).}
\label{Synolakis_1}
\end{center}
\end{figure}
The wave breaking is identified approximately at $t^{\ast} = 17\,s$, with the normalized time $t^{\ast} = t(g/h_0)^{1/2}$, and occurs between gauges $\# 2$ and $\# 3$, which is in agreement with the experiment. The whole breaking process is well reproduced, as well as the subsequent run-up phenomena.
\subsection{Propagation of highly dispersive waves}
The dispersive properties of the numerical model are now assessed through the study of the propagation of periodic waves over a submerged bar, following the experiments of \textit{Dingemans} \cite{Dingemans:1994aa}. The computational domain is a $37.7m$ long and $0.8m$ wide basin. The topography is shown on Fig. \ref{Dingemans1}. The trapezoidal bar extends from $x=10\,m$ to $x=15\,m$ with slopes of $1/20$ at the front and $1/10$ at the back. The initial state is a flow at rest with a reference water depth of $h_0=0.4\,m$.
\begin{figure}
\begin{center}
\begin{fig}
\includegraphics[scale=0.3,angle=0]{Dingemans_1dconfig.eps}
\end{fig}
\caption{Test 4 - Propagation of highly dispersive waves: sketch of the experiment configuration and location of the gauges used for test ${\#}A$.}
\label{Dingemans1}
\end{center}
\end{figure}
Periodic waves are generated at the left boundary, with an amplitude $a$, and a period $T$. Both generating and absorbing layers are set to $5\,m$ at the corresponding boundaries. The two following tests are carried out:
\begin{itemize}
\item[${\#}A$ :] a=0.01\,m \quad , \quad T = 2.02\,s \quad , \quad \text{no wave - breaking}.
\item[${\#}B$ :] a=0.025\,m \quad , \quad T = 2.51\,s \quad , \quad \text{wave - breaking}.
\end{itemize}
For both test ${\#}A$ and ${\#}B$, we set $k=2$ and use a regular triangulation obtained from rectangular elements $\Delta x = \Delta y = 0.125\,m$. The characteristics of the flow are quite complex here, notably due to the high non-linearities induced by the topography. The propagating waves first shoal and steepen over the submerged bar, generating higher-harmonics. These harmonics are progressively released on the downward slope, until encountering deeper waters again. In the first test, the initial amplitude is not large enough to trig the breaking of the waves and the Green-Naghdi equations are resolved in the whole domain. Fig. \ref{Dingemans1} indicates the location of the five wave gauges used for the first test to study the time-evolution of the free surface deformations. The results are plotted on Fig. \ref{Dingemans2}. We observe a very good agreement between analytical and experimental data for the first wave gages. As usual with this set-up, some discrepancies can be observed on the two last gages. Such behavior can be improved with the use of optimized models. In particular, the extension of the present DG approach to the \textit{3-parameters} model of \cite{lannes_marche:2014} is left for future works.
\begin{figure}
\begin{center}
\includegraphics[scale=0.4,angle=0]{Dingemans.eps}
\caption{Test 4 - Propagation of highly dispersive waves: test ${\#}A$, free surface evolution at gauges. Numerical data are denoted in plain lines.}
\label{Dingemans2}
\end{center}
\end{figure}
The test ${\#}B$ corresponds to the experiments of \cite{BejiBattjes:1994aa}. The first reference gage is located at $x=6\,m$, followed by a series of seven gages regularly spaced between $x=11\,m$ and $x=17\,m$. Wave breaking is observed at the level of the flat part of the bump during the propagation and to stabilize the computation, the dispersive terms are turned-off in the vicinity of troubled cells, using the strategy described in \S\ref{Limitation}. Again, a comparison between the numerical results and the experimental data is performed and shown on Fig. \ref{Beji2}
\begin{figure}
\begin{center}
\begin{fig}
\includegraphics[scale=0.35,angle=0]{Beji.eps}
\end{fig}
\caption{Test 4 - Propagation of highly dispersive waves: test ${\#}B$, comparison with experimental data at gauges 1 to 8.}
\label{Beji2}
\end{center}
\end{figure}
\subsection{Periodic waves propagation over an elliptic shoal}
We study now the propagation of a train of monochromatic waves evolving over a varying topography, following the experiment of \textit{Berkhoff} \textit{et al} \cite{berkhoff:1982}. The experimental domain consists of a $20\,m$ wide and $22\,m$ long wave tank. The experimental model reproduces a seabed with a constant slope, forming an angle of $\alpha = 20^\circ$ with the $y$ axis, and deformed by a shoal with an elliptic shape, see Fig. \ref{Elliptic1}. The analytical profile for the topography in rotated coordinates $x_r = x cos(\alpha) - y sin(\alpha) , y_r = x sin(\alpha) + y cos(\alpha)$ is given by $z = z_b + z_s$ where:
\begin{equation}
\begin{split}
z_b(x,y)&=\left\lbrace
\begin{array}{ll}
(5.82 + x_r)/50 \quad \text{if} \quad x_r \geq -5.82 \\
0 \quad \text{elsewhere}
\end{array} \right. \\
z_s(x,y)&=\left\lbrace
\begin{array}{ll}
-0.3 + 0.5 \sqrt{1 - \left( \dfrac{x_r}{3.75}\right)^{2} - \left( \dfrac{y_r}{5}\right)^{2} } \quad \text{if} \quad \left( \dfrac{x_r}{4}\right)^{2} + \left( \dfrac{y_r}{3}\right)^{2} \leq 1\\
0 \quad \text{elsewhere}
\end{array} \right.
\end{split}
\end{equation}
The reference water depth is set to $h_0 = 0.45\,m$, and the corresponding computational domain has dimensions $[-10,12] \times [-10,10]$ (in $m$), with an extension of $5\,m$ at inlet and outlet boundaries for the generation of incident waves and their absorption. The periodic wave train has an amplitude of $a=0.0232\,m$ and a period of $T=1\,s$. Solid wall boundary conditions are used at $y=10\,m$ and $y=-10\,m$. We use an unstructured mesh of $25390$ elements which is refined in the region where the bottom variations are expected to have the greatest impact on the waves transformations (the smallest and largest mesh face's lengths are respectively $0.2\,m$ and $0.5\,m$). The computations are performed with polynomial approximations of order $k=2$.
\begin{figure}
\begin{center}
\includegraphics[scale=0.25,angle=0]{setup2.eps}
\caption{Test 5 - Periodic waves propagation over an elliptic shoal: topography and view of the free surface.}
\label{Elliptic1}
\end{center}
\end{figure}
The data issued from the experiment provides the normalized time-averaged of the free surface measured along several cross sections. In our numerical experiment we focus on the following ones:
\begin{equation}
\begin{split}
&\text{Section 2 :} \quad x = 3\,m \quad , \quad -5\,m \leq y \leq 5\,m \, ,\\
&\text{Section 3 :} \quad x = 5\,m \quad , \quad -5\,m \leq y \leq 5\,m \, ,\\
&\text{Section 5 :} \quad x = 9\,m \quad , \quad -5\,m \leq y \leq 5\,m \, ,\\
&\text{Section 7 :} \quad y = 0\,m \quad , \quad \quad 0\,m \leq x \leq 10\,m \, ,
\end{split}
\end{equation}
allowing a good coverage of the computational domain. Time series of the free surface elevation are hence recorded along these sections, from $t=30\,s$ to $t=50\,s$ and post-processed by mean of the \textit{zero up-crossing} method to isolate single waves and compute the mean wave elevation. The results, classically normalized by the incoming wave amplitude $a$, are shown in Fig. \ref{Elliptic2} and compared with the data.
\begin{figure}
\begin{center}
\begin{fig}
\includegraphics[scale=0.4,angle=0]{Elliptic_shoal.eps}
\end{fig}
\caption{Test 5 - Periodic waves propagation over an elliptic shoal: comparison with experimental data along the four sections.}
\label{Elliptic2}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{fig}
\includegraphics[scale=0.2,angle=0]{topo3.eps} \,
\end{fig}
\caption{Test 6 - Solitary wave propagation over a 3d reef: overview of the mesh in the refined area.}
\label{topo2}
\end{center}
\end{figure}
\subsection{Solitary wave propagation over a 3d reef}
The following test case is based on the experiments of \cite{Swigler:2009aa} and allows to study some complex wave's interactions such as shoaling, refraction, reflection, diffraction, breaking and moving shoreline in a fully two-dimensional context. The experimental domain is a $48.8\;m$ large and $26.5\,m$ wide basin. The topography is a triangular shaped shelf with an island feature located at the offshore point of the shelf. The island is a cone of $6\,m$ diameter and $0.45\,m$ height is also placed on the apex, between $x=14\,m$ and $x=20\,m$ (see Fig. \ref{topo1}). During the experiments, free surface information was recorded via $9$ wave gauges, and the velocity information was recorded with $3$ Acoustic Doppler Velocimeters (ADVs). The complete set up can be found in \cite{Swigler:2009aa} but we also specify the gauge locations in the legend of Fig. \ref{t5gauge_eta}.\\
\begin{figure}[H]
\begin{center}
\begin{fig}
\includegraphics[scale=0.2,angle=0]{topo1.eps} \, \includegraphics[scale=0.2,angle=0]{topo2.eps}
\end{fig}
\caption{Test 6 - Solitary wave propagation over a 3d reef: 3d view of the topography and initial condition.}
\label{topo1}
\end{center}
\end{figure}
Such a geometry motivates the use of an unstructured mesh of $45000$ elements refined in the vicinity of the cone, leading to a smallest face's length of $0.12\,m$ and a largest of $0.51\,m$, see Fig. \ref{topo2}. We set the polynomial order to $k=2$. The reference water depth is $h=0.78\,m$ and we study the propagation and transformations of a solitary wave of relative amplitude $\varepsilon = 0.5$. In addition to the complex processes already stated above, the transformations are quite nonlinear, making the following test particularly interesting. We show on Fig.\ref{t5AC3d1}-\ref{t5AC3d2} some snapshots of the free surface evolution before and after crossing the shelf's apex. We observe that the wave breaks at the apex slightly before $t = 5\,s$, wrapping the cone around $t=6.5\,s$. We also observe refracted waves from the reef slopes and diffracted waves around the cone which converge at the rear side at approximatively $t=8.5\,s$. The snapshots at $t=11.5\,s$ shows the run-up on the beach. Note that according with the positivity preservation property shown in the previous section, the simulation remains stable even with the use of a third order scheme.
We also provide a comparison between the numerical results and the data taken from the experiments for both free surface evolutions at the wave gages locations on Fig. \ref{t5gauge_eta} and the velocity at the ADVs locations, see Fig.\ref{t5gauge_u}. We highlight that even with this relatively moderate level of refinement, the characteristics of the flow are well resolved, especially when compared with those provided for instance in recent works using more refined meshes, see for instance\cite{Kazolea:2014kx, Roeber20121, shi:2012}.
\begin{figure}[H]
\begin{center}
\begin{fig}
\includegraphics[scale=0.18,angle=0]{t5AC_B.eps} \includegraphics[scale=0.18,angle=0]{t5AC_C.eps}\\
\includegraphics[scale=0.18,angle=0]{t5AC_D.eps} \includegraphics[scale=0.18,angle=0]{t5AC_E.eps}\\
\end{fig}
\caption{Test 6 - Solitary wave propagation over a 3d reef - free surface at times $t=4.5, 6.5, 8.5$ and $9.5\,s$.}
\label{t5AC3d1}
\end{center}
\end{figure}
\begin{figure}[H]
\begin{center}
\begin{fig}
\includegraphics[scale=0.18,angle=0]{t5AC_F.eps} \includegraphics[scale=0.18,angle=0]{t5AC_G.eps}\\
\includegraphics[scale=0.18,angle=0]{t5AC_H.eps} \includegraphics[scale=0.18,angle=0]{t5AC_I.eps}
\end{fig}
\caption{Test 6 - Solitary wave propagation over a 3d reef - free surface at times $t=10.5, 11.5, 13$ and $14\,s$.}
\label{t5AC3d2}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{fig}
\includegraphics[scale=0.35,angle=0]{3DreefsB.eps}
\end{fig}
\caption{Test 6 - Solitary wave propagation over a 3d reef - Comparison with experimental solution at gauges 1 to 9.}
\label{t5gauge_eta}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\begin{fig}
\includegraphics[scale=0.35,angle=0]{3DreefsBu.eps}
\end{fig}
\caption{Test 6 - Solitary wave propagation over a 3d reef: $x$ and $y$ velocity components - Comparison with experimental solution at several ADVs.}
\label{t5gauge_u}
\end{center}
\end{figure}
\section{Conclusion}
In this work, we introduce for the first time a fully discontinuous Galerkin formulation designed to approximate the solutions of a multi-dimensional Green-Naghdi model on arbitrary unstructured simplicial meshes. The underlying elliptic-hyperbolic decoupling approach, already investigated in \cite{DuranMarche:2014ab} for the $d=1$ case, allows to regard the dispersive correction simply as an algebraic source term in the NSW equations. Such dispersive correction is obtained as the solution of \textit{scalar} elliptic second order sub-problems, which are approximated using a mixed formulation and \textit{LDG} fluxes. Additionally, the preservation of motionless steady states and of the positivity of the water height, under a suitable time step restriction, are ensured for the whole formulation and for any order of approximation.\\
These properties are assessed through several benchmarks, involving nonlinear wave transformations and severe occurrence of dry areas, and some additional convergence studies are performed. Although the obtained numerical results clearly show very promising abilities for the study of nearshore flows, there are still several issues that need to be addressed in future works. In particular, we use a low-brow approach to stabilize the computations in the vicinity of broken waves. If such a method provides results which are in good agreement with experimental data for the cases under study in this work, we are still working to extend the method of \cite{tissier2} to the $d=2$ case in a robust way.\\
Another important issue may be related to the computational cost of our approach. If we believe in the potential benefits of the use of such a non-conforming high approach, especially concerning the robust approximations of solutions involving strong gradients occurring in the vicinity of breaking waves and possibly strong topography variations, it clearly leads to larger algebraic systems than those obtained with classical continuous approximations. Future works will therefore be devoted to go further in the development and optimization of our approach, and in particular to the reduction of the associated computational cost.
\paragraph{Acknowledgements}
The second author acknowledges partial support from the ANR-13-BS01-0009-01 BOND.
{%
\footnotesize
|
\section{Introduction}
Magnons are the collective excitations of ordered quantum magnets such as ferromagnets or antiferromagnets. In quantum magnetic systems that lack inversion symmetry, the DMI (spin-orbit coupling)\cite{dm} is present and leads to chiral magnons with nontrivial topological properties \cite{alex0,alex1,zhh,shin,shin1, alex1a} similar to electronic systems \cite{yu6,yu7,fdm}. As magnons are uncharged (neutral) quasiparticles, they do not experience a magnetic field in the Lorentz force as in the electronic version of Hall effect. Instead, they exhibit a thermal version of Hall effect in which a temperature gradient $\boldsymbol \nabla T$ transports a heat current $\bold J_Q$ \cite{zhh, alex1, alex0, alex7, alex4, alex44}. The DMI generates a nonzero Berry curvature given by \cite{alex2} $\boldsymbol{\Omega}(\bold k)={\nabla}_{\bf k}\times \bold{A}(\bold k)$, where $\bold{A}(\bold k)$ is a DMI dependent vector potential. The Berry curvature acts as an effective magnetic field by altering the propagation of magnons in the system, thus leads to thermal Hall effect dubbed magnon Hall effect \cite{alex1, alex0}, as well as magnon spin Nernst effect \cite{alex7}. These two phenomena are characterized by two conductivities --- transverse thermal conductivity $\kappa_{xy}$ and transverse spin Nernst conductivity $\alpha_{xy}^s$. They are both directly related to the Berry curvature of the magnon bulk bands reminiscent of Hall conductivity in electronic systems \cite{thou}. However, in contrast to electronic systems, there is no completely filled bands in bosonic systems, so each magnon band contributes a term to $\kappa_{xy}$ and $\alpha_{xy}^s$, and the Chern number of the system simply leads to protected chiral magnon edge states.
The first experimental realization of magnon Hall effect has been reported in three-dimensional (3D) pyrochlore ferromagnetic insulators Lu$_2$V$_2$O$_7$, Ho$_2$V$_2$O$_7$, and In$_2$Mn$_2$O$_7$ \cite{alex1,alex1a}. Quite recently, magnon Hall effect has been observed in 2D kagome magnet Cu(1-3, bdc) \cite{alex6, alex6a}. Both theory and experiment show that the topology of the system leads to a sign change in $\kappa_{xy}$ as a function of temperature or magnetic field on the kagome lattice \cite{alex4,alex6, alex6a} and a sign change in $\kappa_{xy}$ as the magnetic field is reversed on the pyrochlore lattice \cite{alex1, alex1a}. These experimental observations have propelled a possibility of numerous experimentally accessible 2D ferromagnets in different lattices that exhibit nontrivial topological spin excitations. In this regard, magnon Hall transports have been proposed on a single-layer Lieb ferromagnet \cite{xc} with three magnon bulk bands, in which a sign change was also observed in $\kappa_{xy}$. The honeycomb ferromagnet is special as it applies to experimentally accessible graphene sheet \cite{yu6,yu7, cas}. It also forms an example of nontrivial topological energy bands in electronic systems \cite{fdm}. In the context of magnon Hall transports, the author has recently shown that nontrivial magnon bands and magnon Hall effect could be accessible in a single-layer honeycomb quantum ferromagnets, in which a DMI is allowed by the alternating triangular plaquettes of the next-nearest-neighbour sites \cite{sol, sol1}. For the honeycomb lattice, $\kappa_{xy}$ shows no sign change for all parameters of the system \cite{sol1}. Recent study has shown that the proposed honeycomb ferromagnetic system also exhibits spin Nernst effect \cite{kkim}. It has been previously shown that an interfacial contact between two magnon insulators can lead to nontrivial transport properties on the kagome ferromagnet \cite{mook}.
Motivated by these theoretical and experimental realizations of thermal Hall effects of spin excitations and the experimental realizations of bilayer honeycomb-lattice quantum magnetic compounds such as Na$_3$Cu$_2$SbO$_6$ \cite{aat1}, $\beta$-Cu$_2$V$_2$O$_7$ \cite{aat}, and A$_2$IrO$_3$ (A= Na, Li) \cite{aat0,aat00}, we study thermal Hall transports of magnons in bilayer honeycomb magnetic systems. Generalization to other bilayer systems is straightforward. We show that two layers of magnon insulators with the same DMI can be coupled either ferromagnetically or antiferromagnetically. In the former, we show that the presence of an alternating next-nearest-neighbour DMI generates Berry curvatures and dissipationless magnon edge states, which propagate in the same direction on the top and the bottom layers of the bilayer system. We compute the thermal Hall and spin Nernst conductivities.
For the latter case, the spins on the upper\slash lower layer point in opposite direction to those on the lower\slash upper layer, hence the interlayer couplings become antiferromagnetic. In this case, we show that magnon edge states propagate in opposite directions. As a result, thermal Hall and spin Nernst conductivities also show a change sign as the magnetic field is reversed. This result is interesting because it is what is seen experimentally in many ferromagnetic insulators \cite{alex1,alex1a}. As thermal Hall transports of magnons await experimental observation on the honeycomb lattice, our results open new avenue to search for 2D and 3D bilayer systems with nonzero thermal Hall transports.
\section{Bilayer magnon insulator}
Bilayer magnon insulators consist of two single-layer of magnon insulators coupled either ferromagnetically or antiferromagnetically as shown in Fig.~\ref{lattice1}. The Hamiltonian is governed by \begin{align}
H&= H_{FM,\tau}+ H_{DMI,\tau}+H_{ext,\tau}+H_{int.},
\label{h}
\end{align}
where, $H_{\tau}$ represents the single-layer Hamiltonians for the top $\tau=T$ and bottom $\tau=B$ layers respectively, $H_{DMI,\tau}$ represents the DMI interactions on both layers, $H_{ext,\tau}$ is the external magnetic field on each layer, and $H_{int}$ is the interlayer couplings between them. For the honeycomb lattice, they are given by
\begin{align}
&H_{FM,\tau}=-J\sum_{\langle i, j\rangle}{\bf S}_{i}^\tau\cdot{\bf S}_{j}^\tau\\ &H_{DMI,\tau}=D\sum_{\langle \langle i,j\rangle\ra} \nu_{ij}\bold{\hat z}\cdot{\bf S}_{i}^\tau\times{\bf S}_{j}^\tau,\label{model}\\ &H_{ext, \tau}=-h\sum_{i}S_{i,z}^\tau,\\& H_{int.}=-\sum_{ i\in T, j\in B; \alpha}J_{\alpha}{\bf S}_{i}\cdot{\bf S}_{j},
\label{h2}
\end{align}
\begin{figure}[ht]
\centering
\includegraphics[width=.75\linewidth]{unit_copy}
\caption{Color online. Lattice structure of ferromagnetically coupled bilayer honeycomb ferromagnet with various hopping parameters. For antiferromagnetically coupled layers, the spins on the upper layer point in opposite direction to those on the lower layer.}
\label{lattice1}
\end{figure}
where ${\bf S}_{i}$ is the spin moment at site $i$, $J>0$ is a nearest-neighbour (NN) ferromagnetic interaction on each layer, and $D$ is the magnitude of the DMI which is allowed by the next-nearest-neighbour (NNN) triangular plaquettes on the honeycomb lattice, where $\nu_{ij}=\pm 1$ represents hopping from left to right and vice versa on the NNN sites. The Zeeman magnetic field is $h$ in units of $g\mu_B$. The interlayer NN interactions $J_{\alpha}>0$ denotes all the possible interlayer couplings depicted in Fig.~\ref{lattice1}.
\section{Magnon bands}
In this section, we present the band structure of the bilayer honeycomb magnetic system using the linearized HP boson representation of the spin operators \cite{HP}, which is valid at low-temperature when few magnons are thermally excited. In fact, this formalism has been employed frequently in the study of thermal Hall transports of magnons in ferromagnetic systems \cite{zhh, alex0, alex1, alex2, alex4}. It is also supported in recent experimental realizations \cite{alex6, alex6a, alex1}.
\subsection{Ferromagnetically coupled layers}
We first start with ferromagnetically coupled layers, $J_\alpha >0$. Using the linearized HP transformation, the magnon tight binding hopping model is given by
\begin{align}
H_{\tau}&=v_s^\prime\sum_{i} b_{i}^\dagger b_{i} -v_s\sum_{\langle ij\rangle}b_{i}^\dagger b_{j} - v_D\sum_{\langle \langle ij\rangle\ra}i\nu_{ij}b_{i}^\dagger b_{j},\label{hp3}\\
H_{int.}&= \sum_{i\in T, j\in B; \alpha}v_{\alpha}[( b_{i}^\dagger b_{i}+ b_{j}^\dagger b_{j}) -( b_{i}^\dagger b_{j}+ b_{j}^\dagger b_{i})],
\label{hpp3}
\end{align}
where $v_s^\prime=zv_s + h$, $v_s(v_D)= JS(DS)$, and $z=3$ is coordination number of the lattice. The interlayer coupling is $v_{\alpha}=J_{\alpha}S$.
Similar to AB-stacked bilayer graphene\cite{mcc, mcc1,mcc2,mcc3} there are three kinds of interlayer couplings as shown in Fig.~\ref{lattice1}. We identify the bottom layer with two sublattices labeled $A_1$ and $B_1$, and the top layer with $A_2$ and $B_2$. The exchange interaction $J_0$ couples $B_1$ and $A_2$, $J_1$ couples $A_1$ and $B_2$, and $J_2$ couples $A_1$ and $A_2$, $B_1$ and $B_2$. In Fourier space the momentum space Hamiltonian is given by $ H=\sum_{\bold k}\psi^\dagger_{\bold k}\cdot \mathcal{H}_{FM}(\bold k)\cdot\psi_{\bold k},$ with the basis $\psi^\dagger_{\bold k}= (b_{\bold{k}A_1}^{\dagger},\thinspace b_{\bold{k} B_1}^{\dagger},b_{\bold{k} A_2}^{\dagger},\thinspace b_{\bold{k} B_2}^{\dagger})$, where the Bogoliubov Hamiltonian is given by
\begin{figure}[ht]
\centering
\includegraphics[width=4in]{Bands}
\caption{Color online. The magnon bulk bands of the spin-$1/2$ bilayer ferromagnet along $k_y=0$ at $v_s=0.5,~h=0.1,v_0=0.25,~v_1=v_2=0$: $(a)$ $v_{D}=0.0$ $(b)$ $v_{D}=0.05$.}
\label{bands}
\end{figure}
\begin{align}
\mathcal{H}_{FM}(\bold k)=\left(
\begin{array}{cc}
\mathscr A_1(\bold{k})& \mathscr B(\bold{k})\\
\mathscr B^\dagger(\bold{k})& \mathscr A_2(\bold{k})
\end{array}
\right),
\label{honn}
\end{align}
where
\begin{align}
\mathscr A_1(\bold k)=\left(
\begin{array}{cc}\epsilon_{A1}+m_{\bold{k}}&-v_sf_{\bold{k}}\\
-v_sf_{\bold{k}}^*&\epsilon_{B1} -m_{\bold{k}}\\
\end{array}
\right),
\label{a1}
\end{align}
\begin{align}
\mathscr A_2(\bold k)=\left(
\begin{array}{cc}\epsilon_{A2}+m_{\bold{k}}&-v_sf_{\bold{k}}\\
-v_sf_{\bold{k}}^*&\epsilon_{B2} -m_{\bold{k}}\\
\end{array}
\right),
\label{a2}
\end{align}
\begin{align}
\mathscr B(\bold k)=\left(
\begin{array}{cc}
-v_2f_{\bold{k}}&-v_1f_{\bold{k}} \\
-v_0&-v_2f_{\bold{k}}\\
\end{array}
\right),
\label{bb}
\end{align}
where $\epsilon_{A1}=\epsilon_{B2}=v_s^\prime+z(v_1+v_2)$, and $\epsilon_{B1}=\epsilon_{A2}=v_s^\prime+v_0+zv_2$. The lattice factor is
$f_{\bold{k}}= e^{ik_ya/2}\left( 2\cos\sqrt{3}k_xa/2+e^{-3ik_ya/2}\right),$
and the mass is $m_\bold{k}= 4v_D \sin \frac{\sqrt{3}k_x}{2}\left( \cos \frac{\sqrt{3}k_x}{2}-\cos \frac{{3}k_y}{2}\right)$, with $m_\bold{k}=-m_{-\bold{k}}$. Notice that the shifting parameters $\epsilon_{A1} $ and $\epsilon_{B1}$ are the major differences between the bilayer magnon insulator and that of spin orbit coupled AB-stacked bilayer graphene \cite{mak55}. As a result for $v_1=v_2=0$, the eigenvalues of the bilayer magnon insulator yields
\begin{align}
&\epsilon_{\pm}^{(1)}=v_s^\prime \pm \sqrt{m_\bold{k}^2 +|v_sf_\bold{k}|^2},\\
&\epsilon_{\pm}^{(2)}=v_s^\prime+v_0 \pm \sqrt{m_\bold{k}^2 +v_0^2 +|v_sf_\bold{k}|^2}.
\label{eig}
\end{align}
At the Dirac points ${\bf K}_\pm= (\pm 4\pi/3\sqrt{3}a, 0)$, $f_\bold{k}=0$ and $m_\bold{k}=m=3\sqrt{3}v_D$. The lowest band $\epsilon_{-}^{(1)}$ has a Goldstone mode at ${\bf \Gamma}=(0,0)$ when $h=0$. The evolution of the magnon bands are shown in Fig.~\ref{bands}. For $v_D=v_{1,2}=0$ there are three gapless bands at ${\bf K}_\pm$ as opposed to AB-stacked bilayer graphene \cite{mcc, mcc1,mcc2,mcc3}.
\subsection{Antiferromagnetically coupled layers}
The magnon insulator can also be coupled antiferromagnetically. In this case, the spins on the upper layer are designed to point in opposite direction to those on the lower layer, hence the interlayer couplings become antiferromagnetic with $J_\alpha<0$. To capture the correct magnetic excitations, we perform a $\pi$-rotation about the $S_x$-axis on the upper layer, that is $S_{i,j}^x\to S_{i,j}^x$, $S_{i,j}^y\to -S_{i,j}^y$, $S_{i,j}^z\to -S_{i,j}^z$ (see Appendix~\ref{HHPP}). This rotation does not change the ferromagnetic nature of the upper layer $H_{FM,T}$, but it rotates the spins along the new $z$-axis and changes the sign of upper-layer out-of-plane DM interaction, $H_{DMI,T}$ and the upper-layer external magnetic field $H_{ext, T}$. In the HP bosonic representation, the interlayer coupling has the form
\begin{align}
H_{int.}^{AFM}= \sum_{i\in T, j\in B; \alpha}|v_{\alpha}|[( b_{i}^\dagger b_{i}+ b_{j}^\dagger b_{j}) +( b_{i}^\dagger b_{j}^\dagger+ b_{j} b_{i})].
\end{align}
Introducing the Nambu operators $\Psi^\dagger_{\bold k}= (\psi^\dagger_{\bold k},\psi_{-\bold k})$, the momentum space of the total Hamiltonian can be written as $ H=\frac{1}{2}\sum_{\bold k}\Psi^\dagger_{\bold k}\cdot \mathcal{H}_{AFM}(\bold k)\cdot\Psi_{\bold k}+\text{const.}\label{ham1},$ where $\mathcal{H}_{AFM}(\bold k)$ is a $2N \times 2N$ ( $N$ is the number of sublattices) matrix given by
\begin{align}
\mathcal{H}_{AFM}(\bold k)=\left(
\begin{array}{cc}
\mathcal A(\bold{k})& \mathcal B(\bold{k})\\
\mathcal B^*(-\bold{k})& \mathcal A^*(-\bold{k})
\end{array}
\right).
\label{eqn9}
\end{align}
The matrices $\mathcal A(\bold{k})$ and $\mathcal B(\bold{k})$ are given by
\begin{align}
\mathcal{A}(\bold k)=
\begin{pmatrix}
\mathscr A_1^\prime(\bold{k})& 0\\
0&\mathscr A_2^{\prime *}(-\bold{k})
\end{pmatrix}.
\label{honn1}
\end{align}
\begin{align}
\mathcal{B}(\bold k)=
\begin{pmatrix}
0 &\mathscr B^\prime(\bold{k})\\
\mathscr B^{\dagger\prime}(\bold{k})& 0\\
\end{pmatrix}.
\label{honn1}
\end{align}
where $ \mathscr A_1^\prime(\bold{k})= \mathscr A_1(\bold{k})$ and $ \mathscr A_2^\prime(\bold{k})= \mathscr A_2(\bold{k})$ with $\epsilon_{A1}=zv_s+h+z(|v_1|+|v_2|)$, $\epsilon_{B1}=zv_s+h+|v_0|+z|v_2|$, $\epsilon_{A2}=zv_s-h+|v_0|+z|v_2|$, $\epsilon_{B2}=zv_s-h+z(|v_1|+|v_2|)$. Also $\mathscr B^\prime(\bold{k})=\mathscr B(\bold{k})$ with $-v_\alpha\to |v_\alpha|$.
Due to the mass term $\mathcal A(\bold k)\neq \mathcal A^*(-\bold k)$, but $\mathcal B(\bold k)=\mathcal B^*(-\bold k)$.
The Hamiltonian is hermitian but not diagonal. It can be diagonalized by the generalized Bogoliubov transformation
$\Psi_\bold{k}= \mathcal{P}_\bold{k} Q_\bold{k}$,
where $\mathcal{P}_\bold{k}$ is a $2N\times 2N$ matrix and $Q^\dagger_\bold{k}= (\mathcal{Q}_\bold{k}^\dagger,\thinspace \mathcal{Q}_{-\bold{k}})$ with $ \mathcal{Q}_\bold{k}^\dagger=(\alpha_{\bold{k} A1}^{\dagger}\thinspace \alpha_{\bold{k} B1}^{\dagger}\thinspace \alpha_{\bold{k} A2}^{\dagger}\thinspace \alpha_{\bold{k} B2}^{\dagger})$ being the quasiparticle operators. The matrix $\mathcal{P}_\bold{k}$ satisfies the relations,
\begin{align}
&\mathcal{P}_\bold{k}^\dagger \mathcal{H}_{AFM}(\bold{k}) \mathcal{P}_\bold{k}= \epsilon(\bold{k}); \quad \mathcal{P}_\bold{k}^\dagger \eta \mathcal{P}_\bold{k}= \eta,
\label{eqna}
\end{align}
\begin{figure}[ht]
\centering
\includegraphics[width=3in]{BandsA}
\caption{Color online. Magnon bulk bands of spin-$1/2$ antiferromagnetically coupled layers along $k_y=0$. The parameters are the same as Fig.~\ref{bands} respectively.}
\label{bandsA}
\end{figure}
with
$\eta=
\text{diag}(
\mathbf{I}_{N\times N}, -\mathbf{I}_{N\times N} )$, and $\epsilon(\bold{k})$ is the diagonal matrix of the quasiparticle energy eigenvalues.
Using the fact that $\mathcal{P}_\bold{k}^\dagger=\eta \mathcal{P}_\bold{k}^{-1}\eta$, then Eq.~\ref{eqna} is equivalent to saying that we need to diagonalize a non-hermitian Hamiltonian
\begin{align}
\eta\mathcal{H}_{AFM}(\bold{k})=\left(
\begin{array}{cc}
\mathcal A(\bold{k})& \mathcal B(\bold{k})\\
-\mathcal B^*(-\bold{k})& -\mathcal A^*(-\bold{k})
\end{array}
\right),
\label{non}
\end{align}
whose eigenvalues are given by $\eta\epsilon(\bold{k})=[\epsilon_\mu(\bold{k}), -\epsilon_\mu(\bold{k})]$ and the columns of $\mathcal P_\bold{k}$ are the corresponding eigenvectors. We could not find an analytical expression in this case. Figure \ref{bandsA} shows the positive energy bands. At the Dirac points $\bold{K}_\pm$, the eigenvalues can be found analytically. They are given by $\epsilon_1=m+\epsilon_{A1}$, $\epsilon_2=m+\epsilon_{B2}$, $\epsilon_{3,4}= \pm h-m+\sqrt{[(\epsilon_{B1}+\epsilon_{A2})/2]^2-v_0^2}$. A careful calculation also shows that antiferromagnetically coupled bilayer ferromagnets behaves very similar to antiferromagnetically coupled bilayer antiferromagnets with N\'eel states, i.e., $J,J_\alpha<0$. For the most part of this paper, we will focus on the special limit $v_1=v_2=0$ for simplicity. In many cases of physical interest, the magnetic field can induce canting of spins in antiferromagnetic systems. This scenario is analyzed in Appendix B.
\section{Topological magnon transports}
\subsection{Magnon edge states}
We now study the main purpose of this paper --- thermal Hall transports of magnons. As the DMI plays the same role as spin-orbit coupling \cite{mak55}, topological effects in magnon insulators stem from nontrivial properties of the chirality-induced Berry curvatures. This directly implies the propagation of magnon edge states on the edges of each layer. Having diagonalized the Hamiltonians, the Berry curvature of the magnon bulk bands can be written as
\begin{eqnarray}
\Omega_{ij;\mu}(\bold k)=-2\sum_{\mu^\prime\neq \mu}\frac{\text{Im}[ \braket{\mathcal{P}_{\bold{k}\mu}|v_i|\mathcal{P}_{\bold{k}\mu^\prime}}\braket{\mathcal{P}_{\bold{k}\mu^\prime}|v_j|\mathcal{P}_{\bold{k}\mu}}]}{\left(\epsilon_{\bold{k}\mu}-\epsilon_{\bold{k}\mu^\prime}\right)^2},
\label{chern2}
\end{eqnarray}
where $v_{i}=\partial \mathcal{H}_{FM}(\bold k)/\partial k_{i}$ defines the velocity operators with $i,j=x,y$. The columns of $\mathcal{P}_{\bold{k}\mu}$ are the eigenvectors, and $\mu$ labels the bands. For antiferromagnetically coupled layers, $v_{i}=\partial [\eta\mathcal{H}_{AFM}(\bold k)]/\partial k_{i}$ and $\mathcal{P}_{\bold{k}\mu}$ is paraunitary. In the antiferromagnetic case the Berry curvature can be written alternatively as
\begin{align}
\Omega_{ij;\mu}(\bold{k})=-2\text{Im}[\eta\mathcal (\partial_{k_i}\mathcal P_{\bold{k}\mu}^\dagger)\eta(\partial_{k_j}\mathcal P_{\bold{k}\mu})]_{\mu\mu},
\label{bc1}
\end{align}
which can be used for systems with explicit analytical form of $\mathcal P_{\bold{k}\mu}$.
The Berry curvature for ferromagneticcally coupled layers is shown in Fig.~\ref{berry_F} and antiferromagneticcally coupled layers in Fig.~\ref{berry}. In the former, only the Berry curvatures corresponding to the bands $\epsilon_\pm^{(1)}$ are peaked at the Dirac points. This is due to the biased nature of the system. On the other hands, the latter Berry curvatures for each band is suppressed at three corners of the Brillouin zone [see the corresponding bands in Fig.~\ref{bandsA}(b)].
Although magnon Hall transports are induced by the Berry curvatures, the Chern numbers can still be defined for bosonic systems as the integration of the Berry curvature over the first Brillouin zone given by,
\begin{equation}
\mathcal{C}_\mu= \frac{1}{2\pi}\int_{{BZ}} dk_xdk_y~ \Omega_{xy; \mu}(\bold k).
\label{chenn}
\end{equation}
For ferromagnetically coupled layers, the Chern number of the bands $\epsilon_\pm^{(1)}$ and $\epsilon_\pm^{(2)}$ are given by $\mathcal{C}_\mu=[2,-2,0,0]$ and antiferromagnetically coupled layers, we find $\mathcal{C}_\mu=[2,-2,2,-2]$. We see that the sum of the Chern numbers is zero as can also be seen from the Berry curvatures. Similar to AB-stacked bilayer graphene \cite{mak55}, the Chern numbers are double that of single layer \cite{sol}. This implies that the edge states are also double that of single layer [Fig.~\ref{edge} $(i)$] as depicted in Figs.~\ref{edge} $(ii)$ and $(iii)$ and \ref{edgeA} $(i)-(iii)$. We also observe magnon edge states at zero DMI [Figs.~\ref{edge} and \ref{edgeA} $(iv)$], which is not protected by any Chern number similar to AB-stacked bilayer graphene \cite{castt}. The magnon edge states for ferromagnetically and antiferromagnetically coupled layers differ by the direction of propagation as shown schematically in Figs.~\ref{FMedge} and \ref{AFMedge}.
\begin{figure}[ht]
\centering
\subfigure[\label{berry_F}]{\includegraphics[width=.45\linewidth]{Berry_F}}
\subfigure[\label{berry}]{\includegraphics[width=.45\linewidth]{Berry_A1}}
\caption{Color online. Berry curvatures of spin-$1/2$ bilayer honeycomb quantum magnets for bands $\epsilon_{-}^{(1)}$ and $\epsilon_{-}^{(2)}$ (upper panel) and $\epsilon_{+}^{(1)}$ and $\epsilon_{+}^{(2)}$ (lower panel). $(a)$ Ferromagnetic coupling. $(b)$ Antiferromagnetic coupling. The parameters are the same as Figs.~\ref{bands}(b) and \ref{bandsA}(b) respectively. }
\label{berry_A}
\end{figure}
\begin{figure*}[!]
\centering
\subfigure[\label{edge}]{\includegraphics[width=.45\linewidth]{Edge}}
\quad
\subfigure[\label{edgeA}]{\includegraphics[width=.45\linewidth]{Edge_A}}
\caption{Color online. Magnon edge states of bilayer spin-$1/2$ honeycomb ferromagnet in the zigzag geometry for $v_1=v_2=0$ and $v_s=0.5,~h=0.1$. $(a)$ Ferromagnetic coupling, $(i)~v_D=0.1,~v_0=0;~(ii)~v_D=0.1, v_0=0.05; ~(iii)~v_D=0.1, v_0=0.25; ~(iv)~v_D=0.0, v_0=0.5$. $(b)$ Antiferromagnetic coupling, $(i)~v_D=0.1, v_0=0.05; ~(ii)~v_D=0.1, v_0=0.25;~(iii)~v_D=0.1, v_0=0.5; ~(iv)~v_D=0.0, v_0=0.5$. The bands in Figs.~\ref{bands}(b) and \ref{bandsA}(b) correspond to $a~(iii)$ and $b~(ii)$. }
\label{Edge}
\end{figure*}
\begin{figure}[ht]
\centering
\subfigure[\label{FMedge}]{\includegraphics[width=.45\linewidth]{FM}}
\quad
\subfigure[\label{AFMedge}]{\includegraphics[width=.45\linewidth]{AFM}}
\caption{Color online. Schematics of the propagation of chiral magnon edge states. $(a)$ Ferromagnetically coupled layers. $(b)$ Antiferromagnetically coupled layers. }
\end{figure}
\subsection{Magnon Hall and Spin Nernst effects}
Theoretically, thermal Hall effect of magnons is manifested due to the nontrivial topology of the magnon band structures induced by the DMI. The non-vanishing Berry curvatures induce an effective magnetic field in the system, upon the application of a longitudinal temperature gradient $-\boldsymbol \nabla T$. Figure~\ref{mags} shows the schematic representation of magnon Hall effect, in which a longitudinal temperature gradient induces a transverse heat current ${\bf J}_Q$. The propagation of magnon in the bilayer system is deflected by the DMI on the top and bottom layers. This leads to thermal Hall effect \cite{alex0} characterized by the transverse thermal Hall conductivity given by \cite{alex2,shin1}
$
\kappa_{xy}=-\frac{k_B^2 T}{\hbar V}\sum_{\bold{k}}\sum_{\mu=1}^N c_2[ g\left(\epsilon_{\bold{k} \mu}\right)]\Omega_{xy; \mu}(\bold k),$ where $V$ is the volume of the system, $k_B$ is the Boltzmann constant, $T$ is the temperature, $g(\epsilon_{\bold{k}\mu})=[e^{{\epsilon_{\bold{k}\mu}}/k_BT}-1]^{-1}$ is the Bose function, $c_2(x)=(1+x)\left( \ln \frac{1+x}{x}\right)^2-(\ln x)^2-2\text{Li}_2(-x),$ and $\text{Li}_2(x)$ is a dilogarithm.
\begin{figure}[!]
\centering
\subfigure[\label{mags}]{\includegraphics[width=.45\linewidth]{mag1}}
\quad
\subfigure[\label{magt}]{\includegraphics[width=.45\linewidth]{mag}}
\caption{Color online. Schematics of magnon Hall effect $(a)$ and spin Nernst effect $(b)$ in ferromagnetically coupled layers. In magnon Hall effect, the propagation of magnon is deflected by the DMI upon the application of temperature gradient $\nabla T$ and an induced heat current ${\bf J}_Q$; whereas for spin Nernst effect, opposite spins propagate in different directions. Notice that for antiferromagnetically coupled layers, the schematic representation of magnon Hall effect is equivalent to $(b)$, since the upper layer has an opposite DMI as shown in Eq.~\ref{eqn9}. }
\end{figure}
\begin{figure}[!]
\centering
\subfigure[\label{th_f}]{\includegraphics[width=.45\linewidth]{ThM_F}}
\quad
\subfigure[\label{th1_f}]{\includegraphics[width=.45\linewidth]{ThM_Fh.png}}
\quad
\subfigure[\label{sn_f}]{\includegraphics[width=.45\linewidth]{SN_F}}
\quad
\subfigure[\label{sn1_f}]{\includegraphics[width=.45\linewidth]{SN_Fh.png}}
\caption{Color online. Plot of $\kappa_{xy}$ and $\alpha_{xy}^s$ for spin-$1/2$ ferromagnetically coupled layers. $(a)$ $\kappa_{xy}$ vs. $T/J$, $(b)$ Contour plot of $\kappa_{xy}$ in $T/J$ and $h/J$ plane; $(c)$ $\alpha_{xy}^s$ vs. $T/J$; $(d)$ Contour plot of $\alpha_{xy}^s$ in $T/J$ and $h/J$ plane. The parameters for $(a)$ and $(c)$ are $k_B=\hbar=1$, $v_s=v_0=0.5$, $h=\pm 0.5$, $v_1=v_2=0$, and several values of $v_D$. For $(b)$ and $(d)$, $v_s=v_0=0.5; v_D=0.05$. }
\end{figure}
\begin{figure}
\centering
\subfigure[\label{th_a}]{\includegraphics[width=.45\linewidth]{ThM_A.png}}
\quad
\subfigure[\label{sn_a}]{\includegraphics[width=.45\linewidth]{SN_A.png}}
\caption{Color online. $(a)$ Contour plot of $\kappa_{xy}$ in $T/J$ and $h/J$ plane for spin-$1/2$ antiferromagnetically coupled layers. $(b)$ Contour plot of $\alpha_{xy}^s$ in $T/J$ and $h/J$ plane. The parameters are $k_B=J=\hbar=1$, $v_s=0.5~v_0=0.5,v_D=0.05;~v_1=v_2=0$. }
\end{figure}
In what follows, we focus on the simplified model $v_1=v_2=0$ with nonzero $v_0,v_D$. The trend of the thermal conductivity is shown in Fig.~\ref{th_f} for several values of $v_D$ and fixed $v_0=0.5$. At $T=0$, $\kappa_{xy}=0$ and decreases exponentially for large temperature approaching a constant value for very large temperature but never changes sign. Increasing the DMI, $v_D$, decreases $\kappa_{xy}$. The contour plot of $\kappa_{xy}$ in Fig.~\ref{th1_f} shows no sign change by reversing the magnetic field.
As previously mentioned, magnon edge state propagation also carries a transverse spin current which leads to spin Nernst effect depicted in Fig.~\ref{magt}. It can be understood as two copies of magnon Hall effects with opposite spins. It is characterized by a conductivity given by \cite{alex7}
$\alpha_{xy}^s={k_B}V^{-1}\sum_{\bold{k}\mu}c_1\left( n_\mu\right)\Omega_{xy; \mu}(\bold k),$
where $c_1(x)=(1+x)\ln(1+x)-x\ln x$. Due to the Berry curvature, $\alpha_{xy}^s$ has a similar feature as $\kappa_{xy}$ except that it is now positive as shown in Fig.~\ref{sn_f}. The spin Nernst conductivity also vanishes at zero temperature as there are no thermal excitations but the DMI, $v_D$ increases $\alpha_{xy}^s$. We see that $\alpha_{xy}^s$ does not show any sign change for all parameters considered as shown in Fig.~\ref{sn1_f}.
For antiferromagnetically coupled layers, the propagation of magnons are deflected in opposite directions by opposite signs of DMI on each layer (see Eq.~\ref{eqn9}). In addition, the Berry curvature changes sign as the magnetic field is reversed (not shown). Hence, the topology of the system is different from ferromagnetically coupled layers. This is manifested explicitly in the conductivities as shown in Figs.~\ref{th_a} and \ref{sn_a}. The sign change in $\kappa_{xy}$ and $\alpha_{xy}^s$ as the magnetic field is reversed is the manifestation of the band topology of the system for antiferromagnetically coupled layers. A different sign change has been studied on the kagome lattice. In this case, Ref.~[\onlinecite{alex44}] explains the sign change in $\kappa^{xy}$ as a consequence of the sign change in Berry curvature of the highest band and Ref.~[\onlinecite{alex4}] argues that the sign change in $\kappa^{xy}$ is a consequence of the propagation of the magnon edge states, however, with a NNN interaction. The origin of the sign change on the kagome ferromagnets is still not well-understood theoretically. In the present case with antiferromagnetic coupling, reversing the magnetic field flips the spin moments on both layers and thus changes the sign of the spin and bosonic operators in HP transformation (see Appendix~\ref{HHPP}). This has an important consequence on the sign of the DMI on each layer and affects the signs of the conductivities.
\section{Conclusion}
We have presented a detail study of magnon Hall transports in AB-stacked bilayer honeycomb magnon insulators. We show that the interlayer couplings between two ferromagnetic magnon insulators can be treated ferromagnetically or antiferromagnetically. For ferromagnetic coupling, we present explicit calculation based on magnon tight binding model obtained from linearized HP transformation. We show that the system behaves similarly to AB-stacked bilayer graphene and possesses nontrivial topological magnon bulk bands and Berry curvatures. However, for magnon insulators the propagation of edge states give rise to thermal Hall effect and spin Nernst effect. We show that for ferromagnetically coupled layers, the conductivities characterizing these effects are very similar to single-layer honeycomb magnon insulator with no sign change. For antiferromagnetically coupled layers, we show that the topological nature of the magnon bands is different. Magnon edge states propagate in opposite directions and the conductivities of thermal Hall response show a sign change as the magnetic field is reversed.
To the best of our knowledge, thermal Hall effect of magnons has been observed experimentally only in ferromagnetic insulators without inversion symmetry. The Lieb and honeycomb lattices still await experimental observation. In these systems, spin-orbit coupling (DMI) can be induced in many different ways. For the honeycomb lattice, it is also possible to utilize ultracold bosonic atoms trapped in optical lattice at finite but low temperatures, since the bosonic tight binding model is reminiscent of tight binding model in electronic systems, where the Haldane model has been realized experimentally in optical fermionic lattice \cite{jott}. The most important result of our study is that it opens new possibilities to search for bilayer magnon Hall transports in other lattices without inversion symmetry (such as the kagome and pyrochlore lattices), where nontrivial topological effects and magnon Hall effect have been realized experimentally. Our study also generalizes the study of magnon spintronics to bilayer systems. This model can also be studied using the spinon (Schwinger boson) representation. The Hamiltonian in this representation is an 8-band model, which can be regarded as two copies of Holstein-Primakoff Hamiltonians, one for each spin degree of freedom. The magnon bands and the associated magnon edge states are similar to that of gated AB-stacked bilayer graphene \cite{zhe}. Both representations give similar results, however recent experimental results on the kagome magnet Cu(1-3, bdc) confirms that the Holstein-Primakoff boson representation gives a better estimate than the spinon representation \cite{alex6a, alex6}.
\section*{Acknowledgments} Research at Perimeter Institute is supported by the Government of Canada through Industry Canada and by the Province of Ontario through the Ministry of Research
and Innovation.
|
\section{Introduction}
\IEEEPARstart{M}{etcalfe's} law relates to communications networks;
it states that the value of a network is proportional to square of
its size. Not long ago, a group of authors\cite{Briscoe06c:nlogn}
challenged quadratic dependence. That split community into believers
and deniers of Metcalfe's law, and their claim was recently challenged
by statistical data\cite{zhang2015tencent}. However no attempts were
made to establish the truth mathematically, perhaps due to difficulties
with obtaining mathematical definition of ``value''. This paper
establishes a notion of value and analyses two conflicting models
of network. First, traditional model, fails to manifest Metcalfe's
law. Another model, that observes network in a wider context, both
confirms Metcalfe's law and shows its upper boundary.
\section{Network Value}
In lines of Von Neumann\textendash{}Morgenstern utility theorem\cite{von2007theory},
\begin{defn}
Utility of a system is a probability-weighted sum of its value for
all possible events:
\[
U=\sum_{i}^{\infty}\phi(K^{n},e_{i})P(e_{i})
\]
where $U_{S}$ is utility of a system, $\phi$ is a function on n-dimensional
vector of all system properties $K_{S}^{n}=[k_{1},k_{2},...k_{n}]$
given event $e_{i}$, and $P(e_{i})$ is the probability (or relative
frequency), of event $e_{i}$.
The definition is universal because it bases value on a scenario.
Utility of same system in different scenarios differs (and may be
even negative), but it always deterministically follows from system
properties.
Function $\phi$ calculates system utility in case of given event.
In order to support claim that network has size-dependent value, we
need to show that size-dependent component of $\phi$ can be separated
from event-dependent one. Even when we compare systems that differ
in only one parameter, we cannot extract event-independent component
because $a(x,y)=b(x)c(y)$ has no solutions. It shows that network
has no universal size-dependent value.
But it is plausible to assume that there exists a non-empty set of
events $E$ for which we can represent $\phi$ as $\phi=\xi(e_{i})\psi(k)$,
where $\xi(e_{i})$ and $\psi(k)$ are event-dependent and property-dependent
components of $\phi$. Note that $\psi(k)$ is independent of any
event, and therefore total utility can be represented as
\[
U=\psi(k)\sum_{i=1}^{\infty}\xi(e_{i})P(e_{i})
\]
In this case, if systems differ only in $k$, total utility $U$
can be represented as $U=\psi(k)C$, where $C$ is a system-independent
factor. External factors can influence $C$, but as long as our system
has only one varying parameter, its utility is proportional to a function
of that parameter. Now we need to make sure our model meets that criterion.
Information Network is a collection of information consumers and
producers connected by communication channels.
\end{defn}
In order to make size the only property that distinguishes two networks,
we have to add that all nodes and channels are identical, and also
that number of channels is a function of network size. Let's also
discard constraints by assuming that all parts of a network can process
infinite amount of data in no time. Those assumptions are enough to
make the notion of network value for set of events $E$ mathematically
consistent. In the next chapters I study and improve this model.
\section{Network Effect}
Common\cite{wikineteffect} understanding of network effect is both
simple and compelling. The primary function of a network is connecting
users. Therefore the value of a network to one user is proportional
to the number of other users: $Q\propto N-1$ ($\propto$ denotes
proportionality). As a result, total value of a network is proportional
to $N(N-1)=N^{2}-N$. Note that it is identical to the maximum number
of unique directed links between nodes.
Let's analyze that claim. By viewing many to many communication as
a simultaneous mutual broadcasting, we can model the network as a
superposition of broadcast-type networks. See figure 1.
\begin{figure}[tbh]
\begin{centering}
\textsf{\includegraphics[scale=0.165]{sarnoffs2metcalfes}}
\par\end{centering}
\caption{Switched network of size $N$ can be represented as superposition
of $N$ virtual broadcast networks with $N-1$ receivers in each}
\end{figure}
Network effect basically assumes that those networks are independent.
However closer investigation suggests that they are not: though all
virtual transmitters indeed are independent, they share same receivers.
As a result, links compete for nodes, and each link's share of its
terminals equals $\frac{1}{N-1}$. Consequently, in a network with
higher density of links one link has proportionally less value. For
example, when above network expands $X$ times, each link's share
of its terminals decreases $X$ times. To illustrate this with telephone
networks, potential to place a call for each phone increases, but
potential that the call will be answered decreases proportionally.
In other words, the gain is imaginary.
One can argue that sharing is applicable only when node capacity is
severely constrained. But that is not the case. We need $N-1$ links
to make a network of size $N$, and other links are redundant. That
is why net value of a network is always $N$. Network effect suggests
tendency to exponential growth of partially redundant data, not of
value.
\begin{lem}
Direct link in unconstrained network has no value. \end{lem}
\begin{IEEEproof}
In zero-latency, infinite bandwidth network of size $N$, node $X$
has direct connections to every other node. That results in $N-1$
direct links. Node $Y$ has just one connection (of course, to $X$).
As a result, we obtain two subnetworks. Network $A$ has $N-1$ nodes
(all except $Y$), and network $B$ has $2$ nodes: $X$ and $Y$.
Next, we set $X$ to bridge $A$ and $B$. Now $Y$ gets exactly the
same network service as $X$, and the only solution of $V_{lnk}\times(N-1)=V_{lnk}\times1$
is $V_{lnk}=0$.\end{IEEEproof}
\begin{cor}
From 1 follows that all network value is contained in nodes.
\end{cor}
This hints us that information network is identical to a non-distributed
information system, such as the computer. In computer memory, you
always get same amount of memory from $N$ units regardless of how
you connect those units. Notion that a collection of interlinked nodes
can enjoy a non-linear increase in value is mathematically inconsistent,
and there are many ways to prove it. For example we can split every
network node in $n$ parts and connect those parts back to the network,
which, according to network effect, must raise its value $\frac{n^{2}}{n}=n$
times, which is contradiction.
\section{Network Efficacy}
Describing network effect as exponential growth of value might originate
from difficulties to discriminate redundant data from valuable information.
However that neither proves nor disproves Metcalfe's law. The thing
is, describing network as a collection of nodes is not the best way
to model it. A better point of view is that network has value in previous
sense. It provides connectivity between certain parts of (potentially
larger), physical or business system, and it is that larger system
that has value. Here are the corresponding definitions:
\begin{defn}[Information System]
The information system is a collection of complementary subsystems
(nodes), that contribute to overall information value.
\begin{defn}[Telecommunications Network]
The telecommunications network is an apparatus for information transmission
between flexible number of terminals.
\end{defn}
\end{defn}
According to later definitions, the telecommunications network is
just one of possible transmission agents for an information system;
it differs from the system itself. As an illustration, the broadcast
network can be viewed as a complex business that delivers its products
via telecommunications network. According to the definition, Information
System is too complex to derive value from its size. Because of that,
information network is a poor approximation of an information system.
To replace value with something more appropriate for a network, let
me introduce another definition. It is the efficacy of a network.
\begin{defn}[Network Efficacy]
Network Efficacy is the amount of useful data an underlying information
system is able to send through a network of $N$ identical nodes.
It is the product of network size and node communication efficacy
$\zeta$:
\[
\psi=N\zeta
\]
As before, we investigate an unconstrained network with identical
nodes that limits $\zeta$ neither by bandwidth nor by node capacity.
Note that Metcalfe's law holds only when $\zeta$ is proportional
to $N$. If $\zeta$ is constant, network properties are linear in
$N$. As we will see in a moment, $\zeta$ exhibits both of those
behaviors. Now let me describe a model that reveals this.
\end{defn}
\section{Deficit Model}
Consider the following scenario: User $X$ wants to contact certain
members of her family over Skype. These people can be described by
set $A\subseteq\Omega$, where $\Omega$ denotes all nodes of the
information system that corresponds to $X$'s family.
Skype as a network represents an independent set $B$. Those of $\Omega$
members who use Skype comprise the effective network $E=B\cap\Omega$.
Note that the actual size of the Skype network is irrelevant Effective
network size equals $|E|$ ($|...|$ denotes cardinality). $X$ can
contact only those people that belong to $A\cap E$ (the intersection
of her contact list and her relatives in Skype network). Expected
proportion of people from $A$ that Skype allows to contact is independent
of $A$. It is $\frac{|E|}{|\Omega|}$, or effective network size
divided by information system size. Note that $|E|\le|\Omega|$ because
$E\subseteq\Omega$. The same is true for every member of $E$, which
leads to the following:
\begin{thm}[Network Efficacy Theorem]
Network Efficacy is proportional to square of effective network
size divided by information system size, where effective network represents
the networked part of the information system.
\[
\psi=\alpha\frac{N_{E}^{2}}{N_{\Omega}}
\]
where $\alpha$ is size-independent transmission rate, $N_{E}=|E|=|B\cap\Omega|$,
and $N_{\Omega}=|\Omega|$.\end{thm}
\begin{IEEEproof}
To avoid projections to the future, let us investigate what happens
when a network suddenly disconnects part of its nodes: $N_{E}=N_{\Omega}/x$
(1). The disconnected nodes try to contact the network at a constant
rate $\alpha$ and receive errors. Remaining nodes also contact old
address space, therefore their success rate is proportional to remaining
fraction of the network: $\psi=\alpha\frac{N_{E}}{x}$ (2). From 1,
$x=\frac{N_{\Omega}}{N_{E}}$ (3). Substituting 3 into 2 we get $\psi=\frac{\alpha N_{E}^{2}}{N_{\Omega}}$.
The cause of non-serviceable requests is that active information system
is larger than its accessible part. Therefore, regardless of whether
network shrinks, grows, or stays constant, average fraction of satisfied
demand is proportional to the square of networked fraction of the
information system.
\end{IEEEproof}
The next section illustrates that by example.
\section{Heterogeneous Networks}
Deploying a network inside of an information system implies that there
exists an old way of communication between links. In fact, telecommunication
networks constantly replace one another, and most networks are heterogeneous.
\begin{cor}
From Network Efficacy Theorem follows that when there is a default
network $D$ that connects all nodes and a preferred network $K$
that connects part of those nodes,
\[
\psi_{K}+\psi_{D}=\frac{1}{1-n^{2}}
\]
where $\psi_{D,}+\phi_{K}$ denotes the joint capacity of default
and preferred network, and $0\leq n<1$ is the fraction of nodes that
the preferred network connects.
\end{cor}
Let me illustrate above notion by example: Parallel computation cluster
nodes produce synchronous traffic at a constant rate. All traffic
is produced and consumed locally; it is evenly distributed among nodes.
Nodes are connected by a network switch with total flow cap of $1$
Tb/s. An additional, separate network with flow capacity of $2$ Tb/s
is being deployed. The task is to fully load both of networks. When
two networks depend on each other, total traffic is limited to $\frac{1}{1-n}$
of the smaller capacity, where $n$ is fraction of traffic that goes
over the faster network. If faster network took $\frac{2}{3}$ of
the load from the slower one, total capacity would be $3$ Tb/s. As
far, workers have installed faster network on $\frac{2}{3}$ of the
nodes. What is the overall capacity of the network?
\textbf{A}: Efficacy of the faster network $\psi_{F}=\frac{(2/3N_{\Omega})^{2}}{N_{\Omega}}=\frac{4}{9}$
of the slower network. As was already shown, it happens because nodes
connected by the faster network can find a destination inside it only
$\frac{2}{3}$ of the time, therefore $\frac{1}{3}$ of time they
default to the slower network. As a result, total network capacity
is $1/(1-n^{2})=1.8$ Tb/s. To achieve $3$ Tb/s, workers need to
connect $n=\sqrt{1-\frac{1}{\psi_{D}+\psi_{K}}}\approx81.65$\% of
nodes.
\section{Saturated Behavior}
When network size equals to information system size, adding new nodes
results in equal growth of $N_{E}$ and $N_{\Omega}$. As a result,
$\zeta$ stays constant. See figure 2.
Though each node accesses increased address space, each address has
proportionally less value. The same applies to the amount of obtainable
information.
\begin{figure}[tbh]
\includegraphics{saturatedeff}
\caption{After a network catches the size of its underlying information system,
it switches to linear growth}
\end{figure}
\section{Multipurpose Networks}
A multipurpose network is one that serves many information systems.
Total efficacy of a multipurpose network is the sum of its capacities
for all systems. Therefore, overall size of a network may serve as
a rough approximation of its efficacy.
\[
\psi_{mp}^{tot}=\sum_{i}\psi_{i}\propto N^{2}
\]
\section{Discussion}
Studied models did not confirm moderate exponential growth estimates
suggested by \cite{briscoe09f:nlogn-icr}. Their analysis also invalidates
a notion that network effect can lead to exponential growth of utility.
Instead, it suggests that when a network grows by extending existing
or creating new information systems, its value grows linearly, regardless
of the number of spawned systems. One explanation is that simple network
model is just a distorted approximation of a distributed information
system, and as such it has dynamics that is common to information
systems.
However in a more complex setup, when a network grows within a larger
information system, its efficacy raises exponentially until it catches
the size of the underlying information system. That exponential growth
exactly follows Metcalfe's law. Efficacy dramatically affects network
utility, and may play a major role in limiting network value in heterogeneous
environments.
In addition to linear growth of information, model suggests linear
growth of redundancy. As information density is a reciprocal of data
redundancy, network effect dictates that unique content of overall
network traffic is a reciprocal of network size. However the later
effect may be mitigated by limited applicability of network model
to real-world information systems. It can be argued that information
systems built with network model in mind, such as WWW, exhibit more
redundancy than setups based on complementary subsystems. That suggests
promotion of node complementarity as a possible way to reduce redundancy.
\appendices{}
\bibliographystyle{plain}
|
\section{Introduction}\label{sec:introduction}}
\IEEEPARstart{M}{ulti-label} learning has a great many achievements and prospects for its successful application to real-world problems, such as text categorization \cite{yang2009effective,li2015supervised}, gene function classification \cite{barutcuoglu2006hierarchical,cesa2012synergy} and image/video annotation \cite{wang2008automatic,cabral2011matrix}. In the multi-label learning problem, each example can be associated with multiple and nonexclusive labels, and the goal of learning is to allocate the most relevant subset of labels to a new example. In the era of big data, the size of label set is constantly increasing. For example, there are already millions of image tags in Flickr and categories in Wikipedia. Hence, the research challenge is to design scalable yet effective multi-label learning algorithms, which are capable to compromise the conflict between the prodigious number of labels and the limited computation resource.
A straightforward approach for multi-label learning is 1-vs-all or Binary Relevance (BR) \cite{tsoumakas2010mining}, which learns an independent classifier for each label. However, the constant increase on the size of label set makes it computationally infeasible. The prevalent technique to deal with label proliferation problem is to \textit{shrink} the large label space by embedding original high-dimensional label vectors into low-dimensional representations. Different projection mechanisms can be adopted for transforming label vectors, including compressed sensing \cite{hsu2009multi}, principal component analysis \cite{tai2012multilabel}, canonical correlation analysis \cite{zhang2011multi}, singular value decomposition \cite{chen2012feature} and Bloom filters \cite{cisse2013robust}. The predictions made in the low-dimensional label space are then transformed back onto the original high-dimensional label space via a decomposition matrix \cite{yu2014large,sun2014multi} or k-nearest neighbor (kNN) technique \cite{bhatia2015sparse}. Additionally, some works \cite{balasubramanian2012landmark,bi2013efficient} attempt to select a small yet sensible subset of labels to represent the entire label set, and then learn hypotheses regrading this smaller label set.
Aforementioned learning methods successfully remedy the label proliferation problem and have achieved promising performance in different multi-label tasks. However, these methods may be restricted in two aspects. (a) Nearly all of these algorithms explicitly or implicitly assume that \emph{all} the labels in the learning process are given for once, and thus they can only tackle a \textit{static label setting}, which could be easily violated in changing environments. In practice, there is a rapid increase in the volume of labels as the understanding of data goes more in depth, disabling the static label setting in consequence. For example, in social network, users usually belong to different groups or clubs according to their individual characteristics or interests. With the fast development of information techniques and the convenient information transmission, there could emerge new interest groups or clubs, which should be timely and accurately recommended to prospective members. In event detection problem, it is urgent to timely and effectively investigate an emerging new event, which is excluded in the early detection systems. In this way, we can immediately integrate new events into the previous detection system by borrowing the knowledge from past events. Therefore, involvement of constantly emerging labels is very significant for the multi-label learning. (b) Although there are some tricks to adapt classical multi-label algorithms to handle emerging new labels, they could have various disadvantages. More precisely, independently learning for new labels would neglect the knowledge harvested from past labels; integrating new labels and past labels to re-train a new multi-label model requires a huge computation cost, which thus decreases the scalability of the multi-label system, especially when dealing with large scale scenario. As a result, it is challenging to efficiently and accurately model the emergence of the new labels.
Targeting at the both problematic aspects, we define and study streaming label learning (SLL), \emph{i.e.}, labels are arrived on the fly, to learn a model for the newly-arrived $k~(\geq 1)$ labels given a well-trained model for the existing $m$ (usually large) labels within multi-label learning context. The proposed streaming label learning algorithm equips with the capability of modelling newly-arrived labels with the help of the knowledge learned from past labels to timely and effectively respond to the changing and demands of environments. The core of SLL is to explore and exploit the relationships between new labels and past labels. Instead of decomposing the label matrix into label vectors in terms of different data points \cite{tsoumakas2010mining,yu2014large,bhatia2015sparse}, we examine it from the perspective of label space and represent each label through the response values on examples. Based on the idea of ``labels represent themselves", the label structure exploited for labels self-representation stands for relationships between labels, which can be inherited by hypotheses of labels as well. Given the relationships between past labels and newly arrived labels, we can thus easily model the new labels with the help of the well-trained multi-label model on the large number of past labels. We theoretically suggest that the generalization ability of hypotheses of newly arrived labels can be largely improved with the knowledge harvested from past labels. Experimental results on large-scale real-world datasets demonstrate the significance of studying streaming label learning and the effectiveness of the proposed algorithm in timely and effectively learning new labels.
The rest of the paper is organized as follows. In Section 2, we formulate the streaming label learning problem and propose the corresponding mathematical model. The optimization process is elaborated in Section 3 and theoretical analysis is given in Section 4. In Section 5, we present and analyze the experimental results, with concluding remarks stated in Section 6. All detailed proofs are shown in Appendix section.
\section{Problem Formulation}
In this section we present a streaming label learning mechanism to handle the emerging new labels with the help of the knowledge learned from past labels, which is able to explore the previously well-trained multi-label model over a large number of labels and get rid of intensive computation cost. The proposed algorithm seeks to exploit label relationship via label self-representation, which has an important influence on the hypotheses of labels.
We first state multi-label learning (MLL) and introduce frequent notations. Let the given training data set denoted by $\mathcal{D}= \{(\mathbf{x}_1,\mathbf{y}_1),...,(\mathbf{x}_n,\mathbf{y}_n)\}$, where $\mathbf{x}_i\in\mathcal{X}\subseteq\mathbb{R}^d$ is the input feature vector and $\mathbf{y}_i\in\mathcal{Y}\subseteq\{-1,1\}^m$ is the corresponding label vector. Moreover, $y_{ij}=1$ iff the $j$-th label is assigned to the example $\mathbf{x}_i$ and $y_{ij} = -1$ otherwise. Let $\mathbf{X}=[\mathbf{x}_1,...,\mathbf{x}_n]\in\mathbb{R}^{d\times n}$ be the data matrix and $\mathbf{Y}=[\mathbf{y}_1,...,\mathbf{y}_n]\in\{-1,1\}^{m\times n}$ be the label matrix. Given dataset $\mathcal{D}$, multi-label learning aims to learn a function $f: \mathbb{R}^d \rightarrow \{-1,1\}^m$ that generates the prediction on label vector for a test point.
\subsection{Label Relationship}\label{section2.1}
Probe of label relationship is demonstrated to be critical and beneficial in boosting the performance of multi-label learning \cite{bi2011multi,cheng2010bayes,hariharan2010large}. Given the label matrix $\mathbf{Y}=[\mathbf{y}_1,...,\mathbf{y}_n]\in\{-1,1\}^{m\times n}$, where $Y_{ij}$ indicates the response of $i$-th label on example $\mathbf{x}_j$, most works \cite{tsoumakas2010mining,yu2014large,bhatia2015sparse} treat $\mathbf{Y}$ from the perspective of column (example) and investigate different techniques to transform these example-wise vectors. By contrast, we propose to examine the label matrix from the label perspective, namely, considering the row-wise vectors as the representations of different labels. This thus enables us to specify the abstract concept of a label via its responses on $n$ examples. To facilitate the mathematical notations, we equivalently examine the columns of $\mathbf{Y}$'s transpose, denoted as $\mathbf{Y}^* = \mathbf{Y}^T = [\mathbf{y}^*_1,...,\mathbf{y}^*_m]\in\{-1,1\}^{n\times m}$.
In the following, we proceed to introduce two important assumptions for streaming label learning problem.
\textbf{$\bullet$ Label Self-representation.} Given $m$ labels indexed by $l_i(i=1,...,m)$, each of them can be represented by vectors $\{\mathbf{y}_1^*,...,\mathbf{y}_m^*\}$. We employ a valuable assumption of ``labels represent themselves" to model the label relationship. Specifically, a label is assumed to be represented as a combination of other labels. For example, a linear representation is utilized for a given label $l_i$,
\begin{equation}\label{label}
\mathbf{y}_i^* = \sum_{j\neq i}s_{i}^j\mathbf{y}_j^*
\end{equation}
where $\mathbf{s}_i = [s_i^1,...,s_i^m]^T$ is the coefficient vector to reconstruct label $l_i$ and $s_i^i=0$ excludes $l_i$ itself in reconstruction. Moreover, if $s_i^j>0$, then label $l_j$ has positive influence on label $l_i$ in Eq.(\ref{label}), while $s_i^j<0$ implies that label $l_j$ has negative influence on label $l_i$. $\mathbf{s}_i$ is encouraged to be sparse, so that label $l_i$ only has connections with several labels.
\textbf{$\bullet$ Hypotheses of Labels.} Multi-label learning aims to learn better hypotheses of labels with the help of relationship between labels. A simple yet effective approach to formulate the process of multi-label decision $\mathbb{R}^d \rightarrow \{-1,1\}^{m\times n}$ is via using function $f(\mathbf{x};W) = W^T\mathbf{x} = [\bm{w}_1^T\mathbf{x},...,\bm{w}_m^T\mathbf{x}]^T$. The multi-label classifier $W$ can thus be regarded as the composition of classifiers regarding different labels $\{\bm{w}_1,...,\bm{w}_m\}$, where $\bm{w}_i$ is the classifier \emph{w.r.t. } label $l_i$. We assume that the relationship between labels can be inherited by classifiers of different labels. Given label $l_i$ represented by its related labels $\mathcal{N}_i = \{l_j|s_i^j\neq0,j=1,...,m\}$ according to Eq.(\ref{label}), classifier $\bm{w}_i$ w.r.t. label $l_i$ can thus be represented by the classifiers regarding those related labels using the same coefficient vector $\mathbf{s}_i$,
\begin{equation}\label{weight}
\bm{w}_i = \sum_{j\neq i}s_{i}^j\bm{w}_j.
\end{equation}
Broadly speaking, label relationship acts as a regularization of multi-label classifier $W$, which encourages $W$ to be represented by itself as well.
The linear self-representation of labels is a simple yet effective assumption within multi-label learning indeed. There usually exist significant dependencies among labels in multi label learning. With the extension of real datasets, these dependencies are enhanced as well. Thus it is easy for some specific label to investigate a group of ``neighborhood" labels involved in its linear representation among a great many labels. \footnote{We will validate the linear self-representation of labels empirically in Section 6.} Note that our operation over label matrix resembles label selection techniques in \cite{balasubramanian2012landmark}, which also assumes linear self-representation of labels, but they are developed from distinct perspectives. Label selection focuses on selecting a shared subset of labels to recover all the given labels. Nevertheless, the proposed label self-representation aims to accurately represent the label current in progress, which will not be distracted by the reconstruction results of other labels. Besides, we propagate the label relationship exploited through label self-representation into the process of learning multi-label classifiers, instead of independently learning classifiers for the selected label subset. It is instructive to note that the label self-representation operation implicitly encourages $W$ to be low rank, which is also a widely-used assumption in multi-label learning \cite{sun2014multi,yu2014large}.
\subsection{Streaming Label Learning}
Conventional well-trained multi-label learning model on data associated with a large number of labels is difficult to be adapted to the newly arrived labels without computationally intensive re-training, and the label relationship discovered by existing methods cannot be straightforwardly extended with the emerging new labels as well.
This section details the proposed steaming label learning (SLL) mechanism, designed for accommodating emerging new labels. Basically, SLL consists of two steps. For the $k$ newly arrived labels, we first exploit their relationships between $m$ past labels, and then learn their corresponding hypotheses with the help of previously well-trained model regarding past labels and the exploited label relationship.
We first assume there is only one newly-arrived label $(k=1)$ and then extend it into a mini-batch setting ($k\geq 1$). Denote a well-trained multi-label learning model over $m$ labels as $W_m=[\bm{w}_1,...,\bm{w}_m]$, and the label matrix of $n$ examples aligned in label dimension is denoted as $Y_m^*=[\mathbf{y}_1^*,...,\mathbf{y}_m^*]$. Besides, given matrix $S_m$ describing label relationship, we then have $Y^*_m \approx Y^*_m S_m$ and $W_m \approx W_m S_m$.
\textbf{Streaming label learning with one label.} Given a new label $l_{m+1}$ represented by the response vector $\mathbf{y}_{m+1}^*$ on $n$ examples, we assume that it can be represented by the past $m$ labels $\mathbf{y}^*_{m+1} = \sum_{j=1}^m s_{m+1}^j \mathbf{y}^*_j$, where $\mathbf{s}_{m+1}$ is the coefficient vector of the new label $l_{m+1}$, and can be determined by solving the following optimization problem:
\begin{equation}\label{new:weight}
\argmin_{\mathbf{s}_{m+1}\in\mathbb{R}^m} \quad\frac{1}{2}\norm{\mathbf{y}^*_{m+1}-Y^*_m\mathbf{s}_{m+1}}_2^2 + \lambda \norm{\mathbf{s}_{m+1}}_1,
\end{equation}
where the least squares acts as a residual penalty for the label representation and $\lambda>0$ encourages sparsity. In this way, we can obtain the representation $\mathbf{s}_{m+1}$ of the new label $l_{m+1}$.
The label relationship between the new label and past labels can provide us helpful information to learn the hypothesis regarding the new label. According to hypotheses of labels in Eq.(\ref{weight}), we have
\begin{equation}\label{label1}
\bm{w}_{m+1} = W_m \mathbf{s}_{m+1}
\end{equation}
for new label $l_{m+1}$. Eq.(\ref{label1}) actually provides prior information for the new classifier $\bm{w}_{m+1}$ to be learned. By further considering its prediction error, $\bm{w}_{m+1}$ can be learned by minimizing the following objective function:
\begin{multline}\label{new:classifierobj}
J(\bm{w}_{m+1}) = \sum_{i=1}^n\ell(y^*_{m+1,i},\mathbf{x}_i^T\bm{w}_{m+1})\\ + \frac{\beta}{2}\norm{\bm{w}_{m+1}-W_m \mathbf{s}_{m+1}}_2^2
\end{multline}
where $\beta>0$ is a regularization parameter and $\ell(\cdot,\cdot)$ is a loss function to measure the discrepancy between the ground-truth label and the prediction. We choose the $\ell_2$ loss function in our experiment for simplicity though our SLL can adapt to other loss functions, since $\ell_2$ loss is shown to stand out in most cases of multi-label classification tasks comparing to other loss functions, such as logistic loss and $L_2$-hinge loss \cite{yu2014large}. Therefore, the new classifier can be learned subsequently, but integrated with the already-learned knowledge of past labels.
SLL with one new label can be naturally extended into a mini-batch setting, where a mini-batch of new labels instead of one single label is processed at a time.
\textbf{Mini-batch extension.} Given a batch of $k$ new labels $l_{new} = \{l_{m+1},...,l_{m+k}\}$ represented by vectors $Y^*_{new} = \{\mathbf{y}^*_{m+1},...,\mathbf{y}^*_{m+k}\}$, the challenging part is that we need to consider not only the relationships between new labels and past labels, but also those among new labels.
Suppose that each new label is reconstructed with the help of all the other labels (new labels and past labels), i.e. $l_{m+i} = \sum_{j=1}^{m+k} s_{m+i}^j l_j ,j\neq m+i$. According to Eq.(\ref{label}), we can obtain
\begin{equation}
Y^*_{new} = Y^*_{m+k}S_{new}
\end{equation}
with $Y^*_{m+k} = [Y^*_m,Y^*_{new}]$. Then the representation of new labels can be also solved through the following optimization problem as Eq.(\ref{new:weight}),
\begin{equation}\label{online:weight}
\begin{array}{rl}
\arg\displaystyle{\min_{S_{new}}} & \frac{1}{2}\norm{Y^*_{new} - Y^*_{m+k}S_{new}}_F^2+ \lambda\norm{S_{new}}_{1,1}\\
s.t.& (S_{new})_{m+i,i} = 0,\ \forall i = 1,...,k.\\
\end{array}
\end{equation}
where $(S_{new})_{m+i,i}=0$ is to exclude each individual label from its reconstruction. As a result, the representation of new labels can be obtained. Moreover, considering the partitioning of $S_{new} = [S_{new}^{(1)};S_{new}^{(2)}]$, we have $Y^*_{new} = Y^*_{m+k}S_{new} = Y^*_mS_{new}^{(1)} + Y^*_{new}S_{new}^{(2)}$. Thus we can observe that $S_{new}^{(1)}$ corresponds to the representation from $m$ past labels while $S_{new}^{(2)}$ means the interactive representation of the $k$ new labels, which coheres with the assumption we make.
After obtaining $S_{new}$, it can also be employed in the process of learning new classifiers, $W_{new} = [W_m,W_{new}]S_{new} = W_mS_{new}^{(1)} + W_{new}S_{new}^{(2)}$ where $W_{new}$ is the parameter matrix for new labels. Then the optimization function is formulated as
\begin{multline}\label{new:batch:classifier}
\tilde{J}(W_{new}) = \frac{1}{2}\sum_{i=1}^n\norm{\mathbf{y}_{new}^{i}-W_{new}^T\mathbf{x}_i}_2^2\\
+ \frac{\beta}{2}\norm{W_{new}(\mathbf{I}-S_{new}^{(2)})-W_mS_{new}^{(1)}}_F^2
\end{multline}
Note that the adopted loss function in Eq.(\ref{new:batch:classifier}) is decomposable, namely, $\norm{\mathbf{y}_{new}^{i}-W_{new}^T\mathbf{x}_i)}_2^2 = \sum_{j=1}^k (y_{new}^{ij}-\bm{w}_{m+j}^T\mathbf{x}_i)^2$ where $y_{new}^{ij}$ is the $j$-th new label value of the $i$-th example $\mathbf{x}_i$. Besides, when $S_{new}^{(2)}=\bf{0}$, it means that we neglect the relationships within new labels, and only investigate the relationships between new labels and past labels. This case coheres with the single new label scenario, and thus it can be viewed as a special case with batch size 1.
The framework of the proposed algorithm is summarized in Algorithm \ref{alg:online}.
\begin{algorithm}[tb]
\caption{Framework of Streaming Label Learning}
\label{alg:online}
\begin{algorithmic}[1]
\REQUIRE{A well-trained multi-label classifier $W\in\mathbb{R}^{d\times m}$ with respect to $m$ (usually large enough) labels.}
\WHILE{areMoreLabelsAvailable() }
\STATE $\{(\mathbf{x}_i,\mathbf{y}_i)\}_{i=1}^n\leftarrow$ getNewLabelTrainingData()
\STATE $Y^*_{m+k} = [Y_m^*,Y_{new}^*] \leftarrow \mathbf{Y}^T=[\mathbf{y}_1,...,\mathbf{y}_n]^T$
\STATE Label representation structure $S_{new}$: solving Eq.\ref{online:weight},\\
\STATE New labels classifier $W_{new}$: solving Eq.\ref{new:batch:classifier}.\\
\STATE $W \leftarrow [W,W_{new}]$.
\STATE $m \leftarrow m +k$
\ENDWHILE
\ENSURE{multi-label classifier with parameter $W$ in terms of all $m$ labels.}
\end{algorithmic}
\end{algorithm}
\section{Optimization}
In this section, we show the details of optimization in SLL. Basically, the optimization mainly consists of two parts, \emph{i.e.} solving the label representation $S$ and the classifier parameter matrix $W$. Additionally, we introduce a natural method to initialize the proposed steaming label learning algorithm.
\subsection{Optimizing the New Label Representation $S$}
For a single new label (see Problem (\ref{new:weight})), the optimization is unconstrained yet with a non-smooth regularization term. In fact, it is an $\ell_1$-regularized linear least-squares problem or Lasso. This problem has been thoroughly investigated in many literatures, and there exist a great many algorithms to solve it efficiently, such as alternating direction method of multipliers (ADMM) \cite{boyd2011distributed}, least angle regression (LARS) \cite{efron2004least}, grafting \cite{perkins2003online} and feature-sign search algorithm \cite{lee2006efficient}. Also there are some off-the-shelf toolboxes or packages solving it, \emph{e.g. }~CVX \cite{cvx,gb08}, TFOCS \cite{becker2011templates} and SPAMS \cite{mairal2009online}. Note that the optimization of Problem (\ref{new:weight}) depends on the dimension of the label set. To handle the label proliferation problem, we propose to implement clustering trick \cite{bhatia2015sparse} to all labels, then select labels with an appropriate number for each cluster and compose a relatively small label dictionary in order to improve the efficiency.
As for new labels in the mini-batch setting, Problem (\ref{online:weight}) can be efficiently solved over each column of $S_{new}$ using parallel techniques. And each subproblem can also be handled like Eq.(\ref{new:weight}).
\subsection{Optimizing the Classifier $W$ for New Labels }
\textbf{$\bm{w}_{m+1}$ in Problem (\ref{new:classifierobj})}. Various gradient-based or subgradient-based methods can be adopted to minimize Eq.(\ref{new:classifierobj}). Different from \cite{yu2014large}, which seeks to process a large number of labels all at once and has to turn to cheap methods, such as conjugate gradient (CG) method, Eq.(\ref{new:classifierobj}) provides a practical solution to learn multiple labels in a streaming manner, and thus largely alleviates the challenge of a prodigious number of labels on machine load and computational cost. We only concern a $d$-dimension optimization problem, which enables us to adopt more accurate methods, like LBFGS search, with little computational cost. With the squared loss, we can even obtain a closed-form solution of Eq.(\ref{new:classifierobj}),
\begin{equation}\label{eq:solution:l2}
\bm{w}_{m+1} \leftarrow (\mathbf{X}\bx^T+\beta\mathbf{I})^{-1} (\mathbf{X}\mathbf{y}_{m+1}^*+\beta W_m \mathbf{s}_{m+1}).
\end{equation}
which only needs to inverse a $d\times d$ matrix instead of $dr\times dr$ in \cite{yu2014large}, where $r$ is the low rank upper bound.
\textbf{$W_{new}$ in Problem (\ref{new:batch:classifier})}. Similarly, Eq.(\ref{new:batch:classifier}) can also be solved by various gradient-gradient methods. However, the bottleneck is the calculation of gradients and the corresponding Hessian matrix for its cost may be extremely large. Let $\bm{w}_{new} = vec(W_{new})\in\mathbb{R}^{dk}$, where $vec(\cdot)$ is the vectorization of a matrix. Since loss function $\ell$ is separable over each $\bm{w}_{m+i}$, then the gradient and Hessian matrix of $\bm{w}_{new}$ can be calculated using the gradient and Hessian matrix over each $\bm{w}_{m+i}$.
Besides, denoting $D_{new} = ((\mathbf{I}-S_{new}^{(2)})\otimes\mathbf{I})$ and $\mathbf{z}_{new} = vec(W_mS_{new}^{(1)})$, the residual penalty $\frac{1}{2}\norm{W_{new}(\mathbf{I}-S_{new}^{(2)})-W_mS_{new}^{(1)}}_F^2$ can be rewritten as $\frac{1}{2}\norm{D_{new}^T\bm{w}_{new}-\mathbf{z}_{new}}_2^2$, whose gradient and Hessian matrix are easy to calculate. Let $\ell'(a,b) = \frac{\partial}{\partial b}\ell(a,b), \ell''(a,b) = \frac{\partial^2}{\partial b^2}\ell(a,b)$, then the gradient and Hessian-vector multiplication of $\bm{w}_{new}$ are:
\begin{align}
&\nabla \tilde{J}(\bm{w}_{new}) = \mbox{stack}\left\{\sum_{i=1}^n\ell'(y_{new}^{ij},\bm{w}_{m+j}^T\mathbf{x}_i)\mathbf{x}_i\right\}_{j=1}^k \nonumber\\
&\qquad\qquad\qquad+ \beta D_{new}(D_{new}^T\bm{w}_{new}-\mathbf{z}_{new})\nonumber\\
&=vec(\mathbf{X} G + \beta[W_{new}(\mathbf{I}-S_{new}^{(2)})-W_mS_{new}^{(1)}](\mathbf{I}-S_{new}^{(2)})^T)\label{jiandan1}\\
&\nabla^2 \tilde{J}(\bm{w}_{new})\mathbf{z} = \mbox{stack}\left\{\sum_{i=1}^n \ell''(y_{new}^{ij},\bm{w}_{m+j}^T\mathbf{x}_i)\mathbf{x}_i\mathbf{x}_i^T\mathbf{z}_j\right\}_{j=1}^k\nonumber\\
&\qquad\qquad\qquad\quad+ \beta D_{new}D_{new}^T\mathbf{z} \nonumber\\
& = vec(\mathbf{X} H+\beta Z(\mathbf{I}-S_{new}^{(2)})(\mathbf{I}-S_{new}^{(2)})^T)\label{jiandan2}
\end{align}
where $G_{ij} = \ell'(y_{new}^{ij},\bm{w}_{m+j}^T\mathbf{x}_i)$, $\mathbf{z} = vec(Z) = vec([\mathbf{z}_1,...,\mathbf{z}_k])$, $H_{ij} = \ell''(y_{new}^{ij},\bm{w}_{m+j}^T\mathbf{x}_i)\mathbf{x}_i^T\mathbf{z}_j$ and stack$\{\cdot\}$ means stacking the vectors in the set to form a longer vector according to the ascending index order. As a result, the calculation of gradient and Hessian-vector multiplication can be efficiently obtained using Eqs.(\ref{jiandan1})-(\ref{jiandan2}). For the $\ell_2$ loss function, the key factors $G$ and $H$ in Eq.(\ref{jiandan1})-(\ref{jiandan2}) can be more easily calculated,
\begin{equation}\label{GH}
G = \mathbf{X}^T W_{new} - Y_{new}^*;\quad H = \mathbf{I}.
\end{equation}
In SLL, $k$ is usually small and thus we can turn to more refined techniques, such as various line search. However, since the size of Problem (\ref{new:batch:classifier}) is $d\times k$, we propose to adopt cheap methods, such as Conjugate Gradient (CG), when feature dimension $d$ is very large.
\subsection{An Initialization proposal of $W_m$ for Past Labels}
Optimization and implementation of SLL requires an already well-trained model $W_m$ for past $m$ labels, which is critical for the performance over $k$ new labels. In real application, we just utilize an obtained $W_m$ as the input of SLL to model $k$ new labels, without retraining the $m+k$ labels.
Furthermore, although SLL is designed for modelling new labels, one additional merit is that it also provides a solution for the memory limited multi label learning problem. Almost all existing MLL methods need to load the whole feature and label data into the memory, and they may be fairly restrictive to be trained on low-end computation devices at hand since the datasets are tending to be larger and larger. Fortunately, due to the separate two steps of SLL, we can load label data and feature data successively into memory, since they dominate the memory overhead of both two steps respectively. In this way, SLL can basically decrease half of memory need for the training of MLL, together with considering the dependencies among labels.
In this case, based on the two assumptions in Section \ref{section2.1}, one practical proposal for learning the initial classifier $W_m$ over past $m$ labels is in a similar approach as that for SLL, by minimizing the following objective:
\begin{multline}\label{obj:classifierobj}
\mathcal{J}(W_m,S_m) = \frac{1}{2}\sum_{i=1}^n\norm{\mathbf{y}_i-W_m^T\mathbf{x}_i}_2^2 + \lambda_1\norm{S_m}_{1,1} \\+ \frac{\lambda_2}{2}\norm{W_m-W_mS_m}_F^2+\frac{\lambda_3}{2}\norm{Y^*_m-Y^*_mS_m}_F^2
\end{multline}
where $\norm{S_m}_{1,1}=\sum_{i=1}^m\norm{\mathbf{s}_i}_1$ promotes the sparsity of $S_m = [\mathbf{s}_1,...,\mathbf{s}_m]\in\mathbb{R}^{m\times m}$. $\lambda_i>0~(i=1,2,3)$ are the weight parameters. Usually, we expect that the objective function is minimized given a small reconstruction error of labels, thus much weight (\emph{i.e.}~large $\lambda_3$) should be imposed on $\frac{1}{2}\norm{Y^*_m-Y^*_mS_m}_F^2$ term in Eq.(\ref{obj:classifierobj}).
Problem (\ref{obj:classifierobj}) is basically solved with the alternative iteration strategy, i.e. fixing one variable and optimizing the other until convergence. As for solving for $S_m$, it is still a lasso problem, and can adopt the same methods in solving Problem (\ref{online:weight}). As for solving $W_m$, it can be viewed as a special case of Problem (\ref{new:batch:classifier}) with $S_{new}^{(1)} = \bm{0}$, thus Eq.(\ref{new:batch:classifier}) is of size $d\times m$. In this case, especially referred to large scale labels, we may perform cheap updates and obtain a good approximate solution. For example, Conjugate Gradient (CG) can be employed to significantly reduce the computational complexity based on Eq.(\ref{jiandan1}) and (\ref{jiandan2}).
\section{Theoretical Analysis}
In this section, we theoretically analyze the proposed SLL, regarding the following two aspects: (a) the generalization of the designed classifier for new labels; and (b) the difference between the classifier parameter matrix obtained with streaming labels and that without streaming labels.
\subsection{Generalization Error Bounds}
We first analyze excess risk bounds for SLL. In particular, we present a generalization error bound for the new classier learned in the steaming fashion. Moreover, we show that under some circumstances, our SLL can give a tighter bound than the common trace norm or Frobenius norm regularization in the ERM framework.
Since SLL focuses on boosting the performance of new labels by exploring the knowledge from past labels, it needs a well-trained multi-label classifier as the initialization input, denoted as $W_{old}\in\mathbb{R}^{d\times m}$. Suppose $k$ new labels are involved at a time, then SLL is implemented upon a data distribution $\mathcal{D} = \mathcal{X}\times\{-1,1\}^k$, where $\mathcal{X}\in\mathbb{R}^d$ is the feature space. Training data contains $n$ points $(\mathbf{x}_1,\mathbf{y}_1),...,(\mathbf{x}_n,\mathbf{y}_n)$, which are sampled i.i.d. from the distribution $\mathcal{D}$, where $\mathbf{x}_i\in\mathcal{X}$ is the feature vector and $\mathbf{y}_i\in\{-1,1\}^k$ is the ground-truth label vector. Our SLL is based on the proposed label self-representation and hypotheses, which can be viewed as a regularization. And the regularized set can be written as $\mathcal{W}:= \{W\in\mathbb{R}^{d\times k},\norm{W-W_{old}S}_F \leq \varepsilon, \norm{S}_{1,1} \leq \lambda\}$, where $S$ is the representation weight matrix of new labels with a sparsity controlling parameter $\lambda$. For simplicity, we just analyze the scenario where $k$ new labels have no interaction in their representation.
Given the obtained training data, SLL learns a classifier $\hat{W}$ by minimizing the empirical risk over the regularized set $\mathcal{W}$, $\hat{\mathcal{L}}(W) = \frac{1}{n}\sum_{i=1}^n \sum_{l=1}^k\ell(\mathbf{y}^l_i,\mathbf{x}_i,\bm{w}_l)$ and $\hat{W} \in \arg\min_{W\in\mathcal{W}}\hat{\mathcal{L}}(W)$. Define the population risk of an arbitrary $W$ as $\mathcal{L}(W) = \Exp_{(\mathbf{x},\mathbf{y})} [\kern-0.15em[ \sum_{l=1}^k\ell(\mathbf{y}^l,\mathbf{x},\bm{w}_l) ]\kern-0.15em]$, then the goal is to show the learned $\hat{W}$ possesses good generalization, \emph{i.e.}, $\mathcal{L}(\hat{W})\leq\inf_{W\in\mathcal{W}}\mathcal{L}(W) + \epsilon$. We have the following theorem.
\begin{theorem}\label{theorem:1}
Assume we learn a new predictor $W\in\mathbb{R}^{d\times k}$ in terms of $k$ new labels using streaming label learning formulation $\displaystyle{\hat{W} = \arg\min_{W\in\mathcal{W}}\hat{\mathcal{L}}(W)}$ over a set of $n$ training points, where $\mathcal{W}:= \{\norm{W-W_{old}S}_F \leq \varepsilon, \norm{S}_{1,1} \leq \lambda\}$. Then with probability at least $1-\delta$, we have
\begin{equation*}
\mathcal{L}(\hat{W}) \leq \inf_{W\in\mathcal{W}}\mathcal{L}(W) + \mathcal{O}\left((\varepsilon+\lambda c)\sqrt{\frac{k}{n}}\right) + \mathcal{O}\left(k\sqrt{\frac{\log\frac{1}{\delta}}{n}}\right)
\end{equation*}
where $c = \norm{W_{old}}_F$ and we presume (w.l.o.g) that $\Exp_{\mathbf{x}} [\kern-0.15em[\norm{\mathbf{x}}_2^2]\kern-0.15em]\leq 1$.
\end{theorem}
According to Theorem \ref{theorem:1}, the key term of the upper bound is the second term and it depends on the value of $(\varepsilon+\lambda c)$. $\varepsilon$ is related to the accuracy of the self-representation of label hypotheses, then if the representation performs accurately, $\varepsilon$ can be sufficiently small. $\lambda$ controls the coefficient sparsity and is usually small. Moreover, considering $\norm{W}_F\leq\varepsilon+\lambda\norm{W_{old}}_F$ in $\mathcal{W}$, thus SLL might provide a small upper bound of $\norm{W}_F$ or sometimes of $\norm{W}_*$, which means the generalization error bound generated by SLL could be tighter than the common Frobenius or trace norm regularization, when $\varepsilon$ and $\lambda$ are sufficiently small. Proof of Theorem \ref{theorem:1} is referred to Appendix \ref{proof:1}.
\subsection{Streaming Approximation Error Bound}
We now investigate whether the classifier matrix learned by SLL is seriously deviated from the one learned under the conventional multi-label learning setting. Precisely, suppose we have $k$ labels and the classifier matrix is $\hat{W}_k$, and then we learned a new label classifier matrix $\hat{\bm{w}}$ using SLL. However, without SLL we would learn a classifier corresponding to all $k+1$ labels, denoted as $\hat{W}_{k+1}$. The goal is to estimate the difference between $\hat{W}_{k+1}$ and $[\hat{W}_k,\bm{w}]$, which reflects the cost of classifier learned by SLL.
Given the training data $\mathbf{X}=[\mathbf{x}_1,...,\mathbf{x}_n]\in\mathbb{R}^{d\times n}$ and their label matrix $\mathbf{Y}=[\mathbf{y}_1,...,\mathbf{y}_n]\in\{-1,1\}^{(k+1)\times n}$, the classifier parameter matrix $\hat{W}_{k+1}$ is determined in the following optimization:
\begin{equation}\label{deter}
\hat{W}_{k+1} = \argmin_{W\in\mathbb{R}^{d\times (k+1)}}\sum_{i=1}^n \ell(\mathbf{y}_i,\mathbf{x}_i,W) + \frac{\lambda}{2}\norm{W-WS}_F^2,
\end{equation}
where $S$ is the label structure matrix of all $k+1$ labels. Then we have the following theorem:
\begin{theorem}\label{theorem:2}
Given the training data $\{\mathbf{X},\mathbf{Y}\}$ of $k+1$ labels, the classifier matrix $\hat{W}_{k+1}$ is determined in Eq.(\ref{deter}). Assuming the first $k$ labels are also learned using Eq.(\ref{deter}) denoted as $\hat{W}_k$, while the $(k+1)$th label is learned under the streaming label learning framework, denoted as $\hat{\bm{w}}$, then the following inequality holds,
\begin{multline*}
\norm{\hat{W}_{k+1}-[\hat{W}_k,\bm{w}]}_F\leq \frac{2}{\lambda\sigma_1^2(\mathbf{I}-S)+\sigma_1^2(\mathbf{X})}\\
\left(\sqrt{2n\Omega C}+\lambda \tau\sqrt{\norm{\hat{W}}_F^2+\norm{\hat{\bm{w}}}_2^2}\right)
\end{multline*}
where $C = \ell_{2}(\mathbf{Y},\mathbf{X},[\hat{W}_k,\bm{w}])$ is the least squares loss value of the classifier learned by SLL; constant $\tau=\norm{\mathbf{I}-S}_F^2$ and $\sigma_1(\cdot)$ indicates the smallest singular value. Moreover, we presume (w.l.o.g) that $\norm{\mathbf{x}_i}_2^2\leq \Omega$ and $\mathbf{X}$ is of full row rank.
\end{theorem}
As indicated in Theorem \ref{theorem:2}, we present an approximation error bound for $\hat{W}_{k+1}-[\hat{W}_k,\hat{\bm{w}}]$ using the least squares loss for simplicity and it shows that the bound is directly controlled by the loss of SLL. The better the SLL learns the classifier (\emph{i.e.} the smaller $C$ is), the tighter the bound will be. In this way, if we focus on SLL with much effort, then the learned classifier would not tend to be unsatisfying. Proof of Theorem \ref{theorem:2} is referred to Appendix \ref{proof:2}.
\section{Experimental Results}
In this section, we conduct experiments on SLL to demonstrate its effectiveness and efficiency in terms of dealing with new labels.
\textbf{Datasets.} We select 5 benchmark multi-label datasets to implement the setting of SLL, including three small datasets (Bibtex, MediaMill and Delicious) and two large datasets (EURlex and Wiki10). See Table \ref{data} for the details of these datasets.
\begin{table}
\footnotesize
\caption{Data statistics. $d$ and $L$ are the number of features and labels, respectively; $\bar{d}$ and $\bar{L}$ are the average number of nonzero features and positive labels in an instance, respectively. \# means the size of a dataset.}\label{data}
\begin{tabular}{l|r|r|r|r|r|r}
Dataset & $d$ & $L$ & \#train & \#test & $\bar{d}$ & $\bar{L}$ \\ \hline
Bibtex & 1,836 & 159 & 4,880 & 2,515 & 68.74 & 2.40 \\
MadiaMill & 120 & 101 & 30,993 & 12,914 & 120.00 & 4.38 \\
Delicious & 500 & 983 & 12,920 & 3,185 & 18.17 & 19.03 \\
EURlex & 5,000 & 3,993 & 15,539 & 3,809 & 236.69 & 5.31 \\
Wiki10 & 101,938 & 30,938 & 14,146 & 6,616 & 673.45 & 18.64 \\
\hline
\end{tabular}
\end{table}
\textbf{Baseline methods.} We compare the proposed SLL with three competing methods:\\
1). BR \cite{tsoumakas2010mining}. Since our setting focuses on the streaming labels, to our best knowledge, only it can accommodate new labels without retrain the model related to past labels.\\
2). LEML (low rank empirical risk minimization for multi-label learning) \cite{yu2014large}. Since SLL and LEML are both within ERM framework, we analyze the difference of their classification performance.\\
3). SLEEC (sparse local embeddings for extreme classification) \cite{bhatia2015sparse}. Since SLL aims at handling new labels and is capable of scaling to large datasets, we choose the state-of-the-art SLEEC in extreme classification.
\textbf{Evaluation Metrics.} We use three prevalent metrics to measure the performance of all competing methods including our SLL: (a) Hamming loss, which concerns the holistic classification accuracy over all labels, (b) precision at $k$ (P@$k$), which is usually used in tagging or recommendation and only top $k$ predictions are involved in the evaluation, and (c) average AUC, which reflects the ranking performance.
For Lasso-style Problems \ref{new:weight} and \ref{online:weight}, we use a Cholesky-based implementation of the LARS-Lasso algorithm to efficiently solve them with a high accuracy, supported by SPAMS optimization toolbox \cite{mairal2009online}, and implement it in a parallel way. As for obtaining the classifier defined in Problems (\ref{new:classifierobj}) and (\ref{new:batch:classifier}), we propose to use LBFGS line search in small datasets (or use Eq.(\ref{eq:solution:l2}) with moderate $d$ accelerated by GPU) and utilize conjugate gradient descent in large datasets, based on techniques in Eqs.(\ref{jiandan1}) and (\ref{jiandan2}).
\begin{figure*}[!ht]
\centering
\subfloat[]
{\includegraphics[width=0.6\columnwidth]{ys_1}}
\hspace{0.5cm}
\subfloat[]
{\includegraphics[width=0.6\columnwidth]{ys_2}}
\hspace{0.5cm}
\subfloat[]
{\includegraphics[width=0.6\columnwidth]{cai}}
\caption{(a) Accuracy of P@$1$ (Bibtex \& Delicious) and P@$3$ (Mediamill) of SLL in different initial ratio of label size. (b) The average AUC accuracy of SLL in different initial ratio of label size. (c) Comparison of P@$1$ accuracy in all competing methods with 100\% initial ratio.}
\label{fig1}
\end{figure*}
\subsection{Validation of Effectiveness}
\textbf{Single new label.}
Basically SLL relies on an initial multi-label classifier related to the past labels, which can be learned by solving Problem (\ref{obj:classifierobj}). We first compare the classification results with varying label ratios which are involved in the initialization training. Since the streaming labels are processed one by one, only BR can handle this case while other methods (including LEML and SLEEC) have to implement retraining many times, but have identical results with all labels ( ratio = 100\%) involved in initialization. Figure \ref{fig1} shows the P@$3$ accuracy and average AUC results, together with their trends, in various initial label ratios. From these results, we have the following observations. 1) For SLL, more labels involved in the initialization tend to improve its performance. However, with the increase of initial ratio, the improvement would be less obvious; when the ratio is larger than an appropriate value (\emph{e.g. } 70\% for Bibtex, 60\% for Delicious), the performance would keep relatively stable, even sometimes getting worse. Thus selecting the initial label size is very critical when dealing with large datasets, and 50\%$\sim$70\% would be a satisfying option. 2) In terms of all labels (100\% ratio), SLL can yield better accuracies than those of BR and LEML and is competitive with SLEEC. This indicates our used label structure tends to be a stronger regularization in training multi-label classifiers compared to the common Frobenious or trace norm regularization.\\
\begin{figure*}[ht]
\centering
\subfloat[Delicious]
{\includegraphics[width=0.6\columnwidth]{TIME_DELICIOUS.pdf}}
\hspace{3mm}
\subfloat[EURlex]
{\includegraphics[width=0.6\columnwidth]{TIME_EURLEX.pdf}}
\hspace{3mm}
\subfloat[Wiki10]
{\includegraphics[width=0.6\columnwidth]{TIME_WIKI10.pdf}}
\caption{Running time of 100 new labels based on different number of past labels in large datasets.}
\label{time}
\end{figure*}
\begin{table*
\caption{Comparison of learning new labels with different batch sizes by regarding 50\% labels as past labels.}
\label{biao2}
\centering
\footnotesize
\begin{tabular}{cr|rrrr||rrrr||rrrr}%
\multicolumn{2}{c|}{}& \multicolumn{4}{c||}{P@$1$} & \multicolumn{4}{c}{P@$3$} &\multicolumn{4}{c}{P@$5$}\\
\multicolumn{2}{c|}{}& SLL & BR & LEML & SLEEC & SLL &BR& LEML & SLEEC & SLL &BR& LEML & SLEEC \\ \hline
\multirow{5}{*}{Bibtex} & 15 &\textbf{11.51} &10.85 &10.56 &10.89 &\textbf{5.09} &4.26 &4.12 & 4.45&\textbf{2.94} &2.68 &2.66 &2.70\\
& 30 &\textbf{18.20} &17.56 &17.48 &17.89 &\textbf{9.03} &8.42 &8.37 & 8.55&\textbf{5.52} &5.32 &5.27 &5.38\\
& 45 &\textbf{26.18} &25.25 &25.22 &25.84 &\textbf{12.98} &12.32 &12.29 & 12.55&\textbf{8.48} &8.03 &8.04 &8.19\\
& 60 &\textbf{41.11} &40.02 &40.16 &40.80 &\textbf{20.76} &19.78 &19.89 & 20.05&\textbf{13.62} &13.28 &13.32 &13.36\\
& 75 &\textbf{41.45} &40.91 & 40.97&41.31 &\textbf{22.85} &22.37 & 22.55& 22.61&\textbf{15.92} &15.29 &15.44 &15.53\\ \hline
\multirow{5}{*}{Mediamill} & 10 &\textbf{22.16} &21.36 &20.94 &21.75 &\textbf{13.62} &13.37 &13.32 & 13.50&\textbf{10.52} &10.39 &10.36 & 10.47\\
& 20 &\textbf{81.04} &80.27 &80.23 &80.31 &\textbf{53.41} &53.25 &53.21 & 53.36&\textbf{36.64}&36.51 &36.48 & 36.57\\
& 30 &\textbf{84.12} &54.16 &54.12 &83.99 &\textbf{54.26} &54.02 &54.01 & 54.20&\textbf{38.31} &38.06 &38.12 & 38.23\\
& 40 &\textbf{84.35} &84.12 &84.19 &84.27 &\textbf{55.35} &55.05 & 55.14& 55.28&\textbf{39.96} &39.68 &39.72 & 39.87\\
& 50 &\textbf{84.40} &84.22 &84.26 &84.34 &\textbf{55.88} &55.43 &55.66 & 55.80&\textbf{41.90} &41.45 & 41.78& 41.83\\ \hline
\multirow{5}{*}{Delicious} & 100 &\textbf{34.01} & 31.86&32.01 &32.12 &\textbf{23.04} &22.32&22.56& 22.74 &\textbf{18.16} &17.04 &17.22 & 17.32\\
& 200 &\textbf{38.97} & 38.03&38.18 &38.34 &\textbf{31.11} &30.01&30.17& 30.60 &\textbf{25.34} &24.42 &25.00 & 25.05\\
& 300 &\textbf{52.89} &52.13 &52.22 &52.78 &\textbf{43.47} &42.88&43.03& 43.30 &\textbf{36.09} &35.65 & 35.86& 35.97\\
& 400 &\textbf{58.52} &57.84 &58.00 &58.30 &\textbf{49.63} &48.78&49.22& 49.47 &\textbf{43.87} &43.08 &43.20 & 43.35\\
& 500 &\textbf{61.66} &61.23 &61.49 &61.57 &\textbf{54.52} &54.11&54.27& 54.41 &\textbf{48.50} &48.04 &48.12 & 48.33\\ \hline
\multirow{5}{*}{EURlex} & 200 &\textbf{5.12} &4.62 &4.69 & 4.81 & \textbf{2.09} &1.86 &1.91 & 1.96 &\textbf{1.28} & 1.09 &1.17 & 1.21 \\
& 400 &\textbf{11.39} &10.34 &10.55 & 10.70 &\textbf{4.95} &4.22 &4.76 & 4.88 &\textbf{3.05} &2.87 &2.96 & 3.01 \\
& 600 &\textbf{12.48} &11.81 &11.97 & 12.09 &\textbf{5.86} &5.29 &5.54 & 5.77 &\textbf{3.68} & 3.57 &3.62 & 3.66 \\
& 800 &\textbf{14.01} &13.23 &13.77 & 13.90 &\textbf{6.83} &6.42 &6.66 & 6.72 &\textbf{4.26} &4.02 &4.17 & 4.23 \\
&1000 &17.28&17.12 &17.26 & \textbf{17.31} &9.05&8.84 &9.02 & \textbf{9.08} &5.77 & 5.33 &5.65 & \textbf{5.79} \\ \hline
\multirow{5}{*}{Wiki10} & 1k &\textbf{10.19} &9.23 &9.98 &10.02 &\textbf{4.82} &4.55&4.61&4.75 &\textbf{3.19} &3.11 &3.12 & 3.15\\
& 2k &\textbf{15.62} &15.28 &15.36 &15.54 &\textbf{7.95} &7.69&7.77&7.84 &\textbf{5.44} &5.28 &5.33 & 5.38\\
& 3k &\textbf{23.18} &22.71 &22.85 &23.05 &\textbf{12.44} &12.22&12.30&12.38 &\textbf{8.63} &8.43 &8.51 & 8.56\\
& 4k &\textbf{31.11} &30.22 &30.79 &31.00 &\textbf{18.07} &17.34&17.88&18.00 &\textbf{12.52} &12.33 &12.40 & 12.47\\
& 5k &38.69 &38.29 &38.55 &\textbf{38.71} &\textbf{22.22} &21.78&22.10&22.21 &\textbf{15.40} &15.27 &15.32 & 15.37\\ \hline
\end{tabular}
\end{table*}
\noindent
\textbf{Mini-batch new labels.} For handling new labels in the mini-batch fashion, we investigate the results with different batch sizes. Specifically, for each dataset we randomly choose 50\% (for Delicious 483) labels as past labels since it tends to be a sensible option. Then we focus on the following new labels with different batch sizes. Instead of single new label senecio, in this case a batch of new labels can be independently processed within the traditional multi-label learning for LEML and SLEEC, including BR. The average results in various batches of labels are shown in Table \ref{biao2}. As indicated by the results, SLL largely outperforms other methods in dealing with mini-batch new labels. Under this setting, BR, LEML and SLEEC cannot employ the knowledge from past labels and regard the learning of new labels as an independent process. Nevertheless, SLL enables to learn new labels based on the obtained knowledge. Note that with the increase of batch size, the gap between SLL and other methods would shrink since the amount of labels is already large for well train a multi-label model; nevertheless, the training cost will increase accordingly.
\subsection{Efficiency in dealing with new labels}
So far we have shown the superiority of SLL in classification performance, we now evaluate its efficiency in larger datasets. We select 100/200, 1k/2k and 10k/20k labels of Delicious, EURlex and Wiki10 respectively to serve initialization, then focus on the performance of 100 new labels. For SLL, we adopt the same initialization results with LEML. The average running time is showed in Figure \ref{time}, zoomed out using log10 scale. We can see that SLL and BR clearly surpass LEML and SLEEC in running time since they do not need the expensive retraining process. Note that the difference between SLL and BR lies in the label structure probe procedure, and for large datasets we can use clustering trick to reduce the scale of problem (\ref{new:weight}). For example, we select 3k labels for Wiki10 to form the fixed dictionary in Lasso, thus with the increase of label size, SLL can still be comparative with BR in efficiency.
\subsection{Investigated label structure}
Since the adopted label structure plays an significant role in SLL because it helps to train classifiers with a special regularization, we intuitively show the investigated label relationships, \emph{i.e.} for a given label what labels are involved in its reconstruction. We select three labels and their five representation ``neighbors" of Bibtex dataset in Table \ref{structure}. As shown in Table \ref{structure}, we can easily see that some related labels in logic are exactly investigated by our SLL, which intuitively explains the reasons that SLL works. For example, to label ``epitope'', which is a terminology in immunology, some investigated labels are connected with its description, such as ``honogeneous" and ``sequence" while some are also in immunology, including ``lipsome" and ``immunosensor".
\begin{table}[ht
\caption{Some related labels investigated by SLL in Bibtex.}
\label{structure}
\centering
\normalsize
\begin{tabular}{r|l}%
\hline
\multirow{2}{*}{epitope}& sequence; ldl ; homogeneous ; liposome ; \\
& immunosensor \\ \hline
\multirow{2}{*}{fornepomuk} & nepomuk ; langen ; knowledge ; semantics; \\
& knowledgemanagement \\ \hline
\multirow{2}{*}{concept}& formal; requirements; empirical; data; \\
& objectoriented \\ \hline
\end{tabular}
\end{table}
\section{Conclusion}
In this paper we studied the streaming label learning (SLL) framework, which enables to model newly-arrived labels with the help of the knowledge learned from past labels. More precisely, we investigate the relationships among labels by examining label matrix from the perspective of label space and propose the label structure to embed it into the empirical risk minimization (ERM) framework, which regularizes the learning of the new classifier. We showed SLL can provide a tighter generalization bound of new labels and would not lose accuracy because SLL explores and exploits the label relationship. Thus SLL can be viewed as an efficient way to learn new classifiers under multi-label learning framework, but with no need of retraining the whole multi-label model. Experiments comprehensively demonstrated the superiority of SLL to existing multi-label learning methods in terms of handling new labels.
\appendices
\section{Proof of Theorem \ref{theorem:1}} \label{proof:1}
In this section, we detail the proof of Theorem \ref{theorem:1}, following the framework presented in \cite{yu2014large}. In the sequel, we will show that under streaming label learning, a tighter uniform convergence bound for the empirical losses will be obtained in terms of new classifiers. Denote the regularization set of SLL as $\mathcal{W}:= \{W\in\mathbb{R}^{d\times k},\norm{W-W_{old}S}_F \leq \varepsilon, \norm{S}_{1,1} \leq \lambda\}$, where $W_{old}$ is the previous multi-label classifier matrix for past labels and $S$ is the representation weight matrix of new labels with a sparsity inducing parameter $\lambda$.
The goal of the proof is to show with high probability the following inequality holds:
\begin{equation*}\label{proof:bound}
\mathcal{L}(\hat{W}) \leq \hat{\mathcal{L}}(\hat{W}) + \epsilon
\end{equation*}
where $\epsilon$ is a small quantity. $\mathcal{L}(W) = \Exp_{(\mathbf{x},\mathbf{y})} [\kern-0.15em[ \ell(\mathbf{y},f(\mathbf{x};W))]\kern-0.15em] = \Exp_{(\mathbf{x},\mathbf{y})} [\kern-0.15em[ \ell(\mathbf{y},\mathbf{x},W)]\kern-0.15em] = \Exp_{(\mathbf{x},\mathbf{y})} [\kern-0.15em[ \sum_{l=1}^k\ell(\mathbf{y}^l,\mathbf{x},\bm{w}_l) ]\kern-0.15em]$ is the expectation of loss function $\ell$ or the \textit{real} loss. $\hat{\mathcal{L}}(W) = \frac{1}{n}\sum_{i=1}^n \ell(\mathbf{y}_i,\mathbf{x}_i,W) = \frac{1}{n}\sum_{i=1}^n \sum_{l=1}^k\ell(\mathbf{y}^l_i,\mathbf{x}_i,\bm{w}_l)$ is the empirical risk of the training data. Let $\displaystyle{W^* \in \arg\min_{W\in\mathcal{W}}\mathcal{L}(W)}, \displaystyle{\hat{W} \in \arg\min_{W\in\mathcal{W}}\hat{\mathcal{L}}(W)}$, then similar analysis can be implemented to obtain $\hat{\mathcal{L}}(W^*) \leq \mathcal{L}(W^*) + \epsilon$, inducing the ultimate inequality we claim, i.e. $\mathcal{L}(\hat{W})\leq\mathcal{L}(W^*) + 2\epsilon$. Thus we focus on the original uniform convergence bound. Typically, the whole proof of Eq.\ref{proof:bound} can be accomplished within three steps. We will elaborate them in the sequel.
\subsection{Bounding Excess Risk Using Its Supremum}
To probe an appropriate upper bound of the excess risk $\mathcal{L}(\hat{W})-\hat{\mathcal{L}}(\hat{W})$, it is natural to investigate its supremum,
\begin{equation*}
\begin{split}
&\mathcal{L}(\hat{W})-\hat{\mathcal{L}}(\hat{W}) \leq \sup_{W\in\mathcal{W}} \{\mathcal{L}(W) - \hat{\mathcal{L}}(W) \} \\
= &\sup_{W\in\mathcal{W}}\left\{ \Exp_{(\widetilde{\mathbf{x}}_i,\widetilde{\mathbf{y}}_i)}\left[\kern-0.4em\left[\frac{1}{n}\sum_{i=1}^n \ell(\widetilde{\mathbf{y}}_i,\widetilde{\mathbf{x}}_i,W) \right]\kern-0.4em\right] - \frac{1}{n}\sum_{i=1}^n \ell(\mathbf{y}_i,\mathbf{x}_i,W) \right\}\\
\triangleq & g((\mathbf{x}_1,\mathbf{y}_1),...,(\mathbf{x}_n,\mathbf{y}_n))
\end{split}
\end{equation*}
For the decomposable loss function $\ell(\mathbf{y},\mathbf{x},W)= \sum_{l=1}^k\ell(\mathbf{y}^l,\mathbf{x},\bm{w}_l) $, the change in any $(\mathbf{x}_i,\mathbf{y}_i)$ would induce a perturbation of $g((\mathbf{x}_1,\mathbf{y}_1),...,(\mathbf{x}_n,\mathbf{y}_n))$ at most $\mathcal{O}(\frac{k}{n})$. Then by using McDiarmid's inequality, the sum of squared perturbations will be bounded by $\frac{2k^2}{n}$, and thus the excess risk is bounded by a term related to the expectation of $g((\mathbf{x}_1,\mathbf{y}_1),...,(\mathbf{x}_n,\mathbf{y}_n))$, the expected supr\={e}mus deviation. Therefore, with probability at least $1-\delta$, it holds that
\begin{multline*}
\mathcal{L}(\hat{W})-\hat{\mathcal{L}}(\hat{W})\\ \leq \Exp_{(\mathbf{x}_i,\mathbf{y}_i)} [\kern-0.15em[ g((\mathbf{x}_1,\mathbf{y}_1),...,(\mathbf{x}_n,\mathbf{y}_n)) ]\kern-0.15em] + \mathcal{O}\left(k\sqrt{\frac{\log\frac{1}{\delta}}{n}}\right)
\end{multline*}
In the sequel, we will investigate the upper bound of the expected supr\={e}mus deviation.
\subsection{Bounding Expected Supr\={e}mus Deviation by a Rademacher Average}
Now we bound the expected supr\={e}mus deviation using its calculation and some tricks related to the Rademacher complexity. Note that we adopt a Rademacher average introduced in \cite{yu2014large}. And we have
\begin{equation*}
\begin{split}
& \Exp_{(\mathbf{x}_i,\mathbf{y}_i)} [\kern-0.15em[ g((\mathbf{x}_1,\mathbf{y}_1),...,(\mathbf{x}_n,\mathbf{y}_n)) ]\kern-0.15em]\\
= & {\normalsize \Exp\limits_{(\mathbf{x}_i,\mathbf{y}_i)}\left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\left\{ \Exp_{(\widetilde{\mathbf{x}}_i,\widetilde{\mathbf{y}}_i)}\left[\kern-0.4em\left[\frac{1}{n}\sum_{i=1}^n \ell(\widetilde{\mathbf{y}}_i,\widetilde{\mathbf{x}}_i,W)-\ell(\mathbf{y}_i,\mathbf{x}_i,W) \right]\kern-0.4em\right] \right\} \right]\kern-0.4em\right]} \\
\leq & \mathop{\Exp\limits_{(\mathbf{x}_i,\mathbf{y}_i)}}\limits_{(\widetilde{\mathbf{x}}_i,\widetilde{\mathbf{y}}_i)} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\left\{\frac{1}{n}\sum_{i=1}^n \ell(\widetilde{\mathbf{y}}_i,\widetilde{\mathbf{x}}_i,W)- \frac{1}{n}\sum_{i=1}^n \ell(\mathbf{y}_i,\mathbf{x}_i,W) \right\} \right]\kern-0.4em\right] \\
= & \mathop{\Exp_{(\mathbf{x}_i,\mathbf{y}_i)}}\limits_{(\widetilde{\mathbf{x}}_i,\widetilde{\mathbf{y}}_i),\epsilon_i} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\left\{\frac{1}{n}\sum_{i=1}^n \epsilon_i\left(\ell(\widetilde{\mathbf{y}}_i,\widetilde{\mathbf{x}}_i,W)- \ell(\mathbf{y}_i,\mathbf{x}_i,W)\right) \right\} \right]\kern-0.4em\right]\\
\leq & \Exp_{(\mathbf{x}_i,\mathbf{y}_i),(\widetilde{\mathbf{x}}_i,\widetilde{\mathbf{y}}_i),\epsilon_i} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\left\{\frac{1}{n}\sum_{i=1}^n \epsilon_i\ell(\widetilde{\mathbf{y}}_i,\widetilde{\mathbf{x}}_i,W)\right\}\right]\kern-0.4em\right]\\
& +\Exp_{(\mathbf{x}_i,\mathbf{y}_i),(\widetilde{\mathbf{x}}_i,\widetilde{\mathbf{y}}_i),\epsilon_i} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\left\{\frac{1}{n}\sum_{i=1}^n \epsilon_i \ell(\mathbf{y}_i,\mathbf{x}_i,W) \right\} \right]\kern-0.4em\right]\\
= &2 \Exp_{(\mathbf{x}_i,\mathbf{y}_i),\epsilon_i} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\left\{\frac{1}{n}\sum_{i=1}^n \epsilon_i \ell(\mathbf{y}_i,\mathbf{x}_i,W) \right\} \right]\kern-0.4em\right] \\
= & 2\Exp_{(\mathbf{x}_i,\mathbf{y}_i),\epsilon_i} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\left\{\frac{1}{n}\sum_{i=1}^n \epsilon_i \sum_{j=1}^k \ell(\mathbf{y}_i^j,\mathbf{x}_i,\bm{w}_j) \right\} \right]\kern-0.4em\right]\\
\leq & \frac{2C}{n} \Exp_{(\mathbf{x}_i,\mathbf{y}_i),\epsilon_i} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\left\{\sum_{i=1}^n \epsilon_i \sum_{j=1}^k \langle\mathbf{x}_i,\bm{w}_j\rangle \right\} \right]\kern-0.4em\right] \\
= &\frac{2C}{n} \Exp_{\mathbf{x}_i,\epsilon_i} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\left\{\sum_{i=1}^n \sum_{j=1}^k \langle\epsilon_i\mathbf{x}_i,\bm{w}_j\rangle \right\} \right]\kern-0.4em\right] = 2C\mathcal{R}_n(\mathcal{W})
\end{split}
\end{equation*}
where $\epsilon_i(i=1,...,n)$ are the Rademacher variables. The first inequality utilizes the Jensen inequality and the last inequality is based on the assumption that the loss function $\ell$ is bounded and $C$-Lipschitz. Here we adopt a Redemacher complexity defined as follows:
\begin{equation*}
\begin{split}
\mathcal{R}_n(\mathcal{W}) &\triangleq \frac{1}{n} \Exp_{\mathbf{x}_i,\epsilon_i} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\left\{\sum_{i=1}^n \sum_{j=1}^k \langle\epsilon_i\mathbf{x}_i,\bm{w}_j\rangle \right\} \right]\kern-0.4em\right] \\
&= \frac{1}{n} \Exp_{\mathbf{X},\bm{\epsilon}} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\langle W,\mathbf{X}_{\bm{\epsilon}}\rangle \right]\kern-0.4em\right]
\end{split}
\end{equation*}
where $\mathbf{X}_{\bm{\epsilon}} \triangleq [\sum_{i=1}^n\epsilon_i\mathbf{x}_i,...,\sum_{i=1}^n\epsilon_i\mathbf{x}_i]$. Then in the last step, we directly calculate and estimate the Redemacher complexity.
\subsection{Estimating the Redemacher Complexity}
By applying the Cauchy-Schwarz inequality and the Triangle inequality of matrix norms, we obtain
\begin{equation*}
\begin{split}
&\frac{1}{n} \Exp_{\mathbf{X},\bm{\epsilon}} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\langle W,\mathbf{X}_{\bm{\epsilon}}\rangle \right]\kern-0.4em\right] \leq \frac{1}{n} \Exp_{\mathbf{X},\bm{\epsilon}} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\norm{W}_F\norm{\mathbf{X}_{\bm{\epsilon}}}_F \right]\kern-0.4em\right]\\
&= \frac{1}{n} \Exp_{\mathbf{X},\bm{\epsilon}} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}\norm{W-W_{old}S+W_{old}S}_F\norm{\mathbf{X}_{\bm{\epsilon}}}_F \right]\kern-0.4em\right]\\
& \leq \frac{1}{n} \Exp_{\mathbf{X},\bm{\epsilon}} \left[\kern-0.4em\left[ \sup_{W\in\mathcal{W}}(\norm{W-W_{old}S}_F+\norm{W_{old}S}_F)\norm{\mathbf{X}_{\bm{\epsilon}}}_F \right]\kern-0.4em\right]\\
&\leq \frac{\varepsilon+\norm{W_{old}S}_F}{n}\Exp_{\mathbf{X},\bm{\epsilon}} \left[\kern-0.4em\left[ \norm{\mathbf{X}_{\bm{\epsilon}}}_F \right]\kern-0.4em\right]
\leq \frac{\varepsilon+\norm{W_{old}S}_F}{n}\sqrt{\Exp_{\mathbf{X},\bm{\epsilon}} \left[\kern-0.4em\left[ \norm{\mathbf{X}_{\bm{\epsilon}}}_F^2 \right]\kern-0.4em\right]} \\
\end{split}
\end{equation*}
Then we calculate $\Exp_{\mathbf{X},\bm{\epsilon}} \left[\kern-0.4em\left[\norm{\mathbf{X}_{\bm{\epsilon}}}_F^2\right]\kern-0.4em\right]$ as,
\begin{equation*}
\begin{split}
\Exp_{\mathbf{X},\bm{\epsilon}} \left[\kern-0.4em\left[\norm{\mathbf{X}_{\bm{\epsilon}}}_F^2\right]\kern-0.4em\right]& = k\Exp_{\mathbf{X},\bm{\epsilon}} \left[\kern-0.4em\left[ \norm{\sum_{i=1}^n\epsilon_i\mathbf{x}_i}_2^2\right]\kern-0.4em\right] \\
&= k \Exp_{\mathbf{X},\bm{\epsilon}} \left[\kern-0.4em\left[ \sum_{i=1}^n\norm{\mathbf{x}_i}_2^2 + \sum_{i\neq j}\epsilon_i\epsilon_j\langle\mathbf{x}_i,\mathbf{x}_j\rangle\right]\kern-0.4em\right] \\
&= k\sum_{i=1}^n\Exp_{\mathbf{x}_i} [\kern-0.15em[\norm{\mathbf{x}_i}_2^2]\kern-0.15em] \leq kn
\end{split}
\end{equation*}
where the energy of feature $\Exp_{\mathbf{x}} [\kern-0.15em[\norm{\mathbf{x}}_2^2]\kern-0.15em]$ is assumed to be no more than 1 without loss of generality. Thus we can obtain an upper bound of the Redemacher complexity to establish Theorem \ref{theorem:1}:
\begin{equation*}
\begin{split}
\mathcal{R}_n(\mathcal{W}) &\leq \sqrt{\frac{k}{n}}(\varepsilon+\norm{W_{old}S}_F)\\
&\leq \sqrt{\frac{k}{n}}(\varepsilon+\norm{W_{old}}_F\norm{S}_{1,1})\\
&\leq \sqrt{\frac{k}{n}}(\varepsilon+\lambda\norm{W_{old}}_F).
\end{split}
\end{equation*}
\section{Proof of Theorem \ref{theorem:2}}\label{proof:2}
Now we detail the proof of Theorem \ref{theorem:2}, which aims to analyze the cost introduced by SLL, \emph{i.e.}, the perturbation of classifier parameter matrix (including both past labels and new labels) due to the SLL mechanism. Assuming the past label size is $m$ and a new label is given, the goal is to estimate the difference of the classifier matrix between the SLL and the original $m+1$ parameter matrix. Given the training data $\mathbf{X}=[\mathbf{x}_1,...,\mathbf{x}_n]\in\mathbb{R}^{d\times n}$ and their label matrix $\mathbf{Y}=[\mathbf{y}_1,...,\mathbf{y}_n]\in\{-1,1\}^{(m+1)\times n}$, the $m+1$-dimensional classifier parameter matrix is determined by the following optimization:
\begin{equation*}
\hat{Z} = \argmin_{Z\in\mathbb{R}^{d\times (m+1)}}J(Z) = \sum_{i=1}^n \ell(\mathbf{y}_i,\mathbf{x}_i,Z) + \frac{\lambda}{2}\norm{Z-ZS}_F^2,
\end{equation*}
where $S$ is the label structure matrix of all $m+1$ labels.
Denoting $\tilde{Z} = [\hat{W},\hat{\bm{w}}]$, we have $J(\tilde{Z})\geq J(\hat{Z})$, and substitute it into the expression of $J(Z)$, and then we have
\begin{multline*}
\sum_{i=1}^n \ell(\mathbf{y}_i,\mathbf{x}_i,\tilde{Z}) - \sum_{i=1}^n \ell(\mathbf{y}_i,\mathbf{x}_i,\hat{Z})\\ \geq \frac{\lambda}{2}\norm{\hat{Z}-\hat{Z}S}_F^2 - \frac{\lambda}{2}\norm{\tilde{Z}-\tilde{Z}S}_F^2.
\end{multline*}
Denote the approximation error as $\Delta = \hat{Z} - \tilde{Z} = \hat{Z} - [\hat{W},\hat{\bm{w}}]$. Then the right hand side of the above inequality can be rewritten as
\begin{equation*}
\begin{split}
\mbox{right} =& \frac{\lambda}{2}\norm{\tilde{Z}+\Delta - (\tilde{Z}+\Delta )S}_F^2 - \frac{\lambda}{2}\norm{\tilde{Z}-\tilde{Z}S}_F^2\\
=& \frac{\lambda}{2}\norm{\tilde{Z}-\tilde{Z}S}_F^2+\frac{\lambda}{2} \norm{\Delta-\Delta S}_F^2\\
&+ \lambda Tr[(\tilde{Z}-\tilde{Z}S)^T(\Delta-\Delta S)]-\frac{\lambda}{2}\norm{\tilde{Z}-\tilde{Z}S}_F^2\\
=& \frac{\lambda}{2} \norm{\Delta-\Delta S}_F^2 + \lambda \langle\Delta,(\tilde{Z}-\tilde{Z}S)(\mathbf{I}-S)^T\rangle
\end{split}
\end{equation*}
where $\langle\cdot,\cdot\rangle$ is the inner product of two matrices. Similarly, the left hand side can also be rewritten using the approximation error $\Delta$ and for simplicity we consider the least squares loss function,
\begin{equation*}
\begin{split}
\mbox{left}= & \frac{1}{2}\norm{\mathbf{Y} - \tilde{Z}^T\mathbf{X}}_F^2 - \frac{1}{2}\norm{\mathbf{Y} - \hat{Z}^T\mathbf{X}}_F^2 \\
=& \frac{1}{2}\norm{\mathbf{Y} - \tilde{Z}^T\mathbf{X}}_F^2 - \frac{1}{2}\norm{\mathbf{Y} - (\tilde{Z}+\Delta)^T\mathbf{X}}_F^2\\
=& \frac{1}{2}\norm{\mathbf{Y} - \tilde{Z}^T\mathbf{X}}_F^2 - \frac{1}{2}\norm{\mathbf{Y} - \tilde{Z}^T\mathbf{X}}_F^2\\
&- \frac{1}{2}\norm{\Delta^T\mathbf{X}}_F^2 + Tr[(\mathbf{Y} - \tilde{Z}^T\mathbf{X})^T\Delta^T\mathbf{X}]\\
=& - \frac{1}{2}\norm{\Delta^T\mathbf{X}}_F^2 + \langle\mathbf{X}(\mathbf{Y} - \tilde{Z}^T\mathbf{X})^T,\Delta\rangle
\end{split}
\end{equation*}
Thus we obtain
\begin{multline*}
\frac{\lambda}{2} \norm{\Delta-\Delta S}_F^2 + \frac{1}{2}\norm{\Delta^T\mathbf{X}}_F^2 \\
\leq \langle\mathbf{X}(\mathbf{Y} - \tilde{Z}^T\mathbf{X})^T,\Delta\rangle - \lambda \langle\Delta,(\tilde{Z}-\tilde{Z}S)(\mathbf{I}-S)^T\rangle
\end{multline*}
Suppose $d\ll n$ and $\mathbf{X}$ is of full row rank, and denote its smallest singular value as $\sigma_1(\mathbf{X})$, then based on the singular value decomposition, we have
\begin{multline*}
\frac{\lambda}{2} \norm{\Delta-\Delta S}_F^2 + \frac{1}{2}\norm{\Delta^T\mathbf{X}}_F^2\\ \geq \frac{\lambda}{2} \sigma_1^2(\mathbf{I}-S)\norm{\Delta}_F^2 + \frac{1}{2}\sigma_1^2(\mathbf{X})\norm{\Delta}_F^2
\end{multline*}
Thus
\begin{equation*}
\begin{split}
& \frac{1}{2}[\lambda\sigma_1^2(\mathbf{I}-S)+\sigma_1^2(\mathbf{X})]\norm{\Delta}_F^2\\
\leq& \langle\Delta,\mathbf{X}(\mathbf{Y} - \tilde{Z}^T\mathbf{X})^T-\lambda(\tilde{Z}-\tilde{Z}S)(\mathbf{I}-S)^T\rangle\\
\leq& \norm{\Delta}_F \norm{\mathbf{X}(\mathbf{Y} - \tilde{Z}^T\mathbf{X})^T-\lambda(\tilde{Z}-\tilde{Z}S)(\mathbf{I}-S)^T}_F\\
\leq& \norm{\Delta}_F \left(\norm{\mathbf{X}}_F\norm{\mathbf{Y} - \tilde{Z}^T\mathbf{X}}_F+\lambda \norm{\tilde{Z}}_F\norm{\mathbf{I}-S}_F^2\right)
\end{split}
\end{equation*}
Dividing both sides using $\norm{\Delta}_F $, we obtain
\begin{equation*}
\begin{split}
&\frac{1}{2}[\lambda\sigma_1^2(\mathbf{I}-S)+\sigma_1^2(\mathbf{X})]\norm{\Delta}_F\\
\leq& \norm{\mathbf{X}}_F\norm{\mathbf{Y} - \tilde{Z}^T\mathbf{X}}_F+\lambda \norm{\tilde{Z}}_F\norm{\mathbf{I}-S}_F^2\\
\leq& \sqrt{n\Omega}\norm{\mathbf{Y} - \tilde{Z}^T\mathbf{X}}_F+\lambda \norm{\mathbf{I}-S}_F^2 \sqrt{\norm{\hat{W}}_F^2+\norm{\hat{\bm{w}}}_2^2}
\end{split}
\end{equation*}
and it is equivalent to
\begin{multline*}
\norm{\Delta}_F\leq \frac{2}{\lambda\sigma_1^2(\mathbf{I}-S)+\sigma_1^2(\mathbf{X})}\cdot\\ \left(\sqrt{n\Omega}\norm{\mathbf{Y} - \tilde{Z}^T\mathbf{X}}_F+\lambda \norm{\mathbf{I}-S}_F^2 \sqrt{\norm{\hat{W}}_F^2+\norm{\hat{\bm{w}}}_2^2}\right),
\end{multline*}
which completes the proof of Theorem \ref{theorem:2}.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction}
\label{sec:intro}
The study of systems driven far from thermal equilibrium, also known as driven diffusive systems (DDS), has been at the forefront in statistical
mechanics in the last few decades \cite{chowdhury_book}.
These systems have found applications in understanding
transport in superionic conductors \cite{KLS1, KLS2}, protein synthesis
in prokaryotic cells \cite{gibbs_protein1, gibbs_protein2},
traffic flow \cite{traffic}, biophysical transport \cite{chou_bio_transport,frey_bio_transport}, etc. These systems
evolve under local stochastic dynamics and in the long time limit reach a non-equilibrium current carrying stationary state.
Certain surprising features of these non-equilibrium steady states have generated an overwhelming interest among researchers.
For example, these systems may exhibit spontaneous symmetry breaking \cite{ssb_dds}, boundary induced
phase transition \cite{krug_asep, popkov_schutz}, phase separation transition\cite{ABC_model1}, condensation
transition \cite{evans_braz, zrp_review} even
in one dimension.
\par In this work, we discuss reentrant condensation transition in a DDS involving two species
of interacting particles on a one dimensional periodic ring. A reentrant phase transition is said to occur
if by varying a certain parameter,
the system undergoes transition from one phase to another phase and finally reenters the initial phase. Such transitions have been
reported in a variety of equilibrium systems, for example in models of spin glasses \cite{reentrance_spin_glass} and multicomponent
liquid mixtures \cite{reentrance_liquid_mix}. Reentrant transition have also been reported in some non-equilibrium systems.
For example, Antal and Sch\"utz studied a system
of a driven non-equilibrium lattice gas of hard-core particles with next-nearest neighbor interaction in one
dimension \cite{schutz}, where
for attractive interactions, a reentrant transition between high density (HD) phase and maximal current(MC)
phase was observed. A few reaction-diffusion systems \cite{dickman_pcpd,odor_pcpd1} having a
competing dynamics between
diffusion and particle interaction have been observed to show a similar reentrant phase behavior$-$ where
the transition is from absorbing-to-active-to-absorbing phases.
In biological systems, reentrant transitions have been reported experimentally in protein and DNA solutions in presence of
multivalent metal ions \cite{zhang_protein_reentrance}.
Such physical phenomena are vital for understanding biological processes like
protein crystallization and DNA condensation. Reentrant phase behavior has also been observed in driven colloidal
systems \cite{hagan_colloids2} and force induced DNA unzipping transitions \cite{bhattacharjee_DNA_unzipping}.
\par The two species driven diffusive model discussed here was first introduced in \cite{our_model}, where the phase separation transition
in a model of reconstituting $k$-mers can be studied by mapping the model to the two species box particle system. Here
we report that this model can show a reentrant condensation transition where one of the species undergoes fluid-condensate-fluid transition
when the interaction parameter between the two species is varied.
\par The article is organized in the following way: In section 2 we define the model and write down its product measure steady state.
In the next section we study the system in the grand canonical ensemble for a specific choice of diffusion rates
and show that condensation transition in one of the species show a reentrant behavior. In section 4 we compare the reentrant
behavior of this model with that in some single species models and finally we summarize the results in section 5 and conclude
with some discussions.
\begin{figure}
\centering \includegraphics[height=5cm]{model.eps}
\caption{ Dynamics of the two-species box-particle model: The diffusion rate, $u(m,n)$ of species $A$ (red-circles) in general can
depend on the
number of particles
of both the species at that site. Exchange process of species $B$ (blue-circles) can occur only if both the
sites are devoid of the species $A$.
The rate of this process depends on the number of particles of the species $B$ at the two participating sites.
}
\label{reentrant2}
\end{figure}
\section{The model }
\label{sec:model}
We study condensation transition in a box-particle model on a one dimensional periodic lattice of $L$ sites or boxes
labeled as $i=1,2,\cdots,L$
and containing two species of particles, say $A$ and $B$. At a given site $i$, let $m_i$ and $n_i$ be the
number of particles of species $A$ and $B$ respectively. A typical configuration of the model is represented as
$\ {C}= \{ m_1,n_1;m_2,n_2 \cdots m_i,n_i\cdots m_L,n_L\}=\{ m_i,n_i\}$. The dynamics of the model
is the following: from a randomly chosen site $i$, a particle of species $A$ can hop to site $i+1$ with rate $u(m_i,n_i)$ that
in general depends on the number of particles of
both $A$ and $B$. On the other hand, particles of species $B$ can be transferred between two neighboring boxes
$i$ and $i+1$ with rates that depend on the number of particles at both arrival and departure sites. However, in this case
there is an additional restriction that the boxes exchanging particles of species $B$ are devoid of
particles of the species $A$. This condition is crucial for explicit factorization
of the steady state \cite{our_model}. Depending on whether a $B$-type particle is
transferred from site $i$ to $i+1$ or vice-versa, we define $w(n_i, n_{i+1})$ or $w(n_{i+1}, n_{i})$
to be the corresponding rates for the dynamics of species $B$. The above dynamics, also
shown in Figure~\ref{reentrant2} can be
represented in the following way:
\begin{equation}
\{ \cdots m_i,n_i;~m_{i+1},n_{i+1}\cdots\} \xrightarrow{u(m_i,n_i)} \{\cdots m_i-1,n_i;~m_{i+1}+1,n_{i+1}\cdots \},
\label{dyn1}
\end{equation}
\begin{equation}
\{\cdots 0,n_i;~0,n_{i+1} \cdots\} \xrightarrow{w(n_i,n_{i+1})} \{\cdots 0,n_i-1;~0,n_{i+1}+1 \cdots\}
\label{dyn2}
\end{equation}
and
\begin{equation}
\{\cdots 0,n_i;~0,n_{i+1}\cdots\} \xrightarrow{w(n_{i+1},n_i)} \{\cdots 0,n_i +1;~0,n_{i+1}-1 \cdots\}.
\label{dyn3}
\end{equation}
Note that the dynamics conserves the total number of particles of each species.
In this article we would limit to the
situation where $u(m,n) \equiv u(n)$, thus the rate of diffusion of species $A$
depends on the number of particles of species $B$. Such a choice
makes the dynamics of the species $A$ comparable to that of a disordered zero range process \cite{inhomo_zrp},
the background disorder being created by the presence of species $B$ evolve with time. A two
species zero-range process \cite{2sp_zrp1, gross_spohn_03, 2sp_zrp2} has been
previously studied where one of the species obeys a dynamics similar to that of species $A$ in this model. The dynamics of the species $B$ is similar
to that of a misanthrope process \cite{misanthrope_original, misanthrope} and hence
the rate of diffusion depends on the number of particles in both departure and the arrival site. It is necessary to mention here that
in the model of reconstituting $k$-mers, particles of species $A$ corresponds to $0$-particles
and that of species $B$ corresponds to $k$-particles \cite{our_model}.
\par The steady state of the model has a product measure so that the probability
of finding the system in an arbitrary configuration, $\ {C}= \{ m_i,n_i \}$
can be written in the following factorised form:
\begin{equation}
P(\{m_i,n_i\}) = \frac{1}{Q_{M,N}^L} \prod_{i=1}^L f(m_i,n_i) \delta\left(\sum_{i=1}^L m_i-M\right)\delta\left(\sum_{i=1}^L n_i-N\right),
\label{steady_state}
\end{equation}
where the partition function $Q_{M,N}^L$ is given by
\begin{equation}
Q_{M,N}^L = \sum_{\{ m_i\}, \{n_i\}}\prod_{i=1}^Lf(m_i,n_i)
\delta\left(\sum_{i=1}^L m_i-M\right)\delta\left(\sum_{i=1}^L n_i-N\right).
\label{partition_func}
\end{equation}
The $\delta$-functions appearing above make sure that
only those configurations are to be summed over which
the number of particles $M$ and $N$ of species $A$ and $B$ are conserved. Therefore
the densities, $\rho_A=M/L$ and $\rho_B=N/L$ of the two species $A$ and $B$ do not change as
the system evolves over time. Here, we consider factorized rate for species $B$,
\begin{equation}
w(m,n)= w_1(m) w_2(n).
\label{rate1}
\end{equation}
This choice makes the steady
state factorised and the weight factors appearing in Eq.~\eqref{steady_state} are then given by \cite{our_model}
\begin{equation}
f(m,n) = u(n)^{-m} \prod_{i=1}^n\frac{w_2(i-1)}{w_1(i)}.
\label{prod_measure}
\end{equation}
In the next section, we study the nature of condensation transition of the system for certain specific choice of rates.
\section{Condensation and reentrant transition}
\label{sec:cond_transition}
A salient feature of box-particle systems is that they can show condensation transition \cite{zrp_review, evans_pair, misanthrope}. For such systems,
condensation occurs when the
density $\rho$ is larger than a finite critical density $\rho_c$. In the condensed phase there exists
a background critical fluid consisting of $\rho_c L$ particles,
and a condensate carrying $(\rho -\rho_c)L$ particles, $L$ being the system size. For certain choices of diffusion rates, the two
species model discussed here
also shows condensation of one or both the species which can be established by studying the system in the grand
canonical ensemble (GCE) \cite{our_model}. In GCE, let $z$ and $x$ be the fugacities corresponding to the species $A$ and $B$ respectively; then
the grand canonical partition function, $Z_L(z,x)$ obtained from the
canonical partition function (Eq.~\eqref{partition_func}) is given by
\begin{equation}
Z_L(z,x)= \sum_{M=0}^{\infty} \sum_{N=0}^{\infty} z^M x^N Q_{M,N}^L= F(z,x)^L,
\label{grand_part_func}
\end{equation}
where
\begin{equation}
F(z,x)= \sum_{m=0}^{\infty} \sum_{n=0}^{\infty} z^m x^n f(m,n).
\label{grand_f_func}
\end{equation}
Let $z \le z_c$ and $x \le x_c$ define the domain of fugacities for which $F(z,x)<\infty$ and thus the grand canonical measure is valid within this domain.
In GCE, densities are given by
\begin{eqnarray}
\rho_A(z,x)=\frac{z}{F}\frac{\partial F}{\partial z} ~~ {\rm and}~~
\rho_B(z,x)=\frac{x}{F}\frac{\partial F}{\partial x}~.
\label{grand_eq_den}
\end{eqnarray}
If one or both densities become finite
as critical values of fugacities are approached then the system is in a condensate carrying phase. In such a phase a system cannot
accommodate densities that are larger than the ones set by the critical limit and therefore undergoes a condensation transition.
Here, we discuss one of the rates for which the condensation transition of species $A$ shows a reentrant behavior.
Let us set the rate of diffusion of the species $A$ as a step function,
\begin{equation}
u(n) = \left\{\begin{array}{ll}
v ~~~&{\rm for }~ n < k\cr
1~~~ &{\rm otherwise}~.
\end{array}
\right.
\label{diff_A_species}
\end{equation}
Here $k<\infty$ can take only positive integer values and $u(0)=0$. We also consider $w_1$ and $w_2$ to differ by a constant \cite{misanthrope},
\begin{equation}
w_1(n)= \frac{n+2}{n+1} ~~;~~w_2(n)= \frac{n+2}{n+1}-\alpha,
\label{diff_B_species2}
\end{equation}
where $0<\alpha<1$. Defining $\sigma= (3-2 \alpha)/(1-\alpha)$, we characterise the rates of diffusion of
species $A$ and $B$ by three parameters, namely $v$, $k$ and $\sigma$. Using Eq.~\eqref{prod_measure}, we then obtain
\begin{eqnarray}
f(m,n)=\frac{1}{[u(m)]^n}\frac{n!}{(\sigma)_n}(n+1)^2,
\label{eq:fk0}
\end{eqnarray}
where $(\sigma)_n=\sigma(\sigma+1)\cdots(\sigma+n-1)$ is the Pochhammer symbol.
The partition function in the GCE, following Eqs.
\eqref{grand_part_func} and \eqref{grand_f_func}, is $Z_L(z,x)= F(z,x)^L$, with
\begin{equation}
F(z,x)= \frac{z(1-v) G_{k -1 }(x) + (v-z) G_{\infty}(x)}{(1-z)(v-z)},
\label{f_func_choice}
\end{equation}
where
\begin{equation}
G_{k}(x)=\sum_{n=0}^{k} x^n \frac{n!}{(\sigma)_n}(n+1)^2.
\end{equation}
From Eq.~\eqref{f_func_choice}, it is evident that the maximum value of the
fugacities for which $F(z,x)<\infty$ are $z_c=$min$ \{1,v\} $ and $x_c=1$. The densities of the two species
in GCE obtained from Eq.~\eqref{grand_eq_den} are given by
\begin{eqnarray}
\rho_A(z,x)&=& \frac{z}{1-z} + \frac{ z v(1-v)G_{k-1}(x)}{(v-z)[z(1-v) G_{k-1}(x) + (v-z) G_\infty(x)] }\cr
\nonumber \\
{\rm and~}\nonumber\\
\rho_B(z,x) &=& x \frac{z( 1-v) G_{k-1}'(x) + (v-z) G'_\infty(x)}{z ( 1-v) G_{k-1}(x) + (v-z) G_\infty(x)}\nonumber\\.
\label{eq:densities}
\end{eqnarray}
Here, prime ($'$) indicates derivative with respect to $x.$ It has been shown in \cite{our_model} that as $x \to x_c (=1)$, $\rho_B$
may become finite but as $z \to z_c$, $\rho_A$ always diverge. This means that in absence
of $A$-type particles, $i.e.$ when $z=0$, species $B$ can condense but $A$ cannot condense if there are no particles of species $B (x=0)$ in the
system. Therefore
the critical line is given by $x=1$.
\begin{figure}
\centering \includegraphics[height=5.5cm]{f1.eps}
\includegraphics[height=5.5cm]{f2.eps}
\caption{Phase diagram in the $\rho_B$$-$$\rho_A$ plane corresponding to the critical line $x=1$ for (a) $v=2(v>1~regime)$
and (b) $v=0.5(v<1~regime)$. In the inset we chose
the lines $z<z_c$ and plot it in the corresponding density plane. For $v>1$,
the slope of the line PQR is positive and it is zero for $v<1$. The background critical
density is given by point $Q$ \cite{Gross08}.
Here $\sigma=10$ and $k=5.$ }
\label{phase_diag}
\end{figure}
\par In absence of particles of species $A$, the critical density
is $\rho_B^c=4/(\sigma-5)$ and thus condensation of species $B$ occurs when $\sigma>5$ \cite{misanthrope}.
However, when the density of species $A$ is not equal to zero ($z>0$ case), $(\rho_A^c,\rho_B^c)$
depends on all three parameters, $\sigma, k ~{\rm and}~ v.$ For a given value of $k$, the nature of
condensation is found to be different for the regimes $v<1$ and $v>1$ \cite{our_model}. For $\rho_A>0$, the background
critical densities can be found out from the grand canonical measure by using the Gro\ss kinsky theorem \cite{Gross08}. It was
proved in \cite{Gross08} that the normal directions of the critical line in
$\mu_x$$-$$\mu_z$ plane (here chemical potentials are $\mu_{x} = \ln(x)$ and $\mu_{z} = \ln(z)$)
translates to a direction in density plane along which the background density remain invariant.
For the model under study, the critical line is $x=1$ ({\it i.e.} $\mu_x=0$) and thus the normals are defined by
$z=constant$. For illustration, we take $z=0.9 (<z_c=1)$ for $v=2$ and $z=0.4 (<z_c=v)$ for $v=0.5$ in Figure~\ref{phase_diag} and plot the
corresponding line in $\rho_B$$-$$\rho_A$ plane. It approaches the critical point $Q \equiv (\rho_B^c, \rho_A^c)$
with a slope given by the tangent line PQR. The Gro\ss kinsky criteria indicate that the background critical density
along the line QR in the condensate phase is invariant and is given by the point Q. When $v>1$, for any arbitrary
density $(\rho_B, \rho_A)$ on the line QR, $\rho_B>\rho_B^c$ and $\rho_A>\rho_A^c$; both species would therefore have
extra particles which would form a condensate. When $v<1$, the slope of the line PQR is always $zero$, regardless
of the value of $z$ and therefore only species $B$ would form a condensate.
\begin{figure}[h]
\centering \includegraphics[height=6.5cm]{fig_phase.eps}
\caption{Phase diagram of the model in $v-\rho_B$ plane for $k=5$, $\sigma=10$, and $\rho_A=1.5$. The red dashed
line, $v=1$, separates Region I where species $B$ condense and Region II where both $A$ and $B$ can condensate. The blue curve
separates the fluid region (Region III) from regions I and II (also see text). }
\label{reentrant_rB_v_phase_diag}
\end{figure}
\par For $v=1$, from Eq. \eqref{diff_A_species} it is clear that the dynamics of the two species gets decoupled and therefore
species $B$ may condense with $\rho_B^c= 4/(\sigma-5)$ but
species $A$ do not have a condensate. The line $v=1$ thus separates {\it only } species $B$ condensing phase region from other regions.
In Figure~\ref{reentrant_rB_v_phase_diag}, we plot the variation of $\rho_B$ with respect to the coupling parameter $v$ for fixed values
of $\rho_A,~k$ and $\sigma$. Clearly, in the phase diagram, there are three regions describing different phases:
\vskip .1cm
\noindent {\bf Region~I:} Here $v<1$, species $B$ has a condensate but species $A$ remains fluid. However, from Figure~\ref{phase_diag}, it is clear
that for $v<1$, the liquid phase of
$A$ at point R has a density $\rho_A^c$ and hence it remains critical.
\vskip .1cm
\noindent {\bf Region~II:} For intermediate values of $v$, both A and B would condense.
\vskip .1cm
\noindent {\bf Region~III:} For higher values of $v$, the system remains fluid and neither of the two species has a
condensate. Here the fluid of both species is not critical, and is physically different from Region I, where $A$ is a critical fluid.
\vskip .1cm
\par Thus, in the phase diagram (Figure~\ref{reentrant_rB_v_phase_diag}), as we move along $v$-axis, we find
that the system passes successively through regions where species $A$ is fluid-condensate-fluid thereby showing a reentrant feature. For species $B$
the transition is from condensate to fluid phase only, and thus the reentrant behavior does not exist.
\section{Reentrant condensation transition in some single species models}
\label{sec:other_models}
\subsection{Zero-range process}
\label{sec:zrp_reentrant}
Let us consider a zero range process \cite{evans_braz, zrp_review}, on a periodic ring with $L$ boxes or sites and $N$
number of particles. In ZRP, a particle can hop from a site to its neighboring site with a rate that depends only on the
occupation number of the departure site. Each box $i$ can be occupied by any number of particles, but there is an overall
conservation of particles since particles cannot be created or
destroyed. If $n_i$ is the number of particles at an arbitrary site $i$, then $C$=$\{n_1,n_2, \cdots\}=\{n_i\}$ represents
a typical configuration of the system. The particle conserving dynamics is
defined as following: from a randomly chosen site $i$, a particle can jump asymmetrically towards its right site $i+1$ with rate $u(n_i)$
that depends only on the number of particles at $i$.
The steady state for the ZRP can be determined exactly \cite{zrp_review} and can be written in the form of a product measure:
\begin{equation}
P(C)= \frac{1}{Q_{N}^L}\prod_{i=1}^L f(n_i) \delta \left(\sum_{i=1}^L n_i - N\right); ~~~f(n)=\prod_{i=1}^n \frac{1}{u(i)},
\label{eq:zrp1}
\end{equation}
where $Q_{N}^L=\sum_{\{ n_i \}} P(\{n_i\}) \delta \left(\sum_{i=1}^L n_i - N\right)$ is the canonical partition function
and the delta-function ensures that the total number of particles are conserved.
Let us make the following choice of rates
\begin{equation}
u(n)=\frac{n+ a b}{n+ a b^2}.
\label{eq:zrp2}
\end{equation}
Here $0<b<1$ and $a>0$. Note that in the asymptotic limit, where $n \to \infty$, the above equation reads as
\begin{equation}
u(n)= 1+ \frac{\gamma}{n} + O\left(\frac{1}{n^2} \right),
\end{equation}
where $\gamma = a b (1-b)$. It is well known \cite{zrp_review} that in ZRP, with rates having the above functional form,
there is a condensation transition for large densities (for $\rho$ larger than a critical value $\rho_c$) if $\gamma>2$.
Thus, we expect a condensation transition here if $ab(1-b)>2$. However, the critical density $\rho_c$ depends on both $a$ and $b$ and not just
on $\gamma=ab(1-b)$. To calculate $\rho_c$ exactly, we proceed to evaluate the partition function $Z_L(z)$ in the GCE, to obtain,
\begin{equation}
Z_L(z)=F(z)^L,~~~~{\rm with}~~~F(z)= \sum_{n=0}^\infty z^n f(n),
\end{equation}
$z$ being the fugacity. The density in GCE is $\rho(z)= z \frac{F'(z)}{F(z)}$, and the critical density as obtained by following the same
algebraic steps as in \cite{zrp_review} is given by
\begin{equation}
\rho_c= \lim_{z \rightarrow 1} \rho(z)= \lim_{z \rightarrow 1} \frac{z F^{'}(z)}{F(z)}= \frac{1+a b^2}{a b(1-b)-2}.
\label{crit_zrp}
\end{equation}
\iffalse
\begin{equation}
F(z)= \sum_{n=0}^\infty z^n f(n)=\sum_{n=0}^{\infty} z^n \frac{(1+ ab^2)_n (1)_n}{(1+ab)_n n!},
\end{equation}
$(c)_n$ being the Pochhammer symbol. The grand canonical density in this case is then given by $\rho(z)= z \frac{F'(z)}{F(z)}$.
The critical density is then given by,
\begin{equation}
\rho_c= \lim_{z \rightarrow 1} \rho(z)= \lim_{z \rightarrow 1} \frac{z F^{'}(z)}{F(z)}= \frac{1+a b^2}{a b(1-b)-2}.
\label{crit_zrp}
\end{equation}
\fi
From Eq.~\eqref{crit_zrp}, it follows that for any given $a$, the critical density has a minimum at
$b_m=\left(1+\sqrt{1+a}\right)/a$, where $\rho_c=$ $\rho_c^m=2\left(3+\sqrt{1+a}\right)/(a-8)$. Condensation can thus occur only
if $\rho>\rho_c^m$. Further, for any $\rho>\rho_c^m$, the reentrant transition points are given by
\begin{equation}
b_{\mp}=\frac{a\rho \mp \sqrt{a^2 \rho^2-4a(\rho+1)(2\rho+1)}}{2a(\rho+1)}.
\end{equation}
In Figure~\ref{reentrant_zrp}, we plot the critical density $\rho_c$ as a function of $b$ for $a=10$.
Clearly, as $b$ is increased, the system
passes successively from fluid to condensate phase at $b_{-}$ and from condensate to fluid phase at $b_{+}$, thereby showing
a reentrant behavior.
\begin{figure}[htbp]
\centering \includegraphics[height=6cm]{f6.eps}
\caption{ Critical line equation for $a=10$ having a minimum at $b_m$ is shown. As $b$ is increased, the system
passes successively through fluid-condensate-fluid phase
and hence it shows reentrant behavior. The reentrant transition points $b_{\mp}$ for $\rho=8$ are also shown. }
\label{reentrant_zrp}
\end{figure}
\par Note that from Eq. \eqref{crit_zrp}, it is obvious that $\rho_c \sim 1/(\gamma-2)$. This indicates that reentrant transition cannot be
observed in $\gamma-\rho$ plane as $\rho_c$ is a monotonic decreasing function of $\gamma$ (for $\gamma>2$). Thus, in this case, the reentrant
behavior we observe is only because $\gamma$ is quadratic in $b$.
\iffalse
is The minimum value of Condensation occurs when $\rho>\rho_c$.
In Figure~\ref{reentrant_zrp}, we plot the critical density $\rho_c$ as a function of $b$ for $a=10$ and show the corresponding
phase diagram. We find that as $b$ is increased, the system
passes successively through fluid-condensate-fluid phase, thereby showing a reentrant behavior. In this case, the reentrant behavior can be understood
simply because $\gamma$ is quadratic in $b$.}
\fi
\subsection {Models with pair-factorized steady states}
\label{sec:pair_reentrant}
Let us discuss another box-particle model of one species where the steady state has an exact pair-factorized form.
The model \cite{evans_pair, waclaw_pair2} involves biased diffusion
of particles on a one dimensional periodic lattice.
Here, a particle hops from a randomly chosen site $i$ to site $i+1$ with rate $u(n_{i-1},n_i,n_{i+1})$ that depends on the
number of particles at site $i-1$, $i$ and $i+1$. The probability of finding the system in an arbitrary configuration
$C=\{n_i\}$ can be expressed as a product of pair-factorized weights over consecutive sites:
\begin{equation}
P(C)=\frac{1}{Q_{N}^L} \prod_{i=1}^L g(n_i,n_{i+1}) \delta\left(\sum_{i=1}^L n_i-N\right),
\label{pair_eq:1}
\end{equation}
where $Q_{N}^L=\sum_{\{ n_i \}} P(\{n_i\}) \delta \left(\sum_{i=1}^L n_i - N\right)$ is the canonical
partition function. The $\delta$-function ensures
that the total number of particles are conserved. It can be proved \cite{evans_pair, waclaw_pair2} that the steady state master
equation is satisfied if the rate of diffusion satisfies the following condition:
\begin{equation}
u(n_{i-1},n_i,n_{i+1})= \frac{g(n_{i-1},n_i-1) g(n_i-1,n_{i+1})} {g(n_{i-1},n_i) g(n_i,n_{i+1})}.
\label{pair_eq:2}
\end{equation}
We study a particular model with a pair factorized steady state of the form \cite{evans_pair, waclaw_pair2},
\begin{equation}
g(m,n)= \exp\left[(-J+ a U)|m-n|+ \frac{U}{2}(\delta_{m,0}+\delta_{n,0})\right],
\label{pair_eq:3}
\end{equation}
where $a\geq 0$. The corresponding hop rate of particle diffusion
from site $i$ to site $i+1$ obtained from Eq. \eqref{pair_eq:2} is
\begin{equation}
u(n_{i-1},n_i,n_{i+1})=
\begin{cases}
\exp[2a U-2J+U \delta_{n,1}]&\text{if }n_i \leq n_{i-1},n_{i+1}\\
\exp[2 J-2a U +U \delta_{n,1}]&\text{if }n_i > n_{i-1},n_{i+1}\\
\exp[U \delta_{n,1}] & \text{otherwise}.
\end{cases}
\label{pair_eq:4}
\end{equation}
\begin{figure}[htbp]
\centering \includegraphics[height=5.5cm]{f4.eps}
\includegraphics[height=5.5cm]{f5.eps}
\caption { Phase diagram for $J=2$. (a) No reentrance for $a=0$. As we move along the $U$-axis the system makes transition
from fluid phase to condensate phase.
(b) Reentrance for $a=1$. In this case, as $U$ is increased
the system pass through through fluid-condensate-fluid phase and hence the transition is reentrant in nature.}
\label{pair_figure}
\end{figure}
From Eq.~\eqref{pair_eq:4} it follows that if the number of particles at the diffusing site is smaller than the number
of particles in the neighboring sites, the rate of diffusion is low if $J>aU$ and high if $J<aU$. Conversely,
if the number of particles at the departure site is greater than the number of particles at neighboring sites, the rate of diffusion is
low $J<aU$ and high if $J>aU$. Thus the two factors $e^{2 J}$ and $e^{2 aU}$ compete with each other in determining
the evolution of the system. Additionally, a site can get rid of a single particle with rate $e^U$.
Note that when $a=0$, there is no competition and the model is same as
studied in \cite{evans_pair, waclaw_pair2}. For such a model phase transition from a fluid phase to condensate carrying phase
do not show reentrant behavior.
However, when
$a>0$, we find that the system shows a reentrant behavior due to the presence of the competing terms in the interaction.
In GCE, taking $z$ as the fugacity, one can proceed with the same algebraic
steps as in \cite{waclaw_pair2} to find that in the limit $z\rightarrow 1$, the critical density is given by
\begin{equation}
\rho_c = \frac{e^{J_0-aU}-1}{(e^{J_0-aU}- e^{-2(J-J_0)})\times(e^{2(J-J_0)}-1)},
\label{pair_crit_den}
\end{equation}
where
\begin{equation}
J_0=U + a U -ln (e^U -1).
\label{pair_eq:6}
\end{equation}
Note that the critical density, $\rho_c$ is finite provided that $J>J_0$. In Figure~\ref{pair_figure}
we take $J=2$ and plot the phase diagram and the critical density (Eq.~\eqref{pair_crit_den}) for two different
values of the parameter $a$. When $a=0$, the system shows a transition from fluid phase to a condensate carrying phase.
For $a>0$, the transition is of the nature fluid-condensate-fluid and hence shows a reentrant behavior.
\par We would now like to point out certain differences in reentrant transitions in the models discussed above. Firstly, note that although reentrant
transition can be simply observed in some single species interacting particle systems, our model provides a clear example where reentrant transitions
can occur by tuning the interaction parameter between the two species. Secondly, in Region I of the phase diagram (Figure~\ref{reentrant_rB_v_phase_diag}),
species $A$ always remains a critical fluid. This nature is different
from transition in single species models discussed above where the fluid phases are never critical.
\section{Summary and conclusion}
\label{sec:summary}
In summary, we have studied the condensation transition in a box-particle system
of two species $A$ and $B$ respectively. The rate of diffusion of the $A$ species from a given site to its next site depends
on the number of particles of the species $B$ at that site. The dynamics of the species $B$ is a nearest neighbor exchange process,
where by two adjacent sites can exchange particles of the species $B$ among each other. This process can
occur only when are two boxes participating in exchange of particles of the species $B$ are devoid of particles of
$A$ species. We have studied the nature of condensation transition as the interaction parameter that couples the dynamics
of the two species is varied and found that the condensation in species $A$ has a reentrant feature. We have also provided two examples
of single species driven diffusive systems which show reentrant condensation.
\par A condensation transition in a box-particle system corresponds to a phase separation transition in the
corresponding lattice model \cite{psep_criterion}. Therefore the phase separation transition in the corresponding lattice map
\cite{our_model} of the model studied here is also expected to show a reentrant behavior. Finally, we think that it would be interesting
to find out if reentrant phase behavior could also be observed in a more general class of DDS,
namely the finite-range processes which provide cluster factorised steady states \cite{finite_range}.
\section{Acknowledgements}
\label{sec:acknowledgements}
We thank P. K. Mohanty for many useful discussions, critical comments and careful reading of the manuscript. We thankfully acknowledge the referees
for pointing out important references and useful suggestions.
|
\section{Introduction}
\IEEEPARstart{V}{isual} data can be corrupted due to sensory noise or interferential outliers during data acquisition. A fraction of image pixels could be missing sometimes, if the conditions deteriorate further, missing pixels will constantly increase, \emph{e.g.}, terribly-damaged images due to unstable online transmission, cameras covered by noises, photographs overexposed accidently, and outdoor pictures taken behind a screen window. For satisfactory visual recognition, detection, and tracking, corrupted visual data must be recovered during pre-processing~\cite{ijcai13,nuclearnorm,softthreshold,tensor13,MC10,ijcai15}
There has recently been a surge in low rank-based matrix completion methods for visual recovery. Given a matrix $M\in \mathcal{R}^{n \times m}$, where $n$ and $m$ are the width and height of $M$, respectively, ${M_{i,j}=0,\ (i,j) \notin \Omega}$ denote the observed data while the others represent missing data. In general, a low rank matrix $\hat{M}$ can be discovered to approximately represent matrix $M$. The rank of $\hat{M}$ is usually assumed to be lower than any of its two dimensionalities, \emph{i.e.}, $rank(\hat{M})\ll \min(n,m)$.
Although rank minimization provides an approach for recovering the missing observations, it is computationally intractable (NP-hard and non-convex). Under broad conditions, \cite{nuclearnorm} reported that the rank function minimization can be replaced by the trace norm $||X||_* = \sum_k \sigma_k\left(X\right)$ minimization, where $\sigma_k$ is the $k$-$th$ maximum singular value of $X$. Soft-thresholding methods~\cite{softthreshold} are often employed to solve optimization problems with this trace norm regularization. However, trace norms have been improved to better investigate the low rank constraint. For example, \cite{tnuclearnorm} designed a truncated nuclear norm more suitable for matrix completion and~\cite{WNNM} proposed a weighted nuclear norm minimization for image denoising.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.95\linewidth]{./Intro}
\end{center}
\caption{A schematic of the calculation of the proposed DCT norm at two scales. The input image is first densely divided into several $p_i \times p_i$ patches that are then converted to the frequency domain. The DCT norm is returned by accumulating squared values of high-frequency coefficients whose positions are larger than $q_i$ in all the patches.}
\label{Fig:Intro}
\end{figure}
Existing rank minimizing techniques are effective and have delivered promising performance in many visual recovery problems. However, most of these algorithms have only been evaluated for data with at most 80\% missing values~\cite{tensor13}. They may not, therefore, be applicable to data with extremely high numbers of missing values (\emph{e.g.}, 95\%) that occur when data collection devices fail or there is damage to the transmission medium. This extreme visual recovery problem with a very large number of missing pixel values is, therefore, very challenging due to the conflict between the small amount of observed data and the tremendous amount of data that needs to be recovered.
As well as the low rank assumption, smoothness is another important property carried by visual data. The total variation (TV) norm~\cite{TVnorm} is an effective way to exploit smoothness by calculating the differences between neighboring pixels and has been widely used in image processing. However, minimizing local differences via the TV norm risks over-smoothing the image detail and texture~\cite{NLM}.
We consider employing the Discrete Cosine Transformation (DCT~\cite{DCT}) to devise a new smoothness regularizer called the ``DCT norm''. By constraining the frequency coefficients of the data obtained via DCT, the DCT norm flexibly adjusts the degree of smoothness over the data (Fig.\ref{Fig:Intro}). To investigate the local smoothness and multi-scale properties of visual data, we extend the DCT norm to derive local and multi-scale versions, respectively. Our theoretical analysis shows that the DCT norm includes the TV norm as a special case. By combining the classical rank minimization principle with the proposed DCT norm minimization, the resulting model can simultaneously exploit the low rank and smoothness properties of visual data for extreme visual recovery. Experimental results on real-world datasets demonstrate that the proposed DCT norm is highly effective and that the low rank and smoothness issues need to be integrated for successful extreme visual recovery.
\section{Related works}
We first briefly introduce related works on visual completion and TV norm minimization.
Nuclear norm minimization-based matrix completion~\cite{nuclearnorm} is formulated as
\begin{equation}
\hat{X} = \arg\min_X ||X||_*,\quad s.t.\ X_\Omega = M_\Omega,
\label{Fcn:lowrank}
\end{equation}
where $\hat{X}$ is the optimal estimation of $M$. The truncated nuclear norm minimization problem~\cite{tnuclearnorm} is defined as:
\begin{equation}
\hat{X} = \arg\min_X ||X||_r,\quad s.t.\ X_\Omega = M_\Omega,
\label{Fcn:tlowrank}
\end{equation}
where $||X||_r = \sum_{i=r+1}^{\min(n,m)}\sigma_i(X)$ is the sum of $\min(n,m)-r$ minimum singular values.
It is widely known that the TV norm~\cite{TVnorm} is an efficient smoothness regularization that accumulates all the gradients of a given image $X$:
\begin{equation}
||X||_{TV} = \sum_{i,j} \sqrt{|X_{i+1,j}-X_{i,j}|^2+|X_{i,j+1}-X_{i,j}|^2},
\label{Fcn:TVnorm}
\end{equation}
where $i$ and $j$ denote the vertical and horizontal positions of $X$, respectively. Since the TV norm accumulates gradient modules of the entire image, minimizing the TV norm can result in a smooth estimation:
\begin{equation}
\hat{X} = \arg\min_X ||X||_{TV}, \quad s.t.\ X_\Omega = M_\Omega.
\label{Fcn:minTV}
\end{equation}
\cite{jointTV} proposed to simultaneously use the TV and nuclear norms for image recovery:
\begin{equation}
\hat{X} = \arg\min_X||X||_*+\lambda||X||_{TV}, \quad s.t.\ X_\Omega = M_\Omega,
\label{Fcn:joint}
\end{equation}
where $\lambda$ balances the two norms. To the best of our knowledge, Fcn.\ref{Fcn:joint} is the first algorithm to integrate smoothness and nuclear norm regularization.
Since the $||\cdot||_{TV}$ is isotropic and not differentiable, an anisotropic version is proposed in~\cite{aTVnorm} that is easier to minimize:
\begin{equation}
||X||_{anisoTV}=\sum_{i,j}|X_{i+1,j}-X_{i,j}|+|X_{i,j+1}-X_{i,j}|.
\end{equation}
Additionally, in~\cite{LTV}, a modified linear total variation was defined as:
\begin{equation}
||X||_{LTV} = \sum_{i,j}|X_{i+1,j}-X_{i,j}|^2+|X_{i,j+1}-X_{i,j}|^2,
\end{equation}
which leads to a smooth low-rank matrix completion problem:
\begin{equation}
\hat{X} = \arg\min_X||X||_*+\lambda||X||_{LTV}, \quad s.t.\ X_\Omega = M_\Omega.
\label{Fcn:LTVnuclear}
\end{equation}
An ADMM-like optimization scheme~\cite{ADMM} can be adopted to solve Fcn.\ref{Fcn:LTVnuclear}. However, the traditional TV norm only can guarantee an estimation presenting a locally smooth visualization, when in reality a natural image should be smoothed at every scale. In the next section, we propose a multi-scale DCT norm in the frequency domain.
\section{The DCT norm}
A natural image has smooth regions in the spatial domain. Furthermore, there is less high-frequency information than low-frequency information in the image's frequency domain~\cite{DCTdistribution}. This section presents a novel Discrete Cosine Transformation (DCT) norm for smoothing the objective variable, which provides advantages over the TV norm due to its linear and convex properties.
\subsection{The DCT for a 2D matrix}
DCT is widely used in image compression and is an approximate KL transformation~\cite{DCT}. For an arbitrary matrix $X$, its DCT coefficient matrix $\mathcal{C}$ is:
\begin{equation}
\mathcal{C}_{X} = \mathcal{T}\left(X\right) = CXC^{\mathbf T} = \left(C \otimes C\right)vec(X),
\label{Fcn:DCT}
\end{equation}
where $\otimes$ is the Kronecker product, $vec(\cdot)$ denotes the vectorization, and $C$ denotes the transformation matrix with the same size as $X$:
\begin{equation}
C_{ij} = \alpha\left(i\right) cos\left[\frac{(j+0.5)\pi}{N}i\right],
\end{equation}
and
\begin{equation}
\alpha\left(i\right) = \left\{
\begin{aligned}
& \sqrt{1/N}, \ \ i =0\\
& \sqrt{2/N}, \ \ else.
\end{aligned}
\right.
\end{equation}
$\mathcal{C}$ has the same dimensionality with $X$, where $\mathcal{C}_{0,0}$ is the DC (direct current) coefficient which only consists of the overall illumination information~\cite{DC0}, the other coefficients in $\mathcal{C}$ are AC (alternating current) components denote the energies of every frequency levels, \emph{i.e.} the weights of the DCT blocks as showed in Fig.\ref{Fig:Intro}(c). Additionally, it is instructive to note that DCT is a linear lossless transformation and the original data can be restored by $X=\mathcal{T}^{-1}\left(\mathcal{C}_X\right)=\left(C \otimes C\right)^T \mathcal{C}_X$. DCT is also applied to a variety of computer vision tasks, such as image denoising~\cite{BM3D} and image representation~\cite{DCTface}.
Since Fcn.\ref{Fcn:DCT} involves all elements in $X$ we can access the overall structural information of the entire matrix. The DCT and its gradient can be quickly calculated using linear transformations such that the proposed DCT norm-based optimization problems can be efficiently and easily solved.
\subsection{The DCT norm for smoothing}
\textbf{The global smooth DCT norm.} Neighboring pixels in a natural image are generally significantly correlated. On the other hand, the abnormal signal (\emph{e.g.}, noise, missing values, \emph{etc.}) can be seen as a set of external data subject to an i.i.d. distribution. Hence, the frequency distributions of natural images and the abnormal signals are distinct (Fig.\ref{Fig:Intro}). The high-frequency information of the original image is much lower than that of the corrupted observation $M$. Based on this observation, we design a DCT norm in the frequency domain:
\begin{equation}
\begin{aligned}
||X||_{DCT}^q &= ||S_q*\left(CXC^{\mathbf T}\right)||_F^2\\
&= ||vec(S_q)*\left(C\otimes C vec(X)\right)||^2,
\end{aligned}
\label{Fcn:DCTglobal}
\end{equation}
where $*$ denotes the Hadamard product, $||\cdot||_F$ denotes the Frobenius norm, and $S_c$ is a selection mask with parameter $c$ denoting the cut-off position:
\begin{equation}
S_q = \left\{ \begin{array}{l}
S_{ij} = 1, \quad i\leq q\ \&\ j\leq q \\
S_{ij} = 0, \quad otherwise.
\end{array} \right.
\end{equation}
The smoothing-oriented visual recovery problem can thus be formulated as:
\begin{equation}
\hat{X} = \arg\min_X ||X||_{DCT}^q, \quad s.t.\ X_\Omega = M_\Omega.
\label{Fcn:globalsmooth}
\end{equation}
Although we can obtain a smooth estimation by solving the above problem, the optimal solution of Fcn.\ref{Fcn:globalsmooth} will have some deformations as shown in Fig.\ref{Fig:exp}. This is due to some remaining frequency coefficients with positions lower than $q$ still needing to be discarded since they were difficult to estimate. A local smoothness regularization can eliminate these deformations. Therefore, we expand the global smoothed DCT norm to a more comprehensive model that can also represent the local smoothness of the given image.
\noindent \textbf{The locally smooth DCT norm.} Inspired by non-local denoising methods~\cite{NLM,WNNM}, we divide the corrupted observation into several small patches or the locally smooth DCT norm:
\begin{equation}
\begin{aligned}
||X||_{DCT}^{p,q} &= \sum_l ||S_{p,q}*\left(C_p x_p^{(l)} C_p^{\mathbf T}\right)||_F^2\\
&= ||\mathbf{S}_{p,q}*\left(C_p \otimes C_p \mathbf{X}_p\right)||_F^2,
\end{aligned}
\label{Fcn:DCTlocal}
\end{equation}
where $p$ is the scale parameter, $||X||_{DCT}^{p,q}$ will be equivalent to Fcn.\ref{Fcn:DCTglobal} given $p = N$, $S_{p,q}$ and $C_p$ denote the mask and the transformation matrix generated according to p, respectively, $x_p^{(l)}$ is a $p \times p$ matrix denoting the $l$-$th$ patch extracted from $X$, $\mathbf{X}_p = \left[vec(x_p^{(1)}), vec(x_p^{(2)}), ..., vec(x_p^{(l)})\right]$ stacks $x_p^{(l)}$ into a matrix, and $\mathbf{S}_{p,q} = [vec(S_{p,q}),...,vec(S_{p,q})]$, which has the same dimensionality to that of $\mathbf{X}$. Patches extracted at every pixel are overlapping. For an $N\times M$ image, there are $(N-p+1)\times(M-p+1)$ patches with size $p\times p$.
The resulting optimization function w.r.t. the local DCT norm is:
\begin{equation}
\hat{X} = \arg\min_X ||X||_{DCT}^{p,q}, \quad s.t.\ X_\Omega = M_\Omega.
\label{Fcn:localmin}
\end{equation}
\noindent \textbf{The multi-scale DCT norm.} Both the locally smooth DCT norm and the global smooth DCT norm have pros and cons and can be further refined. Inspired by research on the local descriptor~\cite{SIFT,multiregion}, detecting and describing a key point based on multiple scales can capture more geometric information. Hence, we propose to integrate local and global smoothness by integrating DCT norms from multiple image scales:
\begin{equation}
\hat{X} = \arg\min_X \sum_{i=1}^{s} ||X||_{DCT}^{p_i,q_i}, \quad s.t.\ X_\Omega = M_\Omega.
\label{Fcn:multiscale}
\end{equation}
where $s$ indicates the number of scales. If $s = 1$ and $p = N$, Fcn.\ref{Fcn:multiscale} will focus on global smoothness. Given $s = 1$ and $p = 2$, Fcn.\ref{Fcn:multiscale} will be reduced to the local DCT norm minimization problem in Fcn.\ref{Fcn:DCTlocal}.
\subsection{Relationship to the TV Norm}
The TV norm is an efficient smoothing tool (see Fcn.\ref{Fcn:TVnorm}). The following theorem suggests that the TV norm can be regarded as a special case of DCT norm, thereby demonstrating the superiority of the proposed DCT norm.
\begin{myTheo}
Given $X \in \mathbb{R}^{n \times m}$, the optimal solution of Fcn.\ref{Fcn:minTV} with the linear TV norm is exactly that of the DNM problem in Fcn.\ref{Fcn:localmin} with $p$ = 2 and $q$ = 1.
\end{myTheo}
\begin{proof}
For any $2\times 2$ patch $x=[x_{0,0},x_{1,0},x_{0,1},x_{1,1}]^\mathbf{T}$ in $\mathbf{X}$, its DCT coefficient matrix is $\mathcal{C} = \mathcal{T}(x)$ and the local DCT norm is
\begin{equation}
||x||_{DCT}^{2,1} = \mathcal{C}_{0,1}^2+\mathcal{C}_{1,0}^2+\mathcal{C}_{1,1}^2,
\end{equation}
where
\begin{equation}
\begin{aligned}
\mathcal{C}_{0,1} &= \frac{1}{2}\left((x_{0,0}-x_{0,1})+(x_{1,0}-x_{1,1}) \right),\\
\mathcal{C}_{1,0} &= \frac{1}{2}\left((x_{0,0}-x_{1,0})+(x_{0,1}-x_{1,1}) \right),\\
\mathcal{C}_{1,1} &= \frac{1}{2}\left((x_{0,0}+x_{1,1})-(x_{0,1}+x_{1,0}) \right),
\end{aligned}
\end{equation}
and the TV norm of $X$ is calculated as:
\begin{equation}
\begin{aligned}
||x||_{LTV} &= (x_{00}-x_{01})^2+(x_{00}-x_{10})^2\\
&+(x_{10}-x_{11})^2+(x_{01}-x_{11})^2.
\end{aligned}
\end{equation}
Let $f(x) = 4||x||_{DCT}^{2,1}$, $e_1=[1,1,-1,-1]^\mathbf{T}$, $e_2=[1,-1,1,-1]^\mathbf{T}$, and $e_3=[1,-1,-1,1]^\mathbf{T}$. We have
\begin{equation}
\begin{aligned}
f(x)&=\left(e_1^\mathbf{T}x\right)^2+\left(e_2^\mathbf{T}x\right)^2+\left(e_3^\mathbf{T}x\right)^2\\
&=x^\mathbf{T}e_1e_1^\mathbf{T}x+x^\mathbf{T}e_2e_2^\mathbf{T}x+x^\mathbf{T}e_3e_3^\mathbf{T}x\\
&=\sum_{i=1}^3 x^\mathbf{T}e_ie_i^\mathbf{T}x,
\end{aligned}
\end{equation}
Similarly, let $g(x) = ||x||_{LTV}$, $d_1=[1,0,-1,0]^\mathbf{T}$, $d_2=[1,-1,0,0]^\mathbf{T}$, $d_3=[0,1,0,-1]^\mathbf{T}$, and $d_4=[0,0,1,-1]^\mathbf{T}$. We have
\begin{equation}
g(x)=\sum_{i=1}^4 x^\mathbf{T}d_id_i^\mathbf{T}x,
\end{equation}
Consider two minimizations $\min f(x)$ and $\min g(x)$. Their optimal solutions satisfy:
\begin{equation}
\begin{aligned}
\triangledown_x f(x)&=2\sum_{i=1}^3e_ie_i^\mathbf{T}x= \mathcal{E}x=0\\
\triangledown_x g(x)&=2\sum_{i=1}^4d_id_i^\mathbf{T}x= \mathcal{D}x=0,
\end{aligned}
\label{Fcn:nullspace}
\end{equation}
where $\mathcal{E} = 2\sum_{i=1}^3e_ie_i^\mathbf{T}$, $\mathcal{D}=2\sum_{i=1}^4d_id_i^\mathbf{T}$ and their null spaces are equivalent $\mathcal{N}(\mathcal{E}) = \mathcal{N}(\mathcal{D})$.
Furthermore, it is easy to demonstrate the null spaces of $\mathcal{E}$ and $\mathcal{D}$ are identical for any $N \times M$ image by formatting its $\mathcal{E}$ and $\mathcal{D}$ according to Fcn.~\ref{Fcn:nullspace}. Hence, the optimal solution of DNM is equal to that of the TV norm minimization.
\end{proof}
According to Theorem 1, we conclude that $||x||_{DCT}^{2,1}$ is equivalent to the TV norm. Moreover, we suggest that the proposed approach in Fcn.\ref{Fcn:multiscale} has benefits over that of the TV norm since there is flexible to control over the degree of smoothness by operating at multiple scales.
\begin{figure}[h]
\centering
\includegraphics[width=0.95\linewidth]{./remove}
\caption{A comparison of the DCT norm and TV norm in a removal problem. Left to right show corrupted observations and reconstructed images using the TV norm and DCT norm, respectively. }
\label{Fig:inpainting}
\end{figure}
The TV norm is noted for its strong capacity for image inpainting, \emph{e.g.}, text removal~\cite{tnuclearnorm}. This is a challenging task, because the pixels covered by the text are not randomly distributed. Hence, the optimization function with a regularization that encourages global smoothness might be expected to perform well. Since color images have three channels (\emph{i.e.}, red, green, and blue), for fair comparison with the TV norm we simply deal with each channel separately and combine them to obtain the final result. As illustrated in Fig.\ref{Fig:inpainting}, the visual recovery results obtained with the proposed DCT norm are superior to those of the TV norm. The results show that the proposed DCT norm outperforms the TV norm because it takes all pixels into consideration. Here, we use a two-scale DCT norm with $p_1 = 2$, $p_2 = 512$ (image sizes), $q_1 = 1$, and $q_2 = 256$. Obviously, the peak signal-to-noise ratio (PSNR) values of the proposed DCT norm are about 2\emph{dB} higher than that of the TV norm.
\section{DCT norm for visual recovery}
Image completion aims to restore an original image $X$ from its corrupted observation $M$ with an observed region $\Omega$. Recovering missing values in a matrix with limited observed information has recently attracted considerable interest~\cite{tensor13,MC10,tnuclearnorm}.
This problem is commonly addressed with inpainting~\cite{tnuclearnorm} or denoising~\cite{WNNM} methods, especially when the missing ratio is not too high ($\leq 80\%$). The non-local means [et al.2005] and its variations~\cite{WNNM,SAIST} are the current state-of-the-art techniques and exploit the self-similarity characteristic of images. However, when the observation is very corrupt, \emph{e.g.}, the missing ratio is higher than $90\%$ (see Fig.\ref{Fig:exp}), non-local based algorithms are less useful since they cannot find similar patches in the given image.
\begin{table*}[t]
\begin{minipage}{0.5\textwidth}
\centering
\normalsize
\caption{PSNR Comparison (Unit: $dB$) }
\label{Tab:PSNR}
\vspace{0.7em}
\begin{tabular}{c||c|c|c|c|c}
\hline
$\phi$ & SVT & TNNR & LTVNN & SAIST & DNM \\
\hline \hline
90\% & 19.299 & 19.947 & 26.134 & 25.928 & \textbf{28.865}\\
\hline
95\% & 16.984 & 14.125 & 23.408 & 23.682 & \textbf{26.182}\\
\hline
98\% & 15.149 & 7.543 & 20.289 & 21.301 & \textbf{23.377}\\
\hline
99\% & 14.313 & 5.458 & 18.245 & 19.605 & \textbf{21.704}\\
\hline
\end{tabular}
\end{minipage}
\begin{minipage}{0.5\textwidth}
\centering
\normalsize
\caption{SSIM Comparison}
\label{Tab:SSIM}
\vspace{0.7em}
\begin{tabular}{c||c|c|c|c|c}
\hline
$\phi$ & SVT & TNNR & LTVNN & SAIST & DNM \\
\hline \hline
90\% & 0.7170 & 0.6970 & 0.9125 & 0.8876 & \textbf{0.9472}\\
\hline
95\% & 0.6300 & 0.4281 & 0.8647 & 0.8470 & \textbf{0.9161}\\
\hline
98\% &0.5134 & 0.1063 & 0.7917 & 0.7989 & \textbf{0.8712}\\
\hline
99\% &0.4291& 0.0544 & 0.7275 & 0.7624 & \textbf{0.8391}\\
\hline
\end{tabular}
\end{minipage}
\end{table*}
Base on the proposed multi-scale DCT norm and existing rank minimization techniques, we establish an efficient optimization problem that contains the various factors mentioned above:
\begin{equation}
\begin{aligned}
\hat{X} =& \arg\min_X||X||_r+\sum_i \lambda_i ||X||_{DCT}^{p_i,q_i}\\
&+\frac{\gamma}{2}||\mathcal{P}_{\Omega}(X)-\mathcal{P}_{\Omega}(M)||_F^2,
\end{aligned}
\label{Fcn:ObjFcn}
\end{equation}
where $\lambda_i$ denotes the weighting parameter for the DCT norm $||X||_{DCT}^{p_i,q_i}$ of $X$ in the $i$-$th$ scale. $\gamma>0$ is a relaxation factor that converts the original problem into an unconstrained minimization~\cite{tnuclearnorm}. The relaxed problem can be solved using a gradient-based algorithm~\cite{Gradient}.
Fcn.\ref{Fcn:ObjFcn} is naturally designed for the recovery of gray images. However, it is easy to extend Fcn.\ref{Fcn:ObjFcn} to handle color images:
\begin{equation}
\begin{aligned}
\hat{X} =& \sum_c \arg\min_X||X_{(c)}||_r+\sum_c \sum_i \lambda_i ||X_{(c)}||_{DCT}^{p_i,q_i}\\
&+\alpha\sum_c ||X_{(c)}||_{freq}+\frac{\gamma}{2}||\mathcal{P}_{\Omega}(X)-\mathcal{P}_{\Omega}(M)||_F^2,
\end{aligned}
\label{Fcn:colorObjFcn}
\end{equation}
where $X$ is an RGB image and $X_{(c)}$ denotes the $c$-$th$ channel of $X$, \emph{i.e.}, $X_{(1)}$, $X_{(2)}$, $X_{(3)}$ corresponding to the red, green, and blue channels, respectively. An additional regularization with constant weight $\alpha$ is employed in Fcn.\ref{Fcn:colorObjFcn} to encourage DCT coefficients in any two channels of $X$ to be similar:
\begin{equation}
||X_{(c)}||_{freq} = \frac{1}{4}\sum_{i\neq c}||S^c *C\left(X_{(c)}-X_{(i)}\right)C^\mathbf{T}||_F^2,
\end{equation}
where $S^c_{i,j} = 0\ if\ i = 0\ and\ j = 0,\ S^c_{i,j} = 1\ otherwise$ is a mask to remove the DC coefficient because the illumination information between channels is different. Although the value of $||X_{(c)}||_{freq}$ is equal to that calculated in the spatial domain after subtracting the average value of each channel, the calculation in the frequency domain will assign greater weights to low-frequency coefficients. This is important for constructing images of better visual quality. The gradient of the DCT norm is:
\begin{equation}
\begin{aligned}
\mathcal{G}&\left(||X_{(c)}||_{DCT}\right) = \frac{\partial{\sum_i \lambda_i ||X_{(c)}||_{DCT}^{p_i,q_i}}}{\partial{X_{(c)}}}\\
& = \sum_i \lambda_i \mathcal{I} \left( \left(C_{p_i} \otimes C_{p_i}\right)^\mathbf{T} \left[ S_{p_i,q_i} * \left(C_{p_i}\otimes C_{p_i} \mathbf{X}_{p_i} \right) \right] \right),
\end{aligned}
\label{Fcn:DCTgradient}
\end{equation}
where $\mathcal{I}(\cdot)$ is an operation that recovers the stacking matrix $\mathbf{X}_{p_i}$ into the original image and $\mathcal{I}(\mathbf{X}_{p_i}) = X_{(c)}$. The gradient of $||X_{(c)}||_{freq}$ is
\begin{equation}
\begin{aligned}
\mathcal{G}\left(||X_{(c)}||_{freq}\right) &= \frac{\partial{ ||X_{(c)}||_{freq}}}{\partial{X_{(c)}}}\\
= \sum_{i\neq c} &C^\mathbf{T}\left(S^c *C\left(X_{(c)}-X_{(i)}\right)C^\mathbf{T}\right)C.
\end{aligned}
\end{equation}
An iterative TV norm~\cite{iterTV} has also been proposed, and its optimal estimation is superior to that of directly solving the TV norm of the input observation. In this paper, we borrow this iterative strategy~\cite{SAIST,WNNM} and extend the DCT norm to an iterative method:
\begin{equation}
M^{(k+1)}=M^{(k)}+\delta \mathcal{P}_{\Omega}\left(M-X^{(k)} \right),
\label{Fcn:iter}
\end{equation}
where $k$ is the iteration number and $\delta$ is a relaxation parameter (often set to 0.1). We describe the image completion method by exploiting the proposed multi-scale DCT norm as showed in Alg.\ref{Alg:Completion}.
\begin{algorithm}[h]
\caption{DCT norm minimization for visual recovery.}
\label{Alg:Completion}
\begin{algorithmic}[1]
\REQUIRE The corrupted observation $M^0 = M$, $\hat{X}^0 = M$ and the corresponding parameters;
\REPEAT
\STATE $X^k \leftarrow M^k$;
\REPEAT
\FOR{each scale $p_i$}
\STATE Divide $X^k$ into several $p_i \times p_i$ patches;
\STATE Accumulate the gradient of $X^k$ in the $p_i$ scale;
\ENDFOR
\STATE Calculate the gradient, except the nuclear norm;
\STATE Obtain the low-rank approximation;
\UNTIL{convergence}
\STATE $M^{(k+1)} \leftarrow M^{(k)}+\delta \mathcal{P}_{\Omega} \left(M-X^{(k)} \right)$;
\UNTIL{$||M^{(k+1)}-X^{k}||_F \leq \epsilon$}
\ENSURE The estimated image $\hat{X}$;
\end{algorithmic}
\end{algorithm}
\section{Experiments}
\label{Sec:Exp}
\textbf{Experimental setup.} Experiments were carried on eight images~\cite{NCSR} widely used for evaluating the performance of image restoration algorithms (Fig.\ref{Fig:test}). This is a benchmark dataset widely used for evaluating the visual recovery performance, with a variety of scenes.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.95\linewidth]{./tests}
\end{center}
\caption{The test images.}
\label{Fig:test}
\end{figure}
Observations were generated by randomly sampling a small proportion (ranging from 1\% to 10\%) of pixels from the images subject to a Gaussian distribution. We used two standard criteria to evaluate the recovery performance: PSNR and structural similarity (SSIM)~\cite{SSIM}.
\begin{figure*}[t]
\centering
\includegraphics[width=1\linewidth]{./exp}
\caption{The completion results for various corrupted images from top to bottom and their missing pixel rates: Girl (90\%), Butterfly (95\%), Parthenon (98\%), and Plants (99\%). Corrupted observations are shown in the first column and the reconstructed images using SVT, TNNR, LTVNN, SAIST, G-DNM, L-DNM, and the proposed DNM are shown from left to right. DNM results are more visually pleasing than those of previous state-of-the-art methods.}
\label{Fig:exp}
\end{figure*}
\noindent \textbf{Parameter settings.} The proposed completion algorithm has several important parameters: $p_i$, $q_i$ and $\lambda_i$. In the experiments, $p_i$ was set as 2, 8, and 512 (the size of each image) to obtain a better smoothness estimation and $q_1 = 1, q_2 = 4, q_3 = 192$, which denote the locally, blockly, and globally smooth regularizations, respectively. The weight parameter $\alpha$ was set to $10^{-3}$, $\gamma$ was set to 0.5, and $\lambda_i=\lambda = 1.5 \times 10^{-2}$ empirically. For the DNM stop conditions, we set the tolerance $\epsilon$ in Alg.\ref{Alg:Completion} to $10^{-8}$. The parameter $r$ for truncating singular values was set to 192, which is much larger than that in~\cite{tnuclearnorm}. Note that, the proposed algorithm is not sensitive to most of parameters, thus all of the parameters were set empirically which is a common setting in the context of low-rank minimization~\cite{tnuclearnorm}.
\noindent \textbf{Experimental results.} We conducted experiments using different image completion algorithms to compare the performance of the proposed DCT norm with current methods. Therefore, completion experiments were carried using the proposed DNM and state-of-the-art comparison algorithms: SVT~\cite{softthreshold}, TNNR~\cite{tnuclearnorm}, LTVNN~\cite{LTV}, and SAIST~\cite{SAIST}. Most of the comparative methods were conducted using the source code provided by authors. SVT is reported as the baseline since it is the cornerstone of the low rank approach, and TNNR is an enhanced low rank completion algorithm that is known to produce better results. LTVNN embeds the linear TV norm into the low rank minimization. SAIST is a non-local method that improves the estimated results after taking an interpolation as its initial value.
Since TNNR will be degenerated into SVT while the parameter $r$ in $||X||_r$ is set to 0, we set $r = 5$ in TNNR to compare the results of TNNR and SVT. It is interesting to note that TNNR's estimation is better than that of SVT when the missing rate $\phi$ is less than 90\%, while it is inferior to SVT when $90 < \phi$. This is because TNNR retains the largest $r$ singular values, which depict the major structure of the input image, but are most likely reflect the structure of the corrupted image with a considerable missing rate; hence, TNNR is inefficient. Since the proposed DCT norm can conductively recover image structure, we set a larger threshold $r = 192$ in DNM to avoid undesired lines caused by excessive minimization of the nuclear norm. Results are showed in Tab.\ref{Tab:PSNR} and Tab.\ref{Tab:SSIM}. The values are the average of the eight test images, where the highest evaluation result in each case is highlighted in bold. The proposed visual recovery algorithm based on the DCT norm clearly outperforms the others.
We also performed a qualitative comparison of the different completion algorithms. Results on different scenes (\emph{Girl}, \emph{Butterfly}, \emph{Parthenon}, \emph{Plants}) and missing ratios between 90\% and 99\% are shown in Fig.\ref{Fig:exp}. We also report the results exploiting the global and local DCT norms, marked as G-DNM and L-DNM in the figure. G-DNM produces some uncoordinated grids due excessive reduction of necessary high-frequency information, and the results of L-DNM are similar to that of the TV norm. Since images produced by DNM are smooth and natural, they have a better perceptual quality. It is obvious that the result of G-DNM has a clearer overall structure, while its local regions are not smooth enough. The result of L-DNM presents an opposite phenomenon. The multi-scale DNM achieves the best performance by combining them.
The SVT results contain several lines because this algorithm slightly undermines the image structure when shrinking all singular values; therefore, the TNNR estimation is more visually pleasing when the missing pixel rate is not too high. Although the results of LTVNN and SAIST are better than those of the previous two algorithms, LTVNN is not particularly clear and SAIST over-smooths some important structures. Specifically, SAIST post-processes after initializing the image using an interpolation method, which only completes the missing pixels by exploiting neighborhood information; thus, it lacks overall structure. Furthermore, SAIST is useless when the input corrupted image does not possess the initialization produced using interpolation methods. The estimated images of the proposed algorithm are clear, sharp, and visually pleasing because the multi-scale DCT norm makes the image and its patches at different scales smooth and natural. Specifically, it obtains an estimation in which the neighborhood pixels inside are smoothed using the locally smooth DCT norm and it can also obtain an estimation that produces a globally smooth output by exploiting the globally smooth DCT norm.
\section{Discussion and conclusions}
Most existing rank-minimizing techniques do not efficiently handle data with over 90\% missing values. Therefore, we propose a powerful smooth regularization to overcome this problem: the DCT norm. Compared to the traditional TV norm, the proposed scheme involves all the pixel values and can guarantee estimation smoothness at different scales. Moreover, we demonstrate that the TV norm can be regarded as a special case of the DCT norm. By combining the truncated nuclear norm and the proposed scheme we establish an efficient image completion model. Experiments show that the estimated images using the proposed multi-scale DCT norm are more visually pleasing than those produced by the previous state-of-the-art. Additionally, the proposed smooth regularization can be independently embedded into most image processing tasks, \emph{e.g.}, image inpainting and image denoising.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\renewcommand\refname{References}
{\small
\bibliographystyle{ieee}
|
\section{Introduction}
\setlength{\epigraphrule}{0pt}
\epigraph{\hspace{1.2in}{\em Consistency is all I ask!}}{{\sc Tom Stoppard}}
In the past couple of years, deep learning has swept though computer vision like wildfire.
One needs only to buy a GPU, arm oneself with enough training data, and turn the crank to see head-spinning improvements on most computer vision benchmarks.
So it is all the more curious to consider tasks
for which deep learning has {\em not} made much inroad, typically due to the lack of easily obtainable training data.
One such task is {\em dense visual correspondence} -- the problem of estimating a pixel-wise correspondence field between images depicting visually similar objects or scenes. Not only is this a key ingredient for optical flow and stereo matching, but many other computer vision tasks, including recognition, segmentation, depth estimation, etc. could be posed as finding correspondences in a large visual database followed by label transfer.
In cases where the images depict the same physical object/scene across varying viewpoints, such as in stereo matching, there is exciting new work that aims to use the commonality of the scene structure as supervision to learn deep features for correspondence~\cite{agrawal2015learning,deepstereo,dinesh2015,han2015matchnet,vzbontar2015stereo}.
But for computing correspondence {\em across different object/scene instances}, no learning method to date has managed to seriously challenge SIFT flow~\cite{liu2011sift}, the dominant approach for this task.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{teaser_final.pdf
\caption{Estimating a dense correspondence flow field $F_{r_1,r_2}$ between two images $r_1$ and $r_2$ --- essentially, where do pixels of $r_1$ need to go to bring them into correspondence with $r_2$ --- is very difficult. There is a large viewpoint change, and the physical differences between the cars are substantial. We propose to {\em learn} to do this task by training a ConvNet using the concept of cycle consistency in lieu of ground truth. At training time, we find an appropriate 3D CAD model to establish a correspondence 4-cycle, and train the ConvNet to minimize the discrepancy between $\tilde{F}_{s_1\!, s_2}$ and $F_{s_1\!, r_1}\!\circ\!F_{r_1\!, r_2}\!\circ\!F_{r_2\!, s_2}$, where $\tilde{F}_{s_1\!, s_2}$ is known by construction. At test time, no CAD models are used.}
\label{fig:quartet}
\end{figure}
How can we get supervision for dense correspondence between images depicting different object instances, such as images $r_1$ and $r_2$ in Figure~\ref{fig:quartet}?
Our strategy in this paper is to learn the things we don't know by linking them up to the things we do know. In particular, at training time, we use a large dataset of 3D CAD models~\cite{shapenet} to find one that could link the two images, as shown in Figure~\ref{fig:quartet}. Here the dense correspondence between the two views of the same 3D model $s_1$ and $s_2$ can serve as our ground truth supervision (as we know precisely where each shape point goes when rendered in a different viewpoint), but the challenge is to use this information to train a network that can produce correspondence between two real images at test time.
A naive strategy is to train a network to estimate correspondence between the rendered views of the same 3D model, and then hope that the network could generalize to real images as well. Unfortunately, this does not work in practice (see Table~\ref{tab:pck}), likely due to 1) the large visual difference between synthetic and real images and 2) the lack of cross-instance ground truth correspondence for training. Instead, in this paper we utilize the concept of {\em cycle consistency} of correspondence flows~\cite{huang2013consistent,zhou2015flowweb,zhou2015multi} -- the notion that the composition of flow fields for any circular path through the image set should have a zero combined flow. Here, cycle consistency serves as a way to link the correspondence between real images and the rendered views into a single 4-cycle chain. We can then train our correspondence network using cycle consistency as the supervisory signal. The idea is to take advantage of the known synthetic-to-synthetic correspondence as ground-truth anchors that allow cycle consistency to propagate the correct correspondence information from synthetic to real images, without diverging or falling into a trivial solution. Here we could interpret the cycle consistency as a kind of ``meta-supervision'' that operates not on the data directly, but rather on how the data should behave. As we show later, such 3D-guided consistency supervision allows the network to learn cross-instance correspondence that potentially overcomes some of the major difficulties (e.g. significant viewpoint and appearance variations) of previous pairwise matching methods like SIFT flow~\cite{liu2011sift}. Our approach could also be thought of as an extension and a reformulation of FlowWeb~\cite{zhou2015flowweb} as a learning problem, where the image collection is stored implicitly in the network representation.
The main contributions of this paper are: 1) We propose a general learning framework for tasks without direct labels through cycle consistency as an example of ``meta-supervision''; 2) We present the first end-to-end trained deep network for dense cross-instance correspondence; 3) We demonstrate that the widely available 3D CAD models can be used for learning correspondence between 2D images of different object instances.
\section{Related work}
\noindent{\bf Cross-instance pairwise correspondence}
~~The classic SIFT Flow approach~\cite{liu2011sift} proposes an energy minimization framework that computes dense correspondence between different scenes by matching SIFT features~\cite{sift} regularized by smoothness and small displacement priors. Deformable Spatial Pyramid (DSP) Matching~\cite{kim2013deformable}, a recent follow-up to SIFT Flow, greatly speeds up the inference while modestly improving the matching accuracy. Barnes~\etal~\cite{barnes2010generalized} extend the original PatchMatch~\cite{barnes2009patchmatch} algorithm to allow more general-purpose (including cross-instance) matching. Bristow~\etal~\cite{bristow2015dense} build an exemplar-LDA classifier around each pixel, and aggregate the matching responses over all classifiers with additional smoothness priors to obtain dense correspondence estimation. In these same proceedings, Ham~\etal~\cite{ham2016} take advantage of recent developments in object proposals, and utilize local and geometric consistency constraints among object proposals to establish dense semantic correspondence.
\\\\
\noindent{\bf Collection correspondence}~~Traditionally, correspondence has been defined in a pairwise manner, but recent works have tried to pose correspondence as the problem of joint image-set alignment. The classic like on work on Congealing~\cite{learned2006data,huang2007unsupervised} uses sequential optimization to gradually lower the entropy of the intensity distribution of the entire image set by continuously warping each image via a parametric transformation (e.g. affine). RASL~\cite{peng2012rasl}, Collection Flow~\cite{kemelmacher2012collection} and Mobahi ~\etal~\cite{mobahi2014compositional} first estimate a low-rank subspace of the image collection, and then perform joint alignment among images projected onto the subspace. FlowWeb~\cite{zhou2015flowweb} builds a fully-connected graph for the image collection with images as nodes and pairwise flow fields as edges, and establishes globally-consistent dense correspondences by maximizing the cycle consistency among all edges. While achieving state-of-the-art performance, FlowWeb is overly dependent on the initialization quality, and scales poorly with the size of the image collection. Similar to a recent work on joint 3D shape alignment~\cite{huang2013consistent}, Zhou~\etal~\cite{zhou2015multi} tackle the problem by jointly optimizing feature matching and cycle consistency, but formulate it as a low-rank matrix recovery which they solve with a fast alternating minimization method. Virtual View Networks~\cite{carreira2014virtual} leverage annotated keypoints to infer dense correspondence between images connected in a viewpoint graph, and use this graph to align a query image to all the reference images in order to perform single-view 3D reconstruction. Cho~\etal~\cite{cho2015unsupervised} use correspondence consistency among selective search windows in a diverse image collection to perform unsupervised object discovery.
\\\\
\noindent{\bf Deep learning for correspondence}~~Recently, several works have applied convolutional neural networks to learn same-instance dense correspondence. FlowNet~\cite{fischer2015flownet} learns an optical flow CNN with a synthetic Flying Chairs dataset that generalizes well to existing benchmark datasets, yet still falls a bit short of state-of-the-art optical flow methods like DeepFlow~\cite{weinzaepfel2013deepflow} and EpicFlow~\cite{revaud2015epicflow}. Several recent works have also used supervision from reconstructed 3D scene and stereo pairs~\cite{han2015matchnet,vzbontar2015stereo,agrawal2015learning}. However all these approaches are inherently limited to matching images of the same physical object/scene. Long~\etal~\cite{long2014convnets} use deep features learned from large-scale object classification tasks to perform intra-class image alignment, but found it to perform similarly to SIFT flow.
\\\\
{\bf Image-shape correspondence}~~Our work is partially motivated by recent progress in image-shape alignment that allows establishing correspondence between images through intermediate 3D shapes. Aubry~\etal~\cite{Aubry:2014:SCE} learns discriminative patches for matching 2D images to their corresponding 3D CAD models, while Peng~\etal~\cite{peng2015learning} utilizes CAD models to train object detectors with few shots of labeled real images. In cases where depth data is available, deep learning methods have recently been applied to 3D object recognition and alignment between CAD models and RGB-D images~\cite{guptaCVPR15a,song16,wu20153d}. Other works~\cite{DBLP:journals/tog/HuangWK15,shmlg_imageDepth_sig14} leverage image and shape collections for joint pose estimation and refining image-shape alignment, which are further applied to single-view object reconstruction and depth estimation. Although our approach requires 3D CAD models for constructing the training set, the image-shape alignment is jointly learned with the image-image alignment, and no CAD models are required at test time.
\section{Approach}
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{net_final.pdf}
\caption{Overview of our network architecture, which consists of three major components: 1) \textbf{feature encoder} on both input images, 2) \textbf{flow decoder} predicting the dense flow field from the source to the target image and 3) \textbf{matchability decoder} that outputs a probability map indicating whether each pixel in the source image has a correspondence in the target. See~\refsec{network} for more details.}
\label{fig:net}
\end{figure*}
Our goal is to predict a dense flow (or correspondence) field $F_{a,b} : \mathbb R^2 \to \mathbb R^2$ between pairs of images $a$ and $b$. The flow field $F_{a,b}(p) = (p_x - q_x, p_y - q_y)$ computes the relative offset from each point $p$ in image $a$ to a corresponding point $q$ in image $b$. Given that pairwise correspondence might not always be well-defined
(e.g. a side-view car and a frontal-view car do not have many visible parts in common), we additionally compute a matchability map $M_{a,b} : \mathbb R^2 \to [0,1]$ predicting if a correspondence exists $M_{a,b}(p)=1$ or not $M_{a,b}(p)=0$.
We learn both the flow field and the matchability prediction with a convolutional neural network. Both functions are differentiable with respect to the network parameters, which could be directly learned if we had dense annotations for $F_{a,b}$ and $M_{a,b}$ on a large set of real image pairs. However, in practice it is infeasible to obtain those annotations at scale as they are either too time-consuming or ambiguous to annotate.
We instead choose a different route, and learn both functions by placing the supervision on the desired properties of the ground-truth, i.e.
while we do not know what the ground-truth is, we know how it should behave. In this paper, we use {\em cycle consistency} with 3D CAD models as the desired property that will be our supervisory signal. Specifically, for each pair of real training images $r_1$ and $r_2$, we find a 3D CAD model of the same category, and render two synthetic views $s_1$ and $s_2$ in similar viewpoint as $r_1$ and $r_2$, respectively (see \refsec{quartets} for more details). For each training quartet $<s_1,s_2,r_1,r_2>$ we learn to predict flows from $s_1$ to $r_1$ ($F_{s_1, r_1}$) to $r_2$ ($F_{r_1, r_2}$) to $s_2$ ($F_{r_2, s_2}$) that are cycle-consistent with the ground-truth flow from $s_1$ to $s_2$ ($\tilde{F}_{s_1, s_2}$) provided by the rendering engine (similarly for the matchability prediction).
By constructing consistency supervision through 3D CAD models, we aim to learn 2D image correspondences that potentially captures the 3D semantic appearance of the query objects. Furthermore, making $\tilde{F}_{s_1, s_2}$ be ground-truth by construction prevents the cycle-consistency optimization from producing trivial solutions, such as identity flows.
Sections \ref{sec:objective} and \ref{sec:matchability} formally define our training objective for learning correspondence $F$ and matchability $M$, respectively. Section~\ref{sec:lerp} demonstrates how to obtain continuous approximation of discrete maps that allows end-to-end training. Section~\ref{sec:network} describes our network architecture.
\subsection{Learning dense correspondence}
\lblsec{objective}
Given a set of training quartets $\{<s_1, s_2, r_1, r_2>\}$, we train the CNN to minimize the following objective:
\begin{equation}
\sum_{<\!s_1\!,s_2\!,r_1\!,r_2\!>}\!\!\!\mathcal{L}_{flow}\left( \tilde{F}_{s_1\!, s_2},~F_{s_1\!, r_1}\!\circ\!F_{r_1\!, r_2}\!\circ\!F_{r_2\!, s_2} \right),\label{eq:obj}
\end{equation}
where $\tilde{F}_{s_1, s_2}$ refers to the ground-truth flow between two synthetic views, $F_{s_1,r_1}$, $F_{r_1, r_2}$ and $F_{r_2, s_2}$ are predictions made by the CNN along the transitive path. The transitive flow composition $\bar{F}_{a,c} = F_{a,b} \circ F_{b,c}$ is defined as
\begin{equation}
\bar{F}_{a,c}(p) = F_{a,b}(p) + F_{b,c}(p+F_{a,b}(p))~,
\label{eq:comp}
\end{equation}
which is differentiable as long as $F_{a,b}$ and $F_{b,c}$ are differentiable. $\mathcal{L}_{flow}(\tilde{F}_{s_1, s_2}, \bar{F}_{s_1, s_2})$ denotes the truncated Euclidean loss defined as
\begin{align*}
\mathcal{L}_{flow} & (\tilde{F}_{s_1, s_2}, \bar{F}_{s_1, s_2})=\\
&\sum_{p | \tilde{M}_{s_1, s_2}(p)=1}\min(\|\tilde{F}_{s_1,s_2}(p) - \bar{F}_{s_1,s_2}(p)\|^2,T^2)~,
\end{align*}
where $\tilde{M}_{s_1, s_2}(p)$ is the ground-truth matchability map provided by the rendering engine ($\tilde{M}_{s_1, s_2}(p) = 0$ when $p$ is either a background pixel or not visible in $s_2$), and $T=15$ (pixels) for all our experiments. In practice, we found the truncated loss to be more robust to spurious outliers for training, especially during the early stage when the network output tends to be highly noisy.
\subsection{Learning dense matchability}
\lblsec{matchability}
Our training objective for matchability prediction also utilizes the cycle consistency signal:
\begin{equation}
\sum_{<\!s_1\!,s_2\!,r_1\!,r_2\!>}\!\!\!\mathcal{L}_{mat}\left( \tilde{M}_{s_1\!, s_2},~ M_{s_1\!, r_1}\!\circ\!M_{r_1\!, r_2}\!\circ\!M_{r_2\!, s_2} \right),\label{eq:mobj}
\end{equation}
where $\tilde{M}_{s_1,s_2}$ refers to the ground-truth matchability map between the two synthetic views, $M_{s_1,r_1}$, $M_{r_1, r_2}$ and $M_{r_2, s_2}$ are CNN predictions along the transitive path, and $\mathcal{L}_{mat}$ denotes per-pixel cross-entropy loss. The matchability map composition is defined as
\begin{equation}
\bar{M}_{a,c}(p) = M_{a,b}(p) M_{b,c}(p+F_{a,b}(p))~,
\label{eq:mapcomp}
\end{equation}
where the composition depends on both the matchability as well as the flow field.
Due to the multiplicative nature in matchability composition (as opposed to additive in flow composition), we found that training with objective~\ref{eq:mobj} directly results in the network exploiting the clean background in synthetic images, which helps predict a perfect segmentation of the synthetic object in $M_{s_1, r_1}$. Once $M_{s_1, r_1}$ predicts zero values for background points, the network has no incentive to correctly predict the matchability for background points in $M_{r_1,r_2}$, as the multiplicative composition has zero values regardless of the transitive predictions along $M_{r_1, r_2}$ and $M_{r_2, s_2}$.
To address this, we fix $M_{s_1, r_1}=\mathbf{1}$ and $M_{r_2, s_2}=\mathbf{1}$, and only train the CNN to infer $M_{r_1,r_2}$.
This assumes that every pixel in $s_1 (s_2)$ is matchable in $r_1 (r_2)$, and allows the matchability learning to happen between real images. Note that this is still different from directly using $\tilde{M}_{s_1, s_2}$ as supervision for $M_{r_1,r_2}$ as the matchability composition depends on the predicted flow field along the transitive path.
The matchability objective~\ref{eq:mobj} is jointly optimized with the flow objective~\ref{eq:obj} during training, and our final objective can be written as $\sum_{<\!s_1\!,s_2\!,r_1\!,r_2\!>}\mathcal{L}_{flow} + \lambda \mathcal{L}_{mat}$ with $\lambda = 100$.
\subsection{Continuous approximation of discrete maps}
\lblsec{lerp}
An implicit assumption made in our derivation of the transitive composition (Eq.~\ref{eq:comp} and~\ref{eq:mapcomp}) is that $F$ and $M$ are differentiable functions over continuous input, while images inherently consist of discrete pixel grids. To allow end-to-end training with stochastic gradient descent (SGD), we obtain continuous approximation of the full flow field and the matchability map with bilinear interpolation over the CNN predictions on discrete pixel locations. Specifically, for each discrete pixel location $\hat p \in \{1,\ldots,W\} \times \{1,\ldots,H\}$, the network predicts a flow vector $F_{a,b}(\hat p)$ as well as a matchability score $M_{a,b}(\hat p)$, and the approximation over all continuous points $p \in [1, W] \times [1, H]$ is obtained by:
\begin{align*}
F_{a,b}(p) &= \sum_{\hat{p} \in \mathcal{N}_p} (1-|p_x-\hat{p}_x|)(1-|p_y-\hat{p}_y|) F_{a,b}(\hat{p}) \\
M_{a,b}(p) & = \sum_{\hat{p} \in \mathcal{N}_p} (1-|p_x-\hat{p}_x|)(1-|p_y-\hat{p}_y|) M_{a,b}(\hat{p})~,
\end{align*}
where $\mathcal{N}_p$ denotes the four-neighbor pixels (top-left, top-right, bottom-left, bottom-right) of point $p$, or just $p$ if it is one of the discrete pixels. This is equivalent to the differentiable image sampling with a bilinear kernel proposed in~\cite{stn}.
\subsection{Network architecture}
\lblsec{network}
Our network architecture (see \reffig{net}) follows the encoder-decoder design principle with three major components: 1) \textbf{feature encoder} of $8$ convolution layers that extracts relevant features from both input images with shared network weights; 2) \textbf{flow decoder} of $9$ fractionally-strided/up-sampling convolution (uconv) layers that assembles features from both input images, and outputs a dense flow field; 3) \textbf{matchability decoder} of $9$ uconv layers that assembles features from both input images, and outputs a probability map indicating whether each pixel in the source image has a correspondence in the target.
All conv/uconv layers are followed by rectified linear units (ReLUs) except for the last uconv layer of either decoder, and the filter size is fixed to $3\times 3$ throughout the whole network. No pooling layer is used, and the stride is $2$ when increasing/decreasing the spatial dimension of the feature maps. The output of the matchability decoder is further passed to a sigmoid layer for normalization.
During training, we apply the same network to three different input pairs along the cycle ($s_1 \rightarrow r_1, r_1, \rightarrow r_2, $ and $r_2 \rightarrow s_2$), and composite the output to optimize the consistency objectives~\ref{eq:obj} and~\ref{eq:mobj}.
\section{Experimental Evaluation}
In this section, we describe the details of our network training procedure, and evaluate the performance of our network on correspondence and matchability tasks.
\subsection{Training set construction}
\lblsec{quartets}
The 3D CAD models we used for constructing training quartets come from the ShapeNet database~\cite{shapenet}, while the real images are from the PASCAL3D+ dataset~\cite{xiang2014beyond}. For each object instance (cropped from the bounding box and rescaled to $128 \times 128$) in the train split of PASCAL3D+, we render all 3D models under the same camera viewpoint (provided by PASCAL3D+), and only use $K = 20$ nearest models as matches to the object instance based on the HOG~\cite{hog} Euclidean distance. We then construct training quartets each consisting of two real images ($r_1$ and $r_2$) matched to the same 3D model and their corresponding rendered views ($s_1$ and $s_2$). On average, the number of valid training quartets for each category is about $80,000$.
\subsection{Network training}
We train the network in a category-agnostic manner (i.e. a single network for all categories). We first initialize the network (feature encoder + flow decoder pathway) to mimic SIFT flow by randomly sampling image pairs from the training quartets and training the network to minimize the Euclidean loss between the network prediction and the SIFT flow output on the sampled pair\footnote{We also experimented with other initialization strategies (e.g. predicting ground-truth flows between synthetic images), and found that initializing with SIFT flow output works the best.}. Then we fine-tune the whole network end-to-end to minimize the consistency loss defined in Eq.~\ref{eq:obj} and~\ref{eq:mobj}. We use the ADAM solver~\cite{adam} with $\beta_1 = 0.9, \beta_2 = 0.999$, initial learning rate of $0.001$, step size of $50,000$, step multiplier of $0.5$ for $200,000$ iterations. We train with mini-batches of $40$ image pairs during initialization and $10$ quartets during fine-tuning.
We visualize the effect of our cycle-consistency training in Figure~\ref{fig:train}, where we sample some random points in the synthetic image $s_1$, and plot their predicted correspondences along the cycle $s_1 \rightarrow r_1 \rightarrow r_2 \rightarrow s_2$ to compare with the ground-truth in $s_2$. One can see that the transitive trajectories become more and more cycle-consistent with more iterations of training, while individual correspondences along each edge of the cycle also tend to become more semantically plausible.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{cycle_train.pdf}
\caption{Visualizing the effects of consistency training on the network output. The randomly sampled ground-truth correspondences between synthetic images are marked in solid lines, and the correspondence predictions along the cycle (synthetic to real, real to real and real to synthetic) made by our network are marked in dashed lines. One can see that the transitive composition of our network output becomes more and more consistent with the ground-truth as training progresses, while individual correspondences along each edge of the cycle also tend to become more semantically plausible.}
\label{fig:train}
\end{figure*}
\subsection{Feature visualization}
We visualize the features learned by the network using the t-SNE algorithm~\cite{tsne}. Specifically, we extract conv-9 features (i.e. the output of the last encoder layer) from the entire set of car instances in the PASCAL3D+ dataset, and embed them in 2-D with the t-SNE algorithm. Figure~\ref{fig:tsne} visualizes the embedding. Interestingly, while our network is not explicitly trained to perform viewpoint estimation, the embedding layout appears to be viewpoint-sensitive, which implies that the network might implicitly learn that viewpoint is an important cue for correspondence/matchability tasks through our consistency training.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{tsne3.png}
\caption{Conv-9 feature embedding for cars visualized by t-SNE~\cite{tsne}. Interestingly, the overall layout seems to be mainly clustered based on the camera viewpoint, while the network is not explicitly trained to perform viewpoint estimation. This suggests that the network might implicitly learn that viewpoint is an important cue for the correspondence/matchability tasks through our consistency training.}
\label{fig:tsne}
\end{figure*}
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{keypts.pdf}
\caption{Comparison of keypoint transfer performance for different methods on example test image pairs. Overall, our consistency-supervised network (second-to-last row) is able to produce more accurate keypoint transfer results than the baselines. The last column shows a case when SIFT flow performs better than ours.}
\label{fig:key}
\end{figure*}
\begin{table*}[t]
\centering
\scalebox{0.88}{
\begin{tabular}{ccccccccccccc|c}
\toprule
& aero & bike & boat & bottle & bus & car & chair & table & mbike & sofa & train & tv & {\bf mean} \tabularnewline
\midrule
SIFT flow~\cite{liu2011sift}
& $9.8$ & $\mathbf{23.3}$ & $8.9$ & $28.3$ & $28.6$ & $22.4$ & $10.8$ & $\mathbf{13.2}$ & $\mathbf{17.9}$ & $14.2$ & $14.4$ & $42.9$ & $19.6$ \tabularnewline
Long \etal~\cite{long2014convnets}
& $10.4$ & $22.8$ & $7.6$ & $30.8$ & $28.4$ & $21.1$ & $10.2$ & $12.7$ & $13.5$ & $12.9$ & $12.6$ & $38.5$ & $18.5$ \tabularnewline
\midrule
$\text{CNN}_{I2S}$
& $9.1$ & $14.7$ & $5.2$ & $25.9$ & $25.4$ & $23.7$ & $11.9$ & $11.3$ & $13.4$ & $16.8$ & $11.3$ & $45.2$ & $17.8$ \tabularnewline
$\text{CNN}_{init}$
& $8.6$ & $20.3$ & $8.5$ & $29.4$ & $24.3$ & $20.1$ & $9.9$ & $11.6$ & $15.4$ & $11.6$ & $12.5$ & $40.2$ & $17.7$ \tabularnewline
$\text{CNN}_{init} + $ Synthetic ft.
& $10.2$ & $22.2$ & $8.7$ & $30.4$ & $24.5$ & $21.3$ & $10.2$ & $12.1$ & $15.7$ & $12.0$ & $12.8$ & $40.5$ & $18.4$ \tabularnewline
{\bf $\text{CNN}_{init} + $ Consistency ft.}
& $\mathbf{11.3}$ & $22.3$ & $\mathbf{10.1}$ & $\mathbf{40.3}$ & $\mathbf{40.3}$ & $\mathbf{33.3}$ & $\mathbf{15.0}$ & $\mathbf{13.2}$ & 17.2 & $\mathbf{17.4}$ & $\mathbf{16.7}$ & $\mathbf{51.1}$ & $\mathbf{24.0}$ \tabularnewline
\bottomrule
\end{tabular}
}
\vspace{0.07in}
\caption{Keypoint transfer accuracy measured in PCK ($\alpha = 0.1$) on the PASCAL3D+ categories. Overall, our final network (last row) outperforms all baselines (except on ``bicycle'' and ``motorbike''). Notice the performance gap between our initialization ($\text{CNN}_{init}$) and the final network, which highlights the improvement made by cycle-consistency training.}
\label{tab:pck}
\end{table*}
\subsection{Keypoint transfer}
We evaluate the quality of our correspondence output using the keypoint transfer task on the $12$ categories from PASCAL3D+~\cite{xiang2014beyond}. For each category, we exhaustively sample all image pairs from the val split (not seen during training), and determine if a keypoint in the source image is transferred correctly by measuring the Euclidean distance between our correspondence prediction and the annotated ground-truth (if exists) in the target image. A correct transfer means the prediction falls within $\alpha \cdot \max(H,W)$ pixels of the ground-truth with $H$ and $W$ being the image height and width, respectively (both are $128$ pixels in our case). We compute the percentage of correct keypoint transfer (PCK) over all image pairs as the metric, and provide performance comparison for the following methods in Table~\ref{tab:pck}:
\begin{itemize}
\item SIFT flow~\cite{liu2011sift} -- A classic method for dense correspondence using SIFT feature descriptors and hand-designed smoothness and large-displacement priors. We also ran preliminary evaluation on a more recent follow-up based on deformable spatial pyramids~\cite{kim2013deformable}, and found it to perform similarly to SIFT flow.
\item Long \etal~\cite{long2014convnets} -- Similar MRF energy minimization framework as SIFT flow but with deep features learned from the ImageNet classification task.
\item CNN$_{I2S}$ -- Our network trained on real image pairs with correspondence inferred by compositing the output of an off-the-shelf image-to-shape alignment algorithm~\cite{DBLP:journals/tog/HuangWK15} and the ground-truth synthetic correspondence (i.e. obtaining direct supervision for $F_{r_1,r_2}$ through $F_{r_1,s_1} \circ \tilde{F}_{s_1, s_2} \circ F_{s_2,r_2},$ where $F_{r_1,s_1}$ and $F_{s_2,r_2}$ are inferred from~\cite{DBLP:journals/tog/HuangWK15}).
\item CNN$_{init}$ -- Our network trained to mimic SIFT flow.
\item CNN$_{init} + $ Synthetic ft. -- fine-tuning on synthetic image pairs with ground-truth correspondence after initialization with SIFT flow.
\item CNN$_{init} + $ Consistency ft. -- fine-tuning with our objectives~\ref{eq:obj} and~\ref{eq:mobj} after initialization with SIFT flow.
\end{itemize}
Overall, our consistency-supervised network significantly outperforms all other methods (except on ``bicycle" and ``motorbike" where SIFT flow has a slight advantage). Notice the significant improvement over the initial network after consistency fine-tuning. The performance gap between the last two rows of Table~\ref{tab:pck} suggests that consistency supervision is much more effective in adapting to the real image domain than direct supervision from synthetic ground-truth.
Figure~\ref{fig:key} compares sample keypoint transfer results using different methods. In general, our final prediction tends to match the ground-truth much better than the other baselines, and could sometimes overcome substantial viewpoint and appearance variation where previous methods, like SIFT flow, are notoriously error-prone.
\begin{figure*}[t]
\centering
\includegraphics[width=\linewidth]{match.pdf}
\caption{Sample visualization of our matchability prediction. Notice how the prediction varies for the same source image when changing only the target image. The last two columns demonstrate a typical failure mode of our network having trouble localizing the fine boundaries of the matchable regions.}
\label{fig:match}
\end{figure*}
\subsection{Matchability prediction}
\lblsec{exp:match}
We evaluate the performance of matchability prediction using the PASCAL-Part dataset~\cite{chen2014detect}, which provides human-annotated part segment labeling\footnote{For categories without part labels, including boat, chair, table and sofa, we use the foreground segmentation mask instead.}. For each test image pair, a pixel in the source image is deemed matchable if there exists another pixel in the target image that shares the same part label, and all background pixels are unmatchable. We measure the performance by computing the percentage of pixels being classified correctly. For our method, we classify a pixel as matchable if its probability is $> 0.5$ according to the network prediction. To obtain matchability prediction for SIFT flow, we compute the $L_1$ norm of the SIFT feature matching error for each source pixel after the alignment, and a pixel is predicted to be matchable if the error is below a certain threshold (we did grid search on the training set to determine the threshold, and found $1,000$ to perform the best). Table~\ref{tab:match} compares the classification accuracy between our method and SIFT flow prediction (chance performance is $50\%$). Our method significantly outperforms SIFT flow on all categories except ``bicycle'' and ``motorbike'' ($67.8\%$ vs. $57.1\%$ mean accuracy).
We visualize some examples of our matchability prediction in Figure~\ref{fig:match}. Notice how the prediction varies when the target image changes with the source image being the same.
\begin{table*}[t]
\centering
\scalebox{0.88}{
\begin{tabular}{ccccccccccccc|c}
\toprule
& aero & bike & boat & bottle & bus & car & chair & table & mbike & sofa & train & tv & {\bf mean} \tabularnewline
\midrule
SIFT flow~\cite{liu2011sift}
& $66.2$ & $\mathbf{62.7}$ & $49.5$ & $50.5$ & $52.0$ & $64.5$ & $50.7$ & $50.5$ & $\mathbf{80.6}$ & $49.6$ & $58.5$ & $50.2$ & $57.1$ \tabularnewline
Ours
& $\mathbf{75.8}$ & $61.0$ & $\mathbf{66.7}$ & $\mathbf{67.1}$ & $\mathbf{67.3}$ & $\mathbf{72.0}$ & $\mathbf{66.1}$ & $\mathbf{68.4}$ & $68.0$ & $\mathbf{71.2}$ & $\mathbf{64.4}$ & $\mathbf{65.1}$ & $\mathbf{67.8}$ \tabularnewline
\bottomrule
\end{tabular}
}
\vspace{0.07in}
\caption{Performance comparison of matchability prediction between SIFT flow and our method (higher is better). See~\refsec{exp:match} for more details on the experiment setup.}
\label{tab:match}
\end{table*}
\subsection{Shape-to-image segmentation transfer}
\lblsec{app}
\begin{figure}[t]
\centering
\includegraphics[width = \columnwidth]{segment.pdf}
\caption{Visual comparison among different segmentation methods. From left to right: input query image, segmentation by \cite{Krahenbuhl_Koltun_2011}, segmentation transferred using SIFT flow, segmentation transferred using our flow and the retrieved shape whose segmentation is used for transferring. See~\refsec{app} for more details.}
\label{fig:segmentation}
\end{figure}
Although in this paper we are mostly interested in finding correspondence between real images, a nice byproduct of our consistency training is that the network also implicitly learns cross-domain, shape-to-image correspondence, which allows us to transfer per-pixel labels (e.g. surface normals, segmentation masks, etc.) from shapes to real images. As a proof of concept, we ran a toy experiment on the task of segmentation transfer. Specifically, we construct a shape database of about $200$ shapes per category, with each shape being rendered in $8$ canonical viewpoints. Given a query real image, we apply our network to predict the correspondence between the query and each rendered view of the same category, and warp the query image according to the predicted flow field. Then we compare the HOG Euclidean distance between the warped query and the rendered views, and retrieve the rendered view with minimum error whose correspondence to the query image on the foreground region is used for segmentation transfer. Figure~\ref{fig:segmentation} shows sample segmentation using different methods. We can see that our learned flows tend to produce more accurate segmentation transfer than SIFT flow using the same pipeline. In some cases our output can even segment challenging parts such as the bars and wheels of the chairs.
\section{Discussion}
In this paper, we used cycle-consistency as a supervisory signal to learn dense cross-instance correspondences. Not only did we find that this kind of supervision is surprisingly effective, but also that the idea of learning with cycle-consistency could potentially be fairly general.
One could apply the same idea to construct other training scenarios, as long as the ground-truth of one or more edges along the cycle is known. We hope that this work will inspire more efforts to tackle tasks with little or no direct labels by exploiting cycle consistency or other types of indirect or ``meta''-supervision.
\section*{Acknowledgements}
We thank Leonidas Guibas, Shubham Tulsiani, and Saurabh Gupta for helpful discussions. This work was sponsored in part by NSF/Intel VEC 1539099, ONR MURI N000141010934, and a hardware donation by NVIDIA.
{\small
\bibliographystyle{ieee}
|
\subsection{Fluid models for TCP}
Now, we briefly outline the fluid models for the evolution of the average window sizes of the two sets of TCP flows in the congestion avoidance phase for three topologies.
\subsection*{Case I}
This model consists of a single bottleneck link with two distinct sets of TCP flows feeding into a common core router, as shown in Fig.~\ref{Figure.1}(a). The core router has a buffer size of $B$, with link capacity $C$. Thus, for generalised TCP flows, the non-linear, time-delayed, fluid model of the system is given by the following equations:
\\
\begin{align}
\frac{dw_{j}(t)}{dt} = \frac{w_{j}(t-\tau_{j})}{\tau_{j}}\bigg(i_j\left(w_{j}(t)\right)\Big(1-q(t,\tau_1,\tau_2)\Big) - d_j\left((w_{j}(t)\right)q(t,\tau_1,\tau_2)\bigg), \hspace{2ex}j=1,2,
\label{eq:modelb_1}
\end{align}
where $q(t,\tau_1,\tau_2)$ represents the packet loss probability at the core router, and depend on the sending rates of both sets of TCP flows.
\subsection*{Case II}
This model consists of two distinct sets of TCP flows, regulated by a single edge router and feeding into a common core router, as shown in Fig.~\ref{Figure.1}(b). The buffer sizes of the core router and the edge routers are $B_1$ and $B_2$, with link capacities $C_1$ and $C_2$ respectively. Thus, for generalised TCP flows, the non-linear, time-delayed, fluid model of the system is given by the following differential equations:
\begin{align}
\frac{dw_{j}(t)}{dt} = \frac{w_{j}(t-\tau_{j})}{\tau_{j}}\bigg(i_j\left(w_{j}(t)\right)\Big(1-q_1(t,\tau_1,\tau_2)-q_2(t,\tau_1,\tau_2))\Big) - d_j\left((w_{j}(t)\right)\left(q_1(t,\tau_1,\tau_2)+q_2(t,\tau_1,\tau_2)\right)\bigg),
\label{eq:modelb_2}
\end{align}
for $j=1,2,$ and $q_1(t,\tau_1,\tau_2)$ and $q_2(t,\tau_1,\tau_2)$ denote the packet loss probabilities at the edge router and the core router respectively.
\subsection*{Case III}
This model consists of two distinct sets of TCP flows, regulated by two edge routers and feeding into a common core router, as shown in Fig.~\ref{Figure.1}(c). The buffer size at the core router is $B$, with link capacity $C$. The buffer sizes for the edge routers are $B_1$ and $B_2$, with link capacities $C_1$ and $C_2$ respectively. Thus, for generalised TCP flows, the non-linear, time-delayed, fluid model of the system is given by the following equations:\\
\begin{align}
\frac{dw_{j}(t)}{dt} = \frac{w_{j}(t-\tau_{j})}{\tau_{j}}\bigg(i_j\left(w_{j}(t)\right)\Big(1-p_{j}(t-\tau_{j})-q(t,\tau_{1},\tau_{2})\Big) - d_j\left((w_{j}(t)\right)\Big(p_{j}(t-\tau_{j})+q(t,\tau_{1},\tau_{2})\Big)\bigg),
\label{eq:modelb_3}
\end{align}
for $j=1,2.$ The loss probabilities at the two edge routers are $p_1(t)$ and $p_2(t)$. The loss probability at the core router is denoted by $q(t, \tau_1, \tau_2)$. Recall that, the increase and decrease functions are specific to the choice of a particular flavour of TCP. Specifically,~\cite{Raja} has summarised the increase and decrease functions for different TCP flavours including Compound. Since our primary focus is on Compound TCP, we state the increase and decrease functions for Compound as follows:
\begin{equation}
\label{eq:increasedecrease}
i(w(t))=\alpha\left(w(t)\right)^{k-1}, \hspace{1ex}\text{and}\hspace{1ex} d(w(t))=\beta w(t).
\end{equation}
Here, $\alpha$, $k$ are the increase parameters and $\beta$ is the decrease parameter. The default values of these parameters are $\alpha=0.125$, $k=0.75$ and $\beta=0.5$~\cite{Tan}.
\begin{figure*}
\vspace{2mm}
{
\begin{tikzpicture}[scale=0.5]
\draw (0,0) -- (-1.5,0.75);
\draw (0,0) -- (-1.5,0.25);
\draw (0,0) -- (-1.5,-0.25);
\draw (-1.5,-0.75) -- (0,0);
\draw (0,-0.75) -- (3.85,-0.75);
\draw (3.85,-0.75) -- (3.85,0.75);
\draw (3.85,0.75) -- (0,0.75);
\draw (0,0.75) -- (0,-0.75);
\draw (3.65,-0.75) -- (3.65,0.75);
\draw (3.4,-0.75) -- (3.4,0.75);
\draw (3.15,-0.75) -- (3.15,0.75);
\draw (4.65,0) circle(0.75 cm);
\draw [loosely dotted,very thick] (3.0,0) -- (2.4,0);
\draw[->] (5.4,0) -- (6.2,0);
\draw (5.8,0) -- (5.8,2.2);
\draw (5.8,2.2) -- (-1.75,2.2);
\draw (-1.75,2.2) -- (-1.75,0.5);
\draw [->](-1.75,0.5) -- (-1.5,0.5);
\draw (5.8,0) -- (5.8,-2.2);
\draw (5.8,-2.2) -- (-1.75,-2.2);
\draw (-1.75,-2.2) -- (-1.75,-0.5);
\draw [->](-1.75,-0.5) -- (-1.5,-0.5);
\node at (1,0) {$B$};
\node at (4.65,0) {$C$};
\node at (1.5, 1.75) {$\tau_1$};
\node at (1.5,-1.9) {$\tau_2$};
\node at (1.75,-3) {(a)};
\draw (8.5,0) -- (8,0.75);
\draw (8.5,0) -- (8,0.25);
\draw (8.5,0) -- (8,-0.25);
\draw (8.5,0) -- (8,-0.75);
\draw (8.5,-0.75) -- (10.7,-0.75);
\draw (10.7,-0.75) -- (10.7,0.75);
\draw (10.7,0.75) -- (8.5,0.75);
\draw (8.5,0.75) -- (8.5,-0.75);
\draw (10.55,-0.75) -- (10.55,0.75);
\draw (10.35,-0.75) -- (10.35,0.75);
\draw [loosely dotted,very thick,] (10.25,0) -- (9.75,0);
\draw (12.25,0) -- (13,0);
\draw (13,-0.75) -- (15.2,-0.75);
\draw (15.2,-0.75) -- (15.2,0.75);
\draw (15.2,0.75) -- (13,0.75);
\draw (13,0.75) -- (13,-0.75);
\draw (15.05,-0.75) -- (15.05,0.75);
\draw (14.85,-0.75) -- (14.85,0.75);
\draw (11.5,0) circle(0.75 cm);
\draw (16,0) circle(0.75 cm);
\draw [loosely dotted,very thick] (14.8,0) -- (14.4,0);
\draw[->] (16.75,0) -- (17.5,0);
\draw (17.125,0) -- (17.125,2.2);
\draw (17.125,2.2) -- (7.75,2.2);
\draw (7.75,2.2) -- (7.75,0.5);
\draw [->](7.75,0.5) -- (8,0.5);
\draw (17.125,0) -- (17.125,-2.2);
\draw (17.125,-2.2) -- (7.75,-2.2);
\draw (7.75,-2.2) -- (7.75,-0.5);
\draw [->](7.75,-0.5) -- (8,-0.5);
\node at (9.2,0) {$B_1$};
\node at (11.5,0) {$C_1$};
\node at (13.7,0) {$B_2$};
\node at (16.0,0) {$C_2$};
\node at (12.5, 1.75) {$\tau_1$};
\node at (12.5, -1.9) {$\tau_2$};
\node at (12.5,-3){(b)};
\draw (8.5,0) -- (8,0.75);
\draw (8.5,0) -- (8,0.25);
\draw (8.5,0) -- (8,-0.25);
\draw (8.5,0) -- (8,-0.75);
\draw (19.8,0.5) -- (22.5,0.5);
\draw (22.5,0.5) -- (22.5,1.7);
\draw (22.5,1.7) -- (19.8,1.7);
\draw (19.8,1.7) -- (19.8,0.5);
\draw (22.25,0.5) -- (22.25,1.7);
\draw (22,0.5) --(22,1.7);
\draw [loosely dotted,very thick] (22,1.1) -- (21.5,1.1);
\node at (20.4,1.1) {$B_1$};
\node at (23.15,1.1) {$C_1$};
\draw (19.8,-0.5) -- (22.5,-0.5);
\draw (22.5,-0.5)-- (22.5,-1.7);
\draw (22.5,-1.7) -- (19.8,-1.7);
\draw (19.8,-1.7) -- (19.8,-0.5);
\draw (22.25,-0.5) -- (22.25,-1.7);
\draw (22,-0.5) --(22,-1.7);
\draw [loosely dotted,very thick] (22,-1.1) -- (21.5,-1.1);
\draw [loosely dotted,very thick] (27.6,0) -- (27,0);
\node at (20.4,-1.1) {$B_2$};
\node at (23.15,-1.1) {$C_2$};
\node at (24.5,1.75) {$\tau_1$};
\node at (24.5,-1.9) {$\tau_2$};
\node at (26,0) {$B$};
\node at (28.85,0) {$C$};
\node at (24.5,-3) {(c)};
\draw (23.15,1.1) circle(0.6 cm);
\draw (23.15,-1.1) circle(0.6 cm);
\draw (23.75,1.1) -- (25.5,0);
\draw (23.75,-1.1) -- (25.5, 0);
\draw (25.5,0.6) -- (28.2,0.6);
\draw (28.2,0.6) -- (28.2,-0.6);
\draw (28.2,-0.6) -- (25.5,-0.6);
\draw (25.5,-0.6) -- (25.5,0.6);
\draw (27.95,-0.6) -- (27.95,0.6);
\draw (27.7,-0.6) -- (27.7,0.6);
\draw (27.7,-0.6) -- (27.7,0.6);
\draw (28.85,0) circle(0.6 cm);
\draw[->] (29.45,0) -- (31,0);
\draw (18.8,1.7) -- (19.8,1.1);
\draw (18.8,0.5) -- (19.8,1.1);
\draw (18.8,-0.5) -- (19.8,-1.1);
\draw (18.8,-1.7) -- (19.8,-1.1);
\draw (30.225,0) -- (30.225,2.2);
\draw (30.225,2.2) -- (18.5,2.2);
\draw (18.5,2.2) -- (18.5,1.1);
\draw[->] (18.5,1.1) -- (18.9,1.1);
\draw (30.225,0) -- (30.225,-2.2);
\draw (30.225,-2.2) -- (18.5,-2.2);
\draw (18.5,-2.2) -- (18.5,-1.1);
\draw[->] (18.5,-1.1) -- (18.9,-1.1);
\end{tikzpicture}
}
\caption{\emph{Schematic diagrams of three topologies.} (a) Case I, a single bottleneck topology (b) Case II, two routers in tandem and (c) Case III, two routers feeding into one core router.}
\label{Figure.1}
\end{figure*}
\subsection{Packet loss probability}
In this paper, we mainly focus on small buffers with Drop-Tail queue policy for the local stability analysis of the non-linear fluid models of TCP given by \eqref{eq:modelb_1}, \eqref{eq:modelb_2} and \eqref{eq:modelb_3}. We first consider the scenario where a \emph{large} number of long-lived TCP flows having a common round trip time of $\tau$ feed into a router having a buffer size of $B$. The bottleneck link has a capacity $C$. In this scenario, we can approximate the packet loss probability of the router by the blocking probability of an $M/M/1/B$ queue~\cite{Raja}. This gives rise to the following fluid model:
\begin{align}
\label{eq:loss}
p(t)=\bigg(\frac{w(t)}{C\tau}\bigg)^B,
\end{align}
where $w(t)$ represents the average window size of the TCP flows. Using \eqref{eq:loss}, we can then obtain the functional forms of packet loss probabilities for the three scenarios, which we briefly outline as follows:
\subsubsection*{Case I}The fluid model for the loss probability at the core router is given by
\begin{align}
q(t)=\Bigg(\frac{w_1(t) / \tau_1 + w_2(t) / \tau_2}{C}\Bigg)^B.
\label{eq:case1_loss}
\end{align}
\subsubsection*{Case II} The fluid models for the loss probabilities are:
\begin{align}
q_1(t)= \left(\frac{w_{1}(t) / \tau_{1}+w_{2}(t)/ \tau_{2}}{C_2}\right)^{B_2} {\rm and}, \hspace{1ex} q_2(t)= \left(\frac{w_{1}(t) / \tau_{1}+ w_{2}(t) /\tau_{2}}{C_1}\right)^{B_1}.
\label{eq:case2_loss}
\end{align}
\subsubsection*{Case III} Using \eqref{eq:loss}, we can approximate the loss probabilities at various routers as below:
\begin{align}
p_{1}(t) &= \left(\frac{w_{1}(t)}{C_{1}\tau_{1}}\right)^{B_{1}} \hspace{1ex},\hspace{1ex} p_{2}(t)=\left(\frac{w_{2}(t)}{C_{2}\tau_{2}}\right)^{B_{2}},\ {\rm and} \notag \\
q(t) &= \left(\frac{w_{1}(t)/\tau_{1}+w_{2}(t)/\tau_{2}}{C}\right)^{B}.
\label{eq:case3_loss}
\end{align}
Using these functional forms, we now proceed to perform a local stability and bifurcation analysis for the systems given by \eqref{eq:modelb_1}, \eqref{eq:modelb_2} and \eqref{eq:modelb_3}. This would enable us to understand the dynamical properties of the coupled system of Compound TCP with Drop-Tail queue policy to a greater detail.
\section{Introduction}
Network performance, and end-to-end latency are affected by a combination of the choice of TCP, the size of router buffers, and the choice of queue management implemented in Internet routers \cite{Cerf},~\cite{Gettys},~\cite{Nichols}. A major portion of Internet traffic is controlled by the Transmission Control Protocol (TCP)~\cite{Ha},~\cite{Padhye}. There have been proposals for different flavours of TCP and queue management strategies. However, Compound TCP~\cite{Tan} is the default protocol in Windows, and a simple Drop-Tail queue management is commonly implemented in Internet routers. It has been argued that the default large buffer dimensioning rule for router buffers, combined with Drop-Tail, leads to excessive delays in the Internet \cite{Gettys}.
In our recent work~\cite{Jagannathan}, we conducted a performance evaluation of Compound TCP, in a small buffer regime, with particular emphasis on buffer thresholds. One of the key insights obtained therein was the two-fold advantage of having small router buffers. In particular, our analysis showed that small buffers are favourable for ensuring the stability of the system, in addition to reducing queuing delays. Furthermore, our analysis identified that the underlying dynamical systems undergo a Hopf bifurcation, and transit from a locally stable into an unstable regime as the buffer size increases. The Hopf bifurcation alerts us to the emergence of isolated periodic orbits, termed as limit cycles, as a parameter crosses a certain critical value. In addition, we repeatedly observed limit cycles in the queue size dynamics, in numerous packet-level simulations. Fig.~\ref{ns_single} portrays one such instance; indeed, it captures the emergence of limit cycles in the queue size of the core router in a single bottleneck topology. This motivates us to develop an analytical framework under which the emergence of these non-linear oscillations can be better understood. To that end, in this paper, we provide a complete analytical characterisation of the type of the Hopf bifurcation, and prove the orbital stability of the emergent limit cycles.
We consider three different topologies, and focus on analysing the dynamical properties of a fluid model of Compound TCP in conjunction with small Drop-Tail buffers. Our fluid model takes the form of a non-linear, time-delayed dynamical system. The first topology is a generalisation of the single bottleneck topology studied in~\cite{Jagannathan}, and consists of two sets of TCP flows having different round trip times, and feeding into a core router (see Fig.~\ref{Figure.1}(a)). The second topology corresponds to two queues in tandem, and consists of two distinct sets of TCP flows, regulated by a single edge router and feeding into a core router (see Fig.~\ref{Figure.1}(b)). The third topology comprises of two distinct sets of TCP flows, regulated by two separate edge routers, and feeding into a common core router (see Fig.~\ref{Figure.1}(c)).
For each of these cases, we first conduct a local stability analysis and outline necessary and sufficient conditions for local stability, with two simplifying assumptions. In the first scenario, we assume that the network parameters are the same, and that both sets of Compound TCP flows have equal round trip times. In the second scenario, we assume the network parameters to be heterogeneous, and the round trip time of one set of TCP flows to be much larger as compared to the other. If the local stability conditions get marginally violated, our analysis shows that the underlying systems would lose local stability via a Hopf bifurcation. Motivated by this insight, we then analyse only the third topology in greater detail, to better understand the impact of heterogeneous system parameters on local stability. We numerically show through DDE-BIFTOOL~\cite{DDE1},~\cite{DDE2} that, even in the presence of heterogeneous network parameters and different round trip times, the dynamical system undergoes a Hopf bifurcation which leads to the emergence of limit cycles.
As argued in~\cite{Jagannathan}, the emergence of limit cycles in the system dynamics could have a number of detrimental consequences -- for example it could lead to the synchronisation of TCP windows, result in a loss in link utilisation, and cause the downstream traffic to be bursty. Hence, it becomes imperative to study these limit cycles in further detail. To that end, an important contribution of this paper lies in providing an analytical framework to determine the asymptotic orbital stability of the emerging limit cycles. Using Poincar\'{e} normal forms and the center manifold theory, we show that the Hopf bifurcation is indeed supercritical, and hence leads to the emergence of orbitally stable limit cycles. To corroborate our analytical insights, we conduct some packet-level simulations in NS2~\cite{ns2}, to highlight the existence and stability of the limit cycles in the queue size dynamics. Notably, instead of treating any particular system parameter as the bifurcation parameter, we choose a suitably motivated exogenous, non-dimensional parameter as the bifurcation parameter to aid our analysis. The two main advantages of this are: first, it enables us to capture the effects of different system parameters on the system stability in a unified manner and secondly, we need not be concerned about the dimension of the bifurcation parameter.
The rest of the paper is as organised as follows. In section \ref{Models}, we outline the governing fluid models for the three cases we consider. Section \ref{Localstability} deals with local stability analysis of the fluid models. In Section \ref{sec:hopf}, we provide an analytical framework to determine the asymptotic orbital stability of the bifurcating limit cycles, and to characterise the type of the Hopf bifurcation. Packet-level simulations are presented in Section~\ref{simulations} to corroborate some of the analytical insights. Finally, in section \ref{conclusions}, we summarise our contributions.
\begin{figure}
\begin{center}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\psfrag{125}{\begin{scriptsize}$125$\end{scriptsize}}
\psfrag{aaaa}{\begin{scriptsize}Buffer size = 15 pkts\end{scriptsize}}
\psfrag{bbbb}{\begin{scriptsize}Buffer size = 100 pkts\end{scriptsize}}
\psfrag{cccc}{\hspace{-4mm} Queue size (pkts)}
\psfrag{xyz}{\hspace{-8mm}Time (seconds)}
\psfrag{mmmm}{\begin{scriptsize}Round trip time = $10$ ms\end{scriptsize}}
\psfrag{nnnn}{\begin{scriptsize}\hspace{-1mm}Round trip time = $200$ ms\end{scriptsize}}
\includegraphics[width=2.5in,height=3.5in,angle=270]{limitcycles.eps}
\caption{ \emph{Long-lived flows}. 60 long-lived Compound flows over a 2 Mbps link, and feeding into a core router with link capacity 100 Mbps. Observe the emergence of limit cycles in the queue at the core router, for larger buffer thresholds, and larger round trip times.}\vspace{-8mm}
\label{ns_single}
\end{center}
\end{figure}
\section{Models}
\label{Models}
\input{Models.tex}
\section{Local Stability Analysis}
\label{Localstability}
Note that, to perform a local stability and bifurcation analysis for the non-linear models \eqref{eq:modelb_1}, \eqref{eq:modelb_2} and \eqref{eq:modelb_3}, we need to choose an appropriate bifurcation parameter. It can be easily seen that both protocol and network parameters affect the stability of the systems. To that end, instead of treating any of the system parameters as a bifurcation parameter, we introduce an exogenous non-dimensional parameter $\kappa$ as the bifurcation parameter. We choose the non-dimensional parameter in such a manner that it does not affect the equilibrium of the system. Recall that, to conduct the local stability analysis, we primarily focus on Compound TCP with Drop-Tail queues in the small buffer regime. For mathematical tractability, we assume that both sets of TCP flows in all three topologies are regulated by Compound with identical protocol parameters. Further, we consider two simplifying assumptions as briefly outlined below:
\\\\
\emph{Scenario 1:} All network parameters are the same, \emph{i.e.}, $B_1=B_2=B$, and $C_1=C_2=C$. Further, the round trip times of both TCP flow sets are identical, \emph{i.e.}, $\tau_1=\tau_2=\tau.$
\\\\
\emph{Scenario 2:} In this scenario, we assume that all network parameters are distinct and the round trip time of one set of TCP flows is negligible and much smaller as compared to the round trip time of the other set, \emph{i.e}, $\tau_1>>\tau_2$ and $\tau_2\approx0$. Under this assumption, the dynamics of the second set of TCP flows appear almost instantaneous.\\
We now proceed to conduct a detailed local stability analysis to obtain bounds on network, and protocol parameters to ensure stability, for the systems given by \eqref{eq:modelb_1}, \eqref{eq:modelb_2} and \eqref{eq:modelb_3}.
\subsection*{Case I}
The schematic diagram of the topology is presented in Fig.~\ref{Figure.1}(a).
\subsubsection*{Scenario 1} With this assumption, the first model reduces to a single bottleneck link with only one set of TCP flows having round trip time $\tau_1=\tau_2=\tau$. Hence, with the non-dimensional bifurcation parameter $\kappa$, system \eqref{eq:modelb_1} reduces to the following non-linear, first-order, time-delayed differential equation:
\begin{align}
\frac{dw(t)}{dt} = \kappa\frac{w(t-\tau)}{\tau}\bigg(i\left(w(t)\right)\Big(1-q(w(t-\tau))\Big) - d\left((w(t)\right)q(w(t-\tau))\bigg),
\label{eq:modelb_1reduced1}
\end{align}
where $w(t)$ is the \emph{average} window size of the TCP flows. The non-trivial equilibrium $w^{*}$ of system \eqref{eq:modelb_1reduced1} satisfies the following equation
\begin{equation}
i(w^{*})=d(w^{*})q(w^{*}).
\end{equation}
Note that, under the first assumption, the fluid model for the packet loss probability at the core router, given by \eqref{eq:case1_loss} reduces to
\begin{align}
q(w^{*})=\left(\frac{w^{*}}{C\tau}\right)^B,
\end{align}
at equilibrium. A necessary and sufficient condition for this model, with Compound TCP in the small buffer regime is~\cite{Raja}
\begin{align}
\kappa\alpha \left(w^{*}\right)^{k-1}\sqrt{B^{2}-\left(\left(k-2\right)\left(1-q(w^{*})\right)\right)^{2}} < \cos^{-1}\left(\frac{\left(k-2\right)\left(1-q(w^*)\right)}{B}\right).
\label{eq:condition_case1}
\end{align}
\subsubsection*{Scenario 2}
With the introduction of the non-dimensional parameter $\kappa$, system \eqref{eq:modelb_2} becomes
\begin{align}
\dot{w}_1(t) & = \kappa\frac{w_{1}(t-\tau_{1})}{\tau_{1}}\bigg(i\left(w_{1}(t)\right)\left(1-q(t,\tau_1,\tau_2)\right) - d\left((w_{1}(t)\right)q(t,\tau_{1},\tau_{2})\bigg),\notag \\
\dot{w}_2(t) & = \kappa\frac{w_{2}(t)}{\tau_{2}}\bigg(i\left(w_{2}(t)\right)\left(1-q(t,\tau_1,\tau_2)\right) - d\left((w_{2}(t)\right)q(t,\tau_{1},\tau_{2})\bigg).
\label{eq:modelb_1reduced2}
\end{align}
Suppose $(w_1^{*}, w_{2}^{*})$ is a non-trivial equilibrium of \eqref{eq:modelb_1reduced2} and let $u_{1}(t)=w_{1}(t)-w_{1}^{*}$ and $u_{2}(t)=w_{2}(t)-w_{2}^{*}$ be small perturbations about $w_{1}^{*}$ and $w_{2}^{*}$ respectively. Linearising \eqref{eq:modelb_1reduced2} about this equilibrium, we obtain
\begin{align}
\label{eq:linearmodelb_1reduced2}
&\dot{u}_1(t) = -\kappa\left(\mathcal{M}_{1}u_{1}(t)+\mathcal{N}_{1}u_{1}(t-\tau_{1})+\mathcal{P}_{1}u_{2}(t)\right),\notag\\
&\dot{u}_2(t) = -\kappa\left(\big(\mathcal{M}_{2}+\mathcal{N}_{2}\big)u_{2}(t)+\mathcal{P}_{2}u_{1}(t-\tau_{1})\right),
\end{align}
where, the increase and decrease functions for Compound TCP given by \eqref{eq:increasedecrease}, and the functional form of the loss probability at the core router given by~\eqref{eq:case1_loss} yield the following coefficients:
\begin{align}
\label{eq:coefficients}
\mathcal{M}_{j} &=-\frac{\alpha}{\tau_{j}}\left(k-2\right)\ \left(w_{j}^{*}\right)^{k-1}\left(1-\frac{1}{C^B}\left(\frac{w_{1}^{*}}{\tau_1}+\frac{w_{2}^{*}}{\tau_2}\right)^{B}\right),\notag\\
\mathcal{N}_{j} &=\frac{ B \left(w_{j}^{*}\right)^{2}}{\tau_{j}^{2}\left(C\right)^{B}}\left(\alpha \left(w_{j}^{*}\right)^{k-2}+\beta\right)\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B-1},\notag\\
\mathcal{P}_{j}&=\frac{ B \left(w_{j}^{*}\right)^{2}}{\tau_{1}\tau_{2}\left(C\right)^{B}}\left(\alpha \left(w_{j}^{*}\right)^{k-2}+\beta\right)\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B-1},
\end{align}
for $j=1,2$. Looking for exponential solutions, we obtain the characteristic equation for the linearised system \eqref{eq:linearmodelb_1reduced2} as
\begin{align}
\label{eq:linearmodel1a2kappachar}
\lambda^2 + \kappa a\lambda +\kappa b\lambda e^{-\lambda \tau_1} + \kappa^2 c e^{-\lambda \tau_1} + \kappa^2 d=0.
\end{align}
where,
\begin{align}
\label{eq: abcd}
&a = \mathcal{M}_{1}+\mathcal{M}_{2}+\mathcal{N}_{2} \hspace{1ex}, \hspace{1ex} b=\mathcal{N}_{2},\notag\\
&c=\mathcal{M}_{1}\left(\mathcal{M}_{2}+\mathcal{N}_{2}\right), \hspace{2ex}d=\mathcal{N}_{1}\left(\mathcal{M}_{2}+\mathcal{N}_{2}\right)-
\mathcal{P}_{1}\mathcal{P}_{2}.
\end{align}
For system \eqref{eq:modelb_1reduced2} to be locally stable about the equilibrium $(w_{1}^{*},w_{2}^{*})$, all roots of the characteristic equation \eqref{eq:linearmodel1a2kappachar} should lie in the left half of the complex plane. It can be shown that, for negligibly small values of the non-dimensional parameter $\kappa$, the system is stable, \emph{i.e.}, all the roots would have negative real parts. However, as $\kappa$ is increased beyond a critical value, one pair of complex conjugate roots may cross over the imaginary axis, and hence have positive real parts. At this critical value the system would transit into an unstable region and have a pair of purely imaginary roots. To deduce this point, we substitute $\lambda=j\omega$ in \eqref{eq:linearmodel1a2kappachar} and separate real and imaginary parts to get
\begin{align*}
\omega^2 = \frac{\kappa^2(2c-a^2+b^2)}{2}\pm \frac{\kappa^2 \sqrt{(2c-a^2+b^2)^2-4(c^2-d^2)}}{2}.
\end{align*}
\emph{Condition 1}: There exists only one positive value of $\omega^2$ if the following conditions hold
\begin{itemize}
\item[(i)]$(2c-a^2+b^2)>0,\text{and}\hspace{1mm} (2c-a^2+b^2)^2=4(c^2-d^2)$
\item[(ii)]$(2c-a^2+b^2)>0,\text{and}\hspace{1mm} c^2-d^2<0$
\end{itemize}
\emph{Condition 2}: There exists two positive value of $\omega^2$ if the following condition holds
\begin{align*}
(2c-a^2+b^2)>0,\text{and}\hspace{1mm} (c^2-d^2)>0.
\end{align*}
When Condition 1 is satisfied, the system transits from the locally stable regime to instability as $\kappa$ increases beyond a critical value, and never regains stability as $\kappa$ is further increased. On the contrary, when Condition 2 is satisfied, the system may undergo stability switches as $\kappa$ is increased~\cite{Cooke}. In the context of congestion control algorithms, the stability switch phenomenon is an undesirable dynamical feature. Further, we have observed in numerous packet-level simulations that Compound TCP does not exhibit stability switches. Hence, we focus only on the case when Condition 1 is satisfied, and only one positive root of $\omega^2$ exists. This implies that there exists a cross over frequency at which one pair of complex conjugate roots crosses over the imaginary axis, and is given by $\omega=\kappa A$, where
\begin{align*}
A=\sqrt{\frac{(2c-a^2+b^2)}{2}+\frac{\sqrt{(2c-a^2+b^2)^2-4(c^2-d^2)}}{2}}.
\end{align*}
The critical value of $\kappa$ denoted by $\kappa_{c}$, at which this transition occurs, is given by
\begin{align}
\kappa_{c}=\frac{1}{A\tau_1}\cos^{-1}\Big(\frac{A^2(d-ab)-cd}{b^2A^2+c^2}\Big).
\label{eq:kappa_c2}
\end{align}
\subsection*{Case II}
The schematic diagram of the topology is illustrated in Fig.~\ref{Figure.1}(b).
\subsubsection*{Scenario 1}With this assumption, the second model reduces to a single set of TCP flows, regulated by an edge router, and feeding into a core router. Observe that, the loss probabilities at both routers are the same. Hence, with the non-dimensional bifurcation parameter $\kappa$, system \eqref{eq:modelb_1} reduces to the following non-linear, first-order, time-delayed differential equation
\begin{align}
\frac{dw(t)}{dt} = \kappa\frac{w(t-\tau)}{\tau}\bigg(i\left(w(t)\right)\Big(1-p(w(t-\tau))\Big) - d\left((w(t)\right)q(w(t-\tau))\bigg),
\label{eq:modelb2_reduced1}
\end{align}
where $w(t)$ is the average window size of the TCP flows. Using the functional forms of loss probabilities given by \eqref{eq:case2_loss}, we obtain
\begin{align*}
p(w^{*})=2q(w^{*})=2\left(\frac{w^{*}}{C\tau}\right)^B.
\end{align*}
The critical value of $\kappa$, at which system \eqref{eq:modelb2_reduced1} loses its stability, satisfies the following equation
\begin{align}
\kappa_{c}\alpha \left(w^{*}\right)^{k-1}\sqrt{B^{2}-\left(\left(k-2\right)\left(1-p(w^{*})\right)\right)^{2}} = \cos^{-1}\left(\frac{\left(k-2\right)\left(1-p(w^*)\right)}{B}\right).
\label{eq:condition_case2}
\end{align}
\subsubsection*{Scenario 2}With the introduction of the non-dimensional parameter $\kappa$, system \eqref{eq:modelb_2} reduces to
\begin{align}
\frac{dw_{1}(t)}{dt} =& \, \kappa \frac{w_{1}(t-\tau_{1})}{\tau_{1}}\bigg(i\left(w_{1}(t)\right)\Big(1-q_1(t,\tau_1,\tau_2)-q_2(t,\tau_1,\tau_2))\Big) - d\left((w_{1}(t)\right)\left(q_1(t,\tau_1,\tau_2)+q_2(t,\tau_1,\tau_2)\right)\bigg),\notag\\
\frac{dw_{2}(t)}{dt} =& \, \kappa \frac{w_{2}(t)}{\tau_{2}}\bigg(i\left(w_{2}(t)\right)\Big(1-q_1(t,\tau_1,\tau_2)-q_2(t,\tau_1,\tau_2))\Big) - d\left((w_{2}(t)\right)\left(q_1(t,\tau_1,\tau_2)+q_2(t,\tau_1,\tau_2)\right)\bigg).
\label{eq:modelb2_reduced2}
\end{align}
Linearising \eqref{eq:modelb2_reduced2} about its non-trivial equilibrium $(w_1^{*},w_2^{*})$, we obtain
\begin{align}
\label{eq:linearmodelb_2reduced2}
&\dot{u}_1(t) = -\kappa\left(\mathcal{M}_{1}u_{1}(t)+\mathcal{N}_{1}u_{1}(t-\tau_{1})+\mathcal{P}_{1}u_{2}(t)\right),\notag\\
&\dot{u}_2(t) = -\kappa\left(\big(\mathcal{M}_{2}+\mathcal{N}_{2}\big)u_{2}(t)+\mathcal{P}_{2}u_{1}(t-\tau_{1})\right),
\end{align}
where, for Compound TCP, the increase and decrease functions \eqref{eq:increasedecrease}, and the functional forms of the loss probabilities given by~\eqref{eq:case2_loss} yield the following coefficients
\begin{align}
\label{eq:coefficients_2}
\mathcal{M}_{j} & = -\frac{\alpha}{\tau_{j}}\left(k-2\right) \left(w_{j}^{*}\right)^{k-1}\Bigg(1-\bigg(\frac{1}{C_{1}^{B_{1}}}\bigg(\frac{w_1^*}{\tau_1}+\frac{w_2^*}{\tau_2}\bigg)^{B_1} -\frac{1}{C_{2}^{B_{2}}}\bigg(\frac{w_1^*}{\tau_1}+\frac{w_2^*}{\tau_2}\bigg)^{B_2}\Bigg),\notag\\
\mathcal{N}_{j} & = \left(\alpha\left(w_{j}^{*}\right)^{k-1}+\beta w_{j}^{*}\right)\frac{w_{j}^{*}}{\tau_j}\Bigg(\frac{B_{1}}{C_{1}^{B_{1}}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B_{1}-1} + \frac{B_2}{C_{2}^{B_{2}}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B_{2}-1}\Bigg),\notag\\
\mathcal{P}_{j} & = \left(\alpha\left(w_{j}^{*}\right)^{k-1}+\beta w_{j}^{*}\right)\frac{w_{j}^{*}}{\tau_1\tau_2}\Bigg(\frac{B_{1}}{C_{1}^{B_{1}}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B_{1}-1} +\frac{B_2}{C_{2}^{B_{2}}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B_{2}-1}\Bigg),
\end{align}
for $j=1,2.$ Observe that, the linearised system \eqref{eq:linearmodelb_2reduced2} has a similar form as \eqref{eq:linearmodelb_1reduced2}. Hence, conducting a similar kind of analysis as done for system \eqref{eq:modelb2_reduced2}, we obtain the critical value of the non-dimensional parameter $\kappa$, as given by \eqref{eq:kappa_c2}.
\subsection*{Case III}
The schematic diagram for this topology is illustrated in Fig.~\ref{Figure.1}(c).
\subsubsection*{Scenario 1}For Compound TCP in the small buffer regime, the critical value of $\kappa$, denoted by $\kappa_c$, at which system~\eqref{eq:modelb_3} transits into a locally unstable regime, satisfies the following condition:
\begin{align*}
\kappa_c\alpha \left(w^{*}\right)^{k-1}\sqrt{B^{2}
-(k-2)^2\left(1-\left(1+2^B\right)p(w^{*})\right)^{2}} < \cos^{-1}
\left(\frac{(k-2)\left(1-\left(1+2^B\right)p(w^*)\right)}{B}\right).
\end{align*}
\begin{figure*}
\begin{center}
\begin{minipage}[b]{0.45\textwidth}
\label{fig:alpha_kappa}
\psfrag{b}{Non-dimensional parameter, $\kappa$}
\psfrag{T}{Protocol parameter, $\alpha$}
\psfrag{h}{\begin{scriptsize}Hopf condition\end{scriptsize}}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{2}{\begin{scriptsize}$2$\end{scriptsize}}
\psfrag{1}{\begin{scriptsize}$1$\end{scriptsize}}
\psfrag{0.3}{\begin{scriptsize}$0.3$\end{scriptsize}}
\includegraphics[width=2in,height=3in,angle=270]{stabilitychart_alpha_kappa.eps}
\end{minipage}
\hspace*{5mm}
\begin{minipage}[b]{0.45\textwidth}
\label{fig:buffer_kappa}
\psfrag{b}{Non-dimensional parameter, $\kappa$}
\psfrag{T}{Core router buffer, $B$}
\psfrag{h}{\begin{scriptsize}Hopf condition\end{scriptsize}}
\psfrag{0.9}{\begin{scriptsize}$0.9$\end{scriptsize}}
\psfrag{1.2}{\begin{scriptsize}$1.2$\end{scriptsize}}
\psfrag{1.5}{\begin{scriptsize}$1.5$\end{scriptsize}}
\psfrag{16}{\begin{scriptsize}$16$\end{scriptsize}}
\psfrag{22}{\begin{scriptsize}$22$\end{scriptsize}}
\psfrag{28}{\begin{scriptsize}$28$\end{scriptsize}}
\includegraphics[width=2in,height=3in,angle=270]{stabilitychart_buffer_kappa.eps}
\end{minipage}
\caption {\emph{Stability chart.} Hopf condition for~\eqref{eq:modelb_3} with Compound TCP in the small buffer regime with respect to two sets of parameters: (a) the non-dimensional parameter $\kappa$, and the protocol parameter $\alpha$, (b) the non-dimensional parameter $\kappa$, and the buffer size at the core router $B$. The shaded region below the Hopf condition curve represents the stable region.}
\label{fig:charts}
\end{center}
\end{figure*}
\subsubsection*{Scenario 2}With the introduction of the non-dimensional parameter $\kappa$, \eqref{eq:modelb_3} reduces to
\begin{align}
\frac{dw_{1}(t)}{dt} & = \kappa\frac{w_{1}(t-\tau_{1})}{\tau_{1}}\bigg(i\left(w_{1}(t)\right)\Big(1-p_1(t-\tau_1)-q(t,\tau_1,\tau_2))\Big) - d\left((w_{1}(t)\right)\left(p_1(t-\tau_1)+q(t,\tau_1,\tau_2)\right)\bigg),\notag\\
\frac{dw_{2}(t)}{dt} & = \kappa \frac{w_{2}(t)}{\tau_{2}}\bigg(i\left(w_{j}(t)\right)\Big(1-p_2(t)-q(t,\tau_1,\tau_2))\Big) - d\left((w_{j}(t)\right)\left(p_2(t)+q(t,\tau_1,\tau_2)\right)\bigg).
\label{eq:modelb3_reduced2}
\end{align}
Linearising \eqref{eq:modelb3_reduced2} about its equilibrium $(w_{1}^{*},w_{2}^{*})$, we obtain
\begin{align}
\label{eq:linearmodelb_3reduced2}
&\dot{u}_1(t) = -\kappa\left(\mathcal{M}_{1}u_{1}(t)+\mathcal{N}_{1}u_{1}(t-\tau_{1})+\mathcal{P}_{1}u_{2}(t)\right),\notag\\
&\dot{u}_2(t) = -\kappa\left(\big(\mathcal{M}_{2}+\mathcal{N}_{2}\big)u_{2}(t)+\mathcal{P}_{2}u_{1}(t-\tau_{1})\right),
\end{align}
where, for Compound TCP, and the functional forms of the loss probabilities given by~\eqref{eq:case3_loss} yield the following coefficients
\begin{align}
\label{eq:coefficients3}
\mathcal{M}_{j} & = -\frac{\alpha}{\tau_{j}}\left(k-2\right) \left(w_{j}^{*}\right)^{k-1}\Bigg(1-\bigg(\frac{w_j^*}{C_j\tau_j}\bigg)^{B_j} -\frac{1}{C^B}\bigg(\frac{w_1^*}{\tau_1}+\frac{w_2^*}{\tau_2}\bigg)^{B}\Bigg),\notag\\
\mathcal{N}_{j} & = \left(\alpha\left(w_{j}^{*}\right)^{k-1}+\beta w_{j}^{*}\right)\Bigg(\frac{B_{j}}{\tau_j}\left(\frac{w_{j}^{*}}{C_j \tau_j}\right)^{B_j} +\frac{B \left(w_{j}^{*}\right)^{2}}{C^{B}\tau_{j}^{2}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B-1}\Bigg),\notag\\
\mathcal{P}_{j} & = \left(\alpha\left(w_{j}^{*}\right)^{k-1}+\beta w_{j}^{*}\right)\frac{B w_{j}^{*}}{\tau_{1}\tau_{2}\left(C\right)^{B}}\left(\frac{w_{1}^{*}}{\tau_{1}}+\frac{w_{2}^{*}}{\tau_{2}}\right)^{B-1},
\end{align}
for $j=1,2.$ Note that, the linearised system \eqref{eq:linearmodelb_3reduced2} has a similar form as \eqref{eq:linearmodelb_1reduced2}. Hence, a similar kind of local stability analysis would yield the condition on the critical value of the non-dimensional parameter $\kappa$, and the protocol parameters as given by \eqref{eq:kappa_c2}.
\\
\indent For all three scenarios, with the simplifying assumptions, the conditions derived above essentially capture the interdependence among the non-dimensional parameter $\kappa$, and the system parameters to ensure local stability. Observe that, the loss of local stability can be studied with respect to any system parameter. However, we prefer
to choose an exogenous parameter as the bifurcation parameter, to aid our analysis. It can be explicitly shown that, for all the above cases, the system loses local stability via a Hopf bifurcation~\cite{Hassard} if the conditions derived above get violated. We prove this by verifying that the transversality condition of the Hopf spectrum \cite{Hassard}. To verify this, we show that, $\mathrm{Re}(\mathrm{d}\lambda/\mathrm{d}\kappa)\neq0$ at $\kappa=\kappa_c$. In particular, we prove that, $\mathrm{Re}(\mathrm{d}\lambda/\mathrm{d}\kappa)>0$ at $\kappa=\kappa_c$. This implies that, one pair of complex conjugate roots crosses over the imaginary axis from the left half of the complex plane to the right half. Thus, the system undergoes a Hopf bifurcation at $\kappa=\kappa_c$. Hence, $\kappa<\kappa_c$ is a necessary and sufficient condition for local stability, for all the three scenarios.\\
\indent Observe that, deriving a necessary and sufficient condition with heterogeneous network parameters, and different round trip times is analytically complex, for all three scenarios discussed earlier. Hence, we numerically illustrate through DDE-BIFTOOL version 2.03 ~\cite{DDE1},~\cite{DDE2}, that system \eqref{eq:modelb_3} undergoes a Hopf bifurcation if the non-dimensional parameter $\kappa$ is varied beyond a certain critical value. We fix the protocol parameters as follows: $\alpha=0.3$, $\beta=0.5$ and $k=0.75$. Since, we mainly focus on small buffer regime, the buffer sizes of the routers are fixed as: $B_1=10, B_2=15$, and $B=25$. We fix the remaining network parameters as: $C_1=C_2=100$, $C=180$, $\tau_1=1$ and $\tau_2=2$. Now, we vary the non-dimensional parameter $\kappa$ in the range $[0,2]$ and observe that the system undergoes a Hopf bifurcation at $\kappa_c=1$. At this point, the system has one pair of complex conjugate roots on the imaginary axis. Consequently, the system dynamics exhibit limit cycles at $\kappa_c=1$.
\subsubsection*{Stability charts}To obtain insights about the system behaviour at the stability boundary, we now demonstrate some stability charts for system \eqref{eq:modelb_3}. Fig.~\ref{fig:charts}~(a) represents the Hopf condition for system \eqref{eq:modelb_3} in the two parameter space: the non-dimensional parameter $\kappa$, and the protocol parameter $\alpha$. Observe that, if $\kappa$ is increased, $\alpha$ would have to reduce to ensure stability. Fig.~\ref{fig:charts}~(b) illustrates the Hopf condition in the two parameter space: the non-dimensional parameter $\kappa$, and the buffer size at the core router $B$. Observe that, if $\kappa$ is increased, keeping other system parameters fixed, $B$ would have to be decreased accordingly to ensure stability of system~\eqref{eq:modelb_3}. Fig.~\ref{fig:stability_chart} characterises the stability boundary of system~\eqref{eq:modelb_3} with respect to the increase protocol parameters $\alpha$ and $k$. It is evident that, there exists a trade-off between the increase parameters to ensure stability. Hence, we conclude that, both protocol parameters, and network parameters, need to be co-designed carefully to maintain stability of system~\eqref{eq:modelb_3}. If these Hopf conditions get violated, the system would lose stability leading to the emergency of limit cycles in the system dynamics. In the next section, we provide a detailed analytical framework to characterise the \emph{type} of Hopf bifurcation and the asymptotic \emph{orbital stability} of the emergent limit cycles, for system \eqref{eq:modelb_3}.
\section{Hopf Bifurcation Analysis}
\label{sec:hopf}
We have seen that, variation in the exogenous parameter $\kappa $ induces instability in system. Instability in the system could be induced by any of the system parameters. This loss of stability occurs via a Hopf bifurcation which results in limit cycles in the system dynamics which in turn leads to deterministic oscillations in the queue size. Consequently, this results in the overall degradation of the system performance because of loss in link utilisation. To that end, it becomes imperative to study the type of bifurcation and the stability of these emergent limit cycles to a greater detail.
Note that, we have motivated the exogenous, non-dimensional parameter $\kappa$ as the bifurcation parameter. This enables us to capture the effect of the different system parameters on the system stability in a unified manner. The Hopf bifurcation analysis enables us to analyse the system dynamics in its locally unstable regime, in the neighbourhood of the Hopf condition. Using Poincar\'{e} normal forms and the center manifold theory, we present an analytical framework to determine the \emph{type} of the Hopf bifurcation and the orbital stability of the emergent limit cycles. Our analysis closely follows the analysis presented in \cite{Hassard,Kuznetsov,Gaurav}.
Let $\kappa=\kappa_c+\mu$, where $\mu \in \mathbb{R}.$ Observe that, the system undergoes a Hopf bifurcation at $\mu=0$, where $\kappa=\kappa_c$. We can now consider $\mu$ as the bifurcation parameter. An incremental change in $\kappa$ from $\kappa_c$ to $\kappa_c+\mu$ where $\mu>0$, pushes the system to its locally unstable regime.
\emph{Step 1}: Using Taylor series expansion, we segregate the right hand side of \eqref{eq:modelb_3} into linear and non-linear parts. We then cast this equation into the standard form of an operator differential equation.
\emph{Step 2}: At the critical value of the bifurcation parameter, \emph{i.e.} at $\mu=0$, the system has exactly one pair of purely imaginary eigenvalues with non-zero angular velocity. The linear eigenspace spanned by the eigenvectors corresponding these eigenvalues is called the critical eigenspace. The center manifold theorem~\cite[Chapter $5$, Theorem $5.1.$]{Kuznetsov} guarantees the existence of a locally invariant $2-$dimensional manifold which is tangent to the critical eigenspace at the equilibrium of the system.
\emph{Step 3}: Next, we project the system onto its critical eigenspace and its complement at the critical value of the bifurcation parameter. This enables us to capture the dynamics of the system on the center manifold, with the help of an ordinary differential equation in a single complex variable.
\emph{Step 4}: Finally, using Poincar\'{e} normal forms, we evaluate the lyapunov coefficient and the floquet exponent, which characterise the type of the Hopf bifurcation and the asymptotic orbital stability of the emergent limit cycles respectively.
Suppose $(w_1^*,w_2^*)$ is an equilibrium for \eqref{eq:modelb_3}. Let $u_1(t)=w_1(t)-w_1^{\ast}$ and $u_2(t)=w_2(t)-w_2^{\ast}$ be small perturbations about the equilibrium. Thus, a Taylor series expansion of \eqref{eq:modelb_3} about its equilibrium $\left(w_1^{\ast},w_2^{\ast}\right)$ is as follows
\begin{align}
\dot{u_1}(t) = & \, \kappa \Big( \xi_{a} u_1(t)+ \xi_{b} u_1(t-\tau_ 1)+\xi_{d}u_2(t-\tau_2) \Big) +\kappa\Big(\xi_{aa}u_1^2(t)+\xi_{bb}u_1^2(t-\tau_1) +\xi_{dd}u_2^2(t-\tau_2) \notag \\
& + \xi_{ab}u_1(t)u_1(t-\tau_1)+\xi_{ad}u_1(t)u_2(t-\tau_2) + \xi_{bd}u_1(t-\tau_1)u_2(t-\tau_2)\Big) +\kappa\Big(\xi_{aaa}u_1^3(t) \notag \\
& +\xi_{bbb}u_1^3(t-\tau_1)+\xi_{ddd}u_2^3(t-\tau_2) + \xi_{aab}u_1^2(t)u_1(t-\tau_1) +\xi_{aad}u_1^2(t)u_2(t-\tau_2) + \xi_{abb}u_1(t)u_1^2(t-\tau_1) \notag \\
& +\xi_{bbd}u_1^2(t-\tau_1)u_2(t-\tau_2)+
\xi_{add}u_1(t)u_2^2(t-\tau_2)+\xi_{bdd}u_1(t-\tau_1)u_2^2(t-\tau_2) \notag \\
& + \xi_{acd}u_1(t)u_1(t-\tau_1)u_2(t-\tau_2)\Big),\notag \\
\dot{u_2}(t) = & \, \kappa \Big( \chi_{c} u_2(t)+ \chi_{d} u_2(t-\tau_ 2)+\chi_{b}u_1(t-\tau_1) \Big) +\kappa \Big(\chi_{cc}u_2^2(t)+\chi_{dd}u_2^2(t-\tau_2) + \chi_{bb}u_1^2(t-\tau_1)\notag\\
& + \chi_{cd}u_2(t)u_2(t-\tau_2)+\chi_{bc}u_1(t-\tau_1)u_2(t) + \chi_{bd}u_1(t-\tau_1)u_2(t-\tau_2)\Big) +\kappa\Big(\chi_{ccc}u_2^3(t) \notag \\
& +\chi_{ddd}u_2^3(t-\tau_2)+\chi_{bbb}u_1^3(t-\tau_1) + \chi_{ccd}u_2^2(t)u_2(t-\tau_2)+\chi_{bcc}u_1(t-\tau_1)u_2^2(t) + \chi_{cdd}u_2(t)u_2^2(t-\tau_2) \notag \\
& +\chi_{bdd}u_1(t-\tau_1)u_2^2(t-\tau_2) +\chi_{bbc}u_1^2(t-\tau_1)u_2(t)+\chi_{bbd}u_1^2(t-\tau_1)u_2(t-\tau_2)\notag\\
& + \chi_{bcd}u_1(t-\tau_1)u_2(t)u_2(t-\tau_2)\Big).\label{eq:TaylorSeries}
\end{align}
\begin{figure}
\begin{center}
\psfrag{T}{$k$}
\psfrag{b}{$\alpha$}
\psfrag{0.1}{\begin{scriptsize}$0.1$\end{scriptsize}}
\psfrag{1.0}{\begin{scriptsize}$1.0$\end{scriptsize}}
\psfrag{0.5}{\begin{scriptsize}$0.5$\end{scriptsize}}
\psfrag{SSSSSSSSSSSSSS}{\hspace{2mm}\begin{scriptsize}Stable Region\end{scriptsize}}
\psfrag{h}{\begin{scriptsize}Hopf Condition\end{scriptsize}}
\includegraphics[width=1.8in,height=2.6in,angle=270]{stability.eps}
\caption{\emph{Stability chart.} Hopf condition for system \eqref{eq:modelb_3}. The shaded region below the curve denotes the stable region. It is evident that there exists a trade-off between the increase protocol parameters $\alpha$ and $k$. As $\alpha$ increases, $k$ has to be decreased to ensure system stability. }\label{fig:stability_chart}\vspace{-8mm}
\end{center}
\end{figure}
The Taylor series coefficients are given in Table \ref{table:coefficients}. Using the notation $\mathbf{u} = [u_1 \hspace{4pt} u_2]^T$, we reduce equation \eqref{eq:TaylorSeries} to the following form
\begin{align}
\dot{\mathbf{u}}(t) = \mathcal{L}_{\mu}\mathbf{u}_{t} + \mathcal{F}(\mathbf{u}_{t},\mu),\label{eq:nonlinear}
\end{align}
where $t>0, \mu \in \mathbb{R}$. For $\tau>0$,we define
\begin{align*}
\mathbf{u}_{t}(\theta)=\mathbf{u}(t+\theta), \hspace{2ex} \mathbf{u}_{t}:[-\tau,0] \rightarrow \mathbb{R}^2, \hspace{2ex} \theta \in [-\tau,0].
\end{align*}
For this model, without loss of generality, we assume that $\tau_1>\tau_2$. $\mathcal{L}: C[-\tau_1,0] \rightarrow \mathbb{R}^{2}$ denotes a family of continuous and bounded functions parametrised by $\mu$. Here, $C[a,b]$ denotes the set of all continuous functions on the interval $[a,b]$. The operator $\mathcal{F}:C[-\tau_1,0] \rightarrow \mathbb{R}^{2}$ consists of the non-linear terms. Further, we assume that $\mathcal{F}$ is analytic and both $\mathcal{L}$ and $\mathcal{F}$ depend analytically on the bifurcation parameter $\mu$ for small $|\mu|$. The linear operator is
\begin{align}
\label{eq:linearop}
\mathcal{L}_{\mu} \mathbf{u}_{t} = \kappa \begin{bmatrix}
\xi_{a}u_1(t)+\xi_{b}u_1(t-\tau_1) & \xi_{d}u_2(t-\tau_2)\\
\chi_{b}u_1(t-\tau_1) & \chi_{c}u_2(t)+\chi_{d}u_2(t-\tau_2) \\
\end{bmatrix}.
\end{align}
We now cast equation \eqref{eq:nonlinear} into the following standard form of an operator differential equation,
\begin{align}
\label{eq:system}
\dot{\mathbf{u}} = \mathcal{A}(\mu)\mathbf{u}_{t}+\mathcal{R}\mathbf{u}_{t}.
\end{align}
Note that, \eqref{eq:system} has $\mathbf{u}_{t}$ rather than both $\mathbf{u}_{t}$ and $\mathbf{u}$. Now, using the Riesz representation theorem~\cite[Chapter $6,$ Theorem $6.19.$]{Rudin}, we transform the linear problem $(\mathrm{d}/\mathrm{d}t)\mathbf{u}(t)=\mathcal{L}_{\mu}\mathbf{u}_{t}$. The Riesz representation theorem guarantees the existence of an $2\times 2$ matrix-valued measure $\boldsymbol{\eta}(\cdot,\mu):[-\tau_1,0]\rightarrow\mathbb{R}^{4}$, such that each component of $\boldsymbol{\eta}$ has bounded variation and for all $\phi \in C[-\tau_1,0],$
\begin{align*}
\mathcal{L}_{\mu} \phi =& \int_{\theta = -\tau_1}^{0} \mathrm{d} \boldsymbol{\eta}(\theta, \mu) \phi(\theta).
\end{align*}
In particular, we have
\begin{align*}
\mathcal{L}_{\mu} \mathbf{u}_{t} =& \int_{\theta = -\tau_1}^{0} \mathrm{d} \boldsymbol{\eta}(\theta, \mu) \mathbf{u}_{t}(\theta).\notag
\end{align*}
Observe that, for system \eqref{eq:modelb_3}, the matrix $\mathrm{d}\boldsymbol{\eta}$ is \begin{align}
\label{eq:etamatrix}
\mathrm{d} \boldsymbol{\eta}(\theta,\mu) = \kappa \begin{bmatrix}
\xi_{a}\delta(\theta)+\xi_{b}\delta(\theta+\tau_1) & \xi_{d}\delta(\theta+\tau_2)\\
\chi_{b}\delta(\theta+\tau_1) & \chi_{c}\delta(\theta)+\chi_{d}\delta(\theta+\tau_2) \\
\end{bmatrix}\mathrm{d}\theta.
\end{align}
Here, $\delta(\cdot)$ is the Dirac-delta measure. Let $C^{1}[-\tau_1,0]$ denotes the space of all functions defined on $[-\tau_1,0]$, with continuous first derivatives. For $\phi \in C^{1}[-\tau_1,0],$ we then define the following linear and non-linear operators
\begin{align}
\mathcal{A}(\mu) \mathbf{u}_{t}(\theta) &= \left\lbrace \begin{array}{lr}
\frac{\mathrm{d} \mathbf{u}_{t}(\theta)}{\mathrm{d}\theta}, & \theta \in [-\tau_1,0), \\
\mathcal{L}_{\mu}\mathbf{u}_{t}, & \theta = 0.
\end{array} \right. \notag\\
\mathcal{R}\mathbf{u}_{t}(\theta) &= \left \{ \begin{array}{cl}
0, & \theta \in [-\tau_1,0),\\
\mathcal{F}(\mathbf{u}_{t},\mu), & \theta = 0.
\end{array} \right. \label{eq:DLEoperatorR}
\end{align}
Note that, $\mathrm{d}\mathbf{u}_{t}/\mathrm{d}\theta \equiv \mathrm{d}\mathbf{u}_{t}/\mathrm{d}t$. Hence, equation \eqref{eq:nonlinear} can be transformed into \eqref{eq:system}. Further, recall that, $\kappa=\kappa_{c}+\mu$, and the system undergoes bifurcation at the critical point $\mu=0$. Hence, we fix $\mu=0$ to perform the necessary analysis at the point of bifurcation. At $\mu=0$, the system has a pair of complex eigenvalues on the imaginary axis: $\lambda=\pm i\omega_0$, where $\omega_0>0.$ Let $\mathbf{q}(\theta)$ denote the eigenvector for $\mathcal{A}(0)$ corresponding to the eigenvalue $\lambda(0)=i\omega_0$. We assume that $\mathbf{q}(\theta)$ has a form as
\begin{align*}
\mathbf{q}(\theta) &= \begin{bmatrix}1 & \phi_1 \end{bmatrix}^{T}e^{i\omega_0\theta}.
\end{align*}
Now, using
\begin{equation*}
\mathcal{A}(0)\mathbf{q}(\theta) = i\omega_{0}\mathbf{q}(\theta),
\end{equation*}
we obtain $\phi_1$ as
\begin{align*}
\phi_1=\frac{-\kappa \chi_{b}e^{-i\omega_0 \tau_1}}{\kappa \chi_{c}+\kappa \chi_{d}e^{-i\omega_0 \tau_2}-i\omega_0}.
\end{align*}
We now define the following adjoint operator
\begin{equation*}
\mathcal{A}^{\ast}(\mu)\alpha(s) = \begin{cases} \begin{array}{ll}-\frac{\mathrm{d}\alpha(s)}{\mathrm{d}s},&s \in (0,\tau_1],\\\int_{t = -\tau_1}^{0}\mathrm{d}\boldsymbol{\eta}^{T}(t,0)\alpha(-t),&s=0.\end{array}\end{cases}
\end{equation*}
where $\boldsymbol{\eta}^{T}$ denotes the transpose of $\boldsymbol{\eta}$. Observe that, the domains of $\mathcal{A}$ and $\mathcal{A}^{\ast}$ are $C^{1}[-\tau_1,0]$ and $C^{1}[0,\tau_1]$ respectively. Then, $\bar{\lambda}(0)=-i\omega_0$ is an eigenvalue of $\mathcal{A}^\ast$ and for some non-zero vector $\mathbf{p}$, we have
\begin{equation}
\label{eq:A*}
\mathcal{A}^{\ast}(0)\boldsymbol{\mathrm{p}}(\zeta) = -i\omega_{0}\boldsymbol{\mathrm{p}}(\zeta).
\end{equation}
We consider $\mathbf{p}(\theta)$ to have the following form:
\begin{align*}
\mathbf{p}(\theta) &= D\begin{bmatrix}\phi_2 & 1\end{bmatrix}^{T}e^{i\omega_0\theta}.
\end{align*}
Using \eqref{eq:A*}, we obtain $\phi_2$ as
\begin{align*}
\phi_2=\frac{-\kappa \chi_{b}e^{i\omega_0\tau_1}}{\kappa \xi_{a}+\kappa \xi_{b}e^{i\omega_0\tau_1}+i\omega_0}.
\end{align*}
Let us define the inner-product of the functions $\boldsymbol{\psi} \in C[0,\tau_1]$ and $\boldsymbol{\phi} \in C[-\tau_1,0]$ as
\begin{align}
\left\langle \boldsymbol{\psi}, \boldsymbol{\phi} \right\rangle =& \boldsymbol{\bar{\psi}}(0) \boldsymbol{\phi}(0) - \int_{\theta = -\tau_1}^{0} \int_{\zeta=0}^{\theta} \overline{\boldsymbol{\psi}}^{T}(\zeta - \theta) \mathrm{d}\boldsymbol{\eta}(\theta, \mu) \boldsymbol{\phi}(\zeta)\mathrm{d}\zeta. \label{eq:InnerProductDefn}
\end{align}
Using the above definition of inner product, we can easily verify that the eigenvectors $\mathbf{p}$ and $\mathbf{q}$ satisfy the conditions $\left\langle \mathbf{p},\mathbf{q}\right\rangle = 1$ and $\left\langle \mathbf{p},\mathbf{\bar{q}}\right\rangle = 0$ when
\begin{align}
D &=\Big(\phi_{2}\left(1+\kappa \xi_{b}\tau_1 e^{-i\omega_0\tau_1}+\kappa \xi_{d}\phi_1 \tau_2 e^{-i\omega_0 \tau_2}\right)+\phi_{1}\left(1+\kappa \chi_{d}\tau_2 e^{-i\omega_0\tau_2}+
\kappa \chi_{b}\tau_1 e^{-i\omega_0 \tau_1}\right)\Big)^{-1}.
\label{eq:exprn_B}
\end{align}
The critical eigenspace corresponding to the pair of eigenvalues $\pm i\omega_0$, denoted by $T_c$, is now $2-$dimensional and is spanned by $\lbrace\mathrm{Re}\hspace{0.25ex}\boldsymbol{\mathrm{q}}, \hspace{0.5ex}\mathrm{Im}\hspace{0.25ex}\boldsymbol{\mathrm{q}}\rbrace,$ where $\mathrm{Re}\hspace{0.25ex}\boldsymbol{\mathrm{q}}$ and $\mathrm{Im}\hspace{0.25ex}\boldsymbol{\mathrm{q}}$ denote the real and imaginary parts of $\boldsymbol{\mathrm{q}}$ respectively. Further, we denote the complement of the critical eigenspace $T_c$ as $T_{su}$. We now project system \eqref{eq:system} onto $T_c$ and $T_{su}$. For $\mathbf{u}_{t}$, a solution of \eqref{eq:system} at $\mu = 0$, define
\begin{align}
z(t) = \langle \mathbf{p},\mathbf{u}_{t}\rangle ,\hspace{3mm} \text{and} \hspace{3mm} \mathbf{w}(t,\theta) &= \mathbf{u}_{t}(\theta)- 2 \text{Re}\big(z(t)\mathbf{q}(\theta)\big).
\label{eq:eigenbasis}
\end{align}
Recall that, the center manifold, $C_0$ is tangent to the critical eigenspace at the equilibrium. The representation of the center manifold is
\begin{align}
\mathbf{w}(t,\theta) &= \mathbf{w}\big(z(t),\bar{z}(t),\theta\big),\hspace{1ex}\text{where}\notag\\
\mathbf{w}(z,\bar{z},\theta) &= \mathbf{w}_{20}(\theta)\frac{z^2}{2} + \mathbf{w}_{11}(\theta)z\bar{z} + \mathbf{w}_{02}(\theta)\frac{\bar{z}^2}{2} + \cdots.\label{eq:vector_w}
\end{align}
Here, $\mathbf{w}_{ij}(\theta),\hspace{0.5ex}\text{for all}\hspace{0.5ex}i,j\in \lbrace 0,1,2\rbrace$ is a two dimensional vector given as
\begin{align*}
\mathbf{w}_{ij}(\theta)=\begin{bmatrix}
w_{ij1}(\theta) & w_{ij2}(\theta)
\end{bmatrix}^{T}.
\end{align*}
We observe that, $z$ and $\bar{z}$ are the local coordinates on the manifold $C_{0}$ in the direction of the eigenvectors $\mathbf{p}$ and $\bar{\mathbf{p}}$ respectively. Further, note that the existence of the center manifold $C_0$ ensures that equation \eqref{eq:system} can now be reduced to an ordinary differential equation for a single complex variable $z$ on $C_0$. At $\mu = 0$, in the coordinates $\eqref{eq:eigenbasis},$ the dynamics of $z$ can be represented as
\begin{align}
\dot{z}(t) &= \langle \mathbf{p},\mathcal{A}\mathbf{u_{t}} + \mathcal{R}\mathbf{u_{t}}\rangle \notag\\
& = i\omega_{0}z(t) + \bar{\mathbf{p}}(0)\cdot \mathcal{F}\Big(\mathbf{w}(z,\bar{z},\theta)+ 2\text{Re}\big(z(t)\mathbf{q}(\theta)\big)\Big)\notag\\
& = i\omega_{0}z(t) + \bar{\mathbf{p}}(0)\cdot \mathcal{F}_{0}(z,\bar{z}) \notag\\
& = i\omega_{0}z(t) + g(z,\bar{z}).\label{eq:derivative_z}
\end{align}
Now, we can expand the function $g(z,\bar{z})$ in powers of $z$ and $\bar{z}$ as
\begin{align}
g(z,\bar{z}) =&\,\, g_{20}\frac{z^{2}}{2} + g_{11}z \bar{z} + g_{02}\frac{\bar{z}^{2}}{2}+ g_{21}\frac{z^{2}\bar{z}}{2}+ \cdots. \label{eq:vector_g}
\end{align}
We now need to determine the coefficients $\mathbf{w}_{11}(\theta)$, $\mathbf{w}_{20}(\theta)$, $\mathbf{w}_{02}(\theta)$ in equation \eqref{eq:vector_w} to solve the differential equation \eqref{eq:derivative_z} for $z$. Following \cite{Hassard} we can write $\dot{\mathbf{w}} = \dot{\mathbf{u}}_{t} - \dot{z}\mathbf{q} - \dot{\bar{z}}\bar{\mathbf{q}},$ and using \eqref{eq:system} and \eqref{eq:derivative_z} we obtain
\begin{equation*}
\dot{\mathbf{w}} = \begin{cases}
\begin{array}{l l}
\mathcal{A}\mathbf{w} - 2\text{Re}\big(\bar{\mathbf{p}}(0)\cdot\mathcal{F}_0\mathbf{q}(\theta)\big),& \theta \in [-\tau_1,0),\\
\mathcal{A}\mathbf{w} - 2\text{Re}\big(\bar{\mathbf{p}}(0)\cdot\mathcal{F}_0\mathbf{q}(0)\big)+\mathcal{F}_0, & \theta = 0,
\end{array}
\end{cases}
\end{equation*}
which, using \eqref{eq:vector_w}, can be rewritten as
\begin{equation}
\dot{w} = \mathcal{A}\mathbf{w} + \mathbf{H}(z,\bar{z},\theta).\label{eq:derivative_w}
\end{equation}
Here, the function $\mathbf{H}(z,\bar{z},\theta)$ can be expanded in powers of $z$ and $\bar{z}$ as
\begin{equation}
\mathbf{H}(z,\bar{z},\theta) = \mathbf{H}_{20}(\theta)\frac{z^{2}}{2} + \mathbf{H}_{11}(\theta)z\bar{z} + \mathbf{H}_{02}(\theta)\frac{\bar{z}^{2}}{2} + \cdots.\label{eq:vector_H}
\end{equation}
Here, $\mathbf{H}_{ij}(\theta),\hspace{0.5ex}\text{for all}\hspace{0.5ex}i,j\in \lbrace 0,1,2\rbrace$ is a two dimensional vector given as
\begin{align*}
\mathbf{H}_{ij}(\theta)=\begin{bmatrix}
H_{ij1}(\theta) & H_{ij2}(\theta)
\end{bmatrix}^{T}.
\end{align*}
Now, on the center manifold $C_{0}$, near the origin
\begin{equation}
\dot{\mathbf{w}} = \mathbf{w}_{z}\dot{z} + \mathbf{w}_{\bar{z}}\dot{\bar{z}}.\label{eq:derivative1_w}
\end{equation}
We now use equations \eqref{eq:vector_w} and \eqref{eq:derivative_z} to replace $\mathbf{w}_{z}$ and $\dot{z}$ (and their conjugates) and equate this with \eqref{eq:derivative1_w} to get
\begin{align}
(2i\omega_{0}- \mathcal{A})\mathbf{w}_{20}(\theta) &= \mathbf{H}_{20}(\theta),\notag\\
-\mathcal{A}\mathbf{w}_{11}(\theta) &= \mathbf{H}_{11}(\theta),\notag\\
-(2i\omega_{0}+ \mathcal{A})\mathbf{w}_{02}(\theta) &= \mathbf{H}_{02}(\theta)\label{eq:H_operator},
\end{align}
as in \cite{Hassard}. Now, we observe that
\begin{align}
\mathbf{u}_{t}(\theta) =&\,\, \mathbf{w}(z,\bar{z},\theta)+ z\mathbf{q}(\theta) + \bar{z}\bar{\mathbf{q}}(\theta)\notag\\
=&\,\, \mathbf{w}_{20}(\theta)\frac{z^{2}}{2} + \mathbf{w}_{11}(\theta)z\bar{z} + \mathbf{w}_{02}(\theta)\frac{\bar{z}^{2}}{2}+ ze^{i\omega_{0}\theta} + \bar{z}e^{-i\omega_{0}\theta}+ \cdots, \label{eq:vector_u}
\end{align}
from which we obtain $\mathbf{u}_{t}(0)$, $\mathbf{u}_{t}(-\tau_1)$, and $\mathbf{u}_{t}(-\tau_2)$. We now proceed to expand the non-linear terms present in equation \eqref{eq:TaylorSeries} using equation \eqref{eq:vector_u} and retain only the coefficients of $z^2,z\bar{z},\bar{z}^2,z^2\bar{z}$. They are summarised as below:
\begin{align*}
& u_{1,t}^2(0) = z^2 +\bar{z}^2 + 2z\bar{z} +z^2\bar{z}\Big( w_{201}(0) + 2w_{111}(0)\Big) + \cdots,\notag\\
& u_{2,t}^2(0) = \phi_1^2 z^2 +\bar{\phi_1}^{2}\bar{z}^2 + 2\phi_1 \bar{\phi_1} z\bar{z} + z^2\bar{z}\Big( w_{202}(0)\bar{\phi_1} + 2w_{112}(0)\phi_1\Big) + \cdots,\notag\\
& u_{1,t}^2(-\tau_1 ) = z^2e^{-2i\omega_0 \tau_1} +\bar{z}^2 e^{2i\omega_0 \tau_1} + 2z\bar{z}
+z^2\bar{z}\Big( w_{201}(-\tau_1)e^{i\omega_0\tau_1} + 2w_{111}(-\tau_1)e^{-i\omega\tau_1} \Big) + \cdots,\notag\\
& u_{2,t}^2(-\tau_2 ) = \phi_1^2 z^2e^{-2i\omega_0 \tau_2} +\bar{\phi_1}^{2}\bar{z}^2 e^{2i\omega_0 \tau_2}+ 2\phi_1 \bar{\phi_1} z\bar{z}
+z^2\bar{z}\Big( w_{202}(-\tau_2)\bar{\phi_1}e^{i\omega_0\tau_2}
+ 2w_{112}(-\tau_2)\phi_1e^{-i\omega\tau_2} \Big) + \cdots,\notag\\
& u_{1,t}(0)u_{1,t}(-\tau_1 ) = z^2e^{-i\omega_0 \tau_1} +\bar{z}^2 e^{i\omega_0 \tau_1}
+ \Big(e^{i\omega_0 \tau_1}+e^{-i\omega_0 \tau_1}\Big)z\bar{z}
+ z^2\bar{z}\Big(\frac{w_{201}(0)}{2}e^{i\omega_0\tau_1}
+w_{111}(0)e^{-i\omega_0\tau_1}\notag\\
& \hspace*{25mm} +w_{111}(-\tau_1)+\frac{w_{201}(-\tau_1)}{2}\Big)\cdots,\notag\\
& u_{1,t}(0)u_{2,t}(-\tau_2 ) = \phi_1 z^2e^{-i\omega_0 \tau_2} +\bar{\phi_1}\bar{z}^2 e^{i\omega_0 \tau_2}
+ \Big(\bar{\phi_1}e^{i\omega_0 \tau_2}+\phi_1e^{-i\omega_0 \tau_2}\Big)z\bar{z}
+ z^2\bar{z}\Big(\bar{\phi_1}\frac{w_{201}(0)}{2}e^{i\omega_0\tau_2} \notag\\
& \hspace*{25mm} +\phi_1 w_{111}(0)e^{-i\omega_0\tau_2}
+w_{112}(-\tau_2)+\frac{w_{202}(-\tau_2)}{2}\Big)\cdots,\notag\\
& u_{1,t}(-\tau_1)u_{2,t}(-\tau_2 ) = \phi_1 z^2e^{-i\omega_0(\tau_1+ \tau_2)} +\bar{\phi_1}\bar{z}^2 e^{i\omega_0 (\tau_1+\tau_2)}
+ \Big(\bar{\phi_1}e^{i\omega_0 (\tau_2-\tau_1)}+\phi_1e^{-i\omega_0 (\tau_2-\tau_1)}\Big)z\bar{z} \notag\\
& \hspace*{29mm} + z^2\bar{z}\Big(\bar{\phi_1}\frac{w_{201}(-\tau_1)}{2}e^{i\omega_0\tau_2}
+\phi_1 w_{111}(-\tau_1)e^{-i\omega_0\tau_2}
+w_{112}(-\tau_2)e^{-i\omega_0\tau_1}
+\frac{w_{202}(-\tau_2)}{2}e^{i\omega_0\tau_1}\Big)\cdots,\notag\\
\end{align*}
\begin{align*}
& u_{2,t}(0)u_{2,t}(-\tau_2 ) = \phi_1^2 z^2e^{-i\omega_0 \tau_2} +\bar{\phi_1}^2\bar{z}^2 e^{i\omega_0 \tau_2}
+ \phi_1\bar{\phi_1}\Big(e^{i\omega_0 \tau_2}+e^{-i\omega_0 \tau_2}\Big)z\bar{z}
+ z^2\bar{z}\Big(\bar{\phi_1}\frac{w_{202}(0)}{2}e^{i\omega_0\tau_2} \notag \\
& \hspace*{24mm} +\phi_1 w_{112}(0)e^{-i\omega_0\tau_2}
+\phi_1 w_{112}(-\tau_2) + \bar{\phi_1}\frac{w_{202}(-\tau_2)}{2}\Big)\cdots,\notag\\
& u_{2,t}(0)u_{1,t}(-\tau_1 ) = \phi_1 z^2e^{-i\omega_0 \tau_1} +\bar{\phi_1}\bar{z}^2 e^{i\omega_0 \tau_1}
+ \Big(\phi_1 e^{i\omega_0 \tau_1}+\bar{\phi_1}e^{-i\omega_0 \tau_1}\Big)z\bar{z} + z^2\bar{z}\Big(\frac{w_{202}(0)}{2}e^{i\omega_0\tau_1} \notag\\
& \hspace*{24mm} + w_{112}(0)e^{-i\omega_0\tau_1}
+\phi_1 w_{111}(-\tau_1)+\bar{\phi_1}\frac{w_{201}(-\tau_1)}{2}\Big)\cdots,\notag\\
&u_{1,t}^3(0) = 3z^{2}\bar{z}+ \cdots,\notag\\
& u_{2,t}^3(0) = 3\phi_1^2\bar{\phi_1} z^{2}\bar{z}+ \cdots,\notag\\
& u_{1,t}^3(-\tau_1) = 3z^{2}\bar{z}e^{-i\omega \tau_1}+ \cdots,\notag\\
& u_{2,t}^3(-\tau_2) = 3\phi_1^2\bar{\phi_1}z^{2}\bar{z}e^{-i\omega \tau_2}+ \cdots,\notag\\
& u_{1,t}^2(0)u_{1,t}(-\tau_1) =z^2\bar{z} \Big(2e^{-i\omega_0\tau_1}+e^{i\omega_0 \tau_1}\Big)\cdots,\notag\\
& u_{1,t}^2(0)u_{2,t}(-\tau_2) =z^2\bar{z} \Big(2\phi_1 e^{-i\omega_0\tau_2}+\bar{\phi_1}e^{i\omega_0 \tau_2}\Big)\cdots,\notag\\
& u_{1,t}^2(-\tau_1)u_{1,t}(0) =z^2\bar{z} \Big(e^{-2i\omega_0\tau_1}+2\Big)\cdots,\notag\\
& u_{1,t}^2(-\tau_1)u_{2,t}(-\tau_2) =z^2\bar{z} \Big(\bar{\phi_1}e^{-i\omega_0(2\tau_1-\tau_2)} +2\phi_{1}e^{-i\omega_0\tau_2}\Big)\cdots,\notag\\
& u_{2,t}^2(-\tau_2)u_{1,t}(0) =z^2\bar{z} \phi_1 \Big(\phi_1 e^{-2i\omega_0\tau_1}+2\bar{\phi_1}\Big)\cdots,\notag\\
& u_{2,t}^2(-\tau_2)u_{1,t}(-\tau_1) =z^2\bar{z}\phi_1 \Big(\phi_1e^{-i\omega_0(2\tau_2-\tau_1)}+2\bar{\phi_1}e^{-i\omega_0\tau_1}\Big)\cdots,\notag\\
& u_{2,t}^2(0)u_{1,t}(-\tau_1) =z^2\bar{z} \phi_1\Big(2\bar{\phi_1}e^{-i\omega_0 \tau_1}+\phi_1e^{i\omega_0\tau_1}\Big)\cdots,\notag\\
& u_{2,t}^2(0)u_{2,t}(-\tau_2) =z^2\bar{z} \phi_1^2\bar{\phi_1}\Big(2e^{-i\omega_0 \tau_2}+e^{i\omega_0\tau_2}\Big)\cdots,\notag\\
& u_{2,t}^2(-\tau_2)u_{1,t}(0) =z^2\bar{z} \phi_1\Big(\phi_1e^{-2i\omega_0 \tau_2}+2\bar{\phi_1}\Big)\cdots,\notag\\
& u_{1,t}^2(-\tau_1)u_{2,t}(0) =z^2\bar{z} \Big(\bar{\phi_1}e^{-2i\omega_0 \tau_1}+2\phi_1\Big)\cdots,\notag\\
& u_{2,t}^2(-\tau_2)u_{2,t}(0) =z^2\bar{z}\phi_1^2\bar{\phi_1} \Big(2+e^{-2i\omega_0 \tau_2}\Big)\cdots,\notag\\
& u_{1,t}(0)u_{1,t}(-\tau_1)u_{2,t}(-\tau_2) =z^2\bar{z}\Big(\bar{\phi_1}e^{-i\omega_0 (\tau_1-\tau_2)} +\phi_1 e^{i\omega_0 (\tau_1-\tau_2)} +\phi_ 1e^{-i\omega_0 (\tau_1+\tau_2)}\Big)\cdots,\notag\\
& u_{2,t}(0)u_{1,t}(-\tau_1)u_{2,t}(-\tau_2) =z^2\bar{z}\phi_1^2\Big(e^{-i\omega_0 (\tau_1-\tau_2)} + e^{i\omega_0 (\tau_1-\tau_2)} +e^{-i\omega_0 (\tau_1+\tau_2)}\Big)\cdots.
\end{align*}
Using the definition $g(z,\bar{z})=\bar{\mathbf{p}}(0)\cdot \mathcal{F}_{0}(z,\bar{z})$ we then determine the coefficients of $z^2$, $z\bar{z}$, $\bar{z}^2$ and $z^2\bar{z}$, which are outlined below.
\begin{align*}
g_{20} &=2\kappa \bar{D} \Bigg( \bar{\phi}_{2}\Big(\xi_{aa}+\xi_{bb}e^{-2i\omega_0 \tau_1}+\xi_{dd}\phi_1^2 e^{-2i\omega_0 \tau_2}+ \xi_{ab}e^{-i\omega_0 \tau_1}+\xi_{ad}\phi_1 e^{-i\omega_0 \tau_2}+ \xi_{bd}\phi_1 e^{-i\omega_0(\tau_1+ \tau_2)}\Big)\notag\\
&+ \chi_{cc}\phi_1^2 +\chi_{dd}\phi_1^2 e^{-2i\omega_0 \tau_2}+\chi_{bb}e^{-2i\omega_0 \tau_1}+ \chi_{cd}\phi_1^2 e^{-i\omega_0 \tau_2} +\chi_{bc}\phi_1 e^{-i\omega_0 \tau_1}+ \chi_{bd}\phi_1 e^{-i\omega_0(\tau_1+ \tau_2)}\Bigg),\notag\\
\end{align*}
\begin{align}
g_{02} &=2\kappa \bar{D} \Bigg( \bar{\phi}_{2}\Big(\xi_{aa}+\xi_{bb}e^{2i\omega_0 \tau_1}+\xi_{dd}\phi_1^2 e^{2i\omega_0 \tau_2}+ \xi_{ab}e^{i\omega_0 \tau_1} +\xi_{ad}\bar{\phi_1} e^{i\omega_0 \tau_2}+ \xi_{bd}\bar{\phi_1} e^{i\omega_0(\tau_1+ \tau_2)}\Big)\notag\\
&+ \chi_{cc}\bar{\phi_1}^2 +\chi_{dd}\bar{\phi_1}^2 e^{2i\omega_0 \tau_2}+\chi_{bb}e^{2i\omega_0 \tau_1}+ \chi_{cd}\bar{\phi_1}^2 e^{i\omega_0 \tau_2} +\chi_{bc}\bar{\phi_1} e^{i\omega_0 \tau_1}+ \chi_{bd}\bar{\phi_1} e^{i\omega_0(\tau_1+ \tau_2)}\Bigg),\notag\\
g_{11} &=\kappa \bar{D} \Bigg( \bar{\phi}_{2}\Big(2\xi_{aa}+2\xi_{bb}+\xi_{dd}\phi_1\bar{\phi_1}+ \xi_{ab}\left(e^{i\omega_0 \tau_1}+e^{-i\omega_0 \tau_1}\right)+\xi_{ad}\left(\bar{\phi_1}e^{i\omega_0 \tau_2}+\phi_1e^{-i\omega_0 \tau_2}\right)\notag\\
&+ \xi_{bd}\left(\bar{\phi_1}e^{i\omega_0 (\tau_2-\tau_1)}+\phi_1e^{-i\omega_0 (\tau_2-\tau_1)}\right)\Big)+2\phi_1\bar{\phi_1}(\chi_{cc} +\chi_{dd}) +2\chi_{bb}+\chi_{cd}\phi_1\bar{\phi_1}\left(e^{i\omega_0 \tau_2}+e^{-i\omega_0 \tau_2}\right)\notag\\ &+\chi_{bc}\left(\phi_1 e^{i\omega_0 \tau_1}+\bar{\phi_1}e^{-i\omega_0 \tau_1}\right)+ \chi_{bd}\left(\bar{\phi_1}e^{i\omega_0 (\tau_2-\tau_1)}+\phi_1e^{-i\omega_0 (\tau_2-\tau_1)}\right)\Bigg),\notag\\
g_{21}&=2\kappa \bar{D}\Bigg(\bar{\phi_2} \Big(\xi_{aa}\left(w_{201}(0) + 2w_{111}(0)\right)+ \xi_{bb}\left(w_{201}(-\tau_1)e^{i\omega_0\tau_1}+2w_{111}(-\tau_1)e^{-i\omega\tau_1}\right)\notag\\
&+ \xi_{dd}\left(w_{202}(-\tau_2)\bar{\phi_1}e^{i\omega_0\tau_2} +2w_{112}\phi_1(-\tau_2)e^{-i\omega\tau_2}\right)+ \xi_{ab}\big(\frac{w_{201}(0)}{2}e^{i\omega_0\tau_1}+w_{111}(0)e^{-i\omega_0\tau_1}+ w_{111}(-\tau_1)\notag\\
&+\frac{w_{201}(-\tau_1)}{2}\big)+ \xi_{ab}\big(\frac{w_{201}(0)}{2}\bar{\phi_1}e^{i\omega_0\tau_2}+w_{111}(0)\phi_1e^{-i\omega_0\tau_2}+ w_{112}(-\tau_2)+\frac{w_{202}(-\tau_2)}{2}\big)\notag\\
& + \xi_{bd}\big(\frac{w_{201}(-\tau_1)}{2}\bar{\phi_1}e^{i\omega_0\tau_2}+ w_{111}(-\tau_1)\phi_1e^{-i\omega_0\tau_2}+ w_{112}(-\tau_2)e^{-i\omega_0\tau_1}+\frac{w_{202}(-\tau_2)}{2}e^{i\omega_0\tau_1}\big)\notag\\
&+3\xi_{aaa}+3\xi_{bbb}e^{i\omega_0\tau_1}
+3\xi_{ddd}\phi_1^2\bar{\phi_1}e^{-\i\omega_0\tau_2}+\xi_{aab} \big(2e^{-i\omega_0\tau_1}+e^{i\omega_0 \tau_1}\big)+\xi_{aad}\big(2\phi_1 e^{-i\omega_0\tau_2}+\bar{\phi_1}e^{i\omega_0 \tau_2}\big)\notag\\
&+\xi_{abb} \big(e^{-2i\omega_0\tau_1}+2\big)+\xi_{bbd}\big(\bar{\phi_1}e^{-i\omega_0(2\tau_1-\tau_2)}+2\phi_{1}e^{-i\omega_0\tau_2}\big)+ \xi_{add}\phi_1 \big(\phi_1e^{-2i\omega_0\tau_2}+2\bar{\phi_1}\big)\notag\\
&+\xi_{bdd}\phi_1 \big(\phi_1e^{-i\omega_0(2\tau_2-\tau_1)}+\bar{\phi_1}e^{-i\omega_0\tau_1}\big)+\xi_{abd}\big(\bar{\phi_1}e^{-i\omega_0 (\tau_1-\tau_2)}+\phi_1 e^{i\omega_0 (\tau_1-\tau_2)}+\phi_ 1e^{-i\omega_0 (\tau_1+\tau_2)}\big)\Big)\notag\\
&+ \chi_{cc}\left(w_{202}(0)\bar{\phi_1} + 2w_{112}(0)\phi_1\right)+\chi_{dd}\left(w_{202}(-\tau_2)\bar{\phi_1}e^{i\omega_0\tau_2}+2w_{112}(-\tau_2)\phi_1e^{-i\omega\tau_2}\right)\notag\\
&+ \chi_{bb}\left(w_{201}(-\tau_1)e^{i\omega_0\tau_1}+2w_{111}(-\tau_1)e^{-i\omega\tau_1}\right)+ \chi_{cd}\big(\frac{w_{202}(0)}{2}\bar{\phi_1}e^{i\omega_0\tau_2}+\phi_1 w_{112}(0)e^{-i\omega_0\tau_2}\notag\\
&+\phi_1 w_{112}(-\tau_2)+\bar{\phi_1}\frac{w_{202}(-\tau_2)}{2}\big)+\chi_{bc}\big(\frac{w_{202}(0)}{2}e^{i\omega_0\tau_1}+ w_{112}(0)e^{-i\omega_0\tau_1}+\phi_1 w_{111}(-\tau_1)+\bar{\phi_1}\frac{w_{201}(-\tau_1)}{2}\big)\notag\\
&+\chi_{bd}\big(\bar{\phi_1}\frac{w_{201}(-\tau_1)}{2}e^{i\omega_0\tau_2}+\phi_1 w_{111}(-\tau_1)e^{-i\omega_0\tau_2}+w_{112}(-\tau_2)e^{-i\omega_0\tau_1}+\frac{w_{202}(-\tau_2)}{2}e^{i\omega_0\tau_1}\big)+3\chi_{ccc}\phi_1^2\bar{\phi_1}\notag\\ &+3\chi_{ddd}\phi_1^2\bar{\phi_1}e^{-i\omega_0\tau_2}+3\chi_{bbb}e^{-i\omega_0\tau_1}+ \chi_{ccd}\phi_1^2\bar{\phi_1}\big(2e^{-i\omega_0 \tau_2}+e^{i\omega_0\tau_2}\big)+\chi_{bcc}\phi_1 \big(2\bar{\phi_1}e^{-i\omega_0-\tau_1}+\phi_1e^{i\omega_0\tau_1}\big)\notag\\
&+\chi_{cdd} \phi_1^2\bar{\phi_1}\big(e^{-2i\omega_0 \tau_2}+2\big)+\chi_{bdd}\phi_1 \big(\phi_1e^{-i\omega_0(2\tau_2-\tau_1)}+2\bar{\phi_1}e^{-i\omega_0\tau_1}\big)+ \chi_{bbc}\big(\bar{\phi_1}e^{-2i\omega_0 \tau_2}+2\phi_1\big)\notag\\
&+\chi_{bbd}\big(\bar{\phi_1}e^{-i\omega_0(2\tau_1-\tau_2)}+2\phi_{1}e^{-i\omega_0\tau_2}\big)+\chi_{bcd}\phi_1\big(\bar{\phi_1}e^{-i\omega_0 (\tau_1-\tau_2)}+\phi_1 e^{i\omega_0 (\tau_1-\tau_2)}+\bar{\phi_1}e^{-i\omega_0 (\tau_1+\tau_2)}\big)\Bigg).
\label{eq:g21}
\end{align}
Note that, the expression for $g_{21}$ has $\mathbf{w}_{20}(\theta)$ and $\mathbf{w}_{11}(\theta)$ which we need to evaluate. Now, for $\theta \in [-\tau,0)$ from \eqref{eq:vector_H}, we have
\begin{align*}
\mathbf{H}(z,\bar{z},\theta) =& -2\text{Re}\big(\mathbf{\bar{q}}^{\ast}(0)\cdot \mathcal{F}_{0} \mathbf{q}(\theta)\big)\\
=& -g(z,\bar{z})\mathbf{q}(\theta) - \bar{g}(z,\bar{z})\mathbf{\bar{q}}(\theta)\\
=& -\left(g_{20}\frac{z^{2}}{2} + g_{11}z\bar{z} + g_{02}\frac{\bar{z}^{2}}{2}+\cdots \right)\mathbf{q}(\theta) - \left(\bar{g}_{20}\frac{\bar{z}^{2}}{2} + \bar{g}_{11}z\bar{z} + \bar{g}_{02}\frac{z^{2}}{2}+\cdots \right)\mathbf{\bar{q}}(\theta),
\end{align*}
which when compared with \eqref{eq:vector_H} gives
\begin{align}
\label{eq:vector1_H}
\mathbf{H}_{20}(\theta) &= -g_{20}\mathbf{q}(\theta)-\bar{g}_{02}\mathbf{\bar{q}}(\theta),\notag\\
\mathbf{H}_{11}(\theta) &= -g_{11}\mathbf{q}(\theta)-\bar{g}_{11}\mathbf{\bar{q}}(\theta).
\end{align}
Using equations \eqref{eq:DLEoperatorR} and \eqref{eq:H_operator}, we have
\begin{align}\label{eq:w_derivatives}
\dot{\mathbf{w}}_{20}(\theta) &= 2i\omega_{0}\mathbf{w}_{20}(\theta) + g_{20}\mathbf{q}(\theta)+ \bar{g}_{02}\mathbf{\bar{q}}(\theta),\notag\\
\dot{\mathbf{w}}_{11}(\theta) &= g_{11}\mathbf{q}(\theta) + \bar{g}_{11}\mathbf{\bar{q}}(\theta).
\end{align}
Solving the differential equations in \eqref{eq:w_derivatives}, we get
\begin{align}\label{eq:w_vectorssol}
\mathbf{w}_{20}(\theta) &= -\frac{g_{20}}{i\omega_{0}}\boldsymbol{q}(0)e^{i\omega_{0}\theta}-\frac{\bar{g}_{02}}{3i\omega_{0}}\mathbf{\bar{q}}(0)e^{-i\omega_{0}\theta} + \mathbf{e}e^{2i\omega_{0}\theta},\notag\\
\mathbf{w}_{11}(\theta) &= \frac{g_{11}}{i\omega_{0}}\mathbf{q}(0)e^{i\omega_{0}\theta} - \frac{\bar{g}_{11}}{i\omega_{0}}\mathbf{\bar{q}}(0)e^{-i\omega_{0}\theta} + \boldsymbol{f}.
\end{align}
The objective now is to determine $ \boldsymbol{e}$ and $ \boldsymbol{f}$. We define,
\begin{align}
\mathbf{H}(z,\bar{z},0) &= -2\text{Re}\big( \mathbf{\bar{q}}^{\ast}(0)\cdot\mathcal{F}_{0} \mathbf{q}(0)\big)+\mathcal{F}_0,\label{eq:H_0}
\end{align}
where $\boldsymbol{\mathcal{{F}}}_0$ represents the non-linear terms that can be expanded in powers of $z$ as
\begin{equation}
\boldsymbol{\mathcal{F}}=\boldsymbol{\mathcal{F}}_{20}\frac{z^2}{2}+ \boldsymbol{\mathcal{F}}_{11}z\bar{z}+\boldsymbol{\mathcal{F}}_{02}\frac{\bar{z}^2}{2} + \boldsymbol{\mathcal{F}}_{21}\frac{z^2 \bar{z}}{2} + \cdots.
\end{equation}
Substituting the coefficients from the expansion of $\boldsymbol{\mathcal{F}}_0$ gives
\begin{align}
\label{eq:H_coeff}
\mathbf{H}_{20}(0) &= -g_{20}\mathbf{q}(0)-\bar{g}_{02}\mathbf{\bar{q}}(0)+\begin{bmatrix}\mathcal{F}_{201}&\mathcal{F}_{202}\end{bmatrix}^{T},\notag\\
\mathbf{H}_{11}(0) &= -g_{11}\mathbf{q}(0)-\bar{g}_{11}\mathbf{\bar{q}}(0)+\begin{bmatrix}\mathcal{F}_{111}&\mathcal{F}_{112}\end{bmatrix}^{T}.
\end{align}
From \eqref{eq:H_coeff} and \eqref{eq:DLEoperatorR}, we obtain
\begin{align}
&g_{20}\mathbf{q}(0)+\bar{g}_{02}\mathbf{\bar{q}}(0) = \begin{bmatrix}\mathcal{F}_{201}&\mathcal{F}_{202}\end{bmatrix}^{T}+ \begin{bmatrix} (\kappa a_{11}-2i\omega_0)\mathsf{w}_{201}(0) + \kappa a_{12}\mathsf{w}_{201}(-\tau_1)+\kappa a_{13}\mathsf{w}_{202(-\tau_1)}\notag\\
\kappa a_{23}\mathsf{w}_{201}(-\tau)+\kappa(a_{21}- 2i\omega_0)\mathsf{w}_{202}(0)+\kappa a_{22}\mathsf{w}_{202}(-\tau_2)\end{bmatrix},\notag\\
&g_{11}\mathbf{q}(0)+\bar{g}_{11}\mathbf{\bar{q}}(0) = \begin{bmatrix}\mathcal{F}_{111}&\mathcal{F}_{112}\end{bmatrix}^{T}+ \begin{bmatrix}\kappa a_{11}\mathsf{w}_{111}(0)-\kappa a_{12}\mathsf{w}_{111}(-\tau_1)+\kappa a_{13}\mathsf{w}_{112}(-\tau_2)\\ \kappa a_{23}\mathsf{w}_{111}(-\tau_1)+\kappa a_{21} \mathsf{w}_{112}(0)+\kappa a_{22}\mathsf{w}_{112}(-\tau_2)\end{bmatrix}.\label{eq:getE&F}
\end{align}
We substitute $\mathbf{w}_{20}(0),\mathbf{w}_{20}(-\tau),\mathbf{w}_{11}(0)$ and $\mathbf{w}_{11}(-\tau)$ from \eqref{eq:w_vectorssol} in \eqref{eq:getE&F} we get $\mathbf{e}$ and $\mathbf{f}$ of the form
\begin{align}
\boldsymbol{e} = \begin{bmatrix}e_{1}& e_{2}\end{bmatrix}^{T} \quad\text{and}\quad \boldsymbol{f} = \begin{bmatrix}f_{1}& f_{2}\end{bmatrix}^{T}.
\end{align}
Note that, $e_1,e_2,f_1$ and $f_2$ can be derived explicitly in terms of system parameters, which are outlined below:
\begin{align}
e_1 = \frac{Y_2Z_1 - Y_1Z_2}{X_1Y_2-X_2Y_1}, \hspace{3mm} e_2 =\frac{X_1Z_2 - X_2Z_1}{X_1Y_2-X_2Y_1},\hspace{3mm} f_1 = \frac{Q_2R_1 - Q_1R_2}{P_1Q_2-P_2Q_1}, \hspace{3mm} f_2 = \frac{P_1R_2-P_2R_1}{P_1Q_2-P_2Q_1},\notag\\
\end{align}
where,
\begin{align}
X_1 &= \,\, \kappa a_{11}+\kappa a_{12}e^{-2i\omega_0\tau_1}-2i\omega_0,\hspace{5mm} X_2 = \kappa a_{23}e^{-2i\omega _{0}\tau_1},\notag\\
Y_1 &= \,\,\kappa a_{13}e^{-2i\omega_0\tau_2}, \hspace{4mm} Y_2 = \kappa a_{21}+\kappa a_{22}e^{-2i\omega_0\tau_2}-2i\omega_0,\notag\\
P_1 &= \,\, \kappa a_{11}+\kappa a_{12}, \hspace{5mm} P_2 = \kappa a_{23},\hspace{5mm}Q_1 = \,\, \kappa a_{13}, \hspace{5mm} Q_2 = \kappa a_{21}+\kappa a_{22},\notag\\
Z_1 &= \,\, \frac{g_{20}}{i\omega_0}\left(-i\omega_0 + \kappa a_{11}+ \kappa a_{12}e^{-i\omega_0\tau_1}+\kappa a_{13}\phi_1 e^{-i\omega_0\tau_1} \right)+\frac{\bar{g}_{02}}{3i\omega_0}\left(i\omega_0 + \kappa a_{11}+ \kappa a_{12}e^{i\omega_0\tau_1}+\kappa a_{13}\bar{\phi}_1 e^{i\omega_0\tau_1}\right)\notag\\
&- \mathcal{F}_{201},\notag\\
Z_2 &= \,\, \frac{g_{20}}{i\omega_0}\big(-i\omega_0 \phi_1 + \kappa a_{23}e^{i \omega_{0}\tau_1}+\kappa a_{21}\phi_1+ \kappa a_{22}\phi_1 e^{i\omega_0\tau_2}\big)+\frac{\bar{g}_{02}}{3i\omega_0}\big(i\omega_0\bar{\phi}_1 + \kappa a_{23}e^{i \omega_{0}\tau_1}+\kappa a_{21}\bar{\phi}_1\notag\\
&+ \kappa a_{22}\bar{\phi}_1 e^{i\omega_0\tau_2}\big)-\mathcal{F}_{202},\notag
\end{align}
\begin{align}
R_1 &= \,\, \frac{g_{11}}{i\omega_0}\left(i\omega_0 -\kappa a_{11}- \kappa a_{12}e^{-i\omega_0\tau_1}-\kappa a_{13}\phi_1 e^{-i\omega_0\tau_1}\right)+ \frac{\bar{g}_{11}}{i\omega_0}\left(i\omega_0 + \kappa a_{11}+ \kappa a_{12}e^{i\omega_0\tau_1}+\kappa a_{13}\bar{\phi}_1 e^{i\omega_0\tau_1}\right)\notag\\
&-\mathcal{F}_{111},\notag\\
R_2 &= \,\, \frac{g_{11}}{i\omega_0}\big(i\omega_0 \phi_1 \kappa a_{23}e^{i \omega_{0}\tau_1}-\kappa a_{21}\phi_1- \kappa a_{22}\phi_1 e^{i\omega_0\tau_2}\big)+\frac{\bar{g}_{02}}{3i\omega_0}\big(i\omega_0\bar{\phi}_1 + \kappa a_{23}e^{i \omega_{0}\tau_1}+\kappa a_{21}\bar{\phi}_1+ \kappa a_{22}\bar{\phi}_1 e^{i\omega_0\tau_2}\big)\notag\\
&-\mathcal{F}_{112}.
\end{align}
Using $\boldsymbol{e}$ and $\boldsymbol{f}$ we evaluate $\mathbf{w}_{20}$ and $\mathbf{w}_{11}$, using which we compute $g_{21}$. We now have all the terms required for the analysis of Hopf bifurcation as follows, see \cite{Hassard}
\begin{align}
\hspace{-3mm}c_1(0) &= \frac{i}{2\omega_0}\left(g_{20}g_{11}-2|g_{11}|^2-\frac{1}{3}|g_{02}|^2\right)+\frac{g_{21}}{2},\label{eq:cterm}\\
\mu_2 &= -\frac{\text{Re}\big(c_1(0)\big)}{\alpha'(0)},\quad\quad\beta_2 = 2\text{Re}\big(c_1(0)\big),\label{eq:muterm&betaterm}
\end{align}
where $c_1(0)$ is the lyapunov coefficient and $\alpha'(0) =~\text{Re}\left(\mathrm{d}\lambda/\mathrm{d}\kappa\right)|_{\kappa=\kappa_c}$.
The following conditions enable us to verify the type of the Hopf bifurcation, and the asymptotic orbital stability of the limit cycles~\cite{Hassard}.
\begin{itemize}
\item The Hopf bifurcation is \emph{supercritical} if $\mu_2 > 0$ and \emph{sub-critical} if $\mu_2 <0$.
\item The limit cycles are \emph{asymptotically orbitally stable} if $\beta_2< 0$ and \emph{unstable} if $\beta_2>0$.
\end{itemize}
Substituting the expression for $g_{21}$ in \eqref{eq:cterm} yields the expression for $c_1(0)$, which is the lyapunov coefficient. We can then compute $\mu_2$ and $\beta_2$ using \eqref{eq:muterm&betaterm}. We now present a numerical example, and compute the values of $\mu_2$ and $\beta_2$ for Compound TCP in the small buffer regime.
\begin{figure*}
\begin{center}
\begin{minipage}[b]{0.45\textwidth}
\label{fig:phase_0.95}
\psfrag{b}{$w_{2}(t)$}
\psfrag{T}{$w_{2}(t-\tau_2)$}
\psfrag{124}{\begin{scriptsize}$124$\end{scriptsize}}
\psfrag{130}{\begin{scriptsize}$130$\end{scriptsize}}
\psfrag{136}{\begin{scriptsize}$136$\end{scriptsize}}
\includegraphics[width=2in,height=3in,angle=270]{converge.eps}
\end{minipage}
\hspace*{5mm}
\begin{minipage}[b]{0.45\textwidth}
\label{fig:phase_1.05}
\psfrag{b}{$w_2(t)$}
\psfrag{T}{$w_2(t-\tau_2)$}
\psfrag{122}{\begin{scriptsize}$122$\end{scriptsize}}
\psfrag{129}{\begin{scriptsize}$129$\end{scriptsize}}
\psfrag{136}{\begin{scriptsize}$136$\end{scriptsize}}
\includegraphics[width=2in,height=3in,angle=270]{lc.eps}
\end{minipage}
\caption {\emph{Phase portraits.} Emergence of limit cycle in the dynamics of $w_2(t)$ in \eqref{eq:modelb_3}, for Compound TCP in the small buffer regime, with the variation in the non-dimensional parameter $\kappa$. Observe that, (a) Trajectories converge to stable equilibrium for $\kappa=0.95$, (b) Trajectories converge to a stable limit cycle for $\kappa=1.05$.}
\label{fig:phase}
\end{center}
\end{figure*}
\subsection*{Numerical Example:}We first fix the system parameters as follows: $\alpha=0.3$, $k=0.75$, $\beta=0.5$, $B_1=10$, $B_2=15$, $B=25$, $C_1=C_2=100$, $C=180$, $\tau_1=1$, and $\tau_2$. With these parameter values, the system undergoes a Hopf bifurcation at $\kappa_c=1$. We now increase the value of the non-dimensional parameter to $\kappa=1.05$, and push the system beyond the edge of stability. Following the Hopf bifurcation analysis presented above, we compute the required expressions:
\begin{align*}
\mathrm{Re}\left(c_{1}(0)\right) &= -0.0738<0, \hspace{1ex} \alpha'(0)=0.3467>0\notag\\
\mu_2 &=0.2129>0, \hspace{5.5ex}\beta_{2}=-0.1477<0.
\end{align*}
Thus, the Hopf bifurcation is \emph{supercritical} and the emergent limit cycles are asymptotically \emph{orbitally stable}.
\subsection*{Phase portraits and bifurcation diagram:} We present the phase portrait for system \eqref{eq:modelb_3}, for Compound TCP in the small buffer regime, in Fig.~\ref{fig:phase}. First, we fix a point $\alpha=0.3,$ $\kappa=1$, on the stability boundary in the stability chart as shown in Fig.~\ref{fig:charts}~(a). The remaining system parameter values are fixed as mentioned above in the numerical example. We now plot the phase portrait for the window size for the second set of TCP flows, for $\kappa=0.95$ and $\kappa=1.05$ respectively, as shown in Fig.~\ref{fig:phase}. Observe that, for $\kappa=0.95$, the average window size of the second set Compound TCP flows converges to its equilibrium value, as expected. For, $\kappa=1.05$, the average window size exhibits orbitally stable limit cycles, as the system undergoes a Hopf bifurcation at $\kappa=1$. Note that, the average window size of the first set of Compound flows can be shown to exhibit qualitatively similar dynamical behaviour. We now present the bifurcation diagram for system \eqref{eq:modelb_3}, in Fig.~\ref{fig:bifurcation}, obtained from DDE-BIFTOOL version $2.03$. Observe that, the amplitude of the limit cycles increases as $\kappa$ is increased beyond $1$.
\begin{figure}[t]
\begin{center}
\psfrag{T}[0,0.5][0.25,0.75]{Amplitude}
\psfrag{b}{\hspace{-21.5mm}Non-dimensional parameter, $\kappa$}
\psfrag{0.9}{\begin{scriptsize}$0.9$\end{scriptsize}}
\psfrag{1.1}{\begin{scriptsize}$1.1$\end{scriptsize}}
\psfrag{1}{\begin{scriptsize}$1$\end{scriptsize}}
\psfrag{120}{\begin{scriptsize}\hspace{1mm}$120$\end{scriptsize}}
\psfrag{137}{\begin{scriptsize}$137$\end{scriptsize}}
\psfrag{SSSSSSSSSSSSSS}{\hspace{2mm}\begin{scriptsize}Stable Region\end{scriptsize}}
\psfrag{h}{\begin{scriptsize}Hopf Condition\end{scriptsize}}
\includegraphics[width=1.8in,height=2.6in,angle=270]{bif_diag.eps}
\caption{\emph{Bifurcation diagram.} Emergence of limit cycles in the dynamics of $w_2(t)$ at $\kappa=1$ for system \eqref{eq:modelb_3}, with Compound TCP flows in the small buffer regime. The amplitude of the emergent limit cycles increases for further increase in $\kappa$. }\label{fig:bifurcation}\vspace{-8mm}
\end{center}
\end{figure}
\section{Packet-level simulations}
\label{simulations}
In order to corroborate the analytical insights obtained, we conduct some packet-level simulations, for the multiple bottleneck scenario, in NS2 \cite{ns2}.
\begin{figure}
\begin{center}
\psfrag{0}{\begin{scriptsize}$0$\end{scriptsize}}
\psfrag{15}{\begin{scriptsize}$15$\end{scriptsize}}
\psfrag{100}{\begin{scriptsize}$100$\end{scriptsize}}
\psfrag{125}{\begin{scriptsize}$125$\end{scriptsize}}
\psfrag{aaaa}{\begin{scriptsize}Buffer size = 15 pkts\end{scriptsize}}
\psfrag{bbbb}{\begin{scriptsize}Buffer size = 100 pkts\end{scriptsize}}
\psfrag{cccc}{Queue size (pkts)}
\psfrag{xyz}{\hspace{-6mm}Time (seconds)}
\psfrag{mmmm}{\begin{scriptsize}$\tau_1=10$ ms, $\tau_2= 10$ ms\end{scriptsize}}
\psfrag{nnnn}{\begin{scriptsize}\hspace{-1mm}$\tau_1=10$ ms, $\tau_2= 200$ ms\end{scriptsize}}
\includegraphics[width=2.5in,height=3.5in,angle=270]{queuelimitcycles.eps}
\caption{ \emph{Long-lived flows}. Two sets of 60 long-lived Compound flows over a 2 Mbps link, regulated by two edge routers, feeding into a core router with link capacity 180 Mbps. Observe the emergence of limit cycles in the queue at the core router, for larger buffer thresholds, and larger round trip times.}\vspace{-8mm}
\label{ns2}
\end{center}
\end{figure}
The system consists of two distinct sets of $60$ long-lived Compound TCP flows each with an access speed of $2$ Mbps, regulated by two edge routers and feeding into one core router. Each edge router has a link capacity of $100$ Mbps, and the core router has a link capacity of $180$ Mbps. Since our primary focus is on small buffers, we fix the buffer size for each edge router to be $15$ packets, and vary the buffer size of the core router from $15$ packets to $100$. Further, we fix the round trip time of one set of flows to be $10$ ms, and the round trip time of the other set is varied from $10$ ms to $200$ ms. The simulations are illustrated in Fig.~\ref{ns2}. Observe that, if the buffer sizes at all routers are fixed at $15$ packets, the queue at the core router is completely random, and hence stable, since the queue does not exhibit any deterministic oscillations. When the buffer size at the core router is increased to $100$ packets and the round trip time of the second set of flows is $200$ ms, the queue dynamics exhibits limit cycles. Hence, larger queue thresholds are prone to inducing limit cycles, for larger round trip times. These limit cycles in the queue size lead to synchronisation among TCP flows and make the downstream traffic bursty.
\section{Concluding remarks}
We considered three different topologies, and conducted a detailed local stability analysis with two simplifying assumptions, to obtain necessary and sufficient conditions for stability. To aid our analysis, we motivated a suitable non-dimensional bifurcation parameter, and illustrated that, the underlying dynamical systems lose stability if the bifurcation parameter is varied. Further, in the multiple bottleneck scenario, even without any simplifying assumptions on the system parameters, we numerically identified that the system loses stability via a Hopf bifurcation. A key insight obtained was the trade-off between different system parameters to ensure stability, as illustrated through some stability charts. After knowing that a system exhibits a Hopf, it is natural to have a framework to determine the asymptotic orbital stability of the bifurcating limit cycles. To that end, using Poincar\'{e} normal forms and the center manifold theory, we conducted a detailed Hopf bifurcation analysis, in the neighbourhood of the Hopf condition. To corroborate our analytical insights, we conducted some packet-level simulations to highlight the existence and stability of limit cycles in the queue size dynamics as system parameters vary.
The insights obtained in this paper could have important consequences for the modelling and the performance evaluation of communication networks. From a theoretical perspective, this opens many challenging questions centred around the development of accurate fluid models for TCP and queue management policies. From a practical perspective, the emergence of stable limit cycles could have an impact on the end-to-end quality of service -- these issues merit further investigation.
\label{conclusions}
|
\section{Introduction}\label{introduction.sec}
The distribution of ionized gas and the strength and structure of the magnetic field in the GC\footnote{
Throughout this paper we will refer to the inner $\sim\,150\,$pc region centred on \sgr\ as `the GC', which corresponds to the region with strongly enhanced electron densities in NE2001 \citep{cordes2002}.
} have been probed using diffuse synchrotron emission (e.g., \citealt{yusef-zadeh1984}, \citealt{tsuboi1986}, and \citealt{sofue1987}) and extragalactic radio sources \citep{roy2008}. These early measurements found evidence for strong magnetic fields, dense gas, and a new phenomenon: non-thermal radio filaments (NTFs), which are thin strands or sheets of synchrotron-emitting plasma. Over the last few decades many new insights have shed light on this complex environment, but important questions regarding the strength and structure of the magnetic field that pervades the GC remain unanswered (for a critical review see \citealt{ferriere2011}).
We study the properties of the magnetic field in the GC by observing all known radio pulsars within $\approx$ 20\arcmin\ of \sgr.
First we establish that these pulsars lie in the GC and are not merely seen towards it, then we investigate the implications for scattering in the GC.
Radio pulsars have several advantages over source types that have been used previously: pulsars are perfect point sources, pulsar RMs are produced exclusively in the Galactic foreground \citep{noutsos2009}, pulsar distances can be estimated from the delay in pulse arrival times with frequency (quantified by the dispersion measure DM), and the line-of-sight (LOS) magnetic field component in the interstellar medium (ISM) can be estimated from the amount of Faraday rotation of the pulsar signal (quantified by RM).
At a wavelength $\lambda$ the amount of Faraday rotation is equal to $\chi - \chi_0\, =\, \mathrm{RM}\,\lambda^2$, where $\chi$ and $\chi_0$ are the observed and the intrinsic polarization angle of the emission, respectively.
\begin{eqnarray}
\mathrm{RM}\, \left(\mathrm{rad~m}^{-2}\right)\ =\ 0.81 \int_\mathrm{source}^\mathrm{observer} n_\mathrm{e} B_\| \mathrm{d}l\, ,
\label{rm_definition}
\end{eqnarray}
where $n_\mathrm{e}$ is the local electron density in units of cm$^{-3}$, $B_\|$ the LOS component of the magnetic field in units of $\mu\mathrm{G}$, and the path length $l$ is measured in parsec.
If the LOS component of the magnetic field points towards us $B_\|$ is positive, RM increases, and the plane of polarization is rotated counterclockwise.
The dispersion measure DM is defined as
\begin{eqnarray}
\mathrm{DM}\, \left(\mathrm{cm}^{-3}~\mathrm{pc}\right)\ =\ \int_\mathrm{source}^\mathrm{observer} n_\mathrm{e} \mathrm{d}l\, ,
\label{dm_definition}
\end{eqnarray}
and the emission measure $\mathrm{EM}\, \left(\mathrm{cm}^{-6}~\mathrm{pc}\right)\ =\ \int_\mathrm{source}^\mathrm{observer} n_\mathrm{e}^2 \mathrm{d}l$.
The pulsars in our sample have been discovered by \cite{johnston2006} and \cite{deneva2009} and include also the recently discovered magnetar PSR J1745-2900 which is known to lie close to \sgr (\citealt{kennea2013}, \citealt{mori2013}, \citealt{eatough2013}, \citealt{shannon2013}).
This paper is structured as follows: in section~\ref{observations.sec} we describe the observations and their calibration, and in section~\ref{discussion.sec} we interpret our results in terms of the structure of the magnetized ISM in the GC. In section~\ref{summary.sec} we summarise our results.
We will adopt a distance of 8.3 kpc to the GC \citep{reid2014}; at this distance 1 arcminute on the sky corresponds to a distance of 2.4 parsec.
\begin{figure*}
\caption{
Pulse profiles of the target pulsars, showing Stokes $I$ (black), $Q$ (red) and $U$ (green) after correcting for Faraday rotation using the RMs from table~\ref{pulsars.tab}, and $V$ (blue). The Stokes $I$ profile also shows 1-sigma errors.
}
\begin{minipage}[b]{0.42\linewidth}
\resizebox{\hsize}{!}{\includegraphics[width=\linewidth]{j1746-2849.profile.eps}}
\vspace{-3ex}
\end{minipage}
\begin{minipage}[b]{0.42\linewidth}
\resizebox{\hsize}{!}{\includegraphics[width=\linewidth]{j1745-2900.profile.eps}}
\vspace{-3ex}
\end{minipage}
\begin{minipage}[b]{0.42\linewidth}
\resizebox{\hsize}{!}{\includegraphics[width=\linewidth]{j1746-2856.profile.eps}}
\vspace{-3ex}
\end{minipage}
\begin{minipage}[b]{0.42\linewidth}
\resizebox{\hsize}{!}{\includegraphics[width=\linewidth]{j1745-2912.profile.eps}}
\vspace{-3ex}
\end{minipage}
\label{profiles.fig}
\end{figure*}
\begin{figure*}
\caption{Frequency spectra of the target pulsars, showing Stokes $Q$ (top panel) and $U$ (bottom panel) together with the best-fitting models and 68\% confidence intervals.
}
\begin{minipage}[b]{0.42\linewidth}
\resizebox{\hsize}{!}{\includegraphics[width=\linewidth]{j1746-2849.freq_spectra.eps}}
\vspace{-3ex}
\end{minipage}
\begin{minipage}[b]{0.42\linewidth}
\resizebox{\hsize}{!}{\includegraphics[width=\linewidth]{j1745-2900.freq_spectra.eps}}
\vspace{-3ex}
\end{minipage}
\begin{minipage}[b]{0.42\linewidth}
\resizebox{\hsize}{!}{\includegraphics[width=\linewidth]{j1746-2856.freq_spectra.eps}}
\vspace{-3ex}
\end{minipage}
\begin{minipage}[b]{0.42\linewidth}
\resizebox{\hsize}{!}{\includegraphics[width=\linewidth]{j1745-2912.freq_spectra.eps}}
\vspace{-3ex}
\end{minipage}
\label{spectra.fig}
\end{figure*}
\begin{table*}
\centering
\caption{Properties of the pulsars that we detected: Coordinates (Equatorial and Galactic), DM, Stokes $I$ flux density of the brightest phase bin, polarized flux density and intrinsic polarization angle at 5.5 GHz ($\chi_0 = 0$\degr\ towards the west and increases counter-clockwise), flux spectral index $\alpha$, and RM.
For comparison, the Galactic coordinates of \sgr\ are gl, gb = -0.056\degr, -0.046\degr\ \protect\citep{reid1999}.
$\chi^2_4$ lists the value of the chi-square statistic for four degrees of freedom, while ``S/N'' indicates the signal/noise level of a 1D Gaussian random variable which produces the same probability.
$^\star$: DM values were taken from \citet[`J06']{johnston2006}, \citet[`D09']{deneva2009}, and \citet[`E13']{eatough2013}. $^\dagger$: Coordinates from \citet{shannon2013}.
}
\begin{tabular}{lcccc}
\hline
& J1746-2849 & J1745-2900 & J1746-2856 & J1745-2912 \\
\hline
$\rmn{RA}(2000)$ & $17^\rmn{h}46^\rmn{m}03\fs355(8)$ & $17^\rmn{h}45^\rmn{m}40\fs16^\dagger$ & $17^\rmn{h}46^\rmn{m}49\fs856(3)$ & $17^\rmn{h}45^\rmn{m}47\fs830(8)$ \\
$\rmn{Dec.}(2000)$ & $-28\degr50\arcmin13\farcs56(23)$ & $-29\degr00\arcmin29\farcs82^\dagger$ & $-28\degr56\arcmin59\farcs23(10)$ & $-29\degr12\arcmin30\farcs77(23)$ \\
$\rmn{gl}$ & 0.134\degr & -0.056\degr & 0.126\degr & -0.212\degr \\
$\rmn{gb}$ & -0.030\degr & -0.047\degr & -0.233\degr & -0.175\degr \\
DM~(cm$^{-3}~$pc)$^\star$ & 1456 (D09) & 1778 (E13) & 1168 (J06) & 1130 (J06) \\
$I$~(mJy) & 1.16(7) & 23.08(17) & 2.68(6) & 1.30(5)\\
$PI_{5.5}$~(mJy) & 0.48(6) & 5.92(17) & 0.54(4) & 0.22(3)\\
$\chi_{0,5.5}$~($^\circ$) & 46.3(35) & 42.5(8) & 37.3(22) & -84.3(39)\\
$\alpha$ & -0.48(116) & 1.34(28) & -2.18(74) & -3.58(140)\\
RM~(rad~m$^{-2}$) & 10,104(101) & -66,080(24) & 13,253(53) & -535(107)\\
$\chi^2_4$ (``S/N'') & 64 (7) & 779 ($>\,10$) & 166 ($>\,10$) & 54 (6)\\
\hline
\end{tabular}
\label{pulsars.tab}
\end{table*}
\section{Observations}\label{observations.sec}
The targets were observed with the Australia Telescope Compact Array (ATCA) in the 6A configuration between the 20th and the 22nd of April, and on the 6th of May 2015. These observations cover the frequency range between 4.473 and 6.525 GHz with 513 channels that are 4 MHz wide.
PSR J1745-2900 was observed for 2.6 hours in total, other targets were observed for 5.1 hours each.
Because of telescope maintenance all six antennas were only available on the final observing run. On the first three observing runs only 4, 5, and 4 antennas, respectively, were available.
We calibrated the data using the $\textsc{miriad}$ software package \citep{sault1995}: PKS 1934-638 was used to determine the bandpass and absolute flux scale, and PKS 1814-254 (9\degr\ from the targets) was observed approximately every 18 minutes to calibrate the complex gains and polarization leakages. We then transferred the calibration solutions to the target pulsars, and flagged radio frequency interference. PKS 1814-254 rises after the target pulsars; therefore the calibration solution from PKS 1814-254 was copied to the first observation of each target without extrapolation over time.
In our analysis we excluded the two shortest baselines.
Often the pulsar positions listed in the discovery papers had much larger errors than the size of the synthesized beam of our observations ($\approx 4\arcsec \times 1.5\arcsec$),
and we had to search maps of the primary beam to identify the pulsars. We removed a baseline level, calculated from off-pulse bins, before combining the real parts of the visibilities of the four observing runs using the noise variance in Stokes $V$ as weights. Noise-weighted pulse profiles are shown in fig.~\ref{profiles.fig}, and table~\ref{pulsars.tab} lists the pulsars that we detected with the ATCA, their updated coordinates, and the peak total intensity.
Positions were determined after summing the two brightest phase bins from the data collected in May 2015.
The pulsars listed in Table~\ref{pulsars.tab} are bright enough that we could detect them in our most sensitive observing run.
We were not able to detect PSR J1746-2850, even after concatenating visibilities - across all possible pulse phase bin combinations - from our two most sensitive observing runs\footnote{The ATCA depends on pulsar ephemerides to generate pulse profiles; it is not possible to record baseband or pulsar search data. Although we used the most recent available ephemerides, the non-detection of PSR J1746-2850 could also indicate that the ephemeris for this pulsar was no longer accurate enough to fold the data correctly.}, giving a six-sigma limit on the peak flux density of 0.49 mJy.
This non-detection implies that the flux density of this source has decreased by at least a factor of 30 since it was discovered by Deneva et al.
Radio magnetars are known to exhibit large variations in flux density (e.g., \citealt{lazaridis2008}); if PSR J1746-2850 shares characteristics with radio magnetars, as was proposed by Deneva et al., then this could explain our non-detection.
Monitoring observations might be able to detect this source in case it brightens again in the future.
We excluded PSR J1745-2910 from our analysis because of its poorly constrained position and pulse period.
We determined the polarization properties of the pulsars using a new maximum-likelihood-based method that we developed, which we describe in a forthcoming paper (Schnitzeler et al., in prep.). In our method the pulsar is characterized by its polarized flux density and intrinsic polarization angle at 5.5 GHz, its flux spectral index $\alpha$ ($S_\nu \propto \nu^\alpha$) and RM, plus, we allow the noise variances to be off by a scale factor $\eta$. Assuming that the differences between the measurements of Stokes $Q$ and $U$ and the model are described by Gaussian random variables and that the $N_\mathrm{ch}$ frequency channels are independent, we maximize the log likelihood over the parameter space. We search for RMs out to $\pm\, 4\times10^5$ rad~m$^{-2}$, the equivalent of the `maximum RM' from \cite{brentjens2005} in RM synthesis\footnote{Introducing $\delta\lambda^2$ for the channel width in units of wavelength squared, the `maximum RM' satisfies the equation RM$_\mathrm{max}\delta\lambda^2\approx \sqrt{3}$, even though RM synthesis assumes that for all frequency channels RM$\delta\lambda^2\ll1$ (both expressions are from \citealt{brentjens2005}). The difference between the exact formalism by \cite{schnitzeler2015} and the formalism by \cite{brentjens2005} is small for our observing setup and the RM range that we chose.}, and $\alpha$ between -6 and 2, which covers all pulsar spectral indices that were determined by \cite{Lorimer1995} and \cite{Bates2013}.
The difference between the maximum log likelihood and the log likelihood in the absence of any signal follows a chi square distribution \citep{wilks1938} with in our case four degrees of freedom; this enables us to quantify the detection probability of each pulsar. The 68\% confidence limits for the pulsar parameters are based on the region where the log likelihood has decreased by 0.5 (e.g., \citealt{avni1976}).
Fig.~\ref{spectra.fig} shows measured and modelled frequency spectra for the target pulsars, and our maximum likelihood estimates of the polarization properties of each pulsar are listed in table~\ref{pulsars.tab}. Fig.~\ref{map.fig} shows the positions of the pulsars relative to other features in the Galactic centre.
Scintillation of PSR J1745-2900 leads to a systematic and statistically significant variation between the amplitudes of the observed and modelled spectra across the frequency band.
To increase the signal strength in our maximum likelihood estimation we summed the polarization vectors of the brightest two phase bins in the pulse profile of PSR J1746-2856 and PSR J1746-2849, while for PSR J1745-2912 we summed the brightest four phase bins.
If the intrinsic polarization angle $\chi_0$ changes across the pulse profile this leads to depolarization if the polarization vectors of different phase bins are added. To mitigate this effect we added a phase gradient to the polarization vectors of the different phase bins before summing the bins. We varied the gradient between 0\degr/bin and 180\degr/bin, and used the gradient which produced the strongest polarized signal. The largest phase gradient is 29\degr/bin.
\begin{figure}
\caption{Known GC pulsars overplotted on an Effelsberg single-dish total intensity map of the GC at 10.55 GHz \protect\citep{seiradakis1989}. The positions of the Quintuplet and Arches clusters are indicated with pluses, and pulsars that we did not detect are highlighted in orange. On this scale PSR J1745-2900 and \sgr\ lie in the same direction.
}
\resizebox{\hsize}{!}{\rotatebox{-90}{\includegraphics[width=\linewidth]{pgplot_gcpsrsbw_orange.arxiv.ps}}}
\label{map.fig}
\end{figure}
\section{Discussion}\label{discussion.sec}
The large DMs suggest that all target pulsars lie within $\approx$ 150 pc of \sgr\ (based on the free electron density models NE2001 and NE2001thick, \citealt{cordes2002}, \citealt{schnitzeler2012}), within the Central Molecular Zone and the footprint of the Galactic Centre Lobe \citep{sofue1984}.
\cite{johnston2006} and \cite{deneva2009} noticed that the pulse broadening times of the pulsars in our sample are much shorter than predicted by these electron density models.
The absence of strong scattering does not preclude these pulsars from being in the GC: as shown by \cite{lazio1999}, \cite{bower2001}, and \cite{roy2013}, the observed scattering sizes of extragalactic sources within $\approx$ 1\degr\ of \sgr\ can be more than an order of magnitude smaller than predicted by NE2001.
PSRs J1746-2856 and J1746-2849 have the largest RMs measured for any object in the Milky Way beyond 0.1 pc of \sgr\ (\citealt{eatough2013}, \citealt{bower2003}, \citealt{marrone2007}); out of all known pulsars only PSR J1745-2900 has a larger RM.
All NTFs have smaller RMs, and the RMs of almost all extragalactic sources seen towards the GC are an order of magnitude smaller.
To calculate pulsar distances, NE2001 assumes a smoothly varying distribution for $n_\mathrm{e}$ which peaks at 10$\,$cm$^{-3}$.
However, there is evidence for large $n_\mathrm{e}$ variations in the GC.
Electron densities of several hundred cm$^{-3}$ have been measured with the \textit{KAO}, VLA, and the \textit{ISO} satellite (\citealt{colgan1996}, \citealt{lang2001}, \citealt{rodriguez-fernandez2001}, \citealt{cotera2005}).
Each sightline through the GC is thought to pass through at least one molecular cloud \citep{bally1988}; because molecular cloud cores\footnote{
\cite{lazio1999} argue that the ionized edges of molecular clouds could explain why scattering in the GC is patchy. They assumed that gas with a density $n\left(H_2\right) \gtrsim$ 500$\,$cm$^{-3}$ is highly ionized, however, these layers of dense gas are surrounded by more tenuous gas that could shield the inner regions from ionizing radiation (e.g., \citealt{hollenbach1999}).
The role that molecular clouds play in the scattering of background sources could be investigated with radiative transfer models that include chemical networks, e.g., $\textsc{CLOUDY}$ \citep{ferland2013}.
} are not significantly ionized, clouds leave holes in the $n_\mathrm{e}$ distribution with typical diameters of up to 20$\,$pc (FWHM; this value is based on the median diameter of molecular clouds in the CO survey by \citealt{oka2001}).
\cite{simpson2007} derived EMs along a strip at a Galactic longitude of $\approx$ 0.11\degr, and the $n_\mathrm{e}$ values they calculated vary by more than an order of magnitude.
Finally, the amount of diffuse radio emission at 10.55 GHz (fig.$\,$\ref{map.fig}) varies strongly with position, which could point to variations in the amount of free-free emission and therefore to variations in $n_\mathrm{e}$.
These results demonstrate that large $n_\mathrm{e}$ variations exist in the GC. Therefore, one may not rely on the difference between the DMs of two pulsars to estimate their relative LOS positions.
\cite{uchida1985} predicted that winding up of field lines due to differential rotation in the GC leads to an outflow perpendicular to the Galactic Plane, and produces a checkerboard pattern in RM centred on \sgr\ (fig.$\,$3 in \citealt{novak2003} and fig.$\,$14 in \citealt{law2011}).
As discussed by Novak et al. and Law et al.\footnote{Law et al. modelled the magnetized ISM in the GC using observations of polarized diffuse radio emission.
However, diffuse radio emission originates not only in the GC; furthermore, it can suffer from a range of depolarization effects, as discussed by, e.g., \cite{sokoloff1998}. Because of these concerns we prefer to compare our results to those presented by Uchida et al. and Novak et al.}, the signs of the RMs of NTFs agree with one of the two possible patterns.
However, the signs of the RMs of PSRs J1745-2912 and J1746-2856 contradict the particular solution selected by Uchida et al. and by Novak et al.
The sign of RM of PSR J1746-2849 agrees with the model prediction because this pulsar lies at a slightly higher Galactic latitude than \sgr\ \citep{reid1999}, but since this latitude difference is very small, the winding up of field lines along the LOS towards PSR J1746-2849 generates only a small RM.
A large foreground RM could shift the centre of the RM pattern and disrupt the checkerboard pattern.
However, the RMs of PSRs J1746-2856 and J1746-2849 are much larger in magnitude than the RMs of NTFs and extragalactic radio sources, which implies that the foreground contribution cannot explain the difference in sign between the RMs we observe and the RMs Uchida et al. predict.
The contribution by the Galactic foreground could dominate the RM of PSR J1745-2912, and because the RM of PSR J1745-2900 is dominated by the high $n_\mathrm{e}$ and strong magnetic field close to \sgr\ \citep{eatough2013}, we exclude these two pulsars from our analysis.
For the remaining two pulsars we calculate $\langle B_\| \rangle$, the mean LOS component of the magnetic field, weighted with the free electron density:
\begin{eqnarray}
\langle B_\| \rangle \equiv \frac{\int B_\| n_\mathrm{e}\mathrm{d}l}{\int n_\mathrm{e}\mathrm{d}l} = \frac{\mathrm{RM}}{0.81\, \mathrm{DM}}\, ,
\nonumber
\end{eqnarray}
using the definition of DM from equation~\ref{dm_definition}.
If $B_\|$ changes sign along the line of sight then locally $|B_\| |$ will be larger than $|\langle B_\| \rangle|$.
We subtract 670~cm$^{-3}$pc from the observed DMs to correct for the foreground contribution: this is the DM value predicted by NE2001 for the line of sight towards PSR J1745-2900, integrating from the Sun out to 150~pc from Sgr~A$^{\star,}$\footnote{We noticed that along this line of sight DMs do not increase monotonically with distance in NE2001.}.
The extragalactic sources that \cite{roy2008} analysed show large variations in RM, which in large part could be produced by the Galactic foreground.
Estimating the foreground RM from these observations is very difficult, and we did not correct the pulsar RMs for the foreground contribution.
By not correcting for the foreground RM, which is perhaps as high as 1500$\,$\radm, we overestimate $|\langle B_\| \rangle|$ by $\lesssim\, 15\%$.
For PSR J1746-2856 $\langle B_\| \rangle$ = 33$\,\mu$G, and for PSR J1746-2849 $\langle B_\| \rangle$ = 16$\,\mu$G, which are both much higher than the strength of the ordered and turbulent magnetic field in the vicinity of the Sun (1.5-2$\,\mu$G and 3-6$\,\mu$G, respectively: e.g., \citealt{haverkorn2015}).
Strong magnetic fields are not uncommon in the GC: in their analysis of the `Snake' NTF, \cite{gray1995} derived $\langle B_\| \rangle$ $\approx$ 7$\,\mu$G for the foreground ISM.
\cite{larosa2005} analysed maps of the diffuse synchrotron emission in the GC, and showed using minimum-energy arguments that the magnetic field strength $B\, \simeq\, 6\left(k/f\right)^{2/7}\mu$G, where $k$ is the ratio of energies of cosmic ray protons and electrons, and $f$ the filling factor of the synchrotron-emitting gas.
For reasonable values of $k$ and $f$, $B\simeq\, \left(6-80\right)\mu$G \citep{ferriere2011}.
Since the $\langle B_\| \rangle$ we derive are lower limits to $B$, our measurements rule out the weakest pervasive fields that are allowed in the analysis by LaRosa et al.
If a magnetic field with $B \sim 160\,\mu$G \citep{crocker2011} pervades the GC, including the warm ionized medium that is probed by our measurements, then the small $|\langle B_\| \rangle|$ we derive for PSR J1746-2856 implies an angle $\sim$ 12\degr\ between the magnetic field vector and the plane of the sky; towards PSR J1746-2849, which lies closer to the Galactic plane, this angle is even smaller.
If $B_\|$ changes sign along the line of sight then locally the angle between the magnetic field and the plane of the sky can be larger than the values we derived.
PSRs J1746-2849 and J1746-2856 are seen in the direction of the Radio Arc NTF (fig.$\,$\ref{map.fig}), which is known to have RMs between -5500 and -1660$\,$\radm\ (`Source A' at gl, gb = 0.16\degr, -0.13\degr; \citealt{sofue1987}) and between -200 and 1000$\,$\radm\ (`Source A$^\prime$\,' at gl, gb = 0.16\degr, -0.18\degr; \citealt{sofue1987}).
The `Plumes' are the northern and southern extension of the Radio Arc, and show RMs between -500 and 1000$\,$\radm\ (northern plume) and between -1500 and 0$\,$\radm\ (southern plume).
The RMs for the Radio Arc and the Plumes have been published by \cite{inoue1984}, \cite{sofue1987}, \cite{tsuboi1986}, \cite{yusef-zadeh1986a}, and \cite{yusef-zadeh1987}.
The large RM variations that we measured for the GC pulsars and between these pulsars and the different parts of the Radio Arc could be produced in the turbulent GC.
Variations in RM by several thousand \radm\ on scales of arcminutes have also been observed towards NTFs.(\citealt{gray1995}, \citealt{yusef-zadeh1997}, \citealt{lang1999a}, \citealt{lang1999b}).
NTFs like the `Pelican' and the `Ripple' are known to have external Faraday screens, which implies that the observed variations in RM occur outside the NTF, in the ISM of the GC.
Only the GC shows such large RM fluctuations, adding support that our pulsars are in the GC.
This also strengthens the case for PSR J1745-2912 lying in the GC, because its DM is very similar to the DM of PSR J1746-2856.
If PSRs J1746-2856 and J1746-2849 lie along low-scattering `corridors' through the GC this again raises the question why not more pulsars have been detected in the GC.
To answer this question requires a more detailed study of the GC environment and its pulsar population, taking into account the various instrumental effects and selection biases (building on, e.g., \citealt{chennamangalam2014}).
We cannot exclude the possibility that a foreground object in the GC adds a large RM ($\sim$ 10$^4\,$\radm) to PSRs J1746-2856 and J1746-2849.
A helical magnetic field around the Radio Arc could be the origin of this RM, and might produce also the helical features in Stokes $I$ around the Radio Arc that were identified by \cite{yusef-zadeh1987}.
The molecular cloud G0.11-0.11 lies at gl, gb = 0.108\degr, -0.108\degr\ \citep{tsuboi1997}, in between PSRs J1746-2856 and J1746-2849.
Ablated gas from this cloud which is subsequently ionized might be an alternative explanation for the large RMs of these pulsars.
However, the inner part of G0.11-0.11 that emits in the CS $J=1\rightarrow0$ line (which traces gas with densities $n\left(H_2\right) \gtrsim$ 10$^4\,$cm$^{-3}$; \citealt{bally1988}) has a diameter of only 7 pc \citep{tsuboi1997}, while the pulsars are separated by about 29~pc on the sky if they lie in the GC.
Even though a transition region exists between the dense gas in the inner parts of G0.11-0.11 and the low-density environment of the cloud, the ionized halo surrounding G0.11-0.11 would have to be very large for it to contribute to the RMs of both pulsars.
Both scenarios require that the pulsars lie behind the Radio Arc or G0.11-0.11, which places the pulsars at the distance of \sgr\ (\citealt{ponti2010}, \citealt{roy2013}).
f correct, this provides further evidence that long ($>$~100~pc), low-scattering corridors exist in the GC.
\section{Summary}\label{summary.sec}
We studied the properties of the magnetic field in the GC by measuring the amount of Faraday rotation in the signals of four pulsars in the GC.
We observed these pulsars with the Australia Telescope Compact Array at frequencies between 4.5-6.5 GHz.
Our measurements provide more accurate positions for three of the previously known pulsars in the GC.
The non-detection of PSR J1746-2850 implies that the flux density of this pulsar has decreased by at least a factor of thirty since its discovery by \cite{deneva2009}.
To include both the flux spectrum of each pulsar and the variation in sensitivity across the observing band we used a new maximum-likelihood based method that we will describe in an upcoming paper.
PSRs J1746-2856 and J1746-2849 have RMs$\,>\,10^4\,$rad~m$^{-2}$, the second and third largest RMs measured for any pulsar after PSR J1745-2900, and the largest RM measured for any Galactic object beyond $\sim\,0.1\,$pc of \sgr.
Their DMs, RMs, and RM variations place these pulsars within $\approx\,150\,$pc of \sgr, within the Central Molecular Zone and the footprint of the Galactic Centre Lobe.
Based on evidence from the literature we could not pinpoint the pulsar positions more accurately based on only their DMs; the smooth structure in the free electron density that is assumed in NE2001 is an oversimplification.
Differential rotation in the GC, leading to a winding-up of field lines, was predicted to produce a checkerboard pattern in the sign of RM \citep{uchida1985}; the RMs of NTFs agreed with one of the two possible solutions for this pattern (e.g., \citealt{novak2003}).
However, the signs of the RMs of PSRs J1746-2856 and J1746-2849 contradict this particular solution.
By combining the observed DMs and RMs, and correcting for the foreground contribution to DM, we derive LOS lengths of the magnetic field vector between $\sim\,16-33\,\mu$G, which is much larger than the strength of the large-scale and small-scale magnetic field in the vicinity of the Sun \citep{haverkorn2015}.
If the GC, including the warm ionized medium that is probed by our observations, is pervaded by a strong magnetic field, $B\,\sim\,160\,\mu$G \citep{crocker2011}, then our measurements imply that the magnetic field makes an angle $\lesssim\,12$\degr\ with the plane of the sky.
If the direction of the LOS magnetic field component changes sign along the line of sight, this inclination angle will be larger.
Large changes in RM on scales of arcminutes have been observed previously towards NTFs; since the GC is the only location where such large RM variations are known to exist, this provides additional evidence that the target pulsars must lie inside the GC.
The large RMs of PSRs J1746-2856 and J1746-2849 could be produced by foreground objects such as the Radio Arc or an ablated, ionized halo that surrounds the molecular cloud G0.11-0.11.
This implies that these two pulsars must lie close to or even beyond \sgr, more than 100 pc into the GC, which would be additional evidence for the existence of long, low-scattering corridors through the GC.
If this is indeed the case, then future, sensitive observations should be able to detect more pulsars in the GC (e.g., \citealt{chennamangalam2014}, \citealt{eatough2015}).
\section*{Acknowledgements}
We would like to thank the staff at CSIRO Astronomy and Space Science, and in particular Robin Wark, Jamie Stevens, and Phil Edwards, for their support of this project.
Fig.$\,$\ref{map.fig} was prepared using a script provided by Bernd Klein (Max Planck Institute for Radio Astronomy).
The Australia Telescope Compact Array is part of the Australia Telescope National Facility which is funded by the Commonwealth of Australia for operation as a National Facility managed by CSIRO.
|
\section{Introduction}
Over the last decade, several aspects of the human brain functioning have been extensively studied using functional Magnetic Resonance Imaging (fMRI). fMRI has become a very important imaging modality in diagnosing, treating and monitoring several brain disorders such as Schizophrenia \cite{schizo}, Alzheimer's disease \cite{AD} etc. The utility of fMR imaging is mainly because the activity of the brain can be studied at sub-second temporal resolution, and sub-mm spatial resolution, allowing detailed studies. Among the several functional aspects that are studied using fMRI, are cortical mapping of cognitive tasks, networks in cognitive tasks, functional disorders, effects of medication therapies, improvement or deterioration in an affected brain. Hence fMRI is used by various groups for studies of neuroscience and psychiatry.
fMR imaging does not use any external contrast agent, instead it exploits the changes in blood oxygen levels that occur in the brain at regions of neural activity. This change in oxygenation, termed Blood Oxygen Level Dependent (BOLD) signal, indicates the level of neural activity in the brain volume, captured at points (voxels) along a 3D grid. In fMR imaging, the subject is continously scanned over a well-designed period of time, during which the subject follows instructions of either executing a well-designed task or resting. Hence images of the entire brain volume are generated, all along the time-course of the experiment. Thus, if the scanner acquires data every second, and the experiment runs for a course of 100 seconds, one would have 100 volumes of brain data.
The initial works on fMRI analysis focussed on task-specific cortical mapping \cite{friston}, \cite{friston1}, using the Generalized Linear Model approach, where all voxels were processed independently. However, later works involved studying cognitive states, where the entire brain volume was considered as a single pattern. In their pioneering work on decoding cognitive states \cite{decode}, the authors have elaborated the challenges that arise in this approach. The fraction of the number of relevant voxels is typically very small as compared to the entire number of voxels leading to the challenge of feature selection. Besides the number of available samples for a given cognitive state is far less than the dimension of the feature vectors. Hence the search for the optimal feature set, which could be used with a best-suited classifier has been the focus of several works \cite{Pereira,Ryali}.
\section{Related Work}
Classification experiments on six cognitive tasks, performed by five subjects, using activation maps have been reported in the literature \cite{lee}. The activation maps were used to derive feature vectors, from regions that were marked as consistently and exclusively activated for a given task, during training. The Support Vector Machine (SVM) classifier was used for classification. The authors have reported an average classification accuracy of 74.5\% across the five subjects performing six cognitive tasks. However, the drawback of the proposed method is that it requires a training process where experts identify task-specific neural regions to extract feature vectors. In \cite{yong} the authors have addressed the issue of building classifiers that detect a particular cognitive state across different subjects using fMR images. This study also uses training data to extract the relevant brain regions. This is followed by utilizing statistical information of the brain regions to form features. The classification performance of this method is validated in a deception fMRI study, using SVM classifier.
The challenges in using the fMR image sequence for classifying the cognitive tasks is that the dimension of the data is several magnitudes larger than the available samples for training. Hence it is important to choose relevant features or transform them to a domain where classification becomes more efficient. Most of the progress in fMRI analysis has been to determine the optimal combination of features and classifiers.
In this paper, we address the issue of transforming the feature set, and compare the classification performance of the transformed feature set as against using the raw data. The proposed work explores if classification of time series corresponding to different cognitive tasks could be improved. In this work attributes of the time series, such as periodicity, harmonics, and phase are exploited. The study in this paper aims to classify between two cognitive tasks using a combination of feature extraction and iterative classification. The block diagram in Fig. \ref{fig:blk_fig} shows the proposed method. Given a time series of fMR images that correspond to the subject performing certain well-designed cognitive tasks, an expert marks out voxels that are relevant for the classification. The time-series of voxel intensities (here, called raw data) along these regions are separated. The raw data is transformed using transformations such as Fourier transform and Hilbert transform \cite{king} to obtain the complex-valued data after applying random sieve function, whose phase information is processed and utilized as features. The transformed data is given as input to Naive Bayesian (NB) and SVM classifiers, to obtain the final classification.
\begin{figure} [!h]
\begin{center}
\includegraphics[height=2in,width=3in,angle=0]{blkdia.png}
\caption{\small \sl Block diagram of the proposed method \label{fig:blk_fig}}
\end{center}
\end{figure}
\section{Proposed method}
The proposed method is based on the observation that the fMRI data results in voxel intensities that exhibit a quasi-periodic pattern. The assumption of quasi-periodicity is made since the same set of cognitive tasks are repeatedly performed. The phase information, in the spectral analysis, holds key aspects of the features used for classification tasks. Identification of discriminating features from the regions of interest is an important aspect. We have proposed a transformation of fMRI data with random sieve function followed by phase extraction using either Fourier or Hilbert transforms. The utility of classifying data in the transformed domain is motivated by generalizations of analytic signal theory to higher dimensional problems \cite{Bernstein2013}.
Let $X(j)$ be $m$ integers sampled uniformly from the set $\{1,2,\ldots,N\}$. $\gamma(n)$, where $n\in\{1,2,3,\ldots,N\}$ is defined as follows,
\begin{equation}
\gamma(n) =
\begin{dcases}
0, & \exists j \in \{1,2,\ldots,m\} \ni n=X(j)\\
1, & \text{otherwise}
\end{dcases}
\end{equation}
The random sieve function $S_\gamma$ is defined as a mapping from raw voxel data,$f(n)$, to $g(n)$ and $.$ represents coordinate-wise multiplication.:
\begin{equation}
S_\gamma: f(n) \to g(n)
\end{equation}
\begin{equation}
g(n) = f(n) . \gamma(n)
\end{equation}
The Discrete Fourier Transform (DFT) of a sequence $g(n)$ is given by $G(k)$, where $k$ varies from $1$ to $N$,
\begin{equation}
G(k) = \sum_{n=1}^{n = N}{g(n)}{e^{-i2{\pi}k(n-1)/N}}
\end{equation}
This transform results in the complex-valued spectral decomposition. Spectral magnitude and phase are the two important features in the
Fourier representation of the signal. Under certain conditions, the given signal can be reconstructed using only the phase information in the Fourier representation. The proposed work utilizes the phase information for classification of the cognitive tasks. If the phase provides an important information about the features, a shift in phase may enhance or corrupt the discriminative nature of the given features. This leads to Hilbert transform which acts as an asymmetric phase shifting operator. The Discrete Hilbert Transform (DHT) has many forms of representation. It is defined using Inverse Discrete Time Fourier Transform (IDTFT),
\begin{equation}
H(k) = IDTFT\{G(\omega) \sigma_H(\omega)\}
\end{equation}
\begin{equation}
\sigma_H(\omega) =
\begin{dcases}
e^{i{\pi}/2}, & \text{if } -\pi < \omega < 0\\
e^{-i{\pi}/2}, & \text{if } 0 < \omega < \pi\\
0, & \text{otherwise}
\end{dcases}
\end{equation}
This transform also results in complex values, where the imaginary part of the transformed data captures the phase information. The raw voxel data has been transformed by random sieve function before mapped to Fourier and Hilbert domains. The phase information has been captured from the transformed data, in both the cases, and used for classification. The phase information is utilized to classify between the different time series that correspond to distinct cognitive tasks and not used to recover the signal.
Naive Bayes and Support Vector Machine classifiers are used for cognitive task classification in this work. The improvement in classification accuracy is clearly due to the use of transformed data. The sequence of voxel data has the spatial order mix-up as spatially-neighbour voxels may not be sequential neighbours. This mix-up and re-ordering of neighbourhood while generating voxel features suggest the existence of asymmetrical phase shifting which is captured by the use of Hilbert transform. The mix-up is important as it might enhance the classification by creating new features by combining different spatial voxels. The idea behind the proposed method is classification of cognition tasks in the transformed domain using phase information instead of utilizing the raw fMRI data.
The purpose of random sieve function is to reduce the number of data points considered for a good classification performance. Under the assumption that sufficient information is contained in lesser number of voxels, random sieve function is used to reduce the number of voxels in the input data, without creating any structural bias in the selection. The following eight classifier configurations have been compared to bring out the impact of transformation using random sieve function and spectral analysis. $arg(.)$ represents the angle of a given complex number.
\begin{equation}
f(n) \xrightarrow {\text{NB}} C_1
\end{equation}
\begin{equation}
f(n) \xrightarrow {\text{SVM}} C_2
\end{equation}
\begin{equation}
f(n) \xrightarrow {\text{$S_\gamma$}} g(n) \xrightarrow {\text{NB}} C_3
\end{equation}
\begin{equation}
f(n) \xrightarrow {\text{$S_\gamma$}} g(n) \xrightarrow {\text{SVM}} C_4
\end{equation}
\begin{equation}
f(n) \xrightarrow {\text{$S_\gamma$}} g(n) \xrightarrow {\text{arg(DFT)}} arg(G(k)) \xrightarrow {\text{NB}} C_5
\end{equation}
\begin{equation}
f(n) \xrightarrow {\text{$S_\gamma$}} g(n) \xrightarrow {\text{arg(DFT)}} arg(G(k)) \xrightarrow {\text{SVM}} C_6
\end{equation}
\begin{equation}
f(n) \xrightarrow {\text{$S_\gamma$}} g(n) \xrightarrow {\text{arg(DHT)}} arg(H(k)) \xrightarrow {\text{NB}} C_7
\end{equation}
\begin{equation}
f(n) \xrightarrow {\text{$S_\gamma$}} g(n) \xrightarrow {\text{arg(DHT)}} arg(H(k)) \xrightarrow {\text{SVM}} C_8
\end{equation}
The dataset and results are elaborated in the next section.
\section{Data and Results}
\subsection{Star Plus data}
The dataset, called "StarPlus data", was downloaded from the website \cite{star}. The experiment consisted of 54 trials. In one trial of the experiment, the subject was sequentially shown a picture and a sentence, and was told to press a button to determine if the picture correctly matched the sentence. The pictures were geometric arrangement of symbols such as $*$, $+$ and/or $\$$. The sentences were descriptions of the shown picture such as: “It is true that the star is above the plus”. The picture was presented first on the first half of the trials, while on the other half, the sentence was shown first. Snapshots of the brain were collected every 0.5 seconds. The data was marked with 25-30 anatomically defined regions referred to as Regions of Activation (ROA). The data consists of a set of trials. Each trial consists of an average of 5000 voxels for one snapshot totalling to about 270000 voxels for a particular trial over the entire span of time in the experiment. For a particular subject, 40 trials of data are collected for each cognition task. Hence, there are 80 trials of data for each subject. This data is stored for 6 subjects in the experiment for analysis.
\begin{enumerate}
\item{The first stimulus (S or P) was presented at the beginning of the trial.}
\item{The stimulus was removed after 4 seconds, replaced by a blank screen}
\item{The second stimulus was presented after 4 seconds. This was presented for 4 seconds, or until the subject pressed the button, whichever came first.}
\item{A rest period of 15 seconds was allowed after the second stimulus was removed from the screen.}
\end{enumerate}
The available StarPlus data has already been preprocessed to remove artefacts due to head motion, signal drift, and other sources. Experts have marked the relevant anatomical regions in the brain that participate in the mentioned cognitive tasks. The voxels only from the ear-marked areas ('CALC', 'LIPL', 'LT', 'LTRIA', 'LOPER', 'LIPS', and 'LDLPFC' - as given in the web-site \cite{star}) are chosen as features for classification.
\begin{table}[!h]
\caption{Average Classifier Performance}
\centering
\begin{tabular}{c c}
\hline\hline
Classifier Configuration & Correct Classification in \% \\ [0.5ex]
\hline
$C_1$ & 65.6 \\
$C_2$ & 76.4 \\
$C_3$ & 60.8 \\
$C_4$ & 69.1 \\
$C_5$ & 96.8 \\
$C_6$ & 97.5 \\
$C_7$ & 93.7 \\
$C_8$ & 99.0 \\ [1ex]
\hline
\end{tabular}
\label{table:result_avg1}
\end{table}
All simulations have been carried out in MATLAB. The reported trials are limited to single-subject classification. Only activated voxels are considered to reduce the dimension of the input data. The time series data corresponding to each of the cognitive tasks is chosen to create the relevant data sets. For each of the data sets, 40 samples for each of the cognitive tasks is chosen. Leave-one-out cross validation was performed, for example, 79 out of 80 samples constitute the training set, whereas the remaining sample served as the test input. We have tried many values and found $N=14000$ gives the improved performance while defining the random sieve function. The determination of $N$ may be based on cross validation which needs to be investigated. The tabulations in Table \ref{table:result_avg1} compare the classification performance for eight classifier configurations given in the previous section, averaged over all six subjects. Since the Random Sieve Function (RSF) is involved, the experiments have been repeated 50 times and average values are given in Table \ref{table:result_avg1}. The standard deviation lies between 1 and 2 for these cases.
\begin{table}[!h]
\caption{Confusion matrix (subject 5) - Raw data (NB) - $C_1$}
\centering
\begin{tabular}{c c c}
\hline\hline
Class labels & Class-1 & Class-2 \\ [0.5ex]
\hline
Class-1 & 24 & 16 \\
Class-2 & 12 & 28 \\ [1ex]
\hline
\end{tabular}
\label{table:cms5c1}
\end{table}
\begin{table}[!h]
\caption{Confusion matrix (subject 5) - Raw data (SVM) - $C_2$}
\centering
\begin{tabular}{c c c}
\hline\hline
Class labels & Class-1 & Class-2 \\ [0.5ex]
\hline
Class-1 & 29 & 11 \\
Class-2 & 9 & 31 \\ [1ex]
\hline
\end{tabular}
\label{table:cms5c2}
\end{table}
\begin{table}[!h]
\caption{Confusion matrix (subject 5) - Raw data + RSF (NB) - $C_3$}
\centering
\begin{tabular}{c c c}
\hline\hline
Class labels & Class-1 & Class-2 \\ [0.5ex]
\hline
Class-1 & 20 & 20 \\
Class-2 & 10 & 30 \\ [1ex]
\hline
\end{tabular}
\label{table:cms5c3}
\end{table}
\begin{table}[!h]
\caption{Confusion matrix (subject 5) - Raw data + RSF (SVM) - $C_4$}
\centering
\begin{tabular}{c c c}
\hline\hline
Class labels & Class-1 & Class-2 \\ [0.5ex]
\hline
Class-1 & 28 & 12 \\
Class-2 & 9 & 31 \\ [1ex]
\hline
\end{tabular}
\label{table:cms5c4}
\end{table}
\begin{table}[!h]
\caption{Confusion matrix (subject 5) - RSF+DFT (NB) - $C_5$}
\centering
\begin{tabular}{c c c}
\hline\hline
Class labels & Class-1 & Class-2 \\ [0.5ex]
\hline
Class-1 & 38 & 2 \\
Class-2 & 2 & 38 \\ [1ex]
\hline
\end{tabular}
\label{table:cms5c5}
\end{table}
\begin{table}[!h]
\caption{Confusion matrix (subject 5) - RSF+DFT (SVM) - $C_6$}
\centering
\begin{tabular}{c c c}
\hline\hline
Class labels & Class-1 & Class-2 \\ [0.5ex]
\hline
Class-1 & 39 & 1 \\
Class-2 & 1 & 39 \\ [1ex]
\hline
\end{tabular}
\label{table:cms5c6}
\end{table}
\begin{table}[!h]
\caption{Confusion matrix (subject 5) - RSF+DHT (NB) - $C_7$}
\centering
\begin{tabular}{c c c}
\hline\hline
Class labels & Class-1 & Class-2 \\ [0.5ex]
\hline
Class-1 & 35 & 5 \\
Class-2 & 5 & 35 \\ [1ex]
\hline
\end{tabular}
\label{table:cms5c7}
\end{table}
\begin{table}[!h]
\caption{Confusion matrix (subject 5) - RSF+DHT (SVM) - $C_8$}
\centering
\begin{tabular}{c c c}
\hline\hline
Class labels & Class-1 & Class-2 \\ [0.5ex]
\hline
Class-1 & 40 & 0 \\
Class-2 & 0 & 40 \\ [1ex]
\hline
\end{tabular}
\label{table:cms5c8}
\end{table}
\section{Discussion}
Although Fourier and Hilbert Transforms have been applied in several domains \cite{hahn} \cite{alaif} \cite{feldman}, their utility in processing fMRI data for the purpose of classifying between cognitive states, along with random sieve function, is a novel contribution. The applications of these transforms stem from the observation that the cognitive tasks elicit certain periodic patterns. This inherent periodicity in the voxel values can be exploited to distinguish between cognitive tasks. The voxel features are taken as a sequence for classification tasks. Naive Bayesian classifier is a simple classifier for data that is assumed independent. This classifier is proven to be effective in spite of the restrictive assumption of independence, in several applications. The idea behind this classifier is that the posterior is proportional to the prior and that the proportionality factor varies directly with the likelihood. The focus of this paper is to bring out the effect of random sieve function and spectral transformation on the performance of classifiers. The performance of the proposed approach for other transform techniques has to be investigated in the future. The work is concerned with the use of raw data and the transformed data for classification and not the comparison of classification schemes. Classification based on other techniques such as regression coefficients, summary statistic, etc. could also be applied on transformed data.
It is evident, from Table \ref{table:result_avg1}, that the classification accuracy has been improved for the proposed transformation. The average classification accuracy using the raw data, on a Naive Bayesian classifier, stands at 65.6\%. On the other hand, using the transformed data, the classification accuracy obtained on the same Naive Bayesian classifier, is 96.8\% and 93.7\%, respectively for Fourier and Hilbert transformations along with random sieve function. The average classification accuracy using the raw data, on SVM classifier (hard margin), stands at 76.4\%. Outliers in the data would influence the classification boundary in hard-margin SVM. On the other hand, using the transformed data, the classification accuracy obtained on the same SVM classifier, is 97.5\% and 99.0\%, respectively for Fourier and Hilbert transformations along with random sieve function.
The tabulations in Table \ref{table:cms5c1}, \ref{table:cms5c2}, \ref{table:cms5c3}, \ref{table:cms5c4}, \ref{table:cms5c5}, \ref{table:cms5c6}, \ref{table:cms5c7}, and \ref{table:cms5c8}, show the typical confusion matrix obtained for each of the 8 classifier configuration, for Subject-5 data.
The sequence generated from 3D spatial structure implies that there is a mix-up of order when we arrange them as 1D vector. The change in neighbourhood information due to this re-order leads to two cases : adjacent occurrences of voxels do not imply spatial neighbourhood and spatial neighbours may not be closer. The asymmetrical re-ordering of voxel features is an important factor which has not normally been taken into account for classification tasks. By using the transformed data, the proposed approach utilizes this asymmetrical phase-shift explicitly and this leads to improved classification accuracy. The phase components in Fourier domain and the imaginary components of Hilbert transform are given to the classifier. From the classifier point of view, the difference between using the raw data and transformed data is coming from the explicit use of phase information along with random sieve function. The above discussion also opens up the scope of the proposed approach as phase analysis is widely used in communication systems. The future investigations would involve the evaluation of phase algorithms, multivariate methods, other transform techniques, N-fold cross validation, and sieve methods, for fMRI data classification and their theoretical implications.
\section{Conclusion}
In this paper, we have proposed a random sieve function along with the utility of transforms on voxel intensities, for classifying cognitive states. The two cognitive tasks considered here are, "sentence viewing" and "picture viewing". Real datasets, StarPlus data, comprising of fMR data obtained from 6 healthy volunteers are used. The experiment consists of transforming the data using random sieve function, Fourier Transform and Hilbert Transform, prior to using Naive Bayesian and SVM classifiers. The results obtained show an improvement of about 30\% while using Fourier transformed data and about 23\% for Hilbert transformed data, as against using raw voxel intensities.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.